Write An LLVM Backend Tutorial For Cpu0

Release 3.2.3

Chen Chung-Shu [email protected] Anoushe Jamshidi [email protected]

January 17, 2013

CONTENTS

1 2 3 4 5 6 7 8 9

About Getting Started: Installing LLVM and the Cpu0 example code Introduce Cpu0 and LLVM LLVM Backend Structure Adding arithmetic and local pointer support Generating object les Global variables, structs and arrays Control ow statements Function call

3 7 9 33 53 79 93 109 121 165 175 181 183 185

10 ELF Support 11 Porting to LLVM 3.2 12 Todo List 13 Book example code 14 Alternate formats

ii

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Warning: This is a work in progress. If you would like to contribution, please push updates and patches to the main github project available at http://github.com/Jonathan2251/lbd for review.

CONTENTS

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

CONTENTS

CHAPTER

ONE

ABOUT
1.1 Authors

Chen Chung-Shu [email protected] http://jonathan2251.github.com/web/index.html Anoushe Jamshidi [email protected]

1.2 Revision history

Version 3.2.3, Released January 12, 2013 Add chapter Porting to LLVM 3.2. Version 3.2.2, Released January 10, 2013 Add section Full support % and section Verify DIV for operator %. Version 3.2.1, Released January 7, 2013 Add Footnote for references. Reorganize chapters. Version 3.2.0, Released January 1, 2013 Add chapter Function. Move Chapter Installing LLVM and the Cpu0 example code from beginning to Appendix A. Add subsection Install other tools on Linux. Add chapter ELF. Version 3.1.2, Released December 15, 2012 Fix section 6.1 error by add def : Pat<(brcond RC:$cond, bb:$dst), (JNEOp (CMPOp RC:$cond, ZEROReg), bb:$dst)>; in last pattern. Modify section 5.5 Fix bug Cpu0InstrInfo.cpp SW to ST. Correct LW to LD; LB to LDB; SB to STB. Version 3.1.1, Released November 28, 2012 Add Revision history. Correct ldi instruction error (replace ldi instruction with addiu from the beginning and in the all example code). Move ldi instruction change from section of Adjust cpu0 instruction and support type of local variable pointer to Section CPU0 processor architecture. Correct some English & typing errors.

1.3 Licensing
Todo Add info about LLVM documentation licensing.

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

1.4 Preface
The LLVM Compiler Infrastructure provides a versatile structure for creating new backends. Creating a new backend should not be too difcult once you familiarize yourself with this structure. However, the available backend documentation is fairly high level and leaves out many details. This tutorial will provide step-by-step instructions to write a new backend for a new target architecture from scratch. We will use the Cpu0 architecture as an example to build our new backend. Cpu0 is a simple RISC architecture that has been designed for educational purposes. More information about Cpu0, including its instruction set, is available here. The Cpu0 example code referenced in this book can be found here. As you progress from one chapter to the next, you will incrementally build the backends functionality. This tutorial was written using the LLVM 3.1 Mips backend as a reference. Since Cpu0 is an educational architecture, it is missing some key pieces of documentation needed when developing a compiler, such as an Application Binary Interface (ABI). We implement our backend borrowing information from the Mips ABI as a guide. You may want to familiarize yourself with the relevant parts of the Mips ABI as you progress through this tutorial.

1.5 Prerequisites
Readers should be comfortable with the C++ language and Object-Oriented Programming concepts. LLVM has been developed and implemented in C++, and it is written in a modular way so that various classes can be adapted and reused as often as possible. Already having conceptual knowledge of how compilers work is a plus, and if you already have implemented compilers in the past you will likely have no trouble following this tutorial. As this tutorial will build up an LLVM backend stepby-step, we will introduce important concepts as necessary. This tutorial references the following materials. We highly recommend you read these documents to get a deeper understanding of what the tutorial is teaching: The Architecture of Open Source Applications Chapter on LLVM LLVMs Target-Independent Code Generation documentation LLVMs TableGen Fundamentals documentation LLVMs Writing an LLVM Compiler Backend documentation Description of the Tricore LLVM Backend Mips ABI document

1.6 Outline of Sections

sec-llvmstructure: This section introduces the Cpu0 architecture, a high-level view of LLVM, and how Cpu0 will be targeted in in an LLVM backend. This section will run you through the initial steps of building the backend, including initial work on the target description (td), setting up cmake and LLVMBuild les, and target registration. Around 750 lines of source code are added by the end of this section. sec-backendstructure: This section highlights the structure of an LLVM backend using by UML graphs, and we continue to build the Cpu0 backend. Around 2300 lines of source code are added in this section, most of which is common from one LLVM

Chapter 1. About

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

backends to another, regardless of the target architecture. By the end of this section, the Cpu0 LLVM backend will support three instructions to generate some initial assembly output. Adding arithmetic and local pointer support: Over ten C operators and their corresponding LLVM IR instructions are introduced in this section. Around 345 lines of source code, mostly in .td Target Description les, are added. With these 345 lines, the backend can now translate the +, -, *, /, &, |, ^, <<, >>, ! and % C operators into the appropriate Cpu0 assembly code. Use of the llc debug option and of Graphviz as a debug tool are introduced in this section. Generating object les: Object le generation support for the Cpu0 backend is added in this section, as the Target Registration structure is introduced. With 700 lines of additional code, the Cpu0 backend can now generate big and little endian object les. Global variables, structs and arrays: Global variable, struct and array support are added in this section. About 300 lines of source code are added to do this. The Cpu0 supports PIC and static addressing mode, both of which area explained as they are added in this section. Control ow statements: Support for the if, else, while, for, goto ow control statements are added in this section. Around 150 lines of source code added. Function call: This section details the implementation of function calls in the Cpu0 backend. The stack frame, handling incoming & outgoing arguments, and their corresponding standard LLVM functions are introduced in this chapter. Over 700 lines of source code are added in this section. ELF Support: This section details Cpu0 support for the well-known ELF object le format. The ELF format and binutils tools are not a part of LLVM, but are introduced. This section details how to use this ELF tool to verify and analyze the object les created by the Cpu0 backend. Porting to LLVM 3.2: Introduces the changes of the LLVM APIs used by Cpu0 and Mips when updating this guide from LLVM 3.1 to 3.2. sec-appendix-installing: Details how to set up the LLVM source code, development tools, and environment setting for Mac OS X and Linux platforms.

1.6. Outline of Sections

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Chapter 1. About

CHAPTER

TWO

GETTING STARTED: INSTALLING LLVM AND THE CPU0 EXAMPLE CODE

In this chapter, we will run through how to set up LLVM using if you are using Mac OS X or Linux. When discussing Mac OS X, we are using Apples Xcode IDE (version 4.5.1) running on Mac OS X Mountain Lion (version 10.8) to modify and build LLVM from source, and we will be debugging using lldb. We cannot debug our LLVM builds within Xcode at the moment, but if you have experience with this, please contact us and help us build documentation that covers this. For Linux machines, we are building and debugging (using gdb) our LLVM installations on a Fedora 17 system. We will not be using an IDE for Linux, but once again, if you have experience building/ debugging LLVM using Eclipse or other major IDEs, please contact the authors. For information on using cmake to build LLVM, please refer to the Building LLVM with CMake 1 documentation for further information. We are using cmake version 2.8.9. We will install two llvm directories in this chapter. One is the directory llvm/3.1/ which contains the clang, clang++ compiler we will use to translate the C/C++ input le into llvm IR. The other is the directory llvm/3.1.test/cpu0/1 which contains our cpu0 backend program and without clang and clang++.

2.1 Install LLVM 3.1 release on Linux

Our tutorial is based on LLVM 3.1 release. We need LLVM, clang and compiler-rt. In this section, we will show the command line steps to build and install LLVM on your Linux machine. We use the following names to denote paths specic to your enviroment. LLVM_SRC This is the top level directory of the LLVM source tree. Example_SRC This is the top level directory where our example code lives. In our tutorial, the example is a git submodule stored in lbd-example directory. Debug_OBJ This is the top level directory of the LLVM object tree (i.e. the tree where object les and compiled programs will be placed). We build LLVM in debug mode here. INSTALL The resulting binary will be installed in this directory. If you are using IDE, please see appendix for more information. 1. Download the source code.
1

http://www.llvm.org/docs/CMake.html

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

$ $ $ $ $ $ $ $

wget http://llvm.org/releases/3.1/llvm-3.1.src.tar.gz wget http://llvm.org/releases/3.1/clang-3.1.src.tar.gz wget http://llvm.org/releases/3.1/compiler-rt-3.1.src.tar.gz tar xvf llvm-3.1.src.tar.gz tar xvf clang-3.1.src.tar.gz -C llvm-3.1.src/tools/ tar xvf compiler-rt-3.1.src.tar.gz -C llvm-3.1.src/projects/ mv llvm-3.1.src/tools/clang-3.1.src llvm-3.1.src/tools/clang mv llvm-3.1.src/projects/compiler-rt-3.1.src llvm-3.1.src/projects/compiler-rt

2. Build LLVM
$ mkdir obj; cd obj $ ../llvm-3.1.src/configure --prefix=$INSTALL --enable-targets=host $ make -j; make install

3. Set $PATH
$ echo "PATH=$INSTALL/bin:$PATH" > $HOME/.bash_profile

2.2 Install cpu0 debug build on Linux

1. Download the example code.
$ git clone git://github.com/azru0512/lbd-example.git example

2. Make a new directory.

$ mkdir -p cpu0/2/src; cd cpu0/2/src $ cp -rf $LLVM_SRC/* .

3. Update top-level source code.

$ cp -rf $Example_SRC/top-level/* .

4. Update subdirecotry lib/Target.

$ cp -rf $Example_SRC/2/Cpu0/ lib/Target/

5. Remove clang source tree since we dont need it at this moment. Build LLVM in debug mode.
cd .. mkdir debug; cd debug rm -rf ../src/tools/clang/ cmake -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang \ -DLLVM_TARGETS_TO_BUILD="Mips;Cpu0" \ -DCMAKE_BUILD_TYPE=Debug -G "Unix Makefiles" ../src/ $ make $ $ $ $

6. Debug
$ gdb bin/llc

Chapter 2. Getting Started: Installing LLVM and the Cpu0 example code

CHAPTER

THREE

INTRODUCE CPU0 AND LLVM

Before you start, you should know that you can always examine existing LLVM backend code and attempt to port what you nd for your own target architecture . The majority of this code can be found in the /lib/Target directory of your root LLVM directory. As most major RISC instruction set architectures have some similarities, this may be the avenue you might try if you are both an experienced programmer and knowledgable of compiler backends. However, there is a steep learning curve and you may easily get held up debugging your new backend. You can easily spend a lot of time tracing which methods are callbacks of some function, or which are calling some overridden method deep in the LLVM codebase - and with a codebase as large as LLVM, this can easily become a headache. This tutorial will help you work through this process while learning the fundamentals of LLVM backend design. It will show you what is necessary to get your rst backend functional and complete, and it should help you understand how to debug your backend when it does not produce desirable output using the output provided by LLVM. This chapter shows you the cpu0 instruction format rst. Next, the llvm structure is introduced to you by copy and paste the related article from llvm web site. The llvm structure introduced here is extracted from the asop web site. You can read the whole article from the asop web site. After that we will show you how to write register and instruction denitions (Target Description File) which will be used in next chapter.

3.1 Cpu0 Architecture

We copy and redraw gures in english in this section. This web site 1 is chinese version and here 2 is english version.

3.1.1 Introduction
CPU0 is a 32-bit processor which has registers R0 .. R15, IR, MAR, MDR, etc., and its structure is shown below. Uses of each register as follows:

3.1.2 Instruction Set

The CPU0 instruction divided into three types, L-type usually load the saved instruction, A-type arithmetic instructionbased J-type usually jump instruction, the following gure shows the three types of instruction encoding format. The following is the CPU0 processors instruction table format In the second edition of CPU0_v2 we ll the following command:
1 2

http://ccckmit.wikidot.com/ocs:cpu0 http://translate.google.com.tw/translate?js=n&prev=_t&hl=zh-TW&ie=UTF-8&layout=2&eotf=1&sl=zh-CN&tl=en&u=http://ccckmit.wikidot.com/ocs:cpu0

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.1: The structure of the processor of CPU0

Figure 3.2: Cpu0 registers table

10

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.3: CPU0 three instruction formats

3.1.3 Status Register

CPU0 status register contains the state of the N, Z, C, V, and I, T and other interrupt mode bit. Its structure is shown below. When CMP Ra, Rb instruction execution, the state ag will thus change. If Ra> Rb, then the setting state of N = 0, Z = 0. If Ra <Rb, it will set the state of N = 1, Z = 0. If Ra = Rb, then the setting state of N = 0, Z = 1. So conditional jump the JGT, JLT, JGE, JLE, JEQ, JNE instruction jumps N, Z ag in the status register.

3.1.4 Pipeline
CPU0 has three stage pipeline: Instruction fetch, Decode and Execution. 1. Instruction fetch Action 1. The instruction fetch: IR = [PC] Action 2. Update program counter: PC = PC + 4 2. Decode Action 3. Decode: Control unit decodes IR, then set data ow switch and ALU operation mode. 3. Execute Action 4. Execute: Data ow into ALU. After ALU done the operation, the result stored back into destination register.

3.1.5 Instruction Change

We have recognized the ldi instruction is a bad design and replace it with mips instruction addiu. The reason we replace ldi with addiu is that ldi use only one register even though ldi is L type format and has two registers, as 3.7. Mips addiu which allow programmer to do load constant to register like ldi, and add constant to a register. So, its powerful and fully contains the ldi ability.

3.1. Cpu0 Architecture

11

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.4: CPU0 instruction table

12

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.5: CPU0_v2 instruction table

Figure 3.6: CPU0 status register The two instruction formats are shown in gure 3.7 and gure 3.8.

