Intel Architecture

Software Developer’s
Manual
Volume 1:
Basic Architecture

NOTE: The Intel Architecture Software Developer’s Manual consists of

three volumes: Basic Architecture, Order Number 243190; Instruction Set
Reference, Order Number 243191; and the System Programming Guide,
Order Number 243192.
Please refer to all three volumes when evaluating your design needs.

1999
Information in this document is provided in connection with Intel products. No license, express or implied, by estoppel
or otherwise, to any intellectual property rights is granted by this document. Except as provided in Intel’s Terms and
Conditions of Sale for such products, Intel assumes no liability whatsoever, and Intel disclaims any express or implied
warranty, relating to sale and/or use of Intel products including liability or warranties relating to fitness for a particular
purpose, merchantability, or infringement of any patent, copyright or other intellectual property right. Intel products are
not intended for use in medical, life saving, or life sustaining applications.

Intel may make changes to specifications and product descriptions at any time, without notice.

Designers must not rely on the absence or characteristics of any features or instructions marked “reserved” or
“undefined.” Intel reserves these for future definition and shall have no responsibility whatsoever for conflicts or
incompatibilities arising from future changes to them.

Intel’s Intel Architecture processors (e.g., Pentium®, Pentium® II, Pentium® III, and Pentium® Pro processors) may
contain design defects or errors known as errata which may cause the product to deviate from published
specifications. Current characterized errata are available on request.

Contact your local Intel sales office or your distributor to obtain the latest specifications and before placing your
product order.

Copies of documents which have an ordering number and are referenced in this document, or other Intel literature,
may be obtained by calling 1-800-548-4725, or by visiting Intel's literature center at http://www.intel.com.

COPYRIGHT © INTEL CORPORATION 1999

*THIRD-PARTY BRANDS AND NAMES ARE THE PROPERTY OF THEIR RESPECTIVE OWNERS.
 TABLE OF CONTENTS

CHAPTER 1
ABOUT THIS MANUAL
1.1. OVERVIEW OF THE INTEL ARCHITECTURE SOFTWARE DEVELOPER’S MANUAL,
VOLUME 1: BASIC ARCHITECTURE 1-1
1.2. OVERVIEW OF THE INTEL ARCHITECTURE SOFTWARE DEVELOPER’S MANUAL,
VOLUME 2: INSTRUCTION SET REFERENCE 1-3
1.3. OVERVIEW OF THE INTEL ARCHITECTURE SOFTWARE DEVELOPER’S MANUAL,
VOLUME 3: SYSTEM PROGRAMMING GUIDE 1-3
1.4. NOTATIONAL CONVENTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
1.4.1. Bit and Byte Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-5
1.4.2. Reserved Bits and Software Compatibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-6
1.4.3. Instruction Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-7
1.4.4. Hexadecimal and Binary Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-7
1.4.5. Segmented Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-7
1.4.6. Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-8
1.5. RELATED LITERATURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-9

CHAPTER 2
INTRODUCTION TO THE INTEL ARCHITECTURE
2.1. BRIEF HISTORY OF THE INTEL ARCHITECTURE . . . . . . . . . . . . . . . . . . . . . . . . 2-1
2.2. INCREASING INTEL ARCHITECTURE PERFORMANCE AND MOORE’S LAW . 2-4
2.3. BRIEF HISTORY OF THE INTEL ARCHITECTURE FLOATING-POINT UNIT. . . . 2-6
2.4. INTRODUCTION TO THE P6 FAMILY PROCESSOR’S
ADVANCED MICROARCHITECTURE 2-6
2.5. DETAILED DESCRIPTION OF THE P6 FAMILY PROCESSOR
MICROARCHITECTURE 2-9
2.5.1. Memory Subsystem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-9
2.5.2. Fetch/Decode Unit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-11
2.5.3. Instruction Pool (Reorder Buffer). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-11
2.5.4. Dispatch/Execute Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-12
2.5.5. Retirement Unit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-13

CHAPTER 3
BASIC EXECUTION ENVIRONMENT
3.1. MODES OF OPERATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-1
3.2. OVERVIEW OF THE BASIC EXECUTION ENVIRONMENT . . . . . . . . . . . . . . . . . 3-2
3.3. MEMORY ORGANIZATION. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
3.4. MODES OF OPERATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
3.5. 32-BIT VS. 16-BIT ADDRESS AND OPERAND SIZES. . . . . . . . . . . . . . . . . . . . . . 3-4
3.6. REGISTERS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-5
3.6.1. General-Purpose Data Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-6
3.6.2. Segment Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-7
3.6.3. EFLAGS Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-10
3.6.3.1. Status Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-12
3.6.3.2. DF Flag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-13
3.6.4. System Flags and IOPL Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-13
3.7. INSTRUCTION POINTER . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-14
3.8. OPERAND-SIZE AND ADDRESS-SIZE ATTRIBUTES. . . . . . . . . . . . . . . . . . . . . 3-14

iii
TABLE OF CONTENTS

CHAPTER 4
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS
4.1. PROCEDURE CALL TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
4.2. STACK . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
4.2.1. Setting Up a Stack. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-3
4.2.2. Stack Alignment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-3
4.2.3. Address-Size Attributes for Stack Accesses . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-3
4.2.4. Procedure Linking Information. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-4
4.2.4.1. Stack-Frame Base Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-4
4.2.4.2. Return Instruction Pointer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-4
4.3. CALLING PROCEDURES USING CALL AND RET . . . . . . . . . . . . . . . . . . . . . . . . 4-5
4.3.1. Near CALL and RET Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-5
4.3.2. Far CALL and RET Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-6
4.3.3. Parameter Passing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-7
4.3.3.1. Passing Parameters Through the General-Purpose Registers . . . . . . . . . . . .4-7
4.3.3.2. Passing Parameters on the Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-7
4.3.3.3. Passing Parameters in an Argument List . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-7
4.3.4. Saving Procedure State Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-7
4.3.5. Calls to Other Privilege Levels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-8
4.3.6. CALL and RET Operation Between Privilege Levels . . . . . . . . . . . . . . . . . . . . .4-10
4.4. INTERRUPTS AND EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-11
4.4.1. Call and Return Operation for Interrupt or Exception Handling Procedures . . . .4-13
4.4.2. Calls to Interrupt or Exception Handler Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . .4-17
4.4.3. Interrupt and Exception Handling in Real-Address Mode . . . . . . . . . . . . . . . . . .4-17
4.4.4. INT n, INTO, INT 3, and BOUND Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . .4-17
4.5. PROCEDURE CALLS FOR BLOCK-STRUCTURED LANGUAGES. . . . . . . . . . . 4-18
4.5.1. ENTER Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-18
4.5.2. LEAVE Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-24

CHAPTER 5
DATA TYPES AND ADDRESSING MODES
5.1. FUNDAMENTAL DATA TYPES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
5.1.1. Alignment of Words, Doublewords, and Quadwords. . . . . . . . . . . . . . . . . . . . . . .5-2
5.2. NUMERIC, POINTER, BIT FIELD, AND STRING DATA TYPES . . . . . . . . . . . . . . 5-3
5.2.1. Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-3
5.2.2. Unsigned Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5
5.2.3. BCD Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5
5.2.4. Pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5
5.2.5. Bit Fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5
5.2.6. Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5
5.2.7. Floating-Point Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-6
5.2.8. MMX™ Technology Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-6
5.2.9. Streaming SIMD Extensions Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-6
5.3. OPERAND ADDRESSING. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-6
5.3.1. Immediate Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-6
5.3.2. Register Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-7
5.3.3. Memory Operands. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-7
5.3.3.1. Specifying a Segment Selector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-8
5.3.3.2. Specifying an Offset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-9
5.3.3.3. Assembler and Compiler Addressing Modes . . . . . . . . . . . . . . . . . . . . . . . . .5-10
5.3.4. I/O Port Addressing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-11

iv
 TABLE OF CONTENTS

CHAPTER 6
INSTRUCTION SET SUMMARY
6.1. NEW INTEL ARCHITECTURE INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-1
6.1.1. New Instructions Introduced with the Streaming SIMD Extensions . . . . . . . . . . . 6-1
6.1.2. New Instructions Introduced with the MMX™ Technology . . . . . . . . . . . . . . . . . 6-1
6.1.3. New Instructions in the Pentium® Pro Processor . . . . . . . . . . . . . . . . . . . . . . . . 6-2
6.1.4. New Instructions in the Pentium® Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
6.1.5. New Instructions in the Intel486™ Processor . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6.2. INSTRUCTION SET LIST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6.2.1. Integer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6.2.1.1. Data Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-3
6.2.1.2. Binary Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.2.1.3. Decimal Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.2.1.4. Logic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.2.1.5. Shift and Rotate Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-5
6.2.1.6. Bit and Byte Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-6
6.2.1.7. Control Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-7
6.2.1.8. String Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-8
6.2.1.9. Flag Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-9
6.2.1.10. Segment Register Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-9
6.2.1.11. Miscellaneous Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-9
6.2.2. MMX™ Technology Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.2.2.1. MMX™ Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.2.2.2. MMX™ Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.2.2.3. MMX™ Packed Arithmetic Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-10
6.2.2.4. MMX™ Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-11
6.2.2.5. MMX™ Logic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-11
6.2.2.6. MMX™ Shift and Rotate Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-11
6.2.2.7. MMX™ State Management. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-12
6.2.3. Floating-Point Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-12
6.2.3.1. Data Transfer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-12
6.2.3.2. Basic Arithmetic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-13
6.2.3.3. Comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-14
6.2.3.4. Transcendental . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-14
6.2.3.5. Load Constants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-15
6.2.3.6. FPU Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-15
6.2.4. System Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-16
6.2.5. Streaming SIMD Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-17
6.2.5.1. Streaming SIMD Extensions Data Transfer Instructions. . . . . . . . . . . . . . . . 6-17
6.2.5.2. Streaming SIMD Extensions Conversion Instructions. . . . . . . . . . . . . . . . . . 6-17
6.2.5.3. Streaming SIMD Extensions Packed Arithmetic Instructions . . . . . . . . . . . . 6-18
6.2.5.4. Streaming SIMD Extensions Comparison Instructions . . . . . . . . . . . . . . . . . 6-18
6.2.5.5. Streaming SIMD Extensions Logical Instructions . . . . . . . . . . . . . . . . . . . . . 6-18
6.2.5.6. Streaming SIMD Extensions Data Shuffle Instructions . . . . . . . . . . . . . . . . . 6-19
6.2.5.7. Streaming SIMD Extensions Additional SIMD-Integer Instructions. . . . . . . . 6-19
6.2.5.8. Streaming SIMD Extensions Cacheability Control Instructions. . . . . . . . . . . 6-19
6.2.5.9. Streaming SIMD Extensions State Management Instructions . . . . . . . . . . . 6-19
6.3. DATA MOVEMENT INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-20
6.3.1. General-Purpose Data Movement Instructions . . . . . . . . . . . . . . . . . . . . . . . . . 6-20
6.3.1.1. Move Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-20
6.3.1.2. Conditional Move Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-20
6.3.1.3. Exchange Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-21

v
TABLE OF CONTENTS

6.3.2. Stack Manipulation Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-23

6.3.2.1. Type Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-25
6.3.2.2. Simple Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-25
6.3.2.3. Move and Convert . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-26
6.4. BINARY ARITHMETIC INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-26
6.4.1. Addition and Subtraction Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-26
6.4.2. Increment and Decrement Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-26
6.4.3. Comparison and Sign Change Instruction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-27
6.4.4. Multiplication and Divide Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-27
6.5. DECIMAL ARITHMETIC INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-27
6.5.1. Packed BCD Adjustment Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-28
6.5.2. Unpacked BCD Adjustment Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-28
6.6. LOGICAL INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-29
6.7. SHIFT AND ROTATE INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-29
6.7.1. Shift Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-29
6.7.2. Double-Shift Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-31
6.7.3. Rotate Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-32
6.8. BIT AND BYTE INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-34
6.8.1. Bit Test and Modify Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-34
6.8.2. Bit Scan Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-34
6.8.3. Byte Set on Condition Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-34
6.8.4. Test Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-35
6.9. CONTROL TRANSFER INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-35
6.9.1. Unconditional Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-35
6.9.1.1. Jump Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-35
6.9.1.2. Call and Return Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-36
6.9.1.3. Return From Interrupt Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-36
6.9.2. Conditional Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-36
6.9.2.1. Conditional Jump Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-37
6.9.2.2. Loop Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-38
6.9.2.3. Jump If Zero Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-38
6.9.3. Software Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-39
6.10. STRING OPERATIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-39
6.10.1. Repeating String Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-40
6.11. I/O INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-41
6.12. ENTER AND LEAVE INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-41
6.13. EFLAGS INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-42
6.13.1. Carry and Direction Flag Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-42
6.13.2. Interrupt Flag Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-42
6.13.3. EFLAGS Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-42
6.13.4. Interrupt Flag Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-43
6.14. SEGMENT REGISTER INSTRUCTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-43
6.14.1. Segment-Register Load and Store Instructions. . . . . . . . . . . . . . . . . . . . . . . . . .6-43
6.14.2. Far Control Transfer Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-44
6.14.3. Software Interrupt Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-44
6.14.4. Load Far Pointer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-44
6.15. MISCELLANEOUS INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-44
6.15.1. Address Computation Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-44
6.15.2. Table Lookup Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-45
6.15.3. Processor Identification Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-45
6.15.4. No-Operation and Undefined Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-45

vi
 TABLE OF CONTENTS

CHAPTER 7
FLOATING-POINT UNIT
7.1. COMPATIBILITY AND EASE OF USE OF THE INTEL ARCHITECTURE FPU . . . 7-1
7.2. REAL NUMBERS AND FLOATING-POINT FORMATS . . . . . . . . . . . . . . . . . . . . . . 7-2
7.2.1. Real Number System. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3
7.2.2. Floating-Point Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
7.2.2.1. Normalized Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
7.2.2.2. Biased Exponent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
7.2.3. Real Number and Non-number Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
7.2.3.1. Signed Zeros. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-6
7.2.3.2. Normalized and Denormalized Finite Numbers . . . . . . . . . . . . . . . . . . . . . . . 7-6
7.2.3.3. Signed Infinities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
7.2.3.4. NaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
7.2.4. Indefinite . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
7.3. FPU ARCHITECTURE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
7.3.1. FPU Data Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-9
7.3.1.1. Parameter Passing with the FPU Register Stack . . . . . . . . . . . . . . . . . . . . . 7-11
7.3.2. FPU Status Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-12
7.3.2.1. Top of Stack (TOP) Pointer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-12
7.3.2.2. Condition Code Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-12
7.3.2.3. Exception Flags . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-14
7.3.2.4. Stack Fault Flag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-15
7.3.3. Branching and Conditional Moves on FPU Condition Codes . . . . . . . . . . . . . . 7-15
7.3.4. FPU Control Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-16
7.3.4.1. Exception-Flag Masks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-17
7.3.4.2. Precision Control Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-17
7.3.4.3. Rounding Control Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-18
7.3.5. Infinity Control Flag . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-20
7.3.6. FPU Tag Word. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-20
7.3.7. FPU Instruction and Operand (Data) Pointers . . . . . . . . . . . . . . . . . . . . . . . . . . 7-21
7.3.8. Last Instruction Opcode. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-21
7.3.9. Saving the FPU’s State . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-21
7.4. FLOATING-POINT DATA TYPES AND FORMATS . . . . . . . . . . . . . . . . . . . . . . . . 7-24
7.4.1. Real Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-25
7.4.2. Binary Integers. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-27
7.4.3. Decimal Integers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-29
7.4.4. Unsupported Extended-Real Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-30
7.5. FPU INSTRUCTION SET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-31
7.5.1. Escape (ESC) Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-32
7.5.2. FPU Instruction Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-32
7.5.3. Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-32
7.5.4. Load Constant Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-34
7.5.5. Basic Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-35
7.5.6. Comparison and Classification Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-36
7.5.6.1. Branching on the FPU Condition Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-38
7.5.7. Trigonometric Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-38
7.5.8. Pi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-39
7.5.9. Logarithmic, Exponential, and Scale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-40
7.5.10. Transcendental Instruction Accuracy. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-40
7.5.11. FPU Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-41
7.5.12. Waiting Vs. Non-waiting Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-42
7.5.13. Unsupported FPU Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-43

vii
TABLE OF CONTENTS

7.6. OPERATING ON NANS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-43

7.6.1. Operating on NaNs with Streaming SIMD Extensions . . . . . . . . . . . . . . . . . . . .7-44
7.6.2. Uses for Signaling NANs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-45
7.6.3. Uses for Quiet NANs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-45
7.7. FLOATING-POINT EXCEPTION HANDLING . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-46
7.7.1. Arithmetic vs. Non-arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-46
7.7.2. Automatic Exception Handling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-47
7.7.3. Software Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-49
7.7.3.1. Native Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-49
7.7.3.2. MS-DOS* Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-49
7.7.3.3. Typical Floating-Point Exception Handler Actions . . . . . . . . . . . . . . . . . . . . .7-50
7.8. FLOATING-POINT EXCEPTION CONDITIONS . . . . . . . . . . . . . . . . . . . . . . . . . . 7-51
7.8.1. Invalid Operation Exception. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-51
7.8.1.1. Stack Overflow or Underflow Exception (#IS). . . . . . . . . . . . . . . . . . . . . . . . .7-52
7.8.1.2. Invalid Arithmetic Operand Exception (#IA) . . . . . . . . . . . . . . . . . . . . . . . . . .7-52
7.8.2. Divide-By-Zero Exception (#Z) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-53
7.8.3. Denormal Operand Exception (#D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-54
7.8.4. Numeric Overflow Exception (#O) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-54
7.8.5. Numeric Underflow Exception (#U) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-56
7.8.6. Inexact Result (Precision) Exception (#P) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-57
7.8.7. Exception Priority. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-57
7.9. FLOATING-POINT EXCEPTION SYNCHRONIZATION . . . . . . . . . . . . . . . . . . . . 7-58

CHAPTER 8
PROGRAMMING WITH THE INTEL
MMX™ TECHNOLOGY
8.1. OVERVIEW OF THE MMX™ TECHNOLOGY PROGRAMMING ENVIRONMENT 8-1
8.1.1. MMX™ Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-2
8.1.2. MMX™ Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-3
8.1.3. Single Instruction, Multiple Data (SIMD) Execution Model . . . . . . . . . . . . . . . . . .8-4
8.1.4. Memory Data Formats. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-4
8.1.5. Data Formats for MMX™ Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-5
8.2. MMX™ INSTRUCTION SET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-5
8.2.1. Saturation Arithmetic and Wraparound Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-6
8.2.2. Instruction Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-7
8.3. OVERVIEW OF THE MMX™ INSTRUCTION SET . . . . . . . . . . . . . . . . . . . . . . . . . 8-7
8.3.1. Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-7
8.3.2. Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-9
8.3.2.1. Packed Addition and Subtraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-9
8.3.2.2. Packed Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-9
8.3.2.3. Packed Multiply Add . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-9
8.3.3. Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-9
8.3.4. Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-10
8.3.5. Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-10
8.3.6. Shift Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-10
8.3.7. EMMS (Empty MMX™ State) Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-10
8.4. COMPATIBILITY WITH FPU ARCHITECTURE . . . . . . . . . . . . . . . . . . . . . . . . . . 8-11
8.4.1. MMX™ Instructions and the Floating-Point Tag Word . . . . . . . . . . . . . . . . . . . .8-11
8.4.2. Effect of Instruction Prefixes on MMX™ Instructions . . . . . . . . . . . . . . . . . . . . .8-11
8.5. WRITING APPLICATIONS WITH MMX™ CODE . . . . . . . . . . . . . . . . . . . . . . . . . 8-11
8.5.1. Detecting Support for MMX™ Technology Using the CPUID Instruction . . . . . .8-12
8.5.2. Using the EMMS Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-12

viii
 TABLE OF CONTENTS

8.5.3. Interfacing with MMX™ Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-13

8.5.4. Writing Code with MMX™ and Floating-Point Instructions . . . . . . . . . . . . . . . . 8-14
8.5.4.1. RECOMMENDATIONS AND GUIDELINES . . . . . . . . . . . . . . . . . . . . . . . . . 8-14
8.5.5. Using MMX™ Code in a Multitasking Operating System Environment . . . . . . . 8-15
8.5.5.1. COOPERATIVE MULTITASKING OPERATING SYSTEM. . . . . . . . . . . . . . 8-16
8.5.5.2. PREEMPTIVE MULTITASKING OPERATING SYSTEM . . . . . . . . . . . . . . . 8-16
8.5.6. Exception Handling in MMX™ Code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-16
8.5.7. Register Mapping. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-16

CHAPTER 9
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS
9.1. OVERVIEW OF THE STREAMING SIMD EXTENSIONS . . . . . . . . . . . . . . . . . . . . 9-2
9.1.1. SIMD Floating-Point Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
9.1.2. SIMD Floating-Point Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3
9.1.3. Single Instruction, Multiple Data (SIMD) Execution Model . . . . . . . . . . . . . . . . . 9-4
9.1.4. Pentium® III Processor Single Precision Floating-Point Format . . . . . . . . . . . . . 9-4
9.1.5. Memory Data Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
9.1.6. SIMD Floating-Point Register Data Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-5
9.1.7. SIMD Floating-Point Control/Status Register. . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-7
9.1.8. Rounding Control Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-8
9.1.9. Flush-To-Zero . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-8
9.2. STREAMING SIMD EXTENSIONS SET . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-9
9.2.1. Instruction Operands . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-9
9.3. OVERVIEW OF THE STREAMING SIMD EXTENSIONS SET . . . . . . . . . . . . . . . 9-10
9.3.1. Data Movement Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-10
9.3.2. Arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-11
9.3.2.1. Packed/Scalar Addition and Subtraction. . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-11
9.3.2.2. Packed/Scalar Multiplication and Division . . . . . . . . . . . . . . . . . . . . . . . . . . 9-11
9.3.2.3. Packed/Scalar Square Root . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-11
9.3.2.4. Packed Maximum/Minimum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-12
9.3.3. Comparison Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-12
9.3.4. Conversion Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-13
9.3.5. Logical Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-13
9.3.6. Additional SIMD Integer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-14
9.3.7. Shuffle Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-15
9.3.8. State Management Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-16
9.3.9. Cacheability Control Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-17
9.4. COMPATIBILITY WITH FPU ARCHITECTURE. . . . . . . . . . . . . . . . . . . . . . . . . . . 9-19
9.4.1. Effect of Instruction Prefixes on Streaming SIMD Extensions . . . . . . . . . . . . . . 9-19
9.5. WRITING APPLICATIONS WITH STREAMING SIMD EXTENSIONS CODE. . . . 9-21
9.5.1. Detecting Support for Streaming SIMD Extensions Using the
CPUID Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-21
9.5.2. Interfacing with Streaming SIMD Extensions Procedures and Functions . . . . . 9-22
9.5.3. Writing Code with MMX™, Floating-Point, and Streaming SIMD Extensions . . 9-22
9.5.3.1. Cacheability Hint Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-23
9.5.3.2. Recommendations and Guidelines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-25
9.5.4. Using Streaming SIMD Extensions Code in a Multitasking Operating
System Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-25
9.5.4.1. Cooperative Multitasking Operating System. . . . . . . . . . . . . . . . . . . . . . . . . 9-26
9.5.4.2. Preemptive Multitasking Operating System . . . . . . . . . . . . . . . . . . . . . . . . . 9-26
9.5.5. Exception Handling in Streaming SIMD Extensions . . . . . . . . . . . . . . . . . . . . . 9-26

ix
TABLE OF CONTENTS

CHAPTER 10
INPUT/OUTPUT
10.1. I/O PORT ADDRESSING. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-1
10.2. I/O PORT HARDWARE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-1
10.3. I/O ADDRESS SPACE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
10.3.1. Memory-Mapped I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-2
10.4. I/O INSTRUCTIONS. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-3
10.5. PROTECTED-MODE I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
10.5.1. I/O Privilege Level . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-4
10.5.2. I/O Permission Bit Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-5
10.6. ORDERING I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-6

CHAPTER 11
PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION
11.1. PROCESSOR IDENTIFICATION. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-2
11.2. IDENTIFICATION OF EARLIER INTEL ARCHITECTURE PROCESSORS . . . . . 11-4
11.3. CPUID INSTRUCTION EXTENSIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-5
11.3.1. Version Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-5
11.3.2. Control Register Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-7

APPENDIX A
EFLAGS CROSS-REFERENCE

APPENDIX B
EFLAGS CONDITION CODES

APPENDIX C
FLOATING-POINT EXCEPTIONS SUMMARY

APPENDIX D
SIMD FLOATING-POINT EXCEPTIONS SUMMARY

APPENDIX E
GUIDELINES FOR WRITING FPU
EXCEPTIONS HANDLERS
E.1. ORIGIN OF THE MS-DOS* COMPATIBILITY MODE FOR HANDLING
FPU EXCEPTIONS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-2
E.2. IMPLEMENTATION OF THE MS-DOS* COMPATIBILITY MODE IN THE INTEL486™,
PENTIUM®, AND P6 FAMILY PROCESSORS E-3
E.2.1. MS-DOS* Compatibility Mode in the Intel486™ and Pentium® Processors . . . . E-3
E.2.1.1. Basic Rules: When FERR# Is Generated. . . . . . . . . . . . . . . . . . . . . . . . . . . . E-4
E.2.1.2. Recommended External Hardware to Support the
MS-DOS* Compatibility Mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-5
E.2.1.3. No-Wait FPU Instructions Can Get FPU Interrupt in Window. . . . . . . . . . . . . E-7
E.2.2. MS-DOS* Compatibility Mode in the P6 Family Processors . . . . . . . . . . . . . . . . E-9
E.3. RECOMMENDED PROTOCOL FOR MS-DOS* COMPATIBILITY HANDLERS. . E-10
E.3.1. Floating-Point Exceptions and Their Defaults . . . . . . . . . . . . . . . . . . . . . . . . . . E-10
E.3.2. Two Options for Handling Numeric Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . E-11
E.3.2.1. Automatic Exception Handling: Using Masked Exceptions . . . . . . . . . . . . . E-11
E.3.2.2. Software Exception Handling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-12
E.3.3. Synchronization Required for Use of FPU Exception Handlers . . . . . . . . . . . . E-14

x
 TABLE OF CONTENTS

E.3.3.1. Exception Synchronization: What, Why and When. . . . . . . . . . . . . . . . . . . . E-14

E.3.3.2. Exception Synchronization Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-15
E.3.3.3. Proper Exception Synchronization in General . . . . . . . . . . . . . . . . . . . . . . . E-16
E.3.3.4. FPU Exception Handling Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-16
E.3.4. Need for Storing State of IGNNE# Circuit If Using FPU and SMM . . . . . . . . . . E-20
E.3.5. Considerations When FPU Shared Between Tasks . . . . . . . . . . . . . . . . . . . . . E-21
E.3.5.1. Speculatively Deferring FPU Saves, General Overview . . . . . . . . . . . . . . . . E-21
E.3.5.2. Tracking FPU Ownership . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-22
E.3.5.3. Interaction of FPU State Saves and Floating-point Exception Association. . E-22
E.3.5.4. Interrupt Routing From the Kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-25
E.3.5.5. Special Considerations for Operating Systems that Support
Streaming SIMD Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-25
E.4. DIFFERENCES FOR HANDLERS USING NATIVE MODE . . . . . . . . . . . . . . . . . . E-26
E.4.1. Origin with the Intel 286 and Intel 287, and Intel386™ and
Intel 387 Processors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-26
E.4.2. Changes with Intel486™, Pentium®, and P6 Family Processors
with CR0.NE=1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-27
E.4.3. Considerations When FPU Shared Between Tasks Using Native Mode. . . . . . E-27

APPENDIX F
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION HANDLERS
F.1. TWO OPTIONS FOR HANDLING NUMERIC EXCEPTIONS . . . . . . . . . . . . . . . . . F-1
F.2. SOFTWARE EXCEPTION HANDLING . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F-1
F.3. EXCEPTION SYNCHRONIZATION . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F-3
F.4. SIMD FLOATING-POINT EXCEPTIONS AND THE IEEE-754 STANDARD FOR
BINARY FLOATING-POINT COMPUTATIONS F-4
F.4.1. Floating-Point Emulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F-4
F.4.2. Streaming SIMD Extensions Response To Floating-Point Exceptions . . . . . . . . F-6
F.4.2.1. Numeric Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F-7
F.4.2.2. Results of Operations with NaN Operands or a NaN Result for
Streaming SIMD Extensions Numeric Instructions . . . . . . . . . . . . . . . . . . . . . F-7
F.4.2.3. Condition Codes, Exception Flags, and Response for Masked and
Unmasked Numeric Exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . F-10
F.4.3. SIMD Floating-Point Emulation Implementation Example . . . . . . . . . . . . . . . . . F-13

xi
TABLE OF CONTENTS

xii
 TABLE OF FIGURES

Figure 1-1. Bit and Byte Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-6

Figure 2-1. The Processing Units in the P6 Family Processor Microarchitecture
and Their Interface with the Memory Subsystem . . . . . . . . . . . . . . . . . . . . . . .2-7
Figure 2-2. Functional Block Diagram of the P6 Family Processor Microarchitecture . . .2-10
Figure 3-1. P6 Family Processor Basic Execution Environment. . . . . . . . . . . . . . . . . . . . .3-2
Figure 3-2. Three Memory Management Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-3
Figure 3-3. Application Programming Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-6
Figure 3-4. Alternate General-Purpose Register Names . . . . . . . . . . . . . . . . . . . . . . . . . .3-7
Figure 3-5. Use of Segment Registers for Flat Memory Model. . . . . . . . . . . . . . . . . . . . . .3-8
Figure 3-6. Use of Segment Registers in Segmented Memory Model . . . . . . . . . . . . . . . .3-9
Figure 3-7. EFLAGS Register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-11
Figure 4-1. Stack Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-2
Figure 4-2. Stack on Near and Far Calls. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-6
Figure 4-3. Protection Rings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-9
Figure 4-4. Stack Switch on a Call to a Different Privilege Level . . . . . . . . . . . . . . . . . . .4-11
Figure 4-5. Stack Usage on Transfers to Interrupt and Exception Handling Routines . . .4-15
Figure 4-6. Nested Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-20
Figure 4-7. Stack Frame after Entering the MAIN Procedure . . . . . . . . . . . . . . . . . . . . . .4-21
Figure 4-8. Stack Frame after Entering Procedure A . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-22
Figure 4-9. Stack Frame after Entering Procedure B . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-23
Figure 4-10. Stack Frame after Entering Procedure C . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-24
Figure 5-1. Fundamental Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-1
Figure 5-2. SIMD Floating-Point Data Type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-1
Figure 5-3. Bytes, Words, Doublewords and Quadwords in Memory . . . . . . . . . . . . . . . . .5-2
Figure 5-4. Numeric, Pointer, and Bit Field Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . .5-4
Figure 5-5. Memory Operand Address . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-7
Figure 5-6. Offset (or Effective Address) Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-9
Figure 6-1. Operation of the PUSH Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-23
Figure 6-2. Operation of the PUSHA Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-24
Figure 6-3. Operation of the POP Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-24
Figure 6-4. Operation of the POPA Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-25
Figure 6-5. Sign Extension . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-25
Figure 6-6. SHL/SAL Instruction Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-29
Figure 6-7. SHR Instruction Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-30
Figure 6-8. SAR Instruction Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-31
Figure 6-9. SHLD and SHRD Instruction Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-32
Figure 6-10. ROL, ROR, RCL, and RCR Instruction Operations . . . . . . . . . . . . . . . . . . . .6-33
Figure 6-11. Flags Affected by the PUSHF, POPF, PUSHFD, and POPFD instructions . .6-43
Figure 7-1. Binary Real Number System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-3
Figure 7-2. Binary Floating-Point Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-4
Figure 7-3. Real Numbers and NaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-6
Figure 7-4. Relationship Between the Integer Unit and the FPU . . . . . . . . . . . . . . . . . . . .7-9
Figure 7-5. FPU Execution Environment. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-10
Figure 7-6. FPU Data Register Stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-10
Figure 7-7. Example FPU Dot Product Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-12
Figure 7-8. FPU Status Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-13
Figure 7-9. Moving the FPU Condition Codes to the EFLAGS Register. . . . . . . . . . . . . .7-16
Figure 7-10. FPU Control Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-17
Figure 7-11. FPU Tag Word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-20

xiii
TABLE OF FIGURES

Figure 7-12. Contents of FPU Opcode Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-22

Figure 7-13. Protected Mode FPU State Image in Memory, 32-Bit Format . . . . . . . . . . . .7-22
Figure 7-14. Real Mode FPU State Image in Memory, 32-Bit Format . . . . . . . . . . . . . . . .7-23
Figure 7-15. Protected Mode FPU State Image in Memory, 16-Bit Format . . . . . . . . . . . .7-23
Figure 7-16. Real Mode FPU State Image in Memory, 16-Bit Format . . . . . . . . . . . . . . . .7-24
Figure 7-17. Floating-Point Unit Data Type Formats . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-25
Figure 8-1. MMX™ Register Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-2
Figure 8-2. MMX™ Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-3
Figure 8-3. Eight Packed Bytes in Memory (at address 1000H) . . . . . . . . . . . . . . . . . . . . .8-4
Figure 9-1. SIMD Floating-Point Registers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-3
Figure 9-2. Packed Single-FP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-3
Figure 9-3. Four Packed FP Data in Memory (at address 1000H) . . . . . . . . . . . . . . . . . . .9-5
Figure 9-4. SIMD Floating-Point Control/Status Register Format . . . . . . . . . . . . . . . . . . . .9-7
Figure 9-5. Packed Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-9
Figure 9-6. Scalar Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-10
Figure 9-7. Packed Shuffle Operation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-15
Figure 9-8. Unpack High Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-16
Figure 9-9. Unpack Low Operation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-16
Figure 10-1. Memory-Mapped I/O. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-3
Figure 10-2. I/O Permission Bit Map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-5
Figure 11-1. EAX Return Values. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-6
Figure 11-2. CPUID Feature Field Information Bits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-6
Figure 11-3. CR4 Register Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-7
Figure E-1. Recommended Circuit for MS-DOS* Compatibility FPU
Exception Handling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-6
Figure E-2. Behavior of Signals During FPU Exception Handling . . . . . . . . . . . . . . . . . . . E-7
Figure E-3. Timing of Receipt of External Interrupt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-8
Figure E-4. Arithmetic Example Using Infinity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E-12
Figure E-5. General Program Flow for DNA Exception Handler . . . . . . . . . . . . . . . . . . . E-24
Figure E-6. Program Flow for a Numeric Exception Dispatch Routine . . . . . . . . . . . . . . E-24
Figure F-1. Control Flow for Handling Unmasked Floating-Point Exceptions . . . . . . . . . . .F-6

xiv
 TABLE OF TABLES

Table 2-1. Processor Performance Over Time and Other Intel Architecture
Key Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-5
Table 3-1. Effective Operand- and Address-Size Attributes . . . . . . . . . . . . . . . . . . . . . .3-15
Table 4-1. Exceptions and Interrupts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-14
Table 5-1. Default Segment Selection Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-8
Table 6-1. Move Instruction Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-21
Table 6-2. Conditional Move Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-22
Table 6-3. Bit Test and Modify Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-34
Table 6-4. Conditional Jump Instructions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-37
Table 6-5. Information Provided by the CPUID Instruction . . . . . . . . . . . . . . . . . . . . . . .6-45
Table 7-1. Real Number Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-5
Table 7-2. Denormalization Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-7
Table 7-3. FPU Condition Code Interpretation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-14
Table 7-4. Precision Control Field (PC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-17
Table 7-5. Rounding Control Field (RC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-18
Table 7-6. Rounding of Positive Numbers with Masked Overflow . . . . . . . . . . . . . . . . . .7-19
Table 7-7. Rounding of Negative Numbers with Masked Overflow . . . . . . . . . . . . . . . . .7-19
Table 7-8. Length, Precision, and Range of FPU Data Types. . . . . . . . . . . . . . . . . . . . .7-26
Table 7-9. Real Number and NaN Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-27
Table 7-10. Binary Integer Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-28
Table 7-11. Packed Decimal Integer Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-29
Table 7-12. Unsupported Extended-Real Encodings. . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-31
Table 7-13. Data Transfer Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-32
Table 7-14. Floating-Point Conditional Move Instructions . . . . . . . . . . . . . . . . . . . . . . . . .7-33
Table 7-15. Setting of FPU Condition Code Flags for Real Number Comparisons . . . . . .7-37
Table 7-16. Setting of EFLAGS Status Flags for Real Number Comparisons. . . . . . . . . .7-37
Table 7-17. TEST Instruction Constants for Conditional Branching . . . . . . . . . . . . . . . . .7-38
Table 7-18. Rules for Generating QNaNs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-44
Table 7-19. Results of Operations with NaN Operands. . . . . . . . . . . . . . . . . . . . . . . . . . .7-45
Table 7-20. Arithmetic and Non-arithmetic Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . .7-48
Table 7-21. Invalid Arithmetic Operations and the Masked Responses to Them . . . . . . .7-53
Table 7-22. Divide-By-Zero Conditions and the Masked Responses to Them . . . . . . . . .7-54
Table 7-23. Masked Responses to Numeric Overflow. . . . . . . . . . . . . . . . . . . . . . . . . . . .7-55
Table 8-1. Data Range Limits for Saturation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-6
Table 8-2. MMX™ Instruction Set Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-8
Table 8-3. Effect of Prefixes on MMX™ Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-11
Table 9-1. Precision and Range of SIMD Floating-point Datatype . . . . . . . . . . . . . . . . . .9-5
Table 9-2. Real Number and NaN Encodings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-6
Table 9-3. Rounding Control Field (RC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-8
Table 9-4. Streaming SIMD Extensions Behavior with Prefixes . . . . . . . . . . . . . . . . . . .9-20
Table 9-5. SIMD Integer Instructions Behavior with Prefixes. . . . . . . . . . . . . . . . . . . . . .9-20
Table 9-6. Cacheability Control Instruction Behavior with Prefixes . . . . . . . . . . . . . . . . .9-20
Table 9-7. Cache Hints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9-24
Table 10-1. I/O Instruction Serialization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-7
Table 11-1. EAX Input Value and CPUID Return Values . . . . . . . . . . . . . . . . . . . . . . . . .11-5
Table 11-2. New P6-Family Processor Feature Information Returned by
CPUID in EDX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .11-6
Table A-1. EFLAGS Cross-Reference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A-1
Table B-1. EFLAGS Condition Codes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B-1

xv
TABLE OF TABLES

Table C-1. Floating-Point Exceptions Summary. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C-1

Table D-1. Streaming SIMD Extensions Instruction Set Summary . . . . . . . . . . . . . . . . . D-2
Table F-1. ADDPS, ADDSS, SUBPS, SUBSS, MULPS, MULSS, DIVPS, DIVSS . . . . . .F-8
Table F-2. CMPPS.EQ, CMPSS.EQ, CMPPS.ORD, CMPSS.ORD . . . . . . . . . . . . . . . . .F-8
Table F-3. CMPPS.NEQ, CMPSS.NEQ, CMPPS.UNORD, CMPSS.UNORD. . . . . . . . . .F-8
Table F-4. CMPPS.LT, CMPSS.LT, CMPPS.LE, CMPSS.LE . . . . . . . . . . . . . . . . . . . . . .F-8
Table F-5. CMPPS.NLT, CMPSS.NLT, CMPSS.NLT, CMPSS.NLE . . . . . . . . . . . . . . . . .F-8
Table F-6. COMISS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-9
Table F-7. UCOMISS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-9
Table F-8. CVTPS2PI, CVTSS2SI, CVTTPS2PI, CVTTSS2SI . . . . . . . . . . . . . . . . . . . . .F-9
Table F-9. MAXPS, MAXSS, MINPS, MINSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-9
Table F-10. SQRTPS, SQRTSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-9
Table F-11. #I - Invalid Operations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-10
Table F-12. #Z - Divide-by-Zero. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-11
Table F-13. #D - Denormal Operand . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-12
Table F-14. #0 - Numeric Overflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-12
Table F-15. #U - Numeric Underflow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-13
Table F-16. #P - Inexact Result (Precision) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .F-13

xvi
 1
About This Manual
 CHAPTER 1
ABOUT THIS MANUAL

The Intel Architecture Software Developer’s Manual, Volume 1: Basic Architecture (Order
Number 243190) is part of a three-volume set that describes the architecture and programming
environment of all Intel Architecture (IA) processors. The other two volumes in this set are:
• The Intel Architecture Software Developer’s Manual, Volume 2: Instruction Set Reference
(Order Number 243191).
• The Intel Architecture Software Developer’s Manual, Volume 3: System Programming
Guide (Order Number 243192).
The Intel Architecture Software Developer’s Manual, Volume 1, describes the basic architecture
and programming environment of an IA processor; the Intel Architecture Software Developer’s
Manual, Volume 2, describes the instruction set of the processor and the opcode structure. These
two volumes are aimed at application programmers who are writing programs to run under
existing operating systems or executives. The Intel Architecture Software Developer’s Manual,
Volume 3 describes the operating-system support environment of an IA processor, including
memory management, protection, task management, interrupt and exception handling, and
system management mode. It also provides IA processor compatibility information. This
volume is aimed at operating-system and BIOS designers and programmers.

1.1. OVERVIEW OF THE INTEL ARCHITECTURE SOFTWARE

DEVELOPER’S MANUAL, VOLUME 1: BASIC
ARCHITECTURE
The contents of this manual are as follows:
Chapter 1 — About This Manual. Gives an overview of all three volumes of the Intel Archi-
tecture Software Developer’s Manual. It also describes the notational conventions in these
manuals and lists related Intel manuals and documentation of interest to programmers and hard-
ware designers.
Chapter 2 — Introduction to the Intel Architecture. Introduces the IA and the families of
Intel processors that are based on this architecture. It also gives an overview of the common
features found in these processors and brief history of the IA.
Chapter 3 — Basic Execution Environment. Introduces the models of memory organization
and describes the register set used by applications.
Chapter 4 — Procedure Calls, Interrupts, and Exceptions. Describes the procedure stack
and the mechanisms provided for making procedure calls and for servicing interrupts and excep-
tions.

1-1
ABOUT THIS MANUAL

Chapter 5 — Data Types and Addressing Modes. Describes the data types and addressing
modes recognized by the processor.
Chapter 6 — Instruction Set Summary. Gives an overview of all the IA instructions except
those executed by the processor’s floating-point unit. The instructions are presented in function-
ally related groups.
Chapter 7 — Floating-Point Unit. Describes the IA floating-point unit, including the floating-
point registers and data types; gives an overview of the floating-point instruction set; and
describes the processor’s floating-point exception conditions.
Chapter 8 — Programming with Intel MMX™ Technology. Describes the Intel MMX™
technology, including MMX™ registers and data types, and gives an overview of the MMX™
instruction set.
Chapter 9 — Programming with the Streaming SIMD Extensions. Describes the Intel
Streaming SIMD Extensions, including the registers and data types.
Chapter 10 — Input/Output. Describes the processor’s I/O architecture, including I/O port
addressing, the I/O instructions, and the I/O protection mechanism.
Chapter 11 — Processor Identification and Feature Determination. Describes how to deter-
mine the CPU type and the features that are available in the processor.
Appendix A — EFLAGS Cross-Reference. Summarizes how the IA instructions affect the
flags in the EFLAGS register.
Appendix B — EFLAGS Condition Codes. Summarizes how the conditional jump, move, and
byte set on condition code instructions use the condition code flags (OF, CF, ZF, SF, and PF) in
the EFLAGS register.
Appendix C — Floating-Point Exceptions Summary. Summarizes the exceptions that can be
raised by floating-point instructions.
Appendix D — SIMD Floating-Point Exceptions Summary. Provides the Streaming SIMD
Extensions mnemonics, and the exceptions that each instruction can cause.
Appendix E — Guidelines for Writing FPU Exception Handlers. Describes how to design
and write MS-DOS* compatible exception handling facilities for FPU and SIMD floating-point
exceptions, including both software and hardware requirements and assembly-language code
examples. This appendix also describes general techniques for writing robust FPU exception
handlers.
Appendix F — Guidelines for Writing SIMD-FP Exception Handlers. Provides guidelines
for the Streaming SIMD Extensions instructions that can generate numeric (floating-point)
exceptions, and gives an overview of the necessary support for handling such exceptions.

1-2
 ABOUT THIS MANUAL

1.2. OVERVIEW OF THE INTEL ARCHITECTURE SOFTWARE

DEVELOPER’S MANUAL, VOLUME 2: INSTRUCTION SET
REFERENCE
The contents of the Intel Architecture Software Developer’s Manual, Volume 2 are as follows:
Chapter 1 — About This Manual. Gives an overview of all three volumes of the Intel Archi-
tecture Software Developer’s Manual. It also describes the notational conventions in these
manuals and lists related Intel manuals and documentation of interest to programmers and hard-
ware designers.
Chapter 2 — Instruction Format. Describes the machine-level instruction format used for all
IA instructions and gives the allowable encodings of prefixes, the operand-identifier byte
(ModR/M byte), the addressing-mode specifier byte (SIB byte), and the displacement and
immediate bytes.
Chapter 3 — Instruction Set Reference. Describes each of the IA instructions in detail,
including an algorithmic description of operations, the effect on flags, the effect of operand- and
address-size attributes, and the exceptions that may be generated. The instructions are arranged
in alphabetical order. The FPU, MMX™, and Streaming SIMD Extensions instructions are
included in this chapter.
Appendix A — Opcode Map. Gives an opcode map for the IA instruction set.
Appendix B — Instruction Formats and Encodings. Gives the binary encoding of each form
of each IA instruction.
Appendix C — Compiler Intrinsics and Functional Equivalents. Gives the Intel C/C++
compiler intrinsics and functional equivalents for the MMX™ Technology instructions and
Streaming SIMD Extensions.

1.3. OVERVIEW OF THE INTEL ARCHITECTURE SOFTWARE

DEVELOPER’S MANUAL, VOLUME 3: SYSTEM
PROGRAMMING GUIDE
The contents of the Intel Architecture Software Developer’s Manual, Volume 3 are as follows:
Chapter 1 — About This Manual. Gives an overview of all three volumes of the Intel Archi-
tecture Software Developer’s Manual. It also describes the notational conventions in these
manuals and lists related Intel manuals and documentation of interest to programmers and hard-
ware designers.
Chapter 2 — System Architecture Overview. Describes the modes of operation of an IA
processor and the mechanisms provided in the IA to support operating systems and executives,
including the system-oriented registers and data structures and the system-oriented instructions.
The steps necessary for switching between real-address and protected modes are also identified.

1-3
ABOUT THIS MANUAL

Chapter 3 — Protected-Mode Memory Management. Describes the data structures, registers,

and instructions that support segmentation and paging and explains how they can be used to
implement a “flat” (unsegmented) memory model or a segmented memory model.
Chapter 4 — Protection. Describes the support for page and segment protection provided in
the IA. This chapter also explains the implementation of privilege rules, stack switching, pointer
validation, user and supervisor modes.
Chapter 5 — Interrupt and Exception Handling. Describes the basic interrupt mechanisms
defined in the IA, shows how interrupts and exceptions relate to protection, and describes how
the architecture handles each exception type. Reference information for each IA exception is
given at the end of this chapter.
Chapter 6 — Task Management. Describes the mechanisms the IA provides to support multi-
tasking and inter-task protection.
Chapter 7 — Multiple Processor Management. Describes the instructions and flags that
support multiple processors with shared memory, memory ordering, and the advanced program-
mable interrupt controller (APIC).
Chapter 8 — Processor Management and Initialization. Defines the state of an IA processor
and its floating-point and SIMD floating-point units after reset initialization. This chapter also
explains how to set up an IA processor for real-address mode operation and protected-mode
operation, and how to switch between modes.
Chapter 9 — Memory Cache Control. Describes the general concept of caching and the
caching mechanisms supported by the IA. This chapter also describes the memory type range
registers (MTRRs) and how they can be used to map memory types of physical memory.
MTRRs were introduced into the IA with the Pentium® Pro processor. It also presents informa-
tion on using the new cache control and memory streaming instructions introduced with the
Pentium® III processor.
Chapter 10 — MMX™ Technology System Programming. Describes those aspects of the
Intel MMX™ technology that must be handled and considered at the system programming level,
including task switching, exception handling, and compatibility with existing system environ-
ments. The MMX™ technology was introduced into the IA with the Pentium® processor.
Chapter 11 — Streaming SIMD Extensions System Programming. Describes those aspects
of Streaming SIMD Extensions that must be handled and considered at the system programming
level, including task switching, exception handling, and compatibility with existing system
environments. Streaming SIMD Extensions were introduced into the IA with the Pentium®
processor.
Chapter 12 — System Management Mode (SMM). Describes the IA’s system management
mode (SMM), which can be used to implement power management functions.
Chapter 13 — Machine-Check Architecture. Describes the machine-check architecture,
which was introduced into the IA with the Pentium® processor.
Chapter 14 — Code Optimization. Discusses general optimization techniques for program-
ming an IA processor.

1-4
 ABOUT THIS MANUAL

Chapter 15 — Debugging and Performance Monitoring. Describes the debugging registers

and other debug mechanism provided in the IA. This chapter also describes the time-stamp
counter and the performance-monitoring counters.
Chapter 16 — 8086 Emulation. Describes the real-address and virtual-8086 modes of the IA.
Chapter 17 — Mixing 16-Bit and 32-Bit Code. Describes how to mix 16-bit and 32-bit code
modules within the same program or task.
Chapter 18 — Intel Architecture Compatibility. Describes the programming differences
between the Intel 286, Intel386™, Intel486™, Pentium®, and P6 family processors. The differ-
ences among the 32-bit IA processors (the Intel386™, Intel486™, Pentium®, and P6 family
processors) are described throughout the three volumes of the Intel Architecture Software Devel-
oper’s Manual, as relevant to particular features of the architecture. This chapter provides a
collection of all the relevant compatibility information for all IA processors and also describes
the basic differences with respect to the 16-bit IA processors (the Intel 8086 and Intel 286
processors).
Appendix A — Performance-Monitoring Events. Lists the events that can be counted with
the performance-monitoring counters and the codes used to select these events. Both Pentium®
processor and P6 family processor events are described.
Appendix B — Model-Specific Registers (MSRs). Lists the MSRs available in the Pentium®
and P6 family processors and their functions.
Appendix C — Dual-Processor (DP) Bootup Sequence Example (Specific to Pentium®
Processors). Gives an example of how to use the DP protocol to boot two Pentium® processors
(a primary processor and a secondary processor) in a DP system and initialize their APICs.
Appendix D — Multiple-Processor (MP) Bootup Sequence Example (Specific to P6 Family
Processors). Gives an example of how to use of the MP protocol to boot two P6 family proces-
sors in a MP system and initialize their APICs.
Appendix E — Programming the LINT0 and LINT1 Inputs. Gives an example of how to
program the LINT0 and LINT1 pins for specific interrupt vectors.

1.4. NOTATIONAL CONVENTIONS

This manual uses special notation for data-structure formats, for symbolic representation of
instructions, and for hexadecimal numbers. A review of this notation makes the manual easier
to read.

1.4.1. Bit and Byte Order

In illustrations of data structures in memory, smaller addresses appear toward the bottom of the
figure; addresses increase toward the top. Bit positions are numbered from right to left. The
numerical value of a set bit is equal to two raised to the power of the bit position. IA processors
are “little endian” machines; this means the bytes of a word are numbered starting from the least
significant byte. Figure 1-1 illustrates these conventions.

1-5
ABOUT THIS MANUAL

Data Structure
Highest 24 23 16 15 8 7 0 Bit offset
31
Address
28
24
20
16
12
8
4
Lowest
Byte 3 Byte 2 Byte 1 Byte 0 0 Address

Byte Offset

Figure 1-1. Bit and Byte Order

1.4.2. Reserved Bits and Software Compatibility

In many register and memory layout descriptions, certain bits are marked as reserved. When
bits are marked as reserved, it is essential for compatibility with future processors that software
treat these bits as having a future, though unknown, effect. The behavior of reserved bits should
be regarded as not only undefined, but unpredictable. Software should follow these guidelines
in dealing with reserved bits:
• Do not depend on the states of any reserved bits when testing the values of registers which
contain such bits. Mask out the reserved bits before testing.
• Do not depend on the states of any reserved bits when storing to memory or to a register.
• Do not depend on the ability to retain information written into any reserved bits.
• When loading a register, always load the reserved bits with the values indicated in the
documentation, if any, or reload them with values previously read from the same register.

NOTE
Avoid any software dependence upon the state of reserved bits in IA registers.
Depending upon the values of reserved register bits will make software
dependent upon the unspecified manner in which the processor handles these
bits. Programs that depend upon reserved values risk incompatibility with
future processors.

1-6
 ABOUT THIS MANUAL

1.4.3. Instruction Operands

When instructions are represented symbolically, a subset of the IA assembly language is used.
In this subset, an instruction has the following format:
label: mnemonic argument1, argument2, argument3
where:
• A label is an identifier which is followed by a colon.
• A mnemonic is a reserved name for a class of instruction opcodes which have the same
function.
• The operands argument1, argument2, and argument3 are optional. There may be from
zero to three operands, depending on the opcode. When present, they take the form of
either literals or identifiers for data items. Operand identifiers are either reserved names of
registers or are assumed to be assigned to data items declared in another part of the
program (which may not be shown in the example).
When two operands are present in an arithmetic or logical instruction, the right operand is the
source and the left operand is the destination.
For example:
LOADREG: MOV EAX, SUBTOTAL
In this example, LOADREG is a label, MOV is the mnemonic identifier of an opcode, EAX is
the destination operand, and SUBTOTAL is the source operand. Some assembly languages put
the source and destination in reverse order.

1.4.4. Hexadecimal and Binary Numbers

Base 16 (hexadecimal) numbers are represented by a string of hexadecimal digits followed by
the character H (for example, F82EH). A hexadecimal digit is a character from the following
set: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, and F.
Base 2 (binary) numbers are represented by a string of 1s and 0s, sometimes followed by the
character B (for example, 1010B). The “B” designation is only used in situations where confu-
sion as to the type of number might arise.

1.4.5. Segmented Addressing

The processor uses byte addressing. This means memory is organized and accessed as a
sequence of bytes. Whether one or more bytes are being accessed, a byte address is used to
locate the byte or bytes memory. The range of memory that can be addressed is called an
address space.
The processor also supports segmented addressing. This is a form of addressing where a
program may have many independent address spaces, called segments. For example, a program
can keep its code (instructions) and stack in separate segments. Code addresses would always

1-7
ABOUT THIS MANUAL

refer to the code space, and stack addresses would always refer to the stack space. The following
notation is used to specify a byte address within a segment:
Segment-register:Byte-address
For example, the following segment address identifies the byte at address FF79H in the segment
pointed by the DS register:
DS:FF79H
The following segment address identifies an instruction address in the code segment. The CS
register points to the code segment and the EIP register contains the address of the instruction.
CS:EIP

1.4.6. Exceptions
An exception is an event that typically occurs when an instruction causes an error. For example,
an attempt to divide by zero generates an exception. However, some exceptions, such as break-
points, occur under other conditions. Some types of exceptions may provide error codes. An
error code reports additional information about the error. An example of the notation used to
show an exception and error code is shown below.
#PF(fault code)
This example refers to a page-fault exception under conditions where an error code naming a
type of fault is reported. Under some conditions, exceptions which produce error codes may not
be able to report an accurate code. In this case, the error code is zero, as shown below for a
general-protection exception.
#GP(0)
Refer to Chapter 5, Interrupt and Exception Handling, in the Intel Architecture Software Devel-
oper’s Manual, Volume 3, for a list of exception mnemonics and their descriptions.

1-8
 ABOUT THIS MANUAL

1.5. RELATED LITERATURE

The following books contain additional material related to Intel processors:
• Intel Pentium® II Processor Specification Update, Order Number 243337-010.
• Intel Pentium® Pro Processor Specification Update, Order Number 242689.
• Intel Pentium® Processor Specification Update, Order Number 242480.
• AP-485, Intel Processor Identification and the CPUID Instruction, Order Number 241618.
• AP-578, Software and Hardware Considerations for FPU Exception Handlers for Intel
Architecture Processors, Order Number 242415-001.
• Pentium® Pro Processor Family Developer’s Manual, Volume 1: Specifications, Order
Number 242690-001.
• Pentium® Processor Family Developer’s Manual, Order Number 241428.
• Intel486™ Microprocessor Data Book, Order Number 240440.
• Intel486™ SX CPU/Intel487™ SX Math Coprocessor Data Book, Order Number 240950.
• Intel486™ DX2 Microprocessor Data Book, Order Number 241245.
• Intel486™ Microprocessor Product Brief Book, Order Number 240459.
• Intel386™ Processor Hardware Reference Manual, Order Number 231732.
• Intel386™ Processor System Software Writer's Guide, Order Number 231499.
• Intel386™ High-Performance 32-Bit CHMOS Microprocessor with Integrated Memory
Management, Order Number 231630.
• 376 Embedded Processor Programmer’s Reference Manual, Order Number 240314.
• 80387 DX User’s Manual Programmer’s Reference, Order Number 231917.
• 376 High-Performance 32-Bit Embedded Processor, Order Number 240182.
• Intel386™ SX Microprocessor, Order Number 240187.
• Microprocessor and Peripheral Handbook (Vol. 1), Order Number 230843.
• AP-528, Optimizations for Intel’s 32-Bit Processors, Order Number 242816-001.

1-9
ABOUT THIS MANUAL

1-10
 2
Introduction to the
Intel Architecture
 CHAPTER 2
INTRODUCTION TO THE INTEL ARCHITECTURE

A strong case can be made that the exponential growth of both the power and breadth of usage
of the computer has made it the most important force that is reshaping human technology, busi-
ness, and society in the second half of the twentieth century. Further, the computer promises to
continue to dominate technological growth well into the twenty-first century, in part since other
powerful technological forces that are just emerging are strongly dependent on the growth of
computing power for their own existence and growth (such as the Internet, and genetics devel-
opments like recombinant DNA research and development). The Intel Architecture (IA) is
clearly today’s preferred computer architecture, as measured by number of computers in use and
total computing power available in the world. Thus it is hard to overestimate the importance of
the IA.

2.1. BRIEF HISTORY OF THE INTEL ARCHITECTURE

The developments leading to the IA can be traced back through the 8085 and 8080 microproces-
sors to the 4004 microprocessor (the first microprocessor, designed by Intel in 1969). However,
the first actual processor in the IA family is the 8086, quickly followed by a more cost effective
version for smaller systems, the 8088. The object code programs created for these processors
starting in 1978 will still execute on the latest members of the IA family.
The 8086 has 16-bit registers and a 16-bit external data bus, with 20-bit addressing giving a 1-
MByte address space. The 8088 is identical except for a smaller external data bus of 8 bits.
These processors introduced IA segmentation, but only in “Real Mode”; 16-bit registers can act
as pointers to address into segments of up to 64 KBytes in size. The four segment registers hold
the (effectively) 20-bit base addresses of the currently active segments; up to 256 KBytes can
be addressed without switching between segments, and a total address range of 1 MByte is avail-
able.
The Intel 80286 processor introduced the Protected Mode into the IA. This new mode uses the
segment register contents as selectors or pointers into descriptor tables. The descriptors provide
24-bit base addresses, allowing a maximum physical memory size of up to 16 MBytes, support
for virtual memory management on a segment swapping basis, and various protection mecha-
nisms. These include segment limit checking, read-only and execute-only segment options, and
up to four privilege levels to protect operating system code (in several subdivisions, if desired)
from application or user programs. Furthermore, hardware task switching and the local
descriptor tables allow the operating system to protect application or user programs from each
other.
The Intel386™ processor introduced 32-bit registers into the architecture, for use both as oper-
ands for calculations and for addressing. The lower half of each 32-bit register retained the prop-
erties of one of the 16-bit registers of the earlier two generations, to provide complete upward
compatibility. A new virtual-8086 mode was provided to yield greater efficiency when

2-1
INTRODUCTION TO THE INTEL ARCHITECTURE

executing programs created for the 8086 and 8088 processors on the new 32-bit machine. The
32-bit addressing was supported with an external 32-bit address bus, giving a 4-GByte address
space, and also allowed each segment to be as large as 4 GBytes. The original instructions were
enhanced with new 32-bit operand and addressing forms, and completely new instructions were
provided, including those for bit manipulation. The Intel386™ processor also introduced paging
into the IA, with the fixed 4-KByte page size providing a method for virtual memory manage-
ment that was significantly superior compared to using segments for the purpose (it was much
more efficient for operating systems, and completely transparent to the applications without
significant sacrifice of execution speed). Furthermore, the ability to define segments as large as
the 4 GBytes physical address space, together with paging, allowed the creation of protected
“flat model”1 addressing systems in the architecture, including complete implementations of the
widely used mainframe operating system UNIX.
The IA has been and is committed to the task of maintaining backward compatibility at the
object code level to preserve our customers’ very large investment in software, but at the same
time, in each generation of the architecture, the latest most effective microprocessor architecture
and silicon fabrication technologies have been used to produce the fastest, most powerful
processors possible. Intel has worked over the generations to adapt and incorporate increasingly
sophisticated techniques from mainframe architecture into microprocessor architecture. Various
forms of parallel processing have been the most performance enhancing of these techniques, and
the Intel386™ processor was the first IA processor to include a number of parallel stages: six.
These are the Bus Interface Unit (accesses memory and I/O for the other units), the Code
Prefetch Unit (receives object code from the Bus Unit and puts it into a 16-byte queue), the
Instruction Decode Unit (decodes object code from the Prefetch unit into microcode), the
Execution Unit (executes the microcode instructions), the Segment Unit (translates logical
addresses to linear addresses and does protection checks), and the Paging Unit (translates linear
addresses to physical addresses, does page based protection checks, and contains a cache with
information for up to 32 most recently accessed pages).
The Intel486™ processor added more parallel execution capability by (basically) expanding the
Intel386™ processor’s Instruction Decode and Execution Units into five pipelined stages, where
each stage (when needed) operates in parallel with the others on up to five instructions in
different stages of execution. Each stage can do its work on one instruction in one clock, and so
the Intel486™ processor can execute as rapidly as one instruction per CPU clock. An 8-KByte
on-chip L1 cache was added to the Intel486™ processor to greatly increase the percent of
instructions that could execute at the scalar rate of one per clock: memory access instructions
were now included if the operand was in the L1 cache. The Intel486™ processor also for the
first time integrated the floating-point math Unit onto the same chip as the CPU (refer to Section
2.3., “Brief History of the Intel Architecture Floating-Point Unit”) and added new pins, bits, and
instructions to support more complex and powerful systems (L2 cache support and multipro-
cessor support).
Late in the Intel486™ processor generation, Intel incorporated features designed to support
energy savings and other system management capabilities into the IA mainstream with the
Intel486™ SL Enhanced processors. These features were developed in the Intel386™ SL and
Intel486™ SL processors, which were specialized for the rapidly growing battery-operated

1. Requires only one 32-bit address component to access anywhere in the address space.

2-2
 INTRODUCTION TO THE INTEL ARCHITECTURE

notebook PC market. The features include the new System Management Mode, triggered by its
own dedicated interrupt pin, which allows complex system management features (such as power
management of various subsystems within the PC), to be added to a system transparently to the
main operating system and all applications. The Stop Clock and Auto Halt Powerdown features
allow the CPU itself to execute at a reduced clock rate to save power, or to be shut down (with
state preserved) to save even more power.
The Intel Pentium® processor added a second execution pipeline to achieve superscalar perfor-
mance (two pipelines, known as u and v, together can execute two instructions per clock). The
on-chip L1 cache has also been doubled, with 8 KBytes devoted to code, and another 8 KBytes
devoted to data. The data cache uses the MESI protocol to support the more efficient write-back
mode, as well as the write-through mode that is used by the Intel486™ processor. Branch predic-
tion with an on-chip branch table has been added to increase performance in looping constructs.
Extensions have been added to make the virtual-8086 mode more efficient, and to allow for 4-
MByte as well as 4-KByte pages. The main registers are still 32 bits, but internal data paths of
128 and 256 bits have been added to speed internal data transfers, and the burstable external data
bus has been increased to 64 bits. The Advanced Programmable Interrupt Controller (APIC) has
been added to support systems with multiple Pentium® processors, and new pins and a special
mode (dual processing) has been designed in to support glueless two processor systems.
The Intel Pentium® Pro processor introduced “Dynamic Execution.” It has a three-way super-
scalar architecture, which means that it can execute three instructions per CPU clock. It does this
by incorporating even more parallelism than the Pentium® processor. The Pentium® Pro
processor provides Dynamic Execution (micro-data flow analysis, out-of-order execution, supe-
rior branch prediction, and speculative execution) in a superscalar implementation. Three
instruction decode units work in parallel to decode object code into smaller operations called
“micro-ops.” These go into an instruction pool, and (when interdependencies don’t prevent) can
be executed out of order by the five parallel execution units (two integer, two FPU and one
memory interface unit). The Retirement Unit retires completed micro-ops in their original
program order, taking account of any branches. The power of the Pentium® Pro processor is
further enhanced by its caches: it has the same two on-chip 8-KByte L1 caches as does the
Pentium® processor, and also has a 256-KByte L2 cache that is in the same package as, and
closely coupled to, the CPU, using a dedicated 64-bit (“backside”) full clock speed bus. The L1
cache is dual-ported, the L2 cache supports up to 4 concurrent accesses, and the 64-bit external
data bus is transaction-oriented, meaning that each access is handled as a separate request and
response, with numerous requests allowed while awaiting a response. These parallel features for
data access work with the parallel execution capabilities to provide a “non-blocking” architec-
ture in which the processor is more fully utilized and performance is enhanced. The Pentium®
Pro processor also has an expanded 36-bit address bus, giving a maximum physical address
space of 64 GBytes.
The Pentium® II processor added MMX™ instructions to the Pentium® Pro processor architec-
ture, incorporating the new slot 1 and slot 2 packaging techniques. These new packaging tech-
niques moved the L2 cache “off-chip” or “off-die”. The slot 1 and slot 2 package uses a single-
edge connector instead of a socket. The Pentium® II processor expanded the L1 data cache and
L1 instruction cache to 16 KBytes each. The Pentium® II processor has L2 cache sizes of 256
KBytes, 512 KBytes and 1 MByte or 2 MByte (slot 2 only). The slot 1 processor uses a “half
clock speed” backside bus while the slot 2 processor uses a “full clock speed” backside bus. The

2-3
INTRODUCTION TO THE INTEL ARCHITECTURE

Pentium® II processors utilize multiple low-power states such as AutoHALT, Stop-Grant, Sleep,
and Deep Sleep to conserve power during idle times.
The newest processor in the IA is the Pentium® III processor. It is based on the Pentium® Pro
and Pentium® II processors architectures. The Pentium® III processor introduces 70 new instruc-
tions to the IA instruction set. These instructions target existing functional units within the archi-
tecture as well as the new SIMD-floating-point unit. More detailed discussion of the new
features in the Pentium® Pro, Pentium® II, and Pentium® III processors is provided in Section
2.4., “Introduction to the P6 Family Processor’s Advanced Microarchitecture” and Section 2.5.,
“Detailed Description of the P6 FaMILY Processor Microarchitecture”. More detailed hardware
and architectural information on each of the generations of the IA family is available in the sepa-
rate data books for the processor generations (Section 1.5., “Related Literature” in Chapter 1,
About This Manual).

2.2. INCREASING INTEL ARCHITECTURE PERFORMANCE AND

MOORE’S LAW
In the mid-1960s, Intel Chairman of the Board Gordon Moore deduced a principle or “law”
which has continued to be true for over three decades: the computing power and the complexity
(or roughly, the number of transistors per CPU chip) of the silicon integrated circuit micropro-
cessor doubles every one to two years, and the cost per CPU chip is cut in half. This law is the
main explanation for the computer revolution, in which the IA plays such a significant role.

2-4
 INTRODUCTION TO THE INTEL ARCHITECTURE

The table below shows the dramatic increases in performance and transistor count of the IA
processors over their history, as predicted by Moore’s Law, and also summarizes the evolution
of other key features of the architecture.

Table 2-1. Processor Performance Over Time and Other Intel Architecture Key Features
No. of
Date of Perform Max. CPU Transis Main Extern. Max. Caches
Product -ance Frequency -tors on CPU Data Extern. in CPU
Intel Intro- in MIPs1 at Intro- the Die Register Bus Addr. Pack-
Processor duction duction Size2 Size2 Space age3
8086 1978 0.8 8 MHz 29 K 16 16 1 MB None
Intel 286 1982 2.7 12.5 MHz 134 K 16 16 16 MB Note 3
Intel386™ 1985 6.0 20 MHz 275 K 32 32 4 GB Note 3
DX
Intel486™ 1989 20 25 MHz 1.2 M 32 32 4 GB 8KB L1
DX
Pentium® 1993 100 60 MHz 3.1 M 32 64 4 GB 16KB L1
®
Pentium 1995 440 200 MHz 5.5 M 32 64 64 GB 16KB L1;
Pro 256KB or
512KB L2
Pentium II® 1997 466 266 7M 32 64 64 GB 32KB L1;
256KB or
512KB L2
Pentium® 1999 1000 500 8.2 M 32 GP 64 64 GB 32KB L1;
III 128 512KB L2
SIMD-FP

NOTES:
1. Performance here is indicated by Dhrystone MIPs (Millions of Instructions per Second) because even
though MIPs are no longer considered a preferred measure of CPU performance, they are the only
benchmarks that span all six generations of the IA. The MIPs and frequency values given here corre-
spond to the maximum CPU frequency available at product introduction.
2. Main CPU register size and external data bus size are given in bits. Note also that there are 8 and 16-bit
data registers in all of the CPUs, there are eight 80-bit registers in the FPUs integrated into the Intel386™
chip and beyond, and there are internal data paths that are 2 to 4 times wider than the external data bus
for each processor.
3. In addition to the large general-purpose caches listed in the table for the Intel486™ processor (8 KBytes
of combined code and data) and the Intel Pentium® and Pentium® Pro processors (8 KBytes each for
separate code cache and data cache), there are smaller special purpose caches. The Intel 286 has 6
byte descriptor caches for each segment register. The Intel386™ has 8 byte descriptor caches for each
segment register, and also a 32-entry, 4-way set associative Translation Lookaside Buffer (cache) to
store access information for recently used pages on the chip. The Intel486™ has the same caches
described for the Intel386™, as well as its 8K L1 general-purpose cache. The Intel Pentium® and Pen-
tium® Pro processors have their general-purpose caches, descriptor caches, and two Translation Looka-
side Buffers each (one for each 8K L1 cache). The Pentium® II and Pentium® III processors have the
same cache structure as the Pentium® Pro processor except that the size of each cache is 16K.

2-5
INTRODUCTION TO THE INTEL ARCHITECTURE

2.3. BRIEF HISTORY OF THE INTEL ARCHITECTURE FLOATING-

POINT UNIT
The IA Floating-Point Units (FPUs) before the Intel486™ lack the added efficiency of integra-
tion into the CPU, but have provided the option of greatly enhanced floating-point performance
since the beginning of the family. (Since the earlier FPUs were on separate chips, they were
often referred to as numeric processor extensions (NPXs) or math coprocessors (MCPs).) With
each succeeding generation, Intel has made significant increases in the power and flexibility of
the FPU, and yet has maintained complete upward compatibility. The Pentium® Pro Processor
offers compatibility with object code for 8087, Intel 287, Intel 387 DX, Intel 387 SX, and Intel
487 DX math coprocessors and the Intel486™ DX and Pentium® processors.
The 8087 numeric processor extension (NPX) was designed for use in 8086-family systems. The
8086 was the first microprocessor family to partition the processing unit to permit high-perfor-
mance numeric capabilities. The 8087 NPX for this processor family implemented a complete
numeric processing environment in compliance with an early proposal for IEEE Standard 754
for Binary Floating-Point Arithmetic.
With the Intel 287 coprocessor NPX, high-speed numeric computations were extended to 80286
high-performance multitasking and multi-user systems. Multiple tasks using the numeric
processor extension were afforded the full protection of the 80286 memory management and
protection features.
The Intel 387 DX and SX math coprocessors are Intel’s third generation numeric processors.
They implement the final IEEE Standard 754, adding new trigonometric instructions, and using
a new design and CHMOS-III process to allow higher clock rates and require fewer clocks per
instruction. Together, the Intel 387 math coprocessor with additional instructions and the
improved standard brought even more convenience and reliability to numeric programming and
made this convenience and reliability available to applications that need the high-speed and
large memory capacity of the 32-bit environment of the Intel386™ microprocessor.
The Intel486™ processor FPU is an on-chip equivalent of the Intel 387 DX math coprocessor
conforming to both IEEE Standard 754 and the more recent, generalized IEEE Standard 854.
Having the FPU on-chip results in a considerable performance improvement in numeric-inten-
sive computation.
The Pentium® processor FPU has been completely redesigned over the Intel486™ processor
FPU while maintaining conformance to both the IEEE Standard 754 and 854. Faster algorithms
provide at least three times the performance over the Intel486™ processor FPU for common
operations including ADD, MUL, and LOAD. Many applications can achieve five times the
performance of the Intel486™ processor FPU or more with instruction scheduling and pipelined
execution.

2.4. INTRODUCTION TO THE P6 FAMILY PROCESSOR’S

ADVANCED MICROARCHITECTURE
The P6 Family processors (introduced by Intel in 1995) represent the earliest implementation of
most recent processor in the IA family. Like its predecessor, the Pentium® processor (introduced

2-6
 INTRODUCTION TO THE INTEL ARCHITECTURE

by Intel in 1993), the Pentium® Pro processor, with its advanced superscalar microarchitecture,
sets an impressive performance standard. In designing the P6 Family processors, one of the
primary goals of the Intel chip architects was to exceed the performance of the Pentium®
processor significantly while still using the same 0.6-micrometer, four-layer, metal BICMOS
manufacturing process. Using the same manufacturing process as the Pentium® processor meant
that performance gains could only be achieved through substantial advances in the microarchi-
tecture.
The resulting P6 Family processor microarchitecture is a three-way superscalar, pipelined archi-
tecture. The term “three-way superscalar” means that using parallel processing techniques, the
processor is able on average to decode, dispatch, and complete execution of (retire) three
instructions per clock cycle. To handle this level of instruction throughput, the P6 Family
processors use a decoupled, 12-stage superpipeline that supports out-of-order instruction execu-
tion. Figure 2-1 shows a conceptual view of this pipeline, with the pipeline divided into four
processing units (the fetch/decode unit, the dispatch/execute unit, the retire unit, and the instruc-
tion pool). Instructions and data are supplied to these units through the bus interface unit.

System Bus
L2 Cache

Cache Bus

Bus Interface Unit

L1 Instruction L1 Data Cache

Cache

Fetch Load Store

Intel
Fetch/Decode Dispatch/ Architecture
Unit Execute Unit Retire Unit
Registers

Instruction
Pool

Figure 2-1. The Processing Units in the P6 Family Processor Microarchitecture

and Their Interface with the Memory Subsystem

To insure a steady supply of instructions and data to the instruction execution pipeline, the P6
Family processor microarchitecture incorporates two cache levels. The L1 cache provides an 8-

2-7
INTRODUCTION TO THE INTEL ARCHITECTURE

KByte instruction cache and an 8-KByte data cache, both closely coupled to the pipeline. The
L2 cache is a 256-KByte, 512-KByte, or 1-MByte static RAM that is coupled to the core
processor through a full clock-speed 64-bit cache bus.
The centerpiece of the P6 Family processor microarchitecture is an innovative out-of-order
execution mechanism called “dynamic execution.” Dynamic execution incorporates three data-
processing concepts:
• Deep branch prediction.
• Dynamic data flow analysis.
• Speculative execution.
Branch prediction is a concept found in most mainframe and high-speed microprocessor archi-
tectures. It allows the processor to decode instructions beyond branches to keep the instruction
pipeline full. In the P6 Family processors, the instruction fetch/decode unit uses a highly opti-
mized branch prediction algorithm to predict the direction of the instruction stream through
multiple levels of branches, procedure calls, and returns.
Dynamic data flow analysis involves real-time analysis of the flow of data through the processor
to determine data and register dependencies and to detect opportunities for out-of-order instruc-
tion execution. The P6 Family processors dispatch/execute unit can simultaneously monitor
many instructions and execute these instructions in the order that optimizes the use of the
processor’s multiple execution units, while maintaining data integrity. This out-of-order execu-
tion keeps the execution units busy even when cache misses and data dependencies among
instructions occur.
Speculative execution refers to the processor’s ability to execute instructions ahead of the
program counter but ultimately to commit the results in the order of the original instruction
stream. To make speculative execution possible, the P6 Family processors microarchitecture
decouples the dispatching and executing of instructions from the commitment of results. The
processor’s dispatch/execute unit uses data-flow analysis to execute all available instructions in
the instruction pool and store the results in temporary registers. The retirement unit then linearly
searches the instruction pool for completed instructions that no longer have data dependencies
with other instructions or unresolved branch predictions. When completed instructions are
found, the retirement unit commits the results of these instructions to memory and/or the IA
registers (the processor’s eight general-purpose registers and eight floating-point unit data regis-
ters) in the order they were originally issued and retires the instructions from the instruction
pool.
Through deep branch prediction, dynamic data-flow analysis, and speculative execution,
dynamic execution removes the constraint of linear instruction sequencing between the tradi-
tional fetch and execute phases of instruction execution. It allows instructions to be decoded
deep into multi-level branches to keep the instruction pipeline full. It promotes out-of-order
instruction execution to keep the processor’s six instruction execution units running at full
capacity. And finally, it commits the results of executed instructions in original program order
to maintain data integrity and program coherency.
The following section describes the P6 Family processor microarchitecture in greater detail. The
Pentium® Pro processor architecture is the base architecture for the processors that followed it.

2-8
 INTRODUCTION TO THE INTEL ARCHITECTURE

The Pentium® II processor and now the Pentium® III processor are based on the Pentium® Pro
processor architecture. Changes or enhancements to the Pentium® Pro processor architecture are
noted where appropriate.

2.5. DETAILED DESCRIPTION OF THE P6 FAMILY PROCESSOR

MICROARCHITECTURE
Figure 2-2 shows a functional block diagram of the P6 Family processor microarchitecture. In
this diagram, the following blocks make up the four processing units and the memory subsystem
shown in Figure 2-1:
• Memory subsystem—System bus, L2 cache, bus interface unit, instruction cache (L1),
data cache unit (L1), memory interface unit, and memory reorder buffer.
• Fetch/decode unit—Instruction fetch unit, branch target buffer, instruction decoder,
microcode sequencer, and register alias table.
• Instruction pool—Reorder buffer
• Dispatch/execute unit—Reservation station, two integer units, one x87 floating-point unit,
two address generation units, and two SIMD floating-point units.
• Retire unit—Retire unit and retirement register file.

2.5.1. Memory Subsystem

The memory subsystem for the P6 Family processor consists of main system memory, the
primary cache (L1), and the secondary cache (L2). The bus interface unit accesses system
memory through the external system bus. This 64-bit bus is a transaction-oriented bus, meaning
that each bus access is handled as separate request and response operations. While the bus inter-
face unit is waiting for a response to one bus request, it can issue numerous additional requests.
The bus interface unit accesses the close-coupled L2 cache through a 64-bit cache bus. This bus
is also transactional oriented, supporting up to four concurrent cache accesses, and operates at
the full clock speed of the processor.
Access to the L1 caches is through internal buses, also at full clock speed. The 8-KByte L1
instruction cache is four-way set associative; the 8-KByte L1 data cache is dual-ported and two-
way set associative, supporting one load and one store operation per cycle.
Coherency between the caches and system memory are maintained using the MESI (modified,
exclusive, shared, invalid) cache protocol. This protocol fosters cache coherency in single- and
multiple-processor systems. It is also able to detect coherency problems created by self-modi-
fying code.
Memory requests from the processor’s execution units go through the memory interface unit and
the memory order buffer. These units have been designed to support a smooth flow of memory
access requests through the cache and system memory hierarchy to prevent memory access

2-9
INTRODUCTION TO THE INTEL ARCHITECTURE

blocking. The L1 data cache automatically forwards a cache miss on to the L2 cache, and then,
if necessary, the bus interface unit forwards an L2 cache miss to system memory.