Figure 3.7: Cpu0 ldi instruction From gure 3.7 and gure 3.8, you can nd ldi $Ra, 5 can be replaced by addiu $Ra, $zero, 5. And more, addiu can do addiu $Ra, $Rb, 5 which add $Rb and 5 then save to $Ra, but ldi cannot. As a CPU design, its common to redesign CPU instruction when nd a better solution during design the compiler backend for that CPU. So, we add addiu instruction to cpu0. The cpu0 is my brothers work, I will nd time to talk with him.

3.2 LLVM Structure

3.2.1 Overview

gure 3.9 () gure 3.10

3.2. LLVM Structure 13

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.8: Mips addiu instruction format

Figure 3.9: Compilation Pipeline

14

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.10: Compilation Flow LLVM

gure 3.11 1. LLVM IR : 2. LLVM IR : .o le writer 3. : 4. :

Figure 3.11: LLVM

3.2.2 Detail
Following came from AOSA 3 . The most popular design for a traditional static compiler (like most C compilers) is the three phase design whose major components are the front end, the optimizer and the back end (gure 3.12). The front end parses source code, checking it for errors, and builds a language-specic Abstract Syntax Tree (AST) to represent the input code. The AST is optionally converted to a new representation for optimization, and the optimizer and back end are run on the code. The optimizer is responsible for doing a broad variety of transformations to try to improve the codes running time, such as eliminating redundant computations, and is usually more or less independent of language and target. The back end (also known as the code generator) then maps the code onto the target instruction set. In addition to making correct code, it is responsible for generating good code that takes advantage of unusual features of the supported architecture. Common parts of a compiler back end include instruction selection, register allocation, and instruction scheduling. This model applies equally well to interpreters and JIT compilers. The Java Virtual Machine (JVM) is also an implementation of this model, which uses Java bytecode as the interface between the front end and optimizer.
3

http://www.aosabook.org/en/llvm.html

3.2. LLVM Structure

15

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 3.12: Three major components of a Three Phase Compiler The most important win of this classical design comes when a compiler decides to support multiple source languages or target architectures. If the compiler uses a common code representation in its optimizer, then a front end can be written for any language that can compile to it, and a back end can be written for any target that can compile from it, as shown in gure 3.13.

Figure 3.13: Retargetablity With this design, porting the compiler to support a new source language (e.g., Algol or BASIC) requires implementing a new front end, but the existing optimizer and back end can be reused. If these parts werent separated, implementing a new source language would require starting over from scratch, so supporting N targets and M source languages would need N*M compilers. Another advantage of the three-phase design (which follows directly from retargetability) is that the compiler serves a broader set of programmers than it would if it only supported one source language and one target. For an open source project, this means that there is a larger community of potential contributors to draw from, which naturally leads to more enhancements and improvements to the compiler. This is the reason why open source compilers that serve many communities (like GCC) tend to generate better optimized machine code than narrower compilers like FreePASCAL. This isnt the case for proprietary compilers, whose quality is directly related to the projects budget. For example, the Intel ICC Compiler is widely known for the quality of code it generates, even though it serves a narrow audience. A nal major win of the three-phase design is that the skills required to implement a front end are different than those required for the optimizer and back end. Separating these makes it easier for a front-end person to enhance and maintain their part of the compiler. While this is a social issue, not a technical one, it matters a lot in practice, particularly for open source projects that want to reduce the barrier to contributing as much as possible. The most important aspect of its design is the LLVM Intermediate Representation (IR), which is the form it uses to represent code in the compiler. LLVM IR is designed to host mid-level analyses and transformations that you nd in the optimizer section of a compiler. It was designed with many specic goals in mind, including supporting lightweight runtime optimizations, cross-function/interprocedural optimizations, whole program analysis, and aggressive restructuring transformations, etc. The most important aspect of it, though, is that it is itself dened as a rst class language 16 Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

with well-dened semantics. To make this concrete, here is a simple example of a .ll le:
define i32 @add1(i32 %a, i32 %b) { entry: %tmp1 = add i32 %a, %b ret i32 %tmp1 } define i32 @add2(i32 %a, i32 %b) { entry: %tmp1 = icmp eq i32 %a, 0 br i1 %tmp1, label %done, label %recurse recurse: %tmp2 = sub i32 %a, 1 %tmp3 = add i32 %b, 1 %tmp4 = call i32 @add2(i32 %tmp2, i32 %tmp3) ret i32 %tmp4 done: ret i32 %b } This LLVM IR corresponds to this C code, which provides two different ways to add integers: unsigned add1(unsigned a, unsigned b) { return a+b; } // Perhaps not the most efficient way to add two numbers. unsigned add2(unsigned a, unsigned b) { if (a == 0) return b; return add2(a-1, b+1); }

As you can see from this example, LLVM IR is a low-level RISC-like virtual instruction set. Like a real RISC instruction set, it supports linear sequences of simple instructions like add, subtract, compare, and branch. These instructions are in three address form, which means that they take some number of inputs and produce a result in a different register. LLVM IR supports labels and generally looks like a weird form of assembly language. Unlike most RISC instruction sets, LLVM is strongly typed with a simple type system (e.g., i32 is a 32-bit integer, i32** is a pointer to pointer to 32-bit integer) and some details of the machine are abstracted away. For example, the calling convention is abstracted through call and ret instructions and explicit arguments. Another signicant difference from machine code is that the LLVM IR doesnt use a xed set of named registers, it uses an innite set of temporaries named with a % character. Beyond being implemented as a language, LLVM IR is actually dened in three isomorphic forms: the textual format above, an in-memory data structure inspected and modied by optimizations themselves, and an efcient and dense on-disk binary bitcode format. The LLVM Project also provides tools to convert the on-disk format from text to binary: llvm-as assembles the textual .ll le into a .bc le containing the bitcode goop and llvm-dis turns a .bc le into a .ll le. The intermediate representation of a compiler is interesting because it can be a perfect world for the compiler optimizer: unlike the front end and back end of the compiler, the optimizer isnt constrained by either a specic source language or a specic target machine. On the other hand, it has to serve both well: it has to be designed to be easy for a front end to generate and be expressive enough to allow important optimizations to be performed for real targets.

3.2.3 Introduce Target Description File

Note: This section comes from AOSA. http://www.ituring.com.cn/article/4206

3.2. LLVM Structure

17

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

The mix and match approach allows target authors to choose what makes sense for their architecture and permits a large amount of code reuse across different targets. This brings up another challenge: each shared component needs to be able to reason about target specic properties in a generic way. For example, a shared register allocator needs to know the register le of each target and the constraints that exist between instructions and their register operands. LLVMs solution to this is for each target to provide a target description in a declarative domain-specic language (a set of .td les) processed by the tblgen tool. The (simplied) build process for the x86 target is shown in gure 3.14 .

Figure 3.14: Simplied x86 Target Denition The different subsystems supported by the .td les allow target authors to build up the different pieces of their target. For example, the x86 back end denes a register class that holds all of its 32-bit registers named GR32 (in the .td les, target specic denitions are all caps) like this:
def GR32 : RegisterClass<[i32], 32, [EAX, ECX, EDX, ESI, EDI, EBX, EBP, ESP, R8D, R9D, R10D, R11D, R14D, R15D, R12D, R13D]> { ... }

3.2.4 Writing Target Description File

LLVM use target description les (.td) to describe target register le and instruction format. After you writing up those .td les, LLVM generates C++ les (*.inc) by llvm-tblgen. Those *.inc les are C++ les with table driven in concept. 4 is the web site. You should write one XX.td under your target subdirectory, it then includes other target description les dening various target information. The syntax of target description le is pretty similar to C++. Below is Cpu0.td.
//===----------------------------------------------------------------------===// // Target-independent interfaces //===----------------------------------------------------------------------===// include "llvm/Target/Target.td"
4

http://llvm.org/docs/TableGenFundamentals.html

18

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

//===----------------------------------------------------------------------===// // Register File, Calling Conv, Instruction Descriptions //===----------------------------------------------------------------------===// include "Cpu0RegisterInfo.td" include "Cpu0Schedule.td" include "Cpu0InstrInfo.td" def Cpu0InstrInfo : InstrInfo; def Cpu0 : Target { // def Cpu0InstrInfo : InstrInfo as before. let InstructionSet = Cpu0InstrInfo; }

As shown above, Cpu0.td includes Cpu0RegisterInfo.td which denes the Cpu0 register le.
//===----------------------------------------------------------------------===// // Declarations that describe the CPU0 register file //===----------------------------------------------------------------------===// // We have banks of 16 registers each. class Cpu0Reg<string n> : Register<n> { field bits<4> Num; let Namespace = "Cpu0"; } // Cpu0 CPU Registers class Cpu0GPRReg<bits<4> num, string n> : Cpu0Reg<n> { let Num = num; } //===----------------------------------------------------------------------===// // Registers //===----------------------------------------------------------------------===// let Namespace = "Cpu0" in { // General Purpose Registers def ZERO : Cpu0GPRReg< 0, "ZERO">, def AT : Cpu0GPRReg< 1, "AT">, : Cpu0GPRReg< 2, "2">, def V0 def V1 : Cpu0GPRReg< 3, "3">, : Cpu0GPRReg< 4, "4">, def A0 def A1 : Cpu0GPRReg< 5, "5">, def T9 : Cpu0GPRReg< 6, "6">, def S0 : Cpu0GPRReg< 7, "7">, def S1 : Cpu0GPRReg< 8, "8">, def S2 : Cpu0GPRReg< 9, "9">, def GP : Cpu0GPRReg< 10, "GP">, def FP : Cpu0GPRReg< 11, "FP">, : Cpu0GPRReg< 12, "SW">, def SW def SP : Cpu0GPRReg< 13, "SP">, def LR : Cpu0GPRReg< 14, "LR">, : Cpu0GPRReg< 15, "PC">, def PC }

DwarfRegNum<[0]>; DwarfRegNum<[1]>; DwarfRegNum<[2]>; DwarfRegNum<[3]>; DwarfRegNum<[6]>; DwarfRegNum<[7]>; DwarfRegNum<[6]>; DwarfRegNum<[7]>; DwarfRegNum<[8]>; DwarfRegNum<[9]>; DwarfRegNum<[10]>; DwarfRegNum<[11]>; DwarfRegNum<[12]>; DwarfRegNum<[13]>; DwarfRegNum<[14]>; DwarfRegNum<[15]>;

//===----------------------------------------------------------------------===// // Register Classes //===----------------------------------------------------------------------===//

3.2. LLVM Structure

19

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

def CPURegs : RegisterClass<"Cpu0", [i32], 32, (add // Return Values and Arguments V0, V1, A0, A1, // Not preserved across procedure calls T9, // Callee save S0, S1, S2, // Reserved ZERO, AT, GP, FP, SW, SP, LR, PC)>;

In C++ the data layout is declared by class. Declaration tells the variable layout; denition allocates memory for the variable. For example,
class Date { // declare Date int year, month, day; }; Date date; // define(instance) date

Just like C++ class, the keyword class is used for declaring data structure layout. Cpu0Reg<string n> declare a derived class from Register<n> which is declared by llvm already, where n is the argument of type string. In addition to inherited from all the elds of Register class, Cpu0Reg add a new eld Num of type 4 bits. Namespace is same with C++ namespace. Def is used by dene(instance) a concrete variable. As above, we dene a ZERO register which type is Cpu0GPRReg, its eld Num is 0 (4 bits) and eld n is ZERO (declared in Register class). Note the use of let expressions to override values that are initially dened in a superclass. For example, let Namespace = Cpu0 in class Cpu0Reg, will override Namespace declared in Register class. The Cpu0RegisterInfo.td also dene that CPURegs is a variable for type of RegisterClass, where the RegisterClass is a llvm built-in class. The type of RegisterClass is a set/group of Register, so CPURegs variable is dened with a set of Register. Cpu0InstrFormats.td denes Cpu0 instruction format. We dene a base class Cpu0Inst, then dene class FA, FI and FJ for each Cpu0 instruction types.
// Generic Cpu0 Format class Cpu0Inst<dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin, Format f>: Instruction { field bits<32> Inst; Format Form = f; let Namespace = "Cpu0"; let Size = 4; bits<8> Opcode = 0; // Top 8 bits are the opcode field let Inst{31-24} = Opcode; let OutOperandList = outs; let InOperandList = ins; let AsmString let Pattern let Itinerary = asmstr; = pattern; = itin;

// // Attributes specific to Cpu0 instructions...

20

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// bits<4> FormBits = Form.Value; // TSFlags layout should be kept in sync with Cpu0InstrInfo.h. let TSFlags{3-0} = FormBits; let DecoderNamespace = "Cpu0"; field bits<32> SoftFail = 0; } //===----------------------------------------------------------------------===// // Format A instruction class in Cpu0 : <|opcode|ra|rb|rc|cx|> //===----------------------------------------------------------------------===// class FA<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: Cpu0Inst<outs, ins, asmstr, pattern, itin, FrmA> { bits<4> ra; bits<4> rb; bits<4> rc; bits<12> imm12; let Opcode = op; let let let let } //===----------------------------------------------------------------------===// // Format I instruction class in Cpu0 : <|opcode|ra|rb|cx|> //===----------------------------------------------------------------------===// class FL<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: Cpu0Inst<outs, ins, asmstr, pattern, itin, FrmL> { bits<4> ra; bits<4> rb; bits<16> imm16; let Opcode = op; let Inst{23-20} = ra; let Inst{19-16} = rb; let Inst{15-0} = imm16; } //===----------------------------------------------------------------------===// // Format J instruction class in Cpu0 : <|opcode|address|> //===----------------------------------------------------------------------===// class FJ<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: Cpu0Inst<outs, ins, asmstr, pattern, itin, FrmJ> { bits<24> addr; Inst{23-20} Inst{19-16} Inst{15-12} Inst{11-0} = = = = ra; rb; rc; imm12;

3.2. LLVM Structure

21

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

let Opcode = op; let Inst{23-0} = addr; }

The cpu0 instructions td is named to Cpu0InstrInfo.td which contents as follows,

/===- Cpu0InstrInfo.td - Target Description for Cpu0 Target -*- tablegen -*-=// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This file contains the Cpu0 implementation of the TargetInstrInfo class. // //===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===// // Instruction format superclass //===----------------------------------------------------------------------===// include "Cpu0InstrFormats.td" //===----------------------------------------------------------------------===// // Cpu0 profiles and nodes //===----------------------------------------------------------------------===// def SDT_Cpu0Ret : SDTypeProfile<0, 1, [SDTCisInt<0>]>; // Return def Cpu0Ret : SDNode<"Cpu0ISD::Ret", SDT_Cpu0Ret, [SDNPHasChain, SDNPOptInGlue]>; //===----------------------------------------------------------------------===// // Cpu0 Operand, Complex Patterns and Transformations Definitions. //===----------------------------------------------------------------------===// def simm16 : Operand<i32> { let DecoderMethod= "DecodeSimm16"; } // Address operand def mem : Operand<i32> { let PrintMethod = "printMemOperand"; let MIOperandInfo = (ops CPURegs, simm16); let EncoderMethod = "getMemEncoding"; } // Node immediate fits as 16-bit sign extended on target immediate. // e.g. addiu def immSExt16 : PatLeaf<(imm), [{ return isInt<16>(N->getSExtValue()); }]>; // Cpu0 Address Mode! SDNode frameindex could possibily be a match // since load and store instructions from stack used it. def addr : ComplexPattern<iPTR, 2, "SelectAddr", [frameindex], [SDNPWantParent]> ; //===----------------------------------------------------------------------===// // Pattern fragment for load/store //===----------------------------------------------------------------------===// class AlignedLoad<PatFrag Node> : PatFrag<(ops node:$ptr), (Node node:$ptr), [{ LoadSDNode *LD = cast<LoadSDNode>(N); return LD->getMemoryVT().getSizeInBits()/8 <= LD->getAlignment();

22

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

}]>; class AlignedStore<PatFrag Node> : PatFrag<(ops node:$val, node:$ptr), (Node node:$val, node:$ptr), [{ StoreSDNode *SD = cast<StoreSDNode>(N); return SD->getMemoryVT().getSizeInBits()/8 <= SD->getAlignment(); }]>; // Load/Store PatFrags. def load_a : AlignedLoad<load>; def store_a : AlignedStore<store>; //===----------------------------------------------------------------------===// // Instructions specific format //===----------------------------------------------------------------------===// // Arithmetic and logical instructions with 2 register operands. class ArithLogicI<bits<8> op, string instr_asm, SDNode OpNode, Operand Od, PatLeaf imm_type, RegisterClass RC> : FL<op, (outs RC:$ra), (ins RC:$rb, Od:$imm16), !strconcat(instr_asm, "\t$ra, $rb, $imm16"), [(set RC:$ra, (OpNode RC:$rb, imm_type:$imm16))], IIAlu> { let isReMaterializable = 1; } // Move immediate imm16 to register ra. class MoveImm<bits<8> op, string instr_asm, SDNode OpNode, Operand Od, PatLeaf imm_type, RegisterClass RC> : FL<op, (outs RC:$ra), (ins RC:$rb, Od:$imm16), !strconcat(instr_asm, "\t$ra, $imm16"), [(set RC:$ra, (OpNode RC:$rb, imm_type:$imm16))], IIAlu> { let rb = 0; let isReMaterializable = 1; } class FMem<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: FL<op, outs, ins, asmstr, pattern, itin> { bits<20> addr; let Inst{19-16} = addr{19-16}; let Inst{15-0} = addr{15-0}; let DecoderMethod = "DecodeMem"; } // Memory Load/Store let canFoldAsLoad = 1 in class LoadM<bits<8> op, string instr_asm, PatFrag OpNode, RegisterClass RC, Operand MemOpnd, bit Pseudo>: FMem<op, (outs RC:$ra), (ins MemOpnd:$addr), !strconcat(instr_asm, "\t$ra, $addr"), [(set RC:$ra, (OpNode addr:$addr))], IILoad> { let isPseudo = Pseudo; } class StoreM<bits<8> op, string instr_asm, PatFrag OpNode, RegisterClass RC, Operand MemOpnd, bit Pseudo>: FMem<op, (outs), (ins RC:$ra, MemOpnd:$addr), !strconcat(instr_asm, "\t$ra, $addr"), [(OpNode RC:$ra, addr:$addr)], IIStore> { let isPseudo = Pseudo; } // 32-bit load. multiclass LoadM32<bits<8> op, string instr_asm, PatFrag OpNode, bit Pseudo = 0> {

3.2. LLVM Structure

23

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

def #NAME# : LoadM<op, instr_asm, OpNode, CPURegs, mem, Pseudo>; } // 32-bit store. multiclass StoreM32<bits<8> op, string instr_asm, PatFrag OpNode, bit Pseudo = 0> { def #NAME# : StoreM<op, instr_asm, OpNode, CPURegs, mem, Pseudo>; } //===----------------------------------------------------------------------===// // Instruction definition //===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===// // Cpu0I Instructions //===----------------------------------------------------------------------===// /// Load and Store Instructions /// aligned defm LD : LoadM32<0x00, "ld", load_a>; defm ST : StoreM32<0x01, "st", store_a>; /// Arithmetic Instructions (ALU Immediate) //def LDI : MoveImm<0x08, "ldi", add, simm16, immSExt16, CPURegs>; // add defined in include/llvm/Target/TargetSelectionDAG.td, line 315 (def add). def ADDiu : ArithLogicI<0x09, "addiu", add, simm16, immSExt16, CPURegs>; let isReturn=1, isTerminator=1, hasDelaySlot=1, isCodeGenOnly=1, isBarrier=1, hasCtrlDep=1 in def RET : FJ <0x2C, (outs), (ins CPURegs:$target), "ret\t$target", [(Cpu0Ret CPURegs:$target)], IIBranch>; //===----------------------------------------------------------------------===// // Arbitrary patterns that map to one or more instructions //===----------------------------------------------------------------------===// // Small immediates def : Pat<(i32 immSExt16:$in), (ADDiu ZERO, imm:$in)>;

Here we take ADDiu as an example. ADDiu is an instance of class ArithLogicI which is a subclass of class FL.
include "Cpu0InstrFormats.td" // Arithmetic and logical instructions with 2 register operands. class ArithLogicI<bits<8> op, string instr_asm, SDNode OpNode, Operand Od, PatLeaf imm_type, RegisterClass RC> : FL<op, (outs RC:$ra), (ins RC:$rb, Od:$imm16), !strconcat(instr_asm, "\t$ra, $rb, $imm16"), [(set RC:$ra, (OpNode RC:$rb, imm_type:$imm16))], IIAlu> { let isReMaterializable = 1; } def ADDiu def ADDiu : ArithLogicI<0x09, "addiu", add, simm16, immSExt16, CPURegs>; : ArithLogicI<0x09, "addiu", add, simm16, immSExt16, CPURegs>;

/// Arithmetic and logical instructions with 2 register operands. class ArithLogicI<bits<8> op, string instr_asm, SDNode OpNode, Operand Od, PatLeaf imm_type, RegisterClass RC> : FL<op, (outs RC:$ra), (ins RC:$rb, Od:$imm16), !strconcat(instr_asm, "\t$ra, $rb, $imm16"),

24

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

[(set RC:$ra, (OpNode RC:$rb, imm_type:$imm16))], IIAlu> { let isReMaterializable = 1; } So, op = 0x09 instr_asm = addiu OpNode = add Od = simm16 imm_type = immSExt16 RC = CPURegs

Expand with FL further,

: FL<op, (outs RC:$ra), (ins RC:$rb, Od:$imm16), !strconcat(instr_asm, "\t$ra, $rb, $imm16"), [(set RC:$ra, (OpNode RC:$rb, imm_type:$imm16))], IIAlu>

class FL<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: Cpu0Inst<outs, ins, asmstr, pattern, itin, FrmL> { bits<4> ra; bits<4> rb; bits<16> imm16; let Opcode = op; let Inst{23-20} = ra; let Inst{19-16} = rb; let Inst{15-0} = imm16; } So, op = 0x09 outs = CPURegs:$ra ins = CPURegs:$rb,simm16:$imm16 asmstr = "addiu\t$ra, $rb, $imm16" pattern = [(set CPURegs:$ra, (add RC:$rb, immSExt16:$imm16))] itin = IIAlu Members are, ra = CPURegs:$ra rb = CPURegs:$rb imm16 = simm16:$imm16 Opcode = 0x09; Inst{23-20} = CPURegs:$ra; Inst{19-16} = CPURegs:$rb; Inst{15-0} = simm16:$imm16;

Expand with Cpu0Inst further,

class FL<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: Cpu0Inst<outs, ins, asmstr, pattern, itin, FrmL> class Cpu0Inst<dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin, Format f>: Instruction { field bits<32> Inst; Format Form = f;

3.2. LLVM Structure

25

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

let Namespace = "Cpu0"; let Size = 4; bits<8> Opcode = 0; // Top 8 bits are the opcode field let Inst{31-24} = Opcode; let OutOperandList = outs; let InOperandList = ins; let AsmString let Pattern let Itinerary = asmstr; = pattern; = itin;

// // Attributes specific to Cpu0 instructions... // bits<4> FormBits = Form.Value; // TSFlags layout should be kept in sync with Cpu0InstrInfo.h. let TSFlags{3-0} = FormBits; let DecoderNamespace = "Cpu0"; field bits<32> SoftFail = 0; } So, outs = CPURegs:$ra ins = CPURegs:$rb,simm16:$imm16 asmstr = "addiu\t$ra, $rb, $imm16" pattern = [(set CPURegs:$ra, (add RC:$rb, immSExt16:$imm16))] itin = IIAlu f = FrmL Members are, Inst{31-24} = 0x09; OutOperandList = CPURegs:$ra InOperandList = CPURegs:$rb,simm16:$imm16 AsmString = "addiu\t$ra, $rb, $imm16" Pattern = [(set CPURegs:$ra, (add RC:$rb, immSExt16:$imm16))] Itinerary = IIAlu Summary with all members are, // Inherited from parent like Instruction Namespace = "Cpu0"; DecoderNamespace = "Cpu0"; Inst{31-24} = 0x08; Inst{23-20} = CPURegs:$ra; Inst{19-16} = CPURegs:$rb; Inst{15-0} = simm16:$imm16; OutOperandList = CPURegs:$ra InOperandList = CPURegs:$rb,simm16:$imm16 AsmString = "addiu\t$ra, $rb, $imm16" Pattern = [(set CPURegs:$ra, (add RC:$rb, immSExt16:$imm16))] Itinerary = IIAlu

26

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// From Cpu0Inst Opcode = 0x09; // From FL ra = CPURegs:$ra rb = CPURegs:$rb imm16 = simm16:$imm16

Its a lousy process. Similarly, LD and ST instruction denition can be expanded in this way. Please notify the Pattern = [(set CPURegs:$ra, (add RC:$rb, immSExt16:$imm16))] which include keyword add. We will use it in DAG transformations later.

3.2.5 Writing conguration les

You must have to add one LLVMBuild.txt under Target/Cpu0 directory. If you want to support cmake, then CMakeLists.txt needed to be added.
# CMakeLists.txt # Our td all in Cpu0.td, Cpu0RegisterInfo.td and Cpu0InstrInfo.td included in Cpu0.td set(LLVM_TARGET_DEFINITIONS Cpu0.td)

# Generate Cpu0GenRegisterInfo.inc and Cpu0GenInstrInfo.inc which included by your hand code C++ file # Cpu0GenRegisterInfo.inc came from Cpu0RegisterInfo.td, Cpu0GenInstrInfo.inc came from Cpu0InstrInfo tablegen(LLVM Cpu0GenRegisterInfo.inc -gen-register-info) tablegen(LLVM Cpu0GenInstrInfo.inc -gen-instr-info) tablegen(LLVM Cpu0GenSubtargetInfo.inc -gen-subtarget) # Cpu0CommonTableGen must be defined add_public_tablegen_target(Cpu0CommonTableGen) # Cpu0CodeGen should match with LLVMBuild.txt Cpu0CodeGen add_llvm_target(Cpu0CodeGen Cpu0TargetMachine.cpp ) # Should match with "subdirectories = add_subdirectory(TargetInfo) add_subdirectory(MCTargetDesc) MCTargetDesc TargetInfo" in LLVMBuild.txt

;===- ./lib/Target/Cpu0/LLVMBuild.txt --------------------------*- Conf -*--===; ; ; The LLVM Compiler Infrastructure ; ; This file is distributed under the University of Illinois Open Source ; License. See LICENSE.TXT for details. ; ;===------------------------------------------------------------------------===; ; ; This is an LLVMBuild description file for the components in this subdirectory. ; ; For more information on the LLVMBuild system, please see: ; ; http://llvm.org/docs/LLVMBuild.html ; ;===------------------------------------------------------------------------===; # Following comments extracted from http://llvm.org/docs/LLVMBuild.html

3.2. LLVM Structure

27

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

[common] subdirectories =

MCTargetDesc TargetInfo

[component_0] # TargetGroup components are an extension of LibraryGroups, specifically for defining LLVM targets (w type = TargetGroup # The name of the component should always be the name of the target. (should match "def Cpu0 : Target name = Cpu0 # Cpu0 component is located in directory Target/ parent = Target # Whether this target defines an assembly parser, assembly printer, disassembler, and supports JIT co #has_asmparser = 1 #has_asmprinter = 1 #has_disassembler = 1 #has_jit = 1

[component_1] # component_1 is a Library type and name is Cpu0CodeGen. After build it will in lib/libLLVMCpu0CodeGe type = Library name = Cpu0CodeGen # Cpu0CodeGen component(Library) is located in directory Cpu0/ parent = Cpu0 # If given, a list of the names of Library or LibraryGroup components which must also be linked in wh # dependencies for this component. When tools are built, the build system will include the transitive # the tool needs. required_libraries = CodeGen Core MC Cpu0Desc Cpu0Info SelectionDAG Support Target # All LLVMBuild.txt in Target/Cpu0 and subdirectory use add_to_library_groups = Cpu0 add_to_library_groups = Cpu0

LLVMBuild.txt les are written in a simple variant of the INI or conguration le format. Comments are prexed by # in both les. We explain the setting for these 2 les in comments. Please spend a little time to read it. Both CMakeLists.txt and LLVMBuild.txt coexist in sub-directories MCTargetDesc and TargetInfo. Their contents indicate they will generate Cpu0Desc and Cpu0Info libraries. After building, you will nd three libraries: libLLVMCpu0CodeGen.a, libLLVMCpu0Desc.a and libLLVMCpu0Info.a in lib/ of your build directory. For more details please see Building LLVM with CMake 5 and LLVMBuild Guide 6 .