System Bus (External) L2 Cache

Cache Bus

Bus Interface Unit

Next IP
Instruction Fetch Unit Instruction Cache (L1) Unit
Memory
Branch
Reorder
Instruction Decoder Target
Buffer
Buffer
Simple Simple Complex
Instuction Instuction Instuction
Decoder Decoder Decoder Microcode From
Instruction Integer
Sequencer Unit

Register Alias Table

Retirement
Retirement Unit Register File Data Cache
(Intel Arch. Unit (L1)
Reorder Buffer (Instruction Pool) Registers)

Reservation Station

Execution Unit

SIMD FP Floating- Memory

Integer Integer
Unit Point Unit Interface
Unit Unit
(FPU) (FPU) Unit

Internal Data-Results Buses

Figure 2-2. Functional Block Diagram of the P6 Family Processor Microarchitecture

Memory requests to the L2 cache or system memory go through the memory reorder buffer,
which functions as a scheduling and dispatch station. This unit keeps track of all memory
requests and is able to reorder some requests to prevent blocks and improve throughput. For
example, the memory reorder buffer allows loads to pass stores. It also issues speculative loads.
(Stores are always dispatched in order, and speculative stores are never issued.)

2-10
 INTRODUCTION TO THE INTEL ARCHITECTURE

2.5.2. Fetch/Decode Unit

The fetch/decode unit reads a stream of IA instructions from the L1 instruction cache and
decodes them into a series of micro-operations called “micro-ops.” This micro-op stream (still
in the order of the original instruction stream) is then sent to the instruction pool.
The instruction fetch unit fetches one 32-byte cache line per clock from the instruction cache. It
marks the beginning and end of the IA instructions in the cache lines and transmits 16 aligned
bytes to the decoder.
The instruction fetch unit computes the instruction pointer, based on inputs from the branch
target buffer, the exception/interrupt status, and branch-misprediction indications from the
integer execution units. The most important part of this process is the branch prediction
performed by the branch target buffer. Using an extension of Yeh’s algorithm, the 512-entry
branch target buffer looks many instructions ahead of the retirement program counter. Within
this instruction window there may be numerous branches, procedure calls, and returns that must
be correctly predicted if the dispatch/execute unit is to do useful work.
The instruction decoder contains three parallel decoders: two simple-instruction decoders and
one complex instruction decoder. Each decoder converts an IA instruction into one or more
triadic micro-ops (two logical sources and one logical destination per micro-op). Micro-ops are
primitive instructions that are executed by the processor’s six parallel execution units.
Many IA instructions are converted directly into single micro-ops by the simple instruction
decoders, and some instructions are decoded into from one to four micro-ops. The more
complex IA instructions are decoded into sequences of preprogrammed micro-ops obtained
from the microcode instruction sequencer. The instruction decoders also handle the decoding of
instruction prefixes and looping operations. The instruction decoder can generate up to six
micro-ops per clock cycle (one each for the simple instruction decoders and four for the complex
instruction decoder).
The IA’s register set can cause resource stalls due to register dependencies. To solve this
problem, the processor provides 40 internal, general-purpose registers, which are used for the
actual computations. These registers can handle both integer and floating-point values. To allo-
cate the internal registers, the enqueued micro-ops from the instruction decoder are sent to the
register alias table unit, where references to the logical IA registers are converted into internal
physical register references.
In the final step of the decoding process, the allocator in the register alias table unit adds status
bits and flags to the micro-ops to prepare them for out-of-order execution and sends the resulting
micro-ops to the instruction pool.

2.5.3. Instruction Pool (Reorder Buffer)

Prior to entering the instruction pool (known formally as the reorder buffer), the micro-op
instruction stream is in the same order as the IA instruction stream that was sent to the instruc-
tion decoder. No reordering of instructions has taken place.
The reorder buffer is an array of content-addressable memory, arranged into 40 micro-op regis-
ters. It contains micro-ops that are waiting to be executed, as well as those that have already been

2-11
INTRODUCTION TO THE INTEL ARCHITECTURE

executed but not yet committed to machine state. The dispatch/execute unit can execute instruc-
tions from the reorder buffer in any order.

2.5.4. Dispatch/Execute Unit

The dispatch/execute unit is an out-of-order unit that schedules and executes the micro-ops
stored in the reorder buffer according to data dependencies and resource availability and tempo-
rarily stores the results of these speculative executions.
The scheduling and dispatching of micro-ops from the reorder buffer is handled by the reserva-
tion station. It continuously scans the reorder buffer for micro-ops that are ready to be executed
(that is, all the source operands are available) and dispatches them to the available execution
units. The results of a micro-op execution are returned to the reorder buffer and stored along
with the micro-op until it is retired. This scheduling and dispatching process supports classic
out-of-order execution, where micro-ops are dispatched to the execution units strictly according
to data-flow constraints and execution resource availability, without regard to the original
ordering of the instructions. When two or more micro-ops of the same type (for example, integer
operations) are available at the same time, they are executed in a pseudo FIFO order in the
reorder buffer.
Execution of micro-ops is handled by two integer units, two floating-point units, and one
memory-interface unit, allowing up to five micro-ops to be scheduled per clock.
The two integer units can handle two integer micro-ops in parallel. One of the integer units is
designed to handle branch micro-ops. This unit has the ability to detect branch mispredictions
and signal the branch target buffer to restart the pipeline. This operation is handled as follows.
The instruction decoder tags each branch micro-op with both branch destination addresses (the
predicted destination and the fall-through destination). When the integer unit executes the
branch micro-op, it is able to determine whether the predicted or the fall-through destination was
taken. If the predicted branch is taken, then speculatively executed micro-ops are marked usable
and execution continues along the predicted instruction path. If the predicted branch was not
taken, a jump execution unit in the integer unit changes the status of all of the micro-ops
following the branch to remove them from the instruction pool. It then provides the proper
branch destination to the branch target buffer, which in turn restarts the pipeline from the new
target address.
The memory interface unit handles load and store micro-ops. A load access only needs to
specify the memory address, so it can be encoded in one micro-op. A store access needs to
specify both an address and the data to be written, so it is encoded in two micro-ops. The part of
the memory interface unit that handles stores has two ports allowing it to process the address
and the data micro-op in parallel. The memory interface unit can thus execute both a load and a
store in parallel in one clock cycle.
The floating-point execution units are similar to those found in the Pentium® processor. Several
new floating-point instructions have been added to the P6 Family processor to streamline condi-
tional branches and moves.

2-12
 INTRODUCTION TO THE INTEL ARCHITECTURE

2.5.5. Retirement Unit

The retirement unit commits the results of speculatively executed micro-ops to permanent
machine state and removes the micro-ops from the reorder buffer. Like the reservation station,
the retirement unit continuously checks the status of micro-ops in the reorder buffer, looking for
ones that have been executed and no longer have any dependencies with other micro-ops in the
instruction pool. It then retires completed micro-ops in their original program order, taking into
accounts interrupts, exceptions, breakpoints, and branch mispredictions.
The retirement unit can retire three micro-ops per clock. In retiring a micro-op, it writes the
results to the retirement register file and/or memory. The retirement register file contains the IA
registers (eight general-purpose registers and eight floating-point data registers). After the
results have been committed to machine state, the micro-op is removed from the reorder buffer.

2-13
INTRODUCTION TO THE INTEL ARCHITECTURE

2-14
 3
Basic Execution
Environment
 CHAPTER 3
BASIC EXECUTION ENVIRONMENT

This chapter describes the basic execution environment of an Intel Architecture (IA) processor
as seen by assembly-language programmers. It describes how the processor executes instruc-
tions and how it stores and manipulates data. The parts of the execution environment described
here include memory (the address space), the general-purpose data registers, the segment regis-
ters, the EFLAGS register, and the instruction pointer register.
The execution environment for the floating-point unit (FPU) is described in Chapter 7, Floating-
Point Unit.

3.1. MODES OF OPERATION

The IA supports three operating modes: protected mode, real-address mode, and system
management mode. The operating mode determines which instructions and architectural
features are accessible:
• Protected mode. The native state of the processor. In this mode all instructions and archi-
tectural features are available, providing the highest performance and capability. This is
the recommended mode for all new applications and operating systems.
Among the capabilities of protected mode is the ability to directly execute “real-address
mode” 8086 software in a protected, multitasking environment. This feature is called
virtual-8086 mode, although it is not actually a processor mode. Virtual-8086 mode is
actually a protected mode attribute that can be enabled for any task.
• Real-address mode. Provides the programming environment of the Intel 8086 processor
with a few extensions (such as the ability to switch to protected or system management
mode). The processor is placed in real-address mode following power-up or a reset.
• System management mode. A standard architectural feature unique to all Intel
processors, beginning with the Intel386™ SL processor. This mode provides an operating
system or executive with a transparent mechanism for implementing platform-specific
functions such as power management and system security. The processor enters SMM
when the external SMM interrupt pin (SMI#) is activated or an SMI is received from the
advanced programmable interrupt controller (APIC). In SMM, the processor switches to a
separate address space while saving the entire context of the currently running program or
task. SMM-specific code may then be executed transparently. Upon returning from SMM,
the processor is placed back into its state prior to the system management interrupt.
The basic execution environment is the same for each of these operating modes, as is described
in the remaining sections of this chapter.

3-1
BASIC EXECUTION ENVIRONMENT

3.2. OVERVIEW OF THE BASIC EXECUTION ENVIRONMENT

Any program or task running on an IA processor is given a set of resources for executing instruc-
tions and for storing code, data, and state information. These resources (shown in Figure 3-1)
include an address space of up to 236 bytes, a set of general data registers, a set of segment regis-
ters, and a set of status and control registers. When a program calls a procedure, a procedure
stack is added to the execution environment. (Procedure calls and the procedure stack imple-
mentation are described in Chapter 4, Procedure Calls, Interrupts, and Exceptions.)

236 −1

Eight 32-bit General-Purpose

Registers Registers

Six 16-bit Segment Registers Address

Registers Space*

32-bits EFLAGS Register

32-bits EIP (Instruction

Pointer Register)

*The address space can be flat or segmented.

* Physical address space is 2**36-1
* Linear address space is 2**32-1 0
Figure 3-1. P6 Family Processor Basic Execution Environment

3.3. MEMORY ORGANIZATION

The memory that the processor addresses on its bus is called physical memory. Physical
memory is organized as a sequence of 8-bit bytes. Each byte is assigned a unique address, called
a physical address. The physical address space ranges from zero to a maximum of 236 – 1
(64 gigabytes).
Virtually any operating system or executive designed to work with an IA processor will use the
processor’s memory management facilities to access memory. These facilities provide features
such as segmentation and paging, which allow memory to be managed efficiently and reliably.
Memory management is described in detail in Chapter 3, Protected-Mode Memory Manage-
ment, of the Intel Architecture Software Developer’s Manual, Volume 3. The following para-
graphs describe the basic methods of addressing memory when memory management is used.
When employing the processor’s memory management facilities, programs do not directly
address physical memory. Instead, they access memory using any of three memory models: flat,
segmented, or real-address mode.
With the flat memory model (refer to Figure 3-2), memory appears to a program as a single,
continuous address space, called a linear address space. Code (a program’s instructions), data,

3-2
 BASIC EXECUTION ENVIRONMENT

and the procedure stack are all contained in this address space. The linear address space is byte
addressable, with addresses running contiguously from 0 to 236 − 1. An address for any byte in
the linear address space is called a linear address.

Flat Model
Linear Address

Linear
Address
Space*

Segmented Model

Segments

Offset Linear
Address
Space*
Logical
Address Segment Selector

Real-Address Mode Model

Linear Address
Offset Space Divided
Into Equal
Sized Segments
Logical
Address Segment Selector

* The linear address space

can be paged when using the
flat or segmented model.

Figure 3-2. Three Memory Management Models

With the segmented memory model, memory appears to a program as a group of independent
address spaces called segments. When using this model, code, data, and stacks are typically
contained in separate segments. To address a byte in a segment, a program must issue a logical
address, which consists of a segment selector and an offset. (A logical address is often referred
to as a far pointer.) The segment selector identifies the segment to be accessed and the offset
identifies a byte in the address space of the segment. The programs running on an IA processor
can address up to 16,383 segments of different sizes and types, and each segment can be as large
as 236 bytes.
Internally, all the segments that are defined for a system are mapped into the processor’s linear
address space. The processor translates each logical address into a linear address to access a
memory location. This translation is transparent to the application program.

3-3
BASIC EXECUTION ENVIRONMENT

The primary reason for using segmented memory is to increase the reliability of programs and
systems. For example, placing a program’s stack in a separate segment prevents the stack from
growing into the code or data space and overwriting instructions or data, respectively. Placing
the operating system’s or executive’s code, data, and stack in separate segments also protects
them from the application program and vice versa.
With either the flat or segmented model, the IA provides facilities for dividing the linear address
space into pages and mapping the pages into virtual memory. If an operating system/executive
uses the IA’s paging mechanism, the existence of the pages is transparent to an application
program.
The real-address mode model uses the memory model for the Intel 8086 processor, the first IA
processor. It was provided in all the subsequent IA processors for compatibility with existing
programs written to run on the Intel 8086 processor. The real-address mode uses a specific
implementation of segmented memory in which the linear address space for the program and the
operating system/executive consists of an array of segments of up to 64 Kbytes in size each. The
maximum size of the linear address space in real-address mode is 220 bytes. (Refer to Chapter
16, 8086 Emulation, in the Intel Architecture Software Developer’s Manual, Volume 3, for more
information on this memory model.)

3.4. MODES OF OPERATION

When writing code for the Pentium® Pro processor, a programmer needs to know the operating
mode the processor is going to be in when executing the code and the memory model being used.
The relationship between operating modes and memory models is as follows:
• Protected mode. When in protected mode, the processor can use any of the memory
models described in this section. (The real-addressing mode memory model is ordinarily
used only when the processor is in the virtual-8086 mode.) The memory model used
depends on the design of the operating system or executive. When multitasking is imple-
mented, individual tasks can use different memory models.
• Real-address mode. When in real-address mode, the processor only supports the real-
address mode memory model.
• System management mode. When in SMM, the processor switches to a separate address
space, called the system management RAM (SMRAM). The memory model used to
address bytes in this address space is similar to the real-address mode model. (Refer to
Chapter 12, System Management Mode (SMM), in the Intel Architecture Software
Developer’s Manual, Volume 3, for more information on the memory model used in
SMM.)

3.5. 32-BIT VS. 16-BIT ADDRESS AND OPERAND SIZES

The processor can be configured for 32-bit or 16-bit address and operand sizes. With 32-bit
address and operand sizes, the maximum linear address or segment offset is FFFFFFFFH (232),
and operand sizes are typically 8 bits or 32 bits. With 16-bit address and operand sizes, the

3-4
 BASIC EXECUTION ENVIRONMENT

maximum linear address or segment offset is FFFFH (216), and operand sizes are typically 8 bits
or 16 bits.
When using 32-bit addressing, a logical address (or far pointer) consists of a 16-bit segment
selector and a 32-bit offset; when using 16-bit addressing, it consists of a 16-bit segment selector
and a 16-bit offset.
Instruction prefixes allow temporary overrides of the default address and/or operand sizes from
within a program.
When operating in protected mode, the segment descriptor for the currently executing code
segment defines the default address and operand size. A segment descriptor is a system data
structure not normally visible to application code. Assembler directives allow the default
addressing and operand size to be chosen for a program. The assembler and other tools then set
up the segment descriptor for the code segment appropriately.
When operating in real-address mode, the default addressing and operand size is 16 bits. An
address-size override can be used in real-address mode to enable 32-bit addressing; however, the
maximum allowable 32-bit address is still 0000FFFFH (216).

3.6. REGISTERS
The processor provides 16 registers for use in general system and application programing. As
shown in Figure 3-3, these registers can be grouped as follows:
• General-purpose data registers. These eight registers are available for storing operands
and pointers.
• Segment registers. These registers hold up to six segment selectors.
• Status and control registers. These registers report and allow modification of the state of
the processor and of the program being executed.

3-5
BASIC EXECUTION ENVIRONMENT

3.6.1. General-Purpose Data Registers

The 32-bit general-purpose data registers EAX, EBX, ECX, EDX, ESI, EDI, EBP, and ESP are
provided for holding the following items:
• Operands for logical and arithmetic operations
• Operands for address calculations
• Memory pointers
Although all of these registers are available for general storage of operands, results, and
pointers, caution should be used when referencing the ESP register. The ESP register holds the
stack pointer and as a general rule should not be used for any other purpose.

General-Purpose Registers
31 0
EAX
EBX
ECX
EDX
ESI
EDI
EBP
ESP

Segment Registers
15 0
CS
DS
SS
ES
FS
GS

31 Status and Control Registers 0

EFLAGS
31 0
EIP

Figure 3-3. Application Programming Registers

Many instructions assign specific registers to hold operands. For example, string instructions
use the contents of the ECX, ESI, and EDI registers as operands. When using a segmented
memory model, some instructions assume that pointers in certain registers are relative to

3-6
 BASIC EXECUTION ENVIRONMENT

specific segments. For instance, some instructions assume that a pointer in the EBX register
points to a memory location in the DS segment.
The special uses of general-purpose registers by instructions are described in Chapter 6, Instruc-
tion Set Summary, in this volume, and Chapter 3, Instruction Set Reference in the Intel Architec-
ture Software Developer’s Manual, Volume 2. The following is a summary of these special uses:
• EAX—Accumulator for operands and results data.
• EBX—Pointer to data in the DS segment.
• ECX—Counter for string and loop operations.
• EDX—I/O pointer.
• ESI—Pointer to data in the segment pointed to by the DS register; source pointer for string
operations.
• EDI—Pointer to data (or destination) in the segment pointed to by the ES register;
destination pointer for string operations.
• ESP—Stack pointer (in the SS segment).
• EBP—Pointer to data on the stack (in the SS segment).
As shown in Figure 3-4, the lower 16 bits of the general-purpose registers map directly to the
register set found in the 8086 and Intel 286 processors and can be referenced with the names
AX, BX, CX, DX, BP, SP, SI, and DI. Each of the lower two bytes of the EAX, EBX, ECX, and
EDX registers can be referenced by the names AH, BH, CH, and DH (high bytes) and AL, BL,
CL, and DL (low bytes).

General-Purpose Registers
31 16 15 8 7 0 16-bit 32-bit
AH AL AX EAX
BH BL BX EBX
CH CL CX ECX
DH DL DX EDX
BP EBP
SI ESI
DI EDI
SP ESP

Figure 3-4. Alternate General-Purpose Register Names

3.6.2. Segment Registers

The segment registers (CS, DS, SS, ES, FS, and GS) hold 16-bit segment selectors. A segment
selector is a special pointer that identifies a segment in memory. To access a particular segment

3-7
BASIC EXECUTION ENVIRONMENT

in memory, the segment selector for that segment must be present in the appropriate segment
register.
When writing application code, programmers generally create segment selectors with assembler
directives and symbols. The assembler and other tools then create the actual segment selector
values associated with these directives and symbols. If writing system code, programmers may
need to create segment selectors directly. (A detailed description of the segment-selector data
structure is given in Chapter 3, Protected-Mode Memory Management, of the Intel Architecture
Software Developer’s Manual, Volume 3.)
How segment registers are used depends on the type of memory management model that the
operating system or executive is using. When using the flat (unsegmented) memory model, the
segment registers are loaded with segment selectors that point to overlapping segments, each of
which begins at address 0 of the linear address space (as shown in Figure 3-5). These overlap-
ping segments then comprise the linear address space for the program. (Typically, two overlap-
ping segments are defined: one for code and another for data and stacks. The CS segment
register points to the code segment and all the other segment registers point to the data and stack
segment.)
When using the segmented memory model, each segment register is ordinarily loaded with a
different segment selector so that each segment register points to a different segment within the
linear address space (as shown in Figure 3-6). At any time, a program can thus access up to six
segments in the linear-address space. To access a segment not pointed to by one of the segment
registers, a program must first load the segment selector for the segment to be accessed into a
segment register.

Linear Address
Space for Program

Segment Registers Overlapping

Segments
CS of up to
4G Bytes
DS
Beginning at
SS
Address 0
ES
FS
GS
The segment selector in
each segment register
points to an overlapping
segment in the linear
address space.

Figure 3-5. Use of Segment Registers for Flat Memory Model

3-8
 BASIC EXECUTION ENVIRONMENT

Code
Segment
Segment Registers
Data
CS Segment
DS Stack
SS Segment
ES All segments
FS are mapped
GS to the same
linear-address
space
Data
Segment
Data
Segment
Data
Segment

Figure 3-6. Use of Segment Registers in Segmented Memory Model

Each of the segment registers is associated with one of three types of storage: code, data, or
stack). For example, the CS register contains the segment selector for the code segment, where
the instructions being executed are stored. The processor fetches instructions from the code
segment, using a logical address that consists of the segment selector in the CS register and the
contents of the EIP register. The EIP register contains the linear address within the code segment
of the next instruction to be executed. The CS register cannot be loaded explicitly by an appli-
cation program. Instead, it is loaded implicitly by instructions or internal processor operations
that change program control (such as, procedure calls, interrupt handling, or task switching).
The DS, ES, FS, and GS registers point to four data segments. The availability of four data
segments permits efficient and secure access to different types of data structures. For example,
four separate data segments might be created: one for the data structures of the current module,
another for the data exported from a higher-level module, a third for a dynamically created data
structure, and a fourth for data shared with another program. To access additional data segments,
the application program must load segment selectors for these segments into the DS, ES, FS, and
GS registers, as needed.
The SS register contains the segment selector for a stack segment, where the procedure stack is
stored for the program, task, or handler currently being executed. All stack operations use the
SS register to find the stack segment. Unlike the CS register, the SS register can be loaded
explicitly, which permits application programs to set up multiple stacks and switch among them.
Refer to Section 3.3., “Memory Organization” for an overview of how the segment registers are
used in real-address mode.

3-9
BASIC EXECUTION ENVIRONMENT

The four segment registers CS, DS, SS, and ES are the same as the segment registers found in
the Intel 8086 and Intel 286 processors and the FS and GS registers were introduced into the IA
with the Intel386™ family of processors.

3.6.3. EFLAGS Register

The 32-bit EFLAGS register contains a group of status flags, a control flag, and a group of
system flags. Figure 3-7 defines the flags within this register. Following initialization of the
processor (either by asserting the RESET pin or the INIT pin), the state of the EFLAGS register
is 00000002H. Bits 1, 3, 5, 15, and 22 through 31 of this register are reserved. Software should
not use or depend on the states of any of these bits.
Some of the flags in the EFLAGS register can be modified directly, using special-purpose
instructions (described in the following sections). There are no instructions that allow the whole
register to be examined or modified directly. However, the following instructions can be used to
move groups of flags to and from the procedure stack or the EAX register: LAHF, SAHF,
PUSHF, PUSHFD, POPF, and POPFD. After the contents of the EFLAGS register have been
transferred to the procedure stack or EAX register, the flags can be examined and modified using
the processor’s bit manipulation instructions (BT, BTS, BTR, and BTC).
When suspending a task (using the processor’s multitasking facilities), the processor automati-
cally saves the state of the EFLAGS register in the task state segment (TSS) for the task being
suspended. When binding itself to a new task, the processor loads the EFLAGS register with
data from the new task’s TSS.
When a call is made to an interrupt or exception handler procedure, the processor automatically
saves the state of the EFLAGS registers on the procedure stack. When an interrupt or exception
is handled with a task switch, the state of the EFLAGS register is saved in the TSS for the task
being suspended.

3-10
 BASIC EXECUTION ENVIRONMENT

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
I
V V
N O O D I T S Z P C
0 0 0 0 0 0 0 0 0 0 I I I A V R 0 T
A
D C M F P F F F F F F 0 F 0 F 1 F
P F
L

X ID Flag (ID)
X Virtual Interrupt Pending (VIP)
X Virtual Interrupt Flag (VIF)
X Alignment Check (AC)
X Virtual-8086 Mode (VM)
X Resume Flag (RF)
X Nested Task (NT)
X I/O Privilege Level (IOPL)
S Overflow Flag (OF)
C Direction Flag (DF)
X Interrupt Enable Flag (IF)
X Trap Flag (TF)
S Sign Flag (SF)
S Zero Flag (ZF)
S Auxiliary Carry Flag (AF)
S Parity Flag (PF)
S Carry Flag (CF)

S Indicates a Status Flag

C Indicates a Control Flag
X Indicates a System Flag

Reserved bit positions. DO NOT USE.

Always set to values previously read.

Figure 3-7. EFLAGS Register

As the IA has evolved, flags have been added to the EFLAGS register, but the function and
placement of existing flags have remained the same from one family of the IA processors to the
next. As a result, code that accesses or modifies these flags for one family of IA processors
works as expected when run on later families of processors.

3-11
BASIC EXECUTION ENVIRONMENT

3.6.3.1. STATUS FLAGS

The status flags (bits 0, 2, 4, 6, 7, and 11) of the EFLAGS register indicate the results of arith-
metic instructions, such as the ADD, SUB, MUL, and DIV instructions. The functions of the
status flags are as follows:
CF (bit 0) Carry flag. Set if an arithmetic operation generates a carry or a
borrow out of the most-significant bit of the result; cleared otherwise.
This flag indicates an overflow condition for unsigned-integer arith-
metic. It is also used in multiple-precision arithmetic.
PF (bit 2) Parity flag. Set if the least-significant byte of the result contains an
even number of 1 bits; cleared otherwise.
AF (bit 4) Adjust flag. Set if an arithmetic operation generates a carry or a
borrow out of bit 3 of the result; cleared otherwise. This flag is used
in binary-coded decimal (BCD) arithmetic.
ZF (bit 6) Zero flag. Set if the result is zero; cleared otherwise.
SF (bit 7) Sign flag. Set equal to the most-significant bit of the result, which is
the sign bit of a signed integer. (0 indicates a positive value and 1
indicates a negative value.)
OF (bit 11) Overflow flag. Set if the integer result is too large a positive number
or too small a negative number (excluding the sign-bit) to fit in the
destination operand; cleared otherwise. This flag indicates an over-
flow condition for signed-integer (two’s complement) arithmetic.
Of these status flags, only the CF flag can be modified directly, using the STC, CLC, and CMC
instructions. Also the bit instructions (BT, BTS, BTR, and BTC) copy a specified bit into the CF
flag.
The status flags allow a single arithmetic operation to produce results for three different data
types: unsigned integers, signed integers, and BCD integers. If the result of an arithmetic oper-
ation is treated as an unsigned integer, the CF flag indicates an out-of-range condition (carry or
a borrow); if treated as a signed integer (two’s complement number), the OF flag indicates a
carry or borrow; and if treated as a BCD digit, the AF flag indicates a carry or borrow. The SF
flag indicates the sign of a signed integer. The ZF flag indicates either a signed- or an unsigned-
integer zero.
When performing multiple-precision arithmetic on integers, the CF flag is used in conjunction
with the add with carry (ADC) and subtract with borrow (SBB) instructions to propagate a carry
or borrow from one computation to the next.
The condition instructions Jcc (jump on condition code cc), SETcc (byte set on condition code
cc), LOOPcc, and CMOVcc (conditional move) use one or more of the status flags as condition
codes and test them for branch, set-byte, or end-loop conditions.

3-12
 BASIC EXECUTION ENVIRONMENT

3.6.3.2. DF FLAG
The direction flag (DF, located in bit 10 of the EFLAGS register) controls the string instructions
(MOVS, CMPS, SCAS, LODS, and STOS). Setting the DF flag causes the string instructions to
auto-decrement (that is, to process strings from high addresses to low addresses). Clearing the
DF flag causes the string instructions to auto-increment (process strings from low addresses
to high addresses).
The STD and CLD instructions set and clear the DF flag, respectively.

3.6.4. System Flags and IOPL Field

The system flags and IOPL field in the EFLAGS register control operating-system or executive
operations. They should not be modified by application programs. The functions of the
system flags are as follows:
IF (bit 9) Interrupt enable flag. Controls the response of the processor to
maskable interrupt requests. Set to respond to maskable interrupts;
cleared to inhibit maskable interrupts.
TF (bit 8) Trap flag. Set to enable single-step mode for debugging; clear to
disable single-step mode.
IOPL (bits 12, 13) I/O privilege level field. Indicates the I/O privilege level of the currently
running program or task. The current privilege level (CPL) of the
currently running program or task must be less than or equal to the
I/O privilege level to access the I/O address space. This field can only
be modified by the POPF and IRET instructions when operating at a
CPL of 0.
NT (bit 14) Nested task flag. Controls the chaining of interrupted and called
tasks. Set when the current task is linked to the previously executed
task; cleared when the current task is not linked to another task.
RF (bit 16) Resume flag. Controls the processor’s response to debug exceptions.
VM (bit 17) Virtual-8086 mode flag. Set to enable virtual-8086 mode; clear to
return to protected mode.
AC (bit 18) Alignment check flag. Set this flag and the AM bit in the CR0
register to enable alignment checking of memory references; clear
the AC flag and/or the AM bit to disable alignment checking.
VIF (bit 19) Virtual interrupt flag. Virtual image of the IF flag. Used in
conjunction with the VIP flag. (To use this flag and the VIP flag the
virtual mode extensions are enabled by setting the VME flag in
control register CR4.)
VIP (bit 20) Virtual interrupt pending flag. Set to indicate pending interrupts;
or clear when no interrupts are pending. (Software sets and clears this

3-13
BASIC EXECUTION ENVIRONMENT

flag; the processor only reads it.) Used in conjunction with the VIF
flag.
ID (bit 21) Identification flag. The ability of a program to set or clear this flag
indicates support for the CPUID instruction.
Refer to Chapter 3, Protected-Mode Memory Management, in the Intel Architecture Software
Developer’s Manual, Volume 3, for a detail description of these flags.

3.7. INSTRUCTION POINTER

The instruction pointer (EIP) register contains the offset in the current code segment for the next
instruction to be executed. It is advanced from one instruction boundary to the next in straight-
line code or it is moved ahead or backwards by a number of instructions when executing JMP,
Jcc, CALL, RET, and IRET instructions.
The EIP register cannot be accessed directly by software; it is controlled implicitly by control-
transfer instructions (such as JMP, Jcc, CALL, and RET), interrupts, and exceptions. The only
way to read the EIP register is to execute a CALL instruction and then read the value of the
return instruction pointer from the procedure stack. The EIP register can be loaded indirectly by
modifying the value of a return instruction pointer on the procedure stack and executing a return
instruction (RET or IRET). Refer to Section 4.2.4.2., “Return Instruction Pointer” in Chapter 4,
Procedure Calls, Interrupts, and Exceptions.
All IA processors prefetch instructions. Because of instruction prefetching, an instruction
address read from the bus during an instruction load does not match the value in the EIP register.
Even though different processor generations use different prefetching mechanisms, the function
of EIP register to direct program flow remains fully compatible with all software written to run
on IA processors.

3.8. OPERAND-SIZE AND ADDRESS-SIZE ATTRIBUTES

When the processor is executing in protected mode, every code segment has a default operand-
size attribute and address-size attribute. These attributes are selected with the D (default size)
flag in the segment descriptor for the code segment (refer to Chapter 3, Protected-Mode Memory
Management, in the Intel Architecture Software Developer’s Manual, Volume 3). When the D
flag is set, the 32-bit operand-size and address-size attributes are selected; when the flag is clear,
the 16-bit size attributes are selected. When the processor is executing in real-address mode,
virtual-8086 mode, or SMM, the default operand-size and address-size attributes are always 16
bits.
The operand-size attribute selects the sizes of operands that instructions operate on. When the
16-bit operand-size attribute is in force, operands can generally be either 8 bits or 16 bits, and
when the 32-bit operand-size attribute is in force, operands can generally be 8 bits or 32 bits.
The address-size attribute selects the sizes of addresses used to address memory: 16 bits or 32
bits. When the 16-bit address-size attribute is in force, segment offsets and displacements are 16
bits. This restriction limits the size of a segment that can be addressed to 64 KBytes. When the

3-14
 BASIC EXECUTION ENVIRONMENT

32-bit address-size attribute is in force, segment offsets and displacements are 32 bits, allowing
segments of up to 4 GBytes to be addressed.
The default operand-size attribute and/or address-size attribute can be overridden for a particular
instruction by adding an operand-size and/or address-size prefix to an instruction (refer to
Chapter 17, Mixing 16-Bit and 32-Bit Code of the Intel Architecture Software Developer’s
Manual, Volume 3). The effect of this prefix applies only to the instruction it is attached to.
Table 3-1 shows effective operand size and address size (when executing in protected mode)
depending on the settings of the D/B flag and the operand-size and address-size prefixes.

Table 3-1. Effective Operand- and Address-Size Attributes

D Flag in Code Segment
Descriptor 0 0 0 0 1 1 1 1
Operand-Size Prefix 66H N N Y Y N N Y Y
Address-Size Prefix 67H N Y N Y N Y N Y
Effective Operand Size 16 16 32 32 32 32 16 16
Effective Address Size 16 32 16 32 32 16 32 16

NOTES:
Y Yes, this instruction prefix is present.
N No, this instruction prefix is not present.

3-15
BASIC EXECUTION ENVIRONMENT

3-16
 4
Procedure Calls,
Interrupts, and
Exceptions
 CHAPTER 4
PROCEDURE CALLS, INTERRUPTS, AND
EXCEPTIONS

This chapter describes the facilities in the Intel Architecture (IA) for executing calls to proce-
dures or subroutines. It also describes how interrupts and exceptions are handled from the
perspective of an application programmer.

4.1. PROCEDURE CALL TYPES

The processor supports procedure calls in the following two different ways:
• CALL and RET instructions.
• ENTER and LEAVE instructions, in conjunction with the CALL and RET instructions.
Both of these procedure call mechanisms use the procedure stack, commonly referred to simply
as “the stack,” to save the state of the calling procedure, pass parameters to the called procedure,
and store local variables for the currently executing procedure.
The processor’s facilities for handling interrupts and exceptions are similar to those used by the
CALL and RET instructions.

4.2. STACK
The stack (refer to Figure 4-1) is a contiguous array of memory locations. It is contained in a
segment and identified by the segment selector in the SS register. (When using the flat memory
model, the stack can be located anywhere in the linear address space for the program.) A stack
can be up to 4 gigabytes long, the maximum size of a segment.
The next available memory location on the stack is called the top of stack. At any given time,
the stack pointer (contained in the ESP register) gives the address (that is the offset from the base
of the SS segment) of the top of the stack.
Items are placed on the stack using the PUSH instruction and removed from the stack using the
POP instruction. When an item is pushed onto the stack, the processor decrements the ESP
register, then writes the item at the new top of stack. When an item is popped off the stack, the
processor reads the item from the top of stack, then increments the ESP register. In this manner,
the stack grows down in memory (towards lesser addresses) when items are pushed on the stack
and shrinks up (towards greater addresses) when the items are popped from the stack.
A program or operating system/executive can set up many stacks. For example, in multitasking
systems, each task can be given its own stack. The number of stacks in a system is limited by
the maximum number of segments and the available physical memory. When a system sets up

4-1
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

many stacks, only one stack—the current stack—is available at a time. The current stack is the
one contained in the segment referenced by the SS register.

Stack Segment
Bottom of Stack
(Initial ESP Value)

Local Variables
for Calling
Procedure The Stack Can Be
16 or 32 Bits Wide

Parameters
Passed to The EBP register is
Called typically set to point
Procedure to the return
instruction pointer.
Frame Boundary
Return Instruction EBP Register
Pointer
ESP Register
Top of Stack

Pushes Move the Pops Move the

Top Of Stack to Top Of Stack to
Lower Addresses Higher Addresses

Figure 4-1. Stack Structure

The processor references the SS register automatically for all stack operations. For example,
when the ESP register is used as a memory address, it automatically points to an address in the
current stack. Also, the CALL, RET, PUSH, POP, ENTER, and LEAVE instructions all perform
operations on the current stack.

4-2
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

4.2.1. Setting Up a Stack

To set a stack and establish it as the current stack, the program or operating system/executive
must do the following:
1. Establish a stack segment.
2. Load the segment selector for the stack segment into the SS register using a MOV, POP, or
LSS instruction.
3. Load the stack pointer for the stack into the ESP register using a MOV, POP, or LSS
instruction. (The LSS instruction can be used to load the SS and ESP registers in one
operation.)
Refer to Chapter 3, Protected-Mode Memory Management of the Intel Architecture Software
Developer’s Manual, Volume 3, for information on how to set up a segment descriptor and
segment limits for a stack segment.

4.2.2. Stack Alignment

The stack pointer for a stack segment should be aligned on 16-bit (word) or 32-bit (double-word)
boundaries, depending on the width of the stack segment. The D flag in the segment descriptor
for the current code segment sets the stack-segment width (refer to Chapter 3, Protected-Mode
Memory Management of the Intel Architecture Software Developer’s Manual, Volume 3). The
PUSH and POP instructions use the D flag to determine how much to decrement or increment
the stack pointer on a push or pop operation, respectively. When the stack width is 16 bits, the
stack pointer is incremented or decremented in 16-bit increments; when the width is 32 bits, the
stack pointer is incremented or decremented in 32-bit increments.
The processor does not check stack pointer alignment. It is the responsibility of the programs,
tasks, and system procedures running on the processor to maintain proper alignment of stack
pointers. Misaligning a stack pointer can cause serious performance degradation and in some
instances program failures.

4.2.3. Address-Size Attributes for Stack Accesses

Instructions that use the stack implicitly (such as the PUSH and POP instructions) have two
address-size attributes each of either 16 or 32 bits. This is because they always have the implicit
address of the top of the stack, and they may also have an explicit memory address (for example,
PUSH Array1[EBX]). The attribute of the explicit address is determined by the D flag of the
current code segment and the presence or absence of the 67H address-size prefix, as usual.
The address-size attribute of the top of the stack determines whether SP or ESP is used for the
stack access. Stack operations with an address-size attribute of 16 use the 16-bit SP stack pointer
register and can use a maximum stack address of FFFFH; stack operations with an address-size
attribute of 32 bits use the 32-bit ESP register and can use a maximum address of FFFFFFFFH.
The default address-size attribute for data segments used as stacks is controlled by the B flag of

4-3
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

the segment’s descriptor. When this flag is clear, the default address-size attribute is 16; when
the flag is set, the address-size attribute is 32.

4.2.4. Procedure Linking Information

The processor provides two pointers for linking of procedures: the stack-frame base pointer and
the return instruction pointer. When used in conjunction with a standard software procedure-call
technique, these pointers permit reliable and coherent linking of procedures

4.2.4.1. STACK-FRAME BASE POINTER

The stack is typically divided into frames. Each stack frame can then contain local variables,
parameters to be passed to another procedure, and procedure linking information. The stack-
frame base pointer (contained in the EBP register) identifies a fixed reference point within the
stack frame for the called procedure. To use the stack-frame base pointer, the called procedure
typically copies the contents of the ESP register into the EBP register prior to pushing any local
variables on the stack. The stack-frame base pointer then permits easy access to data structures
passed on the stack, to the return instruction pointer, and to local variables added to the stack by
the called procedure.
Like the ESP register, the EBP register automatically points to an address in the current stack
segment (that is, the segment specified by the current contents of the SS register).

4.2.4.2. RETURN INSTRUCTION POINTER

Prior to branching to the first instruction of the called procedure, the CALL instruction pushes
the address in the EIP register onto the current stack. This address is then called the return-
instruction pointer and it points to the instruction where execution of the calling procedure
should resume following a return from the called procedure. Upon returning from a called
procedure, the RET instruction pops the return-instruction pointer from the stack back into the
EIP register. Execution of the calling procedure then resumes.
The processor does not keep track of the location of the return-instruction pointer. It is thus up
to the programmer to insure that stack pointer is pointing to the return-instruction pointer on the
stack, prior to issuing a RET instruction. A common way to reset the stack pointer to the point
to the return-instruction pointer is to move the contents of the EBP register into the ESP register.
If the EBP register is loaded with the stack pointer immediately following a procedure call, it
should point to the return instruction pointer on the stack.
The processor does not require that the return instruction pointer point back to the calling proce-
dure. Prior to executing the RET instruction, the return instruction pointer can be manipulated
in software to point to any address in the current code segment (near return) or another code
segment (far return). Performing such an operation, however, should be undertaken very
cautiously, using only well defined code entry points.

4-4
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

4.3. CALLING PROCEDURES USING CALL AND RET

The CALL instructions allows control transfers to procedures within the current code segment
(near call) and in a different code segment (far call). Near calls usually provide access to local
procedures within the currently running program or task. Far calls are usually used to access
operating system procedures or procedures in a different task. Refer to Chapter 3, Instruction
Set Reference of the Intel Architecture Software Developer’s Manual, Volume 2, for a detailed
description of the CALL instruction.
The RET instruction also allows near and far returns to match the near and far versions of the
CALL instruction. In addition, the RET instruction allows a program to increment the stack
pointer on a return to release parameters from the stack. The number of bytes released from the
stack is determined by an optional argument (n) to the RET instruction. Refer to Chapter 3,
Instruction Set Reference of the Intel Architecture Software Developer’s Manual, Volume 2, for
a detailed description of the RET instruction.

4.3.1. Near CALL and RET Operation

When executing a near call, the processor does the following (refer to Figure 4-2):
1. Pushes the current value of the EIP register on the stack.
2. Loads the offset of the called procedure in the EIP register.
3. Begins execution of the called procedure.
When executing a near return, the processor performs these actions:
1. Pops the top-of-stack value (the return instruction pointer) into the EIP register.
2. (If the RET instruction has an optional n argument.) Increments the stack pointer by the
number of bytes specified with the n operand to release parameters from the stack.
3. Resumes execution of the calling procedure.

4-5
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Stack During Stack During

Stack Near Call Far Call
Stack
Frame
Frame
Before
Before
Call Param 1 Param 1
Call
Param 2 Param 2
Param 3 ESP Before Call Param 3 ESP Before Call
Stack Calling EIP ESP After Call Calling CS
Frame Stack Calling EIP ESP After Call
After Frame
Call After
Call
Stack During Stack During
Near Return Far Return

ESP After Return ESP After Return

Param 1 Param 1
Param 2 Param 2
Param 3 Param 3
Calling EIP ESP Before Return Calling CS
Calling EIP ESP Before Return

Note: On a near or far return, parameters are

released from the stack if the correct
value is given for the n operand in
the RET n instruction.
Figure 4-2. Stack on Near and Far Calls

4.3.2. Far CALL and RET Operation

When executing a far call, the processor performs these actions (refer to Figure 4-2):
1. Pushes current value of the CS register on the stack.
2. Pushes the current value of the EIP register on the stack.
3. Loads the segment selector of the segment that contains the called procedure in the CS
register.
4. Loads the offset of the called procedure in the EIP register.
5. Begins execution of the called procedure.

4-6
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

When executing a far return, the processor does the following:

1. Pops the top-of-stack value (the return instruction pointer) into the EIP register.
2. Pops the top-of-stack value (the segment selector for the code segment being returned to)
into the CS register.
3. (If the RET instruction has an optional n argument.) Increments the stack pointer by the
number of bytes specified with the n operand to release parameters from the stack.
4. Resumes execution of the calling procedure.

4.3.3. Parameter Passing

Parameters can be passed between procedures in any of three ways: through general-purpose
registers, in an argument list, or on the stack.

4.3.3.1. PASSING PARAMETERS THROUGH THE GENERAL-PURPOSE

REGISTERS
The processor does not save the state of the general-purpose registers on procedure calls. A
calling procedure can thus pass up to six parameter to the called procedure by copying the
parameters into any of these registers (except the ESP and EBP registers) prior to executing the
CALL instruction. The called procedure can likewise pass parameters back to the calling proce-
dure through general-purpose registers.

4.3.3.2. PASSING PARAMETERS ON THE STACK

To pass a large number of parameters to the called procedure, the parameters can be placed on
the stack, in the stack frame for the calling procedure. Here, it is useful to use the stack-frame
base pointer (in the EBP register) to make a frame boundary for easy access to the parameters.
The stack can also be used to pass parameters back from the called procedure to the calling
procedure.

4.3.3.3. PASSING PARAMETERS IN AN ARGUMENT LIST

An alternate method of passing a larger number of parameters (or a data structure) to the called
procedure is to place the parameters in an argument list in one of the data segments in memory.
A pointer to the argument list can then be passed to the called procedure through a general-
purpose register or the stack. Parameters can also be passed back to the calling procedure in this
same manner.

4.3.4. Saving Procedure State Information

The processor does not save the contents of the general-purpose registers, segment registers, or
the EFLAGS register on a procedure call. A calling procedure should explicitly save the values

4-7
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

in any of the general-purpose registers that it will need when it resumes execution after a return.
These values can be saved on the stack or in memory in one of the data segments.
The PUSHA and POPA instruction facilitates saving and restoring the contents of the general-
purpose registers. PUSHA pushes the values in all the general-purpose registers on the stack in
the following order: EAX, ECX, EDX, EBX, ESP (the value prior to executing the PUSHA
instruction), EBP, ESI, and EDI. The POPA instruction pops all the register values saved with a
PUSHA instruction (except the ESI value) from the stack to their respective registers.
If a called procedure changes the state of any of the segment registers explicitly, it should restore
them to their former value before executing a return to the calling procedure.
If a calling procedure needs to maintain the state of the EFLAGS register it can save and restore
all or part of the register using the PUSHF/PUSHFD and POPF/POPFD instructions. The
PUSHF instruction pushes the lower word of the EFLAGS register on the stack, while the
PUSHFD instruction pushes the entire register. The POPF instruction pops a word from the
stack into the lower word of the EFLAGS register, while the POPFD instruction pops a double
word from the stack into the register.

4.3.5. Calls to Other Privilege Levels

The IA’s protection mechanism recognizes four privilege levels, numbered from 0 to 3, where
greater numbers mean lesser privileges. The primary reason to use these privilege levels is to
improve the reliability of operating systems. For example, Figure 4-3 shows how privilege
levels can be interpreted as rings of protection.
In this example, the highest privilege level 0 (at the center of the diagram) is used for segments
that contain the most critical code modules in the system, usually the kernel of an operating
system. The outer rings (with progressively lower privileges) are used for segments that contain
code modules for less critical software.
Code modules in lower privilege segments can only access modules operating at higher privi-
lege segments by means of a tightly controlled and protected interface called a gate. Attempts
to access higher privilege segments without going through a protection gate and without having
sufficient access rights causes a general-protection exception (#GP) to be generated.

4-8
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Protection Rings

Operating
System
Kernel Level 0
Operating System
Services (Device
Drivers, Etc.) Level 1

Applications Level 2

Level 3
Highest Lowest
0 1 2 3

Privilege Levels

Figure 4-3. Protection Rings

If an operating system or executive uses this multilevel protection mechanism, a call to a proce-
dure that is in a more privileged protection level than the calling procedure is handled in a
similar manner as a far call (refer to Section 4.3.2., “Far CALL and RET Operation”). The
differences are as follows:
• The segment selector provided in the CALL instruction references a special data structure
called a call gate descriptor. Among other things, the call gate descriptor provides the
following:
— Access rights information.
— The segment selector for the code segment of the called procedure.
— An offset into the code segment (that is, the instruction pointer for the called
procedure).
• The processor switches to a new stack to execute the called procedure. Each privilege level
has its own stack. The segment selector and stack pointer for the privilege level 3 stack are
stored in the SS and ESP registers, respectively, and are automatically saved when a call to
a more privileged level occurs. The segment selectors and stack pointers for the privilege
level 2, 1, and 0 stacks are stored in a system segment called the task state segment (TSS).
The use of a call gate and the TSS during a stack switch are transparent to the calling procedure,
except when a general-protection exception is raised.

4-9
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

4.3.6. CALL and RET Operation Between Privilege Levels

When making a call to a more privileged protection level, the processor does the following (refer
to Figure 4-2):
1. Performs an access rights check (privilege check).
2. Temporarily saves (internally) the current contents of the SS, ESP, CS, and EIP registers.
3. Loads the segment selector and stack pointer for the new stack (that is, the stack for the
privilege level being called) from the TSS into the SS and ESP registers and switches to
the new stack.
4. Pushes the temporarily saved SS and ESP values for the calling procedure’s stack onto the
new stack.
5. Copies the parameters from the calling procedure’s stack to the new stack. (A value in the
call gate descriptor determines how many parameters to copy to the new stack.)
6. Pushes the temporarily saved CS and EIP values for the calling procedure to the new stack.
7. Loads the segment selector for the new code segment and the new instruction pointer from
the call gate into the CS and EIP registers, respectively.
8. Begins execution of the called procedure at the new privilege level.
When executing a return from the privileged procedure, the processor performs these actions:
1. Performs a privilege check.
2. Restores the CS and EIP registers to their values prior to the call.
3. (If the RET instruction has an optional n argument.) Increments the stack pointer by the
number of bytes specified with the n operand to release parameters from the stack. If the
call gate descriptor specifies that one or more parameters be copied from one stack to the
other, a RET n instruction must be used to release the parameters from both stacks. Here,
the n operand specifies the number of bytes occupied on each stack by the parameters. On
a return, the processor increments ESP by n for each stack to step over (effectively
remove) these parameters from the stacks.
4. Restores the SS and ESP registers to their values prior to the call, which causes a switch
back to the stack of the calling procedure.
5. (If the RET instruction has an optional n argument.) Increments the stack pointer by the
number of bytes specified with the n operand to release parameters from the stack (refer to
the explanation in step 3).
6. Resumes execution of the calling procedure.

4-10
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Stack for Stack for

Calling Procedure Called Procedure

Calling SS
Calling ESP
Param 1 Param 1
Stack Frame
Param 2 Param 2 Stack Frame
Before Call
Param 3 ESP Before Call Param 3 After Call
Calling CS
ESP After Call Calling EIP

Calling SS
ESP After Return Calling ESP
Param 1 Param 1
Param 2 Param 2
Param 3 Param 3
Calling CS
ESP Before Return Calling EIP

Note: On a return, parameters are

released on both stacks if the
correct value is given for the n
operand in the RET n instruction.

Figure 4-4. Stack Switch on a Call to a Different Privilege Level

Refer to Chapter 4, Protection of the Intel Architecture Software Developer’s Manual, Volume
3, for detailed information on calls to privileged levels and the call gate descriptor.

4.4. INTERRUPTS AND EXCEPTIONS

The processor provides two mechanisms for interrupting program execution: interrupts and
exceptions:
• An interrupt is an asynchronous event that is typically triggered by an I/O device.
• An exception is a synchronous event that is generated when the processor detects one or
more predefined conditions while executing an instruction. The IA specifies three classes
of exceptions: faults, traps, and aborts.
The processor responds to interrupts and exceptions in essentially the same way. When an inter-
rupt or exception is signaled, the processor halts execution of the current program or task and
switches to a handler procedure that has been written specifically to handle the interrupt or
exception condition. The processor accesses the handler procedure through an entry in the inter-

4-11
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

rupt descriptor table (IDT). When the handler has completed handling the interrupt or exception,
program control is returned to the interrupted program or task.
The operating system, executive, and/or device drivers normally handle interrupts and excep-
tions independently from application programs or tasks. Application programs can, however,
access the interrupt and exception handlers incorporated in an operating system or executive
through assembly-language calls. The remainder of this section gives a brief overview of the
processor’s interrupt and exception handling mechanism. Refer to Chapter 5, Interrupt and
Exception Handling of the Intel Architecture Software Developer’s Manual, Volume 3, for a
detailed description of this mechanism.
The IA defines 17 predefined interrupts and exceptions and 224 user defined interrupts, which
are associated with entries in the IDT. Each interrupt and exception in the IDT is identified with
a number, called a vector. Table 4-1 lists the interrupts and exceptions with entries in the IDT
and their respective vector numbers. Vectors 0 through 8, 10 through 14, and 16 through 19 are
the predefined interrupts and exceptions, and vectors 32 through 255 are the user-defined inter-
rupts, called maskable interrupts.
Note that the processor defines several additional interrupts that do not point to entries in the
IDT; the most notable of these interrupts is the SMI interrupt. Refer to Chapter 5, Interrupt and
Exception Handling of the Intel Architecture Software Developer’s Manual, Volume 3, for more
information about the interrupts and exceptions that the IA supports.
When the processor detects an interrupt or exception, it does one of the following things:
• Executes an implicit call to a handler procedure.
• Executes an implicit call to a handler task.
The Pentium® III processor can generate two types of exceptions:
• Numeric exceptions
• Non-numeric exceptions
When numeric exceptions occur, a processor supporting Streaming SIMD Extensions takes one
of two possible courses of action:
• The processor can handle the exception by itself, producing the most reasonable result and
allowing numeric program execution to continue undisturbed (i.e., masked exception
response).
• A software exception handler can be invoked to handle the exception (i.e., unmasked
exception response).
Each of the numeric exception conditions has corresponding flag and mask bits in the MXCSR
(Streaming SIMD Extensions control status register). If an exception is masked (the corre-
sponding mask bit in MXCSR = 1), the processor takes an appropriate default action and
continues with the computation. If the exception is unmasked (mask bit = 0) and the OS supports
SIMD floating-point exceptions (i.e. CR4.OSXMMEXCPT = 1), a software exception handler
is invoked immediately through SIMD floating-point exception interrupt vector 19. If the excep-
tion is unmasked (mask bit = 0) and the OS does not support SIMD floating-point exceptions
(i.e. CR4.OSXMMEXCPT = 0), an invalid opcode exception is signaled instead of a SIMD

4-12
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

floating-point exception. Refer to Section 9.5.5., “Exception Handling in Streaming SIMD

Extensions”, in Chapter 9, Programming with the Streaming SIMD Extensions for more infor-
mation on handling STREAMING SIMD Extensions exceptions.
Note that because SIMD floating-point exceptions are precise and occur immediately, the situ-
ation does not arise where an x87-FP instruction, an FWAIT instruction, or another Streaming
SIMD Extensions instruction will catch a pending unmasked SIMD floating-point exception.

4.4.1. Call and Return Operation for Interrupt or Exception

Handling Procedures
A call to an interrupt or exception handler procedure is similar to a procedure call to another
protection level (refer to Section 4.3.6., “CALL and RET Operation Between Privilege Levels”).
Here, the interrupt vector references one of two kinds of gates: an interrupt gate or a trap gate.
Interrupt and trap gates are similar to call gates in that they provide the following information:
• Access rights information.
• The segment selector for the code segment that contains the handler procedure.
• An offset into the code segment to the first instruction of the handler procedure.
The difference between an interrupt gate and a trap gate is as follows. If an interrupt or exception
handler is called through an interrupt gate, the processor clears the interrupt enable (IF) flag in
the EFLAGS register to prevent subsequent interrupts from interfering with the execution of the
handler. When a handler is called through a trap gate, the state of the IF flag is not changed.
If the code segment for the handler procedure has the same privilege level as the currently
executing program or task, the handler procedure uses the current stack; if the handler executes
at a more privileged level, the processor switches to the stack for the handler’s privilege level.

4-13
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Table 4-1. Exceptions and Interrupts

Vector No. Mnemonic
Description Source
0 #DE Divide Error DIV and IDIV instructions.
1 #DB Debug Any code or data reference.
2 NMI Interrupt Non-maskable external interrupt.
3 #BP Breakpoint INT 3 instruction.
4 #OF Overflow INTO instruction.
5 #BR BOUND Range Exceeded BOUND instruction.
6 #UD Invalid Opcode (UnDefined UD2 instruction or reserved opcode.1
Opcode)
7 #NM Device Not Available (No Math Floating-point or WAIT/FWAIT instruction.
Coprocessor)
8 #DF Double Fault Any instruction that can generate an
exception, an NMI, or an INTR.
9 CoProcessor Segment Overrun Floating-point instruction.2
(reserved)
10 #TS Invalid TSS Task switch or TSS access.
11 #NP Segment Not Present Loading segment registers or accessing
system segments.
12 #SS Stack Segment Fault Stack operations and SS register loads.
13 #GP General Protection Any memory reference and other
protection checks.
14 #PF Page Fault Any memory reference.
15 (Intel reserved. Do not use.)
16 #MF Floating-Point Error (Math Fault) Floating-point or WAIT/FWAIT instruction.
17 #AC Alignment Check Any data reference in memory.3
18 #MC Machine Check Error codes (if any) and source are model
dependent.4
19 #XF Streaming SIMD Extensions SIMD floating-point numeric exceptions.5
20-31 (Intel reserved. Do not use.)
32-255 Maskable Interrupts External interrupt from INTR pin or INT n
instruction.

1. The UD2 instruction was introduced in the Pentium® Pro processor.

2. IA processors after the Intel386™ processor do not generate this exception.
3. This exception was introduced in the Intel486™ processor.
4. This exception was introduced in the Pentium® processor and enhanced in the Pentium® Pro processor.
5. This exception was introduced in the Pentium® III processor.

4-14
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

If no stack switch occurs, the processor does the following when calling an interrupt or excep-
tion handler (refer to Figure 4-5):
1. Pushes the current contents of the EFLAGS, CS, and EIP registers (in that order) on the
stack.
2. Pushes an error code (if appropriate) on the stack.
3. Loads the segment selector for the new code segment and the new instruction pointer
(from the interrupt gate or trap gate) into the CS and EIP registers, respectively.
4. If the call is through an interrupt gate, clears the IF flag in the EFLAGS register.
5. Begins execution of the handler procedure at the new privilege level.

Stack Usage with No

Privilege-Level Change
Interrupted Procedure’s
and Handler’s Stack

ESP Before
EFLAGS Transfer to Handler
CS
EIP
Error Code ESP After
Transfer to Handler

Stack Usage with

Privilege-Level Change
Interrupted Procedure’s Handler’s Stack
Stack

ESP Before
Transfer to Handler SS
ESP
EFLAGS
CS
EIP
ESP After Error Code
Transfer to Handler

Figure 4-5. Stack Usage on Transfers to Interrupt and Exception Handling Routines

4-15
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

If a stack switch does occur, the processor does the following:

1. Temporarily saves (internally) the current contents of the SS, ESP, EFLAGS, CS, and EIP
registers.
2. Loads the segment selector and stack pointer for the new stack (that is, the stack for the
privilege level being called) from the TSS into the SS and ESP registers and switches to
the new stack.
3. Pushes the temporarily saved SS, ESP, EFLAGS, CS, and EIP values for the interrupted
procedure’s stack onto the new stack.
4. Pushes an error code on the new stack (if appropriate).
5. Loads the segment selector for the new code segment and the new instruction pointer
(from the interrupt gate or trap gate) into the CS and EIP registers, respectively.
6. If the call is through an interrupt gate, clears the IF flag in the EFLAGS register.
7. Begins execution of the handler procedure at the new privilege level.
A return from an interrupt or exception handler is initiated with the IRET instruction. The IRET
instruction is similar to the far RET instruction, except that it also restores the contents of the
EFLAGS register for the interrupted procedure:
When executing a return from an interrupt or exception handler from the same privilege level as
the interrupted procedure, the processor performs these actions:
1. Restores the CS and EIP registers to their values prior to the interrupt or exception.
2. Restores the EFLAGS register.
3. Increments the stack pointer appropriately
4. Resumes execution of the interrupted procedure.
When executing a return from an interrupt or exception handler from a different privilege level
than the interrupted procedure, the processor performs these actions:
1. Performs a privilege check.
2. Restores the CS and EIP registers to their values prior to the interrupt or exception.
3. Restores the EFLAGS register.
4. Restores the SS and ESP registers to their values prior to the interrupt or exception,
resulting in a stack switch back to the stack of the interrupted procedure.
5. Resumes execution of the interrupted procedure.

4-16
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

4.4.2. Calls to Interrupt or Exception Handler Tasks

Interrupt and exception handler routines can also be executed in a separate task. Here, an inter-
rupt or exception causes a task switch to a handler task. The handler task is given its own address
space and (optionally) can execute at a higher protection level than application programs or
tasks.
The switch to the handler task is accomplished with an implicit task call that references a task
gate descriptor. The task gate provides access to the address space for the handler task. As part
of the task switch, the processor saves complete state information for the interrupted program or
task. Upon returning from the handler task, the state of the interrupted program or task is
restored and execution continues. Refer to Chapter 5, Interrupt and Exception Handling, of the
Intel Architecture Software Developer’s Manual, Volume 3, for a detailed description of the
processor’s mechanism for handling interrupts and exceptions through handler tasks.

4.4.3. Interrupt and Exception Handling in Real-Address Mode

When operating in real-address mode, the processor responds to an interrupt or exception with
an implicit far call to an interrupt or exception handler. The processor uses the interrupt or
exception vector number as an index into an interrupt table. The interrupt table contains instruc-
tion pointers to the interrupt and exception handler procedures.
The processor saves the state of the EFLAGS register, the EIP register, the CS register, and an
optional error code on the stack before switching to the handler procedure.
A return from the interrupt or exception handler is carried out with the IRET instruction.
Refer to Chapter 16, 8086 Emulation, of the Intel Architecture Software Developer’s Manual,
Volume 3, for more information on handling interrupts and exceptions in real-address mode.

4.4.4. INT n, INTO, INT 3, and BOUND Instructions

The INT n, INTO, INT 3, and BOUND instructions allow a program or task to explicitly call an
interrupt or exception handler. The INT n instruction uses an interrupt vector as an argument,
which allows a program to call any interrupt handler.
The INTO instruction explicitly calls the overflow exception (#OF) handler if the overflow flag
(OF) in the EFLAGS register is set. The OF flag indicates overflow on arithmetic instructions,
but it does not automatically raise an overflow exception. An overflow exception can only be
raised explicitly in either of the following ways:
• Execute the INTO instruction.
• Test the OF flag and execute the INT n instruction with an argument of 4 (the vector
number of the overflow exception) if the flag is set.
Both the methods of dealing with overflow conditions allow a program to test for overflow at
specific places in the instruction stream.
The INT 3 instruction explicitly calls the breakpoint exception (#BP) handler.

4-17
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

The BOUND instruction explicitly calls the BOUND-range exceeded exception (#BR) handler
if an operand is found to be not within predefined boundaries in memory. This instruction is
provided for checking references to arrays and other data structures. Like the overflow
exception, the BOUND-range exceeded exception can only be raised explicitly with the
BOUND instruction or the INT n instruction with an argument of 5 (the vector number of the
bounds-check exception). The processor does not implicitly perform bounds checks and raise
the BOUND-range exceeded exception.

4.5. PROCEDURE CALLS FOR BLOCK-STRUCTURED

LANGUAGES
The IA supports an alternate method of performing procedure calls with the ENTER (enter
procedure) and LEAVE (leave procedure) instructions. These instructions automatically create
and release, respectively, stack frames for called procedures. The stack frames have predefined
spaces for local variables and the necessary pointers to allow coherent returns from called proce-
dures. They also allow scope rules to be implemented so that procedures can access their own
local variables and some number of other variables located in other stack frames.
The ENTER and LEAVE instructions offer two benefits:
• They provide machine-language support for implementing block-structured languages,
such as C and Pascal.
• They simplify procedure entry and exit in compiler-generated code.

4.5.1. ENTER Instruction

The ENTER instruction creates a stack frame compatible with the scope rules typically used in
block-structured languages. In block-structured languages, the scope of a procedure is the set of
variables to which it has access. The rules for scope vary among languages. They may be based
on the nesting of procedures, the division of the program into separately compiled files, or some
other modularization scheme.
The ENTER instruction has two operands. The first specifies the number of bytes to be reserved
on the stack for dynamic storage for the procedure being called. Dynamic storage is the memory
allocated for variables created when the procedure is called, also known as automatic variables.
The second parameter is the lexical nesting level (from 0 to 31) of the procedure. The nesting
level is the depth of a procedure in a hierarchy of procedure calls. The lexical level is unrelated
to either the protection privilege level or to the I/O privilege level of the currently running
program or task.
The ENTER instruction in the following example, allocates 2 Kbytes of dynamic storage on the
stack and sets up pointers to two previous stack frames in the stack frame for this procedure.

ENTER 2048,3

4-18
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

The lexical nesting level determines the number of stack frame pointers to copy into the new
stack frame from the preceding frame. A stack frame pointer is a doubleword used to access the
variables of a procedure. The set of stack frame pointers used by a procedure to access the
variables of other procedures is called the display. The first doubleword in the display is a
pointer to the previous stack frame. This pointer is used by a LEAVE instruction to undo the
effect of an ENTER instruction by discarding the current stack frame.
After the ENTER instruction creates the display for a procedure, it allocates the dynamic local
variables for the procedure by decrementing the contents of the ESP register by the number of
bytes specified in the first parameter. This new value in the ESP register serves as the initial top-
of-stack for all PUSH and POP operations within the procedure.
To allow a procedure to address its display, the ENTER instruction leaves the EBP register
pointing to the first doubleword in the display. Because stacks grow down, this is actually the
doubleword with the highest address in the display. Data manipulation instructions that specify
the EBP register as a base register automatically address locations within the stack segment
instead of the data segment.
The ENTER instruction can be used in two ways: nested and non-nested. If the lexical level is
0, the non-nested form is used. The non-nested form pushes the contents of the EBP register on
the stack, copies the contents of the ESP register into the EBP register, and subtracts the first
operand from the contents of the ESP register to allocate dynamic storage. The non-nested form
differs from the nested form in that no stack frame pointers are copied. The nested form of the
ENTER instruction occurs when the second parameter (lexical level) is not zero.
The following pseudo code shows the formal definition of the ENTER instruction. STORAGE
is the number of bytes of dynamic storage to allocate for local variables, and LEVEL is the
lexical nesting level.
PUSH EBP;
FRAME_PTR ← ESP;
IF LEVEL > 0
THEN
DO (LEVEL − 1) times
EBP ← EBP − 4;
PUSH Pointer(EBP); (* doubleword pointed to by EBP *)
OD;
PUSH FRAME_PTR;
FI;
EBP ← FRAME_PTR;
ESP ← ESP − STORAGE;
The main procedure (in which all other procedures are nested) operates at the highest lexical
level, level 1. The first procedure it calls operates at the next deeper lexical level, level 2. A level
2 procedure can access the variables of the main program, which are at fixed locations specified
by the compiler. In the case of level 1, the ENTER instruction allocates only the requested
dynamic storage on the stack because there is no previous display to copy.
A procedure which calls another procedure at a lower lexical level gives the called procedure
access to the variables of the caller. The ENTER instruction provides this access by placing a
pointer to the calling procedure’s stack frame in the display.

4-19
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

A procedure which calls another procedure at the same lexical level should not give access to
its variables. In this case, the ENTER instruction copies only that part of the display from the
calling procedure which refers to previously nested procedures operating at higher lexical levels.
The new stack frame does not include the pointer for addressing the calling procedure’s stack
frame.
The ENTER instruction treats a re-entrant procedure as a call to a procedure at the same lexical
level. In this case, each succeeding iteration of the re-entrant procedure can address only its own
variables and the variables of the procedures within which it is nested. A re-entrant procedure
always can address its own variables; it does not require pointers to the stack frames of previous
iterations.
By copying only the stack frame pointers of procedures at higher lexical levels, the ENTER
instruction makes certain that procedures access only those variables of higher lexical levels, not
those at parallel lexical levels (refer to Figure 4-6).

Main (Lexical Level 1)

Procedure A (Lexical Level 2)
Procedure B (Lexical Level 3)

Procedure C (Lexical Level 3)

Procedure D (Lexical Level 4)

Figure 4-6. Nested Procedures

Block-structured languages can use the lexical levels defined by ENTER to control access to the
variables of nested procedures. In Figure 4-6, for example, if procedure A calls procedure B
which, in turn, calls procedure C, then procedure C will have access to the variables of the
MAIN procedure and procedure A, but not those of procedure B because they are at the same
lexical level. The following definition describes the access to variables for the nested procedures
in Figure 4-6.
1. MAIN has variables at fixed locations.
2. Procedure A can access only the variables of MAIN.
3. Procedure B can access only the variables of procedure A and MAIN. Procedure B cannot
access the variables of procedure C or procedure D.
4. Procedure C can access only the variables of procedure A and MAIN. procedure C cannot
access the variables of procedure B or procedure D.
5. Procedure D can access the variables of procedure C, procedure A, and MAIN. Procedure
D cannot access the variables of procedure B.

4-20
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

In Figure 4-7, an ENTER instruction at the beginning of the MAIN procedure creates three
doublewords of dynamic storage for MAIN, but copies no pointers from other stack frames. The
first doubleword in the display holds a copy of the last value in the EBP register before the
ENTER instruction was executed. The second doubleword holds a copy of the contents of the
EBP register following the ENTER instruction. After the instruction is executed, the EBP
register points to the first doubleword pushed on the stack, and the ESP register points to the last
doubleword in the stack frame.
When MAIN calls procedure A, the ENTER instruction creates a new display (refer to Figure
4-8). The first doubleword is the last value held in MAIN’s EBP register. The second double-
word is a pointer to MAIN’s stack frame which is copied from the second doubleword in MAIN’s
display. This happens to be another copy of the last value held in MAIN’s EBP register. Proce-
dure A can access variables in MAIN because MAIN is at level 1. Therefore the base address
for the dynamic storage used in MAIN is the current address in the EBP register, plus four bytes
to account for the saved contents of MAIN’s EBP register. All dynamic variables for MAIN are
at fixed, positive offsets from this value.

Old EBP EBP

Display
Main’s EBP

Dynamic
Storage
ESP

Figure 4-7. Stack Frame after Entering the MAIN Procedure

4-21
 Old EBP
Main’s EBP

Main’s EBP EBP

Display
Main’s EBP
Procedure A’s EBP
Dynamic
Storage
ESP

Figure 4-8. Stack Frame after Entering Procedure A

When procedure A calls procedure B, the ENTER instruction creates a new display (refer to
Figure 4-9). The first doubleword holds a copy of the last value in procedure A’s EBP register.
The second and third doublewords are copies of the two stack frame pointers in procedure A’s
display. Procedure B can access variables in procedure A and MAIN by using the stack frame
pointers in its display.
 PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Old EBP
Main’s EBP

Main’s EBP
Main’s EBP
Procedure A’s EBP

Procedure A’s EBP EBP

Main’s EBP
Display
Procedure A’s EBP
Procedure B’s EBP

Dynamic
Storage
ESP

Figure 4-9. Stack Frame after Entering Procedure B

When procedure B calls procedure C, the ENTER instruction creates a new display for proce-
dure C (refer to Figure 4-10). The first doubleword holds a copy of the last value in procedure
B’s EBP register. This is used by the LEAVE instruction to restore procedure B’s stack frame.
The second and third doublewords are copies of the two stack frame pointers in procedure A’s
display. If procedure C were at the next deeper lexical level from procedure B, a fourth double-
word would be copied, which would be the stack frame pointer to procedure B’s local variables.
Note that procedure B and procedure C are at the same level, so procedure C is not intended to
access procedure B’s variables. This does not mean that procedure C is completely isolated from
procedure B; procedure C is called by procedure B, so the pointer to the returning stack frame
is a pointer to procedure B's stack frame. In addition, procedure B can pass parameters to proce-
dure C either on the stack or through variables global to both procedures (that is, variables in the
scope of both procedures).

4-23
PROCEDURE CALLS, INTERRUPTS, AND EXCEPTIONS

Old EBP
Main’s EBP

Main’s EBP
Main’s EBP
Procedure A’s EBP

Procedure A’s EBP

Main’s EBP
Procedure A’s EBP
Procedure B’s EBP

Procedure B’s EBP EBP

Main’s EBP
Display
Procedure A’s EBP
Procedure C’s EBP

Dynamic
Storage
ESP

Figure 4-10. Stack Frame after Entering Procedure C

4.5.2. LEAVE Instruction

The LEAVE instruction, which does not have any operands, reverses the action of the previous
ENTER instruction. The LEAVE instruction copies the contents of the EBP register into the ESP
register to release all stack space allocated to the procedure. Then it restores the old value of the
EBP register from the stack. This simultaneously restores the ESP register to its original value.
A subsequent RET instruction then can remove any arguments and the return address pushed on
the stack by the calling program for use by the procedure.

4-24
 5
Data Types and
Addressing Modes
 CHAPTER 5
DATA TYPES AND ADDRESSING MODES

This chapter describes data types and addressing modes available to programmers of the Intel
Architecture (IA) processors.

5.1. FUNDAMENTAL DATA TYPES

The fundamental data types of the IA are bytes, words, doublewords, and quadwords (refer to
Figure 5-1). A byte is eight bits, a word is 2 bytes (16 bits), a doubleword is 4 bytes (32 bits),
and a quadword is 8 bytes (64 bits).

7 0
Byte
N
15 87 0
High Low Word
Byte Byte
N+1 N
31 16 15 0
High Word Low Word Doubleword
N+2 N
63 32 31 0
High Doubleword Low Doubleword Quadword

N+4 N
Figure 5-1. Fundamental Data Types

The Pentium® III processor introduced a new data type, a 128-bit packed data type. It is packed
single precision (32 bits) floating-point numbers. These values are the operands for the SIMD
floating-point operations. They are also the operands for the scalar equivalents of these instruc-
tions. Refer to Chapter 5-2, SIMD Floating-Point Data Type for a description of this data type.

127 96 95 64 63 32 31 0

Figure 5-2. SIMD Floating-Point Data Type

5-1
DATA TYPES AND ADDRESSING MODES

Figure 5-2 shows the byte order of each of the fundamental data types when referenced as oper-
ands in memory. The low byte (bits 0 through 7) of each data type occupies the lowest address
in memory and that address is also the address of the operand.

5.1.1. Alignment of Words, Doublewords, and Quadwords

Words, doublewords, and quadwords do not need to be aligned in memory on natural bound-
aries. (The natural boundaries for words, double words, and quadwords are even-numbered
addresses, addresses evenly divisible by four, and addresses evenly divisible by eight, respec-
tively.) However, to improve the performance of programs, data structures (especially stacks)
should be aligned on natural boundaries whenever possible. The reason for this is that the
processor requires two memory accesses to make an unaligned memory access; whereas,
aligned accesses require only one memory access. A word or doubleword operand that crosses
a 4-byte boundary or a quadword operand that crosses an 8-byte boundary is considered
unaligned and requires two separate memory bus cycles to access it; a word that starts on an odd
address but does not cross a word boundary is considered aligned and can still be accessed in
one bus cycle.

EH
7AH DH

Word at Address BH FEH CH

Doubleword at Address AH
Contains FE06H 06H BH Contains 7AFE0636H

36H AH
Byte at Address 9H
1FH 9H
Contains 1FH Quadword at Address 6H
A4H 8H Contains 7AFE06361FA4230BH

Word at Address 6H 23H 7H

Contains 230BH 0BH 6H
5H
4H
Word at Address 2H
Contains 74CBH 74H 3H

Word at Address 1H CBH 2H

Contains CB31H 31H 1H
0H

Figure 5-3. Bytes, Words, Doublewords and Quadwords in Memory

When accessing 128 bit data for the Pentium® III processor, data must be aligned on 16-byte
boundaries. There are instructions that allow for unaligned access, but additional time is
required to receive the data into the cache. If an instruction that expects aligned data is used to
access unaligned data, a general protection fault will occur.

5-2
 DATA TYPES AND ADDRESSING MODES

5.2. NUMERIC, POINTER, BIT FIELD, AND STRING DATA TYPES

Although bytes, words, and doublewords are the fundamental data types of the IA, some instruc-
tions support additional interpretations of these data types to allow operations to be performed
on numeric data types (signed and unsigned integers and BCD integers). Refer to Figure 5-4.
Also, some instructions recognize and operate on additional pointer, bit field, and string data
types. The following sections describe these additional data types.

5.2.1. Integers
Integers are signed binary numbers held in a byte, word, or doubleword. All operations assume
a two’s complement representation. The sign bit is located in bit 7 in a byte integer, bit 15 in a
word integer, and bit 31 in a doubleword integer. The sign bit is set for negative integers and
cleared for positive integers and zero. Integer values range from –128 to +127 for a byte integer,
from –32,768 to +32,767 for a word integer, and from –231 to +231 – 1 for a doubleword integer.

5-3
DATA TYPES AND ADDRESSING MODES

Byte Signed Integer

Sign
7 6 0
Word Signed Integer
Sign
15 14 0
Doubleword Signed Integer
Sign
31 30 0
Byte Unsigned Integer

7 0
Word Unsigned Integer

15 0
Doubleword Unsigned Integer

31 0
BCD Integers
X BCD .... X BCD X BCD
7 43 0
Packed BCD Integers
BCD BCD .... BCD BCD BCD BCD
7 43 0
Near Pointer
Offset or Linear Address
31 0
Far Pointer or Logical Address
Segment Selector Offset
47 32 31 0
Bit Field

Field Length
Least
Significant
Bit

Figure 5-4. Numeric, Pointer, and Bit Field Data Types

5-4
 DATA TYPES AND ADDRESSING MODES

5.2.2. Unsigned Integers

Unsigned integers are unsigned binary numbers contained in a byte, word, or doubleword.
Unsigned integer values range from 0 to 255 for an unsigned byte integer, from 0 to 65,535 for
an unsigned word integer, and from 0 to 232 – 1 for an unsigned doubleword integer. Unsigned
integers are sometimes referred to as ordinals.

5.2.3. BCD Integers

Binary-coded decimal integers (BCD integers) are unsigned 4-bit integers with valid values
ranging from 0 to 9. BCD integers can be unpacked (one BCD digit per byte) or packed (two
BCD digits per byte). The value of an unpacked BCD integer is the binary value of the low half-
byte (bits 0 through 3). The high half-byte (bits 4 through 7) can be any value during addition
and subtraction, but must be zero during multiplication and division.
Packed BCD integers allow two BCD digits to be contained in one byte. Here, the digit in the
high half-byte is more significant than the digit in the low half-byte.

5.2.4. Pointers
Pointers are addresses of locations in memory. The Pentium® Pro processor recognizes two types
of pointers: a near pointer (32 bits) and a far pointer (48 bits). A near pointer is a 32-bit offset
(also called an effective address) within a segment. Near pointers are used for all memory refer-
ences in a flat memory model or for references in a segmented model where the identity of the
segment being accessed is implied. A far pointer is a 48-bit logical address, consisting of a 16-bit
segment selector and a 32-bit offset. Far pointers are used for memory references in a segmented
memory model where the identity of a segment being accessed must be specified explicitly.

5.2.5. Bit Fields

A bit field is a contiguous sequence of bits. It can begin at any bit position of any byte in
memory and can contain up to 32 bits.

5.2.6. Strings
Strings are continuous sequences of bits, bytes, words, or doublewords. A bit string can begin
at any bit position of any byte and can contain up to 232 – 1 bits. A byte string can contain bytes,
words, or doublewords and can range from zero to 232 – 1 bytes (4 gigabytes).

5-5
DATA TYPES AND ADDRESSING MODES

5.2.7. Floating-Point Data Types

The processor’s floating-point instructions recognize a set of real, integer, and BCD integer data
types. Refer to Section 7.4., “Floating-Point Data Types and Formats” in Chapter 7, Floating-
Point Unit for a description of FPU data types.

5.2.8. MMX™ Technology Data Types

IA processors that implement the Intel MMX™ technology recognize a set of packed 64-bit data
types. Refer to Section 8.1.2., “MMX™ Data Types” in Chapter 8, Programming with the Intel
MMX™ Technology for a description of the MMX™ data types.

5.2.9. Streaming SIMD Extensions Data Types

IA processors that implement the Intel Streaming SIMD Extensions recognize a set of 128-bit
data types. Refer to Section 9.1.2., “SIMD Floating-Point Data Types” in Chapter 9, Program-
ming with the Streaming SIMD Extensions for a description of the SIMD floating-point data
types.

5.3. OPERAND ADDRESSING

An IA machine-instruction acts on zero or more operands. Some operands are specified explic-
itly in an instruction and others are implicit to an instruction. An operand can be located in any
of the following places:
• The instruction itself (an immediate operand).
• A register.
• A memory location.
• An I/O port.

5.3.1. Immediate Operands

Some instructions use data encoded in the instruction itself as a source operand. These operands
are called immediate operands (or simply immediates). For example, the following ADD
instruction adds an immediate value of 14 to the contents of the EAX register:
ADD EAX, 14
All the arithmetic instructions (except the DIV and IDIV instructions) allow the source operand
to be an immediate value. The maximum value allowed for an immediate operand varies among
instructions, but can never be greater than the maximum value of an unsigned doubleword
integer (232).