3.3 Target Registration

You must also register your target with the TargetRegistry, which is what other LLVM tools use to be able to lookup and use your target at runtime. The TargetRegistry can be used directly, but for most targets there are helper templates which should take care of the work for you. All targets should declare a global Target object which is used to represent the target during registration. Then, in the targets TargetInfo library, the target should dene that object and use the RegisterTarget template to register the target. For example, the le TargetInfo/Cpu0TargetInfo.cpp register TheCpu0Target for big endian and TheCpu0elTarget for little endian, as follows.
// TargetInfo/Cpu0TargetInfo.cpp ... Target llvm::TheCpu0Target, llvm::TheCpu0elTarget; extern "C" void LLVMInitializeCpu0TargetInfo() { RegisterTarget<Triple::cpu0, /*HasJIT=*/true> X(TheCpu0Target, "cpu0", "Cpu0");
5 6

http://llvm.org/docs/CMake.html http://llvm.org/docs/LLVMBuild.html

28

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

RegisterTarget<Triple::cpu0el, /*HasJIT=*/true> Y(TheCpu0elTarget, "cpu0el", "Cpu0el"); }

Files Cpu0TargetMachine.cpp and MCTargetDesc/Cpu0MCTargetDesc.cpp just dene the empty initialize function since we register nothing in them for this moment.
//===-- Cpu0TargetMachine.cpp - Define TargetMachine for Cpu0 -------------===// ... extern "C" void LLVMInitializeCpu0Target() { } ... //===-- Cpu0MCTargetDesc.cpp - Cpu0 Target Descriptions -------------------===// ... extern "C" void LLVMInitializeCpu0TargetMC() { }

Please see Target Registration 7 for reference.

3.4 Build libraries and td

The llvm3.1 source code is put in /usr/local/llvm/3.1/src and have llvm3.1 release-build in /usr/local/llvm/3.1/congure_release_build. About how to build llvm, please refer 8 . We made a copy from /usr/local/llvm/3.1/src to /usr/local/llvm/3.1.test/cpu0/1/src for working with my Cpu0 target back end. Subdirectories src is for source code and cmake_debug_build is for debug build directory. Except directory src/lib/Target/Cpu0, there are a couple of les modied to support cpu0 new Target. Please check les in src_les_modify/src/. You can search cpu0 without case sensitive to nd the modied les by command,
[Gamma@localhost cmake_debug_build]$ grep -R -i "cpu0" ../src/ ../src/CMakeLists.txt: Cpu0 ../src/lib/Target/LLVMBuild.txt:subdirectories = ARM CellSPU CppBackend Hexagon MBlaze MSP430 Mips Cpu0 PTX PowerPC Sparc X86 XCore ../src/lib/MC/MCExpr.cpp: case VK_Cpu0_GPREL: return "GPREL"; ... ../src/lib/MC/MCELFStreamer.cpp: case MCSymbolRefExpr::VK_Cpu0_TLSGD: ... ../src/lib/MC/MCDwarf.cpp: // AT_language, a 4 byte value. We use DW_LANG_Cpu0 _Assembler as the dwarf2 ../src/lib/MC/MCDwarf.cpp: // MCOS->EmitIntValue(dwarf::DW_LANG_Cpu0_Assembler, 2); ../src/lib/Support/Triple.cpp: case cpu0: return "cpu0"; ... ../src/include/llvm/Support/ELF.h: EM_LATTICEMICO32 = 138, // RISC processor fo r Lattice CPU0 architecture ...

You can update your llvm working copy by,

cp -rf LLVMBackendTutorial/src_files_modified/src/* edir/.
7 8

yourllvm/workingcopy/sourc

http://llvm.org/docs/WritingAnLLVMBackend.html#target-registration http://clang.llvm.org/get_started.html

3.4. Build libraries and td

29

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Now, run the cmake and make command to build td (the following cmake command is for my setting),
[Gamma@localhost cmake_debug_build]$ cmake -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_ C_COMPILER=clang -DCMAKE_BUILD_TYPE=Debug -G "Unix Makefiles" ../src/ -- Targeting Cpu0 ... -- Targeting XCore -- Configuring done -- Generating done -- Build files have been written to: /usr/local/llvm/3.1.test/cpu0/1/cmake_debug _build [Gamma@localhost cmake_debug_build]$ make ... [100%] Built target gtest_main

After build, you can type command llc -version to nd the cpu0 backend,
[Gamma@localhost cmake_debug_build]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug _build/bin/llc --version LLVM (http://llvm.org/): LLVM version 3.1svn DEBUG build with assertions. Built Sep 21 2012 (18:27:58). Default target: x86_64-unknown-linux-gnu Host CPU: penryn Registered arm cellspu cpp cpu0 cpu0el ... Targets: - ARM - STI CBEA Cell SPU [experimental] - C++ backend - Cpu0 - Cpu0el

The llc -version can display cpu0 and cpu0el message, because the following code from le TargetInfo/Cpu0TargetInfo.cpp what in section Target Registration 9 we made. List them as follows again,
// Cpu0TargetInfo.cpp Target llvm::TheCpu0Target, llvm::TheCpu0elTarget; extern "C" void LLVMInitializeCpu0TargetInfo() { RegisterTarget<Triple::cpu0, /*HasJIT=*/true> X(TheCpu0Target, "cpu0", "Cpu0"); RegisterTarget<Triple::cpu0el, /*HasJIT=*/true> Y(TheCpu0elTarget, "cpu0el", "Cpu0el"); }

Now try to do llc command to compile input le ch3.cpp as follows,

// ch3.cpp int main() { return 0; }

First step, compile it with clang and get output ch3.bc as follows,
9

http://jonathan2251.github.com/lbd/llvmstructure.html#target-registration

30

Chapter 3. Introduce Cpu0 and LLVM

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

[Gamma@localhost InputFiles]$ clang -c ch3.cpp -emit-llvm -o ch3.bc

Next step, transfer bitcode .bc to human readable text format as follows,
[Gamma@localhost InputFiles]$ llvm-dis ch3.bc -o ch3.ll // ch3.ll ; ModuleID = ch3.bc target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f3 2:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:6 4-S128" target triple = "x86_64-unknown-linux-gnu" define i32 @main() nounwind uwtable { %1 = alloca i32, align 4 store i32 0, i32* %1 ret i32 0 }

Now, compile ch3.bc into ch3.cpu0.s, we get the error message as follows,
[Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cpu0.s llc: /usr/local/llvm/3.1.test/cpu0/1/src/tools/llc/llc.cpp:456: int main(int, ch ar **): Assertion target.get() && "Could not allocate target machine!" failed. Stack dump: 0. Program arguments: /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cpu0.s Aborted (core dumped)

Currently we just dene target td les (Cpu0.td, Cpu0RegisterInfo.td, ...). According to LLVM structure, we need to dene our target machine and include those td related les. The error message say we didnt dene our target machine.

3.4. Build libraries and td

31

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

32

Chapter 3. Introduce Cpu0 and LLVM

CHAPTER

FOUR

LLVM BACKEND STRUCTURE

This chapter introduces the LLVM backend class inheritance tree and class members rst. Next, we follow the LLVM backend structure and add class implementation one by one in each section. In this chapter, we need some compiler knowledge, like DAG and instruction selection, we will introduce them when needed. At the end of this chapter, we will have a backend being able to compile LLVM IR into cpu0 assembly code. The code added in this chapter are common in every LLVM backend, the only difference are their name. In fact, since Cpu0 is similar to Mips we copy most of the code from the Mips backend. In this chapter, we will focus on the relationship between classes involved in writing a LLVM backend. Once knowing the overall structure, you can quickly create a simple backend from scratch.

4.1 Overview
Note: Following content came from Design and Implementation of a TriCore Backend for the LLVM Compiler Framework 1 Chapter 4. I already got authors approve on the Chinese translation work. :-) As mentioned in Chapter 2, LLVM provides developers with an extensive generic frame work for code generation. It prescribes a xed directory layout and class hierarchy template. Each backend is located in its own subdirectory within lib/Target, where most of the code to be implemented goes. Apart from this, only a handful of the original les in the LLVM source tree have to be modied in order to integrate the new backend with the existing LLVM codebase.

LLVM LLVM lib/Target LLVM LLVM

The target code generator is divided into several components, each of which is discussed in a separate section. For a better overview, each section contains a list of the source les that comprise the respective component

The given class hierarchy species a number of abstract base classes with virtual functions and requires the backend developer to implement subclasses for each of them. Most of these classes do not provide any immediate functionality for code generation, but merely give selected information about the characteristics and properties of the target machine. This makes it possible to keep the bulk of the actual algorithms and procedures fully target-independent by accessing all the required target-specic information through the specied interfaces. LLVM

http://www.opus.ub.uni-erlangen.de/opus/volltexte/2010/1659/pdf/tricore_llvm.pdf

33

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

4.2 Code Generation Process

LLVM LLVM LLVM IR

2

4.2.1 Instruction Selection

LLVM IR SelectionDAG LLVM denition-use using node dening node control ow dependency LLVM DAG DAG instruction selector DAG LLVM DAG DAG instruction selector DAG LLVM
4.2.2 Scheduling and Formation

DAG (list) MachineFunction MachineBlock MachineBlock MachineInstr Scheduler : register pressure SSA ( ABI ) LLVM
4.2.3 SSA-based Machine Code Optimization

SSA SSA
4.2.4 Register Allocation

LLVM spill code : %e2 %d3 %d2 SSA SSA phi copy
4.2.5 Prologue/Epilogue Code Insertion

prologue epilogue (sp) (fp)

2

http://llvm.org/docs/CodeGenerator.html

34

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

4.2.6 Late Machine Code Optimizations

(peephole optimization)
4.2.7 Code Emission

; JIT
4.3 General Target Information

LLVM LLVMTargetMachine LLVM

4.3.1 Target Machine Characteristics

XXTargetMachine LLVM LLVM XXTargetMachine DataLayout : TargetFrameLowering : TargetInstrInfo TargetLowering TargetRegisterInfo

4.3.2 Subtarget Information

4.4 Build Example Code

We build our example code rst, then add components step by step. 1. Prepare the source code.
$ $ $ $ cd cpu0 mkdir -p 3/src; cd 3/src cp -rf ../../2/src/* . cp -rf $Example_SRC/3/1/Cpu0/ lib/Target/

2. Build
$ mkdir debug; cd debug $ cmake -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang \ -DLLVM_TARGETS_TO_BUILD="Mips;Cpu0" \ -DCMAKE_BUILD_TYPE=Debug -G "Unix Makefiles" ../src/

4.4.1 Step 1. Add XXTargetMachine

The role of TargetMachine in the LLVM backend structure is shown in gure 4.1. llc is LLVM static compiler. When you invoke llc command with option -march=XX, it will trigger the corresponding callback functions.

4.3. General Target Information

35

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 4.1: The role of TargetMachine Target Machine is a base class for targets implemented with the LLVM target-independent code generator. You should dene a XXTargetMachine class which inherit class LLVMTargetMachine, where XX is your target name. In our tutorial we name it as Cpu0TargetMachine. Class Cpu0TargetMachine has data members which dene target layout, instruction information, frame/stack ...etc. The Cpu0TargetMachine contents as follows,
//===-- Cpu0TargetMachine.h - Define TargetMachine for Cpu0 -----*- C++ -*-===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This file declares the Cpu0 specific subclass of TargetMachine. // //===----------------------------------------------------------------------===// #ifndef CPU0TARGETMACHINE_H #define CPU0TARGETMACHINE_H #include #include #include #include #include #include #include #include "Cpu0FrameLowering.h" "Cpu0InstrInfo.h" "Cpu0ISelLowering.h" "Cpu0SelectionDAGInfo.h" "Cpu0Subtarget.h" "llvm/Target/TargetMachine.h" "llvm/Target/TargetData.h" "llvm/Target/TargetFrameLowering.h"

namespace llvm { class formatted_raw_ostream; class Cpu0TargetMachine : public LLVMTargetMachine { Cpu0Subtarget Subtarget; const TargetData DataLayout; // Calculates type size & alignment Cpu0InstrInfo InstrInfo; //- Instructions Cpu0FrameLowering FrameLowering; //- Stack(Frame) and Stack direction

36

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Cpu0TargetLowering TLInfo; Cpu0SelectionDAGInfo TSInfo;

//- Stack(Frame) and Stack direction //- Map .bc DAG to backend DAG

public: Cpu0TargetMachine(const Target &T, StringRef TT, StringRef CPU, StringRef FS, const TargetOptions &Options, Reloc::Model RM, CodeModel::Model CM, CodeGenOpt::Level OL, bool isLittle); virtual const Cpu0InstrInfo const *getInstrInfo() { return &InstrInfo; } virtual const TargetFrameLowering *getFrameLowering() { return &FrameLowering; } virtual const Cpu0Subtarget *getSubtargetImpl() const { return &Subtarget; } virtual const TargetData const *getTargetData() { return &DataLayout;}

const

virtual const Cpu0TargetLowering *getTargetLowering() const { return &TLInfo; } virtual const Cpu0SelectionDAGInfo* getSelectionDAGInfo() const { return &TSInfo; } // Pass Pipeline Configuration virtual TargetPassConfig *createPassConfig(PassManagerBase &PM); }; /// Cpu0ebTargetMachine - Cpu032 big endian target machine. /// class Cpu0ebTargetMachine : public Cpu0TargetMachine { virtual void anchor(); public: Cpu0ebTargetMachine(const Target &T, StringRef TT, StringRef CPU, StringRef FS, const TargetOptions &Options, Reloc::Model RM, CodeModel::Model CM, CodeGenOpt::Level OL); }; /// Cpu0elTargetMachine - Cpu032 little endian target machine. /// class Cpu0elTargetMachine : public Cpu0TargetMachine { virtual void anchor(); public: Cpu0elTargetMachine(const Target &T, StringRef TT, StringRef CPU, StringRef FS, const TargetOptions &Options, Reloc::Model RM, CodeModel::Model CM, CodeGenOpt::Level OL); }; } // End llvm namespace #endif

Due to the OO design of LLVM, most member functions are already implemented by base classes. We only need to override some member functions for our target machine.