5-6
 DATA TYPES AND ADDRESSING MODES

5.3.2. Register Operands

Source and destination operands can be located in any of the following registers, depending on
the instruction being executed:
• The 32-bit general-purpose registers (EAX, EBX, ECX, EDX, ESI, EDI, ESP, or EBP).
• The 16-bit general-purpose registers (AX, BX, CX, DX, SI, DI, SP, or BP).
• The 8-bit general-purpose registers (AH, BH, CH, DH, AL, BL, CL, or DL).
• The segment registers (CS, DS, SS, ES, FS, and GS).
• The EFLAGS register.
• System registers, such as the global descriptor table (GDTR) or the interrupt descriptor
table register (IDTR).
Some instructions (such as the DIV and MUL instructions) use quadword operands contained
in a pair of 32-bit registers. Register pairs are represented with a colon separating them. For
example, in the register pair EDX:EAX, EDX contains the high order bits and EAX contains the
low order bits of a quadword operand.
Several instructions (such as the PUSHFD and POPFD instructions) are provided to load and
store the contents of the EFLAGS register or to set or clear individual flags in this register. Other
instructions (such as the Jcc instructions) use the state of the status flags in the EFLAGS register
as condition codes for branching or other decision making operations.
The processor contains a selection of system registers that are used to control memory manage-
ment, interrupt and exception handling, task management, processor management, and debug-
ging activities. Some of these system registers are accessible by an application program, the
operating system, or the executive through a set of system instructions. When accessing a
system register with a system instruction, the register is generally an implied operand of the
instruction.

5.3.3. Memory Operands

Source and destination operands in memory are referenced by means of a segment selector and
an offset (refer to Figure 5-5). The segment selector specifies the segment containing the
operand and the offset (the number of bytes from the beginning of the segment to the first byte
of the operand) specifies the linear or effective address of the operand.

15 0 31 0
Segment Offset (or Linear Address)
Selector

Figure 5-5. Memory Operand Address

5-7
DATA TYPES AND ADDRESSING MODES

5.3.3.1. SPECIFYING A SEGMENT SELECTOR

The segment selector can be specified either implicitly or explicitly. The most common method
of specifying a segment selector is to load it in a segment register and then allow the processor
to select the register implicitly, depending on the type of operation being performed. The
processor automatically chooses a segment according to the rules given in Table 5-1.

Table 5-1. Default Segment Selection Rules

Type of Register Segment
Reference Used Used Default Selection Rule
Instructions CS Code Segment All instruction fetches.
Stack SS Stack Segment All stack pushes and pops.
Any memory reference which uses the ESP or EBP
register as a base register.
Local Data DS Data Segment All data references, except when relative to stack or
string destination.
Destination ES Data Segment Destination of string instructions.
Strings pointed to with
the ES register

When storing data in or loading data from memory, the DS segment default can be overridden
to allow other segments to be accessed. Within an assembler, the segment override is generally
handled with a colon “:” operator. For example, the following MOV instruction moves a value
from register EAX into the segment pointed to by the ES register. The offset into the segment is
contained in the EBX register:
MOV ES:[EBX], EAX;
(At the machine level, a segment override is specified with a segment-override prefix, which is
a byte placed at the beginning of an instruction.) The following default segment selections
cannot be overridden:
• Instruction fetches must be made from the code segment.
• Destination strings in string instructions must be stored in the data segment pointed to by
the ES register.
• Push and pop operations must always reference the SS segment.
Some instructions require a segment selector to be specified explicitly. In these cases, the 16-bit
segment selector can be located in a memory location or in a 16-bit register. For example, the
following MOV instruction moves a segment selector located in register BX into segment
register DS:

MOV DS, BX
Segment selectors can also be specified explicitly as part of a 48-bit far pointer in memory. Here,
the first doubleword in memory contains the offset and the next word contains the segment
selector.

5-8
 DATA TYPES AND ADDRESSING MODES

5.3.3.2. SPECIFYING AN OFFSET

The offset part of a memory address can be specified either directly as an static value (called a
displacement) or through an address computation made up of one or more of the following
components:
• Displacement—An 8-, 16-, or 32-bit value.
• Base—The value in a general-purpose register.
• Index—The value in a general-purpose register.
• Scale factor—A value of 2, 4, or 8 that is multiplied by the index value.
The offset which results from adding these components is called an effective address. Each of
these components can have either a positive or negative (2s complement) value, with the excep-
tion of the scaling factor. Figure 5-6 shows all the possible ways that these components can be
combined to create an effective address in the selected segment.

Base Index Scale Displacement

EAX
EAX None
EBX 1
EBX
ECX
ECX 2 8-bit
EDX
+ EDX +
ESP
EBP
EBP * 3 16-bit
ESI
ESI 4 32-bit
EDI
EDI

Offset = Base + (Index ∗ Scale) + Displacement

Figure 5-6. Offset (or Effective Address) Computation

The uses of general-purpose registers as base or index components are restricted in the following
manner:
• The ESP register cannot be used as an index register.
• When the ESP or EBP register is used as the base, the SS segment is the default segment.
In all other cases, the DS segment is the default segment.
The base, index, and displacement components can be used in any combination, and any of these
components can be null. A scale factor may be used only when an index also is used. Each
possible combination is useful for data structures commonly used by programmers in high-level
languages and assembly language. The following addressing modes suggest uses for common
combinations of address components.

Displacement
A displacement alone represents a direct (uncomputed) offset to the operand. Because the
displacement is encoded in the instruction, this form of an address is sometimes called an abso-
lute or static address. It is commonly used to access a statically allocated scalar operand.

5-9
DATA TYPES AND ADDRESSING MODES

Base
A base alone represents an indirect offset to the operand. Since the value in the base register can
change, it can be used for dynamic storage of variables and data structures.

Base + Displacement
A base register and a displacement can be used together for two distinct purposes:
• As an index into an array when the element size is not 2, 4, or 8 bytes—The displacement
component encodes the static offset to the beginning of the array. The base register holds
the results of a calculation to determine the offset to a specific element within the array.
• To access a field of a record—The base register holds the address of the beginning of the
record, while the displacement is an static offset to the field.
An important special case of this combination is access to parameters in a procedure activation
record. A procedure activation record is the stack frame created when a procedure is entered.
Here, the EBP register is the best choice for the base register, because it automatically selects
the stack segment. This is a compact encoding for this common function.

(Index ∗ Scale) + Displacement

This address mode offers an efficient way to index into a static array when the element size is 2,
4, or 8 bytes. The displacement locates the beginning of the array, the index register holds the
subscript of the desired array element, and the processor automatically converts the subscript
into an index by applying the scaling factor.

Base + Index + Displacement

Using two registers together supports either a two-dimensional array (the displacement holds the
address of the beginning of the array) or one of several instances of an array of records (the
displacement is an offset to a field within the record).

Base + (Index ∗ Scale) + Displacement

Using all the addressing components together allows efficient indexing of a two-dimensional
array when the elements of the array are 2, 4, or 8 bytes in size.

5.3.3.3. ASSEMBLER AND COMPILER ADDRESSING MODES

At the machine-code level, the selected combination of displacement, base register, index
register, and scale factor is encoded in an instruction. All assemblers permit a programmer to
use any of the allowable combinations of these addressing components to address operands.
High-level language (HLL) compilers will select an appropriate combination of these compo-
nents based on the HHL construct a programmer defines.

5-10
 DATA TYPES AND ADDRESSING MODES

5.3.4. I/O Port Addressing

The processor supports an I/O address space that contains up to 65,536 8-bit I/O ports. Ports that
are 16-bit and 32-bit may also be defined in the I/O address space. An I/O port can be addressed
with either an immediate operand or a value in the DX register. Refer to Chapter 10,
Input/Output for more information about I/O port addressing.

5-11
DATA TYPES AND ADDRESSING MODES

5-12
 6
Instruction Set
Summary
 CHAPTER 6
INSTRUCTION SET SUMMARY

This chapter lists all the instructions in the Intel Architecture (IA) instruction set, divided into
three functional groups: integer, floating-point, and system. It also briefly describes each of the
integer instructions.
Brief descriptions of the floating-point instructions are given in Chapter 7, Floating-Point Unit;
brief descriptions of the system instructions are given in the Intel Architecture Software Devel-
oper’s Manual, Volume 3.
Detailed descriptions of all the IA instructions are given in the Intel Architecture Software
Developer’s Manual, Volume 2. Included in this volume are a description of each instruction’s
encoding and operation, the effect of an instruction on the EFLAGS flags, and the exceptions an
instruction may generate.

6.1. NEW INTEL ARCHITECTURE INSTRUCTIONS

The following sections give the IA instructions that were new in the Streaming SIMD Exten-
sions, MMX™ Technology and in the Pentium® Pro, Pentium®, and Intel486™ processors.

6.1.1. New Instructions Introduced with the Streaming SIMD

Extensions
The Intel Streaming SIMD Extensions introduced a new set of instructions to the IA, designed
to enhance the performance of multimedia applications, 3D games and other 3D applications, as
well as other applications. These instructions are recognized by all IA processors that implement
the Streaming SIMD Extensions that are listed in Section 6.2.5., “Streaming SIMD Extensions”.

6.1.2. New Instructions Introduced with the MMX™ Technology

The Intel MMX™ technology introduced a new set of instructions to the IA, designed to
enhance the performance of multimedia applications. These instructions are recognized by all
IA processors that implement the MMX™ technology. The MMX™ instructions are listed in
Section 6.2.2., “MMX™ Technology Instructions”.

6-1
INSTRUCTION SET SUMMARY

6.1.3. New Instructions in the Pentium® Pro Processor

The following instructions are new in the Pentium® Pro processor:
• CMOVcc—Conditional move (refer to Section 6.3.1.2., “Conditional Move Instructions”).
• FCMOVcc—Floating-point conditional move on condition-code flags in EFLAGS register
(refer to Section 7.5.3., “Data Transfer Instructions” in Chapter 7, Floating-Point Unit).
• FCOMI/FCOMIP/FUCOMI/FUCOMIP—Floating-point compare and set condition-code
flags in EFLAGS register (refer to Section 7.5.6., “Comparison and Classification Instruc-
tions” in Chapter 7, Floating-Point Unit).
• RDPMC—Read performance monitoring counters (refer to Chapter 3, Instruction Set
Reference of the Intel Architecture Software Developer’s Manual, Volume 2). (This
instruction is also available in all Pentium® processors that implement the MMX™
technology.)
• UD2—Undefined instruction (refer to Section 6.15.4., “No-Operation and Undefined
Instructions”).

6.1.4.New Instructions in the Pentium® Processor

The following instructions are new in the Pentium® processor:

• CMPXCHG8B (compare and exchange 8 bytes) instruction.

• CPUID (CPU identification) instruction. (This instruction was introduced in the Pentium®
processor and added to later versions of the Intel486™ processor.)
• RDTSC (read time-stamp counter) instruction.
• RDMSR (read model-specific register) instruction.
• WRMSR (write model-specific register) instruction.
• RSM (resume from SMM) instruction.
The form of the MOV instruction used to access the test registers has been removed on the
Pentium® and future IA processors.

6-2
 INSTRUCTION SET SUMMARY

6.1.5. New Instructions in the Intel486™ Processor

The following instructions are new in the Intel486™ processor:
• BSWAP (byte swap) instruction.
• XADD (exchange and add) instruction.
• CMPXCHG (compare and exchange) instruction.
• ΙNVD (invalidate cache) instruction.
• WBINVD (write-back and invalidate cache) instruction.
• INVLPG (invalidate TLB entry) instruction.

6.2. INSTRUCTION SET LIST

This section lists all the IA instructions divided into three major groups: integer, MMX™ tech-
nology, floating-point, and system instructions. For each instruction, the mnemonic and descrip-
tive names are given. When two or more mnemonics are given (for example,
CMOVA/CMOVNBE), they represent different mnemonics for the same instruction opcode.
Assemblers support redundant mnemonics for some instructions to make it easier to read code
listings. For instance, CMOVA (Conditional move if above) and CMOVNBE (Conditional
move is not below or equal) represent the same condition.

6.2.1. Integer Instructions

Integer instructions perform the integer arithmetic, logic, and program flow control operations
that programmers commonly use to write application and system software to run on an IA
processor. In the following sections, the integer instructions are divided into several instruction
subgroups.

6.2.1.1. DATA TRANSFER INSTRUCTIONS

MOV Move
CMOVE/CMOVZ Conditional move if equal/Conditional move if zero
CMOVNE/CMOVNZ Conditional move if not equal/Conditional move if not zero
CMOVA/CMOVNBE Conditional move if above/Conditional move if not below
or equal
CMOVAE/CMOVNB Conditional move if above or equal/Conditional move if
not below
CMOVB/CMOVNAE Conditional move if below/Conditional move if not above
or equal
CMOVBE/CMOVNA Conditional move if below or equal/Conditional move if not above

6-3
INSTRUCTION SET SUMMARY

CMOVG/CMOVNLE Conditional move if greater/Conditional move if not less or equal

CMOVGE/CMOVNL Conditional move if greater or equal/Conditional move if not less
CMOVL/CMOVNGE Conditional move if less/Conditional move if not greater or equal
CMOVLE/CMOVNG Conditional move if less or equal/Conditional move if not greater
CMOVC Conditional move if carry
CMOVNC Conditional move if not carry
CMOVO Conditional move if overflow
CMOVNO Conditional move if not overflow
CMOVS Conditional move if sign (negative)
CMOVNS Conditional move if not sign (non-negative)
CMOVP/CMOVPE Conditional move if parity/Conditional move if parity even
CMOVNP/CMOVPO Conditional move if not parity/Conditional move if parity odd
XCHG Exchange
BSWAP Byte swap
XADD Exchange and add
CMPXCHG Compare and exchange
CMPXCHG8B Compare and exchange 8 bytes
PUSH Push onto stack
POP Pop off of stack
PUSHA/PUSHAD Push general-purpose registers onto stack
POPA/POPAD Pop general-purpose registers from stack
IN Read from a port
OUT Write to a port
CWD/CDQ Convert word to doubleword/Convert doubleword to quadword
CBW/CWDE Convert byte to word/Convert word to doubleword in EAX register
MOVSX Move and sign extend
MOVZX Move and zero extend

6-4
 INSTRUCTION SET SUMMARY

6.2.1.2. BINARY ARITHMETIC INSTRUCTIONS

ADD Integer add
ADC Add with carry
SUB Subtract
SBB Subtract with borrow
IMUL Signed multiply
MUL Unsigned multiply
IDIV Signed divide
DIV Unsigned divide
INC Increment
DEC Decrement
NEG Negate
CMP Compare

6.2.1.3. DECIMAL ARITHMETIC

DAA Decimal adjust after addition
DAS Decimal adjust after subtraction
AAA ASCII adjust after addition
AAS ASCII adjust after subtraction
AAM ASCII adjust after multiplication
AAD ASCII adjust before division

6.2.1.4. LOGIC INSTRUCTIONS

AND And
OR Or
XOR Exclusive or
NOT Not

6.2.1.5. SHIFT AND ROTATE INSTRUCTIONS

SAR Shift arithmetic right
SHR Shift logical right
SAL/SHL Shift arithmetic left/Shift logical left

6-5
INSTRUCTION SET SUMMARY

SHRD Shift right double

SHLD Shift left double
ROR Rotate right
ROL Rotate left
RCR Rotate through carry right
RCL Rotate through carry left

6.2.1.6. BIT AND BYTE INSTRUCTIONS

BT Bit test
BTS Bit test and set
BTR Bit test and reset
BTC Bit test and complement
BSF Bit scan forward
BSR Bit scan reverse
SETE/SETZ Set byte if equal/Set byte if zero
SETNE/SETNZ Set byte if not equal/Set byte if not zero
SETA/SETNBE Set byte if above/Set byte if not below or equal
SETAE/SETNB/SETNC Set byte if above or equal/Set byte if not below/Set byte if not carry
SETB/SETNAE/SETC Set byte if below/Set byte if not above or equal/Set byte if carry
SETBE/SETNA Set byte if below or equal/Set byte if not above
SETG/SETNLE Set byte if greater/Set byte if not less or equal
SETGE/SETNL Set byte if greater or equal/Set byte if not less
SETL/SETNGE Set byte if less/Set byte if not greater or equal
SETLE/SETNG Set byte if less or equal/Set byte if not greater
SETS Set byte if sign (negative)
SETNS Set byte if not sign (non-negative)
SETO Set byte if overflow
SETNO Set byte if not overflow
SETPE/SETP Set byte if parity even/Set byte if parity
SETPO/SETNP Set byte if parity odd/Set byte if not parity
TEST Logical compare

6-6
 INSTRUCTION SET SUMMARY

6.2.1.7. CONTROL TRANSFER INSTRUCTIONS

JMP Jump
JE/JZ Jump if equal/Jump if zero
JNE/JNZ Jump if not equal/Jump if not zero
JA/JNBE Jump if above/Jump if not below or equal
JAE/JNB Jump if above or equal/Jump if not below
JB/JNAE Jump if below/Jump if not above or equal
JBE/JNA Jump if below or equal/Jump if not above
JG/JNLE Jump if greater/Jump if not less or equal
JGE/JNL Jump if greater or equal/Jump if not less
JL/JNGE Jump if less/Jump if not greater or equal
JLE/JNG Jump if less or equal/Jump if not greater
JC Jump if carry
JNC Jump if not carry
JO Jump if overflow
JNO Jump if not overflow
JS Jump if sign (negative)
JNS Jump if not sign (non-negative)
JPO/JNP Jump if parity odd/Jump if not parity
JPE/JP Jump if parity even/Jump if parity
JCXZ/JECXZ Jump register CX zero/Jump register ECX zero
LOOP Loop with ECX counter
LOOPZ/LOOPE Loop with ECX and zero/Loop with ECX and equal
LOOPNZ/LOOPNE Loop with ECX and not zero/Loop with ECX and not equal
CALL Call procedure
RET Return
IRET Return from interrupt
INT Software interrupt
INTO Interrupt on overflow
BOUND Detect value out of range
ENTER High-level procedure entry
LEAVE High-level procedure exit

6-7
INSTRUCTION SET SUMMARY

6.2.1.8. STRING INSTRUCTIONS

MOVS/MOVSB Move string/Move byte string
MOVS/MOVSW Move string/Move word string
MOVS/MOVSD Move string/Move doubleword string
CMPS/CMPSB Compare string/Compare byte string
CMPS/CMPSW Compare string/Compare word string
CMPS/CMPSD Compare string/Compare doubleword string
SCAS/SCASB Scan string/Scan byte string
SCAS/SCASW Scan string/Scan word string
SCAS/SCASD Scan string/Scan doubleword string
LODS/LODSB Load string/Load byte string
LODS/LODSW Load string/Load word string
LODS/LODSD Load string/Load doubleword string
STOS/STOSB Store string/Store byte string
STOS/STOSW Store string/Store word string
STOS/STOSD Store string/Store doubleword string
REP Repeat while ECX not zero
REPE/REPZ Repeat while equal/Repeat while zero
REPNE/REPNZ Repeat while not equal/Repeat while not zero
INS/INSB Input string from port/Input byte string from port
INS/INSW Input string from port/Input word string from port
INS/INSD Input string from port/Input doubleword string from port
OUTS/OUTSB Output string to port/Output byte string to port
OUTS/OUTSW Output string to port/Output word string to port
OUTS/OUTSD Output string to port/Output doubleword string to port

6-8
 INSTRUCTION SET SUMMARY

6.2.1.9. FLAG CONTROL INSTRUCTIONS

STC Set carry flag
CLC Clear the carry flag
CMC Complement the carry flag
CLD Clear the direction flag
STD Set direction flag
LAHF Load flags into AH register
SAHF Store AH register into flags
PUSHF/PUSHFD Push EFLAGS onto stack
POPF/POPFD Pop EFLAGS from stack
STI Set interrupt flag
CLI Clear the interrupt flag

6.2.1.10. SEGMENT REGISTER INSTRUCTIONS

LDS Load far pointer using DS
LES Load far pointer using ES
LFS Load far pointer using FS
LGS Load far pointer using GS
LSS Load far pointer using SS

6.2.1.11. MISCELLANEOUS INSTRUCTIONS

LEA Load effective address
NOP No operation
UB2 Undefined instruction
XLAT/XLATB Table lookup translation
CPUID Processor Identification

6-9
INSTRUCTION SET SUMMARY

6.2.2. MMX™ Technology Instructions

The MMX™ instructions execute on those IA processors that implement the Intel MMX™ tech-
nology. These instructions operate on packed-byte, packed-word, packed-doubleword, and
quadword operands. As with the integer instructions, the following list of MMX™ instructions
is divided into subgroups.

6.2.2.1. MMX™ DATA TRANSFER INSTRUCTIONS

MOVD Move doubleword
MOVQ Move quadword

6.2.2.2. MMX™ CONVERSION INSTRUCTIONS

PACKSSWB Pack words into bytes with signed saturation
PACKSSDW Pack doublewords into words with signed saturation
PACKUSWB Pack words into bytes with unsigned saturation
PUNPCKHBW Unpack high-order bytes from words
PUNPCKHWD Unpack high-order words from doublewords
PUNPCKHDQ Unpack high-order doublewords from quadword
PUNPCKLBW Unpack low-order bytes from words
PUNPCKLWD Unpack low-order words from doublewords
PUNPCKLDQ Unpack low-order doublewords from quadword

6.2.2.3. MMX™ PACKED ARITHMETIC INSTRUCTIONS

PADDB Add packed bytes
PADDW Add packed words
PADDD Add packed doublewords
PADDSB Add packed bytes with saturation
PADDSW Add packed words with saturation
PADDUSB Add packed unsigned bytes with saturation
PADDUSW Add packed unsigned words with saturation
PSUBB Subtract packed bytes
PSUBW Subtract packed words
PSUBD Subtract packed doublewords
PSUBSB Subtract packed bytes with saturation

6-10
 INSTRUCTION SET SUMMARY

PSUBSW Subtract packed words with saturation

PSUBUSB Subtract packed unsigned bytes with saturation
PSUBUSW Subtract packed unsigned words with saturation
PMULHW Multiply packed words and store high result
PMULLW Multiply packed words and store low result
PMADDWD Multiply and add packed words

6.2.2.4. MMX™ COMPARISON INSTRUCTIONS

PCMPEQB Compare packed bytes for equal
PCMPEQW Compare packed words for equal
PCMPEQD Compare packed doublewords for equal
PCMPGTB Compare packed bytes for greater than
PCMPGTW Compare packed words for greater than
PCMPGTD Compare packed doublewords for greater than

6.2.2.5. MMX™ LOGIC INSTRUCTIONS

PAND Bitwise logical and
PANDN Bitwise logical and not
POR Bitwise logical or
PXOR Bitwise logical exclusive or

6.2.2.6. MMX™ SHIFT AND ROTATE INSTRUCTIONS

PSLLW Shift packed words left logical
PSLLD Shift packed doublewords left logical
PSLLQ Shift packed quadword left logical
PSRLW Shift packed words right logical
PSRLD Shift packed doublewords right logical
PSRLQ Shift packed quadword right logical
PSRAW Shift packed words right arithmetic
PSRAD Shift packed doublewords right arithmetic

6-11
INSTRUCTION SET SUMMARY

6.2.2.7. MMX™ STATE MANAGEMENT

EMMS Empty MMX™ state

6.2.3. Floating-Point Instructions

The floating-point instructions are those that are executed by the processor’s floating-point unit
(FPU). These instructions operate on floating-point (real), extended integer, and binary-coded
decimal (BCD) operands. As with the integer instructions, the following list of floating-point
instructions is divided into subgroups.

6.2.3.1. DATA TRANSFER

FLD Load real
FST Store real
FSTP Store real and pop
FILD Load integer
FIST Store integer
FISTP Store integer and pop
FBLD Load BCD
FBSTP Store BCD and pop
FXCH Exchange registers
FCMOVE Floating-point conditional move if equal
FCMOVNE Floating-point conditional move if not equal
FCMOVB Floating-point conditional move if below
FCMOVBE Floating-point conditional move if below or equal
FCMOVNB Floating-point conditional move if not below
FCMOVNBE Floating-point conditional move if not below or equal
FCMOVU Floating-point conditional move if unordered
FCMOVNU Floating-point conditional move if not unordered

6-12
 INSTRUCTION SET SUMMARY

6.2.3.2. BASIC ARITHMETIC

FADD Add real
FADDP Add real and pop
FIADD Add integer
FSUB Subtract real
FSUBP Subtract real and pop
FISUB Subtract integer
FSUBR Subtract real reverse
FSUBRP Subtract real reverse and pop
FISUBR Subtract integer reverse
FMUL Multiply real
FMULP Multiply real and pop
FIMUL Multiply integer
FDIV Divide real
FDIVP Divide real and pop
FIDIV Divide integer
FDIVR Divide real reverse
FDIVRP Divide real reverse and pop
FIDIVR Divide integer reverse
FPREM Partial remainder
FPREMI IEEE Partial remainder
FABS Absolute value
FCHS Change sign
FRNDINT Round to integer
FSCALE Scale by power of two
FSQRT Square root
FXTRACT Extract exponent and significand

6-13
INSTRUCTION SET SUMMARY

6.2.3.3. COMPARISON
FCOM Compare real
FCOMP Compare real and pop
FCOMPP Compare real and pop twice
FUCOM Unordered compare real
FUCOMP Unordered compare real and pop
FUCOMPP Unordered compare real and pop twice
FICOM Compare integer
FICOMP Compare integer and pop
FCOMI Compare real and set EFLAGS
FUCOMI Unordered compare real and set EFLAGS
FCOMIP Compare real, set EFLAGS, and pop
FUCOMIP Unordered compare real, set EFLAGS, and pop
FTST Test real
FXAM Examine real

6.2.3.4. TRANSCENDENTAL
FSIN Sine
FCOS Cosine
FSINCOS Sine and cosine
FPTAN Partial tangent
FPATAN Partial arctangent
F2XM1 2x − 1
FYL2X y∗log2x
FYL2XP1 y∗log2(x+1)

6-14
 INSTRUCTION SET SUMMARY

6.2.3.5. LOAD CONSTANTS

FLD1 Load +1.0
FLDZ Load +0.0
FLDPI Load π
FLDL2E Load log2e
FLDLN2 Load loge2
FLDL2T Load log210
FLDLG2 Load log102

6.2.3.6. FPU CONTROL

FINCSTP Increment FPU register stack pointer
FDECSTP Decrement FPU register stack pointer
FFREE Free floating-point register
FINIT Initialize FPU after checking error conditions
FNINIT Initialize FPU without checking error conditions
FCLEX Clear floating-point exception flags after checking for error
conditions
FNCLEX Clear floating-point exception flags without checking for error
conditions
FSTCW Store FPU control word after checking error conditions
FNSTCW Store FPU control word without checking error conditions
FLDCW Load FPU control word
FSTENV Store FPU environment after checking error conditions
FNSTENV Store FPU environment without checking error conditions
FLDENV Load FPU environment
FSAVE Save FPU state after checking error conditions
FNSAVE Save FPU state without checking error conditions
FRSTOR Restore FPU state
FSTSW Store FPU status word after checking error conditions
FNSTSW Store FPU status word without checking error conditions
WAIT/FWAIT Wait for FPU
FNOP FPU no operation

6-15
INSTRUCTION SET SUMMARY

6.2.4. System Instructions

The following system instructions are used to control those functions of the processor that are
provided to support for operating systems and executives.
LGDT Load global descriptor table (GDT) register
SGDT Store global descriptor table (GDT) register
LLDT Load local descriptor table (LDT) register
SLDT Store local descriptor table (LDT) register
LTR Load task register
STR Store task register
LIDT Load interrupt descriptor table (IDT) register
SIDT Store interrupt descriptor table (IDT) register
MOV Load and store control registers
LMSW Load machine status word
SMSW Store machine status word
CLTS Clear the task-switched flag
ARPL Adjust requested privilege level
LAR Load access rights
LSL Load segment limit
VERR Verify segment for reading
VERW Verify segment for writing
MOV Load and store debug registers
INVD Invalidate cache, no writeback
WBINVD Invalidate cache, with writeback
INVLPG Invalidate TLB Entry
LOCK (prefix) Lock Bus
HLT Halt processor
RSM Return from system management mode (SSM)

6-16
 INSTRUCTION SET SUMMARY

RDMSR Read model-specific register

WRMSR Write model-specific register
RDPMC Read performance monitoring counters
RDTSC Read time stamp counter
SYSENTER Fast System Call, transfers to a flat protected mode kernel at CPL=0.
SYSEXIT Fast System Call, transfers to a flat protected mode kernel at CPL=3.

6.2.5. Streaming SIMD Extensions

The Streaming SIMD Extensions execute on those IA processors that implement the Intel
Streaming SIMD Extensions. These instructions operate on packed single precision floating-
point operands. As with the MMX™ instructions, the following list of Streaming SIMD Exten-
sions is divided into subgroups.

6.2.5.1. STREAMING SIMD EXTENSIONS DATA TRANSFER INSTRUCTIONS

MOVAPS Move aligned packed single-precision floating-point
MOVUPS Move unaligned packed single-precision floating-point
MOVHPS Move unaligned high packed single-precision floating-point
MOVHLPS Move aligned high packed single-precision floating-point to low
packed single-precision floating-point
MOVLPS Move unaligned low packed single-precision floating-point
MOVLHPS Move aligned low packed single-precision floating-point to high
packed single-precision floating-point
MOVMSKPS Move mask packed single-precision floating-point
MOVSS Move scalar single-precision floating-point

6.2.5.2. STREAMING SIMD EXTENSIONS CONVERSION INSTRUCTIONS

CVTPI2PS Convert packed 32-bit integer to packed single-precision floating-point
CVTSI2SS Convert scalar 32-bit integer to scalar single-precision floating-point
CVTPS2PI Convert packed single-precision floating-point to packed 32-bit integer
CVTTPS2PI Convert truncate packed single-precision floating-point to packed 32-bit
integer
CVTSS2SI Convert scalar single-precision floating-point to a 32-bit integer
CVTTSS2SI Convert truncate scalar single-precision floating-point to scalar 32-bit
integer

6-17
INSTRUCTION SET SUMMARY

6.2.5.3. STREAMING SIMD EXTENSIONS PACKED ARITHMETIC

INSTRUCTIONS
ADDPS Add packed single-precision floating-point
SUBPS Subtract packed single-precision floating-point
ADDSS Add scalar single-precision floating-point
SUBSS Subtract scalar single-precision floating-point
MULPS Multiply packed single-precision floating-point
MULSS Multiply scalar single-precision floating-point
DIVPS Divide packed single-precision floating-point
DIVSS Divide scalar single-precision floating-point
SQRTPS Square root packed single-precision floating-point
SQRTSS Square root scalar single-precision floating-point
MAXPS Maximum packed single-precision floating-point
MAXSS Maximum scalar single-precision floating-point
MINPS Minimum packed single-precision floating-point
MINSS Minimum scalar single-precision floating-point

6.2.5.4. STREAMING SIMD EXTENSIONS COMPARISON INSTRUCTIONS

CMPPS Compare packed single-precision floating-point
CMPSS Compare scalar single-precision floating-point
COMISS Compare scalar single-precision floating-point ordered and set
EFLAGS
UCOMISS Unordered compare scalar single-precision floating-point ordered
and set EFLAGS

6.2.5.5. STREAMING SIMD EXTENSIONS LOGICAL INSTRUCTIONS

ANDPS Bit-wise packed logical AND for single-precision floating-point
ANDNPS Bit-wise packed logical AND NOT for single-precision floating-
point
ORPS Bit-wise packed logical OR for single-precision floating-point
XORPS Bit-wise packed logical XOR for single-precision floating-point

6-18
 INSTRUCTION SET SUMMARY

6.2.5.6. STREAMING SIMD EXTENSIONS DATA SHUFFLE INSTRUCTIONS

SHUFPS Shuffle packed single-precision floating-point
UNPCKHPS Unpacked high packed single-precision floating-point
UNPCKLPS Unpacked low packed single-precision floating-point

6.2.5.7. STREAMING SIMD EXTENSIONS ADDITIONAL SIMD-INTEGER

INSTRUCTIONS
PAVGB/PAVGW Average unsigned source sub-operands, without incurring a loss in precision
PEXTRW Extract 16-bit word from MMX™ register
PINSRW Insert 16-bit word into MMX™ register
PMAXUB/PMAXSW Maximum of packed unsigned integer bytes or signed integer words
PMINUB/PMINSW Minimum of packed unsigned integer bytes or signed integer words
PMOVMSKB Move Byte Mask from MMX™ register
PMULHUW Unsigned high packed integer word multiply in MMX™ register
PSADBW Sum of absolute differences
PSHUFW Shuffle packed integer word in MMX™ register

6.2.5.8. STREAMING SIMD EXTENSIONS CACHEABILITY CONTROL

INSTRUCTIONS
MASKMOVQ Non-temporal byte mask store of packed integer in a MMX™ register
MOVNTQ Non-temporal store of packed integer in a MMX™ register
MOVNTPS Non-temporal store of packed single-precision floating-point
PREFETCH Load 32 or greater number of bytes
SFENCE Store Fence

6.2.5.9. STREAMING SIMD EXTENSIONS STATE MANAGEMENT

INSTRUCTIONS
LDMXCSR Load SIMD Floating-Point Control and Status Register
STMXCSR Store SIMD Floating-Point Control and Status Register
FXSAVE Saves floating-point and MMX™ state and SIMD Floating-Point
state to memory
FXRSTOR Loads FP and MMX™ state and SIMD Floating-Point state from
memory

6-19
INSTRUCTION SET SUMMARY

6.3. DATA MOVEMENT INSTRUCTIONS

The data movement instructions move bytes, words, doublewords, or quadwords both between
memory and the processor’s registers and between registers. These instructions are divided into
four groups:
• General-purpose data movement.
• Exchange.
• Stack manipulation.
• Type-conversion.

6.3.1. General-Purpose Data Movement Instructions

The MOV (move) and CMOVcc (conditional move) instructions transfer data between memory
and registers or between registers.

6.3.1.1. MOVE INSTRUCTION

The MOV instruction performs basic load data and store data operations between memory and
the processor’s registers and data movement operations between registers. It handles data trans-
fers along the paths listed in Table 6-1. (Refer to “MOV—Move to/from Control Registers” and
“MOV—Move to/from Debug Registers” in Chapter 3, Instruction Set Reference of the Intel
Architecture Software Developer’s Manual, Volume 2, for information on moving data to and
from the control and debug registers.)
The MOV instruction cannot move data from one memory location to another or from one
segment register to another segment register. Memory-to-memory moves can be performed with
the MOVS (string move) instruction (refer to Section 6.10., “String Operations”).

6.3.1.2. CONDITIONAL MOVE INSTRUCTIONS

The CMOVcc instructions are a group of instructions that check the state of the status flags in
the EFLAGS register and perform a move operation if the flags are in a specified state (or condi-
tion). These instructions can be used to move a 16- or 32-bit value from memory to a general-
purpose register or from one general-purpose register to another. The flag state being tested for
each instruction is specified with a condition code (cc) that is associated with the instruction. If
the condition is not satisfied, a move is not performed and execution continues with the instruc-
tion following the CMOVcc instruction.

6-20
 INSTRUCTION SET SUMMARY

Table 6-1. Move Instruction Operations

Type of Data Movement Source → Destination
From memory to a register Memory location → General-purpose register
Memory location → Segment register
From a register to memory General-purpose register → Memory location
Segment register → Memory location
Between registers General-purpose register → General-purpose register
General-purpose register → Segment register
Segment register → General-purpose register
General-purpose register → Control register
Control register → General-purpose register
General-purpose register → Debug register
Debug register → General-purpose register
Immediate data to a register Immediate → General-purpose register
Immediate data to memory Immediate → Memory location

Table 6-4 shows the mnemonics for the CMOVcc instructions and the conditions being tested
for each instruction. The condition code mnemonics are appended to the letters “CMOV” to
form the mnemonics for the CMOVcc instructions. The instructions listed in Table 6-4 as pairs
(for example, CMOVA/CMOVNBE) are alternate names for the same instruction. The assem-
bler provides these alternate names to make it easier to read program listings.
The CMOVcc instructions are useful for optimizing small IF constructions. They also help elim-
inate branching overhead for IF statements and the possibility of branch mispredictions by the
processor.
These instructions may not be supported on some processors in the Pentium® Pro processor
family. Software can check if the CMOVcc instructions are supported by checking the
processor’s feature information with the CPUID instruction (refer to “CPUID—CPU Identifica-
tion” in Chapter 3, Instruction Set Reference of the Intel Architecture Software Developer’s
Manual, Volume 2).

6.3.1.3. EXCHANGE INSTRUCTIONS

The exchange instructions swap the contents of one or more operands and, in some cases,
performs additional operations such as asserting the LOCK signal or modifying flags in the
EFLAGS register.
The XCHG (exchange) instruction swaps the contents of two operands. This instruction takes
the place of three MOV instructions and does not require a temporary location to save the
contents of one operand location while the other is being loaded. When a memory operand is
used with the XCHG instruction, the processor’s LOCK signal is automatically asserted. This
instruction is thus useful for implementing semaphores or similar data structures for process
synchronization. (Refer to Section 7.1.2., “Bus Locking” of the Intel Architecture Software
Developer’s Manual, Volume 3, for more information on bus locking.)
The BSWAP (byte swap) instruction reverses the byte order in a 32-bit register operand. Bit
positions 0 through 7 are exchanged with 24 through 31, and bit positions 8 through 15 are

6-21
INSTRUCTION SET SUMMARY

exchanged with 16 through 23. Executing this instruction twice in a row leaves the register with
the same value as before. The BSWAP instruction is useful for converting between “big-endian”
and “little-endian” data formats. This instruction also speeds execution of decimal arithmetic.
(The XCHG instruction can be used two swap the bytes in a word.)

Table 6-2. Conditional Move Instructions

Instruction Mnemonic Status Flag States Condition Description
Unsigned Conditional Moves
CMOVA/CMOVNBE (CF or ZF)=0 Above/not below or equal
CMOVAE/CMOVNB CF=0 Above or equal/not below
CMOVNC CF=0 Not carry
CMOVB/CMOVNAE CF=1 Below/not above or equal
CMOVC CF=1 Carry
CMOVBE/CMOVNA (CF or ZF)=1 Below or equal/not above
CMOVE/CMOVZ ZF=1 Equal/zero
CMOVNE/CMOVNZ ZF=0 Not equal/not zero
CMOVP/CMOVPE PF=1 Parity/parity even
CMOVNP/CMOVPO PF=0 Not parity/parity odd
Signed Conditional Moves
CMOVGE/CMOVNL (SF xor OF)=0 Greater or equal/not less
CMOVL/CMOVNGE (SF xor OF)=1 Less/not greater or equal
CMOVLE/CMOVNG ((SF xor OF) or ZF)=1 Less or equal/not greater
CMOVO OF=1 Overflow
CMOVNO OF=0 Not overflow
CMOVS SF=1 Sign (negative)
CMOVNS SF=0 Not sign (non-negative)

The XADD (exchange and add) instruction swaps two operands and then stores the sum of the
two operands in the destination operand. The status flags in the EFLAGS register indicate the
result of the addition. This instruction can be combined with the LOCK prefix (refer to
“LOCK—Assert LOCK# Signal Prefix” in Chapter 3, Instruction Set Reference of the Intel
Architecture Software Developer’s Manual, Volume 2) in a multiprocessing system to allow
multiple processors to execute one DO loop.
The CMPXCHG (compare and exchange) and CMPXCHG8B (compare and exchange 8 bytes)
instructions are used to synchronize operations in systems that use multiple processors. The
CMPXCHG instruction requires three operands: a source operand in a register, another source
operand in the EAX register, and a destination operand. If the values contained in the destination
operand and the EAX register are equal, the destination operand is replaced with the value of
the other source operand (the value not in the EAX register). Otherwise, the original value of the
destination operand is loaded in the EAX register. The status flags in the EFLAGS register

6-22
 INSTRUCTION SET SUMMARY

reflect the result that would have been obtained by subtracting the destination operand from the
value in the EAX register.
The CMPXCHG instruction is commonly used for testing and modifying semaphores. It checks
to see if a semaphore is free. If the semaphore is free it is marked allocated, otherwise it gets the
ID of the current owner. This is all done in one uninterruptible operation. In a single-processor
system, the CMPXCHG instruction eliminates the need to switch to protection level 0 (to disable
interrupts) before executing multiple instructions to test and modify a semaphore. For multiple
processor systems, CMPXCHG can be combined with the LOCK prefix to perform the compare
and exchange operation atomically. (Refer to Section 7.1., “Locked Atomic Operations” of the
Intel Architecture Software Developer’s Manual, Volume 3, for more information on atomic
operations.)
The CMPXCHG8B instruction also requires three operands: a 64-bit value in EDX:EAX, a
64-bit value in ECX:EBX, and a destination operand in memory. The instruction compares the
64-bit value in the EDX:EAX registers with the destination operand. If they are equal, the 64-bit
value in the ECX:EBX register is stored in the destination operand. If the EDX:EAX register
and the destination are not equal, the destination is loaded in the EDX:EAX register. The
CMPXCHG8B instruction can be combined with the LOCK prefix to perform the operation
atomically.

6.3.2. Stack Manipulation Instructions

The PUSH, POP, PUSHA (push all registers), and POPA (pop all registers) instructions move
data to and from the stack. The PUSH instruction decrements the stack pointer (contained in the
ESP register), then copies the source operand to the top of stack (refer to Figure 6-1). It operates
on memory operands, immediate operands, and register operands (including segment registers).
The PUSH instruction is commonly used to place parameters on the stack before calling a proce-
dure. It can also be used to reserve space on the stack for temporary variables.

Stack
Before Pushing Doubleword After Pushing Doubleword
Stack
Growth 31 0 31 0
n ESP
n−4 Doubleword Value ESP
n−8

Figure 6-1. Operation of the PUSH Instruction

The PUSHA instruction saves the contents of the eight general-purpose registers on the stack
(refer to Figure 6-2). This instruction simplifies procedure calls by reducing the number of
instructions required to save the contents of the general-purpose registers. The registers are
pushed on the stack in the following order: EAX, ECX, EDX, EBX, the initial value of ESP
before EAX was pushed, EBP, ESI, and EDI.

6-23
INSTRUCTION SET SUMMARY

Stack
Before Pushing Registers After Pushing Registers
Stack 31 0 31 0
Growth
n
n-4 ESP
n-8 EAX
n - 12 ECX
n - 16 EDX
n - 20 EBX
n - 24 Old ESP
n - 28 EBP
n - 32 ESI
n - 36 EDI ESP

Figure 6-2. Operation of the PUSHA Instruction

The POP instruction copies the word or doubleword at the current top of stack (indicated by the
ESP register) to the location specified with the destination operand, and then increments the ESP
register to point to the new top of stack (refer to Figure 6-3). The destination operand may
specify a general-purpose register, a segment register, or a memory location.

Stack
Before Popping Doubleword After Popping Doubleword
Stack
Growth 31 0 31 0

n
n-4 ESP
n-8 Doubleword Value ESP

Figure 6-3. Operation of the POP Instruction

The POPA instruction reverses the effect of the PUSHA instruction. It pops the top eight words
or doublewords from the top of the stack into the general-purpose registers, except for the ESP
register (refer to Figure 6-4). If the operand-size attribute is 32, the doublewords on the stack are
transferred to the registers in the following order: EDI, ESI, EBP, ignore doubleword, EBX,
EDX, ECX, and EAX. The ESP register is restored by the action of popping the stack. If the
operand-size attribute is 16, the words on the stack are transferred to the registers in the
following order: DI, SI, BP, ignore word, BX, DX, CX, and AX.

6-24
 INSTRUCTION SET SUMMARY

Stack
Before Popping Registers After Popping Registers
Stack 0 31 0 31
Growth
n
n-4 ESP
n-8 EAX
n - 12 ECX
n - 16 EDX
n - 20 EBX
n - 24 Ignored
n - 28 EBP
n - 32 ESI
n - 36 EDI ESP

Figure 6-4. Operation of the POPA Instruction

6.3.2.1. TYPE CONVERSION INSTRUCTIONS

The type conversion instructions convert bytes into words, words into doublewords, and double-
words into quadwords. These instructions are especially useful for converting integers to larger
integer formats, because they perform sign extension (refer to Figure 6-5).
Two kinds of type conversion instructions are provided: simple conversion and move and
convert.

15 0
Before Sign
S N N N N N N N N N N N N N N N
Extension
31 15 0
After Sign
S S S S S S S S S S S S S S S S S N N N N N N N N N N N N N N N
Extension

Figure 6-5. Sign Extension

6.3.2.2. SIMPLE CONVERSION

The CBW (convert byte to word), CWDE (convert word to doubleword extended), CWD
(convert word to doubleword), and CDQ (convert doubleword to quadword) instructions
perform sign extension to double the size of the source operand.

6-25
INSTRUCTION SET SUMMARY

The CBW instruction copies the sign (bit 7) of the byte in the AL register into every bit position
of the upper byte of the AX register. The CWDE instruction copies the sign (bit 15) of the word
in the AX register into every bit position of the high word of the EAX register.
The CWD instruction copies the sign (bit 15) of the word in the AX register into every bit posi-
tion in the DX register. The CDQ instruction copies the sign (bit 31) of the doubleword in the
EAX register into every bit position in the EDX register. The CWD instruction can be used to
produce a doubleword dividend from a word before a word division, and the CDQ instruction
can be used to produce a quadword dividend from a doubleword before doubleword division.

6.3.2.3. MOVE AND CONVERT

The MOVSX (move with sign extension) and MOVZX (move with zero extension) instructions
move the source operand into a register then perform the sign extension.
The MOVSX instruction extends an 8-bit value to a 16-bit value or an 8- or 16-bit value to 32-bit
value by sign extending the source operand, as shown in Figure 6-5. The MOVZX instruction
extends an 8-bit value to a 16-bit value or an 8- or 16-bit value to 32-bit value by zero extending
the source operand.

6.4. BINARY ARITHMETIC INSTRUCTIONS

The binary arithmetic instructions operate on 8-, 16-, and 32-bit numeric data encoded as signed
or unsigned binary integers. Operations include the add, subtract, multiply, and divide as well
as increment, decrement, compare, and change sign (negate). The binary arithmetic instructions
may also be used in algorithms that operate on decimal (BCD) values.

6.4.1. Addition and Subtraction Instructions

The ADD (add integers), ADC (add integers with carry), SUB (subtract integers), and SBB
(subtract integers with borrow) instructions perform addition and subtraction operations on
signed or unsigned integer operands.
The ADD instruction computes the sum of two integer operands.
The ADC instruction computes the sum of two integer operands, plus 1 if the CF flag is set. This
instruction is used to propagate a carry when adding numbers in stages.
The SUB instruction computes the difference of two integer operands.
The SBB instruction computes the difference of two integer operands, minus 1 if the CF flag is
set. This instruction is used to propagate a borrow when subtracting numbers in stages.

6.4.2. Increment and Decrement Instructions

The INC (increment) and DEC (decrement) instructions add 1 to or subtract 1 from an unsigned
integer operand, respectively. A primary use of these instructions is for implementing counters.

6-26
 INSTRUCTION SET SUMMARY

6.4.3. Comparison and Sign Change Instruction

The CMP (compare) instruction computes the difference between two integer operands and
updates the OF, SF, ZF, AF, PF, and CF flags according to the result. The source operands are
not modified, nor is the result saved. The CMP instruction is commonly used in conjunction with
a Jcc (jump) or SETcc (byte set on condition) instruction, with the latter instructions performing
an action based on the result of a CMP instruction.
The NEG (negate) instruction subtracts a signed integer operand from zero. The effect of the
NEG instruction is to change the sign of a two’s complement operand while keeping its
magnitude.

6.4.4. Multiplication and Divide Instructions

The processor provides two multiply instructions, MUL (unsigned multiply) and IMUL signed
multiply), and two divide instructions, DIV (unsigned divide) and IDIV (signed divide).
The MUL instruction multiplies two unsigned integer operands. The result is computed to twice
the size of the source operands (for example, if word operands are being multiplied, the result is
a doubleword).
The IMUL instruction multiplies two signed integer operands. The result is computed to twice
the size of the source operands; however, in some cases the result is truncated to the size of the
source operands (refer to Chapter 3, Instruction Set Reference of the Intel Architecture Software
Developer’s Manual, Volume 2).
The DIV instruction divides one unsigned operand by another unsigned operand and returns a
quotient and a remainder.
The IDIV instruction is identical to the DIV instruction, except that IDIV performs a signed
division.

6.5. DECIMAL ARITHMETIC INSTRUCTIONS

Decimal arithmetic can be performed by combining the binary arithmetic instructions ADD,
SUB, MUL, and DIV (discussed in Section 6.4., “Binary Arithmetic Instructions”) with the
decimal arithmetic instructions. The decimal arithmetic instructions are provided to carry out
the following operations:
• To adjust the results of a previous binary arithmetic operation to produce a valid BCD
result.
• To adjust the operands of a subsequent binary arithmetic operation so that the operation
will produce a valid BCD result.
These instructions operate only on both packed and unpacked BCD values.

6-27
INSTRUCTION SET SUMMARY

6.5.1. Packed BCD Adjustment Instructions

The DAA (decimal adjust after addition) and DAS (decimal adjust after subtraction) instructions
adjust the results of operations performed on packed BCD integers (refer to Section 5.2.3.,
“BCD Integers” in Chapter 5, Data Types and Addressing Modes). Adding two packed BCD
values requires two instructions: an ADD instruction followed by a DAA instruction. The ADD
instruction adds (binary addition) the two values and stores the result in the AL register. The
DAA instruction then adjusts the value in the AL register to obtain a valid, 2-digit, packed BCD
value and sets the CF flag if a decimal carry occurred as the result of the addition.
Likewise, subtracting one packed BCD value from another requires a SUB instruction followed
by a DAS instruction. The SUB instruction subtracts (binary subtraction) one BCD value from
another and stores the result in the AL register. The DAS instruction then adjusts the value in
the AL register to obtain a valid, 2-digit, packed BCD value and sets the CF flag if a decimal
borrow occurred as the result of the subtraction.

6.5.2. Unpacked BCD Adjustment Instructions

The AAA (ASCII adjust after addition), AAS (ASCII adjust after subtraction), AAM (ASCII
adjust after multiplication), and AAD (ASCII adjust before division) instructions adjust the
results of arithmetic operations performed in unpacked BCD values (refer to Section 5.2.3.,
“BCD Integers” in Chapter 5, Data Types and Addressing Modes). All these instructions assume
that the value to be adjusted is stored in the AL register or, in one instance, the AL and AH regis-
ters.
The AAA instruction adjusts the contents of the AL register following the addition of two
unpacked BCD values. It converts the binary value in the AL register into a decimal value and
stores the result in the AL register in unpacked BCD format (the decimal number is stored in the
lower 4 bits of the register and the upper 4 bits are cleared). If a decimal carry occurred as a result
of the addition, the CF flag is set and the contents of the AH register are incremented by 1.
The AAS instruction adjusts the contents of the AL register following the subtraction of two
unpacked BCD values. Here again, a binary value is converted into an unpacked BCD value. If
a borrow was required to complete the decimal subtract, the CF flag is set and the contents of
the AH register are decremented by 1.
The AAM instruction adjusts the contents of the AL register following a multiplication of two
unpacked BCD values. It converts the binary value in the AL register into a decimal value and
stores the least significant digit of the result in the AL register (in unpacked BCD format) and
the most significant digit, if there is one, in the AH register (also in unpacked BCD format).
The AAD instruction adjusts a two-digit BCD value so that when the value is divided with the
DIV instruction, a valid unpacked BCD result is obtained. The instruction converts the BCD
value in registers AH (most significant digit) and AL (least significant digit) into a binary value
and stores the result in register AL. When the value in AL is divided by an unpacked BCD value,
the quotient and remainder will be automatically encoded in unpacked BCD format.

6-28
 INSTRUCTION SET SUMMARY

6.6. LOGICAL INSTRUCTIONS

The logical instructions AND, OR, XOR (exclusive or), and NOT perform the standard Boolean
operations for which they are named. The AND, OR, and XOR instructions require two oper-
ands; the NOT instruction operates on a single operand.

6.7. SHIFT AND ROTATE INSTRUCTIONS

The shift and rotate instructions rearrange the bits within an operand. These instructions fall into
the following classes:
• Shift.
• Double shift.
• Rotate.

6.7.1. Shift Instructions

The SAL (shift arithmetic left), SHL (shift logical left), SAR (shift arithmetic right), SHR (shift
logical right) instructions perform an arithmetic or logical shift of the bits in a byte, word, or
doubleword.
The SAL and SHL instructions perform the same operation (refer to Figure 6-6). They shift the
source operand left by from 1 to 31 bit positions. Empty bit positions are cleared. The CF flag
is loaded with the last bit shifted out of the operand.
.

Initial State
CF Operand

X 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1

After 1-bit SHL/SAL Instruction

0
1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1 0

After 10-bit SHL/SAL Instruction

0
0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1 0 0 0 0 0 0 0 0 0 0

Figure 6-6. SHL/SAL Instruction Operation

The SHR instruction shifts the source operand right by from 1 to 31 bit positions (refer to Figure
6-7). As with the SHL/SAL instruction, the empty bit positions are cleared and the CF flag is
loaded with the last bit shifted out of the operand.

6-29
INSTRUCTION SET SUMMARY

Initial State Operand CF

1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1 X

After 1-bit SHR Instruction

0
0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 1

After 10-bit SHR Instruction

0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0

Figure 6-7. SHR Instruction Operation

The SAR instruction shifts the source operand right by from 1 to 31 bit positions (refer to Figure
6-8). This instruction differs from the SHR instruction in that it preserves the sign of the source
operand by clearing empty bit positions if the operand is positive or setting the empty bits if the
operand is negative. Again, the CF flag is loaded with the last bit shifted out of the operand.
The SAR and SHR instructions can also be used to perform division by powers of 2 (refer to
Chapter 3, Instruction Set Reference of the Intel Architecture Software Developer’s Manual,
Volume 2).

6-30
 INSTRUCTION SET SUMMARY

Initial State (Positive Operand) Operand CF

0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 X

After 1-bit SAR Instruction

0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1

Initial State (Negative Operand)

CF
1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1 X

After 1-bit SAR Instruction

1 1 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 0 0 0 1 1 1

Figure 6-8. SAR Instruction Operation

6.7.2. Double-Shift Instructions

The SHLD (shift left double) and SHRD (shift right double) instructions shift a specified
number of bits from one operand to another (refer to Figure 6-9). They are provided to facilitate
operations on unaligned bit strings. They can also be used to implement a variety of bit string
move operations.

6-31
INSTRUCTION SET SUMMARY

SHLD Instruction
31 0
CF Destination (Memory or Register)

31 0
Source (Register)

SHRD Instruction
31 0
Source (Register)

31 0
Destination (Memory or Register) CF

Figure 6-9. SHLD and SHRD Instruction Operations

The SHLD instruction shifts the bits in the destination operand to the left and fills the empty bit
positions (in the destination operand) with bits shifted out of the source operand. The destination
and source operands must be the same length (either words or doublewords). The shift count can
range from 0 to 31 bits. The result of this shift operation is stored in the destination operand, and
the source operand is not modified. The CF flag is loaded with the last bit shifted out of the desti-
nation operand.
The SHRD instruction operates the same as the SHLD instruction except bits are shifted to the
left in the destination operand, with the empty bit positions filled with bits shifted out of the
source operand.

6.7.3. Rotate Instructions

The ROL (rotate left), ROR (rotate right), RCL (rotate through carry left) and RCR (rotate
through carry right) instructions rotate the bits in the destination operand out of one end and
back through the other end (refer to Figure 6-10). Unlike a shift, no bits are lost during a rotation.
The rotate count can range from 0 to 31.

6-32
 INSTRUCTION SET SUMMARY

ROL Instruction
31 0

CF Destination (Memory or Register)

31 ROR Instruction 0
Destination (Memory or Register) CF

RCL Instruction
31 0
CF Destination (Memory or Register)

RCR Instruction
31 0
Destination (Memory or Register) CF

Figure 6-10. ROL, ROR, RCL, and RCR Instruction Operations

The ROL instruction rotates the bits in the operand to the left (toward more significant bit loca-
tions). The ROR instruction rotates the operand right (toward less significant bit locations).
The RCL instruction rotates the bits in the operand to the left, through the CF flag). This instruc-
tion treats the CF flag as a one-bit extension on the upper end of the operand. Each bit which
exits from the most significant bit location of the operand moves into the CF flag. At the same
time, the bit in the CF flag enters the least significant bit location of the operand.
The RCR instruction rotates the bits in the operand to the right through the CF flag.
For all the rotate instructions, the CF flag always contains the value of the last bit rotated out of
the operand, even if the instruction does not use the CF flag as an extension of the operand. The
value of this flag can then be tested by a conditional jump instruction (JC or JNC).

6-33
INSTRUCTION SET SUMMARY

6.8. BIT AND BYTE INSTRUCTIONS

The bit and byte instructions operate on bit or byte strings. They are divided into four groups:
• Bit test and modify instructions.
• Bit scan instructions.
• Byte set on condition.
• Test.

6.8.1. Bit Test and Modify Instructions

The bit test and modify instructions (refer to Table 6-3) operate on a single bit, which can be in
an operand. The location of the bit is specified as an offset from the least significant bit of the
operand. When the processor identifies the bit to be tested and modified, it first loads the CF flag
with the current value of the bit. Then it assigns a new value to the selected bit, as determined
by the modify operation for the instruction.

Table 6-3. Bit Test and Modify Instructions

Instruction Effect on CF Flag Effect on Selected Bit
BT (Bit Test) CF flag ← Selected Bit No effect
BTS (Bit Test and Set) CF flag ← Selected Bit Selected Bit ← 1
BTR (Bit Test and Reset) CF flag ← Selected Bit Selected Bit ← 0
BTC (Bit Test and Complement) CF flag ← Selected Bit Selected Bit ← NOT (Selected Bit)

6.8.2. Bit Scan Instructions

The BSF (bit scan forward) and BSR (bit scan reverse) instructions scan a bit string in a source
operand for a set bit and store the bit index of the first set bit found in a destination register. The
bit index is the offset from the least significant bit (bit 0) in the bit string to the first set bit. The
BSF instruction scans the source operand low-to-high (from bit 0 of the source operand toward
the most significant bit); the BSR instruction scans high-to-low (from the most significant bit
toward the least significant bit).

6.8.3. Byte Set on Condition Instructions

The SETcc (set byte on condition) instructions set a destination-operand byte to 0 or 1,
depending on the state of selected status flags (CF, OF, SF, ZF, and PF) in the EFLAGS register.
The suffix (cc) added to the SET mnemonic determines the condition being tested for. For
example, the SETO instruction tests for overflow. If the OF flag is set, the destination byte is set
to 1; if OF is clear, the destination byte is cleared to 0. Appendix B, EFLAGS Condition Codes
lists the conditions it is possible to test for with this instruction.

6-34
 INSTRUCTION SET SUMMARY

6.8.4. Test Instruction

The TEST instruction performs a logical AND of two operands and sets the SF, ZF, and PF flags
according to the results. The flags can then be tested by the conditional jump or loop instructions
or the SETcc instructions. The TEST instruction differs from the AND instruction in that it does
not alter either of the operands.

6.9. CONTROL TRANSFER INSTRUCTIONS

The processor provides both conditional and unconditional control transfer instructions to direct
the flow of program execution. Conditional transfers are taken only for specified states of the
status flags in the EFLAGS register. Unconditional control transfers are always executed.

6.9.1. Unconditional Transfer Instructions

The JMP, CALL, RET, INT, and IRET instructions transfer program control to another location
(destination address) in the instruction stream. The destination can be within the same code
segment (near transfer) or in a different code segment (far transfer).

6.9.1.1. JUMP INSTRUCTION

The JMP (jump) instruction unconditionally transfers program control to a destination instruc-
tion. The transfer is one-way; that is, a return address is not saved. A destination operand spec-
ifies the address (the instruction pointer) of the destination instruction. The address can be a
relative address or an absolute address.
A relative address is a displacement (offset) with respect to the address in the EIP register. The
destination address (a near pointer) is formed by adding the displacement to the address in the
EIP register. The displacement is specified with a signed integer, allowing jumps either forward
or backward in the instruction stream.
An absolute address is a offset from address 0 of a segment. It can be specified in either of the
following ways:
• An address in a general-purpose register. This address is treated as a near pointer, which
is copied into the EIP register. Program execution then continues at the new address within
the current code segment.
• An address specified using the standard addressing modes of the processor. Here, the
address can be a near pointer or a far pointer. If the address is for a near pointer, the address
is translated into an offset and copied into the EIP register. If the address is for a far
pointer, the address is translated into a segment selector (which is copied into the CS
register) and an offset (which is copied into the EIP register).
In protected mode, the JMP instruction also allows jumps to a call gate, a task gate, and a task-
state segment.

6-35
INSTRUCTION SET SUMMARY

6.9.1.2. CALL AND RETURN INSTRUCTIONS

The CALL (call procedure) and RET (return from procedure) instructions allow a jump from
one procedure (or subroutine) to another and a subsequent jump back (return) to the calling
procedure.
The CALL instruction transfers program control from the current (or calling procedure) to
another procedure (the called procedure). To allow a subsequent return to the calling procedure,
the CALL instruction saves the current contents of the EIP register on the stack before jumping
to the called procedure. The EIP register (prior to transferring program control) contains the
address of the instruction following the CALL instruction. When this address is pushed on the
stack, it is referred to as the return instruction pointer or return address.
The address of the called procedure (the address of the first instruction in the procedure being
jumped to) is specified in a CALL instruction the same way as it is in a JMP instruction (refer
to Section 6.9.1.1., “Jump Instruction”). The address can be specified as a relative address or an
absolute address. If an absolute address is specified, it can be either a near or a far pointer.
The RET instruction transfers program control from the procedure currently being executed (the
called procedure) back to the procedure that called it (the calling procedure). Transfer of control
is accomplished by copying the return instruction pointer from the stack into the EIP register.
Program execution then continues with the instruction pointed to by the EIP register.
The RET instruction has an optional operand, the value of which is added to the contents of the
ESP register as part of the return operation. This operand allows the stack pointer to be incre-
mented to remove parameters from the stack that were pushed on the stack by the calling
procedure.
Refer to Section 4.3., “Calling Procedures Using CALL and RET” in Chapter 4, Procedure
Calls, Interrupts, and Exceptions for more information on the mechanics of making procedure
calls with the CALL and RET instructions.

6.9.1.3. RETURN FROM INTERRUPT INSTRUCTION

When the processor services an interrupt, it performs an implicit call to an interrupt-handling
procedure. The IRET (return from interrupt) instruction returns program control from an inter-
rupt handler to the interrupted procedure (that is, the procedure that was executing when the
interrupt occurred). The IRET instruction performs a similar operation to the RET instruction
(refer to Section 6.9.1.2., “Call and Return Instructions”) except that it also restores the
EFLAGS register from the stack. The contents of the EFLAGS register are automatically stored
on the stack along with the return instruction pointer when the processor services an interrupt.

6.9.2. Conditional Transfer Instructions

The conditional transfer instructions execute jumps or loops that transfer program control to
another instruction in the instruction stream if specified conditions are met. The conditions for
control transfer are specified with a set of condition codes that define various states of the status
flags (CF, ZF, OF, PF, and SF) in the EFLAGS register.

6-36
 INSTRUCTION SET SUMMARY

6.9.2.1. CONDITIONAL JUMP INSTRUCTIONS

The Jcc (conditional) jump instructions transfer program control to a destination instruction if
the conditions specified with the condition code (cc) associated with the instruction are satisfied
(refer to Table 6-4). If the condition is not satisfied, execution continues with the instruction
following the Jcc instruction. As with the JMP instruction, the transfer is one-way; that is, a
return address is not saved.

Table 6-4. Conditional Jump Instructions

Instruction Mnemonic Condition (Flag States) Description
Unsigned Conditional Jumps
JA/JNBE (CF or ZF)=0 Above/not below or equal
JAE/JNB CF=0 Above or equal/not below
JB/JNAE CF=1 Below/not above or equal
JBE/JNA (CF or ZF)=1 Below or equal/not above
JC CF=1 Carry
JE/JZ ZF=1 Equal/zero
JNC CF=0 Not carry
JNE/JNZ ZF=0 Not equal/not zero
JNP/JPO PF=0 Not parity/parity odd
JP/JPE PF=1 Parity/parity even
JCXZ CX=0 Register CX is zero
JECXZ ECX=0 Register ECX is zero
Signed Conditional Jumps
JG/JNLE ((SF xor OF) or ZF) =0 Greater/not less or equal
JGE/JNL (SF xor OF)=0 Greater or equal/not less
JL/JNGE (SF xor OF)=1 Less/not greater or equal
JLE/JNG ((SF xor OF) or ZF)=1 Less or equal/not greater
JNO OF=0 Not overflow
JNS SF=0 Not sign (non-negative)
JO OF=1 Overflow
JS SF=1 Sign (negative)

The destination operand specifies a relative address (a signed offset with respect to the address
in the EIP register) that points to an instruction in the current code segment. The Jcc instructions
do not support far transfers; however, far transfers can be accomplished with a combination of
a Jcc and a JMP instruction (refer to “Jcc—Jump if Condition Is Met” in Chapter 3, Instruction
Set Reference of the Intel Architecture Software Developer’s Manual, Volume 2).

6-37
INSTRUCTION SET SUMMARY

Table 6-4 shows the mnemonics for the Jcc instructions and the conditions being tested for each
instruction. The condition code mnemonics are appended to the letter “J” to form the mnemonic
for a Jcc instruction. The instructions are divided into two groups: unsigned and signed condi-
tional jumps. These groups correspond to the results of operations performed on unsigned and
signed integers, respectively. Those instructions listed as pairs (for example, JA/JNBE) are alter-
nate names for the same instruction. The assembler provides these alternate names to make it
easier to read program listings.
The JCXZ and JECXZ instructions test the CX and ECX registers, respectively, instead of one
or more status flags. Refer to Section 6.9.2.3., “Jump If Zero Instructions” for more informa-
tion about these instructions.

6.9.2.2. LOOP INSTRUCTIONS

The LOOP, LOOPE (loop while equal), LOOPZ (loop while zero), LOOPNE (loop while not
equal), and LOOPNZ (loop while not zero) instructions are conditional jump instructions that
use the value of the ECX register as a count for the number of times to execute a loop. All the
loop instructions decrement the count in the ECX register each time they are executed and termi-
nate a loop when zero is reached. The LOOPE, LOOPZ, LOOPNE, and LOOPNZ instructions
also accept the ZF flag as a condition for terminating the loop before the count reaches zero.
The LOOP instruction decrements the contents of the ECX register (or the CX register, if the
address-size attribute is 16), then tests the register for the loop-termination condition. If the
count in the ECX register is non-zero, program control is transferred to the instruction address
specified by the destination operand. The destination operand is a relative address (that is, an
offset relative to the contents of the EIP register), and it generally points to the first instruction
in the block of code that is to be executed in the loop. When the count in the ECX register
reaches zero, program control is transferred to the instruction immediately following the
LOOP instruction, which terminates the loop. If the count in the ECX register is zero when the
LOOP instruction is first executed, the register is pre-decremented to FFFFFFFFH, causing the
loop to be executed 232 times.
The LOOPE and LOOPZ instructions perform the same operation (they are mnemonics for the
same instruction). These instructions operate the same as the LOOP instruction, except that they
also test the ZF flag. If the count in the ECX register is not zero and the ZF flag is set, program
control is transferred to the destination operand. When the count reaches zero or the ZF flag is
clear, the loop is terminated by transferring program control to the instruction immediately
following the LOOPE/LOOPZ instruction.
The LOOPNE and LOOPNZ instructions (mnemonics for the same instruction) operate the
same as the LOOPE/LOOPPZ instructions, except that they terminate the loop if the ZF flag
is set.

6.9.2.3. JUMP IF ZERO INSTRUCTIONS

The JECXZ (jump if ECX zero) instruction jumps to the location specified in the destination
operand if the ECX register contains the value zero. This instruction can be used in combination
with a loop instruction (LOOP, LOOPE, LOOPZ, LOOPNE, or LOOPNZ) to test the ECX
register prior to beginning a loop. As described in Section 6.9.2.2., “Loop Instructions”, the loop

6-38
 INSTRUCTION SET SUMMARY

instructions decrement the contents of the ECX register before testing for zero. If the value in
the ECX register is zero initially, it will be decremented to FFFFFFFFH on the first loop instruc-
tion, causing the loop to be executed 232 times. To prevent this problem, a JECXZ instruction
can be inserted at the beginning of the code block for the loop, causing a jump out the loop if
the EAX register count is initially zero. When used with repeated string scan and compare
instructions, the JECXZ instruction can determine whether the loop terminated because the
count reached zero or because the scan or compare conditions were satisfied.
The JCXZ (jump if CX is zero) instruction operates the same as the JECXZ instruction when the
16-bit address-size attribute is used. Here, the CX register is tested for zero.

6.9.3. Software Interrupts

The INT n (software interrupt), INTO (interrupt on overflow), and BOUND (detect value out of
range) instructions allow a program to explicitly raise a specified interrupt or exception, which
in turn causes the handler routine for the interrupt or exception to be called.
The INT n instruction can raise any of the processor’s interrupts or exceptions by encoding the
vector number or the interrupt or exception in the instruction. This instruction can be used to
support software generated interrupts or to test the operation of interrupt and exception handlers.
The IRET instruction (refer to Section 6.9.1.3., “Return From Interrupt Instruction”) allows
returns from interrupt handling routines.
The INTO instruction raises the overflow exception, if the OF flag is set. If the flag is clear,
execution continues without raising the exception. This instruction allows software to access the
overflow exception handler explicitly to check for overflow conditions.
The BOUND instruction compares a signed value against upper and lower bounds, and raises
the “BOUND range exceeded” exception if the value is less than the lower bound or greater than
the upper bound. This instruction is useful for operations such as checking an array index to
make sure it falls within the range defined for the array.

6.10. STRING OPERATIONS

The MOVS (Move String), CMPS (Compare string), SCAS (Scan string), LODS (Load string),
and STOS (Store string) instructions permit large data structures, such as alphanumeric char-
acter strings, to be moved and examined in memory. These instructions operate on individual
elements in a string, which can be a byte, word, or doubleword. The string elements to be oper-
ated on are identified with the ESI (source string element) and EDI (destination string element)
registers. Both of these registers contain absolute addresses (offsets into a segment) that point to
a string element.
By default, the ESI register addresses the segment identified with the DS segment register. A
segment-override prefix allows the ESI register to be associated with the CS, SS, ES, FS, or GS
segment register. The EDI register addresses the segment identified with the ES segment
register; no segment override is allowed for the EDI register. The use of two different segment
registers in the string instructions permits operations to be performed on strings located in
different segments. Or by associating the ESI register with the ES segment register, both the

6-39
INSTRUCTION SET SUMMARY

source and destination strings can be located in the same segment. (This latter condition can also
be achieved by loading the DS and ES segment registers with the same segment selector and
allowing the ESI register to default to the DS register.)
The MOVS instruction moves the string element addressed by the ESI register to the location
addressed by the EDI register. The assembler recognizes three “short forms” of this instruction,
which specify the size of the string to be moved: MOVSB (move byte string), MOVSW (move
word string), and MOVSD (move doubleword string).
The CMPS instruction subtracts the destination string element from the source string element
and updates the status flags (CF, ZF, OF, SF, PF, and AF) in the EFLAGS register according to
the results. Neither string element is written back to memory. The assembler recognizes three
“short forms” of the CMPS instruction: CMPSB (compare byte strings), CMPSW (compare
word strings), and CMPSD (compare doubleword strings).
The SCAS instruction subtracts the destination string element from the contents of the EAX,
AX, or AL register (depending on operand length) and updates the status flags according to the
results. The string element and register contents are not modified. The following “short forms”
of the SCAS instruction specifies the operand length: SCASB (scan byte string), SCASW (scan
word string), and SCASD (scan doubleword string).
The LODS instruction loads the source string element identified by the ESI register into the
EAX register (for a doubleword string), the AX register (for a word string), or the AL register
(for a byte string). The “short forms” for this instruction are LODSB (load byte string), LODSW
(load word string), and LODSD (load doubleword string). This instruction is usually used in a
loop, where other instructions process each element of the string after they are loaded into the
target register.
The STOS instruction stores the source string element from the EAX (doubleword string), AX
(word string), or AL (byte string) register into the memory location identified with the EDI
register. The “short forms” for this instruction are STOSB (store byte string), STOSW (store
word string), and STOSD (store doubleword string). This instruction is also normally used in a
loop. Here a string is commonly loaded into the register with a LODS instruction, operated
on by other instructions, and then stored again in memory with a STOS instruction.
The I/O instructions (refer to Section 6.11., “I/O Instructions”) also perform operations on
strings in memory.

6.10.1. Repeating String Operations

The string instructions described in Section 6.10., “String Operations” perform one iteration of
a string operation. To operate strings longer than a doubleword, the string instructions can be
combined with a repeat prefix (REP) to create a repeating instruction or be placed in a loop.
When used in string instructions, the ESI and EDI registers are automatically incremented or
decremented after each iteration of an instruction to point to the next element (byte, word, or
doubleword) in the string. String operations can thus begin at higher addresses and work toward
lower ones, or they can begin at lower addresses and work toward higher ones. The DF flag in

6-40
 INSTRUCTION SET SUMMARY

the EFLAGS register controls whether the registers are incremented (DF=0) or decremented
(DF=1). The STD and CLD instructions set and clear this flag, respectively.
The following repeat prefixes can be used in conjunction with a count in the ECX register to
cause a string instruction to repeat:
• REP—Repeat while the ECX register not zero.
• REPE/REPZ—Repeat while the ECX register not zero and the ZF flag is set.
• REPNE/REPNZ—Repeat while the ECX register not zero and the ZF flag is clear.
When a string instruction has a repeat prefix, the operation executes until one of the termination
conditions specified by the prefix is satisfied. The REPE/REPZ and REPNE/REPNZ prefixes
are used only with the CMPS and SCAS instructions. Also, note that a A REP STOS instruction
is the fastest way to initialize a large block of memory.

6.11. I/O INSTRUCTIONS

The IN (input from port to register), INS (input from port to string), OUT (output from register
to port), and OUTS (output string to port) instructions move data between the processor’s I/O
ports and either a register or memory.
The register I/O instructions (IN and OUT) move data between an I/O port and the EAX register
(32-bit I/O), the AX register (16-bit I/O), or the AL (8-bit I/O) register. The I/O port being read
or written to is specified with an immediate operand or an address in the DX register.
The block I/O instructions (INS and OUTS) instructions move blocks of data (strings) between
an I/O port and memory. These instructions operate similar to the string instructions (refer to
Section 6.10., “String Operations”). The ESI and EDI registers are used to specify string
elements in memory and the repeat prefixes (REP) are used to repeat the instructions to imple-
ment block moves. The assembler recognizes the following alternate mnemonics for these
instructions: INSB (input byte), INSW (input word), and INSD (input doubleword), and OUTB
(output byte), OUTW (output word), and OUTD (output doubleword).
The INS and OUTS instructions use an address in the DX register to specify the I/O port to be
read or written to.

6.12. ENTER AND LEAVE INSTRUCTIONS

The ENTER and LEAVE instructions provide machine-language support for procedure calls in
block-structured languages, such as C and Pascal. These instructions and the call and return
mechanism that they support are described in detail in Section 4.5., “Procedure Calls for Block-
Structured Languages” in Chapter 4, Procedure Calls, Interrupts, and Exceptions.

6-41
INSTRUCTION SET SUMMARY

6.13. EFLAGS INSTRUCTIONS

The EFLAGS instructions allow the state of selected flags in the EFLAGS register to be read or
modified.

6.13.1. Carry and Direction Flag Instructions

The STC (set carry flag), CLC (clear carry flag), and CMC (complement carry flag) instructions
allow the CF flags in the EFLAGS register to be modified directly. They are typically used to
initialize the CF flag to a known state before an instruction that uses the flag in an operation is
executed. They are also used in conjunction with the rotate-with-carry instructions (RCL and
RCR).
The STD (set direction flag) and CLD (clear direction flag) instructions allow the DF flag in the
EFLAGS register to be modified directly. The DF flag determines the direction in which index
registers ESI and EDI are stepped when executing string processing instructions. If the DF flag
is clear, the index registers are incremented after each iteration of a string instruction; if the DF
flag is set, the registers are decremented.

6.13.2. Interrupt Flag Instructions

The STI (set interrupt flag) and CTI (clear interrupt flag) instructions allow the interrupt IF flag
in the EFLAGS register to be modified directly. The IF flag controls the servicing of hardware-
generated interrupts (those received at the processor’s INTR pin). If the IF flag is set, the
processor services hardware interrupts; if the IF flag is clear, hardware interrupts are masked.

6.13.3. EFLAGS Transfer Instructions

The EFLAGS transfer instructions allow groups of flags in the EFLAGS register to be copied
to a register or memory or be loaded from a register or memory.
The LAHF (load AH from flags) and SAHF (store AH into flags) instructions operate on five of
the EFLAGS status flags (SF, ZF, AF, PF, and CF). The LAHF instruction copies the status flags
to bits 7, 6, 4, 2, and 0 of the AH register, respectively. The contents of the remaining bits in the
register (bits 5, 3, and 1) are undefined, and the contents of the EFLAGS register remain
unchanged. The SAHF instruction copies bits 7, 6, 4, 2, and 0 from the AH register into the SF,
ZF, AF, PF, and CF flags, respectively in the EFLAGS register.
The PUSHF (push flags), PUSHFD (push flags double), POPF (pop flags), and POPFD (pop
flags double) instructions copy the flags in the EFLAGS register to and from the stack. The
PUSHF instruction pushes the lower word of the EFLAGS register onto the stack (refer to
Figure 6-11). The PUSHFD instruction pushes the entire EFLAGS register onto the stack (with
the RF and VM flags read as clear).

6-42
 INSTRUCTION SET SUMMARY

PUSHFD/POPFD

PUSHF/POPF

31 30 29 28 27 26 25 24 23 22 21 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0
I
V V
A N O O D I T S Z P C
0 0 0 0 0 0 0 0 0 0 I I I C V R 0 T A
D M F P F F F F F F 0 F 0 F 1 F
P F
L

Figure 6-11. Flags Affected by the PUSHF, POPF, PUSHFD, and POPFD instructions

The POPF instruction pops a word from the stack into the EFLAGS register. Only bits 11, 10,
8, 7, 6, 4, 2, and 0 of the EFLAGS register are affected with all uses of this instruction. If the
current privilege level (CPL) of the current code segment is 0 (most privileged), the IOPL bits
(bits 13 and 12) also are affected. If the I/O privilege level (IOPL) is greater than or equal to the
CPL, numerically, the IF flag (bit 9) also is affected.
The POPFD instruction pops a doubleword into the EFLAGS register. This instruction can
change the state of the AC bit (bit 18) and the ID bit (bit 21), as well as the bits affected by a
POPF instruction. The restrictions for changing the IOPL bits and the IF flag that were given for
the POPF instruction also apply to the POPFD instruction.

6.13.4. Interrupt Flag Instructions

The CLI (clear interrupt flag) and STI (set interrupt flag) instructions clear and set the interrupt
flag (IF) in the EFLAGS register, respectively. Clearing the IF flag causes external interrupts to
be ignored. The ability to execute these instructions depends on the operating mode of the
processor and the current privilege level (CPL) of the program or task attempting to execute
these instructions.

6.14. SEGMENT REGISTER INSTRUCTIONS

The processor provides a variety of instructions that address the segment registers of the
processor directly. These instructions are only used when an operating system or executive is
using the segmented or the real-address mode memory model.

6.14.1. Segment-Register Load and Store Instructions

The MOV instruction (introduced in Section 6.3.1., “General-Purpose Data Movement Instruc-
tions”) and the PUSH and POP instructions (introduced in Section 6.3.2., “Stack Manipulation
Instructions”) can transfer 16-bit segment selectors to and from segment registers (DS, ES, FS,
GS, and SS). The transfers are always made to or from a segment register and a general-purpose
register or memory. Transfers between segment registers are not supported.

6-43
INSTRUCTION SET SUMMARY

The POP and MOV instructions cannot place a value in the CS register. Only the far control-
transfer versions of the JMP, CALL, and RET instructions (refer to Section 6.14.2., “Far Control
Transfer Instructions”) affect the CS register directly.

6.14.2. Far Control Transfer Instructions

The JMP and CALL instructions (refer to Section 6.9., “Control Transfer Instructions”) both
accept a far pointer as a source operand to transfer program control to a segment other than the
segment currently being pointed to by the CS register. When a far call is made with the CALL
instruction, the current values of the EIP and CS registers are both pushed on the stack.
The RET instruction (refer to Section 6.9.1.2., “Call and Return Instructions”) can be used to
execute a far return. Here, program control is transferred from a code segment that contains a
called procedure back to the code segment that contained the calling procedure. The RET
instruction restores the values of the CS and EIP registers for the calling procedure from the
stack.

6.14.3. Software Interrupt Instructions

The software interrupt instructions INT, INTO, BOUND, and IRET (refer to Section 6.9.3.,
“Software Interrupts”) can also call and return from interrupt and exception handler procedures
that are located in a code segment other than the current code segment. With these instructions,
however, the switching of code segments is handled transparently from the application program.

6.14.4. Load Far Pointer Instructions

The load far pointer instructions LDS (load far pointer using DS), LES (load far pointer using
ES), LFS (load far pointer using FS), LGS (load far pointer using GS), and LSS (load far pointer
using SS) load a far pointer from memory into a segment register and a general-purpose general
register. The segment selector part of the far pointer is loaded into the selected segment register
and the offset is loaded into the selected general-purpose register.

6.15. MISCELLANEOUS INSTRUCTIONS

The following instructions perform miscellaneous operations that are of interest to applications
programmers.

6.15.1. Address Computation Instruction

The LEA (load effective address) instruction computes the effective address in memory (offset
within a segment) of a source operand and places it in a general-purpose register. This instruc-
tion can interpret any of the Pentium® Pro processor’s addressing modes and can perform any
indexing or scaling that may be needed. It is especially useful for initializing the ESI or EDI

6-44
 INSTRUCTION SET SUMMARY

registers before the execution of string instructions or for initializing the EBX register before an
XLAT instruction.

6.15.2. Table Lookup Instructions

The XLAT and XLATB (table lookup) instructions replace the contents of the AL register with
a byte read from a translation table in memory. The initial value in the AL register is interpreted
as an unsigned index into the translation table. This index is added to the contents of the EBX
register (which contains the base address of the table) to calculate the address of the table entry.
These instructions are used for applications such as converting character codes from one
alphabet into another (for example, an ASCII code could be used to look up its EBCDIC equiv-
alent in a table).

6.15.3. Processor Identification Instruction

The CPUID (processor identification) instruction provides information about the processor on
which the instruction is executed. To obtain processor information, a value of from 0 to 2 is
loaded in the EAX register and then the CPUID instruction is executed. The resulting processor
information is placed in the EAX, EBX, ECX, and EDX registers. Table 6-5 shows the informa-
tion that is provided depending on the value initially entered in the EAX register. Refer to
Section 11.1., “Processor Identification” in Chapter 11, Processor Identification and Feature
Determination for detailed information on the output of the CPUID instruction.

Table 6-5. Information Provided by the CPUID Instruction

Initial EAX Value Information Provided about the Processor
0 Maximum CPUID input value.
Vendor identification string (“GenuineIntel”).
1 Version information (family ID, model ID, and stepping ID).
Feature information (identifies the feature set for the processor model).
2 Cache information (about the processor’s internal cache memory).

6.15.4. No-Operation and Undefined Instructions

The NOP (no operation) instruction increments the EIP register to point at the next instruction,
but affects nothing else.
The UD2 (undefined) instruction generates an invalid opcode exception. Intel reserves the
opcode for this instruction for this function. The instruction is provided to allow software to test
an invalid opcode exception handler.

6-45
INSTRUCTION SET SUMMARY

6-46
 7
Floating-Point Unit
 CHAPTER 7
FLOATING-POINT UNIT

The Intel Architecture (IA) Floating-Point Unit (FPU) provides high-performance floating-
point processing capabilities. It supports the real, integer, and BCD-integer data types and the
floating-point processing algorithms and exception handling architecture defined in the IEEE
754 and 854 Standards for Floating-Point Arithmetic. The FPU executes instructions from the
processor’s normal instruction stream and greatly improves the efficiency of IA processors in
handling the types of high-precision floating-point processing operations commonly found in
scientific, engineering, and business applications.
This chapter describes the data types that the FPU operates on, the FPU’s execution environ-
ment, and the FPU-specific instruction set. Detailed descriptions of the FPU instructions are
given in Chapter 3, Instruction Set Reference, in the Intel Architecture Software Developer’s
Manual, Volume 2.

7.1. COMPATIBILITY AND EASE OF USE OF THE INTEL

ARCHITECTURE FPU
The architecture of the IA FPU has evolved in parallel with the architecture of early IA proces-
sors. The first Intel Math Coprocessors (the Intel 8087, Intel 287, and Intel 387) were companion
processors to the Intel 8086/8088, Intel 286, and Intel386™ processors, respectively, and were
designed to improve and extend the numeric processing capability of the IA. The Intel486™ DX
processor for the first time integrated the CPU and the FPU architectures on one chip. The
Pentium® processor’s FPU offered the same architecture as the Intel486™ DX processor’s FPU,
but with improved performance. The Pentium® Pro processor’s FPU further extended the
floating-point processing capability of IA family of processors and added several new instruc-
tions to improve processing throughput.
Throughout this evolution, compatibility among the various generations of FPUs and math
coprocessors has been maintained. For example, the Pentium® Pro processor’s FPU is fully
compatible with the Pentium® and Intel486™ DX processors’s FPUs.
Each generation of the IA FPUs have been explicitly designed to deliver stable, accurate results
when programmed using straightforward “pencil and paper” algorithms, bringing the function-
ality and power of accurate numeric computation into the hands of the general user. The IEEE
754 standard specifically addresses this issue, recognizing the fundamental importance of
making numeric computations both easy and safe to use.
For example, some processors can overflow when two single-precision floating-point numbers
are multiplied together and then divided by a third, even if the final result is a perfectly valid 32-
bit number. The IA FPUs deliver the correctly rounded result. Other typical examples of unde-
sirable machine behavior in straightforward calculations occur when computing financial rate
of return, which involves the expression (1 + i) n or when solving for roots of a quadratic equa-
tion:

7-1
FLOATING-POINT UNIT

2
– b ± b – 4ac
---------------------------------------
2a
If a does not equal 0, the formula is numerically unstable when the roots are nearly coincident
or when their magnitudes are wildly different. The formula is also vulnerable to spurious
over/underflows when the coefficients a, b, and c are all very big or all very tiny. When single-
precision (4-byte) floating-point coefficients are given as data and the formula is evaluated in
the FPU’s normal way, keeping all intermediate results in its stack, the FPU produces impec-
cable single-precision roots. This happens because, by default and with no effort on the
programmer’s part, the FPU evaluates all those sub-expressions with so much extra precision
and range as to overwhelm almost any threat to numerical integrity.
If double-precision data and results were at issue, a better formula would have to be used, and
once again the FPU’s default evaluation of that formula would provide substantially enhanced
numerical integrity over mere double-precision evaluation.
On most machines, straightforward algorithms will not deliver consistently correct results (and
will not indicate when they are incorrect). To obtain correct results on traditional machines
under all conditions usually requires sophisticated numerical techniques that go beyond typical
programming practice. General application programmers using straightforward algorithms will
produce much more reliable programs using the IAs. This simple fact greatly reduces the soft-
ware investment required to develop safe, accurate computation-based products.
Beyond traditional numeric support for scientific applications, the IA processors have built-in
facilities for commercial computing. They can process decimal numbers of up to 18 digits
without round-off errors, performing exact arithmetic on integers as large as 264 (or 1018).
Exact arithmetic is vital in accounting applications where rounding errors may introduce mone-
tary losses that cannot be reconciled.
The Intel FPU’s contain a number of optional numerical facilities that can be invoked by sophis-
ticated users. These advanced features include directed rounding, gradual underflow, and
programmed exception-handling facilities.
These automatic exception-handling facilities permit a high degree of flexibility in numeric
processing software, without burdening the programmer. While performing numeric calcula-
tions, the processor automatically detects exception conditions that can potentially damage a
calculation (for example, X ÷ 0 or X when X < 0). By default, on-chip exception logic handles
these exceptions so that a reasonable result is produced and execution may proceed without
program interruption. Alternatively, the processor can invoke a software exception handler to
provide special results whenever various types of exceptions are detected.

7.2. REAL NUMBERS AND FLOATING-POINT FORMATS

This section describes how real numbers are represented in floating-point format in the IA FPU.
It also introduces terms such as normalized numbers, denormalized numbers, biased exponents,
signed zeros, and NaNs. Readers who are already familiar with floating-point processing tech-
niques and the IEEE standards may wish to skip this section.

7-2
 FLOATING-POINT UNIT

7.2.1. Real Number System

As shown in Figure 7-1, the real-number system comprises the continuum of real numbers from
minus infinity (−∞) to plus infinity (+∞).

Binary Real Number System

-100 -10 -1 0 1 10 100
ςς ςς

Subset of binary real-numbers that can be represented with

IEEE single-precision (32-bit) floating-point format
-100 -10 -1 0 1 10 100
ςς ςς

+10

10.0000000000000000000000
1.11111111111111111111111
Precision 24 Binary Digits

Numbers within this range

cannot be represented.

Figure 7-1. Binary Real Number System

Because the size and number of registers that any computer can have is limited, only a subset of
the real-number continuum can be used in real-number calculations. As shown at the bottom of
Figure 7-1, the subset of real numbers that a particular FPU supports represents an approxima-
tion of the real number system. The range and precision of this real-number subset is determined
by the format that the FPU uses to represent real numbers.

7-3
FLOATING-POINT UNIT

7.2.2. Floating-Point Format

To increase the speed and efficiency of real-number computations, computers or FPUs typically
represent real numbers in a binary floating-point format. In this format, a real number has three
parts: a sign, a significand, and an exponent. Figure 7-2 shows the binary floating-point format
that the IA FPU uses. This format conforms to the IEEE standard.
The sign is a binary value that indicates whether the number is positive (0) or negative (1).
The significand has two parts: a 1-bit binary integer (also referred to as the J-bit) and a binary
fraction. The J-bit is often not represented, but instead is an implied value. The exponent is a
binary integer that represents the base-2 power that the significand is raised to.

Sign
Exponent Significand

Fraction

Integer or J-Bit

Figure 7-2. Binary Floating-Point Format

Table 7-1 shows how the real number 178.125 (in ordinary decimal format) is stored in floating-
point format. The table lists a progression of real number notations that leads to the single-real,
32-bit floating-point format (which is one of the floating-point formats that the FPU supports).
In this format, the significand is normalized (refer to Section 7.2.2.1., “Normalized Numbers”)
and the exponent is biased (refer to Section 7.2.2.2., “Biased Exponent”). For the single-real
format, the biasing constant is +127.

7.2.2.1. NORMALIZED NUMBERS

In most cases, the FPU represents real numbers in normalized form. This means that except for
zero, the significand is always made up of an integer of 1 and the following fraction:
1.fff...ff
For values less than 1, leading zeros are eliminated. (For each leading zero eliminated, the expo-
nent is decremented by one.)

7-4
 FLOATING-POINT UNIT

Table 7-1. Real Number Notation

Notation Value
Ordinary Decimal 178.125
Scientific Decimal 1.78125E102
Scientific Binary 1.0110010001E2111
Scientific Binary 1.0110010001E210000110
(Biased Exponent)
Single-Real Format Sign Biased Exponent Normalized Significand
0 10000110 01100100010000000000000
1. (Implied)

Representing numbers in normalized form maximizes the number of significant digits that can
be accommodated in a significand of a given width. To summarize, a normalized real number
consists of a normalized significand that represents a real number between 1 and 2 and an expo-
nent that specifies the number’s binary point.

7.2.2.2. BIASED EXPONENT

The FPU represents exponents in a biased form. This means that a constant is added to the actual
exponent so that the biased exponent is always a positive number. The value of the biasing
constant depends on the number of bits available for representing exponents in the floating-point
format being used. The biasing constant is chosen so that the smallest normalized number can
be reciprocated without overflow.
(Refer to Section 7.4.1., “Real Numbers” for a list of the biasing constants that the FPU uses for
the various sizes of real data-types.)

7.2.3. Real Number and Non-number Encodings

A variety of real numbers and special values can be encoded in the FPU’s floating-point format.
These numbers and values are generally divided into the following classes:
• Signed zeros.
• Denormalized finite numbers.
• Normalized finite numbers.
• Signed infinities.
• NaNs.
• Indefinite numbers.
(The term NaN stands for “Not a Number.”)

7-5
FLOATING-POINT UNIT

Figure 7-3 shows how the encodings for these numbers and non-numbers fit into the real number
continuum. The encodings shown here are for the IEEE single-precision (32-bit) format, where
the term “S” indicates the sign bit, “E” the biased exponent, and “F” the fraction. (The exponent
values are given in decimal.)
The FPU can operate on and/or return any of these values, depending on the type of computation
being performed. The following sections describe these number and non-number classes.

7.2.3.1. SIGNED ZEROS

Zero can be represented as a +0 or a −0 depending on the sign bit. Both encodings are equal in
value. The sign of a zero result depends on the operation being performed and the rounding
mode being used. Signed zeros have been provided to aid in implementing interval arithmetic.
The sign of a zero may indicate the direction from which underflow occurred, or it may indicate
the sign of an ∞ that has been reciprocated.

7.2.3.2. NORMALIZED AND DENORMALIZED FINITE NUMBERS

Non-zero, finite numbers are divided into two classes: normalized and denormalized. The
normalized finite numbers comprise all the non-zero finite values that can be encoded in a
normalized real number format between zero and ∞. In the single-real format shown in Figure
7-3, this group of numbers includes all the numbers with biased exponents ranging from 1 to
25410 (unbiased, the exponent range is from −12610 to +12710).

NaN NaN
−Denormalized Finite +Denormalized Finite
−∞ −Normalized Finite −0 +0 +Normalized Finite +∞

Real Number and NaN Encodings For 32-Bit Floating-Point Format

S E F S E F
1 0 0 −0 +0 0 0 0

−Denormalized +Denormalized
1 0 0.XXX2 Finite Finite 0 0 0.XXX2
−Normalized +Normalized 0 1...254 Any Value
1 1...254 Any Value Finite Finite
1 255 0 −∞ +∞ 0 255 0

X1 255 1.0XX2 −SNaN +SNaN X1 255 1.0XX2

X1 255 1.1XX −QNaN +QNaN X1 255 1.1XX

NOTES:
1. Sign bit ignored.
2. Fractions must be non-zero.

Figure 7-3. Real Numbers and NaNs

7-6
 FLOATING-POINT UNIT

When real numbers become very close to zero, the normalized-number format can no longer be
used to represent the numbers. This is because the range of the exponent is not large enough to
compensate for shifting the binary point to the right to eliminate leading zeros.
When the biased exponent is zero, smaller numbers can only be represented by making the
integer bit (and perhaps other leading bits) of the significand zero. The numbers in this range are
called denormalized (or tiny) numbers. The use of leading zeros with denormalized numbers
allows smaller numbers to be represented. However, this denormalization causes a loss of preci-
sion (the number of significant bits in the fraction is reduced by the leading zeros).
When performing normalized floating-point computations, an FPU normally operates on
normalized numbers and produces normalized numbers as results. Denormalized numbers
represent an underflow condition.
A denormalized number is computed through a technique called gradual underflow. Table 7-2
gives an example of gradual underflow in the denormalization process. Here the single-real
format is being used, so the minimum exponent (unbiased) is −12610. The true result in this
example requires an exponent of −12910 in order to have a normalized number. Since −12910
is beyond the allowable exponent range, the result is denormalized by inserting leading zeros
until the minimum exponent of −12610 is reached.

Table 7-2. Denormalization Process

Operation Sign Exponent* Significand
True Result 0 −129 1.01011100000...00
Denormalize 0 −128 0.10101110000...00
Denormalize 0 −127 0.01010111000...00
Denormalize 0 −126 0.00101011100...00
Denormal Result 0 −126 0.00101011100...00

NOTE:
* Expressed as an unbiased, decimal number.

In the extreme case, all the significant bits are shifted out to the right by leading zeros, creating
a zero result.
The FPU deals with denormal values in the following ways:
• It avoids creating denormals by normalizing numbers whenever possible.
• It provides the floating-point underflow exception to permit programmers to detect cases
when denormals are created.
• It provides the floating-point denormal operand exception to permit procedures or
programs to detect when denormals are being used as source operands for computations.
When a denormal number in single- or double-real format is used as a source operand and the
denormal exception is masked, the FPU automatically normalizes the number when it is
converted to extended-real format.

7-7
FLOATING-POINT UNIT

7.2.3.3. SIGNED INFINITIES

The two infinities, +∞ and −∞, represent the maximum positive and negative real numbers,
respectively, that can be represented in the floating-point format. Infinity is always represented
by a zero significand (fraction and integer bit) and the maximum biased exponent allowed in the
specified format (for example, 255 10 for the single-real format).
The signs of infinities are observed, and comparisons are possible. Infinities are always inter-
preted in the affine sense; that is, –∞ is less than any finite number and +∞ is greater than any
finite number. Arithmetic on infinities is always exact. Exceptions are generated only when the
use of an infinity as a source operand constitutes an invalid operation.
Whereas denormalized numbers represent an underflow condition, the two infinity numbers
represent the result of an overflow condition. Here, the normalized result of a computation has
a biased exponent greater than the largest allowable exponent for the selected result format.

7.2.3.4. NANS
Since NaNs are non-numbers, they are not part of the real number line. In Figure 7-3, the
encoding space for NaNs in the FPU floating-point formats is shown above the ends of the real
number line. This space includes any value with the maximum allowable biased exponent and
a non-zero fraction. (The sign bit is ignored for NaNs.)
The IEEE standard defines two classes of NaN: quiet NaNs (QNaNs) and signaling NaNs
(SNaNs). A QNaN is a NaN with the most significant fraction bit set; an SNaN is a NaN with
the most significant fraction bit clear. QNaNs are allowed to propagate through most arithmetic
operations without signaling an exception. SNaNs generally signal an invalid operation excep-
tion whenever they appear as operands in arithmetic operations. Exceptions are discussed in
Section 7.7., “Floating-Point Exception Handling”.
Refer to Section 7.6., “Operating on NaNs”, for detailed information on how the FPU handles
NaNs.

7.2.4. Indefinite
For each FPU data type, one unique encoding is reserved for representing the special value
indefinite. For example, when operating on real values, the real indefinite value is a QNaN
(refer to Section 7.4.1., “Real Numbers”). The FPU produces indefinite values as responses
to masked floating-point exceptions.

7.3. FPU ARCHITECTURE

From an abstract, architectural view, the FPU is a coprocessor that operates in parallel with the
processor’s integer unit (refer to Figure 7-4). The FPU gets its instructions from the same
instruction decoder and sequencer as the integer unit and shares the system bus with the integer
unit. Other than these connections, the integer unit and FPU operate independently and in
parallel. (The actual microarchitecture of an IA processor varies among the various families of
processors. For example, the Pentium® Pro processor has two integer units and two FPUs;

7-8
 FLOATING-POINT UNIT

whereas, the Pentium® processor has two integer units and one FPU, and the Intel486™
processor has one integer unit and one FPU.)

Instruction
Decoder and
Sequencer

Integer
FPU
Unit

Data Bus

Figure 7-4. Relationship Between the Integer Unit and the FPU

The instruction execution environment of the FPU (refer to Figure 7-5) consists of 8 data regis-
ters (called the FPU data registers) and the following special-purpose registers:
• The status register.
• The control register.
• The tag word register.
• Instruction pointer register.
• Last operand (data pointer) register.
• Opcode register.
These registers are described in the following sections.

7.3.1. FPU Data Registers

The FPU data registers (shown in Figure 7-5) consist of eight 80-bit registers. Values are stored
in these registers in the extended-real format shown in Figure 7-17. When real, integer, or
packed BCD integer values (in any of the formats shown in Figure 7-17) are loaded from
memory into any of the FPU data registers, the values are automatically converted into
extended-real format (if they are not already in that format). When computation results are
subsequently transferred back into memory from any of the FPU registers, the results can be left
in the extended-real format or converted back into one of the other FPU formats (real, integer,
or packed BCD integers) shown in Figure 7-17.
The FPU instructions treat the eight FPU data registers as a register stack (refer to Figure 7-6).
All addressing of the data registers is relative to the register on the top of the stack. The register
number of the current top-of-stack register is stored in the TOP (stack TOP) field in the FPU
status word. Load operations decrement TOP by one and load a value into the new top-of-stack
register, and store operations store the value from the current TOP register in memory and then

7-9
FLOATING-POINT UNIT

increment TOP by one. (For the FPU, a load operation is equivalent to a push and a store oper-
ation is equivalent to a pop.)

FPU Data Registers

Sign 79 78 64 63 0
R7 Exponent Significand
R6
R5
R4
R3
R2
R1
R0

15 0 47 0
Control FPU Instruction Pointer
Register

Status FPU Operand (Data) Pointer

Register

Tag 10 0
Register
Opcode

Figure 7-5. FPU Execution Environment

If a load operation is performed when TOP is at 0, register wraparound occurs and the new value
of TOP is set to 7. The floating-point stack-overflow exception indicates when wraparound
might cause an unsaved value to be overwritten (refer to Section 7.8.1.1., “Stack Overflow or
Underflow Exception (#IS)”).

FPU Data Register Stack

7
6
Growth
Stack 5 ST(2)
4 ST(1) Top
3 ST(0) 011B
2
1
0

Figure 7-6. FPU Data Register Stack

7-10
 FLOATING-POINT UNIT

Many floating-point instructions have several addressing modes that permit the programmer to
implicitly operate on the top of the stack, or to explicitly operate on specific registers relative to
the TOP. Assemblers supports these register addressing modes, using the expression ST(0), or
simply ST, to represent the current stack top and ST(i) to specify the ith register from TOP in
the stack (0 ≤ i ≤ 7). For example, if TOP contains 011B (register 3 is the top of the stack), the
following instruction would add the contents of two registers in the stack (registers 3 and 5):

FADD ST, ST(2);

Figure 7-7 shows an example of how the stack structure of the FPU registers and instructions
are typically used to perform a series of computations. Here, a two-dimensional dot product is
computed, as follows:
1. The first instruction (FLD value1) decrements the stack register pointer (TOP) and loads
the value 5.6 from memory into ST(0). The result of this operation is shown in snap-shot
(a).
2. The second instruction multiplies the value in ST(0) by the value 2.4 from memory and
stores the result in ST(0), shown in snap-shot (b).
3. The third instruction decrements TOP and loads the value 3.8 in ST(0).
4. The fourth instruction multiplies the value in ST(0) by the value 10.3 from memory and
stores the result in ST(0), shown in snap-shot (c).
5. The fifth instruction adds the value and the value in ST(1) and stores the result in ST(0),
shown in snap-shot (d).
The style of programming demonstrated in this example is supported by the floating-point
instruction set. In cases where the stack structure causes computation bottlenecks, the FXCH
(exchange FPU register contents) instruction can be used to streamline a computation.

7.3.1.1. PARAMETER PASSING WITH THE FPU REGISTER STACK

Like the general-purpose registers in the processor’s integer unit, the contents of the FPU data
registers are unaffected by procedure calls, or in other words, the values are maintained across
procedure boundaries. A calling procedure can thus use the FPU data registers (as well as the
procedure stack) for passing parameter between procedures. The called procedure can reference
parameters passed through the register stack using the current stack register pointer (TOP) and
the ST(0) and ST(i) nomenclature. It is also common practice for a called procedure to leave a
return value or result in register ST(0) when returning execution to the calling procedure or
program.

7-11
FLOATING-POINT UNIT

Computation
Dot Product = (5.6 x 2.4) + (3.8 x 10.3)

Code:
FLD value1 ;(a) value1=5.6
FMUL value2 ;(b) value2=2.4
FLD value3 ; value3=3.8
FMUL value4 ;(c)value4=10.3
FADD ST(1) ;(d)

(a) (b) (c) (d)

R7 R7 R7 R7
R6 R6 R6 R6
R5 R5 R5 R5
R4 5.6 ST(0) R4 13.44 ST(0) R4 13.44 ST(1) R4 13.44 ST(1)
R3 R3 R3 39.14 ST(0) R3 52.58 ST(0)
R2 R2 R2 R2
R1 R1 R1 R1
R0 R0 R0 R0

Figure 7-7. Example FPU Dot Product Computation

7.3.2. FPU Status Register

The 16-bit FPU status register (refer to Figure 7-8) indicates the current state of the FPU. The
flags in the FPU status register include the FPU busy flag, top-of-stack (TOP) pointer, condition
code flags, error summary status flag, stack fault flag, and exception flags. The FPU sets the
flags in this register to show the results of operations.
The contents of the FPU status register (referred to as the FPU status word) can be stored in
memory using the FSTSW/FNSTSW, FSTENV/FNSTENV, and FSAVE/FNSAVE instructions.
It can also be stored in the AX register of the integer unit, using the FSTSW/FNSTSW
instructions.

7.3.2.1. TOP OF STACK (TOP) POINTER

A pointer to the FPU data register that is currently at the top of the FPU register stack is
contained in bits 11 through 13 of the FPU status word. This pointer, which is commonly
referred to as TOP (for top-of-stack), is a binary value from 0 to 7. Refer to Section 7.3.1.,
“FPU Data Registers”, for more information about the TOP pointer.

7.3.2.2. CONDITION CODE FLAGS

The four FPU condition code flags (C0 through C3) indicate the results of floating-point
comparison and arithmetic operations. Table 7-3 summarizes the manner in which the floating-

7-12
 FLOATING-POINT UNIT

point instructions set the condition code flags. These condition code bits are used principally for
conditional branching and for storage of information used in exception handling (refer to
Section 7.3.3., “Branching and Conditional Moves on FPU Condition Codes”).

FPU Busy
Top of Stack Pointer

15 14 13 11 10 9 8 7 6 5 4 3 2 1 0

B C C C C E S P U O Z D I
TOP
3 2 1 0 S F E E E E E E

Condition
Code
Error Summary Status
Stack Fault
Exception Flags
Precision
Underflow
Overflow
Zero Divide
Denormalized Operand
Invalid Operation

Figure 7-8. FPU Status Word

As shown in Table 7-3, the C1 condition code flag is used for a variety of functions. When both
the IE and SF flags in the FPU status word are set, indicating a stack overflow or underflow
exception (#IS), the C1 flag distinguishes between overflow (C1=1) and underflow (C1=0).
When the PE flag in the status word is set, indicating an inexact (rounded) result, the C1 flag is
set to 1 if the last rounding by the instruction was upward. The FXAM instruction sets C1 to the
sign of the value being examined.

The C2 condition code flag is used by the FPREM and FPREM1 instructions to indicate an
incomplete reduction (or partial remainder). When a successful reduction has been completed,
the C0, C3, and C1 condition code flags are set to the three least-significant bits of the quotient
(Q2, Q1, and Q0, respectively). Refer to “FPREM1—Partial Remainder” in Chapter 3, Instruc-
tion Set Reference, of the Intel Architecture Software Developer’s Manual, Volume 2, for more
information on how these instructions use the condition code flags.
The FPTAN, FSIN, FCOS, and FSINCOS instructions set the C2 flag to 1 to indicate that the
source operand is beyond the allowable range of ±263.
Where the state of the condition code flags are listed as undefined in Table 7-3, do not rely on
any specific value in these flags.

7-13
FLOATING-POINT UNIT

Table 7-3. FPU Condition Code Interpretation

Instruction C0 C3 C2 C1
FCOM, FCOMP, FCOMPP, Result of Comparison Operands 0 or #IS
FICOM, FICOMP, FTST, are not
FUCOM, FUCOMP, Comparable
FUCOMPP
FCOMI, FCOMIP, FUCOMI, Undefined. (These instructions set the #IS
FUCOMIP status flags in the EFLAGS register.)
FXAM Operand class Sign
FPREM, FPREM1 Q2 Q1 0=reduction Q0 or #IS
complete
1=reduction
incomplete
F2XM1, FADD, FADDP, Undefined Roundup or #IS
FBSTP, FCMOVcc, FIADD,
FDIV, FDIVP, FDIVR,
FDIVRP, FIDIV, FIDIVR,
FIMUL, FIST, FISTP, FISUB,
FISUBR,FMUL, FMULP,
FPATAN, FRNDINT,
FSCALE, FST, FSTP, FSUB,
FSUBP, FSUBR,
FSUBRP,FSQRT, FYL2X,
FYL2XP1
FCOS, FSIN, FSINCOS, Undefined 1=source Roundup or #IS
FPTAN operand out of (Undefined if
range. C2=1)
FABS, FBLD, FCHS, Undefined 0 or #IS
FDECSTP, FILD, FINCSTP,
FLD, Load Constants, FSTP
(ext. real), FXCH, FXTRACT
FLDENV, FRSTOR Each bit loaded from memory
FFREE, FLDCW,
FCLEX/FNCLEX, FNOP, Undefined
FSTCW/FNSTCW,
FSTENV/FNSTENV,
FSTSW/FNSTSW,
FINIT/FNINIT, 0 0 0 0
FSAVE/FNSAVE

7.3.2.3. EXCEPTION FLAGS

The six exception flags (bits 0 through 5) of the status word indicate that one or more floating-
point exceptions has been detected since the bits were last cleared. The individual exception
flags (IE, DE, ZE, OE, UE, and PE) are described in detail in Section 7.7., “Floating-Point
Exception Handling”, Each of the exception flags can be masked by an exception mask bit in
the FPU control word (refer to Section 7.3.4., “FPU Control Word”). The exception summary
status (ES) flag (bit 7) is set when any of the unmasked exception flags are set. When the ES

7-14
 FLOATING-POINT UNIT

flag is set, the FPU exception handler is invoked, using one of the techniques described in
Section 7.7.3., “Software Exception Handling”. (Note that if an exception flag is masked, the
FPU will still set the flag if its associated exception occurs, but it will not set the ES flag.)
The exception flags are “sticky” bits, meaning that once set, they remain set until explicitly
cleared. They can be cleared by executing the FCLEX/FNCLEX (clear exceptions) instructions,
by reinitializing the FPU with the FINIT/FNINIT or FSAVE/FNSAVE instructions, or by over-
writing the flags with an FRSTOR or FLDENV instruction.
The B-bit (bit 15) is included for 8087 compatibility only. It reflects the contents of the ES flag.

7.3.2.4. STACK FAULT FLAG

The stack fault flag (bit 6 of the FPU status word) indicates that stack overflow or stack under-
flow has occurred. The FPU explicitly sets the SF flag when it detects a stack overflow or under-
flow condition, but it does not explicitly clear the flag when it detects an invalid-arithmetic-
operand condition. When this flag is set, the condition code flag C1 indicates the nature of the
fault: overflow (C1 = 1) and underflow (C1 = 0). The SF flag is a “sticky” flag, meaning that
after it is set, the processor does not clear it until it is explicitly instructed to do so (for example,
by an FINIT/FNINIT, FCLEX/FNCLEX, or FSAVE/FNSAVE instruction).
Refer to Section 7.3.6., “FPU Tag Word” for more information on FPU stack faults.

7.3.3. Branching and Conditional Moves on FPU Condition

Codes
The IA FPU (beginning with the Pentium® Pro processor) supports two mechanisms for
branching and performing conditional moves according to comparisons of two floating-point
values. These mechanism are referred to here as the “old mechanism” and the “new mecha-
nism.”
The old mechanism is available in FPU’s prior to the Pentium® Pro processor and in the
Pentium® Pro processor. This mechanism uses the floating-point compare instructions (FCOM,
FCOMP, FCOMPP, FTST, FUCOMPP, FICOM, and FICOMP) to compare two floating-point
values and set the condition code flags (C0 through C3) according to the results. The contents
of the condition code flags are then copied into the status flags of the EFLAGS register using a
two step process (refer to Figure 7-9):
1. The FSTSW AX instruction moves the FPU status word into the AX register.
2. The SAHF instruction copies the upper 8 bits of the AX register, which includes the
condition code flags, into the lower 8 bits of the EFLAGS register.
When the condition code flags have been loaded into the EFLAGS register, conditional jumps
or conditional moves can be performed based on the new settings of the status flags in the
EFLAGS register.

7-15
FLOATING-POINT UNIT

15 FPU Status Word 0

Condition Status
Code C C C C
Flag 3 2 1 0
C0 CF
C1 (none)
C2 PF FSTSW AX Instruction
C3 ZF 15 AX Register 0
C C C C
3 2 1 0

SAHF Instruction

31 EFLAGS Register 7 0
Z P C
F F 1 F

Figure 7-9. Moving the FPU Condition Codes to the EFLAGS Register

The new mechanism is available only in the Pentium® Pro processor. Using this mechanism, the
new floating-point compare and set EFLAGS instructions (FCOMI, FCOMIP, FUCOMI, and
FUCOMIP) compare two floating-point values and set the ZF, PF, and CF flags in the EFLAGS
register directly. A single instruction thus replaces the three instructions required by the old
mechanism.
Note also that the FCMOVcc instructions (also new in the Pentium® Pro processor) allow condi-
tional moves of floating-point values (values in the FPU data registers) based on the setting of
the status flags (ZF, PF, and CF) in the EFLAGS register. These instructions eliminate the need
for an IF statement to perform conditional moves of floating-point values.

7.3.4. FPU Control Word

The 16-bit FPU control word (refer to Figure 7-10) controls the precision of the FPU and
rounding method used. It also contains the exception-flag mask bits. The control word is cached
in the FPU control register. The contents of this register can be loaded with the FLDCW instruc-
tion and stored in memory with the FSTCW/FNSTCW instructions.
When the FPU is initialized with either an FINIT/FNINIT or FSAVE/FNSAVE instruction, the
FPU control word is set to 037FH, which masks all floating-point exceptions, sets rounding to
nearest, and sets the FPU precision to 64 bits.

7-16
 FLOATING-POINT UNIT

Infinity Control
Rounding Control
Precision Control

15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

P U O Z D I
X RC PC M M M M M M

Exception Masks
Precision
Underflow
Overflow
Zero Divide
Denormalized Operand
Invalid Operation

Reserved

Figure 7-10. FPU Control Word

7.3.4.1. EXCEPTION-FLAG MASKS

The exception-flag mask bits (bits 0 through 5 of the FPU control word) mask the 6 exception
flags in the FPU status word (also bits 0 through 5). When one of these mask bits is set, its corre-
sponding floating-point exception is blocked from being generated.

7.3.4.2. PRECISION CONTROL FIELD

The precision-control (PC) field (bits 8 and 9 of the FPU control word) determines the precision
(64, 53, or 24 bits) of floating-point calculations made by the FPU (refer to Table 7-4). The
default precision is extended precision, which uses the full 64-bit significand available with the
extended-real format of the FPU data registers, but is configurable by the user, compiler, or oper-
ating system. This setting is best suited for most applications, because it allows applications to
take full advantage of the precision of the extended-real format.

Table 7-4. Precision Control Field (PC)

Precision PC Field
*
Single Precision (24-Bits ) 00B
Reserved 01B
*
Double Precision (53-Bits ) 10B
Extended Precision (64-Bits) 11B

NOTE:
*
Includes the implied integer bit.

7-17
FLOATING-POINT UNIT

The double precision and single precision settings, reduce the size of the significand to 53 bits
and 24 bits, respectively. These settings are provided to support the IEEE standard and to allow
exact replication of calculations which were done using the lower precision data types. Using
these settings nullifies the advantages of the extended-real format’s 64-bit significand length.
When reduced precision is specified, the rounding of the significand value clears the unused bits
on the right to zeros.
The precision-control bits only affect the results of the following floating-point instructions:
FADD, FADDP, FSUB, FSUBP, FSUBR, FSUBRP, FMUL, FMULP, FDIV, FDIVP, FDIVR,
FDIVRP, and FSQRT.

7.3.4.3. ROUNDING CONTROL FIELD

The rounding control (RC) field of the FPU control register (bits 10 and 11) controls how the
results of floating-point instructions are rounded. Four rounding modes are supported (refer to
Table 7-5): round to nearest, round up, round down, and round toward zero. Round to nearest is
the default rounding mode and is suitable for most applications. It provides the most accurate
and statistically unbiased estimate of the true result.

Table 7-5. Rounding Control Field (RC)

Rounding RC Field
Mode Setting Description
Round to 00B Rounded result is the closest to the infinitely precise result. If two values
nearest (even) are equally close, the result is the even value (that is, the one with the
least-significant bit of zero).
Round down 01B Rounded result is close to but no greater than the infinitely precise
(toward −∞) result.
Round up 10B Rounded result is close to but no less than the infinitely precise result.
(toward +∞)
Round toward 11B Rounded result is close to but no greater in absolute value than the
zero (Truncate) infinitely precise result.

The round up and round down modes are termed directed rounding and can be used to imple-
ment interval arithmetic. Interval arithmetic is used to determine upper and lower bounds for the
true result of a multistep computation, when the intermediate results of the computation are
subject to rounding.
The round toward zero mode (sometimes called the “chop” mode) is commonly used when
performing integer arithmetic with the FPU.
Whenever possible, the FPU produces an infinitely precise result in the destination format
(single, double, or extended real). However, it is often the case that the infinitely precise result
of an arithmetic or store operation cannot be encoded exactly in the format of the destination
operand.

7-18
 FLOATING-POINT UNIT

For example, the following value (a) has a 24-bit fraction. The least-significant bit of this frac-
tion (the underlined bit) cannot be encoded exactly in the single-real format (which has only a
23-bit fraction):
(a) 1.0001 0000 1000 0011 1001 0111E2 101
To round this result (a), the FPU first selects two representable fractions b and c that most
closely bracket a in value (b < a < c).
(b) 1.0001 0000 1000 0011 1001 011E2 101
(c) 1.0001 0000 1000 0011 1001 100E2 101

The FPU then sets the result to b or to c according to the rounding mode selected in the RC field.
Rounding introduces an error in a result that is less than one unit in the last place to which the
result is rounded.
The rounded result is called the inexact result. When the FPU produces an inexact result, the
floating-point precision (inexact) flag (PE) is set in the FPU status word.
When the overflow exception is masked and the infinitely precise result is between the largest
positive finite value allowed in a particular format and +∞, the FPU rounds the result as shown
in Table 7-6.

Table 7-6. Rounding of Positive Numbers with Masked Overflow

Rounding Mode Result
Rounding to nearest (even) +∞
Rounding toward zero (Truncate) Maximum, positive finite value
Rounding up (toward +∞) +∞
Rounding down) (toward −∞) Maximum, positive finite value

When the overflow exception is masked and the infinitely precise result is between the largest
negative finite value allowed in a particular format and −∞, the FPU rounds the result as shown
in Table 7-7.

Table 7-7. Rounding of Negative Numbers with Masked Overflow

Rounding Mode Result
Rounding to nearest (even) −∞
Rounding toward zero (Truncate) Maximum, negative finite value
Rounding up (toward +∞) Maximum, negative finite value
Rounding down) (toward −∞) −∞

The rounding modes have no effect on comparison operations, operations that produce exact
results, or operations that produce NaN results.

7-19
FLOATING-POINT UNIT

7.3.5. Infinity Control Flag

The infinity control flag (bit 12 of the FPU control word) is provided for compatibility with the
Intel 287 Math Coprocessor; it is not meaningful for the Pentium® Pro processor FPU or for the
Pentium® processor FPU, the Intel486™ processor FPU, or Intel 387 processor NPX. Refer to
Section 7.2.3.3., “Signed Infinities”, for information on how the IA FPUs handle infinity values.

7.3.6. FPU Tag Word

The 16-bit tag word (refer to Figure 7-11) indicates the contents of each the 8 registers in the
FPU data-register stack (one 2-bit tag per register). The tag codes indicate whether a register
contains a valid number, zero, or a special floating-point number (NaN, infinity, denormal, or
unsupported format), or whether it is empty. The FPU tag word is cached in the FPU in the FPU
tag word register. When the FPU is initialized with either an FINIT/FNINIT or
FSAVE/FNSAVE instruction, the FPU tag word is set to FFFFH, which marks all the FPU data
registers as empty.
.

15 0

TAG(7) TAG(6) TAG(5) TAG(4) TAG(3) TAG(2) TAG(1) TAG(0)

TAG Values
00 — Valid
01 — Zero
10 — Special: invalid (NaN, unsupported), infinity, or denormal
11 — Empty

Figure 7-11. FPU Tag Word

Each tag in the FPU tag word corresponds to a physical register (numbers 0 through 7). The
current top-of-stack (TOP) pointer stored in the FPU status word can be used to associate tags
with registers relative to ST(0).
The FPU uses the tag values to detect stack overflow and underflow conditions. Stack overflow
occurs when the TOP pointer is decremented (due to a register load or push operation) to point
to a non-empty register. Stack underflow occurs when the TOP pointer is incremented (due to a
save or pop operation) to point to an empty register or when an empty register is also referenced
as a source operand. A non-empty register is defined as a register containing a zero (01), a valid
value (00), or an special (10) value.
Application programs and exception handlers can use this tag information to check the contents
of an FPU data register without performing complex decoding of the actual data in the register.
To read the tag register, it must be stored in memory using either the FSTENV/FNSTENV or
FSAVE/FNSAVE instructions. The location of the tag word in memory after being saved with
one of these instructions is shown in Figures 7-13 through 7-16.

7-20
 FLOATING-POINT UNIT

Software cannot directly load or modify the tags in the tag register. The FLDENV and FRSTOR
instructions load an image of the tag register into the FPU; however, the FPU uses those tag
values only to determine if the data registers are empty (11B) or non-empty (00B, 01B, or 10B).
If the tag register image indicates that a data register is empty, the tag in the tag register for that
data register is marked empty (11B); if the tag register image indicates that the data register is
non-empty, the FPU reads the actual value in the data register and sets the tag for the register
accordingly. This action prevents a program from setting the values in the tag register to incor-
rectly represent the actual contents of non-empty data registers.

7.3.7. FPU Instruction and Operand (Data) Pointers

The FPU stores pointers to the instruction and operand (data) for the last non-control instruction
executed in two 48-bit registers: the FPU instruction pointer and FPU operand (data) pointer
registers (refer to Figure 7-5). (This information is saved to provide state information for excep-
tion handlers.)
The contents of the FPU instruction and operand pointer registers remain unchanged when any
of the control instructions (FINIT/FNINIT, FCLEX/FNCLEX, FLDCW, FSTCW/FNSTCW,
FSTSW/FNSTSW, FSTENV/FNSTENV, FLDENV, FSAVE/FNSAVE, FRSTOR, and
WAIT/FWAIT) are executed. The contents of the FPU operand register are undefined if the prior
non-control instruction did not have a memory operand.
The pointers stored in the FPU instruction and operand pointer registers consist of an offset
(stored in bits 0 through 31) and a segment selector (stored in bits 32 through 47).
These registers can be accessed by the FSTENV/FNSTENV, FLDENV, FINIT/FNINIT,
FSAVE/FNSAVE and FRSTOR instructions. The FINIT/FNINIT and FSAVE/FNSAVE instruc-
tions clear these registers.
For all the IA FPUs and NPXs except the 8087, the FPU instruction pointer points to any
prefixes that preceded the instruction. For the 8087, the FPU instruction pointer points only to
the actual opcode.

7.3.8. Last Instruction Opcode

The FPU stores the opcode of the last non-control instruction executed in an 11-bit FPU opcode
register. (This information provides state information for exception handlers.) Only the first and
second opcode bytes (after all prefixes) are stored in the FPU opcode register. Figure 7-12 shows
the encoding of these two bytes. Since the upper 5 bits of the first opcode byte are the same for
all floating-point opcodes (11011B), only the lower 3 bits of this byte are stored in the opcode
register.

7.3.9. Saving the FPU’s State

The FSTENV/FNSTENV and FSAVE/FNSAVE instructions store FPU state information in
memory for use by exception handlers and other system and application software. The

7-21
FLOATING-POINT UNIT

FSTENV/FNSTENV instruction saves the contents of the status, control, tag, FPU instruction
pointer, FPU operand pointer, and opcode registers. The FSAVE/FNSAVE instruction stores that
information plus the contents of the FPU data registers. Note that the FSAVE/FNSAVE instruc-
tion also initializes the FPU to default values (just as the FINIT/FNINIT instruction does) after
it has saved the original state of the FPU.

1st Instruction Byte 2nd Instruction Byte

7 2 0 7 0

10 8 7 0

FPU Opcode Register

Figure 7-12. Contents of FPU Opcode Registers

The manner in which this information is stored in memory depends on the operating mode of
the processor (protected mode or real-address mode) and on the operand-size attribute in effect
(32-bit or 16-bit). Refer to Figures 7-13 through 7-16. In virtual-8086 mode or SMM, the real-
address mode formats shown in Figure 7-16 is used. Refer to Chapter 12, System Management
Mode (SMM) of the Intel Architecture Software Developer’s Manual, Volume 3, for special
considerations for using the FPU while in SMM.

32-Bit Protected Mode Format

31 16 15 0
Control Word 0
Status Word 4
Tag Word 8
FPU Instruction Pointer Offset 12
0000 Opcode 10...00 FPU Instruction Pointer Selector 16
FPU Operand Pointer Offset 20
FPU Operand Pointer Selector 24

Reserved

Figure 7-13. Protected Mode FPU State Image in Memory, 32-Bit Format

7-22
 FLOATING-POINT UNIT

32-Bit Real-Address Mode Format

31 16 15 0
Control Word 0
Status Word 4
Tag Word 8
FPU Instruction Pointer 15...00 12
0000 FPU Instruction Pointer 31...16 0 Opcode 10...00 16
Reserved FPU Operand Pointer 15...00 20
0000 FPU Operand Pointer 31...16 000000000000 24

Reserved

Figure 7-14. Real Mode FPU State Image in Memory, 32-Bit Format

16-Bit Protected Mode Format

15 0
Control Word 0
Status Word 2
Tag Word 4
FPU Instruction Pointer Offset 6
FPU Instruction Pointer Selector 8
FPU Operand Pointer Offset 10
FPU Operand Pointer Selector 12

Figure 7-15. Protected Mode FPU State Image in Memory, 16-Bit Format

7-23
FLOATING-POINT UNIT

16-Bit Real-Address Mode and

Virtual-8086 Mode Format
15 0
Control Word 0
Status Word 2
Tag Word 4
FPU Instruction Pointer 15...00 6
IP 19..16 0 Opcode 10...00 8
FPU Operand Pointer 15...00 10
OP 19..16 0 0 0 0 0 0 0 0 0 0 0 0 12

Figure 7-16. Real Mode FPU State Image in Memory, 16-Bit Format

The FLDENV and FRSTOR instructions allow FPU state information to be loaded from
memory into the FPU. Here, the FLDENV instruction loads only the status, control, tag, FPU
instruction pointer, FPU operand pointer, and opcode registers, and the FRSTOR instruction
loads all the FPU registers, including the FPU stack registers.

7.4. FLOATING-POINT DATA TYPES AND FORMATS

The IA FPU recognizes and operates on seven data types, divided into three groups: reals, inte-
gers, and packed BCD integers. Figure 7-17 shows the data formats for each of the FPU data
types. Table 7-8 gives the length, precision, and approximate normalized range that can be repre-
sented of each FPU data type. Denormal values are also supported in each of the real types, as
required by IEEE Standard 854.
With the exception of the 80-bit extended-real format, all of these data types exist in memory
only. When they are loaded into FPU data registers, they are converted into extended-real format
and operated on in that format.

7-24
 FLOATING-POINT UNIT

Single Real
Sign Exp. Fraction
3130 23 22 Implied Integer 0

Double Real
Sign Exponent Fraction
63 62 52 51 Implied Integer 0
Sign
Extended Real
Exponent Fraction
79 78 6463 62 Integer 0
Word Integer
Sign
15 14 0
Short Integer
Sign
31 30 0
Long Integer
Sign
Sign 63 62 0
Packed BCD Integers
X D17 D16 D15 D14 D13 D12 D11 D10 D9 D8 D7 D6 D5 D4 D3 D2 D1 D0
79 78 72 71 4 Bits = 1 BCD Digit 0

Figure 7-17. Floating-Point Unit Data Type Formats

When stored in memory, the least significant byte of an FPU data-type value is stored at the
initial address specified for the value. Successive bytes from the value are then stored in succes-
sively higher addresses in memory. The floating-point instructions load and store memory oper-
ands using only the initial address of the operand.

7.4.1. Real Numbers

The FPU’s three real data types (single-real, double-real, and extended-real) correspond directly
to the single-precision, double-precision, and double-extended-precision formats in the IEEE
standard. The extended-precision format is the format used by the data registers in the FPU.
Table 7-8 gives the precision and range of these data types and Figure 7-17 gives the formats.
For the single-real and double-real formats, only the fraction part of the significand is encoded.
The integer is assumed to be 1 for all numbers except 0 and denormalized finite numbers. For
the extended-real format, the integer is contained in bit 63, and the most-significant fraction bit

7-25
FLOATING-POINT UNIT

is bit 62. Here, the integer is explicitly set to 1 for normalized numbers, infinities, and NaNs,
and to 0 for zero and denormalized numbers.

Table 7-8. Length, Precision, and Range of FPU Data Types

Data Type Length Precision Approximate Normalized Range
(Bits)
Binary Decimal
Binary Real
Single real 32 24 2–126 to 2127 1.18 × 10–38 to 3.40 × 1038
Double real 64 53 2–1022 to 21023 2.23 × 10–308 to 1.79 × 10308
Extended real 80 64 2–16382 to 216383 3.37 × 10–4932 to 1.18 × 104932
Binary Integer
Word integer 16 15 –215 to 215 – 1 –32,768 to 32,767
Short integer 32 31 –231
to 2 31
–1 –2.14 × 109 to 2.14 × 109
Long integer 64 63 –263 to 263 – 1 –9.22 × 1018 to 9.22 × 1018
Packed BCD 80 18 (decimal Not Pertinent (–1018 + 1) to (1018 – 1)
Integers digits)

The exponent of each real data type is encoded in biased format. The biasing constant is 127 for
the single-real format, 1023 for the double-real format, and 16,383 for the extended-real format.
Table 7-9 shows the encodings for all the classes of real numbers (that is, zero, denormalized-
finite, normalized-finite, and ∞) and NaNs for each of the three real data-types. It also gives the
format for the real indefinite value.
When storing real values in memory, single-real values are stored in 4 consecutive bytes in
memory; double-real values are stored in 8 consecutive bytes; and extended-real values are
stored in 10 consecutive bytes.
As a general rule, values should be stored in memory in double-real format. This format
provides sufficient range and precision to return correct results with a minimum of programmer
attention. The single-real format is appropriate for applications that are constrained by memory;
however, it provides less precision and a greater chance of overflow. The single-real format is
also useful for debugging algorithms, because rounding problems will manifest themselves
more quickly in this format. The extended-real format is normally reserved for holding interme-
diate results in the FPU registers and constants. Its extra length is designed to shield final results
from the effects of rounding and overflow/underflow in intermediate calculations. However,
when an application requires the maximum range and precision of the FPU (for data storage,
computations, and results), values can be stored in memory in extended-real format.
The real indefinite value is a QNaN encoding that is stored by several floating-point instructions
in response to a masked floating-point invalid operation exception (refer to Table 7-21).

7-26
 FLOATING-POINT UNIT

Table 7-9. Real Number and NaN Encodings

Class Sign Biased Exponent Significand
1
Integer Fraction
Positive +∞ 0 11..11 1 00..00
+Normals 0 11..10 1 11..11
. . . .
. . . .
0 00..01 1 00..00
+Denormals 0 00..00 0 11.11
. . . .
. . . .
0 00..00 0 00..01
+Zero 0 00..00 0 00..00
Negative −Zero 1 00..00 0 00..00
−Denormals 1 00..00 0 00..01
. . . .
. . . .
1 00..00 0 11..11
−Normals 1 00..01 1 00..00
. . . .
. . . .
1 11..10 1 11..11
−∞ 1 11..11 1 00..00
NaNs SNaN X 11..11 1 0X..XX2
QNaN X 11..11 1 1X..XX
Real Indefinite 1 11..11 1 10..00
(QNaN)
Single-Real: ← 8 Bits → ← 23 Bits →
Double-Real: ← 11 Bits → ← 52 Bits →
Extended-Real ← 15 Bits → ← 63 Bits →
NOTES:
1. Integer bit is implied and not stored for single-real and double-real formats.
2. The fraction for SNaN encodings must be non-zero.

7.4.2. Binary Integers

The FPU’s three binary integer data types (word, short, and long) have identical formats, except
for length. Table 7-8 gives the precision and range of these data types and Figure 7-17 gives the
formats. Table 7-10 gives the encodings of the three binary integer types.

7-27
FLOATING-POINT UNIT

Table 7-10. Binary Integer Encodings

Class Sign Magnitude
Positive Largest 0 11..11
. .
. .
. .
Smallest 0 00..01
Zero 0 00..00
Negative Smallest 1 11..11
. .
. .
. .
. .
Largest 1 00..00
Integer Indefinite 1 00..00
Word Integer: ← 15 bits →
Short Integer: ← 31 Bits →
Long Integer: ← 63 Bits →

The most significant bit of each format is the sign bit (0 for positive and 1 for negative). Nega-
tive values are represented in standard two’s complement notation. The quantity zero is repre-
sented with all bits (including the sign bit) set to zero. Note that the FPU’s word-integer data
type is identical to the word-integer data type used by the processor’s integer unit and the short-
integer format is identical to the integer unit’s doubleword-integer data type.
Word-integer values are stored in 2 consecutive bytes in memory; short-integer values are stored
in 4 consecutive bytes; and long-integer values are stored in 8 consecutive bytes. When loaded
into the FPU’s data registers, all the binary integers are exactly representable in the extended-
real format.
The binary integer encoding 100..00B represents either of two things, depending on the circum-
stances of its use:
• The largest negative number supported by the format (–215, –231, or –263).
• The integer indefinite value.
If this encoding is used as a source operand (as in an integer load or integer arithmetic instruc-
tion), the FPU interprets it as the largest negative number representable in the format being used.
If the FPU detects an invalid operation when storing an integer value in memory with an
FIST/FISTP instruction and the invalid operation exception is masked, the FPU stores the
integer indefinite encoding in the destination operand as a masked response to the exception. In
situations where the origin of a value with this encoding may be ambiguous, the invalid opera-
tion exception flag can be examined to see if the value was produced as a response to an
exception.

7-28
 FLOATING-POINT UNIT

If the integer indefinite is stored in memory and is later loaded back into an FPU data register,
it is interpreted as the largest negative number supported by the format.

7.4.3. Decimal Integers

Decimal integers are stored in a 10-byte, packed BCD format. Table 7-8 gives the precision and
range of this data type and Figure 7-17 shows the format. In this format, the first 9 bytes hold
18 BCD digits, 2 digits per byte (refer to Section 5.2.3., “BCD Integers” in Chapter 5, Data
Types and Addressing Modes). The least-significant digit is contained in the lower half-byte of
byte 0 and the most-significant digit is contained in the upper half-byte of byte 9. The most
significant bit of byte 10 contains the sign bit (0 = positive and 1 = negative). (Bits 0 through 6
of byte 10 are don’t care bits.) Negative decimal integers are not stored in two's complement
form; they are distinguished from positive decimal integers only by the sign bit.
Table 7-11 gives the possible encodings of value in the decimal integer data type.

Table 7-11. Packed Decimal Integer Encodings

Magnitude
Class Sign digit digit digit digit ... digit
Positive
Largest 0 0000000 1001 1001 1001 1001 ... 1001
. . .
. . .
0 0000000 0000 0000 0000 0000 ... 0001
Smallest
Zero 0 0000000 0000 0000 0000 0000 ... 0000
Negative
Zero 1 0000000 0000 0000 0000 0000 ... 0000
1 0000000 0000 0000 0000 0000 ... 0001
Smallest
. . .
. . .
Largest 1 0000000 1001 1001 1001 1001 ... 1001
Decimal 1 1111111 1111 1111 UUUU* UUUU ... UUUU
Integer
Indefinite
← 1 byte → ← 9 bytes →
NOTE:
* UUUU means bit values are undefined and may contain any value.

The decimal integer format exists in memory only. When a decimal integer is loaded in a data
register in the FPU, it is automatically converted to the extended-real format. All decimal inte-
gers are exactly representable in extended-real format.

7-29
FLOATING-POINT UNIT

The packed decimal indefinite encoding is stored by the FBSTP instruction in response to a
masked floating-point invalid operation exception. Attempting to load this value with the FBLD
instruction produces an undefined result.

7.4.4. Unsupported Extended-Real Encodings

The extended-real format permits many encodings that do not fall into any of the categories
shown in Table 7-9. Table 7-12 shows these unsupported encodings. Some of these encodings
were supported by the Intel 287 math coprocessor; however, most of them are not supported by
the Intel 387 math coprocessor, or the internal FPUs in the Intel486™, Pentium®, or Pentium®
Pro processors. These encodings are no longer supported due to changes made in the final
version of IEEE Standard 754 that eliminated these encodings.
The categories of encodings formerly known as pseudo-NaNs, pseudo-infinities, and un-normal
numbers are not supported. The Intel 387 math coprocessor and the internal FPUs in the
Intel486™, Pentium®, and Pentium® Pro processors generate the invalid operation exception
when they are encountered as operands.
The encodings formerly known as pseudo-denormal numbers are not generated by the Intel 387
math coprocessor and the internal FPUs in the Intel486™, Pentium®, and Pentium® Pro proces-
sors; however, they are used correctly when encountered as operands. The exponent is treated
as if it were 00..01B and the mantissa is unchanged. The denormal exception is generated.

7-30
 FLOATING-POINT UNIT

Table 7-12. Unsupported Extended-Real Encodings

Class Sign Biased Exponent Significand
Integer Fraction
Positive 0 11..11 0 11..11
Pseudo-NaNs Quiet . . .
0 11..11 10..00
0 11..11 0 01..11
Signaling . . .
0 11..11 00..01
Positive Reals Pseudo-infinity 0 11..11 0 00..00
0 11..10 0 11..11
Unnormals . . .
0 00..01 00..00
Pseudo-denormals 0 00..00 1 11..11
. . .
0 00..00 00..00
Negative Reals Pseudo-denormals 1 00..00 1 11..11
. . .
1 00..00 00..00
1 11..10 0 11..01
Unnormals . . .
1 00..01 00..00
Pseudo-infinity 1 11..11 0 00..00
Negative 1 11..11 0 01..11
Pseudo-NaNs Signaling . . .
1 11..11 00..01
1 11..11 0 11..11
Quiet . . .
1 11..11 10..00
← 15 bits → ← 63 bits →

7.5. FPU INSTRUCTION SET

The floating-point instructions that the IA FPU supports can be grouped into six functional cate-
gories:

• Data transfer instructions

• Basic arithmetic instructions
• Comparison instructions
• Transcendental instructions
• Load constant instructions
• FPU control instructions

7-31
FLOATING-POINT UNIT

Refer to Section 6.2.3., “Floating-Point Instructions” in Chapter 6, Instruction Set Summary, for
a list of the floating-point instructions by category.
The following section briefly describes the instructions in each category. Detailed descriptions
of the floating-point instructions are given in Chapter 3, Instruction Set Reference, in the Intel
Architecture Software Developer’s Manual, Volume 2.

7.5.1. Escape (ESC) Instructions

All of the instructions in the FPU instruction set fall into a class of instructions known as escape
(ESC) instructions. All of these instructions have a common opcode format, which is slightly
different from the format used by the integer and operating-system instructions.

7.5.2. FPU Instruction Operands

Most floating-point instructions require one or two operands, located on the FPU data-register
stack or in memory. (None of the floating-point instructions accept immediate operands.)
When an operand is located in a data register, it is referenced relative to the ST(0) register (the
register at the top of the register stack), rather than by a physical register number. Often the
ST(0) register is an implied operand.
Operands in memory can be referenced using the same operand addressing methods available
for the integer and system instructions.

7.5.3. Data Transfer Instructions

The data transfer instructions (refer to Table 7-13) perform the following operations:
• Load real, integer, or packed BCD operands from memory into the ST(0) register.
• Store the value in the ST(0) register in memory in real, integer, or packed BCD format.
• Move values between registers in the FPU register stack.

Table 7-13. Data Transfer Instructions

Real Integer Packed Decimal
FLD Load Real FILD Load Integer FBLD Load Packed
Decimal
FST Store Real FIST Store Integer
FSTP Store Real and FISTP Store Integer FBSTP Store Packed
Pop and Pop Decimal and Pop
FXCH Exchange Register
Contents
FCMOVcc Conditional Move

7-32
 FLOATING-POINT UNIT

Operands are normally stored in the FPU data registers in extended-real format (refer to Section
7.3.4.2., “Precision Control Field”). The FLD (load real) instruction pushes a real operand from
memory onto the top of the FPU data-register stack. If the operand is in single- or double-real
format, it is automatically converted to extended-real format. This instruction can also be used
to push the value in a selected FPU data register onto the top of the register stack.
The FILD (load integer) instruction converts an integer operand in memory into extended-real
format and pushes the value onto the top of the register stack. The FBLD (load packed decimal)
instruction performs the same load operation for a packed BCD operand in memory.
The FST (store real) and FIST (store integer) instructions store the value in register ST(0) in
memory in the destination format (real or integer, respectively). Again, the format conversion is
carried out automatically.
The FSTP (store real and pop), FISTP (store integer and pop), and FBSTP (store packed decimal
and pop) instructions store the value in the ST(0) registers into memory in the destination format
(real, integer, or packed BCD), then performs a pop operation on the register stack. A pop oper-
ation causes the ST(0) register to be marked empty and the stack pointer (TOP) in the FPU
control work to be incremented by 1. The FSTP instruction can also be used to copy the value
in the ST(0) register to another FPU register [ST(i)].
The FXCH (exchange register contents) instruction exchanges the value in a selected register in
the stack [ST(i)] with the value in ST(0).
The FCMOVcc (conditional move) instructions move the value in a selected register in the stack
[ST(i)] to register ST(0). These instructions move the value only if the conditions specified with
a condition code (cc) are satisfied (refer to Table 7-14). The conditions being tested with the
FCMOVcc instructions are represented by the status flags in the EFLAGS register. The condi-
tion code mnemonics are appended to the letters “FCMOV” to form the mnemonic for a
FCMOVcc instruction.

Table 7-14. Floating-Point Conditional Move Instructions

Instruction Mnemonic Status Flag States Condition Description
FCMOVB CF=1 Below
FCMOVNB CF=0 Not below
FCMOVE ZF=1 Equal
FCMOVNE ZF=0 Not equal
FCMOVBE (CF or ZF)=1 Below or equal
FCMOVNBE (CF or ZF)=0 Not below nor equal
FCMOVU PF=1 Unordered
FCMOVNU PF=0 Not unordered

7-33
FLOATING-POINT UNIT

Like the CMOVcc instructions, the FCMOVcc instructions are useful for optimizing small IF
constructions. They also help eliminate branching overhead for IF operations and the possibility
of branch mispredictions by the processor.

NOTE
The FCMOVcc instructions may not be supported on some processors in the
Pentium® Pro processor family. Software can check if the FCMOVcc instruc-
tions are supported by checking the processor’s feature information with the
CPUID instruction (refer to “CPUID—CPU Identification” in Chapter 3,
Instruction Set Reference, of the Intel Architecture Software Developer’s
Manual, Volume 2).

7.5.4. Load Constant Instructions

The following instructions push commonly used constants onto the top [ST(0)] of the FPU
register stack:
FLDZ Load +0.0
FLD1 Load +1.0
FLDPI Load π
FLDL2T Load log2 10
FLDL2E Load log2e
FLDLG2 Load log102
FLDLN2 Load loge2
The constant values have full extended-real precision (64 bits) and are accurate to approximately
19 decimal digits. They are stored internally in a format more precise than extended real. When
loading the constant, the FPU rounds the more precise internal constant according to the RC
(rounding control) field of the FPU control word. Refer to Section 7.5.8., “Pi”, for information
on the π constant.

7-34
 FLOATING-POINT UNIT

7.5.5. Basic Arithmetic Instructions

The following floating-point instructions perform basic arithmetic operations on real numbers.
Where applicable, these instructions match IEEE Standard 754:
FADD/FADDP Add real
FIADD Add integer to real
FSUB/FSUBP Subtract real
FISUB Subtract integer from real
FSUBR/FSUBRP Reverse subtract real
FISUBR Reverse subtract real from integer
FMUL/FMULP Multiply real
FIMUL Multiply integer by real
FDIV/FDIVP Divide real
FIDIV Divide real by integer
FDIVR/FDIVRP Reverse divide
FIDIVR Reverse divide integer by real
FABS Absolute value
FCHS Change sign
FSQRT Square root
FPREM Partial remainder
FPREM1 IEEE partial remainder
FRNDINT Round to integral value
FXTRACT Extract exponent and significand
The add, subtract, multiply and divide instructions operate on the following types of operands:
• Two FPU register values.
• A register value and a real or integer value in memory.
Operands in memory can be in single-real, double-real, short-integer, or word-integer format.
They are converted to extended-real format automatically.
Reverse versions of the subtract and divide instructions are provided to foster efficient coding.
For example, the FSUB instruction subtracts the value in a specified FPU register [ST(i)] from
the value in register ST(0); whereas, the FSUBR instruction subtracts the value in ST(0) from
the value in ST(i). The results of both operations are stored in register ST(0). These instructions
eliminate the need to exchange values between register ST(0) and another FPU register to
perform a subtraction or division.
The pop versions of the add, subtract, multiply and divide instructions pop the FPU register
stack following the arithmetic operation.
The FPREM instruction computes the remainder from the division of two operands in the
manner used by the Intel 8087 and Intel 287 math coprocessors; the FPREM1 instructions
computes the remainder is the manner specified in the IEEE specification.
The FSQRT instruction computes the square root of the source operand.
The FRNDINT instructions rounds a real value to its nearest integer value, according to the
current rounding mode specified in the RC field of the FPU control word. This instruction

7-35
FLOATING-POINT UNIT

performs a function similar to the FIST/FISTP instructions, except that the result is saved in a
real format.
The FABS, FCHS, and FXTRACT instructions perform convenient arithmetic operations. The
FABS instruction produces the absolute value of the source operand. The FCHS instruction
changes the sign of the source operand. The FXTRACT instruction separates the source operand
into its exponent and fraction and stores each value in a register in real format.

7.5.6. Comparison and Classification Instructions

The following instructions compare or classify real values:
FCOM/FCOMP/FCOMPP Compare real and set FPU condition code flags.
FUCOM/FUCOMP/FUCOMPP Unordered compare real and set FPU condition code flags.
FICOM/FICOMP Compare integer and set FPU condition code flags.
FCOMI/FCOMIP Compare real and set EFLAGS status flags.
FUCOMI/FUCOMIP Unordered compare real and set EFLAGS status flags.
FTST Test (compare real with 0.0).
FXAM Examine.
Comparison of real values differ from comparison of integers because real values have four
(rather than three) mutually exclusive relationships: less than, equal, greater than, and
unordered.
The unordered relationship is true when at least one of the two values being compared is a NaN
or in an undefined format. This additional relationship is required because, by definition, NaNs
are not numbers, so they cannot have less than, equal, or greater than relationships with other
real values.
The FCOM, FCOMP, and FCOMPP instructions compare the value in register ST(0) with a real
source operand and set the condition code flags (C0, C2, and C3) in the FPU status word
according to the results (refer to Table 7-15). If an unordered condition is detected (one or both
of the values is a NaN or in an undefined format), a floating-point invalid operation exception
is generated.
The pop versions of the instruction pop the FPU register stack once or twice after the comparison
operation is complete.
The FUCOM, FUCOMP, and FUCOMPP instructions operate the same as the FCOM, FCOMP,
and FCOMPP instructions. The only difference is that with the FUCOM, FUCOMP, and
FUCOMPP instructions, if an unordered condition is detected because one or both of the oper-
ands is a QNaN, the floating-point invalid operation exception is not generated.

7-36
 FLOATING-POINT UNIT

Table 7-15. Setting of FPU Condition Code Flags for Real Number Comparisons
Condition C3 C2 C0
ST(0) > Source Operand 0 0 0
ST(0) < Source Operand 0 0 1
ST(0) = Source Operand 1 0 0
Unordered 1 1 1

The FICOM and FICOMP instructions also operate the same as the FCOM and FCOMP instruc-
tions, except that the source operand is an integer value in memory. The integer value is auto-
matically converted into an extended real value prior to making the comparison. The FICOMP
instruction pops the FPU register stack following the comparison operation.
The FTST instruction performs the same operation as the FCOM instruction, except that the
value in register ST(0) is always compared with the value 0.0.
The FCOMI and FCOMIP instructions are new in the Intel Pentium® Pro processor. They
perform the same comparison as the FCOM and FCOMP instructions, except that they set the
status flags (ZF, PF, and CF) in the EFLAGS register to indicate the results of the comparison
(refer to Table 7-16) instead of the FPU condition code flags. The FCOMI and FCOMIP instruc-
tions allow condition branch instructions (Jcc) to be executed directly from the results of their
comparison.

Table 7-16. Setting of EFLAGS Status Flags for Real Number Comparisons
Comparison Results ZF PF CF
ST0 > ST(i) 0 0 0
ST0 < ST(i) 0 0 1
ST0 = ST(i) 1 0 0
Unordered 1 1 1

The FUCOMI and FUCOMIP instructions operate the same as the FCOMI and FCOMIP
instructions, except that they do not generate a floating-point invalid operation exception if the
unordered condition is the result of one or both of the operands being a QNaN. The FCOMIP
and FUCOMIP instructions pop the FPU register stack following the comparison operation.
The FXAM instruction determines the classification of the real value in the ST(0) register (that
is, whether the value is zero, a denormal number, a normal finite number, ∞, a NaN, or an unsup-
ported format) or that the register is empty. It sets the FPU condition code flags to indicate the
classification (refer to “FXAM—Examine” in Chapter 3, Instruction Set Reference, of the Intel
Architecture Software Developer’s Manual, Volume 2). It also sets the C1 flag to indicate the sign
of the value.

7-37
FLOATING-POINT UNIT

7.5.6.1. BRANCHING ON THE FPU CONDITION CODES

The processor does not offer any control-flow instructions that branch on the setting of the
condition code flags (C0, C2, and C3) in the FPU status word. To branch on the state of these
flags, the FPU status word must first be moved to the AX register in the integer unit. The
FSTSW AX (store status word) instruction can be used for this purpose. When these flags are
in the AX register, the TEST instruction can be used to control conditional branching as follows:
1. Check for an unordered result. Use the TEST instruction to compare the contents of the
AX register with the constant 0400H (refer to Table 7-17). This operation will clear the ZF
flag in the EFLAGS register if the condition code flags indicate an unordered result;
otherwise, the ZF flag will be set. The JNZ instruction can then be used to transfer control
(if necessary) to a procedure for handling unordered operands.

Table 7-17. TEST Instruction Constants for Conditional Branching

Order Constant Branch
ST(0) > Source Operand 4500H JZ
ST(0) < Source Operand 0100H JNZ
ST(0) = Source Operand 4000H JNZ
Unordered 0400H JNZ

2. Check ordered comparison result. Use the constants given in Table 7-17 in the TEST
instruction to test for a less than, equal to, or greater than result, then use the corresponding
conditional branch instruction to transfer program control to the appropriate procedure or
section of code.

If a program or procedure has been thoroughly tested and it incorporates periodic checks for
QNaN results, then it is not necessary to check for the unordered result every time a comparison
is made.
Refer to Section 7.3.3., “Branching and Conditional Moves on FPU Condition Codes”, for
another technique for branching on FPU condition codes.
Some non-comparison FPU instructions update the condition code flags in the FPU status word.
To ensure that the status word is not altered inadvertently, store it immediately following a
comparison operation.

7.5.7. Trigonometric Instructions

The following instructions perform four common trigonometric functions:
FSIN Sine
FCOS Cosine
FSINCOS Sine and cosine
FPTAN Tangent
FPATAN Arctangent

7-38
 FLOATING-POINT UNIT

These instructions operate on the top one or two registers of the FPU register stack and they
return their results to the stack. The source operands must be given in radians.
The FSINCOS instruction returns both the sine and the cosine of a source operand value. It oper-
ates faster than executing the FSIN and FCOS instructions in succession.
The FPATAN instruction computes the arctangent of ST(1) divided by ST(0). It is useful for
converting rectangular coordinates to polar coordinates.

7.5.8. Pi
When the argument (source operand) of a trigonometric function is within the range of the func-
tion, the argument is automatically reduced by the appropriate multiple of 2π through the same
reduction mechanism used by the FPREM and FPREM1 instructions. The internal value of π
that the IA FPU uses for argument reduction and other computations is as follows:
π = 0.f ∗ 22
where:
f = C90FDAA2 2168C234 C
(The spaces in the fraction above indicate 32-bit boundaries.)
This internal π value has a 66-bit mantissa, which is 2 bits more than is allowed in the signifi-
cand of an extended-real value. (Since 66 bits is not an even number of hexadecimal digits, two
additional zeros have been added to the value so that it can be represented in hexadecimal
format. The least-significant hexadecimal digit (C) is thus 1100B, where the two least-
significant bits represent bits 67 and 68 of the mantissa.)
This value of π has been chosen to guarantee no loss of significance in a source operand,
provided the operand is within the specified range for the instruction.
If the results of computations that explicitly use π are to be used in the FSIN, FCOS, FSINCOS,
or FPTAN instructions, the full 66-bit fraction of π should be used. This insures that the results
are consistent with the argument-reduction algorithms that these instructions use. Using a
rounded version of π can cause inaccuracies in result values, which if propagated through
several calculations, might result in meaningless results.
A common method of representing the full 66-bit fraction of π is to separate the value into two
numbers (highπ and lowπ) that when added together give the value for π shown earlier in this
section with the full 66-bit fraction:
π = highπ + lowπ
For example, the following two values (given in scientific notation with the fraction in hexadec-
imal and the exponent in decimal) represent the 33 most-significant and the 33 least-significant
bits of the fraction:
highπ (unnormalized)= 0.C90FDAA20 * 2+2
lowπ (unnormalized) = 0.42D184698 * 2−31

7-39
FLOATING-POINT UNIT

These values encoded in standard IEEE double-real format are as follows:

highπ = 400921FB 54400000
lowπ = 3DE0B461 1A600000
(Note that in the IEEE double-real format, the exponents are biased (by 1023) and the fractions
are normalized.)
Similar versions of π can also be written in extended-real format.
When using this two-part π value in an algorithm, parallel computations should be performed
on each part, with the results kept separate. When all the computations are complete, the two
results can be added together to form the final result.
The complications of maintaining a consistent value of π for argument reduction can be avoided,
either by applying the trigonometric functions only to arguments within the range of the
automatic reduction mechanism, or by performing all argument reductions (down to a magni-
tude less than π/4) explicitly in software.

7.5.9. Logarithmic, Exponential, and Scale

The following instructions provide two different logarithmic functions, an exponential function,
and a scale function.
FYL2X Compute log: (y ∗ log2x)
FYL2XP1 Compute log epsilon: (y ∗ log2(x + 1))
F2XM1 Compute exponential: (2x – 1)
FSCALE Scale
The FYL2X and FYL2XP1 instructions perform two different base 2 logarithmic operations.
The FYL2X instruction computes the log of (y ∗ log2(x)). This operation permits the calculation
of the log of any base using the following equation:
logb x = (1/log2 b) ∗ log2 x
The FYL2XP1 instruction computes the log epsilon of (y ∗ log2(x + 1)). This operation provides
optimum accuracy for values of x that may be very close to 0.
The F2XM1 instruction computes the exponential (2x − 1). This instruction only operates on
source values in the range −1.0 to +1.0.
The FSCALE instruction multiplies the source operand by a power of 2.

7.5.10. Transcendental Instruction Accuracy

The algorithms that the Pentium® and Pentium® Pro processors use for the transcendental
instructions (FSIN, FCOS, FSINCOS, FPTAN, FPATAN, F2XM1, FYL2X, and FYL2XP1)
allow a higher level of accuracy than was possible in earlier IA math coprocessors and FPUs.
The accuracy of these instructions is measured in terms of units in the last place (ulp). For a

7-40
 FLOATING-POINT UNIT

given argument x, let f(x) and F(x) be the correct and computed (approximate) function values,
respectively. The error in ulps is defined to be:
f(x) – F(x )
error = --------------------------
-
k – 63
2

–k
where k is an integer such that 1 ≤ 2 f( x) < 2 .
With the Pentium® and Pentium® Pro processors, the worst case error in the transcendental
instructions is less than 1 ulp when rounding to nearest and less than 1.5 ulps when rounding
in other modes. (The instructions fyl2x and fyl2xp1 are two operand instructions and are guar-
anteed to be within 1 ulp only when y = 1.
When y != 1, the maximum ulp error is always within 1.35 ulps in round to nearest mode. The
trigonometric
instructions may use a 66-bit approximation to the true value of pi to reduce the magnitude of
the input argument.
In this case, the final computed result can vary considerably from the true mathematically
precise result.) The instructions are guaranteed to be monotonic, with respect to the input oper-
ands, throughout the domain supported by the instruction. (For the two operand functions,
monotonicity was proved by holding one of the operands constant.)
With the Intel486™ processor and Intel 387 math coprocessor, the worst-case, transcendental-
function error is typically 3 or 3.5 ulps, but is sometimes as large as 4.5 ulps.

7.5.11. FPU Control Instructions

The following instructions control the state and modes of operation of the FPU. They also allow
the status of the FPU to be examined:
FINIT/FNINIT Initialize FPU
FLDCW Load FPU control word
FSTCW/FNSTCW Store FPU control word
FSTSW/FNSTSW Store FPU status word
FCLEX/FNCLEX Clear FPU exception flags
FLDENV Load FPU environment
FSTENV/FNSTENV Store FPU environment
FRSTOR Restore FPU state
FSAVE/FNSAVE Save FPU state
FINCSTP Increment FPU register stack pointer
FDECSTP Decrement FPU register stack pointer
FFREE Free FPU register
FNOP No operation
WAIT/FWAIT Check for and handle pending unmasked FPU exceptions
The FINIT/FNINIT instructions initialize the FPU and its internal registers to default values.

7-41
FLOATING-POINT UNIT

The FLDCW instructions loads the FPU control word register with a value from memory. The
FSTCW/FNSTCW and FSTSW/FNSTSW instructions store the FPU control and status words,
respectively, in memory (or for an FSTSW/FNSTSW instruction in a general-purpose register).
The FSTENV/FNSTENV and FSAVE/FNSAVE instructions save the FPU environment and
state, respectively, in memory. The FPU environment includes all the FPU’s control and status
registers; the FPU state includes the FPU environment and the data registers in the FPU register
stack. (The FSAVE/FNSAVE instruction also initializes the FPU to default values, like the
FINIT/FNINIT instruction, after it saves the original state of the FPU.)
The FLDENV and FRSTOR instructions load the FPU environment and state, respectively,
from memory into the FPU. These instructions are commonly used when switching tasks or
contexts.
The WAIT/FWAIT instructions are synchronization instructions. (They are actually mnemonics
for the same opcode.) These instructions check the FPU status word for pending unmasked FPU
exceptions. If any pending unmasked FPU exceptions are found, they are handled before the
processor resumes execution of the instructions (integer, floating-point, or system instruction)
in the instruction stream. The WAIT/FWAIT instructions are provided to allow synchronization
of instruction execution between the FPU and the processor’s integer unit. Refer to Section 7.9.,
“Floating-Point Exception Synchronization” for more information on the use of the
WAIT/FWAIT instructions.

7.5.12. Waiting Vs. Non-waiting Instructions

All of the floating-point instructions except a few special control instructions perform a wait
operation (similar to the WAIT/FWAIT instructions), to check for and handle pending unmasked
FPU exceptions, before they perform their primary operation (such as adding two real numbers).
These instructions are called waiting instructions. Some of the FPU control instructions, such
as FSTSW/FNSTSW, have both a waiting and a non-waiting versions. The waiting version (with
the “F” prefix) executes a wait operation before it performs its primary operation; whereas, the
non-waiting version (with the “FN” prefix) ignores pending unmasked exceptions. Non-waiting
instructions allow software to save the current FPU state without first handling pending excep-
tions or to reset or reinitialize the FPU without regard for pending exceptions.

NOTE
When operating a Pentium® or Intel486™ processor in MS-DOS compati-
bility mode, it is possible (under unusual circumstances) for a non-waiting
instruction to be interrupted prior to being executed to handle a pending FPU
exception. The circumstances where this can happen and the resulting action
of the processor are described in Section E.2.1.3., “No-Wait FPU Instructions
Can Get FPU Interrupt in Window” in Appendix E, Guidelines for Writing
FPU Exceptions Handlers. When operating a Pentium® Pro processor in MS-
DOS compatibility mode, non-waiting instructions can not be interrupted in
this way (refer to Section E.2.2., “MS-DOS* Compatibility Mode in the P6
Family Processors” in Appendix E, Guidelines for Writing FPU Exceptions
Handlers).