4.4. Build Example Code

37

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

After writing up our target description les, llvm-tblgen generates C++ les included by others. Take Cpu0InstrInfo.td as an example, llvm-tblgen generates Cpu0GenInstrInfo.inc from Cpu0InstrInfo.td. Then Cpu0InstrInfo.h denes GET_INSTRINFO_HEADER marco so that it can get code it needs from Cpu0GenInstrInfo.inc Below is the code fragment from Cpu0GenInstrInfo.inc. Code between #if def GET_INSTRINFO_HEADER and #endif // GET_INSTRINFO_HEADER will be extracted by Cpu0InstrInfo.h.
//- Cpu0GenInstInfo.inc which generate from Cpu0InstrInfo.td #ifdef GET_INSTRINFO_HEADER #undef GET_INSTRINFO_HEADER namespace llvm { struct Cpu0GenInstrInfo : public TargetInstrInfoImpl { explicit Cpu0GenInstrInfo(int SO = -1, int DO = -1); }; } // End llvm namespace #endif // GET_INSTRINFO_HEADER //- Cpu0InstrInfo.h extracts above code by defining GET_INSTRINFO_HEADER #define GET_INSTRINFO_HEADER #include "Cpu0GenInstrInfo.inc" namespace llvm { class Cpu0InstrInfo : public Cpu0GenInstrInfo { Cpu0TargetMachine &TM; public: explicit Cpu0InstrInfo(Cpu0TargetMachine &TM); }; }

Reference Write An LLVM Backend web site 3 .

$ cp -rf $Example_SRC/3/1/Cpu0 lib/Target

We add the following les in this section, and we have to modify CMakele.txt to include thhose new added C++ les. 1. Cpu0TargetMachine.[h,cpp] 2. Cpu0Subtarget.[h,cpp] 3. Cpu0InstrInfo.[h,cpp] 4. Cpu0FrameLowering.[h,cpp] 5. Cpu0TargetLowering.[h,cpp] 6. Cpu0SelectionDAGInfo.[h,cpp]
# CMakeLists.txt ... add_llvm_target(Cpu0CodeGen Cpu0ISelLowering.cpp Cpu0InstrInfo.cpp Cpu0FrameLowering.cpp Cpu0Subtarget.cpp Cpu0TargetMachine.cpp Cpu0SelectionDAGInfo.cpp )
3

http://llvm.org/docs/WritingAnLLVMBackend.html#target-machine

38

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Please follow commands below to build example code for this section. 1. Prepare the source code.
$ $ $ $ cd cpu0 mkdir -p 3/1/src; cd 3/1/src cp -rf ../../../2/src/* . cp -rf $Example_SRC/3/1/Cpu0/ lib/Target/

2. Build
$ mkdir debug; cd debug $ cmake -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang \ -DLLVM_TARGETS_TO_BUILD="Mips;Cpu0" \ -DCMAKE_BUILD_TYPE=Debug -G "Unix Makefiles" ../src/

4.4.2 Step 2. Add RegisterInfo

class Cpu0InstrInfo has a data member of class Cpu0RegisterInfo. This section adds class Cpu0RegisterInfo, then includes it in Cpu0InstrInfo.h and Cpu0TargetMachine.h. Finally, we modify CMakele.txt to add our new added le Cpu0RegisterInfo.cpp.
//- Cpu0InstrInfo.h class Cpu0InstrInfo : public Cpu0GenInstrInfo { Cpu0TargetMachine &TM; const Cpu0RegisterInfo RI; public: explicit Cpu0InstrInfo(Cpu0TargetMachine &TM); virtual const Cpu0RegisterInfo &getRegisterInfo() const; public: }; //- Cpu0InstrInfo.cpp Cpu0InstrInfo::Cpu0InstrInfo(Cpu0TargetMachine &tm) : TM(tm), RI(*TM.getSubtargetImpl(), *this) {} const Cpu0RegisterInfo &Cpu0InstrInfo::getRegisterInfo() const { return RI; } // Cpu0TargetMachine.h virtual const Cpu0RegisterInfo *getRegisterInfo() return &InstrInfo.getRegisterInfo(); }

const {

# CMakeLists.txt ... add_llvm_target(Cpu0CodeGen ... Cpu0RegisterInfo.cpp ... )

Now, lets replace 3/1/Cpu0 with 3/2/Cpu0 of adding register class denition and rebuild.

4.4. Build Example Code

39

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

1. Prepare the source code.

$ $ $ $ cd cpu0/3 mkdir -p 2/src; cd 2/src cp -rf ../../1/src/* . cp -rf $Example_SRC/3/2/Cpu0/ lib/Target/

2. Build
$ mkdir debug; cd debug $ cmake -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_C_COMPILER=clang \ -DLLVM_TARGETS_TO_BUILD="Mips;Cpu0" \ -DCMAKE_BUILD_TYPE=Debug -G "Unix Makefiles" ../src/

Now, lets run llc to compile a LLVM bitcode to see what happen. 1. Prepare LLVM bitcode.
$ clang -c example/input/ch3.cpp -emit-llvm -o ch3.bc

2. Compile
$ llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cpu0.s

llc: /cpu0/3/2/src/lib/CodeGen/LLVMTargetMachine.cpp:78: llvm::LLVMTargetMachine::LLVMTargetMachine(c Stack dump: 0. Program arguments: ./bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cp Aborted

Look into the assertion in above error message, it tell us that we dont have a ASMPrinter for our target to print assembly. We will add it in next section.

4.4.3 Step 3. Add AsmPrinter

First, wee modify Cpu0.td to add support for ASMPrinter. MCAsmWriter, we explictly say we will support MCAsmWriter. LLVM now has a newer ASMPrinter called

class Proc<string Name, list<SubtargetFeature> Features> : Processor<Name, Cpu0GenericItineraries, Features>; def : Proc<"cpu032", [FeatureCpu032]>; def Cpu0AsmWriter : AsmWriter { string AsmWriterClassName = "InstPrinter"; bit isMCAsmWriter = 1; } def Cpu0 : Target { let InstructionSet = Cpu0InstrInfo; let AssemblyWriters = [Cpu0AsmWriter]; }

llvm-tblgen will generate Cpu0GenAsmWrite.inc included by Cpu0InstPrinter.cpp which will be added soon. Cpu0GenAsmWrite.inc implements two member functions: printInstruction and getRegisterName for class Cpu0InstPrinter. Both of them are auto-generated from target description les we added before, i.e., Cpu0InstrInfo.td and Cpu0RegisterInfo.td. The last thing we have to do is adding class Cpu0InstPrinter, then include Cpu0GenAsmWrite.inc in it.

40

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

//- Cpu0InstPrinter.cpp #include "Cpu0GenAsmWriter.inc" void Cpu0InstPrinter::printRegName(raw_ostream &OS, unsigned RegNo) const { //- getRegisterName(RegNo) defined in Cpu0GenAsmWriter.inc which came from //- Cpu0.td indicate. OS << $ << StringRef(getRegisterName(RegNo)).lower(); } void Cpu0InstPrinter::printInst(const MCInst *MI, raw_ostream &O, StringRef Annot) { //- printInstruction(MI, O) defined in Cpu0GenAsmWriter.inc which came from //- Cpu0.td indicate. printInstruction(MI, O); printAnnotation(O, Annot); }

Next, add les below into new subdirectory MCTargetDesc. 1. Cpu0AsmPrinter.[h,cpp] 2. Cpu0MCInstLower.[h,cpp] 3. Cpu0BaseInfo.[h,cpp] 4. Cpu0FixupKinds.[h,cpp] 5. Cpu0MCAsmInfo.[h,cpp] We register our target MCAsmInfo and InstPrinter in Cpu0MCTargetDesc.cpp.
//- Cpu0MCTargetDesc.cpp static MCAsmInfo *createCpu0MCAsmInfo(const Target &T, StringRef TT) { MCAsmInfo *MAI = new Cpu0MCAsmInfo(T, TT); MachineLocation Dst(MachineLocation::VirtualFP); MachineLocation Src(Cpu0::SP, 0); MAI->addInitialFrameState(0, Dst, Src); return MAI; } static MCInstPrinter *createCpu0MCInstPrinter(const Target &T, unsigned SyntaxVariant, const MCAsmInfo &MAI, const MCInstrInfo &MII, const MCRegisterInfo &MRI, const MCSubtargetInfo &STI) { return new Cpu0InstPrinter(MAI, MII, MRI); } extern "C" void LLVMInitializeCpu0TargetMC() { // Register the MC asm info. RegisterMCAsmInfoFn X(TheCpu0Target, createCpu0MCAsmInfo); RegisterMCAsmInfoFn Y(TheCpu0elTarget, createCpu0MCAsmInfo); // Register the MCInstPrinter. TargetRegistry::RegisterMCInstPrinter(TheCpu0Target, createCpu0MCInstPrinter); TargetRegistry::RegisterMCInstPrinter(TheCpu0elTarget,

4.4. Build Example Code

41

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

createCpu0MCInstPrinter); }

Then we register AsmPrinter in Cpu0AsmPrinter.cpp.

//- Cpu0AsmPrinter.cpp // Force static initialization. extern "C" void LLVMInitializeCpu0AsmPrinter() { RegisterAsmPrinter<Cpu0AsmPrinter> X(TheCpu0Target); RegisterAsmPrinter<Cpu0AsmPrinter> Y(TheCpu0elTarget); }

This sections adds two subdirectories: InstPrinter and MCTargetDesc. You have to modify top-level LLVMBuild.txt and CMakele.txt accordingly, and add LLVMBuild.txt and CMakele.txt into those subdirectories also.
//top-level LLVMBuild.txt

[common] subdirectories = InstPrinter MCTargetDesc TargetInfo [component_0] ... # Please enable asmprinter has_asmprinter = 1 ... [component_1] # Add AsmPrinter Cpu0AsmPrinter required_libraries = AsmPrinter CodeGen Core MC Cpu0AsmPrinter Cpu0Desc Cpu0Info

O.K., lets build example code, then run llc as before. You will see error below:
$ ./bin/llc -march=cpu0 ../../2/debug/ch3.bc ./bin/llc: target does not support generation of this file type!

llc fails to compile LLVM IR into target machine code since we dont implement class Cpu0DAGToDAGISel yet. Before moving on, we will introduce LLVM code generation sequence, DAG, and instruction selection in next three sections.

4.5 LLVM Code Generation Sequence

Following content came from Design and Implementation of a TriCore Backend for the LLVM Compiler Framework 6 section 4.2 Code Generation Process. gure 4.2 gives an overview of LLVM code generation sequence. LLVM IR is a Static Single Assignment (SSA) 4 based representation. LLVM provides innite virtual registers which can hold value of primitive type (integral, oating point, or pointer). So, every operand can save in different virtual register in LLVM SSA representation. Comment is prexed with ; in LLVM representation. Following is a LLVM IR example.
; store i32 type of 0 to virtual register %a, %a is ; pointer type which point to i32 value store i32 %b, i32* %c ; store %b contents to %c point to, %b isi32 type virtual
4

store i32 0, i32* %a

http://en.wikipedia.org/wiki/Static_single_assignment_form

42

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 4.2: LLVM code generation sequence

%a1 = load i32* %a %a3 = add i32 %a2, 1

; register, %c is pointer type which point to i32 value. ; load the memory value where %a point to and assign the ; memory value to %a1 ; add %a2 and 1 and save to %a3

We explain the code generation process as below. If you dont feel comfortable, please check tricore_llvm.pdf section 4.2 rst. You can read The LLVM Target-Independent Code Generator from 5 and LLVM Language Reference Manual from 6 before go ahead, but we think read section 4.2 of tricore_llvm.pdf is enough. We suggest you read the web site documents as above only when you are still not quite understand, even though you have read this section and next 2 sections article for DAG and Instruction Selection. Design and Implementation of a TriCore Backend for the LLVM Compiler Framework 6 section 4.2. 1. Instruction Selection
// In this stage, transfer the llvm opcode into machine opcode, but the operand // still is llvm virtual operand. store i16 0, i16* %a // store 0 of i16 type to where virtual register %a // point to => addiu i16 0, i32* %a

2. Scheduling and Formation

// // st st %d In this stage, reorder the instructions sequence for optimization in instructions cycle or in register pressure. i32 %a, i16* %b, i16 5 ; st %a to *(%b+5) %b, i32* %c, i16 0 = ld i32* %c

// Transfer above instructions order as follows. In RISC architecture like

5 6

http://llvm.org/docs/CodeGenerator.html http://llvm.org/docs/LangRef.html

4.5. LLVM Code Generation Sequence

43

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

=>

// // st st %d

Mips, since the "ld %c" depends on previous instruction "st %c", it must wait more than 1 cycles. Therefore, the ld cannot follow st immediately. %b, i32* %c, i16 0 i32 %a, i16* %b, i16 5 = ld i32* %c, i16 0

=>

// If without reordering instructions, a instruction nop which does nothing // must be inserted as a bubble, contribute one instruction cycle more than // optimization. // Actually, Mips is scheduled with hardware dynamically and will insert nop // between st and ld instructions if compiler didnt insert nop. st i32 %a, i16* %b, i16 5 st %b, i32* %c, i16 0 nop %d = ld i32* %c, i16 0 // // // // // %a %b st st // // %a st %b st Minimum register pressure Suppose %c is alive after this basic block (meaning %c will be used after this basic block), %a and %b are not dead after this one. The following non-reordering version need 3 registers at least. = add i32 1, i32 0 = add i32 2, i32 0 %a, i32* %c, 1 %b, i32* %c, 2 The reordering version need 2 registers only (by allocate %a and %b in the same register) = add i32 1, i32 0 %a, i32* %c, 1 = add i32 2, i32 0 %b, i32* %c, 2

=>

3. SSA-based Machine Code Optimization For example, common expression remove, shown in next section DAG. 4. Register Allocation Allocate real register for virtual register. 5. Prologue/Epilogue Code Insertion Explain in section Add Prologue/Epilogue functions 6. Late Machine Code Optimizations Any last-minute peephole optimizations of the nal machine code can be applied during this phase. For example, replace x = x * 2 by x = x < 1 for integer operand. 7. Code Emission Finally, the completed machine code is emitted. For static compilation, the end result is an assembly code le; for JIT compilation, the opcodes of the machine instructions are written into memory.