7-42
 FLOATING-POINT UNIT

7.5.13. Unsupported FPU Instructions

The Intel 8087 instructions FENI and FDISI and the Intel 287 math coprocessor instruction
FSETPM perform no function in the Intel 387 math coprocessor, or the Intel486™, Pentium®,
or Pentium® Pro processors. If these opcodes are detected in the instruction stream, the FPU
performs no specific operation and no internal FPU states are affected.

7.6. OPERATING ON NANS

As was described in Section 7.2.3.4., “NaNs”, the FPU supports two types of NaNs: SNaNs and
QNaNs. An SNaN is any NaN value with its most-significant fraction bit set to 0 and at least
one other fraction bit set to 1. (If all the fraction bits are set to 0, the value is an ∞.) A QNaN is
any NaN value with the most-significant fraction bit set to 1. The sign bit of a NaN is not inter-
preted.
As a general rule, when a QNaN is used in one or more arithmetic floating-point instructions, it
is allowed to propagate through a computation. An SNaN on the other hand causes a floating-
point invalid operation exception to be signaled. SNaNs are typically used to trap or invoke an
exception handler. They must be inserted by software; that is, the FPU never generates an SNaN
as a result.
The floating-point invalid operation exception has a flag and a mask bit associated with it in the
FPU status and control registers, respectively (refer to Section 7.7., “Floating-Point Exception
Handling”). The mask bit determines how the FPU handles an SNaN value. If the floating-point
invalid operation mask bit is set, the SNaN is converted to a QNaN by setting the most-signifi-
cant fraction bit of the value to 1. The result is then stored in the destination operand and the
floating-point invalid operation flag is set. If the invalid operation mask is clear, a floating-point
invalid operation fault is signaled and no result is stored in the destination operand.
When a real operation or exception delivers a QNaN result, the value of the result depends on
the source operands, as shown in Table 7-18.
Except for the rules given at the beginning of this section for encoding SNaNs and QNaNs, soft-
ware is free to use the bits in the significand of a NaN for any purpose. Both SNaNs and QNaNs
can be encoded to carry and store data, such as diagnostic information.

7-43
FLOATING-POINT UNIT

Table 7-18. Rules for Generating QNaNs

Source Operands QNaN Result
An SNaN and a QNaN. The QNaN source operand.
Two SNaNs. The SNaN with the larger significand converted
into a QNaN.
Two QNaNs. The QNaN with the larger significand.
An SNaN and a real value. The SNaN converted into a QNaN.
A QNaN and a real value. The QNaN source operand.
Neither source operand is a NaN and a floating- The default QNaN real indefinite.
point invalid operation exception is signaled.

7.6.1. Operating on NaNs with Streaming SIMD Extensions

The information presented in Section 7.6., “Operating on NaNs”, is applicable to the floating-
point operations in the Streaming SIMD Extensions which operate on data in the floating-point
registers. Specific differences are noted in this section.
The invalid operation exception has a flag and a mask bit associated with it in MXCSR. The
mask bit determines how the an SNaN value is handled. If the invalid operation mask bit is set,
the SNaN is converted to a QNaN by setting the most-significant fraction bit of the value to 1.
The result is then stored in the destination operand and the invalid operation flag is set. If the
invalid operation mask is clear, an invalid operation fault is signaled and no result is stored in
the destination operand.
When a real operation or exception delivers a QNaN result, the value of the result depends on
the source operands, as shown in Table 7-19. The exceptions to the behavior described in
Table 7-19 are the MINPS and MAXPS instructions. If only one source is a NaN for these
instructions, the Src2 operand (either NaN or real value) is written to the result; this differs from
the behavior for other instructions as defined in Table 7-19, which is to always write the NaN to
the result, regardless of which source operand contains the NaN. This approach for
MINPS/MAXPS allows NaN data to be screened out of the bounds-checking portion of an algo-
rithm. If instead of this behavior, it is required that the NaN source operand be returned, the
min/max functionality can be emulated using a sequence of instructions: comparison followed
by AND, ANDN and OR.
In general Src1 and Src2 relate to an Katmai New Instruction instruction as follows:
ADDPS Src1, Src2/m128
Except for the rules given at the beginning of this section for encoding SNaNs and QNaNs, soft-
ware is free to use the bits in the significand of a NaN for any purpose. Both SNaNs and QNaNs
can be encoded to carry and store data, such as diagnostic information.

7-44
 FLOATING-POINT UNIT

Table 7-19. Results of Operations with NaN Operands

Source Operands NaN Result
(invalid operation exception is masked)
An SNaN and a QNaN. Src1 NaN (converted to QNaN if Src1 is an SNaN).
Two SNaNs. Src1 NaN (converted to QNaN )
Two QNaNs. Src1 QNaN
An SNaN and a real value. The SNaN converted into a QNaN.
A QNaN and a real value. The QNaN source operand.
An SNaN/QNaN value ( for instructions which take The SNaN converted into a QNaN/the source QNaN.
only one operand ie. RCPPS, RCPSS,
RSQRTPS, RSQRTSS)
Neither source operand is a NaN and a floating- The default QNaN real indefinite.
point invalid operation exception is signaled.

7.6.2. Uses for Signaling NANs

By unmasking the invalid operation exception, the programmer can use signaling NaNs to trap
to the exception handler. The generality of this approach and the large number of NaN values
that are available provide the sophisticated programmer with a tool that can be applied to a
variety of special situations.
For example, a compiler can use signaling NaNs as references to uninitialized (real) array
elements. The compiler can preinitialize each array element with a signaling NaN whose signif-
icand contained the index (relative position) of the element. Then, if an application program
attempts to access an element that it had not initialized, it can use the NaN placed there by the
compiler. If the invalid operation exception is unmasked, an interrupt will occur, and the excep-
tion handler will be invoked. The exception handler can determine which element has been
accessed, since the operand address field of the exception pointers will point to the NaN, and the
NaN will contain the index number of the array element.

7.6.3. Uses for Quiet NANs

Quiet NaNs are often used to speed up debugging. In its early testing phase, a program often
contains multiple errors. An exception handler can be written to save diagnostic information in
memory whenever it was invoked. After storing the diagnostic data, it can supply a quiet NaN
as the result of the erroneous instruction, and that NaN can point to its associated diagnostic area
in memory. The program will then continue, creating a different NaN for each error. When the
program ends, the NaN results can be used to access the diagnostic data saved at the time the
errors occurred. Many errors can thus be diagnosed and corrected in one test run.
In embedded applications which use computed results in further computations, an undetected
QNaN can invalidate all subsequent results. Such applications should therefore periodically
check for QNaNs and provide a recovery mechanism to be used if a QNaN result is detected.

7-45
FLOATING-POINT UNIT

7.7. FLOATING-POINT EXCEPTION HANDLING

The FPU detects six classes of exception conditions while executing floating-point instructions:

• Invalid operation (#I)

— Stack overflow or underflow (#IS)
— Invalid arithmetic operation (#IA)
• Divide-by-zero (#Z)
• Denormalized operand (#D)
• Numeric overflow (#O)
• Numeric underflow (#U)
• Inexact result (precision) (#P)
The nomenclature of “#” symbol followed by one or two letters (for example, #IS) is used in this
manual to indicate exception conditions. It is merely a short-hand form and is not related to
assembler mnemonics.
Each of the six exception classes has a corresponding flag bit in the FPU status word and a mask
bit in the FPU control word (refer to Section 7.3.2., “FPU Status Register” and Section 7.3.4.,
“FPU Control Word”, respectively). In addition, the exception summary (ES) flag in the status
word indicates when any of the exceptions has been detected, and the stack fault (SF) flag (also
in the status word) distinguishes between the two types of invalid operation exceptions.
When the FPU detects a floating-point exception, it sets the appropriate flags in the FPU status
word, then takes one of two possible courses of action:
• Handles the exception automatically, producing a predefined (and often times usable
result), while allowing program execution to continue undisturbed.
• Invokes a software exception handler to handle the exception.
The following sections describe how the FPU handles exceptions (either automatically or by
calling a software exception handler), how the FPU detects the various floating-point excep-
tions, and the automatic (masked) response to the floating-point exceptions.

7.7.1. Arithmetic vs. Non-arithmetic Instructions

When dealing with floating-point exceptions, it is useful to distinguish between arithmetic
instructions and non-arithmetic instructions. Non-arithmetic instructions have no operands
or do not make substantial changes to their operands. Arithmetic instructions do make signifi-
cant changes to their operands; in particular, they make changes that could result in a floating-
point exception being signaled. Table 7-20 lists the non-arithmetic and arithmetic instructions.
It should be noted that some non-arithmetic instructions can signal a floating-point stack (fault)
exception, but this exception is not the result of an operation on an operand.

7-46
 FLOATING-POINT UNIT

7.7.2. Automatic Exception Handling

If the FPU detects an exception condition for a masked exception (an exception with its mask
bit set), it sets the exception flag for the exception and delivers a predefined (default) response
and continues executing instructions. The masked (default) responses to exceptions have been
chosen to deliver a reasonable result for each exception condition and are generally satisfactory
for most floating-point applications. By masking or unmasking specific floating-point excep-
tions in the FPU control word, programmers can delegate responsibility for most exceptions to
the FPU and reserve the most severe exception conditions for software exception handlers.
Because the exception flags are “sticky,” they provide a cumulative record of the exceptions that
have occurred since they were last cleared. A programmer can thus mask all exceptions, run a
calculation, and then inspect the exception flags to see if any exceptions were detected during
the calculation.

7-47
FLOATING-POINT UNIT

Table 7-20. Arithmetic and Non-arithmetic Instructions

Non-arithmetic Instructions Arithmetic Instructions
FABS F2XM1
FCHS FADD/FADDP
FCLEX FBLD
FDECSTP FBSTP
FFREE FCOM/FCOMP/FCOMPP
FINCSTP FCOS
FINIT/FNINIT FDIV/FDIVP/FDIVR/FDIVRP
FLD (register-to-register) FIADD
FLD (extended format from memory) FICOM/FICOMP
FLD constant FIDIV/FIDIVR
FLDCW FILD
FLDENV FIMUL
FNOP FIST/FISTP
FRSTOR FISUB/FISUBR
FSAVE/FNSAVE FLD (conversion)
FST/FSTP (register-to-register) FMUL/FMULP
FSTP (extended format to memory) FPATAN
FSTCW/FNSTCW FPREM/FPREM1
FSTENV/FNSTENV FPTAN
FSTSW/FNSTSW FRNDINT
WAIT/FWAIT FSCALE
FXAM FSIN
FXCH FSINCOS
FSQRT
FST/FSTP (conversion)
FSUB/FSUBP/FSUBR/FSUBRP
FTST
FUCOM/FUCOMP/FUCOMPP
FXTRACT
FYL2X/FYL2XP1

7-48
 FLOATING-POINT UNIT

Note that when exceptions are masked, the FPU may detect multiple exceptions in a single
instruction, because it continues executing the instruction after performing its masked response.
For example, the FPU can detect a denormalized operand, perform its masked response to this
exception, and then detect numeric underflow.

7.7.3. Software Exception Handling

The FPU in the Pentium® Pro, Pentium®, and Intel486™ processors provides two different
modes of operation for invoking a software exception handler for floating-point exceptions:
native mode and MS-DOS compatibility mode. The mode of operation is selected with the NE
flag of control register CR0. (Refer to Chapter 2, System Architecture Overview, in the Intel
Architecture Software Developer’s Manual, Volume 3, for more information about the NE flag.)

7.7.3.1. NATIVE MODE

The native mode for handling floating-point exceptions is selected by setting the NE flag in
control register CR0 to 1. In this mode, if the FPU detects an exception condition while
executing a floating-point instruction and the exception is unmasked (the mask bit for the excep-
tion is cleared), the FPU sets the flag for the exception and the ES flag in the FPU status word.
It then invokes the software exception handler through the floating-point-error exception (#MF,
vector 16), immediately before execution of any of the following instructions in the processor’s
instruction stream:
• The next floating-point instruction, unless it is one of the non-waiting instructions
(FNINIT, FNCLEX, FNSTSW, FNSTCW, FNSTENV, and FNSAVE).
• The next WAIT/FWAIT instruction.
• The next MMX™ instruction.
If the next floating-point instruction in the instruction stream is a non-waiting instruction, the
FPU executes the instruction without invoking the software exception handler.

7.7.3.2. MS-DOS* COMPATIBILITY MODE

If the NE flag in control register CR0 is set to 0, the MS-DOS compatibility mode for handling
floating-point exceptions is selected. In this mode, the software exception handler for floating-
point exceptions is invoked externally using the processor’s FERR#, INTR, and IGNNE# pins.
This method of reporting floating-point errors and invoking an exception handler is provided to
support the floating-point exception handling mechanism used in PC systems that are running
the MS-DOS or Windows* 95 operating system.

7-49
FLOATING-POINT UNIT

The MS-DOS compatibility mode is typically used as follows to invoke the floating-point
exception handler:
1. If the FPU detects an unmasked floating-point exception, it sets the flag for the exception
and the ES flag in the FPU status word.
2. If the IGNNE# pin is deasserted, the FPU then asserts the FERR# pin either immediately,
or else delayed (deferred) until just before the execution of the next waiting floating-point
instruction or MMX™ instruction. Whether the FERR# pin is asserted immediately or
delayed depends on the type of processor, the instruction, and the type of exception.
3. If a preceding floating-point instruction has set the exception flag for an unmasked FPU
exception, the processor freezes just before executing the next WAIT instruction, waiting
floating-point instruction, or MMX™ instruction. Whether the FERR# pin was asserted at
the preceding floating-point instruction or is just now being asserted, the freezing of the
processor assures that the FPU exception handler will be invoked before the new floating-
point (or MMX™) instruction gets executed.
4. The FERR# pin is connected through external hardware to IRQ13 of a cascaded, program-
mable interrupt controller (PIC). When the FERR# pin is asserted, the PIC is programmed
to generate an interrupt 75H.
5. The PIC asserts the INTR pin on the processor to signal the interrupt 75H.
6. The BIOS for the PC system handles the interrupt 75H by branching to the interrupt 2
(NMI) interrupt handler.
7. The interrupt 2 handler determines if the interrupt is the result of an NMI interrupt or a
floating-point exception.
8. If a floating-point exception is detected, the interrupt 2 handler branches to the floating-
point exception handler.
If the IGNNE# pin is asserted, the processor ignores floating-point error conditions. This pin is
provided to inhibit floating-point exceptions from being generated while the floating-point
exception handler is servicing a previously signaled floating-point exception.
Appendix E, Guidelines for Writing FPU Exceptions Handlers, describes the MS-DOS compat-
ibility mode in much greater detail. This mode is somewhat more complicated in the Intel486™
and Pentium® processor implementations, as described in Appendix E, Guidelines for Writing
FPU Exceptions Handlers.

7.7.3.3. TYPICAL FLOATING-POINT EXCEPTION HANDLER ACTIONS

After the floating-point exception handler is invoked, the processor handles the exception in the
same manner that it handles non-FPU exceptions. (The floating-point exception handler is
normally part of the operating system or executive software.) A typical action of the exception
handler is to store FPU state information in memory (with the FSTENV/FNSTENV or
FSAVE/FNSAVE instructions) so that it can evaluate the exception and formulate an appropriate
response (refer to Section 7.3.9., “Saving the FPU’s State”).

7-50
 FLOATING-POINT UNIT

Other typical exception handler actions include:

• Examining stored FPU state information (control, status, and tag words, and FPU
instruction and operand pointers) to determine the nature of the error.
• Correcting the condition that caused the error.
• Clearing the exception bits in the status word.
• Returning to the interrupted program and resuming normal execution.
If the faulting floating-point instruction is followed by one or more non-floating-point instruc-
tions, it may not be useful to re-execute the faulting instruction. Refer to Section 7.9., “Floating-
Point Exception Synchronization”, for more information on synchronizing floating-point excep-
tions.
In cases where the handler needs to restart program execution with the faulting instruction, the
IRET instruction cannot be used directly. The reason for this is that because the exception is not
generated until the next floating-point or WAIT/FWAIT instruction following the faulting
floating-point instruction, the return instruction pointer on the stack may not point to the faulting
instruction. To restart program execution at the faulting instruction, the exception handler must
obtain a pointer to the instruction from the saved FPU state information, load it into the return
instruction pointer location on the stack, and then execute the IRET instruction.
In lieu of writing recovery procedures, the exception handler can do the following:
• Increment an exception counter for later display or printing.
• Print or display diagnostic information (such as, the FPU environment and registers).
• Halt further program execution.
Refer to Section E.3.3.4., “FPU Exception Handling Examples” in Appendix E, Guidelines for
Writing FPU Exceptions Handlers for general examples of floating-point exception handlers
and for specific examples of how to write a floating-point exception handler when using the MS-
DOS compatibility mode.

7.8. FLOATING-POINT EXCEPTION CONDITIONS

The following sections describe the various conditions that cause a floating-point exception to
be generated and the masked response of the FPU when these conditions are detected. Chapter
3, Instruction Set Reference, in the Intel Architecture Software Developer’s Manual, Volume 2,
lists the floating-point exceptions that can be signaled for each floating-point instruction.

7.8.1. Invalid Operation Exception

The floating-point invalid operation exception occurs in response to two general types of oper-
ations:
• Stack overflow or underflow (#IS).
• Invalid arithmetic operand (#IA).

7-51
FLOATING-POINT UNIT

The flag for this exception (IE) is bit 0 of the FPU status word, and the mask bit (IM) is bit 0 of
the FPU control word. The stack fault flag (SF) of the FPU status word indicates the type of
operation caused the exception. When the SF flag is set to 1, a stack operation has resulted in
stack overflow or underflow; when the flag is cleared to 0, an arithmetic instruction has encoun-
tered an invalid operand. Note that the FPU explicitly sets the SF flag when it detects a stack
overflow or underflow condition, but it does not explicitly clear the flag when it detects an
invalid-arithmetic-operand condition. As a result, the state of the SF flag can be 1 following an
invalid-arithmetic-operation exception, if it was not cleared from the last time a stack overflow
or underflow condition occurred. Refer to Section 7.3.2.4., “Stack Fault Flag”, for more infor-
mation about the SF flag.

7.8.1.1. STACK OVERFLOW OR UNDERFLOW EXCEPTION (#IS)

The FPU tag word keeps track of the contents of the registers in the FPU register stack (refer to
Section 7.3.6., “FPU Tag Word”). It then uses this information to detect two different types of
stack faults:
• Stack overflow—an instruction attempts to write a value into a non-empty FPU register
• Stack underflow—an instruction attempts to read a value from an empty FPU register.
When the FPU detects stack overflow or underflow, it sets the IE flag (bit 0) and the SF flag (bit
6) in the FPU status word to 1. It then sets condition-code flag C1 (bit 9) in the FPU status word
to 1 if stack overflow occurred or to 0 if stack underflow occurred.
If the invalid operation exception is masked, the FPU then returns the real, integer, or BCD-
integer indefinite value to the destination operand, depending on the instruction being executed.
This value overwrites the destination register or memory location specified by the instruction.
If the invalid operation exception is not masked, a software exception handler is invoked (refer
to Section 7.7.3., “Software Exception Handling”) and the top-of-stack pointer (TOP) and
source operands remain unchanged.
The term stack overflow comes from the condition where the a program has pushed eight values
onto the FPU register stack and the next value pushed on the stack causes a stack wraparound
to a register that already contains a value. The term stack underflow refers to the opposite condi-
tion from stack overflow. Here, a program has popped eight values from the FPU register stack
and the next value popped from the stack causes stack wraparound to an empty register.

7.8.1.2. INVALID ARITHMETIC OPERAND EXCEPTION (#IA)

The FPU is able to detect a variety of invalid arithmetic operations that can be coded in a
program. These operations generally indicate a programming error, such as dividing ∞ by ∞.
Table 7-21 lists the invalid arithmetic operations that the FPU detects. This group includes the
invalid operations defined in IEEE Standard 854.
When the FPU detects an invalid arithmetic operand, it sets the IE flag (bit 0) in the FPU status
word to 1. If the invalid operation exception is masked, the FPU then returns an indefinite value
to the destination operand or sets the floating-point condition codes, as shown in Table 7-21. If
the invalid operation exception is not masked, a software exception handler is invoked (refer to

7-52
 FLOATING-POINT UNIT

Section 7.7.3., “Software Exception Handling”) and the top-of-stack pointer (TOP) and source
operands remain unchanged.

Table 7-21. Invalid Arithmetic Operations and the Masked Responses to Them
Condition Masked Response
Any arithmetic operation on an operand that is in an Return the real indefinite value to the destination
unsupported format. operand.
Any arithmetic operation on a SNaN. Return a QNaN to the destination operand (refer
to Section 7.6., “Operating on NaNs”).
Compare and test operations: one or both operands Set the condition code flags (C0, C2, and C3) in
are NaNs. the FPU status word to 111B (not comparable).
Addition: operands are opposite-signed infinities. Return the real indefinite value to the destination
Subtraction: operands are like-signed infinities. operand.
Multiplication: ∞ by 0; 0 by ∞. Return the real indefinite value to the destination
operand.
Division: ∞ by ∞; 0 by 0. Return the real indefinite value to the destination
operand.
Remainder instructions FPREM, FPREM1: modulus Return the real indefinite; clear condition code
(divisor) is 0 or dividend is ∞. flag C2 to 0.
Trigonometric instructions FCOS, FPTAN, FSIN, Return the real indefinite; clear condition code
FSINCOS: source operand is ∞. flag C2 to 0.
FIST/FISTP instruction when input operand <> Return MAXNEG to destination operand.
MAXINT for destination operand size.
FSQRT: negative operand (except FSQRT (–0) = –0); Return the real indefinite value to the destination
FYL2X: negative operand (except FYL2X (–0) = –∞); operand.
FYL2XP1: operand more negative than –1.
FBSTP: source register is empty or it contains a NaN, Store BCD integer indefinite value in the
∞, or a value that cannot be represented in 18 destination operand.
decimal digits.
FXCH: one or both registers are tagged empty. Load empty registers with the real indefinite
value, then perform the exchange.

7.8.2. Divide-By-Zero Exception (#Z)

The FPU reports a floating-point zero-divide exception whenever an instruction attempts to
divide a finite non-zero operand by 0. The flag (ZE) for this exception is bit 2 of the FPU status
word, and the mask bit (ZM) is bit 2 of the FPU control word. The FDIV, FDIVP, FDIVR,
FDIVRP, FIDIV, and FIDIVR instructions and the other instructions that perform division inter-
nally (FYL2X and FXTRACT) can report the divide-by-zero exception.
When a divide-by-zero exception occurs and the exception is masked, the FPU sets the ZE flag
and returns the values shown in Table 7-21. If the divide-by-zero exception is not masked, the
ZE flag is set, a software exception handler is invoked (refer to Section 7.7.3., “Software Excep-
tion Handling”), and the top-of-stack pointer (TOP) and source operands remain unchanged.

7-53
FLOATING-POINT UNIT

Table 7-22. Divide-By-Zero Conditions and the Masked Responses to Them

Condition Masked Response
Divide or reverse divide operation Returns an ∞ signed with the exclusive OR of the sign of the two
with a 0 divisor. operands to the destination operand.
FYL2X instruction. Returns an ∞ signed with the opposite sign of the non-zero
operand to the destination operand.
FXTRACT instruction. ST(1) is set to –∞; ST(0) is set to 0 with the same sign as the
source operand.

7.8.3. Denormal Operand Exception (#D)

The FPU signals the denormal operand exception under the following conditions:
• If an arithmetic instruction attempts to operate on a denormal operand (refer to Section
7.2.3.2., “Normalized and Denormalized Finite Numbers”).
• If an attempt is made to load a denormal single- or double-real value into an FPU register.
(If the denormal value being loaded is an extended-real value, the denormal operand
exception is not reported.)
The flag (DE) for this exception is bit 1 of the FPU status word, and the mask bit (DM) is bit 1
of the FPU control word.
When a denormal operand exception occurs and the exception is masked, the FPU sets the DE
flag, then proceeds with the instruction. The denormal operand in single- or double-real format
is automatically normalized when converted to the extended-real format. Operating on denormal
numbers will produce results at least as good as, and often better than, what can be obtained
when denormal numbers are flushed to zero. In fact, subsequent operations will benefit from the
additional precision of the internal extended-real format. Most programmers mask this excep-
tion so that a computation may proceed, then analyze any loss of accuracy when the final result
is delivered.
When a denormal operand exception occurs and the exception is not masked, the DE flag is set
and a software exception handler is invoked (refer to Section 7.7.3., “Software Exception
Handling”). The top-of-stack pointer (TOP) and source operands remain unchanged. When
denormal operands have reduced significance due to loss of low-order bits, it may be advisable
to not operate on them. Precluding denormal operands from computations can be accomplished
by an exception handler that responds to unmasked denormal operand exceptions.

7.8.4. Numeric Overflow Exception (#O)

The FPU reports a floating-point numeric overflow exception (#O) whenever the rounded result
of an arithmetic instruction exceeds the largest allowable finite value that will fit into the real
format of the destination operand. For example, if the destination format is extended-real (80
bits), overflow occurs when the rounded result falls outside the unbiased range of −1.0 ∗ 216384
to 1.0 ∗ 216384 (exclusive). Numeric overflow can occur on arithmetic operations where the result
is stored in an FPU data register. It can also occur on store-real operations (with the FST and

7-54
 FLOATING-POINT UNIT

FSTP instructions), where a within-range value in a data register is stored in memory in a single-
or double-real format. The overflow threshold range for the single-real format is −1.0 ∗ 2128 to
1.0 ∗ 2128; the range for the double-real format is −1.0 ∗ 21024 to 1.0 ∗ 21024.
The numeric overflow exception cannot occur when overflow occurs when storing values in an
integer or BCD integer format. Instead, the invalid-arithmetic-operand exception is signaled.
The flag (OE) for the numeric overflow exception is bit 3 of the FPU status word, and the mask
bit (OM) is bit 3 of the FPU control word.
When a numeric overflow exception occurs and the exception is masked, the FPU sets the OE
flag and returns one of the values shown in Table 7-23. The value returned depends on the
current rounding mode of the FPU (refer to Section 7.3.4.3., “Rounding Control Field”).
.
Table 7-23. Masked Responses to Numeric Overflow
Rounding Mode Sign of True Result Result
To nearest + +∞
– –∞
Toward –∞ + Largest finite positive number
– –∞
Toward +∞ + +∞
– Largest finite negative number
Toward zero + Largest finite positive number
– Largest finite negative number

The action that the FPU takes when numeric overflow occurs and the numeric overflow excep-
tion is not masked, depends on whether the instruction is supposed to store the result in memory
or on the register stack.
If the destination is a memory location, the OE flag is set and a software exception handler is
invoked (refer to Section 7.7.3., “Software Exception Handling”). The top-of-stack pointer
(TOP) and source and destination operands remain unchanged.
If the destination is the register stack, the exponent of the rounded result is divided by 224576 and
the result is stored along with the significand in the destination operand. Condition code bit C1
in the FPU status word (called in this situation the “round-up bit”) is set if the significand was
rounded upward and cleared if the result was rounded toward 0. After the result is stored, the
OE flag is set and a software exception handler is invoked.
The scaling bias value 24,576 is equal to 3 ∗ 213. Biasing the exponent by 24,576 normally trans-
lates the number as nearly as possible to the middle of the extended-real exponent range so that,
if desired, it can be used in subsequent scaled operations with less risk of causing further
exceptions.
When using the FSCALE instruction, massive overflow can occur, where the result is too large
to be represented, even with a bias-adjusted exponent. Here, if overflow occurs again, after the
result has been biased, a properly signed ∞ is stored in the destination operand.

7-55
7.8.5. Numeric Underflow Exception (#U)
The FPU reports a floating-point numeric underflow exception (#U) whenever the rounded
result of an arithmetic instruction is “tiny” (that is, less than the smallest possible normalized,
finite value that will fit into the real format of the destination operand). For example, if the desti-
nation format is extended-real (80 bits), underflow occurs when the rounded result falls in the
unbiased range of −1.0 ∗ 2−16382 to 1.0 ∗ 2−16382 (exclusive). Like numeric overflow, numeric
underflow can occur on arithmetic operations where the result is stored in an FPU data register.
It can also occur on store-real operations (with the FST and FSTP instructions), where a within-
range value in a data register is stored in memory in a single- or double-real format. The under-
flow threshold range for the single-real format is −1.0 ∗ 2−126 to 1.0 ∗ 2−126; the range for the
double-real format is −1.0 ∗ 2−1022 to 1.0 ∗ 2−1022. (The numeric underflow exception cannot
occur when storing values in an integer or BCD integer format.)
The flag (UE) for the numeric-underflow exception is bit 4 of the FPU status word, and the mask
bit (UM) is bit 4 of the FPU control word.
When a numeric-underflow exception occurs and the exception is masked, the FPU denormal-
izes the result (refer to Section 7.2.3.2., “Normalized and Denormalized Finite Numbers”). If
the denormalized result is exact, the FPU stores the result in the destination operand, without
setting the UE flag. If the denormal result is inexact, the FPU sets the UE flag, then goes on to
handle the inexact result exception condition (refer to Section 7.8.6., “Inexact Result (Precision)
Exception (#P)”). It is important to note that if numeric-underflow is masked, a numeric-under-
flow exception is signaled only if the denormalized result is inexact. If the denormalized result
is exact, no flags are set and no exceptions are signaled.
The action that the FPU takes when numeric underflow occurs and the numeric-underflow
exception is not masked, depends on whether the instruction is supposed to store the result in
memory or on the register stack.
If the destination is a memory location, the UE flag is set and a software exception handler is
invoked (refer to Section 7.7.3., “Software Exception Handling”). The top-of-stack pointer
(TOP) and source and destination operands remain unchanged.
If the destination is the register stack, the exponent of the rounded result is multiplied by
224576 and the product is stored along with the significand in the destination operand. Condition
code bit C1 in the FPU the status register (acting here as a “round-up bit”) is set if the significand
was rounded upward and cleared if the result was rounded toward 0. After the result is stored,
the UE flag is set and a software exception handler is invoked.
The scaling bias value 24,576 is the same as is used for the overflow exception and has the same
effect, which is to translate the result as nearly as possible to the middle of the extended-real
exponent range.
When using the FSCALE instruction, massive underflow can occur, where the result is too tiny
to be represented, even with a bias-adjusted exponent. Here, if underflow occurs again, after the
result has been biased, a properly signed 0 is stored in the destination operand.
 FLOATING-POINT UNIT

7.8.6. Inexact Result (Precision) Exception (#P)

The inexact result exception (also called the precision exception) occurs if the result of an oper-
ation is not exactly representable in the destination format. For example, the fraction 1/3 cannot
be precisely represented in binary form. This exception occurs frequently and indicates that
some (normally acceptable) accuracy has been lost. The exception is supported for applications
that need to perform exact arithmetic only. Because the rounded result is generally satisfactory
for most applications, this exception is commonly masked. Note that the transcendental instruc-
tions [FSIN, FCOS, FSINCOS, FPTAN, FPATAN, F2XM1, FYL2X, and FYL2XP1] by nature
produce inexact results.

The inexact result exception flag (PE) is bit 5 of the FPU status word, and the mask bit (PM) is
bit 5 of the FPU control word.
If the inexact result exception is masked when an inexact result condition occurs and a numeric
overflow or underflow condition has not occurred, the FPU sets the PE flag and stores the
rounded result in the destination operand. The current rounding mode determines the method
used to round the result (refer to Section 7.3.4.3., “Rounding Control Field”). The C1 (round-
up) bit in the FPU status word indicates whether the inexact result was rounded up (C1 is set) or
“not rounded up” (C1 is cleared). In the “not rounded up” case, the least-significant bits of the
inexact result are truncated so that the result fits in the destination format.
If the inexact result exception is not masked when an inexact result occurs and numeric overflow
or underflow has not occurred, the FPU performs the same operation described in the previous
paragraph and, in addition, invokes a software exception handler (refer to Section 7.7.3., “Soft-
ware Exception Handling”).
If an inexact result occurs in conjunction with numeric overflow or underflow, one of the
following operations is carried out:
• If an inexact result occurs along with masked overflow or underflow, the OE or UE flag
and the PE flag are set and the result is stored as described for the overflow or underflow
exceptions (refer to Section 7.8.4., “Numeric Overflow Exception (#O)” or Section 7.8.5.,
“Numeric Underflow Exception (#U)”). If the inexact result exception is unmasked, the
FPU also invokes the software exception handler.
• If an inexact result occurs along with unmasked overflow or underflow and the destination
operand is a register, the OE or UE flag and the PE flag are set, the result is stored as
described for the overflow or underflow exceptions, and the software exception handler is
invoked.
• If an inexact result occurs along with unmasked overflow or underflow and the destination
operand is a memory location, the inexact result condition is ignored.

7.8.7. Exception Priority

The processor handles exceptions according to a predetermined precedence. When an instruc-
tion generates two or more exception conditions, the exception precedence sometimes results in
the higher-priority exception being handled and the lower-priority exceptions being ignored. For

7-57
FLOATING-POINT UNIT

example, dividing an SNaN by zero can potentially signal an invalid-arithmetic-operand excep-

tion (due to the SNaN operand) and a divide-by-zero exception. Here, if both exceptions are
masked, the FPU handles the higher-priority exception only (the invalid-arithmetic-operand
exception), returning a real indefinite to the destination. Alternately, a denormal operand or
inexact result exception can accompany a numeric underflow or overflow exception, with both
exceptions being handled.
The precedence for floating-point exceptions is as follows:
1. Invalid operation exception, subdivided as follows:
a. Stack underflow.
b. Stack overflow.
c. Operand of unsupported format.
d. SNaN operand.
2. QNaN operand. Though this is not an exception, the handling of a QNaN operand has
precedence over lower-priority exceptions. For example, a QNaN divided by zero results
in a QNaN, not a zero-divide exception.
3. Any other invalid operation exception not mentioned above or a divide-by-zero exception.
4. Denormal operand exception. If masked, then instruction execution continues, and a
lower-priority exception can occur as well.
5. Numeric overflow and underflow exceptions in conjunction with the inexact result
exception.
6. Inexact result exception.
Invalid operation, zero divide, and denormal operand exceptions are detected before a floating-
point operation begins, whereas overflow, underflow, and precision errors are not detected until
a true result has been computed. When a pre-operation exception is detected, the FPU register
stack and memory have not yet been updated, and appear as if the offending instructions has not
been executed. When a post-operation exception is detected, the register stack and memory
may be updated with a result (depending on the nature of the error).
For more information on the order in which multiple exceptions or interrupts are serviced, refer
to Section 5.7., “Priority Among Simultaneous Exceptions and Interrupts”, in Chapter 5, Inter-
rupt and Exception Handling, of the Intel Architecture Software Developer’s Manual, Volume 3.

7.9. FLOATING-POINT EXCEPTION SYNCHRONIZATION

Because the integer unit and FPU are separate execution units, it is possible for the processor to
execute floating-point, integer, and system instructions concurrently. No special programming
techniques are required to gain the advantages of concurrent execution. (Floating-point instruc-
tions are placed in the instruction stream along with the integer and system instructions.)
However, concurrent execution can cause problems for floating-point exception handlers.

7-58
 FLOATING-POINT UNIT

This problem is related to the way the FPU signals the existence of unmasked floating-point
exceptions. (Special exception synchronization is not required for masked floating-point excep-
tions, because the FPU always returns a masked result to the destination operand.)
When a floating-point exception is unmasked and the exception condition occurs, the FPU stops
further execution of the floating-point instruction and signals the exception event. On the next
occurrence of a floating-point instruction or a WAIT/FWAIT instruction in the instruction
stream, the processor checks the ES flag in the FPU status word for pending floating-point
exceptions. It floating-point exceptions are pending, the FPU makes an implicit call (traps) to
the floating-point software exception handler. The exception handler can then execute recovery
procedures for selected or all floating-point exceptions.
Synchronization problems occur in the time frame between when the exception is signaled and
when it is actually handled. Because of concurrent execution, integer or system instructions can
be executed during this time frame. It is thus possible for the source or destination operands for
a floating-point instruction that faulted to be overwritten in memory, making it impossible for
the exception handler to analyze or recover from the exception.
To solve this problem, an exception synchronizing instruction (either a floating-point instruction
or a WAIT/FWAIT instruction) can be placed immediately after any floating-point instruction
that might present a situation where state information pertaining to a floating-point exception
might be lost or corrupted. Floating-point instructions that store data in memory are prime candi-
dates for synchronization. For example, the following three lines of code have the potential for
exception synchronization problems:

FILD COUNT ; Floating-point instruction

INC COUNT ; Integer instruction
FSQRT ; Subsequent floating-point instruction
In this example, the INC instruction modifies the result of a floating-point instruction (FILD).
If an exception is signaled during the execution of the FILD instruction, the result stored in the
COUNT memory location might be overwritten before the exception handler is called.
Rearranging the instructions, as follows, so that the FSQRT instruction follows the FILD
instruction, synchronizes the exception handling and eliminates the possibility of the exception
being handled incorrectly.

FILD COUNT ; Floating-point instruction

FSQRT ; Subsequent floating-point instruction synchronizes
; any exceptions generated by the FILD instruction.
INC COUNT ; Integer instruction
The FSQRT instruction does not require any synchronization, because the results of this instruc-
tion are stored in the FPU data registers and will remain there, undisturbed, until the next
floating-point or WAIT/FWAIT instruction is executed. To absolutely insure that any exceptions
emanating from the FSQRT instruction are handled (for example, prior to a procedure call), a
WAIT instruction can be placed directly after the FSQRT instruction.
Note that some floating-point instructions (non-waiting instructions) do not check for pending
unmasked exceptions (refer to Section 7.5.11., “FPU Control Instructions”). They include the
FNINIT, FNSTENV, FNSAVE, FNSTSW, FNSTCW, and FNCLEX instructions. When an

7-59
FLOATING-POINT UNIT

FNINIT, FNSTENV, FNSAVE, or FNCLEX instruction is executed, all pending exceptions are
essentially lost (either the FPU status register is cleared or all exceptions are masked). The
FNSTSW and FNSTCW instructions do not check for pending interrupts, but they do not
modify the FPU status and control registers. A subsequent “waiting” floating-point instruction
can then handle any pending exceptions.

7-60
 8
Programming With
the Intel MMX™
Technology
 CHAPTER 8
PROGRAMMING WITH THE INTEL
MMX™ TECHNOLOGY

The Intel MMX™ technology comprises a set of extensions to the Intel Architecture (IA) that
are designed to greatly enhance the performance of advanced media and communications appli-
cations. These extensions (which include new registers, data types, and instructions) are
combined with a single-instruction, multiple-data (SIMD) execution model to accelerate the
performance of applications such as motion video, combined graphics with video, image
processing, audio synthesis, speech synthesis and compression, telephony, video conferencing,
and 2D and 3D graphics, which typically use compute-intensive algorithms to perform repeti-
tive operations on large arrays of simple, native data elements.
The MMX™ technology defines a simple and flexible software model, with no new mode or
operating-system visible state. All existing software will continue to run correctly, without
modification, on IA processors that incorporate the MMX™ technology, even in the presence
of existing and new applications that incorporate this technology.
The following sections of this chapter describe the MMX™ technology’s basic programming
environment, including the MMX™ register set, data types, and instruction set. Detailed
descriptions of the MMX™ instructions are provided in Chapter 3, Instruction Set Reference, of
the Intel Architecture Software Developer’s Manual, Volume 2. The manner in which the
MMX™ technology is integrated into the IA system programming model is described in
Chapter 10, MMX™ Technology System Programming, in the Intel Architecture Software Devel-
oper’s Manual, Volume 3.

8.1. OVERVIEW OF THE MMX™ TECHNOLOGY PROGRAMMING

ENVIRONMENT
MMX™ technology provides the following new extensions to the IA programming environ-
ment.
• Eight MMX™ registers (MM0 through MM7).
• Four MMX™ data types (packed bytes, packed words, packed doublewords, and
quadword).
• The MMX™ instruction set.
The MMX™ registers and data types are described in the following sections. Refer to Section
8.3., “Overview of the MMX™ Instruction Set”, for an overview of the MMX™ instructions.

8-1
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.1.1. MMX™ Registers

The MMX™ register set consists of eight 64-bit registers (refer to Figure 8-1). The MMX™
instructions access the MMX™ registers directly using the register names MM0 through MM7.
These registers can only be used to perform calculations on MMX™ data types; they cannot be
used to address memory. Addressing of MMX™ instruction operands in memory is handled by
using the standard IA addressing modes and general-purpose registers (EAX, EBX, ECX, EDX,
EBP, ESI, EDI, and ESP).

63 0
MM7
MM6
MM5
MM4

MM3
MM2
MM1
MM0

3006044

Figure 8-1. MMX™ Register Set

Although the MMX™ registers are defined in the IA as separate registers, they are aliased to the
registers in the FPU data register stack (R0 through R7). (Refer to Chapter 10, MMX™ Tech-
nology System Programming, in the Intel Architecture Software Developer’s Manual, Volume 3,
for a more detailed discussion of the aliasing of MMX™ registers.)

8-10
 PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.1.2. MMX™ Data Types

The MMX™ technology defines the following new 64-bit data types (refer to Figure 8-2):
Packed bytes Eight bytes packed into one 64-bit quantity.
Packed words Four (16-bit) words packed into one 64-bit quantity.
Packed doublewords Two (32-bit) doublewords packed into one 64-bit quantity.
Quadword One 64-bit quantity.
The bytes in the packed bytes data type are numbered 0 through 7, with byte 0 being contained
in the least significant bits of the data type (bits 0 through 7) and byte 7 being contained in the
most significant bits (bits 56 through 63). The words in the packed words data type are
numbered 0 through 4, with word 0 being contained in the bits 0 through 15 of the data type and
word 4 being contained in bits 48 through 63. The doublewords in a packed doublewords data
type are numbered 0 and 1, with doubleword 0 being contained in bits 0 through 31 and double-
word 1 being contained in bits 32 through 63.

Packed bytes (8x8 bits)

63 56 55 48 47 40 39 32 31 24 23 16 15 8 7 0

Packed word (4x16 bits)

63 48 47 32 31 16 15 0

Packed doublewords (2x32 bits)

63 32 31 0

Quadword (64 bits)

63 0

3006002

Figure 8-2. MMX™ Data Types

The MMX™ instructions move the packed data types (packed bytes, packed words, or packed
doublewords) and the quadword data type to-and-from memory or to-and-from the IA general-
purpose registers in 64-bit blocks. However, when performing arithmetic or logical operations
on the packed data types, the MMX™ instructions operate in parallel on the individual bytes,

8-11
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

words, or doublewords contained in a 64-bit MMX™ register, as described in the following

section (Section 8.1.3., “Single Instruction, Multiple Data (SIMD) Execution Model”).
When operating on the bytes, words, and doublewords within packed data types, the MMX™
instructions recognize and operate on both signed and unsigned byte integers, word integers, and
doubleword integers.

8.1.3. Single Instruction, Multiple Data (SIMD) Execution Model

The MMX™ technology uses the single instruction, multiple data (SIMD) technique for
performing arithmetic and logical operations on the bytes, words, or doublewords packed into
64-bit MMX™ registers. For example, the PADDSB instruction adds 8 signed bytes from the
source operand to 8 signed bytes in the destination operand and stores 8 byte-results in the desti-
nation operand. This SIMD technique speeds up software performance by allowing the same
operation to be carried out on multiple data elements in parallel. The MMX™ technology
supports parallel operations on byte, word, and doubleword data elements when contained in
MMX™ registers.
The SIMD execution model supported in the MMX™ technology directly addresses the needs
of modern media, communications, and graphics applications, which often use sophisticated
algorithms that perform the same operations on a large number of small data types (bytes, words,
and doublewords). For example, most audio data is represented in 16-bit (word) quantities. The
MMX™ instructions can operate on 4 of these words simultaneously with one instruction. Video
and graphics information is commonly represented as palletized 8-bit (byte) quantities. Here,
one MMX™ instruction can operate on 8 of these bytes simultaneously.

8.1.4. Memory Data Formats

When stored in memory the bytes, words, and doublewords in the packed data types are stored
in consecutive addresses, with the least significant byte, word, or doubleword being stored in
the lowest address and the more significant bytes, words, or doubleword being stored at consec-
utively higher addresses (refer to Figure 8-3). The ordering bytes, words, or doublewords in
memory is always little endian. That is, the bytes with the lower addresses are less significant
than the bytes with the higher addresses.

63 56 55 48 47 40 39 32 31 24 23 16 15 8 7 0
Byte 7 Byte 6 Byte 5 Byte 4 Byte 3 Byte 2 Byte 1 Byte 0

Memory Address 1008h Memory Address 1000h

3006045

Figure 8-3. Eight Packed Bytes in Memory (at address 1000H)

8-10
 PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.1.5. Data Formats for MMX™ Registers

Values in MMX™ registers have the same format as a 64-bit quantity in memory. MMX™
registers have two data access modes: 64-bit access mode and 32-bit access mode.
The 64-bit access mode is used for 64-bit memory access, 64-bit transfer between MMX™
registers, all pack, logical and arithmetic instructions, and some unpack instructions.
The 32-bit access mode is used for 32-bit memory access, 32-bit transfer between integer regis-
ters and MMX™ registers, and some unpack instructions.

8.2. MMX™ INSTRUCTION SET

The MMX™ instruction set consists of 57 instructions, grouped into the following categories:
• Data transfer instructions
• Arithmetic instructions
• Comparison instructions
• Conversion instructions
• Logical instructions
• Shift instructions
• Empty MMX™ state instruction (EMMS)
When operating on packed data within an MMX™ register, the data is cast by the type specified
by the instruction. For example, the PADDB (add packed bytes) instruction treats the packed
data in an MMX™ register as 8 packed bytes; whereas, the PADDW (add packed words)
instruction treats the packed data as 4 packed words. Refer to Section 9.3.6., “Additional SIMD
Integer Instructions”, in Chapter 9, Programming with the Streaming SIMD Extensions, for
additional SIMD integer instructions added with the Streaming SIMD Extensions.

8-11
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.2.1. Saturation Arithmetic and Wraparound Mode

The MMX™ technology supports a new arithmetic capability known as saturating arithmetic.
Saturation is best defined by contrasting it with wraparound mode.
In wraparound mode, results that overflow or underflow are truncated and only the lower (least
significant) bits of the result are returned; that is, the carry is ignored.
In saturation mode, results of an operation that overflow or underflow are clipped (saturated) to
a data-range limit for the data type (refer to Table 8-1). The result of an operation that exceeds
the range of a data-type saturates to the maximum value of the range. A result that is less than
the range of a data type saturates to the minimum value of the range. This method of handling
overflow and underflow is useful in many applications, such as color calculations.

Table 8-1. Data Range Limits for Saturation

Data Type Lower Limit Upper Limit
Hexadecimal Decimal Hexadecimal Decimal
Signed Byte 80H -128 7FH 127
Signed Word 8000H -32,768 7FFFH 32,767
Unsigned Byte 00H 0 FFH 255
Unsigned Word 0000H 0 FFFFH 65,535

For example, when the result exceeds the data range limit for signed bytes, it is saturated to 7FH
(FFH for unsigned bytes). If a value is less than the data range limit, it is saturated to 80H for
signed bytes (00H for unsigned bytes).
Saturation provides a useful feature of avoiding wraparound artifacts. In the example of color
calculations, saturation causes a color to remain pure black or pure white without allowing for
and inversion.
MMX™ instructions do not indicate overflow or underflow occurrence by generating excep-
tions or setting flags.

8-10
 PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.2.2. Instruction Operands

All MMX™ instructions, except the EMMS instruction, reference and operate on two operands:
the source and destination operands. The first operand is the destination and the second operand
is the source. The destination operand may also be a second source operand for the operation.
The instruction overwrites the destination operand with the result.
For example, a two-operand instruction would be decoded as:
DEST (first operand) ← DEST (first operand) OPERATION SRC (second operand)
The source operand for all the MMX™ instructions (except the data transfer instructions), can
reside either in memory or in an MMX™ register. The destination operand resides in an
MMX™ register.
For data transfer instructions, the source and destination operands can also be an integer register
(for the MOVD instruction) or memory location (for both the MOVD and MOVQ instructions).

8.3. OVERVIEW OF THE MMX™ INSTRUCTION SET

Table 8-2 shows the instructions in the MMX™ instruction set. The following sections give a
brief overview of each group of instructions in the MMX™ instruction set and the instructions
within each group. Refer to Section 9.3.6., “Additional SIMD Integer Instructions”, in Chapter
9, Programming with the Streaming SIMD Extensions, for additional SIMD integer instructions
added with the Streaming SIMD Extensions.

8.3.1. Data Transfer Instructions

The MOVD (Move 32 Bits) instruction transfers 32 bits of packed data from memory to
MMX™ registers and visa versa, or from integer registers to MMX™ registers and visa versa.
The MOVQ (Move 64 Bits) instruction transfers 64-bits of packed data from memory to
MMX™ registers and vise versa, or transfers data between MMX™ registers.

8-11
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

Table 8-2. MMX™ Instruction Set Summary

Category Wraparound Signed Unsigned
Saturation Saturation
Arithmetic Addition PADDB, PADDW, PADDSB, PADDUSB,
PADDD PADDSW PADDUSW
Subtraction PSUBB, PSUBW, PSUBSB, PSUBUSB,
PSUBD PSUBSW PSUBUSW
Multiplication PMULL, PMULH
Multiply and Add PMADD
Comparison Compare for Equal PCMPEQB,
PCMPEQW,
PCMPEQD
Compare for PCMPGTPB,
Greater Than PCMPGTPW,
PCMPGTPD
Conversion Pack PACKSSWB, PACKUSWB
PACKSSDW
Unpack High PUNPCKHBW,
PUNPCKHWD,
PUNPCKHDQ
Unpack Low PUNPCKLBW,
PUNPCKLWD,
PUNPCKLDQ
Packed Full Quadword
Logical And PAND
And Not PANDN
Or POR
Exclusive OR PXOR
Shift Shift Left Logical PSLLW, PSLLD PSLLQ
Shift Right Logical PSRLW, PSRLD PSRLQ
Shift Right PSRAW, PSRAD
Arithmetic
Doubleword Transfers Quadword Transfers
Data Transfer Register to Register MOVD MOVQ
Load from Memory MOVD MOVQ
Store to Memory MOVD MOVQ
Empty EMMS
MMX™ State

8-10
 PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.3.2. Arithmetic Instructions

The arithmetic instructions perform addition, subtraction, multiplication, and multiply/add oper-
ations on packed data types.

8.3.2.1. PACKED ADDITION AND SUBTRACTION

The PADDSB, PADDSW, and PADDWD (packed add) and PSUBB, PSUBW, and PSUBD
(packed subtract) instructions add or subtract the signed or unsigned data elements of the source
operand to or from the destination operand in wrap- around mode. These instructions support
packed byte, packed word, and packed doubleword data types.
The PADDSB and PADDSW (packed add with saturation) and PSUBSB and PSUBSW (packed
subtract with saturation) instructions add or subtract the signed data elements of the source
operand to or from the signed data elements of the destination operand and saturate the result to
the limits of the signed data-type range. These instructions support packed byte and packed word
data types.
The PADDUSB and PADDUSW (packed add unsigned with saturation) and PSUBUSB and
PSUBUSW (packed subtract unsigned with saturation) instructions add or subtract the unsigned
data elements of the source operand to or from the unsigned data elements of the destination
operand and saturate the result to the limits of the unsigned data-type range. These instructions
support packed byte and packed word data types.

8.3.2.2. PACKED MULTIPLICATION

Packed multiplication instructions perform four multiplications on pairs of signed 16-bit oper-
ands, producing 32-bit intermediate results. Users may choose the low-order or high-order parts
of each 32-bit result.
The PMULHW (packed multiply high) and PMULLW (packed multiply low) instructions
multiply the signed words of the source and destination operands and write the high-order or
low-order 16 bits of each of the results to the destination operand.

8.3.2.3. PACKED MULTIPLY ADD

The PMADDWD (packed multiply and add) instruction calculates the products of the signed
words of the source and destination operands. The four intermediate 32-bit doubleword products
are summed in pairs to produce two 32-bit doubleword results.

8.3.3. Comparison Instructions

The PCMPEQB, PCMPEQW, and PCMPEQD (packed compare for equal) and PCMPGTB,
PCMPGTW, and PCMPGTD (packed compare for greater than) instructions compare the corre-
sponding data elements in the source and destination operands for equality or value greater than,
respectively. These instructions generate a mask of ones or zeros which are written to the desti-
nation operand. Logical operations can use the mask to select elements. This can be used to

8-11
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

implement a packed conditional move operation without a branch or a set of branch instructions.
No flags are set.
These instructions support packed byte, packed word and packed doubleword data types.

8.3.4. Conversion Instructions

The conversion instructions convert the data elements within a packed data type.
The PACKSSWB and PACKSSDW (packed with signed saturation) instruction converts signed
words into signed bytes or signed doublewords into signed words, in signed saturation mode.
The PACKUSWB (packed with unsigned saturation) instruction converts signed words into
unsigned bytes, in unsigned saturation mode.
The PUNPCKHBW, PUNPCKHWD, and PUNPCKHDQ (unpack high packed data) and
PUNPCKLBW, PUNPCKLWD, and PUNPCKLDQ (unpack low packed data) instructions
convert bytes to words, words to doublewords, or doublewords to quadwords.

8.3.5. Logical Instructions

The PAND (bitwise logical AND), PANDN (bitwise logical AND NOT), POR (bitwise logical
OR), and PXOR (bitwise logical exclusive OR) instructions perform bitwise logical operations
on 64-bit quantities.

8.3.6. Shift Instructions

The logical shift left, logical shift right and arithmetic shift right instructions shift each element
by a specified number of bits. The logical left and right shifts also enable a 64-bit quantity (quad-
word) to be shifted as one block, assisting in data type conversions and alignment operations.
The PSLLW and PSLLD (packed shift left logical) and PSRLW and PSRLD (packed shift right
logical) instructions perform a logical left or right shift, and fill the empty high or low order bit
positions with zeros. These instructions support packed word, packed doubleword, and quad-
word data types.
The PSRAW and PSRAD (packed shift right arithmetic) instruction performs an arithmetic right
shift, copying the sign bit into empty bit positions on the upper end of the operand. This instruc-
tion supports packed word and packed doubleword data types.

8.3.7. EMMS (Empty MMX™ State) Instruction

The EMMS instruction empties the MMX™ state. This instruction must be used to clear the
MMX™ state (empty the floating-point tag word) at the end of an MMX™ routine before
calling other routines that can execute floating-point instructions.

8-10
 PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.4. COMPATIBILITY WITH FPU ARCHITECTURE

The MMX™ state is aliased upon the IA floating-point state. No new state or mode is added to
support the MMX™ technology. The same floating-point instructions that save and restore the
floating-point state also handle the MMX™ state (for example, during context switching).
MMX™ technology uses the same interface techniques between the floating-point architecture
and the operating system (primarily for task switching purposes). For more details, refer to
Chapter 10, MMX™ Technology System Programming, in the Intel Architecture Software Devel-
oper’s Manual, Volume 3.

8.4.1. MMX™ Instructions and the Floating-Point Tag Word

After each MMX™ instruction, the entire floating-point tag word is set to Valid (00s). The
Empty MMX™ state (EMMS) instruction sets the entire floating-point tag word to Empty (11s).
Chapter 10, MMX™ Technology System Programming, in the Intel Architecture Software Devel-
oper’s Manual, Volume 3, describes the effects of floating-point and MMX™ instructions on the
floating-point tag word. For details on floating-point tag word, refer to Section 7.3.6., “FPU Tag
Word” in Chapter 7, Floating-Point Unit.

8.4.2. Effect of Instruction Prefixes on MMX™ Instructions

Table 8-3 details the effect of an instruction prefix on an MMX™ instruction.

Table 8-3. Effect of Prefixes on MMX™ Instructions

Prefix Type Effect of Prefix
Address size (67H) Affects MMX™ instructions with a memory operand.
Ignored by MMX™ instructions without a memory operand.
Operand size (66H) Reserved.
Segment override Affects MMX™ instructions with a memory operand.
Ignored by MMX™ instructions without a memory operand.
Repeat Reserved.
Lock (F0H) Generates an invalid opcode exception.

Refer to Section 2.2., “Instruction Prefixes” in Chapter 2, Instruction Format of the Intel Archi-
tecture Software Developer’s Manual, Volume 2, for detailed information on prefixes.

8.5. WRITING APPLICATIONS WITH MMX™ CODE

The following sections give guidelines for writing applications code using the MMX™ tech-
nology.

8-11
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.5.1. Detecting Support for MMX™ Technology Using the

CPUID Instruction
Use the CPUID instruction to determine whether the processor supports the MMX™ instruction
set (refer to Section 3.2., “Instruction Reference” in Section 3, “Instruction Set Reference” of
the Intel Architecture Software Developer’s Manual, Volume 2, for a detailed description of the
CPUID instruction). When the support for MMX™ technology is detected by the CPUID
instruction, it is signaled by setting bit 23 (MMX™ technology bit) in the feature flags to 1. In
general, two versions of the routine can be created: one with scalar instructions and one with
MMX™ instructions. The application will call the appropriate routine depending on the results
of the CPUID instruction. If support for MMX™ technology is detected, then the MMX™
routine is called; if no support for the MMX™ technology exists, the application calls the scalar
routine.

NOTE
The CPUID instruction will continue to report the existence of the MMX™
technology if the CR0.EM bit is set (which signifies that the CPU is
configured to generate exception interrupt 7 that can be used to emulate
floating-point instructions). In this case, executing an MMX™ instruction
results in an invalid opcode exception.
Example 8-1 illustrates how to use the CPUID instruction. This example does not represent the
entire CPUID sequence, but shows the portion used for detection of MMX™ technology.

Example 8-1. Partial Routine for Detecting MMX™ Technology with the CPUID Instruction
; identify existence of CPUID instruction

; identify Intel processor

movEAX, 1; request for feature flags

CPUID ; 0Fh, 0A2h CPUID instruction
testEDX, 00800000h; Is IA MMX technology bit (Bit 23 of EDX)
; in feature flags set?
jnz MMX_Technology_Found

8.5.2. Using the EMMS Instruction

When integrating an MMX™ routine into an application running under an existing operating
system, programmers need to take special precautions, similar to those when writing floating-
point code.
When an MMX™ instruction executes, the floating-point tag word is marked valid (00s). Subse-
quent floating-point instructions that will be executed may produce unexpected results because
the floating-point stack seems to contain valid data. The EMMS instruction marks the floating-

8-10
 PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

point tag word as empty. Therefore, it is imperative to use the EMMS instruction at the end of
every MMX™ routine, if the next routine may contain FPU code.
The EMMS instruction must be used in each of the following cases:
• When an application using the floating-point instructions calls an MMX™ technology
library/DLL. (Use the EMMS instruction at the end of the MMX™ code.)
• When an application using MMX™ instructions calls a floating-point library/DLL. (Use
the EMMS instruction before calling the floating-point code.)
• When a switch is made between MMX™ code in a task/thread and other tasks/threads in
cooperative operating systems, unless it is certain that more MMX™ instructions will be
executed before any FPU code.
If the EMMS instruction is not used when trying to execute a floating-point instruction, the
following may occur:
• Depending on the exception mask bits of the floating-point control word, a floating- point
exception event may be generated.
• A “soft exception” may occur. In this case floating-point code continues to execute, but
generates incorrect results. This happens when the floating-point exceptions are masked
and no visible exceptions occur. The internal exception handler (microcode, not user
visible) loads a NaN (Not a Number) with an exponent of 11..11B onto the floating-point
stack. The NaN is used for further calculations, yielding incorrect results.
• A potential error may occur only if the operating system does NOT manage floating-point
context across task switches. These operating systems are usually cooperative operating
systems. It is imperative that the EMMS instruction execute at the end of all the MMX™
routines that may enable a task switch immediately after they end execution (explicit yield
API or implicit yield API).
• The EMMS instruction is not returned when mixing MMX™ technology instructions and
Streaming SIMD Extensions. Refer to Section 9.4., “Compatibility with FPU Archi-
tecture” in Chapter 9.4., Compatibility with FPU Architecture, of the Intel Architecture
Software Developer’s Manual, Volume 3, for more detailed information.

8.5.3. Interfacing with MMX™ Code

The MMX™ technology enables direct access to all the MMX™ registers. This means that all
existing interface conventions that apply to the use of the processor’s general-purpose registers
(EAX, EBX, etc.) also apply to use of MMX™ register.
An efficient interface to MMX™ routines might pass parameters and return values through the
MMX™ registers or through a combination of memory locations (via the stack) and MMX™
registers. Such an interface would have to be written in assembly language since passing param-
eters through MMX™ registers is not currently supported by any existing C compilers. Do not
use the EMMS instruction when the interface to the MMX™ code has been defined to retain
values in the MMX™ register.

8-11
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

If a high-level language, such as C, is used, the data types could be defined as a 64-bit structure
with packed data types.
When implementing usage of MMX™ instructions in high-level languages other approaches
can be taken, such as:
• Passing parameters to an MMX™ routine by passing a pointer to a structure via the integer
stack.
• Returning a value from a function by returning the pointer to a structure.

8.5.4. Writing Code with MMX™ and Floating-Point Instructions

The MMX™ technology aliases the MMX™ registers on the floating-point registers. The main
reason for this is to enable MMX™ technology to be fully compatible and transparent to existing
software environments (operating systems and applications). This way operating systems will
be able to include new applications and drivers that use the MMX™ technology.
An application can contain both floating-point and MMX™ code. However, the user is discour-
aged from causing frequent transitions between MMX™ and floating-point instructions by
mixing MMX™ code and floating-point code.

8.5.4.1. RECOMMENDATIONS AND GUIDELINES

Do not mix MMX™ code and floating-point code at the instruction level for the following
reasons:
• The TOS (top of stack) value of the floating-point status word is set to 0 after each
MMX™ instruction. This means that the floating-point code loses its pointer to its
floating-point registers if the code mixes MMX™ instructions within a floating-point
routine.
• An MMX™ instruction write to an MMX™ register writes ones (11s) to the exponent part
of the corresponding floating-point register.
• Floating-point code that uses register contents that were generated by the MMX™ instruc-
tions may cause floating-point exceptions or incorrect results. These floating-point
exceptions are related to undefined floating-point values and floating-point stack usage.
• All MMX™ instructions (except EMMS) set the entire tag word to the valid state (00s in
all tag fields) without preserving the previous floating-point state.
• Frequent transitions between the MMX™ and floating-point instructions may result in
significant performance degradation in some implementations.

8-10
If the application contains floating-point and MMX™ instructions, follow these guidelines:
• Partition the MMX™ technology module and the floating-point module into separate
instruction streams (separate loops or subroutines) so that they contain only instructions of
one type.
• Do not rely on register contents across transitions.
• When the MMX™ state is not required, empty the MMX™ state using the EMMS
instruction.
• Exit the floating-point code section with an empty stack.

Example 8-2. Floating-point (FP) and MMX™ Code

FP_code:
..
.. (*leave the FPU stack empty*)
MMX_code:
..
EMMS (*mark the FPU tag word as empty*)

FP_code 1:
..
.. (*leave the FPU stack empty*)

8.5.5. Using MMX™ Code in a Multitasking Operating System

Environment
An application needs to identify the nature of the multitasking operating system on which it
runs. Each task retains its own state which must be saved when a task switch occurs. The
processor state (context) consists of the general-purpose registers and the floating-point and
MMX™ registers.
Operating systems can be classified into two types:
• Cooperative multitasking operating system.
• Preemptive multitasking operating system.
The behavior of the two operating-system types in context switching is described in Section
10.4., “Designing Operating System Task and Context Switching Facilities” in Chapter 10,
MMX™ Technology System Programming, of the Intel Architecture Software Developer’s
Manual, Volume 3.
PROGRAMMING WITH THE INTEL MMX™ TECHNOLOGY

8.5.5.1. COOPERATIVE MULTITASKING OPERATING SYSTEM

Cooperative multitasking operating systems do not save the FPU or MMX™ state when
performing a context switch. Therefore, the application needs to save the relevant state before
relinquishing direct or indirect control to the operating system.

8.5.5.2. PREEMPTIVE MULTITASKING OPERATING SYSTEM

Preemptive multitasking operating systems are responsible for saving and restoring the FPU and
MMX™ state when performing a context switch. Therefore, the application does not have to
save or restore the FPU and MMX™ state.

8.5.6. Exception Handling in MMX™ Code

MMX™ instructions generate the same type of memory-access exceptions as other IA instruc-
tions. Some examples are: page fault, segment not present, and limit violations. Existing excep-
tion handlers can handle these types of exceptions. They do not have to be modified.
Unless there is a pending floating-point exception, MMX™ instructions do not generate
numeric exceptions. Therefore, there is no need to modify existing exception handlers or add
new ones.
If a floating-point exception is pending, the subsequent MMX™ instruction generates a numeric
error exception (interrupt 16 and/or FERR#). The MMX™ instruction resumes execution upon
return from the exception handler.

8.5.7. Register Mapping

The MMX™ registers and their tags are mapped to physical locations of the floating-point regis-
ters and their tags. Register aliasing and mapping is described in more detail in Chapter 10,
MMX™ Technology System Programming Model, in the Intel Architecture Software Devel-
oper’s Manual, Volume 3.

8-10
 9
Programming With
the Streaming SIMD
Extensions
 CHAPTER 9
PROGRAMMING WITH THE STREAMING SIMD
EXTENSIONS

The Intel Streaming SIMD Extensions comprise a set of extensions to the Intel Architecture (IA)
that is designed to greatly enhance the performance of advanced media and communications
applications. These extensions (which include new registers, data types, and instructions) are
combined with a single-instruction, multiple-data (SIMD) execution model to accelerate the
performance of applications. Applications that typically use compute-intensive algorithms to
perform repetitive operations on large arrays of simple, native data elements benefit the most.
Applications that require regular access to large amount of data also benefit from the Streaming
SIMD Extensions prefetching and streaming stores capabilities.
Examples of these types of applications include:
• motion video
• combined graphics with video
• image processing
• audio synthesis
• speech recognition, synthesis, and compression
• telephony
• video conferencing
• 2D and 3D graphics.
The Streaming SIMD Extensions define a simple and flexible software model. This new mode
introduces a new operating-system visible state. To enhance performance and yield more
concurrent execution, a new set of registers has been added. All existing software will continue
to run correctly without modification on IA processors that incorporate the Streaming SIMD
Extensions, even in the presence of existing and new applications that incorporate this tech-
nology.
The following sections of this chapter describe the Streaming SIMD Extensions’ basic program-
ming environment, including the SIMD floating-point register set, data types, and instruction
set. Detailed descriptions of the Streaming SIMD Extensions are provided in Chapter 3, Instruc-
tion Set Reference, of the Intel Architecture Software Developer’s Manual, Volume 2. The
manner in which the Streaming SIMD Extensions are integrated into the IA system program-
ming model is described in Chapter 10, MMX™ Technology System Programming, in the Intel
Architecture Software Developer’s Manual, Volume 3.

9-1
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9.1. OVERVIEW OF THE STREAMING SIMD EXTENSIONS

The Streaming SIMD Extensions introduce new, general-purpose, floating-point instructions,
that operate on a new set of eight 128-bit SIMD floating-point registers. This set enables the
programmer to develop algorithms that can finely-mix packed, single-precision, floating-point
and integer using both Streaming SIMD Extensions and MMX™ instructions respectively. In
addition to these instructions, Streaming SIMD Extensions also provide new instructions to
control cacheability of all MMX™ technology and 32-bit IA data types. These instructions
include the ability to stream data to memory without polluting the caches, and the ability to
prefetch data before it is actually used.
The Streaming SIMD Extensions provide the following new extensions to the IA programming
environment:
• Eight SIMD floating-point registers (XMM0 through XMM7).
• SIMD floating-point data type - 128-bit, packed floating-point.
• The Streaming SIMD Extensions set.
The SIMD floating-point registers and data types are described in the following sections. Refer
to Section 9.3., “Overview of the Streaming SIMD Extensions Set”, for an overview of the
Streaming SIMD Extensions.

9.1.1. SIMD Floating-Point Registers

The IA Streaming SIMD Extensions provide eight 128-bit general-purpose registers, each of
which can be directly addressed. These registers are new, and require support from the operating
system to use them.
The SIMD floating-point registers hold packed 128-bit data. The Streaming SIMD Extensions
access the SIMD floating-point registers directly using the register names XMM0 to XMM7
(Figure 9-1). SIMD floating-point registers can be used to perform calculations on data; they
cannot be used to address memory. Addressing is accomplished by using the integer registers
and standard IA addressing modes and general-purpose registers (EAX, EBX, ECX, EDX, EBP,
ESI, EDI, and ESP).
There is a new control/status register MXCSR, that is used to mask/unmask numerical exception
handling, to set rounding modes, to set flush-to-zero mode, and to view status flags.

9-2
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

XMM7
XMM6
XMM5
XMM4
XMM3
XMM2
XMM1
XMM0

Figure 9-1. SIMD Floating-Point Registers

MMX™ registers are mapped onto the floating-point registers. Transitioning from MMX™
operations to floating-point operations required executing the EMMS instruction. Since SIMD
floating-point registers are a separate register file, MMX™ instructions and floating-point
instructions can be mixed with Streaming SIMD Extensions without execution of a special
instruction such as EMMS.

9.1.2. SIMD Floating-Point Data Types

The principal data type of the IA Streaming SIMD Extensions is a packed, single-precision,
floating-point operand, specifically:
• Four 32-bit single-precision (SP), floating-point numbers (Figure 9-2)
The new SIMD-integer instructions will operate on the packed byte, word or doubleword data
types. The new prefetch instruction works on typeless data of size 32 bytes or greater.

127 96 95 65 63 32 31 0

Packed Single-FP

Figure 9-2. Packed Single-FP

9-3
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

The 32-bit, single-precision, floating-point numbers (doublewords) are numbered 0 through 3,

with 0 being contained in the least significant 32-bits (doubleword) of the register.
The Streaming SIMD Extensions move the packed data types (single-precision, floating-point -
doublewords) to-and-from memory in 64-bit or 128-bit blocks. However, when performing
arithmetic or logical operations on the packed data types, the Streaming SIMD Extensions
operate in parallel on the individual doublewords contained in the SIMD floating-point regis-
ters, as described in the following, Section 9.1.3., “Single Instruction, Multiple Data (SIMD)
Execution Model”.
The new SIMD-integer instructions follow the conventions of the MMX™ instructions and
operate on data in the MMX™ registers, not the SIMD floating-point 128-bit registers (refer to
Section 8.1.1., “MMX™ Registers” and Section 8.1.2., “MMX™ Data Types” in Chapter 8,
Programming with the Intel MMX™ Technology).

9.1.3. Single Instruction, Multiple Data (SIMD) Execution Model

The Streaming SIMD Extensions use the Single Instruction, Multiple Data (SIMD) technique
for performing arithmetic and logical operations on the single-precision, floating-point values
in the 128-bit SIMD floating-point registers. This technique speeds up software performance by
processing multiple data elements in parallel, using a single instruction. The Streaming SIMD
Extensions support operations on packed, single-precision, floating-point data types, and the
additional SIMD Integer instructions support operations on packed quadrate data types (byte,
word, or doubleword).
This approach was chosen because most media processing applications have the following char-
acteristics:
• inherently parallel;
• wide dynamic range, hence floating-point based;
• regular and re-occurring memory access patterns;
• localized re-occurring operations performed on the data;
• data independent control flow.
The Streaming SIMD Extensions are 100% compatible with the IEEE Standard 754 for Binary
Floating-Point Arithmetic. The Streaming SIMD Extensions are accessible from all IA execu-
tion modes: Protected mode, Real-address mode, and Virtual 8086 mode.

9.1.4. Pentium® III Processor Single Precision Floating-Point

Format
The Pentium® III processor’s SIMD floating-point instructions operate on a 32-bit single preci-
sion floating-point number. For specific information and details on real numbers and special
values represented by the IEEE single precision (32-bit) format, and how the Pentium® III

9-4
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

processor operates on these values, refer to Section 7.2., “Real Numbers and Floating-Point
Formats” in Chapter 7, Floating-Point Unit.

9.1.5. Memory Data Formats

The IA Streaming SIMD Extensions introduces a new packed 128-bit data type that consists of
four, single-precision, floating-point numbers. The 128 bits are numbered 0 through 127. Bit 0
is the least significant bit (LSB), and bit 127 is the most significant bit (MSB). Bytes in the new
data type format have consecutive memory addresses. The ordering is always little endian, that
is, the bytes with the lower addresses are less significant than the bytes with the higher addresses
(Figure 9-3).

Byte 15 Byte 0

15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0

Memory Address 1016d Memory Address 1000d

Figure 9-3. Four Packed FP Data in Memory (at address 1000H)

9.1.6. SIMD Floating-Point Register Data Formats

Values in SIMD floating-point registers have the same format as a 128-bit quantity in memory.
They have two data access modes: 128-bit access mode and 32-bit access mode. The data type
corresponds directly to the single-precision format in the IEEE standard. Table 9-1 gives the
precision and range of this data type. Only the fraction part of the significand is encoded. The
integer is assumed to be 1 for all numbers, except 0 and denormalized finite numbers. The expo-
nent of the single-precision data type is encoded in biased format. The biasing constant is 127
for the single-precision format.

Table 9-1. Precision and Range of SIMD Floating-point Datatype

Data Type Length Precision Approximate Normalized Range
(Bits)
Binary Decimal
single-precision 32 24 –126
2 to 2 127
1.18 × 10 –38
to 1.70 × 1038

Table 9-2 shows the encodings for all the classes of real numbers (that is, zero, denormalized-
finite, normalized-finite, and ∞) and NaNs for the single-real data-type. It also gives the format
for the real indefinite value, which is a QNaN encoding that is generated by several Streaming
SIMD Extensions in response to a masked, floating-point, invalid operation exception.

9-5
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

When storing real values in memory, single-real values are stored in 4 consecutive bytes in
memory. The 128-bit access mode is used for 128-bit memory accesses, 128-bit transfers
between SIMD floating-point registers, and all logical, unpack and arithmetic instructions. The
32-bit access mode is used for 32-bit memory access, 32-bit transfers between SIMD floating-
point registers, and all arithmetic instructions.

Table 9-2. Real Number and NaN Encodings

Class Sign Biased Exponent Significand
1
Integer Fraction
Positive +∞ 0 11..11 1 00..00
+Normals 0 11..10 1 11..11
. . . .
. . . .
0 00..01 1 00..00
+Denormals 0 00..00 0 11.11
. . . .
. . . .
0 00..00 0 00..01
+Zero 0 00..00 0 00..00
Negative −Zero 1 00..00 0 00..00
−Denormals 1 00..00 0 00..01
. . . .
. . . .
1 00..00 0 11..11
−Normals 1 00..01 1 00..00
. . . .
. . . .
1 11..10 1 11..11
-∞ 1 11..11 1 00..00
NaNs SNaN X 11..11 1 0X..XX2
QNaN X 11..11 1 1X..XX
Real Indefinite 1 11..11 1 10..00
(QNaN)
Single ← 8 Bits → ← 23 Bits →

NOTES:
1. Integer bit is implied and not stored for single-real and double-real formats.
2. The fraction for SNaN encodings must be non-zero.

9-6
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9.1.7. SIMD Floating-Point Control/Status Register

The control/status register is used to enable masked/unmasked numerical exception handling, to
set rounding modes, to set flush-to-zero mode, and to view status flags. The contents of this
register can be loaded with the LDMXCSR and FXRSTOR instructions and stored in memory
with the STMXCSR and FXSAVE instructions. Figure 9-4 shows the format and encoding of
the fields in the MXCSR.

31-16 15 10 5 0
Reserved F R R P U O Z D I R P U O Z D I
Z C C M M M M M M s E E E E E E
v
d

Figure 9-4. SIMD Floating-Point Control/Status Register Format

Bits 5-0 indicate whether a SIMD floating-point numerical exception has been detected. They
are “sticky” flags, and can be cleared by using the LDMXCSR instruction to write zeroes to
these fields. If an LDMXCSR instruction clears a mask bit and sets the corresponding exception
flag bit, an exception will not be immediately generated. The exception will occur only upon the
next Streaming SIMD Extensions to cause this type of exception. Streaming SIMD Extensions
use only one exception flag for each exception. There is no provision for individual exception
reporting within a packed data type. In situations where multiple identical exceptions occur
within the same instruction, the associated exception flag is updated and indicates that at least
one of these conditions happened. These flags are cleared upon reset.
Bits 12-7 configure numerical exception masking; an exception type is masked if the corre-
sponding bit is set, and it is unmasked if the bit is clear. These enables are set upon reset,
meaning that all numerical exceptions are masked.
Bits 14-13 encode the rounding control, which provides for the common round to nearest mode,
as well as directed rounding and true chop (refer to Section 9.1.8., “Rounding Control Field”).
The rounding control is set to round to nearest upon reset.
Bit 15 (FZ) is used to turn on the Flush-To-Zero mode (refer to Section 9.1.9., “Flush-To-Zero”).
This bit is cleared upon reset, disabling the Flush-To-Zero mode.
The other bits of MXCSR (bits 31-16 and bit 6) are defined as reserved and cleared; attempting
to write a non-zero value to these bits, using either the FXRSTOR or LDMXCSR instructions,
will result in a general protection exception.

9-7
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9.1.8. Rounding Control Field

The rounding control (RC) field of MXCSR (bits 13 and 14) controls how the results of floating-
point instructions are rounded. Four rounding modes are supported: round to nearest, round up,
round down, and round toward zero (refer to Table 9-3). Round to nearest is the default rounding
mode and is suitable for most applications. It provides the most accurate and statistically unbi-
ased estimate of the true result.

Table 9-3. Rounding Control Field (RC)

Rounding RC Field
Mode Setting Description
Round to 00B Rounded result is the closest to the infinitely precise result. If two values
nearest (even) are equally close, the result is the even value (that is, the one with the
least-significant bit of zero).
Round down 01B Rounded result is close to but no greater than the infinitely precise
(toward −∞) result.
Round up 10B Rounded result is close to but no less than the infinitely precise result.
(toward +∞)
Round toward 11B Rounded result is close to but no greater in absolute value than the
zero (Truncate) infinitely precise result.

The round up and round down modes are termed directed rounding and can be used to imple-
ment interval arithmetic. Interval arithmetic is used to determine upper and lower bounds for the
true result of a multistep computation, when the intermediate results of the computation are
subject to rounding.
The round toward zero mode (sometimes called the “chop” mode) is commonly used when
performing integer arithmetic with the processor.

9.1.9. Flush-To-Zero
Turning on the Flush-To-Zero mode has the following effects during underflow situations:
• Zero results are returned with the sign of the true result
• Precision and underflow exception flags are set
The IEEE mandated masked response to underflow is to deliver the denormalized result (i.e.,
gradual underflow); consequently, the Flush-To-Zero mode is not compatible with IEEE Stan-
dard 754. It is provided primarily for performance reasons. At the cost of a slight precision loss,
faster execution can be achieved for applications where underflows are common. Underflow for
Flush-To-Zero is defined to occur when the exponent for a computed result, prior to denormal-
ization scaling, falls in the denormal range; this is regardless of whether a loss of accuracy has
occurred. Unmasking the underflow exception takes precedence over Flush-To-Zero mode; this
means that an exception handler will be invoked for a Streaming SIMD Extensions instruction
that generates an underflow condition while this exception is unmasked, regardless of whether
Flush-To-Zero is enabled.

9-8
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9.2. STREAMING SIMD EXTENSIONS SET

The Streaming SIMD Extensions set consists of 70 instructions, grouped into the following cate-
gories:
• Data movement instructions
• Arithmetic instructions
• Comparison instructions
• Conversion instructions
• Logical instructions
• Additional SIMD integer instructions
• Shuffle instructions
• State management instructions
• Cacheability control instructions

9.2.1. Instruction Operands

The IA Streaming SIMD Extensions supply a rich set of instructions that operate on either all or
the least significant pairs of packed data operands in parallel. The packed instructions operate
on a pair of operands as shown in Figure 9-5 while scalar instructions always operate on the least
significant pair of the two operands as shown in Figure 9-6; for scalar operations, the three upper
components from the first operand are passed through to the destination. In general, the address
of a memory operand has to be aligned on a 16-byte boundary for all instructions, except for
unaligned loads and stores.