4.5.1 DAG (Directed Acyclic Graph)

Many important techniques for local optimization begin by transforming a basic block into DAG. For example, the basic block code and its corresponding DAG as DAG example.

44

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 4.3: DAG example If b is not live on exit from the block, then we can do common expression remove to get the following code.
a = b + c d = a - d c = d + c

As you can imagine, the common expression remove can apply in IR or machine code. DAG like a tree which opcode is the node and operand (register and const/immediate/offset) is leaf. It can also be represented by list as prex order in tree. For example, (+ b, c), (+ b, 1) is IR DAG representation.

4.5.2 Instruction Selection

In the compiler backend, we need to translate compiler intermediate representation , i.e., IR code into machine code at instruction selection stage 7 . gure 4.4 shows a simple example to illustrate the before and after instruction selection.

Figure 4.4: IR and its corresponding machine instruction For machine instruction selection, the better solution is represent IR and machine instruction by DAG. In gure 4.5, we skip the register leaf. The rj + rk is IR DAG representation (for symbol notation, not llvm SSA form). ADD is machine instruction. The IR DAG and machine instruction DAG can also represented as list. For example, (+ ri, rj), (- ri, 1) are lists for IR DAG; (ADD ri, rj), (SUBI ri, 1) are lists for machine instruction DAG.
7

http://en.wikipedia.org/wiki/Instruction_selection

4.5. LLVM Code Generation Sequence

45

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 4.5: Instruction DAG representation Now, lets recall the ADDiu instruction dened on Cpu0InstrInfo.td in the previous chapter. And It will expand to the following Pattern as mentioned in section Write td (Target Description) of the previous chapter as follows,
def ADDiu : ArithLogicI<0x09, "addiu", add, simm16, immSExt16, CPURegs>;

Pattern = [(set CPURegs:$ra, (add RC:$rb, immSExt16:$imm16))]

This pattern meaning the IR DAG node add can translate into machine instruction DAG node ADDiu by pattern match mechanism. Similarly, the machine instruction DAG node LD and ST can be got from IR DAG node load and store. Some cpu/fpu (oating point processor) has multiply-and-add oating point instruction, fmadd. It can be represented by DAG list (fadd (fmul ra, rc), rb). For this implementation, we can assign fmadd DAG pattern to instruction td as follows,
def FMADDS : AForm_1<59, 29, (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB), "fmadds $FRT, $FRA, $FRC, $FRB", [(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC), F4RC:$FRB))]>;

Similar with ADDiu, [(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC), F4RC:$FRB))] is the pattern which include node fmul and node fadd. Now, for the following basic block notation IR and llvm SSA IR code,
d = a * c e = d + b ...

46

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

%d = fmul %a, %c %e = fadd %d, %b ...

The llvm SelectionDAG Optimization Phase (is part of Instruction Selection Process) prefered to translate this 2 IR DAG node (fmul %a, %b) (fadd %d, %c) into one machine instruction DAG node (fmadd %a, %c, %b), than translate them into 2 machine instruction nodes fmul and fadd.
%e = fmadd %a, %c, %b ...

As you can see, the IR notation representation is easier to read then llvm SSA IR form. So, we use the notation form in this book sometimes. For the following basic block code,
a = b + c d = a - d %e = fmadd %a, %c, %b // in notation IR form // in llvm SSA IR form

We can apply backendstructure_f7 Instruction tree pattern to get the following machine code,
load load add load sub fmadd rb, M(sp+8); // assume b allocate in sp+8, sp is stack point register rc, M(sp+16); ra, rb, rc; rd, M(sp+24); rd, ra, rd; re, ra, rc, rb;

4.5.3 Step 4. Add Cpu0DAGToDAGISel

The IR DAG to machine instruction DAG transformation is introduced in the previous section. Now, lets check what IR DAG node the le ch3.bc has. List ch3.ll as follows,
// ch3.ll define i32 @main() nounwind uwtable { %1 = alloca i32, align 4 store i32 0, i32* %1 ret i32 0 }

As above, ch3.ll use the IR DAG node store, ret. Actually, it also use add for sp (stack point) register adjust. So, the denitions in Cpu0InstInfo.td as follows is enough. IR DAG is dened in le include/llvm/Target/TargetSelectionDAG.td.
/// Load and Store Instructions /// aligned defm LD : LoadM32<0x00, "ld", defm ST : StoreM32<0x01, "st",

load_a>; store_a>;

/// Arithmetic Instructions (ALU Immediate) //def LDI : MoveImm<0x08, "ldi", add, simm16, immSExt16, CPURegs>; // add defined in include/llvm/Target/TargetSelectionDAG.td, line 315 (def add). def ADDiu : ArithLogicI<0x09, "addiu", add, simm16, immSExt16, CPURegs>; let isReturn=1, isTerminator=1, hasDelaySlot=1, isCodeGenOnly=1, isBarrier=1, hasCtrlDep=1 in

4.5. LLVM Code Generation Sequence

47

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

def RET : FJ <0x2C, (outs), (ins CPURegs:$target), "ret\t$target", [(Cpu0Ret CPURegs:$target)], IIBranch>;

Add class Cpu0DAGToDAGISel (Cpu0ISelDAGToDAG.cpp) to CMakeLists.txt, and add following fragment to Cpu0TargetMachine.cpp,
// Cpu0TargetMachine.cpp ... // Install an instruction selector pass using // the ISelDag to gen Cpu0 code. bool Cpu0PassConfig::addInstSelector() { PM->add(createCpu0ISelDag(getCpu0TargetMachine())); return false; } // Cpu0ISelDAGToDAG.cpp /// createCpu0ISelDag - This pass converts a legalized DAG into a /// CPU0-specific DAG, ready for instruction scheduling. FunctionPass *llvm::createCpu0ISelDag(Cpu0TargetMachine &TM) { return new Cpu0DAGToDAGISel(TM); }

This version adding the following code in Cpu0InstInfo.cpp to enable debug information which called by llvm at proper time.
// Cpu0InstInfo.cpp ... MachineInstr* Cpu0InstrInfo::emitFrameIndexDebugValue(MachineFunction &MF, int FrameIx, uint64_t Offset, const MDNode *MDPtr, DebugLoc DL) const { MachineInstrBuilder MIB = BuildMI(MF, DL, get(Cpu0::DBG_VALUE)) .addFrameIndex(FrameIx).addImm(0).addImm(Offset).addMetadata(MDPtr); return &*MIB; }

Build 3/4, run it, we nd the error message in 3/3 is gone. The new error message for 3/4 as follows,
[Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cpu0.s Target didnt implement TargetInstrInfo::storeRegToStackSlot! UNREACHABLE executed at /usr/local/llvm/3.1.test/cpu0/1/src/include/llvm/Target/ TargetInstrInfo.h:390! Stack dump: 0. Program arguments: /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cpu0.s 1. Running pass Function Pass Manager on module ch3.bc. 2. Running pass Prologue/Epilogue Insertion & Frame Finalization on function @main Aborted (core dumped)

4.5.4 Step 5. Add Prologue/Epilogue

Following came from Design and Implementation of a TriCore Backend for the LLVM Compiler Framework section 4.4.2.
5

For some target architectures, some aspects of the target architectures register set are dependent upon variable factors and have to be determined at runtime. As a consequence, they cannot be generated statically from a TableGen descrip48 Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

tion although that would be possible for the bulk of them in the case of the TriCore backend. Among them are the following points: Callee-saved registers. Normally, the ABI species a set of registers that a function must save on entry and restore on return if their contents are possibly modied during execution. Reserved registers. Although the set of unavailable registers is already dened in the TableGen le, TriCoreRegisterInfo contains a method that marks all non-allocatable register numbers in a bit vector. The following methods are implemented: emitPrologue() inserts prologue code at the beginning of a function. Thanks to TriCores context model, this is a trivial task as it is not required to save any registers manually. The only thing that has to be done is reserving space for the functions stack frame by decrementing the stack pointer. In addition, if the function needs a frame pointer, the frame register %a14 is set to the old value of the stack pointer beforehand. emitEpilogue() is intended to emit instructions to destroy the stack frame and restore all previously saved registers before returning from a function. However, as %a10 (stack pointer), %a11 (return address), and %a14 (frame pointer, if any) are all part of the upper context, no epilogue code is needed at all. All cleanup operations are performed implicitly by the ret instruction. eliminateFrameIndex() is called for each instruction that references a word of data in a stack slot. All previous passes of the code generator have been addressing stack slots through an abstract frame index and an immediate offset. The purpose of this function is to translate such a reference into a registeroffset pair. Depending on whether the machine function that contains the instruction has a xed or a variable stack frame, either the stack pointer %a10 or the frame pointer %a14 is used as the base register. The offset is computed accordingly. Figure 3.9 demonstrates for both cases how a stack slot is addressed. If the addressing mode of the affected instruction cannot handle the address because the offset is too large (the offset eld has 10 bits for the BO addressing mode and 16 bits for the BOL mode), a sequence of instructions is emitted that explicitly computes the effective address. Interim results are put into an unused address register. If none is available, an already occupied address register is scavenged. For this purpose, LLVMs framework offers a class named RegScavenger that takes care of all the details.

Figure 4.6: Addressing of a variable a located on the stack. If the stack frame has a variable size, slot must be addressed relative to the frame pointer We will explain the Prologue and Epilogue further by example code. So for the following llvm IR code, Cpu0 back end will emit the corresponding machine instructions as follows,

4.5. LLVM Code Generation Sequence

49

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

define i32 @main() nounwind uwtable { %1 = alloca i32, align 4 store i32 0, i32* %1 ret i32 0 } .section .mdebug.abi32 .previous .file "ch3.bc" .text .globl main .align 2 .type main,@function .ent main main: .frame .mask .set .set # BB#0: addiu addiu st addiu ret .set .set .end $tmp1: .size main, ($tmp1)-main $sp, $sp, -8 $2, $zero, 0 $2, 4($sp) $sp, $sp, 8 $lr macro reorder main $sp,8,$lr 0x00000000,0 noreorder nomacro # %entry

# @main

LLVM get the stack size by parsing IR and counting how many virtual registers is assigned to local variables. After that, it call emitPrologue(). This function will emit machine instructions to adjust sp (stack pointer register) for local variables since we dont use fp (frame pointer register). For our example, it will emit the instructions,
addiu $sp, $sp, -8

The emitEpilogue will emit addiu $sp, $sp, 8, 8 is the stack size. Since Instruction Selection and Register Allocation occurs before Prologue/Epilogue Code Insertion, eliminateFrameIndex() is called after machine instruction and real register allocated. It translate the frame index of local variable (%1 and %2 in the following example) into stack offset according the frame index order upward (stack grow up downward from high address to low address, 0($sp) is the top, 52($sp) is the bottom) as follows,
define i32 @main() %1 = alloca %2 = alloca ... store i32 0, store i32 5, ... ret i32 0 => # BB#0: addiu $tmp1: addiu st nounwind uwtable { i32, align 4 i32, align 4 i32* %1 i32* %2, align 4

$sp, $sp, -56 $3, $zero, 0 $3, 52($sp) // %1 is the first frame index local variable, so allocate // in 52($sp)

50

Chapter 4. LLVM Backend Structure

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

addiu st ... ret

$2, $zero, 5 $2, 48($sp) // %2 is the second frame index local variable, so // allocate in 48($sp) $lr

After add these Prologue and Epilogue functions, and build with 3/5/Cpu0. Now we are ready to compile our example code ch3.bc into cpu0 assembly code. Following is the command and output le ch3.cpu0.s,
[Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch3.bc -o ch3.cpu0.s [Gamma@localhost InputFiles]$ cat ch3.cpu0.s .section .mdebug.abi32 .previous .file "ch3.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .frame $sp,8,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: # %entry addiu $sp, $sp, -8 addiu $2, $zero, 0 st $2, 4($sp) addiu $sp, $sp, 8 ret $lr .set macro .set reorder .end main $tmp1: .size main, ($tmp1)-main

4.6 Summary
We have nished a simple assembler for cpu0 which only support addiu, st and ret 3 instructions. We are satised with this result. But you may think After so many codes we program, and just get the 3 instructions. The point is we have created a frame work for cpu0 target machine (please look back the llvm back end structure class inherit tree early in this chapter). Until now, we have around 3050 lines of source code with comments which include les *.cpp, *.h, *.td, CMakeLists.txt and LLVMBuild.txt. It can be counted by command wc find dir -name *.cpp for les *.cpp, *.h, *.td, *.txt. LLVM front end tutorial have 700 lines of source code without comments totally. Dont feel down with this result. In reality, write a back end is warm up slowly but run fast. Clang has over 500,000 lines of source code with comments in clang/lib directory which include C++ and Obj C support. Mips back end has only 15,000 lines with comments. Even the complicate X86 CPU which CISC outside and RISC inside (micro instruction), has only 45,000 lines with comments. In next chapter, we will show you that add a new instruction support is as easy as 123.