X1 (SP) X2 (SP) X3 (SP) X4 (SP)

Y1 (SP) Y2 (SP) Y3 (SP) Y4 (SP)

OP OP OP OP

X1 op Y1 (SP) X2 op Y2 (SP) X3 op Y3 (SP) X4 op Y4 (SP)

Figure 9-5. Packed Operations

9-9
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

X1 (SP) X2 (SP) X3 (SP) X4 (SP)

Y1 (SP) Y2 (SP) Y3 (SP) Y4 (SP)

OP

X1 (SP) X2 (SP) X3 (SP) X4 op Y4 (SP)

Figure 9-6. Scalar Operations

9.3. OVERVIEW OF THE STREAMING SIMD EXTENSIONS SET

Appendix D, SIMD Floating-Point Exceptions Summary shows the instructions in the
Streaming SIMD Extensions set. The following sections give a brief overview of each group of
instructions in the Streaming SIMD Extensions set and the instructions within each group.

9.3.1. Data Movement Instructions

The MOVAPS (Move aligned packed, single-precision, floating-point) instruction transfers 128
bits of packed data from memory to SIMD floating-point registers and vice versa, or between
SIMD floating-point registers. The memory address is aligned to 16-byte boundary; otherwise,
a general protection exception will occur.
The MOVUPS (Move unaligned packed, single-precision, floating-point) instruction transfers
128 bits of packed data from memory to SIMD floating-point registers and vice versa, or
between SIMD floating-point registers. No assumption is made for alignment.
The MOVHPS (Move unaligned, high packed, single-precision, floating-point) instruction
transfers 64 bits of packed data from memory to the upper two fields of a SIMD floating-point
register and vice versa. The lower two fields are left unchanged.
The MOVHLPS (Move high to low packed, single-precision, floating-point) instruction trans-
fers the upper 64-bits of the source register into the lower 64-bits of the 128-bit destination
register. The upper 64-bits of the destination register are left unchanged.
The MOVLHPS (Move low to high packed, single-precision, floating-point) instruction trans-
fers the lower 64-bits of the source register into the upper 64-bits of the 128-bit destination
register. The lower 64-bits of the destination register are left unchanged.

9-10
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

The MOVLPS (Move unaligned, low packed, single-precision, floating-point) instruction trans-
fers 64 bits of packed data from memory to the lower two fields of a SIMD floating-point
register and vice versa. The upper two fields are left unchanged.
The MOVMSKPS (Move mask packed, single-precision, floating-point) instruction transfers
the most significant bit of each of the four, packed, single-precision, floating-point numbers to
an IA integer register. This 4-bit value can then be used as a condition to perform branching.
The MOVSS (Move scalar single-precision, floating-point) instruction transfers the least signif-
icant 32 bits from memory to a SIMD floating-point register or vice versa, and between regis-
ters.

9.3.2. Arithmetic Instructions

9.3.2.1. PACKED/SCALAR ADDITION AND SUBTRACTION

The ADDPS (Add packed, single-precision, floating-point) and SUBPS (Subtract packed,
single-precision, floating-point) instructions add or subtract four pairs of packed, single-preci-
sion, floating-point operands.
The ADDSS (Add scalar single-precision, floating-point) and SUBSS (Subtract scalar single-
precision, floating-point) instructions add or subtract the least significant pair of packed, single-
precision, floating-point operands; the upper three fields are passed through from the source
operand.

9.3.2.2. PACKED/SCALAR MULTIPLICATION AND DIVISION

The MULPS (Multiply packed, single-precision, floating-point) instruction multiplies four pairs
of packed, single-precision, floating-point operands.
The MULSS (Multiply scalar single-precision, floating-point) instruction multiplies the least
significant pair of packed, single-precision, floating-point operands; the upper three fields are
passed through from the source operand.
The DIVPS (Divide packed, single-precision, floating-point) instruction divides four pairs of
packed, single-precision, floating-point operands.
The DIVSS (Divide scalar single-precision, floating-point) instruction divides the least signifi-
cant pair of packed, single-precision, floating-point operands; the upper three fields are passed
through from the source operand.

9.3.2.3. PACKED/SCALAR SQUARE ROOT

The SQRTPS (Square root packed, single-precision, floating-point) instruction returns the
square root of the packed four single-precision, floating-point numbers from the source to a
destination register.

9-11
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

The SQRTSS (Square root scalar single-precision, floating-point) instruction returns the square
root of the least significant component of the packed, single-precision, floating-point numbers
from source to a destination register; the upper three fields are passed through from the source
operand.

9.3.2.4. PACKED MAXIMUM/MINIMUM

The MAXPS (Maximum packed, single-precision, floating-point) instruction returns the
maximum of each pair of packed, single-precision, floating-point numbers into the destination
register. (destreg = {MAX xmm1[1], xmm2[1]; MAX xmm1[2], xmm2[2]; MAX xmm1[3],
xmm2[3]; MAX xmm1[4], xmm2[4]})
The MAXSS (Maximum scalar single-precision, floating-point) instructions returns the
maximum of the least significant pair of packed, single-precision, floating-point numbers into
the destination register; the upper three fields are passed through from the source operand, to the
destination register.
The MINPS (Minimum packed, single-precision, floating-point) instruction returns the
minimum of each pair of packed, single-precision, floating-point numbers into the destination
register. (destreg = {MIN xmm1[1], xmm2[1]; MIN xmm1[2], xmm2[2]; MIN xmm1[3],
xmm2[3]; MIN xmm1[4], xmm2[4]})
The MINSS (Minimum scalar single-precision, floating-point) instruction returns the minimum
of the least significant pair of packed, single-precision, floating-point numbers into the destina-
tion register; the upper three fields are passed through from the source operand, to the destina-
tion register.

9.3.3. Comparison Instructions

The CMPPS (Compare packed, single-precision, floating-point) instruction compares four pairs
of packed, single-precision, floating-point numbers using the immediate operand as a predicate,
returning per SP field an all "1" 32-bit mask or an all "0" 32-bit mask as a result. The instruction
supports a full set of 12 conditions: equal, less than, less than equal, greater than, greater than or
equal, unordered, not equal, not less than, not less than or equal, not greater than, not greater
than or equal, ordered.
The CMPSS (Compare scalar single-precision, floating-point) instruction compares the least
significant pairs of packed, single-precision, floating-point numbers using the immediate
operand as a predicate (same as CMPPS), returning per SP field an all "1" 32-bit mask or an all
"0" 32-bit mask as a result.
The COMISS (Compare scalar single-precision, floating-point ordered and set EFLAGS)
instruction compares the least significant pairs of packed, single-precision, floating-point
numbers, and sets the ZF, PF, and CF bits in the EFLAGS register (the OF, SF, and AF bits are
cleared).
The UCOMISS (Unordered compare scalar single-precision, floating-point ordered and set
EFLAGS) instruction compares the least significant pairs of packed, single-precision, floating-

9-12
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

point numbers, and sets the ZF, PF, and CF bits in the EFLAGS register as described above (the
OF, SF, and AF bits are cleared).

9.3.4. Conversion Instructions

These instructions support packed and scalar conversions between 128-bit SIMD floating-point
registers and either 64-bit integer MMX™ registers or 32-bit integer IA32 registers. The packed
versions behave identically to original MMX™ instructions, in the presence of x87-FP instruc-
tions, including:
• Transition from x87-FP to MMX™ technology (TOS=0, FP valid bits set to all valid).
• MMX™ instructions write ones (1s) to the exponent part of the corresponding x87-FP
register.
• Use of EMMS for transition from MMX™ technology to x87-FP.
The CVTPI2PS (Convert packed 32-bit integer to packed, single-precision, floating-point)
instruction converts two 32-bit signed integers in an MMX™ register to the two least significant
single-precision, floating-point numbers. When the conversion is inexact, the rounded value
according to the rounding mode in MXCSR is returned. The upper two significant numbers in
the destination register are retained.
The CVTSI2SS (Convert scalar 32-bit integer to scalar single-precision, floating-point) instruc-
tion converts a 32-bit signed integer in an MMX™ register to the least significant single-preci-
sion, floating-point number. When the conversion is inexact, the rounded value according to the
rounding mode in MXCSR is returned. The upper three significant numbers in the destination
register are retained.
The CVTPS2PI (Convert packed, single-precision, floating-point to packed 32-bit integer)
instruction converts the two least significant single-precision, floating-point numbers to two 32-
bit signed integers in an MMX™ register. When the conversion is inexact, the rounded value
according to the rounding mode in MXCSR is returned. The CVTTPS2PI (Convert truncate
packed, single-precision, floating-point to packed 32-bit integer) instruction is similar to
CVTPS2PI, except if the conversion is inexact, in which case the truncated result is returned.
The CVTSS2SI (Convert scalar single-precision, floating-point to a 32-bit integer) instruction
converts the least significant single-precision, floating-point number to a 32-bit signed integer
in an IA 32-bit integer register. When the conversion is inexact, the rounded value according to
the rounding mode in MXCSR is returned. The CVTTSS2SI (Convert truncate scalar single-
precision, floating-point to scalar 32-bit integer) instruction is similar to CVTSS2SI, except if
the conversion is inexact, the truncated result is returned.

9.3.5. Logical Instructions

The ANDPS (Bit-wise packed logical AND for single-precision, floating-point) instruction
returns a bitwise AND between the two operands.

9-13
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

The ANDNPS (Bit-wise packed logical AND NOT for single-precision, floating-point) instruc-
tion returns a bitwise AND NOT between the two operands.
The ORPS (Bit-wise packed logical OR for single-precision, floating-point) instruction returns
a bitwise OR between the two operands.
The XORPS (Bit-wise packed logical XOR for single-precision, floating-point) instruction
returns a bitwise XOR between the two operands.

9.3.6. Additional SIMD Integer Instructions

Similar to the conversion instructions discussed in Section 9.3.4., “Conversion Instructions”,
these SIMD Integer instructions also behave identically to original MMX™ instructions, in the
presence of x87-FP instructions.
The PAVGB/PAVGW (Average unsigned source sub-operands, without incurring a loss in preci-
sion) instructions add the unsigned data elements of the source operand to the unsigned data
elements of the destination register. The results of the add are then each independently shifted
right by one bit position. The high order bits of each element are filled with the carry bits of the
sums. To prevent cumulative round-off errors, an averaging is performed. The low order bit of
each final shifted result is set to 1 if at least one of the two least significant bits of the interme-
diate unshifted shifted sum is 1.
The PEXTRW (Extract 16-bit word from MMX™ register) instruction moves the word in an
MMX™ register selected by the two least significant bits of the immediate operand to the lower
half of a 32-bit integer register; the upper word in the integer register is cleared.
The PINSRW (Insert 16-bit word into MMX™ register) instruction moves the lower word in a
32-bit integer register or 16-bit word from memory into one of the four word locations in an
MMX™ register, selected by the two least significant bits of the immediate operand.
The PMAXUB/PMAXSW (Maximum of packed unsigned integer bytes or signed integer
words) instructions return the maximum of each pair of packed elements into the destination
register.
The PMINUB/PMINSW (Minimum of packed unsigned integer bytes or signed integer words)
instructions return the minimum of each pair of packed data elements into the destination
register.
The PMOVMSKB (Move Byte Mask from MMX™ register) instruction returns an 8-bit mask
formed of the most significant bits of each byte of its source operand in an MMX™ register to
an IA integer register.
The PMULHUW (Unsigned high packed integer word multiply in MMX™ register) instruction
performs an unsigned multiply on each word field of the two source MMX™ registers, returning
the high word of each result to an MMX™ register.
The PSADBW (Sum of absolute differences) instruction computes the absolute difference for
each pair of sub-operand byte sources, and then accumulates the eight differences into a single
16-bit result.

9-14
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

The PSHUFW (Shuffle packed integer word in MMX™ register) instruction performs a full
shuffle of any source word field to any result word field, using an 8-bit immediate operand.

9.3.7. Shuffle Instructions

The SHUFPS (Shuffle packed, single-precision, floating-point) instruction is able to shuffle any
of the packed four single-precision, floating-point numbers from one source operand to the
lower two destination fields; the upper two destination fields are generated from a shuffle of any
of the four SP FP numbers from the second source operand (Figure 9-7). By using the same
register for both sources, SHUFPS can return any combination of the four SP FP numbers from
this register.

X4 X3 X2 X1

Y4 Y3 Y2 Y1

{Y4 ... Y1} {Y4 ... Y1} {X4 ... X1} {X4 ... X1}

Figure 9-7. Packed Shuffle Operation

The UNPCKHPS (Unpacked high packed, single-precision, floating-point) instruction performs

an interleaved unpack of the high-order data elements of first and second packed, single-preci-
sion, floating-point operands. It ignores the lower half part of the sources (Figure 9-8). When
unpacking from a memory operand, the full 128-bit operand is accessed from memory, but only
the high order 64 bits are utilized by the instruction.

9-15
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

X4 X3 X2 X1

Y4 Y3 Y2 Y1

Y4 X4 Y3 X3

Figure 9-8. Unpack High Operation

The UNPCKLPS (Unpacked low packed, single-precision, floating-point) instruction performs

an interleaved unpack of the low-order data elements of first and second packed, single-preci-
sion, floating-point operands. It ignores the higher half part of the sources (Figure 9-9). When
unpacking from a memory operand, the full 128-bit operand is accessed from memory, but only
the low order 64 bits are utilized by the instruction.

X4 X3 X2 X1

Y4 Y3 Y2 Y1

Y2 X2 Y1 X1

Figure 9-9. Unpack Low Operation

9.3.8. State Management Instructions

The LDMXCSR (Load SIMD Floating-Point Control and Status Register) instruction loads the
SIMD floating-point control and status register from memory. STMXCSR (Store SIMD

9-16
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

Floating-Point Control and Status Register) instruction stores the Streaming SIMD Extensions
control and status word to memory.
The FXSAVE instruction saves FP and MMX™ state and SIMD floating-point state to memory.
Unlike FSAVE, FXSAVE it does not clear the x87-FP state. FXRSTOR loads FP and MMX™
state and SIMD floating-point state from memory.

9.3.9. Cacheability Control Instructions

Data referenced by a programmer can have temporal (data will be used again) or spatial (data
will be in adjacent locations, e.g. same cache line) locality. Some multimedia data types, such
as the display list in a 3D graphics application, are referenced once and not reused in the imme-
diate future. We will refer to this data type as non-temporal data. Thus, the programmer does not
want the application’s cached code and data to be overwritten by this non-temporal data. The
cacheability control instructions enable the programmer to control caching so that non-temporal
accesses will minimize cache pollution.
In addition, the execution engine needs to be fed such that it does not become stalled waiting for
data. Streaming SIMD Extensions allow the programmer to prefetch data long before its final
use. These instructions are not architectural since they do not update any architectural state and
are specific to each implementation. The programmer may have to tune his application for each
implementation to take advantage of these instructions. These instructions merely provide a hint
to the hardware, and they will not generate exceptions or faults. Excessive use of prefetch
instructions may degrade processor performance due to resource allocation.
The following three instructions provide programmatic control for minimizing cache pollution
when writing data to memory from either the MMX™ registers or the SIMD floating-point regis-
ters.
• The MASKMOVQ (Non-temporal byte mask store of packed integer in an MMX™
register) instruction stores data from an MMX™ register to the location specified by the
(DS) EDI register. The most significant bit in each byte of the second MMX™ mask
register is used to selectively write the data of the first register on a per-byte basis. The
instruction is implicitly weakly-ordered, with all of the characteristics of the WC memory
type; successive non-temporal stores may not write memory in program-order, do not
write-allocate (i.e., the processor will not fetch the corresponding cache line into the cache
hierarchy, prior to performing the store), write combine/collapse, and minimize cache
pollution.
• The MOVNTQ (Non-temporal store of packed integer in an MMX™ register) instruction
stores data from an MMX™ register to memory. The instruction is implicitly weakly-
ordered, does not write-allocate, and minimizes cache pollution.
• The MOVNTPS (Non-temporal store of packed, single-precision, floating-point)
instruction stores data from a SIMD floating-point register to memory. The memory
address must be aligned to a 16-byte boundary; if it is not aligned, a general protection
exception will occur. The instruction is implicitly weakly-ordered, does not write-allocate,
and minimizes cache pollution.

9-17
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

The non-temporal store instructions (MOVNTPS, MOVNTQ, and MASKMOVQ) minimize

cache pollution while writing data. The main difference between a non-temporal store and a
regular cacheable store is in the write-allocation policy. The memory type of the region being
written to can override the non-temporal hint, leading to the following considerations. If the
programmer specifies a non-temporal store to:
• Uncacheable memory, the store behaves like an uncacheable store; the non-temporal hint is
ignored, and the memory type for the region is retained. Uncacheable as referred to here
means that the region being written to has been mapped with either a UC or WP memory
type. If the memory region has been mapped as WB, WT, or WC, the non-temporal store
will implement weakly-ordered (WC) semantic behavior.
• Cacheable memory, two cases may result. If the data is:
• Present in the cache hierarchy, the hint is ignored and the cache line is updated
normally. A given processor may choose different ways to implement this; some
examples include: updating data in-place in the cache hierarchy while preserving
the memory type semantics assigned to that region, or evicting the data from the
caches and writing the new non-temporal data to memory (with WC semantics).
• Not present in the cache hierarchy, and the destination region is mapped as WB,
WT, or WC, the transaction will be weakly-ordered, and is subject to all WC
memory semantics; consequently, the programmer is responsible for maintaining
coherency. The non-temporal store will not write allocate (i.e., the processor will
not fetch the corresponding cache line into the cache hierarchy, prior to
performing the store). Different implementations may choose to collapse and
combine these stores prior to issuing them to memory.
In general, WC semantics require software to ensure coherency, with respect to other processors
and other system agents (such as graphics cards). Appropriate use of synchronization and a
fencing operation (refer to SFENCE, below) must be performed for producer-consumer usage
models. Fencing ensures that all system agents have global visibility of the stored data. For
instance, failure to fence may result in a written cache line staying within a processor, and the
line would not be visible to other agents. For processors that implement non-temporal stores by
updating data in-place that already resides in the cache hierarchy, the destination region should
also be mapped as WC. Otherwise, if mapped as WB or WT, there is the potential for speculative
processor reads to bring the data into the caches. In this case, non-temporal stores would then
update in place, and data would not be flushed from the processor by a subsequent fencing oper-
ation.
The memory type visible on the bus in the presence of memory type aliasing is implementation-
specific. As one possible example, the memory type written to the bus may reflect the memory
type for the first store to this line, as seen in program order; other alternatives are possible. This
behavior should be considered reserved, and dependence on the behavior of any particular
implementation risks future incompatibility.
The PREFETCH (Load 32 or greater number of bytes) instructions load either non-temporal
data or temporal data in the specified cache level. This access and the cache level are specified
as a hint. The prefetch instructions do not affect functional behavior of the program and will be
implementation-specific.

9-18
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

For more information on prefetch hints, refer to Section 9.5.3.1., “Cacheability Hint Instruc-
tions”. For even more detailed information, refer to Chapter 6, “Optimizing Cache Utilization
for Pentium® III Processors”, in the Intel Architecture Optimization Reference Manual (Order
Number 245127-001).
The SFENCE (Store Fence) instruction guarantees that every store instruction that precedes the
store fence instruction in program order is globally visible before any store instruction that
follows the fence. The SFENCE instruction provides an efficient way of ensuring ordering
between routines that produce weakly-ordered results and routines that consume this data.
The use of weakly-ordered memory types can be important under certain data sharing relation-
ships, such as a producer-consumer relationship. The use of weakly-ordered memory can make
the assembling of data more efficient, but care must be taken to ensure that the consumer obtains
the data that the producer intended it to see.

9.4. COMPATIBILITY WITH FPU ARCHITECTURE

The Streaming SIMD Extensions introduce a new state in the architecture. It is not aliased onto
the floating-point registers as are the MMX™ instructions. New instructions must be used to
save/restore the state of a Pentium® III processor.
The interface for context switching is discussed in detail in Section 11.5., “Saving and Restoring
the Streaming SIMD Extensions state” and Section 11.6., “Designing Operating System Task
and Context Switching Facilities” in Chapter 11, Streaming SIMD Extensions System Program-
ming, of the Intel Architecture Software Developer’s Manual, Volume 3.

9.4.1. Effect of Instruction Prefixes on Streaming SIMD

Extensions
The Streaming SIMD Extensions use prefixes as specified in Table 9-4, Table 9-5, and Table
9-6. The effect of multiple prefixes (more than one prefix from a group) is unpredictable and
may vary from processor to processor. Applying a prefix, in a manner not defined in this docu-
ment, is considered reserved behavior. For example, Table 9-4 shows general behavior for most
Streaming SIMD Extensions; however, the application of a prefix (Repeat, Repeat NE, Operand
Size) is reserved for the following instructions: ANDPS, ANDNPS, COMISS, FXRSTOR,
FXSAVE, ORPS, LDMXCSR, MOVAPS, MOVHPS, MOVLPS, MOVMSKPS, MOVNTPS,
MOVUPS, SHUFPS, STMXCSR, UCOMISS, UNPCKHPS, UNPCKLPS, XORPS.

9-19
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

Table 9-4. Streaming SIMD Extensions Behavior with Prefixes

Prefix Type Effect on Streaming SIMD Extensions
Address Size Prefix (67H) Affects Streaming SIMD Extensions with memory operand.
Ignored by Streaming SIMD Extensions without memory operand.
Operand Size (66H) Reserved and may result in unpredictable behavior.
Segment Override Affects Streaming SIMD Extensions with memory operand.
(2EH,36H,3EH,26H,64H,65H) Ignored by Streaming SIMD Extensions without memory operand.
Repeat Prefix (F3H) Affects Streaming SIMD Extensions.
Repeat NE Prefix(F2H) Reserved and may result in unpredictable behavior.
Lock Prefix (0F0H) Generates invalid opcode exception.

Table 9-5. SIMD Integer Instructions Behavior with Prefixes

Prefix Type Effect on MMX™ Instructions
Address Size Prefix (67H) Affects MMX™ instructions with mem. operand.
Ignored by MMX™ instructions without mem. operand.
Operand Size (66H) Reserved and may result in unpredictable behavior.
Segment Override Affects MMX™ instructions with mem. operand.
(2EH,36H,3EH,26H,64H,65H) Ignored by MMX™ instructions without mem operand.
Repeat Prefix (F3H) Reserved and may result in unpredictable behavior.
Repeat NE Prefix(F2H) Reserved and may result in unpredictable behavior.
Lock Prefix (0F0H) Generates invalid opcode exception.

Table 9-6. Cacheability Control Instruction Behavior with Prefixes

Prefix Type Effect on Streaming SIMD Extensions
Address Size Prefix (67H) Affects cacheability control instruction with a mem. operand.
Ignored by cacheability control instruction w/o a mem. operand.
Operand Size (66H) Reserved and may result in unpredictable behavior.
Segment Override Affects cacheability control instructions with mem. operand.
(2EH,36H,3EH,26H,64H,65H) Ignored by cacheability control instruction without mem operand.
Repeat Prefix(F3H) Reserved and may result in unpredictable behavior.
Repeat NE Prefix(F2H) Reserved and may result in unpredictable behavior.
Lock Prefix (0F0H) Generates an invalid opcode exception for all cacheability
instructions.

9-20
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9.5. WRITING APPLICATIONS WITH STREAMING SIMD

EXTENSIONS CODE
The following sections give guidelines for writing applications code using the Streaming SIMD
Extensions.

9.5.1. Detecting Support for Streaming SIMD Extensions Using

the CPUID Instruction
Use the CPUID instruction to determine whether the processor supports the Streaming SIMD
Extensions set (refer to Section 3.2., “Instruction Reference” in Chapter 3, Instruction Set Refer-
ence of the Intel Architecture Software Developer’s Manual, Volume 2, for a detailed description
of the CPUID instruction). When support for the Streaming SIMD Extensions is detected by the
CPUID instruction, it is signaled by setting bit 25 (Streaming SIMD Extensions bit) in the
feature flags to 1. This only determines the Streaming SIMD Extensions are present. There are
other support considerations related to Streaming SIMD Extensions.
The Streaming SIMD Extensions extensions can be divided into four categories:
• Single-precision, packed/scalar floating-point
• Additional SIMD Integer Instructions
• State management (i.e., FXSAVE/FXRSTOR)
• Cacheability control, further subdivided as:
• streaming stores for both packed FP (MOVNTPS) and integer MMX™
(MASKMOVQ and MOVNTQ) instructions.
• PREFETCH and SFENCE, which are not constrained to work with any specific
data type.
In order for an application to use SIMD floating-point extensions, the following conditions must
exist, otherwise an invalid opcode exception (Int 6) is generated:
• CR0.EM(bit 2) = 0 (emulation disabled)
• CR4.OSFXSR(bit 9) = 1 (OS supports saving SIMD floating-point state during context
switches)
• CPUID.XMM(EDX bit 25) = 1 (processor supports Streaming SIMD Extensions)
To verify support for the additional SIMD Integer instructions, including the corresponding
cacheability control instructions, the application needs only to check that CPUID.XMM is set
to 1. The SIMD integer instructions behave otherwise identically to the original MMX™
instructions; this implies that they will generate an invalid opcode exception if CR0.EM is set,
but will not generate an exception if CR4.OSFXSR is disabled/cleared.
To verify support for the PREFETCH and SFENCE instructions, the application needs only to
check that CPUID.XMM is set to 1; these instructions are not affected by CR0.EM or
CR4.OSFXSR.

9-21
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

For full details on how to determine what support is present for the Streaming SIMD Extensions,
please refer to the Intel Processor Identification and the CPUID Instruction Application Note
(AP-485), order number 241618-008.

9.5.2. Interfacing with Streaming SIMD Extensions Procedures

and Functions
The Streaming SIMD Extensions allow direct access to all SIMD floating-point registers. All
existing interface conventions that apply to the use of other general registers (for example: EAX,
EBX) will also apply to SIMD floating-point register usage.
An efficient interface to the Streaming SIMD Extensions routines might pass parameters and
return values through the SIMD floating-point registers or through a combination of memory
locations (view the stack) and SIMD floating-point registers. The three common IA-32 calling
conventions (cdecl, stdcall, and fastcall) have been extended to support the new register set for
Streaming SIMD Extensions in the following ways:
• The first three __m128 parameters are passed in registers xmm0, xmm1, and xmm2 (“args
in registers”). Additional __m128 parameters are passed on the stack as usual.
• __m128 return values are passed in xmm0.
• Registers xmm0 through xmm7 are caller-save.
The caller must reserve the space in the argument block where the first three __m128 parameters
would normally appear. These locations are generally left empty by the caller, but can be used
by the callee as “homes” for the xmm0, xmm1, and xmm2 registers if necessary.
New versions of the stdarg.h and varargs.h headers are provided with the Intel C/C++ compiler.
These new implementations support variable argument lists containing __m128 data (i.e., where
padding may have been inserted as required for aligned parameters as described above). The
new convention requires that functions with variable argument lists be prototyped before calls
are made to them, and that, for this case only, the caller must fill the locations on the stack for
data in registers xmm0, xmm1, and xmm2. Callers to non-prototyped functions with variable
argument lists with __m128 data must pass parameters both on the stack and in registers.

9.5.3. Writing Code with MMX™, Floating-Point, and Streaming

SIMD Extensions
The SIMD floating-point registers are separate from the FP / MMX™ registers. An application
can use Streaming SIMD Extensions and MMX™ instructions or Streaming SIMD Extensions
and x87-FP instructions simultaneously, without any penalty. An application can use x87-FP for
operations that need double or extended precision arithmetic, or for accessing any of the x87-FP
trigonometric instructions.
The restrictions on the simultaneous use of x87-FP and MMX™ instructions continue to exist,
because they share the same architectural registers. The user still needs to perform an EMMS
instruction when switching from MMX™ code to x87-FP code. However, the EMMS instruc-

9-22
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

tion is not necessary when integrating a Streaming SIMD Extensions module with existing
MMX™ technology modules or existing x87-FP modules. Streaming SIMD Extensions also do
not affect the floating-point tag word (FTW), floating-point control word (FCW), floating-point
status word (FSW) or floating-point exception state (FIP, FOP, FCS, FDS and FDP).
The SIMD integer instructions that are included in Streaming SIMD Extensions behave identi-
cally to original MMX™ instructions, in the presence of x87-FP instructions; this includes:
• Transition from x87-FP to MMX™ technology (TOS=0, FP valid bits set to all valid).
• MMX™ instructions write ones (1s) to the exponent part of the corresponding x87-FP
register.
• Use of EMMS for transition from MMX™ technology to x87-FP.
The Streaming SIMD Extensions that follow this behavior are: CVTPI2PS, CVTPS2PI,
CVTTPS2PI, MASKMOVQ, MOVNTQ, PEXTRW, PINSRW, PMOVMSKB, PMULHUW,
PSHUFW.

9.5.3.1. CACHEABILITY HINT INSTRUCTIONS

The Pentium® III processor’s cacheability control instructions enable the programmer to control
caching and prefetching of data. When correctly used, these instructions can significantly
improve application performance.
The PREFETCH instruction can minimize the latency of data access in performance-critical
sections of application code by allowing data to be fetched in advance of actual usage. The
instruction fetches 32 aligned bytes (or more, depending on the implementation) containing the
addressed byte, to a location in the processor cache hierarchy as specified by the temporal
locality hint (Table 9-7). In this table, cache level 0 is closest to the processor and cache level 2
is farthest from the processor. The hints specify fetch of either temporal or non-temporal data.
Subsequent accesses to temporal data are treated like normal accesses, while those to non-
temporal data will continue to minimize cache pollution. If the data is already present in a level
of the cache hierarchy that is closer to the processor, the PREFETCH instruction will not result
in any data movement.

9-23
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

Table 9-7. Cache Hints

HINTS ACTIONS
T0 Temporal data - fetch data into all levels of cache hierarchy
(L1 or L2 on Pentium® III)
T1 Temporal data - fetch data into level 2 cache and higher
(L2 on Pentium® III)
T2 Temporal data - fetch data into level 2 cache and higher
(L2 on Pentium® III)
NTA Non-temporal data - fetch data into location close to the processor, minimizing cache
pollution (for level 1 cache)
(L1 on Pentium® III)

The PREFETCH instruction does not change the user-visible semantics of a program, although
it may affect the performance of a program. The operation of this instruction is implementation-
dependent and can be overloaded to a subset of the hints (for example, T0, T1, and T2 may have
the same behavior) or altogether ignored by an implementation. The programmer will have to
tune his application for each implementation to take advantage of these instructions. These
instructions do not generate exceptions or faults. Excessive usage of prefetch instructions may
be throttled by the processor. For more detailed information on prefetch hints, refer to Chapter
6, “Optimizing Cache Utilization for Pentium® III Processors”, in the Intel Architecture Opti-
mization Reference Manual (Order Number 245127-001).
Some common usage models that may be affected in this way by weakly-ordered stores are:
• library functions, which use weakly-ordered memory to write results
• compiler-generated code, which also benefit from writing weakly-ordered results
• hand-crafted code
The degree to which a consumer of data knows that the data is weakly-ordered can vary for these
cases. As a result, the SFENCE instruction should be used to ensure ordering between routines
that produce weakly-ordered data and routines that consume this data. The SFENCE instruction
provides a performance-efficient way to ensure ordering, by guaranteeing that every store
instruction that precedes the store fence instruction in program order is globally visible before
any store instruction that follows the fence.

9-24
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9.5.3.2. RECOMMENDATIONS AND GUIDELINES

For more specific information relating to these recommendations and guidelines, such as port
assignments, prefetch instruction details, etc, refer to the Intel Architecture Optimization Refer-
ence Manual (Order Number 245127-001).
• Balance the limitations of the architecture.
a. Schedule instructions to resolve dependencies.
b. Intermix SIMD floating-point operations that utilize port 0 and port 1.
c. Do not issue consecutive instructions that utilize the same port.
• Use the reciprocal instructions followed by iteration for increased accuracy. These instruc-
tions yield reduced accuracy but execute much faster. If reduced accuracy is acceptable,
use them with no iteration. If near full accuracy is needed, use a Newton-Raphson
iteration. If full accuracy is needed, divide and square root provides this but slows down
performance.
• Exceptions
a. Mask exceptions to achieve higher performance. Unmasked exceptions may cause a
reduction in the retirement rate.
b. Utilize the Flush-to-Zero mode for higher performance to avoid the penalty of dealing
with denormals and underflows.
• Incorporate the prefetch instruction whenever possible.
• Try to emulate conditional moves by masked compares and logicals instead of using
conditional jumps.
• Utilize MMX™ instructions if the computations can be done in SIMD-integer, or for
shuffling data or copying data that is not used later in SIMD floating-point computations.
• If the algorithm requires extended precision, conversion to SIMD floating-point code is not
advised, because the SIMD floating-point instructions are single-precision.

9.5.4. Using Streaming SIMD Extensions Code in a Multitasking

Operating System Environment
An application needs to identify the nature of the multitasking operating system on which it
runs. Each task retains its own state that must be saved when a task switch occurs. The processor
state (context) consists of the integer registers, floating-point unit registers, and SIMD floating-
point registers. The STMXCSR and FXSAVE instructions store SIMD floating-point state in
memory for use by exception handlers and other system and application software. The
STMXCSR instruction saves the contents of the SIMD floating-point control/status register. The
FXSAVE instruction saves the x87-FP state (status, control, tag, instruction pointer, data pointer,
opcode and stack registers) and SIMD floating-point state (status/control, tag and data registers).
An application needs to verify that the processor supports FXSAVE prior to using this instruc-
tion. For a processor that implements FXSAVE but not Streaming SIMD Extensions, this can be

9-25
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

done by checking the CPUID.FXSR bit; for a processor that does implement Streaming SIMD
Extensions, use the approach described in Section 9.5.1., “Detecting Support for Streaming
SIMD Extensions Using the CPUID Instruction”. For even more detailed information, refer to
the Intel Processor Identification and the CPUID Instruction Application Note (AP-485), order
number 241618-008 and Identifying Support for Streaming SIMD Extensions in the Processor
and Operating System (AP-900).
The operating systems can be classified into two types:
• Cooperative multitasking operating systems
• Preemptive multitasking operating systems

9.5.4.1. COOPERATIVE MULTITASKING OPERATING SYSTEM

This type of multitasking operating system does not save the FP and MMX™ state and SIMD
floating-point state when performing a context switch. Therefore, the application needs to save
the relevant state before relinquishing direct or indirect control to the operating system.

9.5.4.2. PREEMPTIVE MULTITASKING OPERATING SYSTEM

This type of multitasking operating system saves the FP and MMX™ state and SIMD floating-
point state when performing a context switch. Therefore, the application does not have to save
or restore SIMD floating-point state.

9.5.5. Exception Handling in Streaming SIMD Extensions

Streaming SIMD Extensions can generate two kinds of exceptions:
• Non-numeric exceptions
• Numeric exceptions
Streaming SIMD Extensions can generate the same type of memory access exceptions as the IA
instructions do. Some examples are: page fault, segment not present, and limit violations.
Existing exception handlers can handle these types of exceptions without any code modifica-
tion. The Streaming SIMD Extensions PREFETCH instruction hints will not generate any kind
of exception and instead will be ignored.
Streaming SIMD Extensions can generate the same six numeric exceptions that x87-FP instruc-
tions can generate. All SIMD floating-point numeric exceptions are reported independently of
x87-FP numeric exceptions. Independent masking and unmasking of SIMD floating-point
numeric exceptions is achieved by setting/resetting specific bits in the MXCSR register.
The application must ensure that the OS can support unmasked SIMD floating-point exceptions
before unmasking them. (Use the approach described in Section 9.5.1., “Detecting Support for
Streaming SIMD Extensions Using the CPUID Instruction”. For even more detailed informa-
tion, refer to the Intel Processor Identification and the CPUID Instruction Application Note
(AP-485), order number 241618-008 and Identifying Support for Streaming SIMD Extensions

9-26
 PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

in the Processor and Operating System (AP-900).) If an application unmasks exceptions using
either FXRSTOR or LDMXCSR without the required OS support being enabled, an invalid
opcode fault, instead of a SIMD floating-point exception, will be generated on the first faulting
Streaming SIMD Extensions.
SIMD floating-point numeric exceptions are precise and occur as soon as the instruction
completes execution. They will not catch pending x87 floating-point exceptions and will not
cause assertion of FERR# (independent of the value of CR0.NE). In addition, they ignore the
assertion/de-assertion of IGNNE#.
For more details on SIMD floating-point exceptions and exception handlers, refer to Section
4.4., “Interrupts and Exceptions”, in Chapter 4, Procedure Calls, Interrupts, and Exceptions,
Appendix D, SIMD Floating-Point Exceptions Summary, and Appendix E, Guidelines for
Writing FPU Exceptions Handlers.

9-27
PROGRAMMING WITH THE STREAMING SIMD EXTENSIONS

9-28
 10
Input/Output
 CHAPTER 10
INPUT/OUTPUT

In addition to transferring data to and from external memory, Intel Architecture (IA) processors
can also transfer data to and from input/output ports (I/O ports). I/O ports are created in system
hardware by circuity that decodes the control, data, and address pins on the processor. These I/O
ports are then configured to communicate with peripheral devices. An I/O port can be an input
port, an output port, or a bidirectional port. Some I/O ports are used for transmitting data, such
as to and from the transmit and receive registers, respectively, of a serial interface device. Other
I/O ports are used to control peripheral devices, such as the control registers of a disk controller.
This chapter describes the processor’s I/O architecture. The topics discussed include:
• I/O port addressing.
• I/O instructions.
• I/O protection mechanism.

10.1. I/O PORT ADDRESSING

The processor allows I/O ports to be accessed in either of two ways:
• Through a separate I/O address space.
• Through memory-mapped I/O.
Accessing I/O ports through the I/O address space is handled through a set of I/O instructions
and a special I/O protection mechanism. Accessing I/O ports through memory-mapped I/O is
handled with the processors general-purpose move and string instructions, with protection
provided through segmentation or paging. I/O ports can be mapped so that they appear in the
I/O address space or the physical-memory address space (memory mapped I/O) or both.
One benefit of using the I/O address space is that writes to I/O ports are guaranteed to be
completed before the next instruction in the instruction stream is executed. Thus, I/O writes to
control system hardware cause the hardware to be set to its new state before any other instruc-
tions are executed. Refer to Section 10.6. for more information on serializing of I/O operations.

10.2. I/O PORT HARDWARE

From a hardware point of view, I/O addressing is handled through the processor’s address lines.
For Pentium® Pro, Pentium II, and Pentium® III processors, a special memory-I/O transaction
on the system bus indicates whether the address lines are being driven with a memory address
or an I/O address; for Pentium and earlier IA processors, the M/IO pin indicates a memory
address (1) or an I/O address (0). When the separate I/O address space is selected, it is the
responsibility of the hardware to decode the memory-I/O bus transaction to select I/O ports
rather than memory.
Data is transmitted between the processor and an I/O device through the data lines.

10-1
INPUT/OUTPUT

10.3. I/O ADDRESS SPACE

The processor’s I/O address space is separate and distinct from the physical-memory address
space. The I/O address space consists of 216 (64K) individually addressable 8-bit I/O ports,
numbered 0 through FFFFH. I/O port addresses 0F8H through 0FFH are reserved. Do not assign
I/O ports to these addresses. The result of an attempt to address beyond the I/O address space
limit of FFFFH is implementation-specific; refer to the Developer’s Manuals for specific
processors for more details.
Any two consecutive 8-bit ports can be treated as a 16-bit port, and any four consecutive ports
can be a 32-bit port. In this manner, the processor can transfer 8, 16, or 32 bits to or from a device
in the I/O address space. Like words in memory, 16-bit ports should be aligned to even addresses
(0, 2, 4, ...) so that all 16 bits can be transferred in a single bus cycle. Likewise, 32-bit ports
should be aligned to addresses that are multiples of four (0, 4, 8, ...). The processor supports data
transfers to unaligned ports, but there is a performance penalty because one or more extra bus
cycle must be used.
The exact order of bus cycles used to access unaligned ports is undefined and is not guaranteed
to remain the same in future IA processors. If hardware or software requires that I/O ports be
written to in a particular order, that order must be specified explicitly. For example, to load a
word-length I/O port at address 2H and then another word port at 4H, two word-length writes
must be used, rather than a single doubleword write at 2H.
Note that the processor does not mask parity errors for bus cycles to the I/O address space.
Accessing I/O ports through the I/O address space is thus a possible source of parity errors.

10.3.1. Memory-Mapped I/O

I/O devices that respond like memory components can be accessed through the processor’s
physical-memory address space (refer to Figure 10-1). When using memory-mapped I/O, any
of the processor’s instructions that reference memory can be used to access an I/O port located
at a physical-memory address. For example, the MOV instruction can transfer data between any
register and a memory-mapped I/O port. The AND, OR, and TEST instructions may be used to
manipulate bits in the control and status registers of a memory-mapped peripheral devices.
When using memory-mapped I/O, caching of the address space mapped for I/O operations must
be prevented. With the Pentium® Pro, Pentium II, and Pentium® III processors, caching of I/O
accesses can be prevented by using memory type range registers (MTRRs) to map the address
space used for the memory-mapped I/O as uncacheable (UC). Refer to Chapter 9, Memory
Cache Control, in the Intel Architecture Software Developer’s Manual, Volume 3, for a complete
discussion of the MTRRs.
The Pentium and Intel486™ processors do not support MTRRs. Instead, they provide the KEN#
pin, which when held inactive (high) prevents caching of all addresses sent out on the system
bus. To use this pin, external address decoding logic is required to block caching in specific
address spaces.

10-2
 INPUT/OUTPUT

Physical Memory
FFFF FFFFH

EPROM

I/O Port
I/O Port
I/O Port

RAM

Figure 10-1. Memory-Mapped I/O

All the IA processors that have on-chip caches also provide the PCD (page-level cache disable)
flag in page table and page directory entries. This flag allows caching to be disabled on a page-
by-page basis. Refer to Chapter 3.6.4., Page-Directory and Page-Table Entries in Chapter 3,
Protected-Mode Memory Management, in the Intel Architecture Software Developer’s Manual,
Volume 3.

10.4. I/O INSTRUCTIONS

The processor’s I/O instructions provide access to I/O ports through the I/O address space.
(These instructions cannot be used to access memory-mapped I/O ports.) There are two groups
of I/O instructions:
• Those which transfer a single item (byte, word, or doubleword) between an I/O port and a
general-purpose register.
• Those which transfer strings of items (strings of bytes, words, or doublewords) between an
I/O port and memory.
The register I/O instructions IN (input from I/O port) and OUT (output to I/O port) move data
between I/O ports and the EAX register (32-bit I/O), the AX register (16-bit I/O), or the AL
(8-bit I/O) register. The address of the I/O port can be given with an immediate value or a value
in the DX register.
The string I/O instructions INS (input string from I/O port) and OUTS (output string to I/O port)
move data between an I/O port and a memory location. The address of the I/O port being
accesses is given in the DX register; the source or destination memory address is given in the
DS:ESI or ES:EDI register, respectively.

10-3
INPUT/OUTPUT

When used with one of the repeat prefixes (such as REP), the INS and OUTS instructions
perform string (or block) input or output operations. The repeat prefix REP modifies the INS
and OUTS instructions to transfer blocks of data between an I/O port and memory. Here, the ESI
or EDI register is incremented or decremented (according to the setting of the DF flag in the
EFLAGS register) after each byte, word, or doubleword is transferred between the selected I/O
port and memory.
Refer to the individual references for the IN, INS, OUT, and OUTS instructions in Chapter 3,
Instruction Set Reference, of the Intel Architecture Software Developer’s Manual, Volume 2, for
more information on these instructions.

10.5. PROTECTED-MODE I/O

When the processor is running in protected mode, the following protection mechanisms regulate
access to I/O ports:
• When accessing I/O ports through the I/O address space, two protection devices control
access:
— The I/O privilege level (IOPL) field in the EFLAGS register.
— The I/O permission bit map of a task state segment (TSS).
• When accessing memory-mapped I/O ports, the normal segmentation and paging
protection and the MTRRs (in processors that support them) also affect access to I/O ports.
Refer to Chapter 4, Protection, and Chapter 9, Memory Cache Control, in the Intel Archi-
tecture Software Developer’s Manual, Volume 3, for a complete discussion of memory
protection.
The following sections describe the protection mechanisms available when accessing I/O ports
in the I/O address space with the I/O instructions.

10.5.1. I/O Privilege Level

In systems where I/O protection is used, the IOPL field in the EFLAGS register controls access
to the I/O address space by restricting use of selected instructions. This protection mechanism
permits the operating system or executive to set the privilege level needed to perform I/O. In a
typical protection ring model, access to the I/O address space is restricted to privilege levels 0
and 1. Here, kernel and the device drivers are allowed to perform I/O, while less privileged
device drivers and application programs are denied access to the I/O address space. Application
programs must then make calls to the operating system to perform I/O.
The following instructions can be executed only if the current privilege level (CPL) of the
program or task currently executing is less than or equal to the IOPL: IN, INS, OUT, OUTS, CLI
(clear interrupt-enable flag), and STI (set interrupt-enable flag). These instructions are called
I/O sensitive instructions, because they are sensitive to the IOPL field. Any attempt by a less
privileged program or task to use an I/O sensitive instruction results in a general-protection

10-4
 INPUT/OUTPUT

exception (#GP) being signaled. Because each task has its own copy of the EFLAGS register,
each task can have a different IOPL.
The I/O permission bit map in the TSS can be used to modify the effect of the IOPL on I/O sensi-
tive instructions, allowing access to some I/O ports by less privileged programs or tasks (refer
to Section 10.5.2.).
A program or task can change its IOPL only with the POPF and IRET instructions; however,
such changes are privileged. No procedure may change the current IOPL unless it is running at
privilege level 0. An attempt by a less privileged procedure to change the IOPL does not result
in an exception; the IOPL simply remains unchanged.
The POPF instruction also may be used to change the state of the IF flag (as can the CLI and
STI instructions); however, the POPF instruction in this case is also I/O sensitive. A procedure
may use the POPF instruction to change the setting of the IF flag only if the CPL is less than or
equal to the current IOPL. An attempt by a less privileged procedure to change the IF flag does
not result in an exception; the IF flag simply remains unchanged.

10.5.2. I/O Permission Bit Map

The I/O permission bit map is a device for permitting limited access to I/O ports by less privi-
leged programs or tasks and for tasks operating in virtual-8086 mode. The I/O permission bit
map is located in the TSS (refer to Figure 10-2) for the currently running task or program. The
address of the first byte of the I/O permission bit map is given in the I/O map base address field
of the TSS. The size of the I/O permission bit map and its location in the TSS are variable.

Task State Segment (TSS)

31 24 23 0
Last byte of bit 1 1 1 1 1 1 1 1
map must be
followed by a byte
with all bits set
I/O Permission Bit Map

I/O Map Base 64H

I/O base map must
not exceed DFFFH.

Figure 10-2. I/O Permission Bit Map

Because each task has its own TSS, each task has its own I/O permission bit map. Access to indi-
vidual I/O ports can thus be granted to individual tasks.

10-5
INPUT/OUTPUT

If in protected mode and the CPL is less than or equal to the current IOPL, the processor allows
all I/O operations to proceed. If the CPL is greater than the IOPL or if the processor is operating
in virtual-8086 mode, the processor checks the I/O permission bit map to determine if access to
a particular I/O port is allowed. Each bit in the map corresponds to an I/O port byte address. For
example, the control bit for I/O port address 29H in the I/O address space is found at bit position
1 of the sixth byte in the bit map. Before granting I/O access, the processor tests all the bits corre-
sponding to the I/O port being addressed. For a doubleword access, for example, the processors
tests the four bits corresponding to the four adjacent 8-bit port addresses. If any tested bit is set,
a general-protection exception (#GP) is signaled. If all tested bits are clear, the I/O operation is
allows to proceed.
Because I/O port addresses are not necessarily aligned to word and doubleword boundaries, the
processor reads two bytes from the I/O permission bit map for every access to an I/O port. To
prevent exceptions from being generated when the ports with the highest addresses are accessed,
an extra byte needs to included in the TSS immediately after the table. This byte must have all
of its bits set, and it must be within the segment limit.
It is not necessary for the I/O permission bit map to represent all the I/O addresses. I/O addresses
not spanned by the map are treated as if they had set bits in the map. For example, if the TSS
segment limit is 10 bytes past the bit-map base address, the map has 11 bytes and the first 80 I/O
ports are mapped. Higher addresses in the I/O address space generate exceptions.
If the I/O bit map base address is greater than or equal to the TSS segment limit, there is no I/O
permission map, and all I/O instructions generate exceptions when the CPL is greater than the
current IOPL. The I/O bit map base address must be less than or equal to DFFFH.

10.6. ORDERING I/O

When controlling I/O devices it is often important that memory and I/O operations be carried
out in precisely the order programmed. For example, a program may write a command to an I/O
port, then read the status of the I/O device from another I/O port. It is important that the status
returned be the status of the device after it receives the command, not before.
When using memory-mapped I/O, caution should be taken to avoid situations in which the
programmed order is not preserved by the processor. To optimize performance, the processor
allows cacheable memory reads to be reordered ahead of buffered writes in most situations.
Internally, processor reads (cache hits) can be reordered around buffered writes. When using
memory-mapped I/O, therefore, is possible that an I/O read might be performed before the
memory write of a previous instruction. The recommended method of enforcing program
ordering of memory-mapped I/O accesses with the Pentium® Pro, Pentium II, and Pentium® III
processors is to use the MTRRs to make the memory mapped I/O address space uncacheable;
for the Pentium and Intel486™ processors, either the #KEN pin or the PCD flags can be used
for this purpose (refer to Section 10.3.1.). When the target of a read or write is in an uncacheable
region of memory, memory reordering does not occur externally at the processor’s pins (that is,
reads and writes appear in-order). Designating a memory mapped I/O region of the address
space as uncacheable insures that reads and writes of I/O devices are carried out in program

10-6
 INPUT/OUTPUT

order. Refer to Chapter 9, Memory Cache Control, in the Intel Architecture Software Devel-
oper’s Manual, Volume 3, for more information on using MTRRs.
Another method of enforcing program order is to insert one of the serializing instructions, such
as the CPUID instruction, between operations. Refer to Chapter 7, Multiple-Processor Manage-
ment, in the Intel Architecture Software Developer’s Manual, Volume 3, for more information on
serialization of instructions.
It should be noted that the chip set being used to support the processor (bus controller, memory
controller, and/or I/O controller) may post writes to uncacheable memory which can lead to out-
of-order execution of memory accesses. In situations where out-of-order processing of memory
accesses by the chip set can potentially cause faulty memory-mapped I/O processing, code must
be written to force synchronization and ordering of I/O operations. Serializing instructions can
often be used for this purpose.
When the I/O address space is used instead of memory-mapped I/O, the situation is different in
two respects:
• The processor never buffers I/O writes. Therefore, strict ordering of I/O operations is
enforced by the processor. (As with memory-mapped I/O, it is possible for a chip set to
post writes in certain I/O ranges.)
• The processor synchronizes I/O instruction execution with external bus activity (refer to
Table 10-1).

Table 10-1. I/O Instruction Serialization

Processor Delays Execution of … Until Completion of …

Instruction Being Current

Executed Instruction? Next Instruction? Pending Stores? Current Store?

IN Yes Yes
INS Yes Yes
REP INS Yes Yes
OUT Yes Yes Yes
OUTS Yes Yes Yes
REP OUTS Yes Yes Yes

10-7
 11
Processor
Identification
and Feature
Determination
 CHAPTER 11
PROCESSOR IDENTIFICATION AND FEATURE
DETERMINATION

When writing software intended to run on several different types of Intel Architecture (IA)
processors, it is generally necessary to identify the type of processor present in a system and the
processor features that are available to an application. This chapter describes how to identify the
processor that is executing the code and determine the features the processor supports. It also
shows how to determine if an FPU or NPX is present. For more information about processor
identification and supported features, refer to the following documents:
• AP-485, Intel Processor Identification and the CPUID Instruction
• For a complete list of the features that are available for the different IA processors, refer to
Chapter 18, Intel Architecture Compatibility of the Intel Architecture Software Developer’s
Manual, Volume 3: System Programming Guide.

11-1
PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

11.1. PROCESSOR IDENTIFICATION

The CPUID instruction returns the processor type for the processor that executes the instruction.
It also indicates the features that are present in the processor, including the existence of an
on-chip FPU. The following information can be obtained with this instruction:
• The highest operand value the instruction responds to (2 for the Pentium® Pro processors
and 1 for the Pentium® processors and recent Intel486™ processors).
• The processor’s family identification (ID) number, model ID, and stepping ID.
• The presence of an on-chip FPU.
• Support for or the presence of the following architectural extensions and enhancements:
— Virtual-8086 mode enhancements.
— Debugging extensions.
— Page-size extensions.
— Read time stamp counter (RDTSC) instruction.
— Read model specific registers (RDMSR) and write model specific registers (WRMSR)
instructions.
— Physical address extension.
— Machine check exceptions.
— Compare and exchange 8 bytes instruction (CMPXCHG8B).
— On-chip, advanced programmable interrupt controller (APIC).
— Memory-type range registers (MTRRs).
— Page global flag.
— Machine check architecture.
— Conditional move instruction (CMOVcc).
— MMX™ technology.
• Cache and TLB information.
To use this instruction, a source operand value of 0, 1, or 2 is placed in the EAX register.
Processor identification and feature information is then returned in the EAX, EBX, ECX, and
EDX registers. Refer to Section 3.2., “Instruction Reference” in Chapter 3, Instruction Set
Reference of the Intel Architecture Software Developer’s Manual, Volume 2, for more detailed
information about the instruction.

11-2
 PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

AP-485, Intel Processor Identification and the CPUID Instruction (Order Number 241618),
provides additional information and example source code for use in identifying IA processors.
It also contains guidelines for using the CPUID instruction to help maintain the widest range of
software compatibility. The following guidelines are among the most important, and should
always be followed when using the CPUID instruction to determine available features:
• Always begin by testing for the “GenuineIntel,” message in the EBX, EDX, and ECX
registers when the CPUID instruction is executed with EAX equal to 0. If the processor is
not genuine Intel, the feature identification flags may have different meanings than are
described in “CPUID—CPU Identification” in Chapter 3, Instruction Set Reference of the
Intel Architecture Software Developer’s Manual, Volume 2.
• Do not assume a value of 1 in a feature identification flag indicates that a given feature is
present. For future feature identification flags, a value of 1 may indicate that the specific
feature is not present.
• Test feature identification flags individually and do not make assumptions about undefined
bits.
Note that the CPUID instruction will cause the invalid opcode exception (#UD) if executed on
a processor that does not support it. The CPUID instruction application note provides a code
sequence to test the validity of the CPUID instruction. Also, this test code (for CPUID valid) is
not reliable when executed in virtual-8086 mode. To avoid this, if the test code is written to run
in real-address mode, the SMSW instruction must be used to read the PE bit from the MSW
(lower half of CR0). If PE flag is set to 1, the Real Mode code is actually being executed in
virtual-8086 mode, and the test sequence cannot be guaranteed to return reliable information.
(Note that the new version of the CPUID application note (AP-485, Intel Processor Identifica-
tion and the CPUID Instruction (Order Number 241618-005)), explains this virtual-8086
problem, but the older versions of the application note do not.)

11-3
11.2. IDENTIFICATION OF EARLIER INTEL ARCHITECTURE
PROCESSORS
The CPUID instruction is only available in the Pentium® Pro, Pentium®, and recent Intel486™
processors. For the earlier IA processors (including the earlier Intel486™ processors), several
other architectural features can be exploited to identify the processor.
The settings of bits 12 and 13 (IOPL), 14 (NT), and 15 (reserved) in the EFLAGS register (refer
to Figure 3-7, Section 3.6.3., “EFLAGS Register”, in Chapter 3, Basic Execution Environment)
is different for Intel’s 32-bit processors than for the Intel 8086 and Intel 286 processors. By
examining the settings of these bits (with the PUSHF/PUSHFD and POP/POPFD instructions),
an application program can determine whether the processor is an 8086, Intel286, or one of the
Intel 32-bit processors:
• 8086 processor — Bits 12 through 15 of the EFLAGS register are always set.
• Intel 286 processor — Bits 12 through 15 are always clear in real-address mode.
• 32-bit processors — In real-address mode, bit 15 is always clear and bits 12 through 14
have the last value loaded into them. In protected mode, bit 15 is always clear, bit 14 has
the last value loaded into it, and the IOPL bits depends on the current privilege level
(CPL). The IOPL field can be changed only if the CPL is 0.
Other EFLAG register bits that can be used to differentiate between the 32-bit processors:
• Bit 18 (AC) — Implemented only on the Pentium® Pro, Pentium®, and Intel486™
processors. The inability to set or clear this bit distinguishes an Intel386™ processor from
the other Intel 32-bit processors.
• Bit 21 (ID) — Determines if the processor is able to execute the CPUID instruction. The
ability to set and clear this bit indicates that the processor is a Pentium ® Pro, Pentium®, or
later version Intel486™ processor.
To determine whether an FPU or NPX is present in a system, applications can write to the
FPU/NPX status and control registers using the FNINIT instruction and then verify the correct
values are read back using the FNSTENV instruction.
After determining that an FPU or NPX is present, its type can then be determined. In most cases,
the processor type will determine the type of FPU or NPX; however, an Intel386™ processor is
compatible with either an Intel 287 or Intel 387 math coprocessor. The method the coprocessor
uses to represent ∞ (after the execution of the FINIT, FNINIT, or RESET instruction) indicates
which coprocessor is present. The Intel 287 math coprocessor uses the same bit representation
for +∞ and −∞; whereas, the Intel 387 math coprocessor uses different representations for +∞
and −∞.
 PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

11.3. CPUID INSTRUCTION EXTENSIONS

The CPUID instructions of all P6-family processors behave identically. The CPUID instruction
is described in detail in the application note, AP-485, Intel Processor Identification and the
CPUID Instruction. This section describes processor-specific information returned by the
CPUID instruction.
The CPUID instruction’s behavior varies depending upon the contents of the EAX register when
the instruction is executed. Table 11-1 shows the interaction between the value in EAX before
the call to CPUID and the value that CPUID returns.

Table 11-1. EAX Input Value and CPUID Return Values

EAX CPUID Return Values
0 EAX Maximum CPUID input value

EBX 756E6547H ‘uneG’ (G in BL)

ECX 6C65746EH ‘letn’ (n in CL)

EDX 49656E69H ‘leni’ (i in DL)

1 EAX Version information (Type, Family, Model, Stepping)

EBX Reserved

ECX Reserved

EDX Feature Information

2 EAX Cache Information

EBX Cache Information

ECX Cache Information

EDX Cache Information

Refer to the CPUID application note, AP-485, for details on cache information. AP-485 is avail-
able from the following web site: http://developer.intel.com/design/pro/applnots/ap485.htm.
In addition, the following two new cache descriptors are defined for P6-family processors with
Model > 3:
1M L2 Cache 4-way set associative 32-byte line size 44h
2M L2 Cache 4-way set associative 32-byte line size 45h

11.3.1. Version Information

When the CPUID instruction is executed with a 1 in EAX, it returns version and feature infor-
mation. Figure 11-1 shows the version information bit fields returned by CPUID in EAX. The
233, 266, and 300 MHz Pentium® II processors are indicated by a “6” in the Family ID and a

11-5
PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

“3” in the Model ID field. Future P6-family processors are indicated by a “6” in the Family ID
and a value greater than “3” in the Model ID field.

31 12 11 08 07 04 03 00
Reserved (0) Family ID Model ID Stepping ID

Figure 11-1. EAX Return Values

Figure 11-2 shows the feature information bit fields returned by CPUID in EAX.

3 26 25 24 23 22 18 17 16 15 14 13 12 11 10 09 08 07 06 05 04 03 02 01 00
-
19
XMM

FXSR

MMX

rsvd

PN

PSE-36

PAT

CMOV

MCA

PGE

MTRR

SEP

rsvd

APIC

CX8

MCE

PAE

MSR

TSC

PSE

DE

VME

FPU
Reserved
(0)

Figure 11-2. CPUID Feature Field Information Bits

Table 11-2 describes the bit representations for the new P6-family processor features.

Table 11-2. New P6-Family Processor Feature Information Returned by CPUID in EDX
Bit Feature Value Description Notes
11 SEP 1 Fast System Call Indicates whether the processor supports the Fast
System Call instructions SYSENTER and SYSEXIT.
16 PAT 1 Page Attribute Indicates whether the processor supports the Page
Table Attribute Table. This feature augments the Memory
Type Range Registers (MTRRs), allowing an
operating system to specify attributes of memory on a
page granularity through a linear address.
17 PSE-36 1 36-bit Page Size Indicates whether the processor supports 4 MB pages
Extension that are capable of addressing physical memory
beyond 4 GB. This feature indicates that the up-per
four bits of the physical address of the 4-MB page is
encoded by bits 13-16 of the page directory entry.
18 PN 1 Processor Indicates whether the processor supports the 96-bit
Number Processor Number feature.
19-22 rsvd 0 Reserved These bits are reserved for future use. The contents of
these fields are not defined and should not be relied
upon or altered.
23 MMX 1 MMX-technology Indicates whether the processor supports the MMX
technology instruction set and architecture.
24 FXSR 1 Fast floating- Indicates whether the processor supports the
point save and FXSAVE and FXRSTOR instructions for fast save and
restore restore of the floating-point context. Presence of this
bit also indicates that CR4.OSFXSR is available,
allowing an operating system to indicate that it uses
the fast save/restore instructions.
25 XMM 1 Streaming SIMD Indicates whether the processor supports the
Extension Streaming SIMD Extensions instruction set.

11-6
 PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

11.3.2. Control Register Extensions

The control registers (CR0, CR1, CR2, CR3, and CR4) determine the operating mode of the
processor and the characteristics of the currently executing task. A new field has been added to
CR4, which contains a group of flags used to enable several architectural extensions as depicted
in Figure 11-3.

31 10 09 08 07 06 05 04 03 02 01 00
Reserved (set to 0) OSFXSR PCE PGE MCE PAE PSE DE TSD PVI VME

Figure 11-3. CR4 Register Extensions

The new field at bit 9 (OSFXSR) is set by the operating system to indicate that it uses the
FXSAVE/FXRSTOR instructions for saving/restoring FP/MMX™ state during context
switches. This bit defaults to clear (zero) at processor initialization.

11-7
PROCESSOR IDENTIFICATION AND FEATURE DETERMINATION

11-8
 A
EFLAGS
Cross-Reference
 APPENDIX A
EFLAGS CROSS-REFERENCE

The cross-reference in Table A-1 summarizes how the flags in the processor’s EFLAGS register
are affected by each instruction. For detailed information on how flags are affected, refer to
Chapter 3, Instruction Set Reference of the Intel Architecture Software Developer’s Manual, Vol-
ume 2. The following codes describe how the flags are affected:

T Instruction tests flag.

M Instruction modifies flag (either sets or resets depending on operands).
0 Instruction resets flag.
1 Instruction sets flag.
— Instruction’s effect on flag is undefined.
R Instruction restores prior value of flag.
Blank Instruction does not affect flag.

Table A-1. EFLAGS Cross-Reference

Instruction OF SF ZF AF PF CF TF IF DF NT RF
AAA — — — TM — M
AAD — M M — M —
AAM — M M — M —
AAS — — — TM — M
ADC M M M M M TM
ADD M M M M M M
AND 0 M M — M 0
ARPL M
BOUND
BSF/BSR — — M — — —
BSWAP
BT/BTS/BTR/BTC — — — — — M

A-1
EFLAGS CROSS-REFERENCE

Table A-1. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
CALL
CBW
CLC 0
CLD 0
CLI 0
CLTS
CMC M
CMOVcc T T T T T
CMP M M M M M M
CMPS M M M M M M T
CMPXCHG M M M M M M
CMPXCHG8B M
CPUID
COMISS 1 1 1 1 1 1
CWD
DAA — M M TM M TM
DAS — M M TM M TM
DEC M M M M M
DIV — — — — — —
ENTER
ESC
FCMOVcc T T T
FCOMI, FCOMIP, M M M
FUCOMI, FUCOMIP
HLT
IDIV — — — — — —
IMUL M — — — — M
IN
INC M M M M M
INS T
INT 0 0
INTO T 0 0
INVD

A-2
 EFLAGS CROSS-REFERENCE

Table A-1. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
INVLPG
IRET R R R R R R R R R T
Jcc T T T T T
JCXZ
JMP
LAHF
LAR M
LDS/LES/LSS/LFS/LGS
LEA
LEAVE
LGDT/LIDT/LLDT/LMSW
LOCK
LODS T
LOOP
LOOPE/LOOPNE T
LSL M
LTR
MOV
MOV control, debug, test — — — — — —
MOVS T
MOVSX/MOVZX
MUL M — — — — M
NEG M M M M M M
NOP
NOT
OR 0 M M — M 0
OUT
OUTS T
POP/POPA
POPF R R R R R R R R R R
PUSH/PUSHA/PUSHF
RCL/RCR 1 M TM
RCL/RCR count — TM

A-3
EFLAGS CROSS-REFERENCE

Table A-1. EFLAGS Cross-Reference (Contd.)

Instruction OF SF ZF AF PF CF TF IF DF NT RF
RDMSR
RDPMC
RDTSC
REP/REPE/REPNE
RET
ROL/ROR 1 M M
ROL/ROR count — M
RSM M M M M M M M M M M M
SAHF R R R R R
SAL/SAR/SHL/SHR 1 M M M — M M
SAL/SAR/SHL/SHR count — M M — M M
SBB M M M M M TM
SCAS M M M M M M T
SETcc T T T T T
SGDT/SIDT/SLDT/SMSW
SHLD/SHRD — M M — M M
STC 1
STD 1
STI 1
STOS T
STR
SUB M M M M M M
TEST 0 M M — M 0
UCOMISS 1 1 1 1 1 1
UD2
VERR/VERRW M
WAIT
WBINVD
WRMSR
XADD M M M M M M
XCHG
XLAT
XOR 0 M M — M 0

A-4
 B
EFLAGS
Condition Codes
 APPENDIX B
EFLAGS CONDITION CODES

Table B-1 gives all the condition codes that can be tested for by the CMOVcc, FCMOVcc, Jcc
and SETcc instructions. The condition codes refer to the setting of one or more status flags (CF,
OF, SF, ZF, and PF) in the EFLAGS register. The “Mnemonic” column gives the suffix (cc) add-
ed to the instruction to specific the test condition. The “Condition Tested For” column describes
the condition specified in the “Status Flags Setting” column. The “Instruction Subcode” column
gives the opcode suffix added to the main opcode to specify a test condition.

Table B-1. EFLAGS Condition Codes

Instruction
Mnemonic (cc) Condition Tested For Subcode Status Flags Setting

O Overflow 0000 OF = 1
NO No overflow 0001 OF = 0
B Below 0010 CF = 1
NAE Neither above nor equal
NB Not below 0011 CF = 0
AE Above or equal
E Equal 0100 ZF = 1
Z Zero
NE Not equal 0101 ZF = 0
NZ Not zero
BE Below or equal 0110 (CF OR ZF) = 1
NA Not above
NBE Neither below nor equal 0111 (CF OR ZF) = 0
A Above
S Sign 1000 SF = 1
NS No sign 1001 SF = 0
P Parity 1010 PF = 1
PE Parity even
NP No parity 1011 PF = 0
PO Parity odd
Instruction
Mnemonic Meaning Subcode Condition Tested
L Less 1100 (SF xOR OF) = 1
NGE Neither greater nor equal
NL Not less 1101 (SF xOR OF) = 0
GE Greater or equal

B-1
EFLAGS CONDITION CODES

Table B-1. EFLAGS Condition Codes (Contd.)

Instruction
Mnemonic (cc) Condition Tested For Subcode Status Flags Setting

LE Less or equal 1110 ((SF XOR OF) OR ZF) = 1

NG Not greater
NLE Neither less nor equal 1111 ((SF XOR OF) OR ZF) = 0
G Greater

Many of the test conditions are described in two different ways. For example, LE (less or equal)
and NG (not greater) describe the same test condition. Alternate mnemonics are provided to
make code more intelligible.
The terms “above” and “below” are associated with the CF flag and refer to the relation between
two unsigned integer values. The terms “greater” and “less” are associated with the SF and OF
flags and refer to the relation between two signed integer values.

B-2
 C
Floating-Point
Exceptions Summary
 APPENDIX C
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-1 lists the floating-point instruction mnemonics in alphabetical order. For each
mnemonic, it summarizes the exceptions that the instruction may cause. Refer to Section 7.8.,
“Floating-Point Exception Conditions” in Chapter 7, Floating-Point Unit for a detailed discus-
sion of the floating-point exceptions. The following codes indicate the floating-point excep-
tions:

#IS Invalid operation exception for stack underflow or stack overflow.

#IA Invalid operation exception for invalid arithmetic operands and
unsupported formats.
#D Denormal operand exception.
#Z Divide-by-zero exception.
#O Numeric overflow exception.
#U Numeric underflow exception.
#P Inexact result (precision) exception.

Table C-1. Floating-Point Exceptions Summary

Mnemonic Instruction #IS #IA #D #Z #O #U #P
X–1
F2XM1 2 Y Y Y Y Y
FABS Absolute value Y
FADD(P) Add real Y Y Y Y Y Y
FBLD BCD load Y
FBSTP BCD store and pop Y Y Y
FCHS Change sign Y
FCLEX Clear exceptions
FCMOVcc Floating-point conditional move Y
FCOM, FCOMP, FCOMPP Compare real Y Y Y
FCOMI, FCOMIP, FUCOMI, Compare real and set EFLAGS Y Y
FUCOMIP
FCOS Cosine Y Y Y Y Y
FDECSTP Decrement stack pointer
FDIV(R)(P) Divide real Y Y Y Y Y Y Y

C-1
FLOATING-POINT EXCEPTIONS SUMMARY

Table C-1. Floating-Point Exceptions Summary (Contd.)

Mnemonic Instruction #IS #IA #D #Z #O #U #P
FFREE Free register
FIADD Integer add Y Y Y Y Y Y
FICOM(P) Integer compare Y Y Y
FIDIV Integer divide Y Y Y Y Y Y
FIDIVR Integer divide reversed Y Y Y Y Y Y Y
FILD Integer load Y
FIMUL Integer multiply Y Y Y Y Y Y
FINCSTP Increment stack pointer
FINIT Initialize processor
FIST(P) Integer store Y Y Y
FISUB(R) Integer subtract Y Y Y Y Y Y
FLD extended or stack Load real Y
FLD single or double Load real Y Y Y
FLD1 Load + 1.0 Y
FLDCW Load Control word Y Y Y Y Y Y Y
FLDENV Load environment Y Y Y Y Y Y Y
FLDL2E Load log2e Y
FLDL2T Load log210 Y
FLDLG2 Load log102 Y
FLDLN2 Load loge2 Y
FLDPI Load π Y
FLDZ Load + 0.0 Y
FMUL(P) Multiply real Y Y Y Y Y Y
FNOP No operation
FPATAN Partial arctangent Y Y Y Y Y
FPREM Partial remainder Y Y Y Y
FPREM1 IEEE partial remainder Y Y Y Y
FPTAN Partial tangent Y Y Y Y Y
FRNDINT Round to integer Y Y Y Y
FRSTOR Restore state Y Y Y Y Y Y Y
FSAVE Save state
FSCALE Scale Y Y Y Y Y Y
FSIN Sine Y Y Y Y Y

C-2
 FLOATING-POINT EXCEPTIONS SUMMARY

Table C-1. Floating-Point Exceptions Summary (Contd.)

Mnemonic Instruction #IS #IA #D #Z #O #U #P
FSINCOS Sine and cosine Y Y Y Y Y
FSQRT Square root Y Y Y Y
FST(P) stack or extended Store real Y
FST(P) single or double Store real Y Y Y Y Y Y
FSTCW Store control word
FSTENV Store environment
FSTSW (AX) Store status word
FSUB(R)(P) Subtract real Y Y Y Y Y Y
FTST Test Y Y Y
FUCOM(P)(P) Unordered compare real Y Y Y
FWAIT CPU Wait
FXAM Examine
FXCH Exchange registers Y
FXTRACT Extract Y Y Y Y
FYL2X Y ⋅ log2X Y Y Y Y Y Y Y
FYL2XP1 Y ⋅ log2(X + 1) Y Y Y Y Y

C-3
 D
SIMD Floating-Point
Exceptions Summary
 APPENDIX D
SIMD FLOATING-POINT EXCEPTIONS SUMMARY

Table D-1 lists the Streaming SIMD Extensions mnemonics in alphabetical order. For each
mnemonic, it summarizes the exceptions that the instruction may cause. Refer to Section 9.5.5.,
“Exception Handling in Streaming SIMD Extensions” in Chapter 9, Programming with the
Streaming SIMD Extensions for a detailed discussion of the various exceptions that can occur
when executing Streaming SIMD Extensions.
The following codes indicate the exceptions associated with execution of an instruction that uti-
lizes the 128-bit Streaming SIMD Extensions registers.

#I Invalid operation exception for invalid arithmetic operands and

unsupported formats.
#D Denormal operand exception.
#Z Divide-by-zero exception.
#O Numeric overflow exception.
#U Numeric underflow exception.
#P Inexact result (precision) exception.

D-1
SIMD FLOATING-POINT EXCEPTIONS SUMMARY

Table D-1. Streaming SIMD Extensions Instruction Set Summary

Mnemonic Instruction #I # # # # #
D Z O U P

ADDPS Packed add Y Y Y Y Y

ADDSS Scalar add Y Y Y Y Y
ANDNPS Packed logical INVERT and
AND
ANDPS Packed logical AND
CMPPS Packed compare Y Y
CMPSS Scalar compare Y Y
COMISS Scalar ordered compare lower Y Y
SP FP numbers and set the
status flags
CVTPI2PS Convert two 32-bit signed Y
integers from MM2/Mem to two
SP FP.
CVTPS2PI Convert lower 2 SP FP from Y Y
XMM/Mem to 2 32-bit signed
integers in MM using rounding
specified by MXCSR.
CVTSI2SS Convert one 32-bit signed Y
integer from Integer Reg/Mem
to one SP FP.
CVTSS2SI Convert one SP FP from Y Y
XMM/Mem to one 32-bit signed
integer using rounding mode
specified by MXCSR, and move
the result to an integer register.
CVTTPS2PI Convert lower 2 SP FP from Y Y
XMM2/Mem to 2 32-bit signed
integers in MM1 using truncate.
CVTTSS2SI Convert lowest SP FP from Y Y
XMM/Mem to one 32-bit signed
integer using truncate, and
move the result to an integer
register.
DIVPS Packed divide Y Y Y Y Y Y
DIVSS Scalar divide Y Y Y Y Y Y
FXRSTOR Load FP and Streaming SIMD
Extensions state
FXSAVE Store FP and Streaming SIMD
Extensions state
LDMXCSR Load control/status word
MAXPS Packed maximum Y Y
MAXSS Scalar maximum Y Y
MINPS Packed minimum Y Y
MINSS Scalar minimum Y Y

D-2
 SIMD FLOATING-POINT EXCEPTIONS SUMMARY

Mnemonic Instruction #I # # # # #
D Z O U P

MOVAPS Move aligned packed data

MOVHPS Move high 64 bits

MOVLPS Move low 64 bits

MOVMSKPS Move mask to r32
MOVSS Move scalar
MOVUPS Move unaligned packed data
MULPS Packed multiply Y Y Y Y Y
MULSS Scalar multiply Y Y Y Y Y
ORPS Packed OR
RCPPS Packed reciprocal
RCPSS Scalar reciprocal
RSQRTPS Packed reciprocal square root
RSQRTSS Scalar reciprocal square root
SHUFPS Shuffle
SQRTPS Square Root of the packed SP Y Y Y
FP numbers
SQRTSS Scalar square root Y Y Y
STMXCSR Store control/status word
SUBPS Packed subtract Y Y Y Y Y
SUBSS Scalar subtract Y Y Y Y Y
UCOMISS Unordered compare lower SP Y Y
FP numbers and set the status
flags
UNPCKHPS Interleave SP FP numbers
UNPCKLPS Interleave SP FP numbers
XORPS Packed XOR

D-3
SIMD FLOATING-POINT EXCEPTIONS SUMMARY

D-4
 E
Guidelines for
Writing FPU
Exception Handlers
 APPENDIX E
GUIDELINES FOR WRITING FPU
EXCEPTIONS HANDLERS

As described in Chapter 7, Floating-Point Unit, the Intel Architecture (IA) supports two mech-
anisms for accessing exception handlers to handle unmasked FPU exceptions: native mode and
MS-DOS compatibility mode. The primary purpose of this appendix is to provide detailed in-
formation to help software engineers design and write FPU exception-handling facilities to run
on PC systems that use the MS-DOS compatibility modeI for handling FPU exceptions. Some
of the information in this appendix will also be of interest to engineers who are writing native-
mode FPU exception handlers. The information provided is as follows:

• Discussion of the origin of the MS-DOS* FPU exception handling mechanism and its
relationship to the FPU’s native exception handling mechanism.

• Description of the IA flags and processor pins that control the MS-DOS FPU exception
handling mechanism.

• Description of the external hardware typically required to support MS-DOS exception

handling mechanism.

• Description of the FPU’s exception handling mechanism and the typical protocol for FPU
exception handlers.

• Code examples that demonstrate various levels of FPU exception handlers.

• Discussion of FPU considerations in multitasking environments.

• Discussion of native mode FPU exception handling.

The information given is oriented toward the most recent generations of IA processors, starting
with the Intel486™. It is intended to augment the reference information given in Chapter 7,
Floating-Point Unit.
A more extensive version of this appendix is available in the application note AP-578, Software
and Hardware Considerations for FPU Exception Handlers for Intel Architecture Processors
(Order Number 242415-001), which is available from Intel.

NOTES
I Microsoft Windows* 95 and Windows* 3.1 (and earlier versions) operating systems use almost the same
FPU exception handling interface as the operating system. The recommendations in this appendix for a
MS-DOS* compatible exception handler thus apply to all three operating systems.

E-1
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

E.1. ORIGIN OF THE MS-DOS* COMPATIBILITY MODE FOR

HANDLING FPU EXCEPTIONS
The first generations of IA processors (starting with the Intel 8086 and 8088 processors and go-
ing through the Intel 286 and Intel386™ processors) did not have an on-chip floating-point unit.
Instead, floating-point capability was provided on a separate numeric coprocessor chip. The first
of these numeric coprocessors was the Intel 8087, which was followed by the Intel 287 and Intel
387 numeric coprocessors.
To allow the 8087 to signal floating-point exceptions to its companion 8086 or 8088, the 8087
has an output pin, INT, which it asserts when an unmasked floating-point exception occurs. The
designers of the 8087 recommended that the output from this pin be routed through a program-
mable interrupt controller (PIC) such as the Intel 8259A to the INTR pin of the 8086 or 8088.
The accompanying interrupt vector number could then be used to access the floating-point ex-
ception handler.
However, the original IBM PC design and MS-DOS operating system used a different mecha-
nism for handling the INT output from the 8087. It connected the INT pin directly to the NMI
input pin of the 8086 or 8088. The NMI interrupt handler then had to determine if the interrupt
was caused by a floating-point exception or another NMI event. This mechanism is the origin
of what is now called the “MS-DOS compatibility mode.” The decision to use this latter float-
ing-point exception handling mechanism came about because when the IBM PC was first de-
signed, the 8087 was not available. When the 8087 did become available, other functions had
already been assigned to the eight inputs to the PIC. One of these functions was a BIOS video
interrupt, which was assigned to interrupt number 16 for the 8086 and 8088.
The Intel 286 processor created the “native mode” for handling floating-point exceptions by pro-
viding a dedicated input pin (ERROR#) for receiving floating-point exception signals and a ded-
icated interrupt number, 16. Interrupt 16 was used to signal floating-point errors (also called
math faults). It was intended that the ERROR# pin on the Intel 286 be connected to a corre-
sponding ERROR# pin on the Intel 287 numeric coprocessor. When the Intel 287 signals a float-
ing-point exception using this mechanism, the Intel 286 generates an interrupt 16, to invoke the
floating-point exception handler.
To maintain compatibility existing PC software, the native floating-point exception handling
mode of the Intel 286 and 287 was not used in the IBM PC AT* system design. Instead, the ER-
ROR# pin on the Intel 286 was tied permanently high, and the ERROR# pin from the Intel 287
was routed to a second (cascaded) PIC. The resulting output of this PIC was routed through an
exception handler and eventually caused an interrupt 2 (NMI interrupt). Here the NMI interrupt
was shared with PC AT’s new parity checking feature. Interrupt 16 remained assigned to the
BIOS video interrupt handler. The external hardware for the MS-DOS compatibility mode must
prevent the Intel 286 processor from executing past the next FPU instruction when an unmasked
exception has been generated. To do this, it asserts the BUSY# signal into the Intel 286 when
the ERROR# signal is asserted by the Intel 287.
The Intel386™ processor and its companion Intel 387 numeric coprocessor provided the same
hardware mechanism for signaling and handling floating-point exceptions as the Intel 286 and
287 processors. And again, to maintain compatibility with existing MS-DOS software, basically

E-2
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

the same MS-DOS compatibility floating-point exception handling mechanism that was used in
the PC AT was used in PCs based on the Intel386™.