4.6. Summary

51

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

52

Chapter 4. LLVM Backend Structure

CHAPTER

FIVE

ADDING ARITHMETIC AND LOCAL POINTER SUPPORT

This chapter add more cpu0 arithmetic instructions support rst. The logic operation not support and translation in section Operator not !. The section Display llvm IR nodes with Graphviz will show you the DAG optimization steps and their corresponding llc display options. These DAG optimization steps result can be displayed by the graphic tool of Graphviz which supply very useful information with graphic view. You will appreciate Graphviz support in debug, we think. In section Adjust cpu0 instructions, we adjust cpu0 instructions to support some data type for C language. The section Local variable pointer introduce you the local variable pointer translation. Finally, section Operator mod, % take care the C operator %.

5.1 Support arithmetic instructions

Run the 3/5/Cpu0 llc with input le ch4_1_1.bc will get the error as follows,
// ch4_1_1.cpp int main() { int a = 5; int b = 2; int c = 0; c = a + b; return c; } [Gamma@localhost 3]$ clang -c ch4_1_1.cpp -emit-llvm -o ch4_1_1.bc [Gamma@localhost 3]$ llvm-dis ch4_1_1.bc -o ch4_1_1.ll [Gamma@localhost 3]$ cat ch4_1.ll ; ModuleID = ch4_1_1.bc target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64: 64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128n8:16:32:64-S128" target triple = "x86_64-unknown-linux-gnu" define %1 = %a = %b = %c = i32 @main() alloca i32, alloca i32, alloca i32, alloca i32, nounwind uwtable { align 4 align 4 align 4 align 4

53

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

store i32 0, i32* %1 store i32 5, i32* %a, align 4 store i32 2, i32* %b, align 4 store i32 0, i32* %c, align 4 %2 = load i32* %a, align 4 %3 = load i32* %b, align 4 %4 = add nsw i32 %2, %3 store i32 %4, i32* %c, align 4 %5 = load i32* %c, align 4 ret i32 %5 } [Gamma@localhost 3]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/ llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_1_1.bc -o ch4_1_1. cpu0.s LLVM ERROR: Cannot select: 0x30da480: i32 = add 0x30da280, 0x30da380 [ORD=7] [ID=17] 0x30da280: i32,ch = load 0x30da180, 0x30d9b80, 0x30d9880<LD4[%a]> [ORD=5] [ID=15] 0x30d9b80: i32 = FrameIndex<1> [ORD=2] [ID=5] 0x30d9880: i32 = undef [ORD=1] [ID=3] 0x30da380: i32,ch = load 0x30da180, 0x30d9e80, 0x30d9880<LD4[%b]> [ORD=6] [ID=14] 0x30d9e80: i32 = FrameIndex<2> [ORD=3] [ID=7] 0x30d9880: i32 = undef [ORD=1] [ID=3]

This error says we have not instructions to translate IR DAG node add. The ADDiu instruction is dened for node add with operands of 1 register and 1 immediate. This node add is for 2 registers. So, appending the following code to Cpu0InstrInfo.td and Cpu0Schedule.td in 4/1/Cpu0,
// Cpu0InstrInfo.td /// Arithmetic Instructions (3-Operand, R-Type) def CMP : CmpInstr<0x10, "cmp", IIAlu, CPURegs, 1>; def ADD : ArithLogicR<0x13, "add", add, IIAlu, CPURegs, 1>; def SUB : ArithLogicR<0x14, "sub", sub, IIAlu, CPURegs, 1>; def MUL : ArithLogicR<0x15, "mul", mul, IIImul, CPURegs, 1>; def DIV : ArithLogicR<0x16, "div", sdiv, IIIdiv, CPURegs, 1>; : ArithLogicR<0x18, "and", and, IIAlu, CPURegs, 1>; def AND : ArithLogicR<0x19, "or", or, IIAlu, CPURegs, 1>; def OR def XOR : ArithLogicR<0x1A, "xor", xor, IIAlu, CPURegs, 1>; /// def def def def Shift Instructions ROL : ArithLogicR<0x1C, ROR : ArithLogicR<0x1D, SHL : ArithLogicR<0x1E, : ArithLogicR<0x1F, SHR

"rol", "ror", "shl", "shr",

rotl, IIAlu, CPURegs, 1>; rotr, IIAlu, CPURegs, 1>; shl, IIAlu, CPURegs, 1>; sra, IIAlu, CPURegs, 1>;

// Cpu0Schedule.td ... def ALU : FuncUnit; def IMULDIV : FuncUnit; //===------------------------------------------------------------------===// // Instruction Itinerary classes used for Cpu0 //===------------------------------------------------------------------===// ... def IIImul : InstrItinClass; : InstrItinClass; def IIIdiv

54

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

def IIPseudo

: InstrItinClass;

//===------------------------------------------------------------------===// // Cpu0 Generic instruction itineraries. //===------------------------------------------------------------------===// // http://llvm.org/docs/doxygen/html/structllvm_1_1InstrStage.html def Cpu0GenericItineraries : ProcessorItineraries<[ALU, IMULDIV], [], [ ... InstrItinData<IIImul , [InstrStage<17, [IMULDIV]>]>, InstrItinData<IIIdiv , [InstrStage<38, [IMULDIV]>]> ]>;

In RISC CPU like Mips, the multiply/divide function unit and add/sub/logic unit are designed from two different hardware circuits, and more, their data path is separate. We think the cpu0 is the same even though no explanation in its web site. So, these two function units can be executed at same time (instruction level parallelism). Reference 1 for instruction itineraries. Now, lets build 4/1/Cpu0 and run with input le ch4_1_2.cpp. This version can process +, -, *, /, &, |, ^, <<, and >> operators in C language. The corresponding llvm IR instructions are add, sub, mul, sdiv, and, or, xor, shl, ashr. IR instruction sdiv stand for signed div while udiv is for unsigned div. The ashr instruction (arithmetic shift right) returns the rst operand shifted to the right a specied number of bits with sign extension. In brief, we call ashr is shift with sign extension ll.
Example: <result> = ashr i32 4, 1 <result> = ashr i8 -2, 1 <result> = ashr i32 1, 32 ; yields {i32}:result = 2 ; yields {i8}:result = -1 ; undefined

The C operator >> for negative operand is dependent on implementation. Most compiler translate it into shift with sign extension ll, for example, Mips sra is the instruction. Following is the Micosoft web site explanation, Note: >>, Microsoft Specic The result of a right shift of a signed negative quantity is implementation dependent. Although Microsoft C++ propagates the most-signicant bit to ll vacated bit positions, there is no guarantee that other implementations will do likewise. In addition to ashr, the other instruction shift with zero lled lshr in llvm (Mips implement lshr with instruction srl) has the following meaning.
Example: <result> = lshr i8 -2, 1 ; yields {i8}:result = 0x7FFFFFFF

In llvm, IR node sra is dened for ashr IR instruction, node srl is dened for lshr instruction (I dont know why dont use ashr and lshr as the IR node name directly). We assume Cpu0 shr instruction is shift with zero lled, and dene it with IR DAG node srl. But at that way, Cpu0 will fail to compile x >> 1 in case of x is signed integer because clang and most compilers translate it into ashr, which meaning shift with sign extension ll. Similarly, Cpu0 div instruction, has the same problem. We assume Cpu0 div instruction is for sdiv which can take care both positive and negative integer, but it will fail for divide operation / on unsigned integer operand in C. If we consider the x >> 1 denition is x = x/2. In case of x is unsigned int, range x is 0 ~ 4G-1 (0 ~ 0xFFFFFFFF) in 32 bits register, implement shift >> 1 by shift with zero lled is correct and satisfy the denition x = x/2, but shift with sign extension ll is not correct for range 2G ~ 4G-1. In case of x is signed int, range x is -2G ~ 2G-1, implement x >> 1 by shift with sign extension ll is correct for the denition, but shift with zero lled is not correct for range x is -2G ~ -1. So, if x = x/2 is denition for x >> 1, in order to satisfy the denition in both unsigned and signed integer of x, we need those two instructions, shift with zero lled and shift with sign extension ll.
1

http://llvm.org/docs/doxygen/html/structllvm_1_1InstrStage.html

5.1. Support arithmetic instructions

55

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Again, consider the x << 1 denition is x = x*2. We apply the x << 1 with shift 1 bit to left and ll the least bit with 0. In case of x is unsigned int, x << 1 satisfy the denition in range 0 ~ 2G-1, and x is overow when x > 2G-1 (no need to care what the register value is because overow). In case of x is signed int, x << 1 is correct for -1G ~ 1G-1; and x is overow for -2G ~ -1G-1 or 1G ~ 2G-1. So, implementation by shift 1bit to left and ll the least bit with 0 satisfy the denition x = x*2 for x << 1, no matter operand x is signed or unsigned int. Micorsoft implementation references as 2 . The sub-section ashr Instruction and sub-section lshr Instruction of 3 . The 4/1 version just add 40 lines code in td les. With these 40 lines code, it process 9 operators more for C language and their corresponding llvm IR instructions. The arithmetic instructions are easy to implement by add the denition in td le only.

5.2 Operator not !

Files ch4_2.cpp and ch4_2.bc are the C source code for not boolean operator and its corresponding llvm IR. List them as follows,
// ch4_2.cpp int main() { int a = 5; int b = 0; b = !a; return b; } ; ModuleID = ch4_2.bc target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128" target triple = "i386-apple-macosx10.8.0" define i32 @main() nounwind ssp { entry: %retval = alloca i32, align 4 %a = alloca i32, align 4 %b = alloca i32, align 4 store i32 0, i32* %retval store i32 5, i32* %a, align 4 store i32 0, i32* %b, align 4 %0 = load i32* %a, align 4 // a = %0 %tobool = icmp ne i32 %0, 0 // ne: stand for not egual %lnot = xor i1 %tobool, true %conv = zext i1 %lnot to i32 store i32 %conv, i32* %b, align 4 %1 = load i32* %b, align 4 ret i32 %1 }

As above comment, b = !a, translate to (xor (icmp ne i32 %0, 0), true). The %0 is the virtual register of variable a and the result of (icmp ne i32 %0, 0) is 1 bit size. To prove the translation is correct. Lets assume %0 != 0 rst, then the
2 3

http://msdn.microsoft.com/en-us/library/336xbhcz%28v=vs.80%29.aspx http://llvm.org/docs/LangRef.html.

56

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

(icmp ne i32 %0, 0) = 1 (or true), and (xor 1, 1) = 0. When %0 = 0, (icmp ne i32 %0, 0) = 0 (or false), and (xor 0, 1) = 1. So, the translation is correct. Now, lets run ch4_2.bc with 4/2/Cpu0 with llc -debug option to get result as follows,
118-165-16-22:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -debug -relocation-model=pic -filetype=asm ch4_3.bc -o ch4_3.cpu0.s ... === main Initial selection DAG: BB#0 main:entry SelectionDAG has 20 nodes: ... 0x7fbfc282c510: <multiple use> 0x7fbfc282c510: <multiple use> 0x7fbfc282bc10: <multiple use> 0x7fbfc282c610: ch = setne [ORD=5] 0x7fbfc282c710: i1 = setcc 0x7fbfc282c510, 0x7fbfc282bc10, 0x7fbfc282c610 [ORD=5] 0x7fbfc282c810: i1 = Constant<-1> [ORD=6] 0x7fbfc282c910: i1 = xor 0x7fbfc282c710, 0x7fbfc282c810 [ORD=6] 0x7fbfc282ca10: i32 = zero_extend 0x7fbfc282c910 [ORD=7] ...

Replacing.3 0x7fbfc282c910: i1 = xor 0x7fbfc282c710, 0x7fbfc282c810 [ORD=6] With: 0x7fbfc282ec10: i1 = setcc 0x7fbfc282c510, 0x7fbfc282bc10, 0x7fbfc282e910 Optimized lowered selection DAG: BB#0 main:entry SelectionDAG has 17 nodes: ... 0x7fbfc282c510: <multiple use> 0x7fbfc282c510: <multiple use> 0x7fbfc282bc10: <multiple use> 0x7fbfc282e910: ch = seteq 0x7fbfc282ec10: i1 = setcc 0x7fbfc282c510, 0x7fbfc282bc10, 0x7fbfc282e910 0x7fbfc282ca10: i32 = zero_extend 0x7fbfc282ec10 [ORD=7] ... Type-legalized selection DAG: BB#0 main:entry SelectionDAG has 18 nodes: ... 0x7fbfc282c510: <multiple use> 0x7fbfc282c510: <multiple use> 0x7fbfc282bc10: <multiple use> 0x7fbfc282e910: ch = seteq [ID=-3] 0x7fbfc282c610: i32 = setcc 0x7fbfc282c510, 0x7fbfc282bc10, 0x7fbfc282e910 [ID=-3]

5.2. Operator not !

57

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

0x7fbfc282c710: i32 = Constant<1> [ID=-3] 0x7fbfc282c810: i32 = and 0x7fbfc282c610, 0x7fbfc282c710 [ID=-3] ...