E.2. IMPLEMENTATION OF THE MS-DOS* COMPATIBILITY MODE

IN THE INTEL486™, PENTIUM®, AND P6 FAMILY
PROCESSORS
Beginning with the Intel486™ processor, the IA provided a dedicated mechanism for enabling
the MS-DOS compatibility mode for FPU exceptions and for generating external FPU-excep-
tion signals while operating in this mode. The following sections describe the implementation
of the MS-DOS compatibility mode in Intel486™, Pentium® processors, and P6 family proces-
sors. Also described is the recommended external hardware to support this mode of operation.

E.2.1. MS-DOS* Compatibility Mode in the Intel486™ and

Pentium® Processors
In the Intel486™, several things were done to enhance and speed up the numeric coprocessor,
now called the floating-point unit (FPU). The most important enhancement was that the FPU
was included in the same chip as the processor, for increased speed in FPU computations and
reduced latency for FPU exception handling. Also, for the first time, the MS-DOS compatibility
mode was built into the chip design, with the addition of the NE bit in control register CR0 and
the addition of the FERR# (Floating-point ERRor) and IGNNE# (IGNore Numeric Error) pins.
The NE bit selects the native FPU exception handling mode (NE = 1) or the MS-DOS compat-
ibility mode (NE = 0). When native mode is selected, all signaling of floating-point exceptions
is handled internally in the Intel486™ chip, resulting in the generation of an interrupt 16.
When MS-DOS compatibility mode is selected the FERRR# and IGNNE# pins are used to sig-
nal floating-point exceptions. The FERR# output pin, which replaces the ERROR# pin from the
previous generations of IA numeric coprocessors, is connected to a PIC. A new input signal,
IGNNE#, is provided to allow the FPU exception handler to execute FPU instructions, if de-
sired, without first clearing the error condition and without triggering the interrupt a second
time. This IGNNE# feature is needed to replicate the capability that was provided on MS-DOS
compatible Intel 286 and Intel 287 and Intel386™ and Intel 387 systems by turning off the
BUSY# signal, when inside the FPU exception handler, before clearing the error condition.
Note that Intel, in order to provide Intel486™ processors for market segments which had no
need for an FPU, created the “SX” versions. These Intel486™ SX processors did not contain the
floating-point unit. Intel also produced Intel 487 SX processors for end users who later decided
to upgrade to a system with an FPU. These Intel 487 SX processors are similar to standard
Intel486™ processors with a working FPU on board. Thus, the external circuitry necessary to
support the MS-DOS compatibility mode for Intel 487 SX processors is the same as for standard
Intel486™ DX processors.
The Pentium® and P6 family processors offer the same mechanism (the NE bit and the FERR#
and IGNNE# pins) as the Intel486™ processors for generating FPU exceptions in MS-DOS

E-3
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

compatibility mode. The actions of these mechanisms are slightly different and more straight-
forward for the P6 family processors, as described in Section E.2.2., “MS-DOS* Compatibility
Mode in the P6 Family Processors”.
For Pentium® and P6 family processors, it is important to note that the special DP (Dual Pro-
cessing) mode for Pentium® processors and also the more general Intel MultiProcessor Specifi-
cation for systems with multiple Pentium® or P6 family processors support FPU exception
handling only in the native mode. Intel does not recommend using the MS-DOS compatibility
FPU mode for systems using more than one processor.

E.2.1.1. BASIC RULES: WHEN FERR# IS GENERATED

When MS-DOS compatibility mode is enabled for the Intel486™ or Pentium® processors (NE
bit is set to 0) and the IGNNE# input pin is de-asserted, the FERR# signal is generated as fol-
lows:
1. When an FPU instruction causes an unmasked FPU exception, the processor (in most
cases) uses a “deferred” method of reporting the error. This means that the processor does
not respond immediately, but rather freezes just before executing the next WAIT or FPU
instruction (except for “no-wait” instructions, which the FPU executes regardless of an
error condition).
2. When the processor freezes, it also asserts the FERR# output.
3. The frozen processor waits for an external interrupt, which must be supplied by external
hardware in response to the FERR# assertion.
4. In MS-DOS* compatibility systems, FERR# is fed to the IRQ13 input in the cascaded PIC.
The PIC generates interrupt 75H, which then branches to interrupt 2, as described earlier in
this appendix for systems using the Intel 286 and Intel 287 or Intel386™ and Intel 387
processors.
The deferred method of error reporting is used for all exceptions caused by the basic arithmetic
instructions (including FADD, FSUB, FMUL, FDIV, FSQRT, FCOM and FUCOM), for preci-
sion exceptions caused by all types of FPU instructions, and for numeric underflow and over-
flow exceptions caused by all types of FPU instructions except stores to memory.
Some FPU instructions with some FPU exceptions use an “immediate” method of reporting er-
rors. Here, the FERR# is asserted immediately, at the time that the exception occurs. The imme-
diate method of error reporting is used for FPU stack fault, invalid operation and denormal
exceptions caused by all transcendental instructions, FSCALE, FXTRACT, FPREM and others,
and all exceptions (except precision) when caused by FPU store instructions. Like deferred error
reporting, immediate error reporting will cause the processor to freeze just before executing the
next WAIT or FPU instruction if the error condition has not been cleared by that time.
Note that in general, whether deferred or immediate error reporting is used for an FPU exception
depends both on which exception occurred and which instruction caused that exception. A com-
plete specification of these cases, which applies to both the Pentium® and the Intel486™ pro-
cessors, is given in Section 5.1.2.1., “Program-Error Exceptions”, in Chapter 5, Interrupt and
Exception Handling, of the Intel Architecture Software Developer’s Manual, Volume 3.

E-4
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

If NE=0 but the IGNNE# input is active while an unmasked FPU exception is in effect, the pro-
cessor disregards the exception, does not assert FERR#, and continues. If IGNNE# is then de-
asserted and the FPU exception has not been cleared, the processor will respond as described
above. (That is, an immediate exception case will assert FERR# immediately. A deferred excep-
tion case will assert FERR# and freeze just before the next FPU or WAIT instruction.) The as-
sertion of IGNNE# is intended for use only inside the FPU exception handler, where it is needed
if one wants to execute non-control FPU instructions for diagnosis, before clearing the exception
condition. When IGNNE# is asserted inside the exception handler, a preceding FPU exception
has already caused FERR# to be asserted, and the external interrupt hardware has responded,
but IGNNE# assertion still prevents the freeze at FPU instructions. Note that if IGNNE# is left
active outside of the FPU exception handler, additional FPU instructions may be executed after
a given instruction has caused an FPU exception. In this case, if the FPU exception handler ever
did get invoked, it could not determine which instruction caused the exception.
To properly manage the interface between the processor’s FERR# output, its IGNNE# input, and
the IRQ13 input of the PIC, additional external hardware is needed. A recommended configu-
ration is described in the following section.

E.2.1.2. RECOMMENDED EXTERNAL HARDWARE TO SUPPORT THE

MS-DOS* COMPATIBILITY MODE
Figure E-1 provides an external circuit that will assure proper handling of FERR# and IGNNE#
when an FPU exception occurs. In particular, it assures that IGNNE# will be active only inside
the FPU exception handler without depending on the order of actions by the exception handler.
Some hardware implementations have been less robust because they have depended on the ex-
ception handler to clear the FPU exception interrupt request to the PIC (FP_IRQ signal) before
the handler causes FERR# to be de-asserted by clearing the exception from the FPU itself. Fig-
ure E-2 shows the details of how IGNNE# will behave when the circuit in Figure E-1 is im-
plemented. The temporal regions within the FPU exception handler activity are described as
follows:
1. The FERR# signal is activated by an FPU exception and sends an interrupt request through
the PIC to the processor’s INTR pin.
2. During the FPU interrupt service routine (exception handler) the processor will need to
clear the interrupt request latch (Flip Flop #1). It may also want to execute non-control
FPU instructions before the exception is cleared from the FPU. For this purpose the
IGNNE# must be driven low. Typically in the PC environment an I/O access to Port 0F0H
clears the external FPU exception interrupt request (FP_IRQ). In the recommended circuit,
this access also is used to activate IGNNE#. With IGNNE# active the FPU exception
handler may execute any FPU instruction without being blocked by an active FPU
exception.
3. Clearing the exception within the FPU will cause the FERR# signal to be deactivated and
then there is no further need for IGNNE# to be active. In the recommended circuit, the
deactivation of FERR# is used to deactivate IGNNE#. If another circuit is used, the
software and circuit together must assure that IGNNE# is deactivated no later than the exit
from the FPU exception handler.

E-5
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

FF #1

Intel486,
Pentium®, or
FF #2
Pentium Pro
processor

FP_IRQ

Legend:
FF #n Flip Flop #n
CLR Clear or Reset
Figure E-1. Recommended Circuit for MS-DOS* Compatibility FPU Exception Handling
In the circuit in Figure E-1, when the FPU exception handler accesses I/O port 0F0H it clears
the IRQ13 interrupt request output from Flip Flop #1 and also clocks out the IGNNE# signal
(active) from Flip Flop #2. So the handler can activate IGNNE#, if needed, by doing this 0F0H
access before clearing the FPU exception condition (which de-asserts FERR#). However, the

E-6
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

circuit does not depend on the order of actions by the FPU exception handler to guarantee the
correct hardware state upon exit from the handler. Flip Flop #2, which drives IGNNE# to the
processor, has its CLEAR input attached to the inverted FERR#. This ensures that IGNNE# can
never be active when FERR# is inactive. So if the handler clears the FPU exception condition
before the 0F0H access, IGNNE# does not get activated and left on after exit from the handler.

0F0H Address
Decode

Figure E-2. Behavior of Signals During FPU Exception Handling

E.2.1.3. NO-WAIT FPU INSTRUCTIONS CAN GET FPU INTERRUPT IN

WINDOW
The Pentium® and Intel486™ processors implement the “no-wait” floating-point instructions
(FNINIT, FNCLEX, FNSTENV, FNSAVE, FNSTSW, FNSTCW, FNENI, FNDISI or
FNSETPM) in the MS-DOS compatibility mode in the following manner. (Refer to Section
7.5.11., “FPU Control Instructions” and Section 7.5.12., “Waiting Vs. Non-waiting Instruc-
tions” in Chapter 7, Floating-Point Unit, for a discussion of the no-wait instructions.)
If an unmasked numeric exception is pending from a preceding FPU instruction, a member of
the no-wait class of instructions will, at the beginning of its execution, assert the FERR# pin in
response to that exception just like other FPU instructions, but then, unlike the other FPU in-
structions, FERR# will be de-asserted. This de-assertion was implemented to allow the no-wait
class of instructions to proceed without an interrupt due to any pending numeric exception.
However, the brief assertion of FERR# is sufficient to latch the FPU exception request into most
hardware interface implementations (including Intel’s recommended circuit).
All the FPU instructions are implemented such that during their execution, there is a window in
which the processor will sample and accept external interrupts. If there is a pending interrupt,
the processor services the interrupt first before resuming the execution of the instruction. Con-
sequently, it is possible that the no-wait floating-point instruction may accept the external inter-
rupt caused by it’s own assertion of the FERR# pin in the event of a pending unmasked numeric

E-7
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

exception, which is not an explicitly documented behavior of a no-wait instruction. This process
is illustrated in Figure E-3.

Exception Generating
Floating-Point
Instruction

Assertion of FERR#
by the Processor Start of the “No-Wait”
Floating-Point
Instruction
System
Dependent
Delay
Case 1 External Interrupt
Sampling Window
Assertion of INTR Pin
by the System
Case 2
Window Closed

Figure E-3. Timing of Receipt of External Interrupt

Figure E-3 assumes that a floating-point instruction that generates a “deferred” error (as defined
in the Section E.2.1.1., “Basic Rules: When FERR# Is Generated”), which asserts the FERR#
pin only on encountering the next floating-point instruction, causes an unmasked numeric ex-
ception. Assume that the next floating-point instruction following this instruction is one of the
no-wait floating-point instructions. The FERR# pin is asserted by the processor to indicate the
pending exception on encountering the no-wait floating-point instruction. After the assertion of
the FERR# pin the no-wait floating-point instruction opens a window where the pending exter-
nal interrupts are sampled.
Then there are two cases possible depending on the timing of the receipt of the interrupt via the
INTR pin (asserted by the system in response to the FERR# pin) by the processor.
Case 1 If the system responds to the assertion of FERR# pin by the no-wait floating-point
instruction via the INTR pin during this window then the interrupt is serviced first,
before resuming the execution of the no-wait floating-point instruction.
Case 2 If the system responds via the INTR pin after the window has closed then the inter-
rupt is recognized only at the next instruction boundary.
There are two other ways, in addition to Case 1 above, in which a no-wait floating-point instruc-
tion can service a numeric exception inside its interrupt window. First, the first floating-point
error condition could be of the “immediate” category (as defined in Section E.2.1.1., “Basic
Rules: When FERR# Is Generated”) that asserts FERR# immediately. If the system delay before

E-8
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

asserting INTR is long enough, relative to the time elapsed before the no-wait floating-point in-
struction, INTR can be asserted inside the interrupt window for the latter. Second, consider two
no-wait FPU instructions in close sequence, and assume that a previous FPU instruction has
caused an unmasked numeric exception. Then if the INTR timing is too long for an FERR# sig-
nal triggered by the first no-wait instruction to hit the first instruction’s interrupt window, it
could catch the interrupt window of the second.
The possible malfunction of a no-wait FPU instruction explained above cannot happen if the in-
struction is being used in the manner for which Intel originally designed it. The no-wait instruc-
tions were intended to be used inside the FPU exception handler, to allow manipulation of the
FPU before the error condition is cleared, without hanging the processor because of the FPU er-
ror condition, and without the need to assert IGNNE#. They will perform this function correctly,
since before the error condition is cleared, the assertion of FERR# that caused the FPU error
handler to be invoked is still active. Thus the logic that would assert FERR# briefly at a no-wait
instruction causes no change since FERR# is already asserted. The no-wait instructions may also
be used without problem in the handler after the error condition is cleared, since now they will
not cause FERR# to be asserted at all.
If a no-wait instruction is used outside of the FPU exception handler, it may malfunction as ex-
plained above, depending on the details of the hardware interface implementation and which
particular processor is involved. The actual interrupt inside the window in the no-wait instruc-
tion may be blocked by surrounding it with the instructions: PUSHFD, CLI, no-wait, then
POPFD. (CLI blocks interrupts, and the push and pop of flags preserves and restores the original
value of the interrupt flag.) However, if FERR# was triggered by the no-wait, its latched value
and the PIC response will still be in effect. Further code can be used to check for and correct
such a condition, if needed. Section E.3.5., “Considerations When FPU Shared Between Tasks”
discusses an important example of this type of problem and gives a solution.

E.2.2. MS-DOS* Compatibility Mode in the P6 Family Processors

When bit NE=0 in CR0, the MS-DOS compatibility mode of the P6 family processors provides
FERR# and IGNNE# functionality that is almost identical to the Intel486™ and Pentium® pro-
cessors. The same external hardware described in Section E.2.1.2., “Recommended External
Hardware to Support the MS-DOS* Compatibility Mode” is recommended for the P6 family
processors as well as the two previous generations. The only change to MS-DOS compatibility
FPU exception handling with the P6 family processors is that all exceptions for all FPU instruc-
tions cause immediate error reporting. That is, FERR# is asserted as soon as the FPU detects an
unmasked exception; there are no cases in which error reporting is deferred to the next FPU or
WAIT instruction. (As is discussed in Section E.2.1.1., “Basic Rules: When FERR# Is Generat-
ed”, most exception cases in the Intel486™ and Pentium® processors are of the deferred type.)
Although FERR# is asserted immediately upon detection of an unmasked FPU error, this cer-
tainly does not mean that the requested interrupt will always be serviced before the next instruc-
tion in the code sequence is executed. To begin with, the P6 family processors executes several
instructions simultaneously. There also will be a delay, which depends on the external hardware
implementation, between the FERR# assertion from the processor and the responding INTR as-
sertion to the processor. Further, the interrupt request to the PICs (IRQ13) may be temporarily
blocked by the operating system, or delayed by higher priority interrupts, and processor re-

E-9
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

sponse to INTR itself is blocked if the operating system has cleared the IF bit in EFLAGS. Note
that Streaming SIMD Extensions numeric exceptions will not cause assertion of FERR# (inde-
pendent of the value of CR0.NE). In addition they ignore the assertion /de-assertion of IGNNE#.
However, just as with the Intel486™ and Pentium® processors, if the IGNNE# input is inactive,
a floating-point exception which occurred in the previous FPU instruction and is unmasked
causes the processor to freeze immediately when encountering the next WAIT or FPU instruc-
tion (except for no-wait instructions). This means that if the FPU exception handler has not al-
ready been invoked due to the earlier exception (and therefore, the handler not has cleared that
exception state from the FPU), the processor is forced to wait for the handler to be invoked and
handle the exception, before the processor can execute another WAIT or FPU instruction.
As explained in Section E.2.1.3., “No-Wait FPU Instructions Can Get FPU Interrupt in Win-
dow”, if a no-wait instruction is used outside of the FPU exception handler, in the Intel486™
and Pentium® processors, it may accept an unmasked exception from a previous FPU instruction
which happens to fall within the external interrupt sampling window that is opened near the be-
ginning of execution of all FPU instructions. This will not happen in the P6 family processors,
because this sampling window has been removed from the no-wait group of FPU instructions.

E.3. RECOMMENDED PROTOCOL FOR MS-DOS*

COMPATIBILITY HANDLERS
The activities of numeric programs can be split into two major areas: program control and arith-
metic. The program control part performs activities such as deciding what functions to perform,
calculating addresses of numeric operands, and loop control. The arithmetic part simply adds,
subtracts, multiplies, and performs other operations on the numeric operands. The processor is
designed to handle these two parts separately and efficiently. An FPU exception handler, if a sys-
tem chooses to implement one, is often one of the most complicated parts of the program control
code.

E.3.1. Floating-Point Exceptions and Their Defaults

The FPU can recognize six classes of floating-point exception conditions while executing float-
ing-point instructions:
1. #I — Invalid operation
#IS — Stack fault
#IA — IEEE standard invalid operation
2. #Z — Divide-by-zero
3. #D — Denormalized operand
4. #O — Numeric overflow
5. #U — Numeric underflow
6. #P — Inexact result (precision)

E-10
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

For complete details on these exceptions and their defaults, refer to Section 7.7., “Floating-Point
Exception Handling” and Section 7.8., “Floating-Point Exception Conditions” in Chapter 7,
Floating-Point Unit.

E.3.2. Two Options for Handling Numeric Exceptions

Depending on options determined by the software system designer, the processor takes one of
two possible courses of action when a numeric exception occurs:

• The FPU can handle selected exceptions itself, producing a default fix-up that is
reasonable in most situations. This allows the numeric program execution to continue
undisturbed. Programs can mask individual exception types to indicate that the FPU should
generate this safe, reasonable result whenever the exception occurs. The default exception
fix-up activity is treated by the FPU as part of the instruction causing the exception; no
external indication of the exception is given (except that the instruction takes longer to
execute when it handles a masked exception.) When masked exceptions are detected, a
flag is set in the numeric status register, but no information is preserved regarding where or
when it was set.

• Alternatively, a software exception handler can be invoked to handle the exception. When
a numeric exception is unmasked and the exception occurs, the FPU stops further
execution of the numeric instruction and causes a branch to a software exception handler.
The exception handler can then implement any sort of recovery procedures desired for any
numeric exception detectable by the FPU.

E.3.2.1. AUTOMATIC EXCEPTION HANDLING: USING MASKED

EXCEPTIONS
Each of the six exception conditions described above has a corresponding flag bit in the FPU
status word and a mask bit in the FPU control word. If an exception is masked (the correspond-
ing mask bit in the control word = 1), the processor takes an appropriate default action and con-
tinues with the computation. The processor has a default fix-up activity for every possible
exception condition it may encounter. These masked-exception responses are designed to be
safe and are generally acceptable for most numeric applications.
For example, if the Inexact result (Precision) exception is masked, the system can specify
whether the FPU should handle a result that cannot be represented exactly by one of four modes
of rounding: rounding it normally, chopping it toward zero, always rounding it up, or always
down. If the Underflow exception is masked, the FPU will store a number that is too small to be
represented in normalized form as a denormal (or zero if it’s smaller than the smallest denor-
mal). Note that when exceptions are masked, the FPU may detect multiple exceptions in a single
instruction, because it continues executing the instruction after performing its masked response.
For example, the FPU could detect a denormalized operand, perform its masked response to this
exception, and then detect an underflow.
As an example of how even severe exceptions can be handled safely and automatically using the
default exception responses, consider a calculation of the parallel resistance of several values
using only the standard formula (refer to Figure E-4). If R1 becomes zero, the circuit resistance

E-11
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

becomes zero. With the divide-by-zero and precision exceptions masked, the processor will pro-
duce the correct result. FDIV of R1 into 1 gives infinity, and then FDIV of (infinity +R2 +R3)
into 1 gives zero.

R1 R2 R3

1
Equivalent Resistance =
1 1 1
+ +
R1 R2 R3

Figure E-4. Arithmetic Example Using Infinity

By masking or unmasking specific numeric exceptions in the FPU control word, programmers
can delegate responsibility for most exceptions to the processor, reserving the most severe ex-
ceptions for programmed exception handlers. Exception-handling software is often difficult to
write, and the masked responses have been tailored to deliver the most reasonable result for each
condition. For the majority of applications, masking all exceptions yields satisfactory results
with the least programming effort. Certain exceptions can usefully be left unmasked during the
debugging phase of software development, and then masked when the clean software is actually
run. An invalid operation exception for example, typically indicates a program error that must
be corrected.
The exception flags in the FPU status word provide a cumulative record of exceptions that have
occurred since these flags were last cleared. Once set, these flags can be cleared only by execut-
ing the FCLEX/FNCLEX (clear exceptions) instruction, by reinitializing the FPU with
FINIT/FNINIT or FSAVE/FNSAVE, or by overwriting the flags with an FRSTOR or FLDENV
instruction. This allows a programmer to mask all exceptions, run a calculation, and then inspect
the status word to see if any exceptions were detected at any point in the calculation.

E.3.2.2. SOFTWARE EXCEPTION HANDLING

If the FPU in or with an IA processor (Intel 286 and onwards) encounters an unmasked excep-
tion condition, with the system operated in the MS-DOS compatibility mode and with IGNNE#
not asserted, a software exception handler is invoked through a PIC and the processor’s INTR
pin. The FERR# (or ERROR#) output from the FPU that begins the process of invoking the ex-
ception handler may occur when the error condition is first detected, or when the processor en-
counters the next WAIT or FPU instruction. Which of these two cases occurs depends on the
processor generation and also on which exception and which FPU instruction triggered it, as dis-

E-12
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

cussed earlier in Section E.1., “Origin of the MS-DOS* Compatibility Mode for Handling FPU
Exceptions” and Section E.2., “Implementation of the MS-DOS* Compatibility Mode in the
Intel486™, Pentium®, and P6 family processors” The elapsed time between the initial error sig-
nal and the invocation of the FPU exception handler depends of course on the external hardware
interface, and also on whether the external interrupt for FPU errors is enabled. But the architec-
ture ensures that the handler will be invoked before execution of the next WAIT or floating-point
instruction since an unmasked floating-point exception causes the processor to freeze just before
executing such an instruction (unless the IGNNE# input is active, or it is a no-wait FPU instruc-
tion).
The frozen processor waits for an external interrupt, which must be supplied by external hard-
ware in response to the FERR# (or ERROR#) output of the processor (or coprocessor), usually
through IRQ13 on the “slave” PIC, and then through INTR. Then the external interrupt invokes
the exception handling routine. Note that if the external interrupt for FPU errors is disabled
when the processor executes an FPU instruction, the processor will freeze until some other (en-
abled) interrupt occurs if an unmasked FPU exception condition is in effect. If NE = 0 but the
IGNNE# input is active, the processor disregards the exception and continues. Error reporting
via an external interrupt is supported for MS-DOS compatibility. Chapter 18, Intel Architecture
Compatibility of the Intel Architecture Software Developer’s Manual, Volume 3, contains further
discussion of compatibility issues.
The references above to the ERROR# output from the FPU apply to the Intel 387 and Intel 287
math coprocessors (NPX chips). If one of these coprocessors encounters an unmasked exception
condition, it signals the exception to the Intel 286 or Intel386™ processor using the ERROR#
status line between the processor and the coprocessor. Refer to Section E.1., “Origin of the MS-
DOS* Compatibility Mode for Handling FPU Exceptions”, in this appendix, and Chapter 18,
Intel Architecture Compatibility, in the Intel Architecture Software Developer’s Manual, Volume
3 for differences in FPU exception handling.
The exception-handling routine is normally a part of the systems software. The routine must
clear (or disable) the active exception flags in the FPU status word before executing any float-
ing-point instructions that cannot complete execution when there is a pending floating-point ex-
ception. Otherwise, the floating-point instruction will trigger the FPU interrupt again, and the
system will be caught in an endless loop of nested floating-point exceptions, and hang. In any
event, the routine must clear (or disable) the active exception flags in the FPU status word after
handling them, and before IRET(D). Typical exception responses may include:

• Incrementing an exception counter for later display or printing.

• Printing or displaying diagnostic information (e.g., the FPU environment and registers).

• Aborting further execution, or using the exception pointers to build an instruction that will
run without exception and executing it.
Applications programmers should consult their operating system’s reference manuals for the ap-
propriate system response to numerical exceptions. For systems programmers, some details on
writing software exception handlers are provided in Chapter 5, Interrupt and Exception Han-
dling, in the Intel Architecture Software Developer’s Manual, Volume 3, as well as in Section
E.3.3.4., “FPU Exception Handling Examples” in this appendix.

E-13
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

As discussed in Section E.2.1.2., “Recommended External Hardware to Support the MS-DOS*

Compatibility Mode”, some early FERR# to INTR hardware interface implementations are less
robust than the recommended circuit. This is because they depended on the exception handler
to clear the FPU exception interrupt request to the PIC (by accessing port 0F0H) before the han-
dler causes FERR# to be de-asserted by clearing the exception from the FPU itself. To eliminate
the chance of a problem with this early hardware, Intel recommends that FPU exception han-
dlers always access port 0F0H before clearing the error condition from the FPU.

E.3.3. Synchronization Required for Use of FPU Exception

Handlers
Concurrency or synchronization management requires a check for exceptions before letting the
processor change a value just used by the FPU. It is important to remember that almost any nu-
meric instruction can, under the wrong circumstances, produce a numeric exception.

E.3.3.1. EXCEPTION SYNCHRONIZATION: WHAT, WHY AND WHEN

Exception synchronization means that the exception handler inspects and deals with the excep-
tion in the context in which it occurred. If concurrent execution is allowed, the state of the pro-
cessor when it recognizes the exception is often not in the context in which it occurred. The
processor may have changed many of its internal registers and be executing a totally different
program by the time the exception occurs. If the exception handler cannot recapture the original
context, it cannot reliably determine the cause of the exception or to recover successfully from
the exception. To handle this situation, the FPU has special registers updated at the start of each
numeric instruction to describe the state of the numeric program when the failed instruction was
attempted. This provides tools to help the exception handler recapture the original context, but
the application code must also be written with synchronization in mind. Overall, exception syn-
chronization must ensure that the FPU and other relevant parts of the context are in a well de-
fined state when the handler is invoked after an unmasked numeric exception occurs.
When the FPU signals an unmasked exception condition, it is requesting help. The fact that the
exception was unmasked indicates that further numeric program execution under the arithmetic
and programming rules of the FPU will probably yield invalid results. Thus the exception must
be handled, and with proper synchronization, or the program will not operate reliably.
For programmers in higher-level languages, all required synchronization is automatically pro-
vided by the appropriate compiler. However, for assembly language programmers exception
synchronization remains the responsibility of the programmer. It is not uncommon for a pro-
grammer to expect that their numeric program will not cause numeric exceptions after it has
been tested and debugged, but in a different system or numeric environment, exceptions may
occur regularly nonetheless. An obvious example would be use of the program with some num-
bers beyond the range for which it was designed and tested. Example E-1 and Example E-2 in
Section E.3.3.2., “Exception Synchronization Examples” shows a more subtle way in which un-
expected exceptions can occur.

E-14
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

As described in Section E.3.1., “Floating-Point Exceptions and Their Defaults”, depending on

options determined by the software system designer, the processor can perform one of two pos-
sible courses of action when a numeric exception occurs.

• The FPU can provide a default fix-up for selected numeric exceptions. If the FPU performs
its default action for all exceptions, then the need for exception synchronization is not
manifest. However, code is often ported to contexts and operating systems for which it was
not originally designed. Example E-1 and Example E-2, below, illustrate that it is safest to
always consider exception synchronization when designing code that uses the FPU.

• Alternatively, a software exception handler can be invoked to handle the exception. When
a numeric exception is unmasked and the exception occurs, the FPU stops further
execution of the numeric instruction and causes a branch to a software exception handler.
When an FPU exception handler will be invoked, synchronization must always be
considered to assure reliable performance.
Example E-1 and Example E-2, below, illustrate the need to always consider exception synchro-
nization when writing numeric code, even when the code is initially intended for execution with
exceptions masked.

E.3.3.2. EXCEPTION SYNCHRONIZATION EXAMPLES

In the following examples, three instructions are shown to load an integer, calculate its square
root, then increment the integer. The synchronous execution of the FPU will allow both of these
programs to execute correctly, with INC COUNT being executed in parallel in the processor, as
long as no exceptions occur on the FILD instruction. However, if the code is later moved to an
environment where exceptions are unmasked, the code in Example E-1 will not work correctly:

Example E-1. Incorrect Error Synchronization

FILD COUNT; FPU instruction
INC COUNT; integer instruction alters operand
FSQRT ; subsequent FPU instruction -- error
; from previous FPU instruction detected here

Example E-2. Proper Error Synchronization

FILD COUNT; FPU instruction
FSQRT ; subsequent FPU instruction -- error from
; previous FPU instruction detected here
INC COUNT; integer instruction alters operand
In some operating systems supporting the FPU, the numeric register stack is extended to mem-
ory. To extend the FPU stack to memory, the invalid exception is unmasked. A push to a full
register or pop from an empty register sets SF (Stack Fault flag) and causes an invalid operation
exception. The recovery routine for the exception must recognize this situation, fix up the stack,
then perform the original operation. The recovery routine will not work correctly in Example
E-1. The problem is that the value of COUNT is incremented before the exception handler is

E-15
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

invoked, so that the recovery routine will load an incorrect value of COUNT, causing the pro-
gram to fail or behave unreliably.

E.3.3.3. PROPER EXCEPTION SYNCHRONIZATION IN GENERAL

As explained in Section E.2.1.2., “Recommended External Hardware to Support the MS-DOS*
Compatibility Mode”, if the FPU encounters an unmasked exception condition a software ex-
ception handler is invoked before execution of the next WAIT or floating-point instruction. This
is because an unmasked floating-point exception causes the processor to freeze immediately be-
fore executing such an instruction (unless the IGNNE# input is active, or it is a no-wait FPU
instruction). Exactly when the exception handler will be invoked (in the interval between when
the exception is detected and the next WAIT or FPU instruction) is dependent on the processor
generation, the system, and which FPU instruction and exception is involved.
To be safe in exception synchronization, one should assume the handler will be invoked at the
end of the interval. Thus the program should not change any value that might be needed by the
handler (such as COUNT in Example E-1 and Example E-2) until after the next FPU instruction
following an FPU instruction that could cause an error. If the program needs to modify such a
value before the next FPU instruction (or if the next FPU instruction could also cause an error),
then a WAIT instruction should be inserted before the value is modified. This will force the han-
dling of any exception before the value is modified. A WAIT instruction should also be placed
after the last floating-point instruction in an application so that any unmasked exceptions will
be serviced before the task completes.

E.3.3.4. FPU EXCEPTION HANDLING EXAMPLES

There are many approaches to writing exception handlers. One useful technique is to consider
the exception handler procedure as consisting of “prologue,” “body,” and “epilogue” sections of
code.
In the transfer of control to the exception handler due to an INTR, NMI, or SMI, external inter-
rupts have been disabled by hardware. The prologue performs all functions that must be protect-
ed from possible interruption by higher-priority sources. Typically, this involves saving registers
and transferring diagnostic information from the FPU to memory. When the critical processing
has been completed, the prologue may re-enable interrupts to allow higher-priority interrupt
handlers to preempt the exception handler. The standard “prologue” not only saves the registers
and transfers diagnostic information from the FPU to memory but also clears the floating-point
exception flags in the status word. Alternatively, when it is not necessary for the handler to be
re-entrant, another technique may also be used. In this technique, the exception flags are not
cleared in the “prologue” and the body of the handler must not contain any floating-point in-
structions that cannot complete execution when there is a pending floating-point exception. (The
no-wait instructions are discussed in Section 7.5.12., “Waiting Vs. Non-waiting Instructions” in
Chapter 7, Floating-Point Unit.) Note that the handler must still clear the exception flag(s) be-
fore executing the IRET. If the exception handler uses neither of these techniques the system will
be caught in an endless loop of nested floating-point exceptions, and hang.
The body of the exception handler examines the diagnostic information and makes a response
that is necessarily application-dependent. This response may range from halting execution, to

E-16
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

displaying a message, to attempting to repair the problem and proceed with normal execution.
The epilogue essentially reverses the actions of the prologue, restoring the processor so that nor-
mal execution can be resumed. The epilogue must not load an unmasked exception flag into the
FPU or another exception will be requested immediately.
The following code examples show the ASM386/486 coding of three skeleton exception han-
dlers, with the save spaces given as correct for 32-bit protected mode. They show how prologues
and epilogues can be written for various situations, but the application dependent exception han-
dling body is just indicated by comments showing where it should be placed.
The first two are very similar; their only substantial difference is their choice of instructions to
save and restore the FPU. The trade-off here is between the increased diagnostic information
provided by FNSAVE and the faster execution of FNSTENV. (Also, after saving the original
contents, FNSAVE re-initializes the FPU, while FNSTENV only masks all FPU exceptions.)
For applications that are sensitive to interrupt latency or that do not need to examine register
contents, FNSTENV reduces the duration of the “critical region,” during which the processor
does not recognize another interrupt request. (Refer to Section 7, “Floating-Point Unit” in Chap-
ter 7, Floating-Point Unit, for a complete description of the FPU save image.) If the processor
supports Streaming SIMD Extensions and the operating system supports it, the FXSAVE in-
struction should be used instead of FNSAVE. If the FXSAVE instruction is used, the save area
should be increased to 512 bytes and aligned to 16 bytes to save the entire state. These steps will
ensure that the complete context is saved.
After the exception handler body, the epilogues prepare the processor to resume execution from
the point of interruption (i.e., the instruction following the one that generated the unmasked ex-
ception). Notice that the exception flags in the memory image that is loaded into the FPU are
cleared to zero prior to reloading (in fact, in these examples, the entire status word image is
cleared).
Example E-3 and Example E-4 assume that the exception handler itself will not cause an un-
masked exception. Where this is a possibility, the general approach shown in Example E-5 can
be employed. The basic technique is to save the full FPU state and then to load a new control
word in the prologue. Note that considerable care should be taken when designing an exception
handler of this type to prevent the handler from being reentered endlessly.

Example E-3. Full-State Exception Handler

SAVE_ALLPROC
;
; SAVE REGISTERS, ALLOCATE STACK SPACE FOR FPU STATE IMAGE
PUSHEBP
.
.
MOV EBP, ESP
SUB ESP, 108 ; ALLOCATES 108 BYTES (32-bit PROTECTED MODE SIZE)
;SAVE FULL FPU STATE, RESTORE INTERRUPT ENABLE FLAG (IF)
FNSAVE[EBP-108]
PUSH [EBP + OFFSET_TO_EFLAGS] ; COPY OLD EFLAGS TO STACK TOP
POPFD ; RESTORE IF TO VALUE BEFORE FPU EXCEPTION

E-17
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

;
; APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE
;
; CLEAR EXCEPTION FLAGS IN STATUS WORD (WHICH IS IN MEMORY)
; RESTORE MODIFIED STATE IMAGE
MOVBYTE PTR [EBP-104], 0H
FRSTOR[EBP-108]
; DE-ALLOCATE STACK SPACE, RESTORE REGISTERS
MOVESP, EBP
.
.
POPEBP
;
; RETURN TO INTERRUPTED CALCULATION
IRETD
SAVE_ALLENDP

Example E-4. Reduced-Latency Exception Handler

SAVE_ENVIRONMENTPROC
;
; SAVE REGISTERS, ALLOCATE STACK SPACE FOR FPU ENVIRONMENT
PUSHEBP
.
.
MOV EBP, ESP
SUB ESP, 28 ; ALLOCATES 28 BYTES (32-bit PROTECTED MODE SIZE)
;SAVE ENVIRONMENT, RESTORE INTERRUPT ENABLE FLAG (IF)
FNSTENV[EBP-28]
PUSH [EBP + OFFSET_TO_EFLAGS] ; COPY OLD EFLAGS TO STACK TOP
POPFD ; RESTORE IF TO VALUE BEFORE FPU EXCEPTION

;
; APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE
;
; CLEAR EXCEPTION FLAGS IN STATUS WORD (WHICH IS IN MEMORY)
; RESTORE MODIFIED ENVIRONMENT IMAGE
MOV BYTE PTR [EBP-24], 0H
FLDENV[EBP-28]
; DE-ALLOCATE STACK SPACE, RESTORE REGISTERS
MOV ESP, EBP
.
.
POP EBP
;
; RETURN TO INTERRUPTED CALCULATION
IRETD

E-18
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

SAVE_ENVIRONMENT ENDP

Example E-5. Reentrant Exception Handler

.
.
LOCAL_CONTROL DW ?; ASSUME INITIALIZED
.
.
REENTRANTPROC
;
; SAVE REGISTERS, ALLOCATE STACK SPACE FOR FPU STATE IMAGE
PUSH EBP
.
.
MOV EBP, ESP
SUB ESP, 108 ; ALLOCATES 108 BYTES (32-bit PROTECTED MODE SIZE)

; SAVE STATE, LOAD NEW CONTROL WORD, RESTORE INTERRUPT ENABLE FLAG (IF)
FNSAVE[EBP-108]
FLDCW LOCAL_CONTROL
PUSH [EBP + OFFSET_TO_EFLAGS] ; COPY OLD EFLAGS TO STACK TOP
POPFD ; RESTORE IF TO VALUE BEFORE FPU EXCEPTION

.
.
;
; APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE. AN
UNMASKED EXCEPTION

; GENERATED HERE WILL CAUSE THE EXCEPTION HANDLER TO BE REENTERED.

; IF LOCAL STORAGE IS NEEDED, IT MUST BE ALLOCATED ON THE STACK.
;
.
.
; CLEAR EXCEPTION FLAGS IN STATUS WORD (WHICH IS IN MEMORY)
; RESTORE MODIFIED STATE IMAGE
MOV BYTE PTR [EBP-104], 0H
FRSTOR[EBP-108]
; DE-ALLOCATE STACK SPACE, RESTORE REGISTERS
MOV ESP, EBP
.
.
POP EBP
;
; RETURN TO POINT OF INTERRUPTION
IRETD
REENTRANT ENDP

E-19
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

E.3.4. Need for Storing State of IGNNE# Circuit If Using FPU and
SMM
The recommended circuit (refer to Figure E-1) for MS-DOS compatibility FPU exception han-
dling for Intel486™ processors and beyond contains two flip flops. When the FPU exception
handler accesses I/O port 0F0H it clears the IRQ13 interrupt request output from Flip Flop #1
and also clocks out the IGNNE# signal (active) from Flip Flop #2. The assertion of IGNNE#
may be used by the handler if needed to execute any FPU instruction while ignoring the pending
FPU errors. The problem here is that the state of Flip Flop #2 is effectively an additional (but
hidden) status bit that can affect processor behavior, and so ideally should be saved upon enter-
ing SMM, and restored before resuming to normal operation. If this is not done, and also the
SMM code saves the FPU state, AND an FPU error handler is being used which relies on
IGNNE# assertion, then (very rarely) the FPU handler will nest inside itself and malfunction.
The following example shows how this can happen.
Suppose that the FPU exception handler includes the following sequence:
FNSTSWsave_sw ; save the FPU status word
; using a no-wait FPU instruction
OUT0F0H, AL; clears IRQ13 & activates IGNNE#
....
FLDCW new_cw; loads new CW ignoring FPU errors,
; since IGNNE# is assumed active; or any
; other FPU instruction that is not a no-wait
; type will cause the same problem
....
FCLEX ; clear the FPU error conditions & thus turn off FERR# & reset the IGNNE# FF
The problem will only occur if the processor enters SMM between the OUT and the FLDCW
instructions. But if that happens, AND the SMM code saves the FPU state using FNSAVE, then
the IGNNE# Flip Flop will be cleared (because FNSAVE clears the FPU errors and thus de-as-
serts FERR#). When the processor returns from SMM it will restore the FPU state with FR-
STOR, which will re-assert FERR#, but the IGNNE# Flip Flop will not get set. Then when the
FPU error handler executes the FLDCW instruction, the active error condition will cause the
processor to re-enter the FPU error handler from the beginning. This may cause the handler to
malfunction.
To avoid this problem, Intel recommends two measures:
1. Do not use the FPU for calculations inside SMM code. (The normal power management,
and sometimes security, functions provided by SMM have no need for FPU calculations; if
they are needed for some special case, use scaling or emulation instead.) This eliminates
the need to do FNSAVE/FRSTOR inside SMM code, except when going into a 0 V
suspend state (in which, in order to save power, the CPU is turned off completely, requiring
its complete state to be saved.)
2. The system should not call upon SMM code to put the processor into 0 V suspend while
the processor is running FPU calculations, or just after an interrupt has occurred. Normal
power management protocol avoids this by going into power down states only after timed
intervals in which no system activity occurs.

E-20
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

E.3.5. Considerations When FPU Shared Between Tasks

The IA allows speculative deferral of floating-point state swaps on task switches. This feature
allows postponing an FPU state swap until an FPU instruction is actually encountered in another
task. Since kernel tasks rarely use floating-point, and some applications do not use floating-point
or use it infrequently, the amount of time saved by avoiding unnecessary stores of the floating-
point state is significant. Speculative deferral of FPU saves does, however, place an extra burden
on the kernel in three key ways:
1. The kernel must keep track of which thread owns the FPU, which may be different from
the currently executing thread.
2. The kernel must associate any floating-point exceptions with the generating task. This
requires special handling since floating-point exceptions are delivered asynchronous with
other system activity.
3. There are conditions under which spurious floating-point exception interrupts are
generated, which the kernel must recognize and discard.

E.3.5.1. SPECULATIVELY DEFERRING FPU SAVES, GENERAL OVERVIEW

In order to support multitasking, each thread in the system needs a save area for the general-pur-
pose registers, and each task that is allowed to use floating-point needs an FPU save area large
enough to hold the entire FPU stack and associated FPU state such as the control word and status
word. (Refer to Section 7.3.9., “Saving the FPU’s State” in Chapter 7, Floating-Point Unit, for
a complete description of the FPU save image.) If the processor and the operating system sup-
port Streaming SIMD Extensions, the save area should be large enough and aligned correctly to
hold FPU and Streaming SIMD Extensions state.
On a task switch, the general-purpose registers are swapped out to their save area for the sus-
pending thread, and the registers of the resuming thread are loaded. The FPU state does not need
to be saved at this point. If the resuming thread does not use the FPU before it is itself suspended,
then both a save and a load of the FPU state has been avoided. It is often the case that several
threads may be executed without any usage of the FPU.
The processor supports speculative deferral of FPU saves via interrupt 7 “Device Not Available”
(DNA), used in conjunction with CR0 bit 3, the “Task Switched” bit (TS). (Refer to Section 2.5.,
“Control Registers”, in Chapter 2, System Architecture Overview of the Intel Architecture Soft-
ware Developer’s Manual, Volume 3.) Every task switch via the hardware supported task switch-
ing mechanism (refer to Section 6.3., “Task Switching” in Chapter 6, Task Management of the
Intel Architecture Software Developer’s Manual, Volume 3) sets TS. Multi-threaded kernels that
use software task switchingI can set the TS bit by reading CR0, ORing a “1” intoII bit 3, and
writing back CR0. Any subsequent floating-point instructions (now being executed in a new
thread context) will fault via interrupt 7 before execution. This allows a DNA handler to save
the old floating-point context and reload the FPU state for the current thread. The handler should

NOTES
I In a software task switch, the operating system uses a sequence of instructions to save the suspending
thread’s state and restore the resuming thread’s state, instead of the single long non-interruptible task
switch operation provided by the IA.

E-21
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

clear the TS bit before exit using the CLTS instruction. On return from the handler the faulting
thread will proceed with its floating-point computation.
Some operating systems save the FPU context on every task switch, typically because they also
change the linear address space between tasks. The problem and solution discussed in the fol-
lowing sections apply to these operating systems also.

E.3.5.2. TRACKING FPU OWNERSHIP

Since the contents of the FPU may not belong to the currently executing thread, the thread iden-
tifier for the last FPU user needs to be tracked separately. This is not complicated; the kernel
should simply provide a variable to store the thread identifier of the FPU owner, separate from
the variable that stores the identifier for the currently executing thread. This variable is updated
in the DNA exception handler, and is used by the DNA exception handler to find the FPU save
areas of the old and new threads. A simplified flow for a DNA exception handler is then:
1. Use the “FPU Owner” variable to find the FPU save area of the last thread to use the FPU.
2. Save the FPU contents to the old thread’s save area, typically using an FNSAVE or
FXSAVE instruction.
3. Set the FPU Owner variable to the identify the currently executing thread.
4. Reload the FPU contents from the new thread’s save area, typically using an FRSTOR or
FXRSTOR instruction.
5. Clear TS using the CLTS instruction and exit the DNA exception handler.
While this flow covers the basic requirements for speculatively deferred FPU state swaps, there
are some additional subtleties that need to be handled in a robust implementation.

E.3.5.3. INTERACTION OF FPU STATE SAVES AND FLOATING-POINT

EXCEPTION ASSOCIATION
Recall these key points from earlier in this document: When considering floating-point excep-
tions across all implementations of the IA, and across all floating-point instructions, an floating-
point exception can be initiated from any time during the excepting floating-point instruction,
up to just before the next floating-point instruction. The “next” floating-point instruction may
be the FNSAVE used to save the FPU state for a task switch. In the case of “no-wait:” instruc-
tions such as FNSAVE, the interrupt from a previously excepting instruction (NE=0 case) may
arrive just before the no-wait instruction, during, or shortly thereafter with a system dependent
delay. Note that this implies that an floating-point exception might be registered during the state
swap process itself, and the kernel and floating-point exception interrupt handler must be pre-
pared for this case.

NOTES
II Although CR0, bit 2, the emulation flag (EM), also causes a DNA exception, do not use the EM bit as a
surrogate for TS. EM means that no floating-point unit is available and that floating-point instructions
must be emulated. Using EM to trap on task switches is not compatible with IA MMX™ technology. If the
EM flag is set, MMX™ instructions raise the invalid opcode exception.

E-22
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

A simple way to handle the case of exceptions arriving during FPU state swaps is to allow the
kernel to be one of the FPU owning threads. A reserved thread identifier is used to indicate ker-
nel ownership of the FPU. During an floating-point state swap, the “FPU owner” variable should
be set to indicate the kernel as the current owner. At the completion of the state swap, the vari-
able should be set to indicate the new owning thread. The numeric exception handler needs to
check the FPU owner and discard any numeric exceptions that occur while the kernel is the FPU
owner. A more general flow for a DNA exception handler that handles this case is shown in Fig-
ure E-5.
Numeric exceptions received while the kernel owns the FPU for a state swap must be discarded
in the kernel without being dispatched to a handler. A flow for a numeric exception dispatch rou-
tine is shown in Figure E-6.
It may at first glance seem that there is a possibility of floating-point exceptions being lost be-
cause of exceptions that are discarded during state swaps. This is not the case, as the exception
will be re-issued when the floating-point state is reloaded. Walking through state swaps both
with and without pending numeric exceptions will clarify the operation of these two handlers.
Case #1: FPU State Swap Without Numeric Exception
Assume two threads A and B, both using the floating-point unit. Let A be the thread to have most
recently executed a floating-point instruction, with no pending numeric exceptions. Let B be the
currently executing thread. CR0.TS was set when thread A was suspended. When B starts to ex-
ecute a floating-point instruction the instruction will fault with the DNA exception because TS
is set.
At this point the handler is entered, and eventually it finds that the current FPU Owner is not the
currently executing thread. To guard the FPU state swap from extraneous numeric exceptions,
the FPU Owner is set to be the kernel. The old owner’s FPU state is saved with FNSAVE, and
the current thread’s FPU state is restored with FRSTOR. Before exiting, the FPU owner is set to
thread B, and the TS bit is cleared.
On exit, thread B resumes execution of the faulting floating-point instruction and continues.
Case #2: FPU State Swap with Discarded Numeric Exception
Again, assume two threads A and B, both using the floating-point unit. Let A be the thread to
have most recently executed a floating-point instruction, but this time let there be a pending nu-
meric exception. Let B be the currently executing thread. When B starts to execute a floating-
point instruction the instruction will fault with the DNA exception and enter the DNA handler.
(If both numeric and DNA exceptions are pending, the DNA exception takes precedence, in or-
der to support handling the numeric exception in its own context.)

E-23
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

DNA Handler Entry

<other handler setup code>

Current Thread
same as
FPU Owner? Yes

No

FPU Owner = Kernel

Use FNSAVE or FXSAVE to Old

Thread’s FP Save Area
(may cause numeric exception)
<handler final cleanup>

Use FRSTOR or FXRSTOR from CLTS (clears CR0.TS)

Current Thread’s FP Save Area

Exit DNA Handler

<other handler code>

FPU Owner = Current Thread

Figure E-5. General Program Flow for DNA Exception Handler

Numeric Exception Entry

Is Kernel
FPU Owner? Yes

No
Normal Dispatch to
Numeric Exception Handler Exit

Figure E-6. Program Flow for a Numeric Exception Dispatch Routine

When the FNSAVE starts, it will trigger an interrupt via FERR# because of the pending numeric
exception. After some system dependent delay, the numeric exception handler is entered. It may

E-24
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

be entered before the FNSAVE starts to execute, or it may be entered shortly after execution of
the FNSAVE. Since the FPU Owner is the kernel, the numeric exception handler simply exits,
discarding the exception. The DNA handler resumes execution, completing the FNSAVE of the
old floating-point context of thread A and the FRSTOR of the floating-point context for thread
B.
Thread A eventually gets an opportunity to handle the exception that was discarded during the
task switch. After some time, thread B is suspended, and thread A resumes execution. When
thread A starts to execute a floating-point instruction, once again the DNA exception handler is
entered. B’s FPU state is stored, and A’s FPU state is restored. Note that in restoring the FPU
state from A’s save area, the pending numeric exception flags are reloaded in to the floating-
point status word. Now when the DNA exception handler returns, thread A resumes execution
of the faulting floating-point instruction just long enough to immediately generate a numeric ex-
ception, which now gets handled in the normal way. The net result is that the task switch and
resulting FPU state swap via the DNA exception handler causes an extra numeric exception
which can be safely discarded.

E.3.5.4. INTERRUPT ROUTING FROM THE KERNEL

In MS-DOS, an application that wishes to handle numeric exceptions hooks interrupt 16 by plac-
ing its handler address in the interrupt vector table, and exiting via a jump to the previous inter-
rupt 16 handler. Protected mode systems that run MS-DOS programs under a subsystem can
emulate this exception delivery mechanism. For example, assume a protected mode O.S. that
runs with CR.NE = 1, and that runs MS-DOS programs in a virtual machine subsystem. The
MS-DOS program is set up in a virtual machine that provides a virtualized interrupt table. The
MS-DOS application hooks interrupt 16 in the virtual machine in the normal way. A numeric
exception will trap to the kernel via the real INT 16 residing in the kernel at ring 0. The INT 16
handler in the kernel then locates the correct MS-DOS virtual machine, and reflects the interrupt
to the virtual machine monitor. The virtual machine monitor then emulates an interrupt by jump-
ing through the address in the virtualized interrupt table, eventually reaching the application’s
numeric exception handler.

E.3.5.5. SPECIAL CONSIDERATIONS FOR OPERATING SYSTEMS THAT

SUPPORT STREAMING SIMD EXTENSIONS
Operating systems that support Streaming SIMD Extensions instructions introduced with the
Pentium® III processor should use the FXSAVE and FXRSTOR instructions to save and restore
the new SIMD floating-point instruction register state as well as the floating-point state. Such
operating systems must consider the following issues:
1. Enlarged state save area: the FNSAVE/FRSTOR instructions operate on a 94-byte or
108-byte memory region, depending on whether they are executed in 16-bit or 32-bit
mode. The FXSAVE/FXRSTOR instructions operate on a 512-byte memory region.
2. Alignment requirements: the FXSAVE/FXRSTOR instructions require the memory
region on which they operate to be 16-byte aligned (refer to the individual instruction
instructions descriptions in Chapter 3, Instruction Set Reference, in the Intel Architecture

E-25
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

Software Developer’s Manual, Volume 2, for information about exceptions generated if the
memory region is not aligned).
3. Maintaining compatibility with legacy applications/libraries: The operating system
changes to support Streaming SIMD Extensions must be invisible to legacy applications or
libraries that deal only with floating-point instructions. The layout of the memory region
operated on by the FXSAVE/FXRSTOR instructions is different from the layout for the
FNSAVE/FRSTOR instructions. Specifically, the format of the FPU tag word and the
length of the various fields in the memory region is different. Care must be taken to return
the FPU state to a legacy application (e.g., when reporting FP exceptions) in the format it
expects.
4. Instruction semantic differences: There are some semantic differences between the way
the FXSAVE and FSAVE/FNSAVE instructions operate. The FSAVE/FNSAVE instruc-
tions clear the FPU after they save the state while the FXSAVE instruction saves the
FPU/Streaming SIMD Extensions state but does not clear it. Operating systems that use
FXSAVE to save the FPU state before making it available for another thread (e.g., during
thread switch time) should take precautions not to pass a “dirty” FPU to another appli-
cation.

E.4. DIFFERENCES FOR HANDLERS USING NATIVE MODE

The 8087 has a pin INT which it asserts when an unmasked exception occurs. But there is no
interrupt input pin in the 8086 or 8088 dedicated to its attachment, nor an interrupt vector num-
ber in the 8086 or 8088 specific for an FPU error assertion. But beginning with the Intel 286 and
Intel 287 hardware connections were dedicated to support the FPU exception, and interrupt vec-
tor 16 assigned to it.

E.4.1. Origin with the Intel 286 and Intel 287, and Intel386™ and
Intel 387 Processors
The Intel 286 and Intel 287, and Intel386™ and Intel 387 processor/coprocessor pairs are each
provided with ERROR# pins that are recommended to be connected between the processor and
FPU. If this is done, when an unmasked FPU exception occurs, the FPU records the exception,
and asserts its ERROR# pin. The processor recognizes this active condition of the ERROR# sta-
tus line immediately before execution of the next WAIT or FPU instruction (except for the no-
wait type) in its instruction stream, and branches to the routine at interrupt vector 16. Thus an
FPU exception will be handled before any other FPU instruction (after the one causing the error)
is executed (except for no-wait instructions, which will be executed without triggering the FPU
exception interrupt, but it will remain pending).
Using the dedicated interrupt 16 for FPU exception handling is referred to as the native mode.
It is the simplest approach, and the one recommended most highly by Intel.

E-26
 GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

E.4.2. Changes with Intel486™, Pentium®, and P6 Family

Processors with CR0.NE=1
With these latest three generations of the IA, more enhancements and speedup features have
been added to the corresponding FPUs. Also, the FPU is now built into the same chip as the pro-
cessor, which allows further increases in the speed at which the FPU can operate as part of the
integrated system. This also means that the native mode of FPU exception handling, selected by
setting bit NE of register CR0 to 1, is now entirely internal.
If an unmasked exception occurs during an FPU instruction, the FPU records the exception in-
ternally, and triggers the exception handler through interrupt 16 immediately before execution
of the next WAIT or FPU instruction (except for no-wait instructions, which will be executed as
described in Section E.4.1., “Origin with the Intel 286 and Intel 287, and Intel386™ and Intel
387 Processors”).
An unmasked numerical exception causes the FERR# output to be activated even with NE=1,
and at exactly the same point in the program flow as it would have been asserted if NE were
zero. However, the system would not connect FERR# to a PIC to generate INTR when operating
in the native, internal mode. (If the hardware of a system has FERR# connected to trigger IRQ13
in order to support MS-DOS, but an O/S using the native mode is actually running the system,
it is the O/S’s responsibility to make sure that IRQ13 is not enabled in the slave PIC.) With this
configuration a system is immune to the problem discussed in Section E.2.1.3., “No-Wait FPU
Instructions Can Get FPU Interrupt in Window”, where for Intel486™ and Pentium® processors
a no-wait FPU instruction can get an FPU exception.

E.4.3. Considerations When FPU Shared Between Tasks Using

Native Mode
The protocols recommended in Section E.3.5., “Considerations When FPU Shared Between
Tasks” for MS-DOS compatibility FPU exception handlers that are shared between tasks may
be used without change with the native mode. However, the protocols for a handler written spe-
cifically for native mode can be simplified, because the problem of a spurious floating-point ex-
ception interrupt occurring while the kernel is executing cannot happen in native mode.
The problem as actually found in practical code in a MS-DOS compatibility system happens
when the DNA handler uses FNSAVE to switch FPU contexts. If an FPU exception is active,
then FNSAVE triggers FERR# briefly, which usually will cause the FPU exception handler to
be invoked inside the DNA handler. In native mode, neither FNSAVE nor any other no-wait in-
structions can trigger interrupt 16. (As discussed above, FERR# gets asserted independent of the
value of the NE bit, but when NE=1, the O/S should not enable its path through the PIC.) An-
other possible (very rare) way a floating-point exception interrupt could occur while the kernel
is executing is by an FPU immediate exception case having its interrupt delayed by the external
hardware until execution has switched to the kernel. This also cannot happen in native mode be-
cause there is no delay through external hardware.
Thus the native mode FPU exception handler can omit the test to see if the kernel is the FPU
owner, and the DNA handler for a native mode system can omit the step of setting the kernel as
the FPU owner at the handler’s beginning. Since however these simplifications are minor and

E-27
GUIDELINES FOR WRITING FPU EXCEPTIONS HANDLERS

save little code, it would be a reasonable and conservative habit (as long as the MS-DOS com-
patibility mode is widely used) to include these steps in all systems.
Note that the special DP (Dual Processing) mode for Pentium® Processors, and also the more
general Intel MultiProcessor Specification for systems with multiple Pentium® or P6 family pro-
cessors, support FPU exception handling only in the native mode. Intel does not recommend us-
ing the MS-DOS compatibility mode for systems using more than one processor.

E-28
 F
Guidelines for
Writing SIMD
Floating-Point
Exception Handlers
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

APPENDIX F
GUIDELINES FOR WRITING SIMD FLOATING-
POINT EXCEPTION HANDLERS

Most of the information on Streaming SIMD Extensions instructions can be found in Chapter 9,
Programming with the Streaming SIMD Extensions. Exceptions in Streaming SIMD Extensions
are specifically presented in Section 9.5.5., “Exception Handling in Streaming SIMD Exten-
sions”
This appendix considers only the Streaming SIMD Extensions instructions that can generate nu-
meric (floating-point) exceptions, and gives an overview of the necessary support for handling
such exceptions. This appendix does not address RSQRTSS, RSQRTPS, RCPSS, RCPPS, or
any unlisted instruction. For detailed information on which instructions generate numeric excep-
tions, and a listing of those exceptions, refer to Appendix D, SIMD Floating-Point Exceptions
Summary. Non-numeric exceptions are handled in a way similar to that for the standard IA-32
instructions.

F.1. TWO OPTIONS FOR HANDLING NUMERIC EXCEPTIONS

Just as for FPU floating-point exceptions, the processor takes one of two possible courses of ac-
tion when a Streaming SIMD Extensions instruction raises a floating-point exception.

• If the exception being raised is masked (by setting the corresponding mask bit in the
MXCSR to 1), then a default result is produced, which is acceptable in most situations. No
external indication of the exception is given, but the corresponding exception flags in the
MXCSR are set, and may be examined later. Note though that for packed operations, an
exception flag that is set in the MXCSR will not tell which of the four sets of sub-operands
caused the event to occur.

• If the exception being raised is not masked (by setting the corresponding mask bit in the
MXCSR to 0), a software exception handler previously registered by the user will be
invoked through the SIMD floating-point exception vector 19. This case is discussed
below in Section F.2., “Software Exception Handling”.

F.2. SOFTWARE EXCEPTION HANDLING

The exception handling routine reached via interrupt vector 19 is usually part of the system soft-
ware (the operating system kernel). Note that an interrupt descriptor table (IDT) entry must have
been previously set up for this vector (refer to Chapter 5, Interrupt and Exception Handling, in
the Intel Architecture Software Developer’s Manual, Volume 3). Some compilers use specific
run-time libraries to assist in floating-point exception handling. If any FPU floating-point oper-
ations are going to be performed that might raise floating-point exceptions, then the exception
handling routine must either disable all floating-point exceptions (for example, loading a local

F-1
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

control word with FLDCW), or it must be implemented as re-entrant (for the case of FPU ex-
ceptions, refer to Example E-5 in Appendix E, Guidelines for Writing FPU Exceptions Han-
dlers). If this is not the case, the routine has to clear the status flags for FPU exceptions, or to
mask all FPU floating-point exceptions. For Streaming SIMD Extensions floating-point excep-
tions though, the exception flags in MXCSR do not have to be cleared, even if they remain un-
masked (they may still be cleared). Exceptions are in this case precise and occur immediately,
and a Streaming SIMD Extensions exception status flag that is set when the corresponding ex-
ception is unmasked will not generate an exception.
Typical actions performed by this low-level exception handling routine are:

• incrementing an exception counter for later display or printing

• printing or displaying diagnostic information (e.g. the MXCSR and XMM registers)

• aborting further execution, or using the exception pointers to build an instruction that will
run without exception and executing it

• storing information about the exception in a data structure that will be passed to a higher
level user exception handler
In most cases (and this applies also to the Streaming SIMD Extensions instructions), there will
be three main components of a low-level floating-point exception handler: a “prologue”, a
“body”, and an “epilogue”.
The prologue performs functions that must be protected from possible interruption by higher-
priority sources - typically saving registers and transferring diagnostic information from the pro-
cessor to memory. When the critical processing has been completed, the prologue may re-enable
interrupts to allow higher-priority interrupt handlers to preempt the exception handler (assuming
that the interrupt handler was called through an interrupt gate, meaning that the processor
cleared the interrupt enable (IF) flag in the EFLAGS register - refer to Section 4.4.1., “Call and
Return Operation for Interrupt or Exception Handling Procedures” in Chapter 4, Procedure
Calls, Interrupts, and Exceptions).
The body of the exception handler examines the diagnostic information and makes a response
that is application-dependent. It may range from halting execution, to displaying a message, to
attempting to fix the problem and then proceeding with normal execution, to setting up a data
structure, calling a higher-level user exception handler and continuing execution upon return
from it. This latter case will be assumed in Section F.4., “SIMD Floating-Point Exceptions and
the IEEE-754 Standard for Binary Floating-Point Computations” below.
Finally, the epilogue essentially reverses the actions of the prologue, restoring the processor
state so that normal execution can be resumed.
The following example represents a typical exception handler. To link it with Example F-2 that
will follow in Section F.4.3., “SIMD Floating-Point Emulation Implementation Example”, as-
sume that the body of the handler (not shown here in detail) passes the saved state to a routine
that will examine in turn all the sub-operands of the excepting instruction, invoking a user float-
ing-point exception handler if a particular set of sub-operands raises an unmasked (enabled) ex-
ception, or emulating the instruction otherwise.

F-2
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Example F-1. SIMD Floating-Point Exception Handler

SIMD_FP_EXC_HANDLER PROC
;
;;; PROLOGUE
; SAVE REGISTERS
PUSH EBP ; SAVE EBP
PUSH EAX ; SAVE EAX
...
MOV EBP, ESP ; SAVE ESP in EBP
SUB ESP, 512 ; ALLOCATE 512 BYTES
AND ESP, 0fffffff0h ; MAKE THE ADDRESS 16-BYTE ALIGNED
FXSAVE [ESP] ; SAVE FP, MMX, AND SIMD FP STATE
PUSH [EBP+EFLAGS_OFFSET] ; COPY OLD EFLAGS TO STACK TOP
POPD ;RESTORE THE INTERRUPT ENABLE FLAG IF
;TO VALUE BEFORE SIMD FP EXCEPTION
;
;;; BODY
; APPLICATION-DEPENDENT EXCEPTION HANDLING CODE GOES HERE
LDMXCSR LOCAL_MXCSR ; LOAD LOCAL FPU CW IF NEEDED
...
...
;
;;; EPILOGUE
FXRSTOR [ESP] ; RESTORE MODIFIED STATE IMAGE
MOV ESP, EBP ; DE-ALLOCATE STACK SPACE
...
POP EAX ; RESTORE EAX
POP EBP ; RESTORE EBP
IRET ; RETURN TO INTERRUPTED CALCULATION
SIMD_FP_EXC_HANDLER ENDP

F.3. EXCEPTION SYNCHRONIZATION

A Streaming SIMD Extensions instruction can execute in parallel with other similar instruc-
tions, with integer instructions, and with floating-point or MMX™ instructions. Exception syn-
chronization may therefore be necessary, similarly to the situation described in Section E.3.3.,
“Synchronization Required for Use of FPU Exception Handlers” in Appendix E, Guidelines for
Writing FPU Exceptions Handlers). Careful coding will ensure proper synchronization in case
a floating-point exception handler is invoked, and will lead to reliable performance.

F-3
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

F.4. SIMD FLOATING-POINT EXCEPTIONS AND THE IEEE-754

STANDARD FOR BINARY FLOATING-POINT
COMPUTATIONS
The Streaming SIMD Extensions are 100% compatible with the ANSI/IEEE Standard 754-
1985, IEEE Standard for Binary Floating-Point Arithmetic, satisfying all of its mandatory re-
quirements (when the flush-to-zero mode is not enabled). But a programming environment that
includes the Streaming SIMD Extensions instructions will comply with both the obligatory and
the strongly recommended requirements of the IEEE Standard 754 regarding floating-point ex-
ception handling, only as a combination of hardware and software (which is acceptable). The
standard states that a user should be able to request a trap on any of the five floating-point ex-
ceptions (note that the denormal exception is an IA addition), and it also specifies the values (op-
erands or result) to be delivered to the exception handler.
The main issue is that for Streaming SIMD Extensions instructions that raise post-computation
exceptions (traps: overflow, underflow, or inexact), unlike for IA-32 FPU instructions, the pro-
cessor does not provide the result recommended by the IEEE standard to the user handler. If a
user program needs the result of an instruction that generated a post-computation exception, it
is the responsibility of the software to produce this result by emulating the faulting Streaming
SIMD Extensions instruction. Another issue is that the standard does not specify explicitly how
to handle multiple floating-point exceptions that occur simultaneously. For packed operations,
a logical OR of the flags that would be set by each sub-operation is used to set the exception
flags in the MXCSR. The following subsections present one possible way to solve these prob-
lems.

F.4.1. Floating-Point Emulation

Every operating system must provide a kernel level floating-point exception handler (a template
was presented in Section F.2., “Software Exception Handling” above). In the following, assume
that a user mode floating-point exception filter is supplied for Streaming SIMD Extensions ex-
ceptions (for example as part of a library of C functions), that a user program can invoke in order
to handle unmasked exceptions. The user mode floating-point exception filter (not shown here)
has to be able to emulate the subset of Streaming SIMD Extensions instructions that can gener-
ate numeric exceptions, and has to be able to invoke a user provided floating-point exception
handler for floating-point exceptions. When a floating-point exception that is not masked is
raised by a Streaming SIMD Extensions instruction, the low-level floating-point exception han-
dler will be called. This low-level handler may in turn call the user mode floating-point excep-
tion filter. The filter function receives the original operands of the excepting instruction, as no
results are provided by the hardware, whether a pre-computation or a post-computation excep-
tion has occurred. The filter will unpack the operands into up to four sets of sub-operands, and
will submit them one set at a time to an emulation function (that will be presented in Example
F-2 in Section F.4.3., “SIMD Floating-Point Emulation Implementation Example”, below). The
emulation function will examine the sub-operands, and will possibly redo the necessary calcu-
lation.

F-4
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Two cases are possible:

• If an unmasked (enabled) exception occurs in this process, the emulation function will
return to its caller (the filter function) with the appropriate information. The filter will
invoke a (previously registered) user floating-point exception handler for this set of sub-
operands, and will record the result upon return from the user handler (provided the user
handler allows continuation of the execution).

• If no unmasked (enabled) exception occurs, the emulation function will determine and will
return to its caller the result of the operation for the current set of sub-operands (it has to
be IEEE compliant). The filter function will record the result (plus any new flag settings).
The user level filter function will then call the emulation function for the next set of sub-oper-
ands (if any). When done, the partial results will be packed (if the excepting instruction has a
packed floating-point result, which is true for most Streaming SIMD Extensions numeric in-
structions) and the filter will return to the low-level exception handler, which in turn will return
from the interruption, allowing execution to continue. Note that the instruction pointer (EIP) has
to be altered to point to the instruction following the excepting instruction, in order to continue
execution correctly.
If a user mode floating-point exception filter is not provided, then all the work for decoding the
excepting instruction, reading its operands, emulating the instruction for the components of the
result that do not correspond to unmasked floating-point exceptions, and providing the com-
pounded result will have to be performed by the user provided floating-point exception handler.
Actual emulation will have to take place for one operand or pair of operands for scalar opera-
tions, and for all four operands or pairs of operands for packed operations. The steps to perform
are the following:

• the excepting instruction has to be decoded and the operands have to be read from the
saved context

• the instruction has to be emulated for each (pair of) sub-operand(s); if no floating-point
exception occurs, the partial result has to be saved; if a masked floating-point exception
occurs, the masked result has to be produced through emulation and saved, and the
appropriate status flags have to be set; if an unmasked floating-point exception occurs, the
result has to be generated by the user provided floating-point exception handler, and the
appropriate status flags have to be set

• the four partial results have to be combined and written to the context that will be restored
upon application program resumption

F-5
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

A diagram of the control flow in handling an unmasked floating-point exception is presented

below.

User Application

Low-Level Floating-Point Exception Handler

User Level Floating-Point Exception Filter

User Floating-Point Exception Handler

Figure F-1. Control Flow for Handling Unmasked Floating-Point Exceptions

From the user level floating-point filter, Example F-2 in Section F.4.3., “SIMD Floating-Point
Emulation Implementation Example”will present only the floating-point emulation part. In or-
der to understand the actions involved, the expected response to exceptions has to be known for
all the Streaming SIMD Extensions numeric instructions in two situations: with exceptions en-
abled (unmasked result), and with exceptions disabled (masked result). The latter can be found
in Section 4.4., “Interrupts and Exceptions”, in Chapter 4, Procedure Calls, Interrupts, and Ex-
ceptions. The response to NaN operands that do not raise an exception is specified in Section
9.1.6., “SIMD Floating-Point Register Data Formats”. Operating on NaNs from the same
source. It is also given in more detail in the next subsection, along with the unmasked and
masked responses to floating-point exceptions.

F.4.2. Streaming SIMD Extensions Response To Floating-Point

Exceptions
This subsection specifies the unmasked response expected from the Streaming SIMD Exten-
sions instructions that raise floating-point exceptions. The masked response is given in parallel,
as it is necessary in the emulation process of the instructions that raise unmasked floating-point
exceptions. The response to NaN operands is also included in more detail than in Section 9.1.6.,
“SIMD Floating-Point Register Data Formats”. For floating-point exception priority, refer to

F-6
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Section 5.7., “Priority Among Simultaneous Exceptions and Interrupts” in Chapter 5, Interrupt
and Exception Handling.
Note that some floating-point instructions (non-waiting instructions) do not check for pending
unmasked exceptions (refer to Section 7.5.11., “FPU Control Instructions”, in Chapter 7,
Floating-Point Unit). They include the FNINIT, FNSTENV, FNSAVE, FNSTSW, FNSTCW,
and FNCLEX instructions. When an FNINIT, FNSTENV, FNSAVE, or FNCLEX instruction is
executed, all pending exceptions are essentially lost (either the FPU status register is cleared or
all exceptions are masked). The FNSTSW and FNSTCW instructions do not check for pending
interrupts, but they do not modify the FPU status and control registers. A subsequent “waiting”
floating-point instruction can then handle any pending exceptions.

F.4.2.1. NUMERIC EXCEPTIONS

There are six classes of numeric (floating-point) exception conditions that can occur: Invalid op-
eration (#I), Divide-by-Zero (#Z), Denormal Operand (#D), Numeric Overflow (#O), Numeric
Underflow (#U), and Inexact Result (precision) (#P). #I, #Z, #D are pre-computation exceptions
(floating-point faults), detected before the arithmetic operation. #O, #U, #P are post-computa-
tion exceptions (floating-point traps).
Users can control how the exceptions are handled by setting the mask/unmask bits in MXCSR.
Masked exceptions are handled by the processor or by software if they are combined with un-
masked exceptions occurring in the same instruction. Unmasked exceptions are usually handled
by the low-level exception handler, in conjunction with user-level software.

F.4.2.2. RESULTS OF OPERATIONS WITH NAN OPERANDS OR A NAN

RESULT FOR STREAMING SIMD EXTENSIONS NUMERIC
INSTRUCTIONS
The tables below specify the response of the Streaming SIMD Extensions technology instruc-
tions to NaN inputs, or to other inputs that lead to NaN results.
These results will be referenced by subsequent tables. Most operations do not raise an invalid
exception for quiet NaN operands, but even so, they will have higher precedence over raising
some exception.
Note that the single-precision QNaN Indefinite value is 0xffc00000, and the Integer Indefinite
value is 0x80000000 (not a floating-point number, but it can be the result of a conversion in-
struction from floating-point to integer).
For an unmasked exception, no result will be provided to the user handler. If a user registered
floating-point exception handler is invoked, it may provide a result for the excepting instruction,
that will be used if execution of the application code is continued after returning from the inter-
ruption.
In Tables F-1 through Table F-10, the specified operands cause an invalid exception, unless the
unmasked result is marked with ‘(not an exception)’. In this latter case, the unmasked and
masked results are the same.

F-7
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Table F-1. ADDPS, ADDSS, SUBPS, SUBSS, MULPS, MULSS, DIVPS, DIVSS
Source Operands Masked Result Unmasked Result
SNaN1 op SNaN2 SNaN1 | 0x00400000 None
SNaN1 op QNaN2 SNaN1 | 0x00400000 None
QNaN1 op SNaN2 QNaN1 None
QNaN1 op QNaN2 QNaN1 QNaN1 (not an exception)
SNaN op real value SNaN | 0x00400000 None
Real value op SNaN SNaN | 0x00400000 None
QNaN op real value QNaN QNaN (not an exception)
Real value op QNaN QNaN QNaN (not an exception)
Neither source operand is SNaN, Single-Precision QNaN Indefinite None
but #I is signaled (e.g. for Inf - Inf,
Inf * 0, Inf / Inf, 0/0)

Note 1. SNaN | 0x00400000 is a quiet NaN obtained from the signaling NaN given as input
Note 2. Operations involving only quiet NaNs do not raise a floating-point exception

Table F-2. CMPPS.EQ, CMPSS.EQ, CMPPS.ORD, CMPSS.ORD

Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) 0x00000000 0x00000000 (not an exception)
Opd1 op NaN (any Opd1) 0x00000000 0x00000000 (not an exception)

Table F-3. CMPPS.NEQ, CMPSS.NEQ, CMPPS.UNORD, CMPSS.UNORD

Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) 0x11111111 0x11111111 (not an exception)
Opd1 op NaN (any Opd1) 0x11111111 0x11111111 (not an exception)

Table F-4. CMPPS.LT, CMPSS.LT, CMPPS.LE, CMPSS.LE

Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) 0x00000000 None
Opd1 op NaN (any Opd1) 0x00000000 None

Table F-5. CMPPS.NLT, CMPSS.NLT, CMPSS.NLT, CMPSS.NLE

Source Operands Masked Result Unmasked Result
NaN op Opd2 (any Opd2) 0x11111111 None
Opd1 op NaN (any Opd1) 0x11111111 None

F-8
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Table F-6. COMISS

Source Operands Masked Result Unmasked Result
SNaN op Opd2 (any Opd2) OF,SF,AF=000 ZF,PF,CF=111 None
Opd1 op SNaN (any Opd1) OF,SF,AF=000 ZF,PF,CF=111 None
QNaN op Opd2 (any Opd2) OF,SF,AF=000 ZF,PF,CF=111 None
Opd1 op QNaN (any Opd1) OF,SF,AF=000 ZF,PF,CF=111 None

Table F-7. UCOMISS

Source Operands Masked Result Unmasked Result
SNaN op Opd2 (any Opd2) OF,SF,AF=000 ZF,PF,CF=111 None
Opd1 op SNaN (any Opd1) OF,SF,AF=000 ZF,PF,CF=111 None
QNaN op Opd2
(any Opd2 ≠ SNaN) OF,SF,AF=000 ZF,PF,CF=111 OF,SF,AF=000 ZF,PF,CF=111
(not an exception)
Opd1 op QNaN
(any Opd1 ≠ SNaN) OF,SF,AF=000 ZF,PF,CF=111 OF,SF,AF=000 ZF,PF,CF=111
(not an exception)

Table F-8. CVTPS2PI, CVTSS2SI, CVTTPS2PI, CVTTSS2SI

Source Operand Masked Result Unmasked Result
SNaN 0x80000000 (Integer Indefinite) None
QNaN 0x80000000 (Integer Indefinite) None

Table F-9. MAXPS, MAXSS, MINPS, MINSS

Source Operands Masked Result Unmasked Result
Opd1 op NaN2 (any Opd1) NaN2 None
NaN1 op Opd2 (any Opd2) Opd2 None

Note: SNaN and QNaN operands raise an Invalid Operand fault

Table F-10. SQRTPS, SQRTSS

Source Operand Masked Result Unmasked Result
QnaN QNaN QNaN (not an exception)
SNaN SNaN | 0x00400000 None
Source operand is not SNaN, but
#I is signaled
(e.g. for sqrt (-1.0)) Single-Precision QNaN Indefinite None

Note: SNaN | 0x00400000 is a quiet NaN obtained from the signaling NaN given as input

F-9
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

F.4.2.3. CONDITION CODES, EXCEPTION FLAGS, AND RESPONSE FOR

MASKED AND UNMASKED NUMERIC EXCEPTIONS
In the following, the masked response is what the processor provides when a masked exception
is raised by a Streaming SIMD Extensions numeric instruction. The same response is provided
by the floating-point emulator for Streaming SIMD Extensions numeric instructions, when cer-
tain components of the quadruple input operands generate exceptions that are masked (the em-
ulator also generates the correct answer, as specified by the IEEE standard wherever applicable,
in the case when no floating-point exception occurs). The unmasked response is what the emu-
lator provides to the user handler for those components of the quadruple input operands of the
Streaming SIMD Extensions instructions that raise unmasked exceptions. Note that for pre-
computation exceptions (floating-point faults), no result is provided to the user handler. For
post-computation exceptions (floating-point traps), a result is also provided to the user handler,
as specified below.
In the following tables, the result is denoted by ’res’, with the understanding that for the actual
instruction, the destination coincides with the first source operand (except for COMISS and
UCOMISS, whose destination is the EFLAGS register).

Table F-11. #I - Invalid Operations

Instruction Condition Masked Response Unmasked
Response and
Exception Code
ADDPS src1 or src2 = SNaN Refer to Table F-1 for NaN src1, src2
operands, #IA=1 unchanged, #IA=1
ADDSS src1=+Inf, src2 = -Inf or res = QNaN Indefinite,
src1=-Inf, src2 = +Inf #IA=1
SUBPS src1 or src2 = SNaN Refer to Table F-1 for NaN src1, src2
operands, #IA=1 unchanged, #IA=1
SUBSS src1=+Inf, src2 = +Inf or res = QNaN Indefinite,
src1=-Inf, src2 = -Inf #IA=1
MULPS src1 or src2 = SNaN Refer to Table F-1 for NaN src1, src2
operands, #IA=1 unchanged, #IA=1
MULSS src1=±Inf, src2 = ±0 or res = QNaN Indefinite,
src1=±0, src2 = ±Inf #IA=1
DIVPS src1 or src2 = SNaN Refer to Table F-1 for NaN src1, src2
operands, #IA=1 unchanged, #IA=1
DIVSS src1=±Inf, src2 = ±Inf or res = QNaN Indefinite,
src1=±0, src2 = ±0 #IA=1

F-10
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Table F-11. #I - Invalid Operations

Instruction Condition Masked Response Unmasked
Response and
Exception Code
SQRTPS src = SNaN Refer to Table F-10 for NaN src unchanged,
operands, #IA=1 #IA=1
SQRTSS src < 0 res = QNaN Indefinite,
(note that -0 < 0 is false) #IA=1
MAXPS src1 = NaN or src2 = NaN res = src2, #IA=1 src1, src2
MAXSS unchanged, #IA=1
MINP src1 = NaN or src2 = NaN res = src2, #IA=1 src1, src2
MINSS unchanged, #IA=1
CMPPS.LT src1 = NaN or src2 = NaN Refer to Table F-4 and src1, src2
CMPPS.LE Table F-5 for NaN unchanged, #IA=1
CMPPS.NLT operands, #IA=1
CMPPS.NLE
CMPSS.LT
CMPSS.LE
CMPSS.NLT
CMPSS.NLE
COMISS src1 = NaN or src2 = NaN Refer to Table F-6 for NaN src1, src2, EFLAGS
operands unchanged,#IA=1
UCOMISS src1 = SNaN or src2 = SNaN Refer to Table F-7 for NaN src1, src2, EFLAGS
operands unchanged,#IA=1
CVTPS2PI src = NaN, ±Inf, res = Integer Indefinite src unchanged,
CVTSS2SI |(src)rnd | > 0x7fffffff #IA=1 #IA=1
CVTTPS2PI src = NaN, ±Inf, res = Integer Indefinite src unchanged,
CVTTSS2SI |(src)rz | > 0x7fffffff #IA=1 #IA=1

Note 1. rnd signifies the user rounding mode from MXCSR, and rz signifies the rounding mode toward zero
(truncate), when rounding a floating-point value to an integer. For more information, refer to Table 9-3 in
Section 9.1.8., “Rounding Control Field”, of Chapter 9, Programming with the Streaming SIMD Exten-
sions.
Note 2. For NAN encodings, see Table 9-2, Chapter 9, Programming with the Streaming SIMD Extensions.

Table F-12. #Z - Divide-by-Zero

Instruction Condition Masked Response Unmasked
Response and
Exception Code
DIVPS src1 = finite non-zero (normal, or res = ±Inf src1, src2
DIVSS denormal) #ZE=1 unchanged, #ZE=1
src2 = ±0

F-11
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Table F-13. #D - Denormal Operand

Instruction Condition Masked Response Unmasked Response and
Exception Code
ADDPS src1 = denormal or res = result rounded to the src1, src2 unchanged,
SUBPS src2 = denormal destination precision and #DE=1
MULPS #DE=1 using the bounded
DIVPS exponent, but only if no
SQRTPS unmasked post-
MAXPS computation exception (SQRT only has 1 src)
MINPS occurs
CMPPS
ADDSS
SUBSS
MULSS
DIVSS
SQRTSS
MAXSS
MINSS
CMPSS
COMISS
UCOMISS

Note: For denormal encodings, see Table 9-2, Chapter 9, Programming with the Streaming SIMD Exten-
sions.

Table F-14. #0 - Numeric Overflow

Instruction Condition Masked Response Unmasked Response and
Exception Code
ADDPS rounded result Rounding Sign Result & Status Flags res = (result calculated with
SUBPS > largest single- unbounded exponent and
MULPS precision finite To + #OE=1, #PE=1 rounded to the destination
DIVPS normal value nearest - res = + ∞ precision) / 2192
ADDSS res = – ∞ #OE=1
SUBSS #PE=1 if the result is
MULSS Toward + #OE=1, #PE=1 inexact
DIVSS –∞ - res = 1.11…1 * 2127
res = – ∞
Toward + #OE=1, #PE=1
+∞ - res = + ∞
res = -1.11…1 * 2127
Toward + #OE=1, #PE=1
0 - res = 1.11…1 * 2127
res = -1.11…1 * 2127

F-12
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

Table F-15. #U - Numeric Underflow

Instruction Condition Masked Response Unmasked Response
and Exception Code
ADDPS result calculated with #UE=1 and #PE=1, res = (result calculated with
SUBPS unbounded exponent and but only if the result is unbounded exponent and
MULPS rounded to the destination exact rounded to the destination
DIVPS precision < smallest single- precision) * 2192
ADDSS precision finite normal value res = ±0, denormal, or #UE=1
SUBSS normal #PE=1 if the result is
MULSS inexact
DIVSS

Table F-16. #P - Inexact Result (Precision)

Instruction Condition Masked Response Unmasked Response and Exception
Code
ADDPS the result is not res = result rounded only if no underflow/overflow condition
SUBPS exactly to the destination occurred, or if the corresponding
MULPS representable in the precision and using exceptions are masked:
DIVPS destination format the bounded set #OE if masked overflow and set result
SQRTPS exponent, but only if as described above for masked overflow;
CVTPI2PS no unmasked set #UE if masked underflow and set
CVTPS2PI underflow or overflow result as described above for masked
CVTTPS2PI conditions occur (This underflow;
ADDSS exception can occur if neither underflow nor overflow, res = the
SUBSS in the presence of a result rounded to the destination precision
MULSS masked underflow or and using the bounded exponent
DIVSS overflow) set #PE=1
SQRTSS #PE=1
CVTSI2SS
CVTSS2SI
CVTTSS2SI

F.4.3. SIMD Floating-Point Emulation Implementation Example

The sample code listed below may be considered as being part of a user-level floating-point ex-
ception filter for Streaming SIMD Extensions numeric instructions. It is assumed that the filter
function is invoked by a low-level exception handler (reached via interrupt vector 19 when an
unmasked floating-point exception occurs), and that it operates as explained in Section F.4.1.,
“Floating-Point Emulation” The sample code does the emulation for the add, subtract, multiply,
and divide operations. For this, it uses C code and IA-32 FPU operations (readability, and not
efficiency was the primary goal). Operations corresponding to other Streaming SIMD Exten-
sions numeric instructions have to be emulated, but only place holders for them are included.
The example assumes that the emulation function receives a pointer to a data structure specify-
ing a number of input parameters: the operation that caused the exception, a set of two sub-op-
erands (unpacked, of type float), the rounding mode (the precision is always single), exception
masks (having the same relative bit positions as in the MXCSR but starting from bit 0 in an un-
signed integer), and a flush-to-zero indicator. The output parameters are a floating-point result
(of type float), the cause of the exception (identified by constants not explicitly defined below),
and the exception status flags. The corresponding C definition is:

F-13
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

typedef struct {
unsigned int operation; // Streaming SIMD Extensions operation: ADDPS, ADDSS, ...
float operand1_fval; // first operand value
float operand2_fval; // second operand value (if any)
float result_fval; // result value (if any)
unsigned int rounding_mode; // rounding mode
unsigned int exc_masks; // exception masks, in the order P, U, O, Z, D, I
unsigned int exception_cause; // exception cause
unsigned int status_flag_inexact; // inexact status flag
unsigned int status_flag_underflow; // underflow status flag
unsigned int status_flag_overflow; // overflow status flag
unsigned int status_flag_divide_by_zero; // divide by zero status flag
unsigned int status_flag_denormal_operand; // denormal operand status flag
unsigned int status_flag_invalid_operation; // invalid operation status flag
unsigned int ftz; // flush-to-zero flag
} EXC_ENV;
The arithmetic operations exemplified are emulated as follows:
1. Perform the operation using IA-32 FPU instructions, with exceptions disabled, the original
user rounding mode, and single precision; this will reveal invalid, denormal, or divide-by-
zero exceptions (if there are any); store the result in memory as a double precision value
(whose exponent range is large enough to look like “unbounded” to the result of the single
precision computation).
2. If no unmasked exceptions were detected, determine if the result is tiny (less than the
smallest normal number that can be represented in single precision format), or huge
(greater than the largest normal number that can be represented in single precision format);
if an unmasked overflow or underflow occur, calculate the scaled result that will be handed
to the user exception handler, as specified by the IEEE-754 Standard for Binary Floating-
Point Computations.
3. If no exception was raised above, calculate the result with “bounded” exponent; if the
result was tiny, it will require denormalization (shifting right the significand while incre-
menting the exponent to bring it into the admissible range of [-126,+127] for single
precision floating-point numbers); the result obtained in step A above cannot be used
because it might incur a double rounding error (it was rounded to 24 bits in step A, and
might have to be rounded again in the denormalization process); the way to overcome this
is to calculate the result as a double precision value, and then to store it to memory in
single precision format - rounding first to 53 bits in the significand, and then to 24 will
never cause a double rounding error (exact properties exist that state when double-
rounding error does not occur, but for the elementary arithmetic operations, the rule of
thumb is that if we round an infinitely precise result to 2p+1 bits and then again to p bits,
the result is the same as when rounding directly to p bits, which means that no double
rounding error occurs).
4. If the result is inexact and the inexact exceptions are unmasked, the result calculated in
step C will be delivered to the user floating-point exception handler.
5. Finally, the flush-to-zero case is dealt with if the result is tiny.

F-14
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

The emulation function returns RAISE_EXCEPTION to the filter function if an exception has
to be raised (the exception_cause field will indicate the cause); otherwise, the emulation func-
tion returns DO_NOT_ RAISE_EXCEPTION. In the first case, the result will be provided by
the user exception handler called by the filter function. In the second case, it is provided by the
emulation function. The filter function has to collect all the partial results, and to assemble the
scalar or packed result that will be used if execution is to be continued.

Example F-2. SIMD Floating-Point Emulation

// masks for individual status word bits
#define PRECISION_MASK 0x20
#define UNDERFLOW_MASK 0x10
#define OVERFLOW_MASK 0x08
#define ZERODIVIDE_MASK 0x04
#define DENORMAL_MASK 0x02
#define INVALID_MASK 0x01

// 32-bit constants
static unsigned ZEROF_ARRAY[] = {0x00000000};
#define ZEROF *(float *) ZEROF_ARRAY
// +0.0
static unsigned NZEROF_ARRAY[] = {0x80000000};
#define NZEROF *(float *) NZEROF_ARRAY
// -0.0
static unsigned POSINFF_ARRAY[] = {0x7f800000};
#define POSINFF *(float *)POSINFF_ARRAY
// +Inf
static unsigned NEGINFF_ARRAY[] = {0xff800000};
#define NEGINFF *(float *)NEGINFF_ARRAY
// -Inf

// 64-bit constants
static unsigned MIN_SINGLE_NORMAL_ARRAY [] = {0x00000000, 0x38100000};
#define MIN_SINGLE_NORMAL *(double *)MIN_SINGLE_NORMAL_ARRAY
// +1.0 * 2^-126
static unsigned MAX_SINGLE_NORMAL_ARRAY [] = {0x70000000, 0x47efffff};
#define MAX_SINGLE_NORMAL *(double *)MAX_SINGLE_NORMAL_ARRAY
// +1.1...1*2^127
static unsigned TWO_TO_192_ARRAY[] = {0x00000000, 0x4bf00000};
#define TWO_TO_192 *(double *)TWO_TO_192_ARRAY
// +1.0 * 2^192
static unsigned TWO_TO_M192_ARRAY[] = {0x00000000, 0x33f00000};
#define TWO_TO_M192 *(double *)TWO_TO_M192_ARRAY
// +1.0 * 2^-192

// auxiliary functions
static int isnanf (float f); // returns 1 if f is a NaN, and 0 otherwise

F-15
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

static float quietf (float f); // converts a signaling NaN to a quiet NaN, and
// leaves a quiet NaN unchanged

// emulation of Streaming SIMD Extensions instructions using

// C code and IA-32 FPU instructions

unsigned int
simd_fp_emulate (EXC_ENV *exc_env)

float opd1; // first operand of the add, subtract, multiply, or divide

float opd2; // second operand of the add, subtract, multiply, or divide
float res; // result of the add, subtract, multiply, or divide
double dbl_res24; // result with 24-bit significand, but "unbounded" exponent
// (needed to check tininess, to provide a scaled result to
// an underflow/overflow trap handler, and in flush-to-zero mode)
double dbl_res; // result in double precision format (needed to avoid a
// double rounding error when denormalizing)
unsigned int result_tiny;
unsigned int result_huge;
unsigned short int sw; // 16 bits
unsigned short int cw; // 16 bits

// have to check first for faults (V, D, Z), and then for traps (O, U, I)

// initialize FPU (floating-point exceptions are masked)

_asm {
fninit;
}

result_tiny = 0;
result_huge = 0;

switch (exc_env->operation) {

case ADDPS:
case ADDSS:
case SUBPS:
case SUBSS:
case MULPS:
case MULSS:
case DIVPS:
case DIVSS:

F-16
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

opd1 = exc_env->operand1_fval;
opd2 = exc_env->operand2_fval;

// execute the operation and check whether the invalid, denormal, or

// divide by zero flags are set and the respective exceptions enabled

// set control word with rounding mode set to exc_env->rounding_mode,

// single precision, and all exceptions disabled
switch (exc_env->rounding_mode) {
case ROUND_TO_NEAREST:
cw = 0x003f; // round to nearest, single precision, exceptions masked
break;
case ROUND_DOWN:
cw = 0x043f; // round down, single precision, exceptions masked
break;
case ROUND_UP:
cw = 0x083f; // round up, single precision, exceptions masked
break;
case ROUND_TO_ZERO:
cw = 0x0c3f; // round to zero, single precision, exceptions masked
break;
default:
;
}
__asm {
fldcw WORD PTR cw;
}

// compute result and round to the destination precision, with

// "unbounded" exponent (first IEEE rounding)
switch (exc_env->operation) {

case ADDPS:
case ADDSS:
// perform the addition
__asm {
fnclex;
// load input operands
fld DWORD PTR opd1; // may set the denormal or invalid status flags
fld DWORD PTR opd2; // may set the denormal or invalid status flags
faddp st(1), st(0); // may set the inexact or invalid status flags
// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

case SUBPS:

F-17
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

case SUBSS:
// perform the subtraction
__asm {
fnclex;
// load input operands
fld DWORD PTR opd1; // may set the denormal or invalid status flags
fld DWORD PTR opd2; // may set the denormal or invalid status flags
fsubp st(1), st(0); // may set the inexact or invalid status flags
// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

case MULPS:
case MULSS:
// perform the multiplication
__asm {
fnclex;
// load input operands
fld DWORD PTR opd1; // may set the denormal or invalid status flags
fld DWORD PTR opd2; // may set the denormal or invalid status flags
fmulp st(1), st(0); // may set the inexact or invalid status flags
// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

case DIVPS:
case DIVSS:
// perform the division
__asm {
fnclex;
// load input operands
fld DWORD PTR opd1; // may set the denormal or invalid status flags
fld DWORD PTR opd2; // may set the denormal or invalid status flags
fdivp st(1), st(0); // may set the inexact, divide by zero, or
// invalid status flags
// store result
fstp QWORD PTR dbl_res24; // exact
}
break;

default:
; // will never occur

F-18
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

// read status word

__asm {
fstsw WORD PTR sw;
}

if (sw & ZERODIVIDE_MASK)

sw = sw & ~DENORMAL_MASK; // clear D flag for (denormal / 0)

// if invalid flag is set, and invalid exceptions are enabled, take trap
if (!(exc_env->exc_masks & INVALID_MASK) && (sw & INVALID_MASK)) {
exc_env->status_flag_invalid_operation = 1;
exc_env->exception_cause = INVALID_OPERATION;
return (RAISE_EXCEPTION);
}

// checking for NaN operands has priority over denormal exceptions; also fix for the
// differences in treating two NaN inputs between the Streaming SIMD Extensions
// instructions and other IA-32 instructions
if (isnanf (opd1) || isnanf (opd2)) {

if (isnanf (opd1) && isnanf (opd2))

exc_env->result_fval = quietf (opd1);
else
exc_env->result_fval = (float)dbl_res24; // exact

if (sw & INVALID_MASK) exc_env->status_flag_invalid_operation = 1;

return (DO_NOT_RAISE_EXCEPTION);
}

// if denormal flag is set, and denormal exceptions are enabled, take trap
if (!(exc_env->exc_masks & DENORMAL_MASK) && (sw & DENORMAL_MASK)) {
exc_env->status_flag_denormal_operand = 1;
exc_env->exception_cause = DENORMAL_OPERAND;
return (RAISE_EXCEPTION);
}

// if divide by zero flag is set, and divide by zero exceptions are

// enabled, take trap (for divide only)
if (!(exc_env->exc_masks & ZERODIVIDE_MASK) && (sw & ZERODIVIDE_MASK)) {
exc_env->status_flag_divide_by_zero = 1;
exc_env->exception_cause = DIVIDE_BY_ZERO;
return (RAISE_EXCEPTION);
}

// done if the result is a NaN (QNaN Indefinite)

res = (float)dbl_res24;
if (isnanf (res)) {

F-19
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

exc_env->result_fval = res; // exact

exc_env->status_flag_invalid_operation = 1;
return (DO_NOT_RAISE_EXCEPTION);
}

// dbl_res24 is not a NaN at this point

if (sw & DENORMAL_MASK) exc_env->status_flag_denormal_operand = 1;

// Note: (dbl_res24 == 0.0 && sw & PRECISION_MASK) cannot occur

if (-MIN_SINGLE_NORMAL < dbl_res24 && dbl_res24 < 0.0 ||
0.0 < dbl_res24 && dbl_res24 < MIN_SINGLE_NORMAL) {
result_tiny = 1;
}

// check if the result is huge

if (NEGINFF < dbl_res24 && dbl_res24 < -MAX_SINGLE_NORMAL ||
MAX_SINGLE_NORMAL < dbl_res24 && dbl_res24 < POSINFF) {
result_huge = 1;
}

// at this point, there are no enabled I, D, or Z exceptions; the instr.

// might lead to an enabled underflow, enabled underflow and inexact,
// enabled overflow, enabled overflow and inexact, enabled inexact, or
// none of these; if there are no U or O enabled exceptions, re-execute
// the instruction using IA-32 double precision format, and the
// user’s rounding mode; exceptions must have been disabled before calling
// this function; an inexact exception may be reported on the 53-bit
// fsubp, fmulp, or on both the 53-bit and 24-bit conversions, while an
// overflow or underflow (with traps disabled) may be reported on the
// conversion from dbl_res to res

// check whether there is an underflow, overflow, or inexact trap to be

// taken

// if the underflow traps are enabled and the result is tiny, take
// underflow trap
if (!(exc_env->exc_masks & UNDERFLOW_MASK) && result_tiny) {
dbl_res24 = TWO_TO_192 * dbl_res24; // exact
exc_env->status_flag_underflow = 1;
exc_env->exception_cause = UNDERFLOW;
exc_env->result_fval = (float)dbl_res24; // exact
if (sw & PRECISION_MASK) exc_env->status_flag_inexact = 1;
return (RAISE_EXCEPTION);
}

// if overflow traps are enabled and the result is huge, take

F-20
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

// overflow trap
if (!(exc_env->exc_masks & OVERFLOW_MASK) && result_huge) {
dbl_res24 = TWO_TO_M192 * dbl_res24; // exact
exc_env->status_flag_overflow = 1;
exc_env->exception_cause = OVERFLOW;
exc_env->result_fval = (float)dbl_res24; // exact
if (sw & PRECISION_MASK) exc_env->status_flag_inexact = 1;
return (RAISE_EXCEPTION);
}

// set control word with rounding mode set to exc_env->rounding_mode,

// double precision, and all exceptions disabled
cw = cw | 0x0200; // set precision to double
__asm {
fldcw WORD PTR cw;
}

switch (exc_env->operation) {

case ADDPS:
case ADDSS:
// perform the addition
__asm {
// load input operands
fld DWORD PTR opd1; // may set the denormal status flag
fld DWORD PTR opd2; // may set the denormal status flag
faddp st(1), st(0); // rounded to 53 bits, may set the inexact
// status flag
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

case SUBPS:
case SUBSS:
// perform the subtraction
__asm {
// load input operands
fld DWORD PTR opd1; // may set the denormal status flag
fld DWORD PTR opd2; // may set the denormal status flag
fsubp st(1), st(0); // rounded to 53 bits, may set the inexact
// status flag
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

F-21
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

case MULPS:
case MULSS:
// perform the multiplication
__asm {
// load input operands
fld DWORD PTR opd1; // may set the denormal status flag
fld DWORD PTR opd2; // may set the denormal status flag
fmulp st(1), st(0); // rounded to 53 bits, exact
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

case DIVPS:
case DIVSS:
// perform the division
__asm {
// load input operands
fld DWORD PTR opd1; // may set the denormal status flag
fld DWORD PTR opd2; // may set the denormal status flag
fdivp st(1), st(0); // rounded to 53 bits, may set the inexact
// status flag
// store result
fstp QWORD PTR dbl_res; // exact, will not set any flag
}
break;

default:
; // will never occur

// calculate result for the case an inexact trap has to be taken, or

// when no trap occurs (second IEEE rounding)
res = (float)dbl_res;
// may set P, U or O; may also involve denormalizing the result

// read status word

__asm {
fstsw WORD PTR sw;
}

// if inexact traps are enabled and result is inexact, take inexact trap
if (!(exc_env->exc_masks & PRECISION_MASK) &&
((sw & PRECISION_MASK) || (exc_env->ftz && result_tiny))) {
exc_env->status_flag_inexact = 1;

F-22
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

exc_env->exception_cause = INEXACT;
if (result_tiny) {
exc_env->status_flag_underflow = 1;

// if ftz = 1 and result is tiny, result = 0.0

// (no need to check for underflow traps disabled: result tiny and
// underflow traps enabled would have caused taking an underflow
// trap above)
if (exc_env->ftz) {
if (res > 0.0)
res = ZEROF;
else if (res < 0.0)
res = NZEROF;
// else leave res unchanged
}
}
if (result_huge) exc_env->status_flag_overflow = 1;
exc_env->result_fval = res;
return (RAISE_EXCEPTION);
}

// if it got here, then there is no trap to be taken; the following must

// hold: ((the MXCSR U exceptions are disabled or
//
// the MXCSR underflow exceptions are enabled and the underflow flag is
// clear and (the inexact flag is set or the inexact flag is clear and
// the 24-bit result with unbounded exponent is not tiny)))
// and (the MXCSR overflow traps are disabled or the overflow flag is
// clear) and (the MXCSR inexact traps are disabled or the inexact flag
// is clear)
//
// in this case, the result has to be delivered (the status flags are
// sticky, so they are all set correctly already)

// read status word to see if result is inexact

__asm {
fstsw WORD PTR sw;
}

if (sw & UNDERFLOW_MASK) exc_env->status_flag_underflow = 1;

if (sw & OVERFLOW_MASK) exc_env->status_flag_overflow = 1;
if (sw & PRECISION_MASK) exc_env->status_flag_inexact = 1;

// if ftz = 1, and result is tiny (underflow traps must be disabled),

// result = 0.0
if (exc_env->ftz && result_tiny) {
if (res > 0.0)

F-23
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

res = ZEROF;
else if (res < 0.0)
res = NZEROF;
// else leave res unchanged

exc_env->status_flag_inexact = 1;
exc_env->status_flag_underflow = 1;
}

exc_env->result_fval = res;
if (sw & ZERODIVIDE_MASK) exc_env->status_flag_divide_by_zero = 1;
if (sw & DENORMAL_MASK) exc_env->status_flag_denormal= 1;
if (sw & INVALID_MASK) exc_env->status_flag_invalid_operation = 1;
return (DO_NOT_RAISE_EXCEPTION);

break;

case CMPPS:
case CMPSS:

...

break;

case COMISS:
case UCOMISS:

...

break;

case CVTPI2PS:
case CVTSI2SS:

...

break;

case CVTPS2PI:
case CVTSS2SI:
case CVTTPS2PI:
case CVTTSS2SI:

...

break;

F-24
 GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

case MAXPS:
case MAXSS:
case MINPS:
case MINSS:

...

break;

case SQRTPS:
case SQRTSS:

...

break;

case UNSPEC:

...

break;

default:
...

F-25
GUIDELINES FOR WRITING SIMD FLOATING-POINT EXCEPTION

F-26
 INDEX

Numerics B
16-bit B (default size) flag, segment descriptor .3-14, 4-3
address size . . . . . . . . . . . . . . . . . . . . . . . . .3-4 Base (operand addressing) . . . . . . . . . . .5-9, 5-10
operand size . . . . . . . . . . . . . . . . . . . . . . . . .3-4 Basic execution environment . . . . . . . . . . . . . . 3-2
32-bit B-bit, FPU status word . . . . . . . . . . . . . . . . . . 7-15
address size . . . . . . . . . . . . . . . . . . . . . . . . .3-4 BCD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
operand size . . . . . . . . . . . . . . . . . . . . . . . . .3-4 BCD integers . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
FPU encoding . . . . . . . . . . . . . . . . . . . . . . 7-29
packed. . . . . . . . . . . . . . . . . . . . . . . . .5-5, 6-28
A relationship to status flags. . . . . . . . . . . . . 3-12
AAA instruction. . . . . . . . . . . . . . . . . . . . . . . . .6-28 unpacked. . . . . . . . . . . . . . . . . . . . . . .5-5, 6-28
AAD instruction . . . . . . . . . . . . . . . . . . . . . . . .6-28 BH register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
AAM instruction . . . . . . . . . . . . . . . . . . . . . . . .6-28 Bias value
AAS instruction. . . . . . . . . . . . . . . . . . . . . . . . .6-28 numeric overflow . . . . . . . . . . . . . . . . . . . . 7-55
AC (alignment check) flag, EFLAGS register. .3-13 numeric underflow. . . . . . . . . . . . . . . . . . . 7-56
Access rights, segment descriptor . . . . . . 4-9, 4-13 Biased exponent. . . . . . . . . . . . . . . . . . . . . . . . 7-5
ADC instruction . . . . . . . . . . . . . . . . . . . . . . . .6-26 Binary numbers . . . . . . . . . . . . . . . . . . . . . . . . 1-7
ADD instruction . . . . . . . . . . . . . . . . . . . . . . . .6-26 Binary-coded decimal (see BCD)
Address size attribute Bit fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
code segment . . . . . . . . . . . . . . . . . . . . . . .3-14 Bit order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
description of . . . . . . . . . . . . . . . . . . . . . . .3-14 BOUND instruction . . . . . . . . . . . . 4-17, 6-39, 6-44
of stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-3 BOUND range exceeded exception (#BR) . . . 4-18
Address sizes. . . . . . . . . . . . . . . . . . . . . . . . . . .3-4 BP register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
Addressing modes Branch prediction . . . . . . . . . . . . . . . . . . . . . . . 2-8
assembler . . . . . . . . . . . . . . . . . . . . . . . . . .5-10 Branching, on FPU condition codes . . . .7-15, 7-38
base . . . . . . . . . . . . . . . . . . . . . . . . . . 5-9, 5-10 BSF instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-34
base plus displacement . . . . . . . . . . . . . . .5-10 BSR instruction. . . . . . . . . . . . . . . . . . . . . . . . 6-34
base plus index plus displacement . . . . . . .5-10 BSWAP instruction . . . . . . . . . . . . . . . . . .6-3, 6-21
base plus index time scale plus BT instruction . . . . . . . . . . . . . . . . 3-10, 3-12, 6-34
displacement . . . . . . . . . . . . . . . . . . . . .5-10 BTC instruction . . . . . . . . . . . . . . . 3-10, 3-12, 6-34
displacement. . . . . . . . . . . . . . . . . . . . . . . . .5-9 BTR instruction . . . . . . . . . . . . . . . 3-10, 3-12, 6-34
effective address. . . . . . . . . . . . . . . . . . . . . .5-9 BTS instruction . . . . . . . . . . . . . . . 3-10, 3-12, 6-34
immediate operands . . . . . . . . . . . . . . . . . . .5-6 Bus interface unit . . . . . . . . . . . . . . . . . . . . . . . 2-9
index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-9 BX register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
index times scale plus displacement . . . . .5-10 Byte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
memory operands. . . . . . . . . . . . . . . . . . . . .5-7 Byte order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5
register operands . . . . . . . . . . . . . . . . . . . . .5-7
scale factor . . . . . . . . . . . . . . . . . . . . . . . . . .5-9
specifying a segment selector . . . . . . . . . . .5-8 C
specifying an offset . . . . . . . . . . . . . . . . . . . .5-9 C1 flag, FPU status word . . 7-13, 7-52, 7-55, 7-57
Addressing, segments . . . . . . . . . . . . . . . . . . . .1-7 C2 flag, FPU status word . . . . . . . . . . . . . . . . 7-13
Advanced programmable interrupt controller Call gate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-8
(see APIC) CALL instruction . . . 3-14, 4-4, 4-5, 4-9, 6-36, 6-44
AF (adjust) flag, EFLAGS register . . . . . . . . . .3-12 Calls (see Procedure calls)
AH register . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-7 CBW instruction . . . . . . . . . . . . . . . . . . . . . . . 6-26
Alignment CDQ instruction . . . . . . . . . . . . . . . . . . . . . . . 6-26
of words, doublewords, and quadwords . . . .5-2 CF (carry) flag, EFLAGS register . . . . . . . . . . 3-12
AND instruction . . . . . . . . . . . . . . . . . . . . . . . .6-29 CH register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
APIC, presence of . . . . . . . . . . . . . . . . . . . . . .11-2 CLC instruction . . . . . . . . . . . . . . . . . . . .3-12, 6-42
Arctangent, FPU operation. . . . . . . . . . . . . . . .7-38 CLD instruction . . . . . . . . . . . . . . . . . . . .3-13, 6-42
Arithmetic instructions, FPU. . . . . . . . . . . . . . .7-46 CLI instruction. . . . . . . . . . . . . . . . . . . . .6-43, 10-4
Assembler, addressing modes. . . . . . . . . . . . .5-10 CMC instruction . . . . . . . . . . . . . . . . . . .3-12, 6-42
AX register . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-7 CMOVcc instructions . . . . . . . . . . . . . . . .6-2, 6-20

INDEX-1
INDEX

CMP instruction . . . . . . . . . . . . . . . . . . . . . . . .6-27 DE (denormal operand exception) flag,

CMPS instruction . . . . . . . . . . . . . . . . . . 3-13, 6-40 FPU status word. . . . . . . . . . .7-14, 7-54
CMPXCHG instruction . . . . . . . . . . . . . . . 6-3, 6-22 DEC instruction. . . . . . . . . . . . . . . . . . . . . . . . 6-26
CMPXCHG8B instruction . . . . . . . . 6-2, 6-22, 11-2 Decimal integers, FPU
Code segment . . . . . . . . . . . . . . . . . . . . . . . . . .3-9 description of. . . . . . . . . . . . . . . . . . . . . . . 7-29
Compare encodings . . . . . . . . . . . . . . . . . . . . . . . . . 7-29
compare and exchange . . . . . . . . . . . . . . .6-22 Deep branch prediction . . . . . . . . . . . . . . . . . . 2-8
integers . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-27 Denormal number
real numbers, FPU . . . . . . . . . . . . . . . . . . .7-37 (see Denormalized finite number)
strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-40 Denormal operand exception (#D) . . . . . . . . . 7-54
Compatibility Denormalization process . . . . . . . . . . . . . . . . . 7-7
software . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-6 Denormalized finite number . . . . . . . 7-6, 7-25, 9-5
Condition code flags, FPU status word DF (direction) flag, EFLAGS register . . . . . . . 3-13
branching on . . . . . . . . . . . . . . . . . . . . . . . .7-15 DH register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
conditional moves on . . . . . . . . . . . . . . . . .7-15 DI register . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
description of . . . . . . . . . . . . . . . . . . . . . . .7-12 Dispatch/execute unit . . . . . . . . . . . . . . . . . . . 2-12
interpretation of. . . . . . . . . . . . . . . . . . . . . .7-14 Displacement (operand addressing). . . . .5-9, 5-10
use of . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-36 DIV instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-27
Conditional moves, on FPU condition codes . .7-15 Division-by-zero exception (#Z) . . . . . . . . . . . 7-53
Constants (floating-point) Double-extended-precision, IEEE
descriptions of. . . . . . . . . . . . . . . . . . . . . . .7-34 floating-point format . . . . . . . . .7-25, 9-5
Cosine, FPU operation. . . . . . . . . . . . . . . . . . .7-38 Double-precision, IEEE
CPUID instruction. . . . . . . . . 6-2, 6-45, 11-2, 11-4 floating-point format . . . . . . . . .7-25, 9-5
CS register . . . . . . . . . . . . . . . . . . . . . . . . . 3-7, 3-9 Double-real floating-point format . . . . . . .7-25, 9-5
CTI instruction . . . . . . . . . . . . . . . . . . . . . . . . .6-42 Doubleword . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-1
Current privilege level (see CPL) DS register . . . . . . . . . . . . . . . . . . . . . . . . .3-7, 3-9
Current stack . . . . . . . . . . . . . . . . . . . . . . . 4-2, 4-4 DX register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
CWD instruction . . . . . . . . . . . . . . . . . . . . . . . .6-26 Dynamic data flow analysis . . . . . . . . . . . . . . . 2-8
CWDE instruction. . . . . . . . . . . . . . . . . . . . . . .6-26 Dynamic execution . . . . . . . . . . . . . . . . . . . . . . 2-8
CX register . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-7
E
D EAX register . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
DAA instruction . . . . . . . . . . . . . . . . . . . . . . . .6-28 EBP register . . . . . . . . . . . . . . . . . . . . 3-6, 4-4, 4-7
DAS instruction . . . . . . . . . . . . . . . . . . . . . . . .6-28 EBX register . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
Data pointer, FPU . . . . . . . . . . . . . . . . . . . . . .7-21 ECX register . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
Data segment. . . . . . . . . . . . . . . . . . . . . . . . . . .3-9 EDI register. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
Data types EDX register . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
alignment of words, doublewords, Effective address . . . . . . . . . . . . . . . . . . . . . . . 5-9
and quadwords . . . . . . . . . . . . . . . . . . . .5-2 EFLAGS Condition Codes . . . . . . . . . . . . . . . . B-1
BCD integers . . . . . . . . . . . . . . . . . . . 5-5, 6-28 EFLAGS register . . . . . . . . . . . . . . . . . . . . . . 3-10
bit fields. . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5 restoring from procedure stack . . . . . . . . . . 4-8
byte . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-1 saving on a procedure call . . . . . . . . . . . . . 4-8
doubleword . . . . . . . . . . . . . . . . . . . . . . . . . .5-1 status flags . . . . . . . . . . . . . . . 7-15, 7-16, 7-37
FPU BCD decimal. . . . . . . . . . . . . . . . . . . .7-29 EIP register. . . . . . . . . . . . . . . . . . . . . . . .3-8, 3-14
FPU integer. . . . . . . . . . . . . . . . . . . . . . . . .7-27 EMMS instruction . . . . . . . . . . . . . . . . . .8-10, 8-12
FPU real number . . . . . . . . . . . . . . . . . . . .7-25 ENTER instruction . . . . . . . . . . . . . . . . .4-18, 6-41
fundamental data types . . . . . . . . . . . . . . . .5-1 ES register . . . . . . . . . . . . . . . . . . . . . . . . .3-7, 3-9
integers . . . . . . . . . . . . . . . . . . . 5-3, 6-26, 6-27 ES (exception summary) flag,
packed bytes. . . . . . . . . . . . . . . . . . . . . . . . .8-3 FPU status word. . . . . . . . . . . . . . . 7-59
packed doublewords. . . . . . . . . . . . . . . . . . .8-3 ESC instructions, FPU . . . . . . . . . . . . . . . . . . 7-32
packed words . . . . . . . . . . . . . . . . . . . . . . . .8-3 ESI register. . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-6
pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5 ESP register . . . . . . . . . . . . . . . . 3-6, 4-1, 4-3, 4-4
quadword . . . . . . . . . . . . . . . . . . . . . . . 5-1, 8-3 Exception flags, FPU status word. . . . . . . . . . 7-14
strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5 Exception handler. . . . . . . . . . . . . . . . . . . . . . 4-11
unsigned integers . . . . . . . . . . . 5-5, 6-26, 6-27 Exception priority, FPU exceptions. . . . . . . . . 7-57
word . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-1 Exception-flag masks, FPU control word . . . . 7-17

INDEX-2
 INDEX

Exceptions FLD instruction . . . . . . . . . . . . . . . . . . . . . . . . 7-33

BOUND range exceeded (#BR) . . . . . . . . .4-18 FLD1 instruction . . . . . . . . . . . . . . . . . . . . . . . 7-34
description of . . . . . . . . . . . . . . . . . . . . . . .4-11 FLDCW instruction . . . . . . . . . . . . . . . . .7-16, 7-42
implicit call to handler . . . . . . . . . . . . . . . . . .4-1 FLDENV instruction . . . . . . 7-15, 7-21, 7-24, 7-42
in real-address mode . . . . . . . . . . . . . . . . .4-17 FLDL2E instruction . . . . . . . . . . . . . . . . . . . . . 7-34
notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1-8 FLDL2T instruction . . . . . . . . . . . . . . . . . . . . . 7-34
overflow exception (#OF) . . . . . . . . . . . . . .4-17 FLDLG2 instruction. . . . . . . . . . . . . . . . . . . . . 7-34
summary of . . . . . . . . . . . . . . . . . . . . . . . . .4-14 FLDLN2 instruction. . . . . . . . . . . . . . . . . . . . . 7-34
vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-12 FLDPI instruction . . . . . . . . . . . . . . . . . . . . . . 7-34
Exponent FLDSW instruction . . . . . . . . . . . . . . . . . . . . . 7-42
floating-point number . . . . . . . . . . . . . . . . . .7-4 FLDZ instruction . . . . . . . . . . . . . . . . . . . . . . . 7-34
Exponential, FPU operation . . . . . . . . . . . . . . .7-40 Floating-point data types . . . . . . . . . . . . . . . . 7-24
Extended real Floating-point exceptions
encodings, unsupported . . . . . . . . . . . . . . .7-30 automatic handling . . . . . . . . . . . . . . . . . . 7-47
floating-point format . . . . . . . . . . . . . . 7-25, 9-5 denormal operand exception. . . . . . . . . . . 7-54
division-by-zero . . . . . . . . . . . . . . . . . . . . . 7-53
exception conditions . . . . . . . . . . . . . . . . . 7-51
F exception priority . . . . . . . . . . . . . . . . . . . . 7-57
F2XM1 instruction . . . . . . . . . . . . . . . . . . . . . .7-40 inexact result (precision) . . . . . . . . . . . . . . 7-57
FABS instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 invalid arithmetic operand . . . . . . . . .7-51, 7-52
FADD instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 numeric overflow . . . . . . . . . . . . . . . . . . . . 7-54
FADDP instruction . . . . . . . . . . . . . . . . . . . . . .7-35 numeric underflow. . . . . . . . . . . . . . . . . . . 7-56
Far call software handling . . . . . . . . . . . . . . . . . . . 7-49
description of . . . . . . . . . . . . . . . . . . . . . . . .4-5 stack overflow . . . . . . . . . . . . . . . . . .7-13, 7-52
operation. . . . . . . . . . . . . . . . . . . . . . . . . . . .4-6 stack underflow . . . . . . . . . . . . 7-13, 7-51, 7-52
Far pointer summary of . . . . . . . . . . . . . . . . . . . . . . . . 7-46
16-bit addressing . . . . . . . . . . . . . . . . . . . . .3-4 synchronization . . . . . . . . . . . . . . . . . . . . . 7-58
32-bit addressing . . . . . . . . . . . . . . . . . . . . .3-4 Floating-point format
description of . . . . . . . . . . . . . . . . . . . . 3-3, 5-5 biased exponent . . . . . . . . . . . . . . . . . . . . . 7-5
FBSTP instruction . . . . . . . . . . . . . . . . . . . . . .7-33 description of. . . . . . . . . . . . . . . . . . . . . . . 7-24
FCHS instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 exponent . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
FCLEX/FNCLEX instructions . . . . . . . . . . . . . .7-15 fraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
FCMOVcc instructions . . . . . . . . . . 6-2, 7-16, 7-33 real number system. . . . . . . . . . . . . . . . . . . 7-2
FCOM instruction . . . . . . . . . . . . . . . . . . 7-15, 7-36 real numbers . . . . . . . . . . . . . . . . . . . . . . . 7-25
FCOMI instruction . . . . . . . . . . . . . 6-2, 7-16, 7-36 sign . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
FCOMIP instruction . . . . . . . . . . . . 6-2, 7-16, 7-36 significand . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
FCOMP instruction. . . . . . . . . . . . . . . . . 7-15, 7-36 FMUL instruction . . . . . . . . . . . . . . . . . . . . . . 7-35
FCOMPP instruction . . . . . . . . . . . . . . . 7-15, 7-36 FMULP instruction . . . . . . . . . . . . . . . . . . . . . 7-35
FCOS instruction . . . . . . . . . . . . . . . . . . 7-13, 7-38 FNOP instruction . . . . . . . . . . . . . . . . . . . . . . 7-41
FDIV instruction . . . . . . . . . . . . . . . . . . . . . . . .7-35 FPATAN instruction . . . . . . . . . . . . . . . .7-38, 7-39
FDIVP instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 FPREM instruction . . . . . . . . . . . . 7-13, 7-35, 7-39
FDIVR instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 FPREM1 instruction . . . . . . . . . . . 7-13, 7-35, 7-39
FDIVRP instruction. . . . . . . . . . . . . . . . . . . . . .7-35 FPTAN instruction . . . . . . . . . . . . . . . . . . . . . 7-13
Feature determination, of processor . . . . . . . .11-1 FPU
Fetch/decode unit. . . . . . . . . . . . . . . . . . . . . . .2-11 architecture . . . . . . . . . . . . . . . . . . . . . . . . . 7-8
FIADD instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 compatibility with Intel Architecture FPUs and
FICOM instruction . . . . . . . . . . . . . . . . . 7-15, 7-36 math coprocessors . . . . . . . . . . . . . . . . 7-1
FICOMP instruction . . . . . . . . . . . . . . . . 7-15, 7-36 floating-point format . . . . . . . . . . . . . . .7-2, 7-4
FIDIV instruction. . . . . . . . . . . . . . . . . . . . . . . .7-35 IEEE standards . . . . . . . . . . . . . . . . . . . . . . 7-1
FIDIVR instruction . . . . . . . . . . . . . . . . . . . . . .7-35 presence of . . . . . . . . . . . . . . . . . . . . . . . . 11-2
FILD instruction . . . . . . . . . . . . . . . . . . . . . . . .7-33 transcendental instruction accuracy . . . . . 7-40
FIMUL instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 FPU control word
FINIT/FNINIT instructions . 7-15, 7-16, 7-20, 7-41 description of. . . . . . . . . . . . . . . . . . . . . . . 7-16
FIST instruction . . . . . . . . . . . . . . . . . . . . . . . .7-33 exception-flag masks . . . . . . . . . . . . . . . . 7-17
FISTP instruction . . . . . . . . . . . . . . . . . . . . . . .7-33 PC field . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-17
FISUB instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 RC field . . . . . . . . . . . . . . . . . . . . . . . .7-18, 9-8
FISUBR instruction. . . . . . . . . . . . . . . . . . . . . .7-35 FPU data pointer . . . . . . . . . . . . . . . . . . . . . . 7-21
Flat memory model . . . . . . . . . . . . . . . . . . 3-2, 3-8 FPU data registers . . . . . . . . . . . . . . . . . . . . . . 7-9

INDEX-3
INDEX

FPU instruction pointer. . . . . . . . . . . . . . . . . . .7-21 G

FPU instructions General-purpose registers . . . . . . . . . . . . .3-5, 3-6
arithmetic vs. non-arithmetic instructions . .7-46 parameter passing . . . . . . . . . . . . . . . . . . . 4-7
instruction set . . . . . . . . . . . . . . . . . . . . . . .7-31 GS register . . . . . . . . . . . . . . . . . . . . . . . . .3-7, 3-9
operands. . . . . . . . . . . . . . . . . . . . . . . . . . .7-32
overview . . . . . . . . . . . . . . . . . . . . . . . . . . .7-31
unsupported . . . . . . . . . . . . . . . . . . . . . . . .7-43 H
FPU integer Hexadecimal numbers . . . . . . . . . . . . . . . . . . . 1-7
description of . . . . . . . . . . . . . . . . . . . . . . .7-27
encodings . . . . . . . . . . . . . . . . . . . . . . . . . .7-28
FPU last opcode. . . . . . . . . . . . . . . . . . . . . . . .7-21 I
FPU register stack ID (identification) flag, EFLAGS register. . . . . 3-14
description of . . . . . . . . . . . . . . . . . . . . . . . .7-9 IDIV instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-27
parameter passing . . . . . . . . . . . . . . . . . . .7-11 IE (invalid operation exception) flag,
FPU state FPU status word. . . . . . . . . . .7-14, 7-52
image . . . . . . . . . . . . . . . . . . . . . . . . 7-22, 7-23 IEEE 754 and 854 standards for
saving . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-21 floating-point arithmetic . . . . . . . . . . 7-1
FPU status word IF (interrupt enable) flag,
condition code flags . . . . . . . . . . . . . . . . . .7-12 EFLAGS register . . . . . 3-13, 4-13, 10-5
DE flag . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-54 Immediate operands. . . . . . . . . . . . . . . . . . . . . 5-6
description of . . . . . . . . . . . . . . . . . . . . . . .7-12 IMUL instruction . . . . . . . . . . . . . . . . . . . . . . . 6-27
exception flags . . . . . . . . . . . . . . . . . . . . . .7-14 IN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-3
OE flag . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-54 IN instruction. . . . . . . . . . . . . . . . . 6-41, 10-3, 10-4
PE flag . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-13 INC instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-26
TOP field. . . . . . . . . . . . . . . . . . . . . . . . . . . .7-9 Indefinite
FPU tag word . . . . . . . . . . . . . . . . . . . . . . . . . .7-20 description of. . . . . . . . . . . . . . . . . . . . . . . . 7-8
Fraction, floating-point number . . . . . . . . . . . . .7-4 integer . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-28
FRNDINT instruction . . . . . . . . . . . . . . . . . . . .7-35 packed BCD decimal. . . . . . . . . . . . . . . . . 7-30
FRSTOR instruction . . . . . . 7-15, 7-21, 7-24, 7-42 real . . . . . . . . . . . . . . . . . . . . . . . . . . .7-27, 9-6
FS register . . . . . . . . . . . . . . . . . . . . . . . . . 3-7, 3-9 Index (operand addressing) . . . . . . . . . . .5-9, 5-10
FSAVE/FNSAVE instructions. . . .7-12, 7-15, 7-20, Inexact result (precision) exception (#P) . . . . 7-57
7-21, 7-42 Inexact result, FPU . . . . . . . . . . . . . . . . . . . . . 7-19
FSCALE instruction . . . . . . . . . . . . . . . . . . . . .7-40 Infinity control flag, FPU control word. . . . . . . 7-20
FSIN instruction . . . . . . . . . . . . . . . . . . . 7-13, 7-38 Infinity, floating-point format . . . . . . . . . . . . . . . 7-8
FSINCOS instruction . . . . . . . . . . . . . . . 7-13, 7-39 INIT pin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-10
FSQRT instruction . . . . . . . . . . . . . . . . . . . . . .7-35 Input/output (see I/O)
FST instruction . . . . . . . . . . . . . . . . . . . . . . . . .7-33 INS instruction . . . . . . . . . . . . . . . 6-41, 10-3, 10-4
FSTCW/FNSTCW instructions. . . . . . . . 7-16, 7-42 Instruction decoder . . . . . . . . . . . . . . . . . . . . . 2-11
FSTENV/FNSTENV instructions .7-12, 7-20, 7-21, Instruction operands . . . . . . . . . . . . . . . . . . . . . 1-7
7-42 Instruction pointer (EIP register). . . . . . . . . . . 3-14
FSTP instruction. . . . . . . . . . . . . . . . . . . . . . . .7-33 Instruction pointer, FPU . . . . . . . . . . . . . . . . . 7-21
FSTSW/FNSTSW instructions . . . . . . . . 7-12, 7-42 Instruction pool (reorder buffer) . . . . . . . . . . . 2-11
FSUB instruction . . . . . . . . . . . . . . . . . . . . . . .7-35 Instruction prefixes (see Prefixes)
FSUBP instruction . . . . . . . . . . . . . . . . . . . . . .7-35 Instruction set
FSUBR instruction . . . . . . . . . . . . . . . . . . . . . .7-35 binary arithmetic instructions. . . . . . . . . . . 6-26
FSUBRP instruction . . . . . . . . . . . . . . . . . . . . .7-35 bit scan instructions. . . . . . . . . . . . . . . . . . 6-34
FTST instruction. . . . . . . . . . . . . . . . . . . 7-15, 7-36 bit test and modify instructions . . . . . . . . . 6-34
FUCOM instruction. . . . . . . . . . . . . . . . . . . . . .7-36 byte-set-on-condition instructions . . . . . . . 6-34
FUCOMI instruction . . . . . . . . . . . . 6-2, 7-16, 7-36 control transfer instructions . . . . . . . . . . . . 6-35
FUCOMIP instruction . . . . . . . . . . . 6-2, 7-16, 7-36 data movement instructions . . . . . . . . . . . 6-20
FUCOMP instruction . . . . . . . . . . . . . . . . . . . .7-36 decimal arithmetic instructions . . . . . . . . . 6-27
FUCOMPP instruction . . . . . . . . . . . . . . 7-15, 7-36 EFLAGS instructions. . . . . . . . . . . . . . . . . 6-42
FXAM instruction . . . . . . . . . . . . . . . . . . 7-13, 7-36 floating-point instructions . . . . . . . . .6-10, 6-12
FXCH instruction . . . . . . . . . . . . . . . . . . . . . . .7-33 integer instructions . . . . . . . . . . . . . . . . . . . 6-3
FXTRACT instruction . . . . . . . . . . . . . . . . . . . .7-35 I/O instructions . . . . . . . . . . . . . . . . . . . . . 6-41
FYL2X instruction. . . . . . . . . . . . . . . . . . . . . . .7-40 lists of . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2
FYL2XP1 instruction . . . . . . . . . . . . . . . . . . . .7-40 logical instructions. . . . . . . . . . . . . . . . . . . 6-29

INDEX-4
 INDEX

MMX instructions . . . . . . . . . . . . . . . . . . . . .8-5 ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-6

processor identification instruction . . . . . . .6-45 protected mode I/O . . . . . . . . . . . . . . . . . . 10-4
repeating string operations . . . . . . . . . . . . .6-40 I/O privilege level (see IOPL)
rotate instructions . . . . . . . . . . . . . . . . . . . .6-32 I/O sensitive instructions. . . . . . . . . . . . . . . . . 10-4
segment register instructions . . . . . . . . . . .6-43
shift instructions . . . . . . . . . . . . . . . . . . . . .6-29
software interrupt instructions. . . . . . . . . . .6-39 J
string operation instructions . . . . . . . . . . . .6-39 J-bit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-4
summary . . . . . . . . . . . . . . . . . . . . . . . . . . . .6-1 Jcc instructions . . . . . . . . . . . . . . . 3-12, 3-14, 6-37
system instructions . . . . . . . . . . . . . . . . . . .6-16 JMP instruction . . . . . . . . . . . . . . . 3-14, 6-35, 6-44
test instruction. . . . . . . . . . . . . . . . . . . . . . .6-35
type conversion instructions . . . . . . . . . . . .6-25 L
INT instruction . . . . . . . . . . . . . . . . . . . . 4-17, 6-44
L1 (level 1) cache . . . . . . . . . . . . . . . . . . . .2-7, 2-9
Integers . . . . . . . . . . . . . . . . . . . . . 5-3, 6-26, 6-27
L2 (level 2) cache . . . . . . . . . . . . . . . . . . . .2-7, 2-9
Integer, FPU data type
LAHF instruction . . . . . . . . . . . . . . . . . . .3-10, 6-42
description of . . . . . . . . . . . . . . . . . . . . . . .7-27
Last instruction opcode, FPU . . . . . . . . . . . . . 7-21
indefinite . . . . . . . . . . . . . . . . . . . . . . . . . . .7-28
LDS instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-44
Inter-privilege level call
LEA instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-44
description of . . . . . . . . . . . . . . . . . . . . . . . .4-8
LEAVE instruction. . . . . . . . . . . . . 4-18, 4-24, 6-41
operation. . . . . . . . . . . . . . . . . . . . . . . . . . .4-10
LES instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-44
Inter-privilege level return
LGS instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-44
description of . . . . . . . . . . . . . . . . . . . . . . . .4-8
Linear address . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
operation. . . . . . . . . . . . . . . . . . . . . . . . . . .4-10
Linear address space
Interrupt gate . . . . . . . . . . . . . . . . . . . . . . . . . .4-13
defined . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-2
Interrupt handler. . . . . . . . . . . . . . . . . . . . . . . .4-11
maximum size . . . . . . . . . . . . . . . . . . . . . . . 3-2
Interrupt vector . . . . . . . . . . . . . . . . . . . . . . . . .4-12
LOCK signal . . . . . . . . . . . . . . . . . . . . . . . . . . 6-21
Interrupts
LODS instruction . . . . . . . . . . . . . . . . . .3-13, 6-40
description of . . . . . . . . . . . . . . . . . . . . . . .4-11
Log epsilon, FPU operation . . . . . . . . . . . . . . 7-40
implicit call to an interrupt handler
Log (base 2), FPU computation . . . . . . . . . . . 7-40
procedure . . . . . . . . . . . . . . . . . . . . . . .4-13
Logical address . . . . . . . . . . . . . . . . . . . . . . . . 3-3
implicit call to an interrupt handler task. . . .4-17
LOOP instructions . . . . . . . . . . . . . . . . . . . . . 6-38
in real-address mode . . . . . . . . . . . . . . . . .4-17
LOOPcc instructions. . . . . . . . . . . . . . . .3-12, 6-38
maskable . . . . . . . . . . . . . . . . . . . . . . . . . .4-12
LSS instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-44
summary of . . . . . . . . . . . . . . . . . . . . . . . . .4-14
user-defined . . . . . . . . . . . . . . . . . . . . . . . .4-12
vector . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-12 M
INTn instruction . . . . . . . . . . . . . . . . . . . . . . . .6-39 Maskable interrupts . . . . . . . . . . . . . . . . . . . . 4-12
INTO instruction . . . . . . . . . . . . . . 4-17, 6-39, 6-44 Masked responses
Invalid arithmetic operand exception (#IA), FPU to denormal operand exception. . . . . . . . . 7-54
description of . . . . . . . . . . . . . . . . . . . . . . .7-52 to division-by-zero exception. . . . . . . . . . . 7-54
masked response to . . . . . . . . . . . . . . . . . .7-53 to FPU stack overflow or
Invalid operation exception . . . . . . . . . . . . . . .7-51 underflow exception. . . . . . . . . . . . . . . 7-52
INVD instruction . . . . . . . . . . . . . . . . . . . . . . . . .6-3 to inexact result (precision) exception. . . . 7-57
INVLPG instruction. . . . . . . . . . . . . . . . . . . . . . .6-3 to invalid arithmetic operation . . . . . . . . . . 7-53
IOPL (I/O privilege level) field, to numeric overflow exception. . . . . . . . . . 7-55
EFLAGS register . . . . . . . . . . 3-13, 10-4 to numeric underflow exception . . . . . . . . 7-56
IRET instruction 3-14, 4-16, 4-17, 6-36, 6-44, 10-5 Masks, exception-flags, FPU control word . . . 7-17
I/O address space . . . . . . . . . . . . . . . . . . . . . .10-2 Memory
I/O instructions order buffer . . . . . . . . . . . . . . . . . . . . . . . . 2-10
overview of . . . . . . . . . . . . . . . . . . . . 6-41, 10-3 organization. . . . . . . . . . . . . . . . . . . . . .3-2, 3-3
serialization. . . . . . . . . . . . . . . . . . . . . . . . .10-6 subsystem . . . . . . . . . . . . . . . . . . . . . . . . . . 2-9
I/O map base . . . . . . . . . . . . . . . . . . . . . . . . . .10-5 Memory interface unit . . . . . . . . . . . . . . . . . . . . 2-9
I/O permission bit map . . . . . . . . . . . . . . . . . . .10-5 Memory operands. . . . . . . . . . . . . . . . . . . . . . . 5-7
I/O ports Memory-mapped I/O. . . . . . . . . . . . . . . .10-1, 10-2
addressing . . . . . . . . . . . . . . . . . . . . . . . . .10-1 MESI (modified, exclusive, shared, invalid)
defined . . . . . . . . . . . . . . . . . . . . . . . . . . . .10-1 cache protocol . . . . . . . . . . . . . . . . . 2-9
hardware. . . . . . . . . . . . . . . . . . . . . . . 9-7, 10-1
memory-mapped I/O. . . . . . . . . . . . . . . . . .10-2

INDEX-5
INDEX

Microarchitecture SNaNs vs. QNaNs . . . . . . . . . . . . . . . . . . . 7-8

detailed description. . . . . . . . . . . . . . . . . . . .2-9 Near call
overview . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-6 description of. . . . . . . . . . . . . . . . . . . . . . . . 4-5
Micro-ops . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2-11 operation . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-5
MM0, MM1, MM2, MM3, MM4, MM5, MM6, Near pointer
MM7 registers . . . . . . . . . . . . . . . . . .8-2 description of. . . . . . . . . . . . . . . . . . . . . . . . 5-5
MMX technology Near return
arithmetic instructions . . . . . . . . . . . . . . . . . .8-9 operation . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-5
comparison instructions . . . . . . . . . . . . . . . .8-9 Near return operation . . . . . . . . . . . . . . . . . . . . 4-6
compatibility with FPU architecture. . . . . . .8-11 NEG instruction . . . . . . . . . . . . . . . . . . . . . . . 6-27
conversion instructions . . . . . . . . . . . . . . . .8-10 Non-arithmetic instructions, FPU . . . . . . . . . . 7-46
CPUID instruction . . . . . . . . . . . . . . . . . . . .11-2 Non-number encodings, FPU . . . . . . . . . . . . . . 7-5
data transfer instructions . . . . . . . . . . . . . . .8-7 Non-waiting instructions . . . . . . . . . . . . .7-42 , 7-49
data types . . . . . . . . . . . . . . . . . . . . . . . . . . .8-3 NOP instruction . . . . . . . . . . . . . . . . . . . . . . . 6-45
detecting MMX technology with Normalized finite number . . . . . . . . . . . . . .7-4, 7-6
CPUID instruction . . . . . . . . . . . . . . . . .8-11 NOT instruction. . . . . . . . . . . . . . . . . . . . . . . . 6-29
detecting with CPUID instruction . . . . . . . .8-12 Notation
effect of instruction prefixes on bit and byte order . . . . . . . . . . . . . . . . . . . . 1-5
MMX instructions. . . . . . . . . . . . . . . . . .8-11 exceptions . . . . . . . . . . . . . . . . . . . . . . . . . . 1-8
EMMS instruction . . . . . . . . . . . . . . . . . . . .8-10 hexadecimal and binary numbers . . . . . . . . 1-7
exception handling in MMX code . . . . . . . .8-16 instruction operands . . . . . . . . . . . . . . . . . . 1-7
instruction operands . . . . . . . . . . . . . . . . . . .8-7 reserved bits . . . . . . . . . . . . . . . . . . . . . . . . 1-6
instruction set . . . . . . . . . . . . . . . . . . . . 8-5, 8-7 segmented addressing . . . . . . . . . . . . . . . . 1-7
interfacing with MMX code . . . . . . . . . . . . .8-13 Notational conventions . . . . . . . . . . . . . . . . . . . 1-5
introduction to . . . . . . . . . . . . . . . . . . . . . . . .8-1 NT (nested task) flag, EFLAGS register . . . . . 3-13
logical instructions . . . . . . . . . . . . . . . . . . .8-10 Numeric overflow exception (#O) . . . . . .7-13, 7-54
memory data formats . . . . . . . . . . . . . . . . . .8-4 Numeric underflow exception (#U) . . . . .7-13, 7-56
mixing MMX and floating-point
instructions . . . . . . . . . . . . . . . . . . . . . .8-14
programming environment (overview) . . . . .8-1 O
register data formats. . . . . . . . . . . . . . . . . . .8-5 OE (numeric overflow exception) flag,
register mapping . . . . . . . . . . . . . . . . . . . . .8-16 FPU status word. . . . . . . . . . .7-14, 7-55
registers . . . . . . . . . . . . . . . . . . . . . . . . . . . .8-2 OF (overflow) flag, EFLAGS register . . .3-12, 4-17
saturation arithmetic . . . . . . . . . . . . . . . . . . .8-6 Offset (operand addressing). . . . . . . . . . . . . . . 5-9
shift instructions . . . . . . . . . . . . . . . . . . . . .8-10 Operand
SIMD execution environment . . . . . . . . . . . .8-4 FPU instructions . . . . . . . . . . . . . . . . . . . . 7-32
support for, determing. . . . . . . . . . . . . . . . .11-2 instruction . . . . . . . . . . . . . . . . . . . . . . . . . . 1-7
using MMX code in a multitasking Operand addressing, modes . . . . . . . . . . . . . . 5-6
operating system environment . . . . . . .8-15 Operand sizes . . . . . . . . . . . . . . . . . . . . . . . . . 3-4
using the EMMS instruction . . . . . . . . . . . .8-12 Operand-size attribute
wraparound mode. . . . . . . . . . . . . . . . . . . . .8-6 code segment . . . . . . . . . . . . . . . . . . . . . . 3-14
Modes, operating . . . . . . . . . . . . . . . . . . . . . . . .3-4 description of. . . . . . . . . . . . . . . . . . . . . . . 3-14
MOV instruction . . . . . . . . . . . . . . . . . . . 6-20, 6-43 Operating modes . . . . . . . . . . . . . . . . . . . . . . . 3-4
MOVD instruction . . . . . . . . . . . . . . . . . . . . . . . .8-7 OR instruction. . . . . . . . . . . . . . . . . . . . . . . . . 6-29
MOVQ instruction. . . . . . . . . . . . . . . . . . . . . . . .8-7 Ordering I/O . . . . . . . . . . . . . . . . . . . . . . . . . . 10-6
MOVS instruction . . . . . . . . . . . . . . . . . . 3-13, 6-40 OUT instruction. . . . . . . . . . . . . . . 6-41, 10-3, 10-4
MOVSX instruction. . . . . . . . . . . . . . . . . . . . . .6-26 OUTS instruction . . . . . . . . . . . . . 6-41, 10-3, 10-4
MOVZX instruction . . . . . . . . . . . . . . . . . . . . . .6-26 Overflow exception (#OF). . . . . . . . . . . . . . . . 4-17
MTRRs (memory type range registers) Overflow, FPU exception
presence of . . . . . . . . . . . . . . . . . . . . . . . . .11-2 (see Numeric overflow exception)
MUL instruction . . . . . . . . . . . . . . . . . . . . . . . .6-27 Overflow, FPU stack. . . . . . . . . . . . . . . .7-51, 7-52

N P
NaN P6 family processors
description of . . . . . . . . . . . . . . . . . . . . 7-5, 7-8 microarchitecture. . . . . . . . . . . . . . . . . .2-6, 2-9
encoding of . . . . . . . . . . . . . . . . . . . . . 7-6, 7-27 Packed BCD integers . . . . . . . . . . . . . . . . . . . . 5-5
operating on . . . . . . . . . . . . . . . . . . . . . . . .7-43

INDEX-6
 INDEX

Packed bytes data type . . . . . . . . . . . . . . . . . . .8-3 stack pointer . . . . . . . . . . . . . . . . . . . . . . . . 4-1

Packed decimal indefinite . . . . . . . . . . . . . . . .7-30 stack segment . . . . . . . . . . . . . . . . . . . . . . . 4-1
Packed doublewords data type . . . . . . . . . . . . .8-3 stack-frame base pointer, EBP register . . . 4-4
Packed words data type. . . . . . . . . . . . . . . . . . .8-3 switching . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-9
Parameter passing top of stack . . . . . . . . . . . . . . . . . . . . . . . . . 4-1
argument list . . . . . . . . . . . . . . . . . . . . . . . . .4-7 width . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
FPU register stack . . . . . . . . . . . . . . . . . . .7-11 Processor identification
on procedure stack . . . . . . . . . . . . . . . . . . . .4-7 earlier Intel architecture processors . . . . . 11-4
on the procedure stack . . . . . . . . . . . . . . . . .4-7 using CPUID instruction . . . . . . . . . . . . . . 11-2
through general-purpose registers . . . . . . . .4-7 Processorstateinformation,savingonaprocedurecall
PC (precision) field, FPU control word . . . . . . .7-17 4-7
PE (inexact result exception) flag, Protected mode
FPU status word 7-13, 7-14, 7-19, 7-57 description of. . . . . . . . . . . . . . . . . . . . . . . . 3-4
PF (parity) flag, EFLAGS register . . . . . . . . . .3-12 I/O . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-4
Physical address space . . . . . . . . . . . . . . . . . . .3-2 Pseudo-denormal number . . . . . . . . . . . . . . . 7-30
Physical memory . . . . . . . . . . . . . . . . . . . . . . . .3-2 Pseudo-infinity . . . . . . . . . . . . . . . . . . . . . . . . 7-30
Pi Pseudo-NaN . . . . . . . . . . . . . . . . . . . . . . . . . . 7-30
description of FPU constant . . . . . . . . . . . .7-39 PUSH instruction . . . . . . . . . . 4-1, 4-3, 6-23, 6-43
Pointers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-5 PUSHA instruction . . . . . . . . . . . . . . . . . .4-8, 6-23
POP instruction . . . . . . . . . . . 4-1, 4-3, 6-24, 6-43 PUSHF instruction . . . . . . . . . . . . . 3-10, 4-8, 6-42
POPA instruction . . . . . . . . . . . . . . . . . . . 4-8, 6-24 PUSHFD instruction . . . . . . . . . . . . 3-10, 4-8, 6-42
POPF instruction . . . . . . . . . 3-10, 4-8, 6-42, 10-5
POPFD instruction . . . . . . . . . . . . . 3-10, 4-8, 6-42
Privilege levels Q
description of . . . . . . . . . . . . . . . . . . . . . . . .4-9 QNaN
inter-privilege level calls . . . . . . . . . . . . . . . .4-8 description of. . . . . . . . . . . . . . . . . . . . . . . . 7-8
stack switching . . . . . . . . . . . . . . . . . . . . . .4-13 operating on . . . . . . . . . . . . . . . . . . . . . . . 7-43
Procedure calls rules for generating . . . . . . . . . . . . . . . . . . 7-44
description of . . . . . . . . . . . . . . . . . . . . . . . .4-5 Quadword . . . . . . . . . . . . . . . . . . . . . . . . . .5-1, 8-3
far call . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-5 Quiet NaN (see QNaN)
for block-structured languages . . . . . . . . . .4-18
inter-privilege level call . . . . . . . . . . . . . . . .4-10 R
linking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-4
RC (rounding control) field, FPU
near call . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-5
control word . . . . . . . . . . . . . . .7-18, 9-8
overview . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-1
RCL instruction . . . . . . . . . . . . . . . . . . . . . . . . 6-33
procedure stack . . . . . . . . . . . . . . . . . . . . . .4-1
RCR instruction . . . . . . . . . . . . . . . . . . . . . . . 6-33
return instruction pointer (EIP register). . . . .4-4
RDMSR instruction . . . . . . . . . . . . . . . . . .6-2, 11-2
saving procedure state information. . . . . . . .4-7
RDPMC instruction . . . . . . . . . . . . . . . . . . . . . . 6-2
stack switching . . . . . . . . . . . . . . . . . . . . . . .4-9
RDTSC instruction . . . . . . . . . . . . . . . . . .6-2, 11-2
to exception handler procedure . . . . . . . . .4-13
Real numbers
to exception task. . . . . . . . . . . . . . . . . . . . .4-17
encoding . . . . . . . . . . . . . . . . . . . 7-5, 7-6, 7-27
to interrupt handler procedure . . . . . . . . . .4-13
floating-point format . . . . . . . . . . . . . . . . . 7-25
to interrupt task . . . . . . . . . . . . . . . . . . . . . .4-17
indefinite . . . . . . . . . . . . . . . . . . . . . . .7-27, 9-6
to other privilege levels . . . . . . . . . . . . . . . . .4-8
notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5
types of . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-1
system. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-2
Procedure stack
Real-address mode . . . . . . . . . . . . . . . . . . . . . 3-4
address-size attribute . . . . . . . . . . . . . . . . . .4-3
handling exceptions in. . . . . . . . . . . . . . . . 4-17
alignment of stack pointer. . . . . . . . . . . . . . .4-3
handling interrupts in. . . . . . . . . . . . . . . . . 4-17
current stack . . . . . . . . . . . . . . . . . . . . . 4-2, 4-4
description of . . . . . . . . . . . . . . . . . . . . . . . .4-1 Register operands . . . . . . . . . . . . . . . . . . . . . . 5-7
EIP register (return instruction pointer). . . . .4-4 Register stack, FPU . . . . . . . . . . . . . . . . . . . . . 7-9
maximum size. . . . . . . . . . . . . . . . . . . . . . . .4-1 Registers
number allowed . . . . . . . . . . . . . . . . . . . . . .4-1 EFLAGS register . . . . . . . . . . . . . . . . . . . . 3-10
passing parameters on . . . . . . . . . . . . . . . . .4-7 EIP register . . . . . . . . . . . . . . . . . . . . . . . . 3-14
popping values from . . . . . . . . . . . . . . . . . . .4-1 general-purpose registers . . . . . . . . . . .3-5, 3-6
segment registers . . . . . . . . . . . . . . . . .3-5, 3-7
procedure linking information . . . . . . . . . . . .4-4
Related literature . . . . . . . . . . . . . . . . . . . . . . . 1-9
pushing values on. . . . . . . . . . . . . . . . . . . . .4-1
REP/REPE/REPZ/REPNE/REPNZ
return instruction pointer . . . . . . . . . . . . . . . .4-4
prefixes. . . . . . . . . . . . . . . . . .6-40, 10-4
SS register . . . . . . . . . . . . . . . . . . . . . . . . . .4-1
Reserved bits . . . . . . . . . . . . . . . . . . . . . . . . . . 1-6

INDEX-7
RESET pin . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-10 SIMD (single-instruction, multiple-data)
RET instruction. . . . . . . . 3-14, 4-4, 4-5, 6-36, 6-44 execution model . . . . . . . . . . . . . . . . 8-4
Retirement unit. . . . . . . . . . . . . . . . . . . . . . . . .2-13 Sine, FPU operation . . . . . . . . . . . . . . . . . . . . 7-38
Return instruction pointer . . . . . . . . . . . . . . . . . .4-4 Single-precision, IEEE
Returns, from procedure calls floating-point format . . . . . . . . .7-25, 9-5
exception handler, return from . . . . . . . . . .4-13 Single-real floating-point format . . . . . . . .7-25, 9-5
far return . . . . . . . . . . . . . . . . . . . . . . . . . . . .4-6 SNaN
interrupt handler, return from . . . . . . . . . . .4-13 description of. . . . . . . . . . . . . . . . . . . . . . . . 7-8
Returns, from procedures calls operating on . . . . . . . . . . . . . . . . . . . . . . . 7-43
inter-privilege level return . . . . . . . . . . . . . .4-10 typical uses of . . . . . . . . . . . . . . . . . . . . . . 7-43
near return . . . . . . . . . . . . . . . . . . . . . . . . . .4-5 SP register . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-7
RF (resume) flag, EFLAGS register . . . . . . . . .3-13 Speculative execution. . . . . . . . . . . . . . . . .2-7, 2-8
ROL instruction . . . . . . . . . . . . . . . . . . . . . . . .6-33 SS register . . . . . . . . . . . . . . . . . . . . . 3-7, 3-9, 4-1
ROR instruction . . . . . . . . . . . . . . . . . . . . . . . .6-33 Stack alignment . . . . . . . . . . . . . . . . . . . . . . . . 4-3
Rounding Stack fault, FPU . . . . . . . . . . . . . . . . . . . . . . . 7-15
control, RC field of FPU control word . 7-18, 9-8 Stack overflow and underflow
modes, FPU . . . . . . . . . . . . . . . . . . . . 7-18, 9-8 exceptions (#IS), FPU . . . . . . . . . . 7-52
results, FPU . . . . . . . . . . . . . . . . . . . . . . . .7-19 Stack overflow exception, FPU. . . . . . . .7-13, 7-51
RSM instruction . . . . . . . . . . . . . . . . . . . . . . . . .6-2 Stack pointer (ESP register) . . . . . . . . . . . . . . . 4-1
Stack segment . . . . . . . . . . . . . . . . . . . . . . . . . 3-9
Stack switching
S on calls to interrupt and
SAHF instruction . . . . . . . . . . . . . . . . . . 3-10, 6-42 exception handlers. . . . . . . . . . . . . . . . 4-13
SAL instruction . . . . . . . . . . . . . . . . . . . . . . . . .6-29 on inter-privilege level calls . . . . . . . .4-10, 4-16
SAR instruction . . . . . . . . . . . . . . . . . . . . . . . .6-30 Stack underflow exception, FPU . . . . . .7-13, 7-51
Saturation arithmetic (MMX instructions) . . . . . .8-6 Stack (see Procedure stack)
Saving the FPU state . . . . . . . . . . . . . . . . . . . .7-21 Stack-frame base pointer, EBP register . . . . . . 4-4
SBB instruction. . . . . . . . . . . . . . . . . . . . . . . . .6-26 Status flags, EFLAGS register . . 3-12, 7-15, 7-16,
Scale (operand addressing) . . . . . . . . . . . 5-9, 5-10 7-37
Scale, FPU operation . . . . . . . . . . . . . . . . . . . .7-40 STC instruction . . . . . . . . . . . . . . . . . . . .3-12, 6-42
Scaling bias value . . . . . . . . . . . . . . . . . 7-55, 7-56 STD instruction . . . . . . . . . . . . . . . . . . . .3-13, 6-42
SCAS instruction . . . . . . . . . . . . . . . . . . 3-13, 6-40 STI instruction. . . . . . . . . . . . . . . . 6-42, 6-43, 10-4
Segment registers STOS instruction . . . . . . . . . . . . . . . . . .3-13, 6-40
description of . . . . . . . . . . . . . . . . . . . . 3-5, 3-7 Strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-5
Segment selector ST(0), top-of-stack register. . . . . . . . . . . . . . . 7-11
description of . . . . . . . . . . . . . . . . . . . . 3-3, 3-7 SUB instruction. . . . . . . . . . . . . . . . . . . . . . . . 6-26
specifying . . . . . . . . . . . . . . . . . . . . . . . . . . .5-8 Superscalar . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-7
Segmented addressing . . . . . . . . . . . . . . . . . . .1-7 Synchronization, of floating-point exceptions . 7-58
Segmented memory model . . . . . . . . . . . . 3-3, 3-8 System flags, EFLAGS register . . . . . . . . . . . 3-13
Segments System management mode (SSM) . . . . . . . . . 3-4
defined . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-3
maximum number . . . . . . . . . . . . . . . . . . . . .3-3
Serialization of I/O instructions. . . . . . . . . . . . .10-6 T
SETcc instructions . . . . . . . . . . . . . . . . . 3-12, 6-34 Tangent, FPU operation . . . . . . . . . . . . . . . . . 7-38
SF (sign) flag, EFLAGS register. . . . . . . . . . . .3-12 Task gate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-17
SF (stack fault) flag, FPU status word . . 7-15, 7-52 Task state segment (see TSS)
SHL instruction. . . . . . . . . . . . . . . . . . . . . . . . .6-29 Tasks
SHLD instruction . . . . . . . . . . . . . . . . . . . . . . .6-32 exception handler . . . . . . . . . . . . . . . . . . . 4-17
SHR instruction . . . . . . . . . . . . . . . . . . . . . . . .6-29 interrupt handler . . . . . . . . . . . . . . . . . . . . 4-17
SHRD instruction . . . . . . . . . . . . . . . . . . . . . . .6-32 TEST instruction . . . . . . . . . . . . . . . . . . . . . . . 6-35
SI register. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3-7 TF (trap) flag, EFLAGS register . . . . . . . . . . . 3-13
Signaling NaN (see SNaN) Tiny number . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-7
Signed infinity. . . . . . . . . . . . . . . . . . . . . . . . . . .7-8 TOP (stack TOP) field, FPU status word . . . . . 7-9
Signed zero . . . . . . . . . . . . . . . . . . . . . . . . . . . .7-6 Transcendental instruction accuracy . . . . . . . 7-40
Significand Trap gate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-13
of floating-point number . . . . . . . . . . . . . . . .7-4 TSS
Sign, floating-point number . . . . . . . . . . . . . . . .7-4 I/O map base. . . . . . . . . . . . . . . . . . . . . . . 10-5
I/O permission bit map . . . . . . . . . . . . . . . 10-5
 INDEX

saving state of EFLAGS register . . . . . . . .3-10

U
UD2 instruction. . . . . . . . . . . . . . . . . . . . . 6-2, 6-45
UE (numeric overflow exception) flag,
FPU status word . . . . . . . . . . 7-14, 7-56
Underflow, FPU exception
(see Numeric underflow exception)
Underflow, FPU stack . . . . . . . . . . . . . . 7-51, 7-52
Underflow, numeric . . . . . . . . . . . . . . . . . . . . . .7-7
Un-normal number . . . . . . . . . . . . . . . . . . . . . .7-30
Unsigned integers . . . . . . . . . . . . . 5-5, 6-26, 6-27
Unsupported floating-point formats . . . . . . . . .7-30
Unsupported FPU instructions . . . . . . . . . . . . .7-43

V
Vector (see Interrupt vector)
VIF (virtual interrupt) flag, EFLAGS register . .3-13
VIP (virtual interrupt pending) flag,
EFLAGS register . . . . . . . . . . . . . . .3-13
Virtual 8086 mode
description of . . . . . . . . . . . . . . . . . . . . . . .3-13
memory model . . . . . . . . . . . . . . . . . . . . . . .3-4
VM (virtual 8086 mode) flag, EFLAGS register 3-13

W
Waiting instructions . . . . . . . . . . . . . . . . . . . . .7-42
WAIT/FWAIT instructions. . . . . . . . . . . . 7-42, 7-59
WBINVD instruction . . . . . . . . . . . . . . . . . . . . . .6-3
Word. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5-1
Wraparound mode (MMX instructions) . . . . . . .8-6
WRMSR instruction . . . . . . . . . . . . . . . . . 6-2, 11-2

X
XADD instruction . . . . . . . . . . . . . . . . . . . 6-3, 6-22
XCHG instruction . . . . . . . . . . . . . . . . . . . . . . .6-21
XLAT/XLATB instruction . . . . . . . . . . . . . . . . .6-45
XOR instruction . . . . . . . . . . . . . . . . . . . . . . . .6-29

Z
ZE (division-by-zero exception) flag,
FPU status word . . . . . . . . . . . . . . .7-14
Zero, floating-point format . . . . . . . . . . . . . . . . .7-6
ZF (zero) flag, EFLAGS register . . . . . . . . . . .3-12

INDEX-9