The (setcc %1, %2, setne) and (xor %3, -1) in Initial selection DAG stage corresponding (icmp %1, %2, ne) and (xor %3, 1) in ch4_2.bc. The argument in xor is 1 bit size (1 and -1 are same, they are all represented by 1). The (zero_extend %4) of Initial selection DAG corresponding (zext i1 %lnot to i32) of ch4_2.bc. As above it translate 2 DAG nodes (setcc %1, %2, setne) and (xor %3, -1) into 1 DAG node (setcc %1, %2, seteq) in Optimized lowered selection DAG stage. This translation is right since for 1 bit size, (xor %3, 1) and (not %3) has same result, and (not (setcc %1, %2, setne)) is equal to (setcc %1, %2, seteq). In Optimized lowered selection DAG stage, it also translate (zero_extern i1 %lnot to 32) into (and %lnot, 1). (zero_extern i1 %lnot to 32) just expand the %lnot to i32 32 bits result, so translate into (and %lnot, 1) is correct. Finally, translate (setcc %1, %2, seteq) into (xor (xor %1, %2), (ldi $0, 1) in Instruction selection stage by the rule dened in Cpu0InstrInfo.td as follows,
// Cpu0InstrInfo.td ... def : Pat<(not CPURegs:$in), (XOR CPURegs:$in, (LDI ZERO, 1))>; // setcc patterns multiclass SeteqPats<RegisterClass RC, Instruction XOROp, Register ZEROReg> { def : Pat<(seteq RC:$lhs, RC:$rhs), (XOROp (XOROp RC:$lhs, RC:$rhs), (LDI ZERO, 1))>; } defm : SeteqPats<CPURegs, XOR, ZERO>;

After xor, the (and %4, 1) is translated into (and $2, (ldi $3, 1)) which is dened before already. List the asm le ch4_3.cpu0.s code fragment as below, you can check it with the nal result.
118-165-16-22:InputFiles Jonathan$ cat ch4_2.cpu0.s ... # BB#0: # %entry addiu $sp, $sp, -16 $tmp1: .cfi_def_cfa_offset 16 addiu $2, $zero, 0 st $2, 12($sp) addiu $3, $zero, 5 st $3, 8($sp) st $2, 4($sp) ld $3, 8($sp) xor $2, $3, $2 ldi $3, 1 xor $2, $2, $3 addiu $3, $zero, 1 and $2, $2, $3 st $2, 4($sp) addiu $sp, $sp, 16 ret $lr ...

58

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

5.3 Display llvm IR nodes with Graphviz

The previous section, display the DAG translation process in text on terminal by llc -debug option. The llc also support the graphic display. The section Install other tools on iMac mentioned the web for llc graphic display information. The llc graphic display with tool Graphviz is introduced in this section. The graphic display is more readable by eye than display text in terminal. Its not necessary, but it help a lot especially when you are tired in tracking the DAG translation process. List the llc graphic support options from the sub-section SelectionDAG Instruction Selection Process of web 4 as follows, Note: The llc Graphviz DAG display options -view-dag-combine1-dags displays the DAG after being built, before the rst optimization pass. -view-legalize-dags displays the DAG before Legalization. -view-dag-combine2-dags displays the DAG before the second optimization pass. -view-isel-dags displays the DAG before the Select phase. -view-sched-dags displays the DAG before Scheduling. By tracking llc -debug, you can see the DAG translation steps as follows,
Initial selection DAG Optimized lowered selection DAG Type-legalized selection DAG Optimized type-legalized selection DAG Legalized selection DAG Optimized legalized selection DAG Instruction selection Selected selection DAG Scheduling ...

Lets run llc with option -view-dag-combine1-dags, and open the output result with Graphviz as follows,
118-165-12-177:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -view-dag-combine1-dags -march=cpu0 -relocation-model=pic -filetype=asm ch4_2.bc -o ch4_2.cpu0.s Writing /tmp/llvm_84ibpm/dag.main.dot... done. 118-165-12-177:InputFiles Jonathan$ Graphviz /tmp/llvm_84ibpm/dag.main.dot

It will show the /tmp/llvm_84ibpm/dag.main.dot as llc option -view-dag-combine1-dags graphic view. From llc option -view-dag-combine1-dags graphic view, we can see the -view-dag-combine1-dags option is for Initial selection DAG. We list the other view options and their corresponding DAG translation stage as follows,
-view-dag-combine1-dags: Initial selection DAG -view-legalize-dags: Optimized type-legalized selection DAG -view-dag-combine2-dags: Legalized selection DAG -view-isel-dags: Optimized legalized selection DAG -view-sched-dags: Selected selection DAG

The -view-isel-dags is important and often used by an llvm backend writer because it is the DAG before instruction selection. The backend programmer need to know what is the DAG for writing the pattern match instruction in target description le .td.
4

http://llvm.org/docs/CodeGenerator.html

5.3. Display llvm IR nodes with Graphviz

59

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 5.1: llc option -view-dag-combine1-dags graphic view

60

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

5.4 Adjust cpu0 instructions

We decide add instructions udiv and sra to avoid compiler errors for C language operators / in unsigned int and >> in signed int as section Support arithmetic instructions mentioned. To support these 2 operators, we only need to add these code in Cpu0InstrInfo.td as follows,
// Cpu0InstsInfo.td ... def UDIV : ArithLogicR<0x17, "udiv", udiv, IIIdiv, CPURegs, 1>; ... /// Shift Instructions // work, its for ashr llvm IR instruction def SRA : shift_rotate_imm32<0x1B, 0x00, "sra", sra>;

To use addiu only instead of ldi, change Cpu0InstrInfo.td as follows,

// Cpu0InstsInfo.td ... //def LDI : MoveImm<0x08, "ldi", add, simm16, immSExt16, CPURegs>; ... // setcc patterns multiclass SeteqPats<RegisterClass RC, Instruction XOROp> { def : Pat<(seteq RC:$lhs, RC:$rhs), (XOROp (XOROp RC:$lhs, RC:$rhs), (ADDiu ZERO, 1))>; } defm : SeteqPats<CPURegs, XOR>;

Run ch4_4.cpp with code 4/4/Cpu0 which support udiv, sra, and use addiu only instead of ldi, will get the result as follows,
// ch4_4.cpp int main() { int a = 1; int b = 2; int k = 0; unsigned int a1 = -5, f1 = 0; f1 = a1 / b; k = (a >> 2); return k; } 118-165-13-40:InputFiles Jonathan$ clang -c ch4_4.cpp -emit-llvm -o ch4_4.bc 118-165-13-40:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_4.bc -o ch4_4.cpu0.s 118-165-13-40:InputFiles Jonathan$ cat ch4_4.cpu0.s ... addiu $sp, $sp, -24 addiu $2, $zero, 0 ... udiv $2, $3, $2 st $2, 0($sp) ld $2, 16($sp) sra $2, $2, 2

5.4. Adjust cpu0 instructions

61

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

...

5.4.1 Use cpu0 ofcial LDI instead of ADDiu

According cpu0 web site instruction denition. There is no addiu instruction denition. We add addiu instruction because we nd this instruction is more powerful and reasonable than ldi instruction. We highlight this change in section CPU0 processor architecture. Even with that, we show you how to replace our addiu with ldi according the cpu0 original design. 4/4_2/Cpu0 is the code changes for use ldi instruction. This changes replace addiu with ldi in Cpu0InstrInfo.td and modify Cpu0FrameLowering.cpp as follows,
// Cpu0InstrInfo.td ... /// Arithmetic Instructions (ALU Immediate) def LDI : MoveImm<0x08, "ldi", add, simm16, immSExt16, CPURegs>; // add defined in include/llvm/Target/TargetSelectionDAG.td, line 315 (def add). //def ADDiu : ArithLogicI<0x09, "addiu", add, simm16, immSExt16, CPURegs>; ... // Small immediates def : Pat<(i32 immSExt16:$in), (LDI ZERO, imm:$in)>; // hi/lo relocs def : Pat<(Cpu0Hi tglobaladdr:$in), (SHL (LDI ZERO, tglobaladdr:$in), 16)>; // Expect cpu0 add LUi support, like Mips //def : Pat<(Cpu0Hi tglobaladdr:$in), (LUi tglobaladdr:$in)>; def : Pat<(Cpu0Lo tglobaladdr:$in), (LDI ZERO, tglobaladdr:$in)>; def : Pat<(add CPURegs:$hi, (Cpu0Lo tglobaladdr:$lo)), (ADD CPURegs:$hi, (LDI ZERO, tglobaladdr:$lo))>; // gp_rel relocs def : Pat<(add CPURegs:$gp, (Cpu0GPRel tglobaladdr:$in)), (ADD CPURegs:$gp, (LDI ZERO, tglobaladdr:$in))>; def : Pat<(not CPURegs:$in), (XOR CPURegs:$in, (LDI ZERO, 1))>; // Cpu0FrameLowering.cpp ... void Cpu0FrameLowering::emitPrologue(MachineFunction &MF) const { ... // Adjust stack. if (isInt<16>(-StackSize)) { // ldi fp, (-stacksize) // add sp, sp, fp BuildMI(MBB, MBBI, dl, TII.get(Cpu0::LDI), Cpu0::FP).addReg(Cpu0::FP) .addImm(-StackSize); BuildMI(MBB, MBBI, dl, TII.get(Cpu0::ADD), SP).addReg(SP).addReg(Cpu0::FP); } ... } void Cpu0FrameLowering::emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const {

62

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

... // Adjust stack. if (isInt<16>(-StackSize)) { // ldi fp, (-stacksize) // add sp, sp, fp BuildMI(MBB, MBBI, dl, TII.get(Cpu0::LDI), Cpu0::FP).addReg(Cpu0::FP) .addImm(-StackSize); BuildMI(MBB, MBBI, dl, TII.get(Cpu0::ADD), SP).addReg(SP).addReg(Cpu0::FP); } ... }

As above code, we use add IR binary instruction (1 register operand and 1 immediate operand, and the register operand is xed with ZERO) in our solution since we didnt nd the move IR unary instruction. This code is correct since all the immediate value is translated into ldi Zero, imm/address. And (add CPURegs:$gp, $imm16) is translated into (ADD CPURegs:$gp, (LDI ZERO, $imm16)). Lets run 4/4_2/Cpu0 with ch4_4.cpp to get the correct result below. As you will see, addiu $sp, $sp, -24 will be replaced with the pair instructions of ldi $fp, -24 and add $sp, $sp, $fp. Since the $sp pointer adjustment is so frequently occurs (it occurs in every function entry and exit point), we reserve the $fp to the pair of stack adjustment instructions ldi and add. If we didnt reserve the dedicate registers $fp and $sp, it need to save and restore them in the stack adjustment. It meaning more instructions running cost in this. Anyway, the pair of ldi and add to adjust stack pointer is double in cost compete to addiu, thats the benet we mentioned in section CPU0 processor architecture.
118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_4.bc -o ch4_4.cpu0.s 118-165-66-82:InputFiles Jonathan$ cat ch4_4.cpu0.s .section .mdebug.abi32 .previous .file "ch4_4.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .cfi_startproc .frame $sp,24,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: ldi $fp, -24 add $sp, $sp, $fp $tmp1: .cfi_def_cfa_offset 24 ldi $2, 0 st $2, 20($sp) ldi $3, 1 st $3, 16($sp) ldi $3, 2 st $3, 12($sp) st $2, 8($sp) ldi $3, -5 st $3, 4($sp) st $2, 0($sp) ld $2, 12($sp) ld $3, 4($sp)

5.4. Adjust cpu0 instructions

63

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

udiv $2, $3, $2 st $2, 0($sp) ld $2, 16($sp) sra $2, $2, 2 st $2, 8($sp) ldi $fp, 24 add $sp, $sp, $fp ret $lr .set macro .set reorder .end main $tmp2: .size main, ($tmp2)-main .cfi_endproc

5.5 Local variable pointer

To support pointer to local variable, add this code fragment in Cpu0InstrInfo.td and Cpu0InstPrinter.cpp as follows,
// Cpu0InstrInfo.td ... def mem_ea : Operand<i32> { let PrintMethod = "printMemOperandEA"; let MIOperandInfo = (ops CPURegs, simm16); let EncoderMethod = "getMemEncoding"; } ... class EffectiveAddress<string instr_asm, RegisterClass RC, Operand Mem> : FMem<0x09, (outs RC:$ra), (ins Mem:$addr), instr_asm, [(set RC:$ra, addr:$addr)], IIAlu>; ... // FrameIndexes are legalized when they are operands from load/store // instructions. The same not happens for stack address copies, so an // add op with mem ComplexPattern is used and the stack address copy // can be matched. Its similar to Sparc LEA_ADDRi def LEA_ADDiu : EffectiveAddress<"addiu\t$ra, $addr", CPURegs, mem_ea> { let isCodeGenOnly = 1; } // Cpu0InstPrinter.cpp ... void Cpu0InstPrinter:: printMemOperandEA(const MCInst *MI, int opNum, raw_ostream &O) { // when using stack locations for not load/store instructions // print the same way as all normal 3 operand instructions. printOperand(MI, opNum, O); O << ", "; printOperand(MI, opNum+1, O); return; }

Run ch4_5.cpp with code 4/5/Cpu0 which support pointer to local variable, will get result as follows,
// ch4_5.cpp int main() {

64

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

int b = 3; int* p = &b; return *p; } 118-165-66-82:InputFiles Jonathan$ clang -c ch4_5.cpp -emit-llvm -o ch4_5.bc 118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_5.bc -o ch4_5.cpu0.s 118-165-66-82:InputFiles Jonathan$ cat ch4_5.cpu0.s .section .mdebug.abi32 .previous .file "ch4_5.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .cfi_startproc .frame $sp,16,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: addiu $sp, $sp, -16 $tmp1: .cfi_def_cfa_offset 16 addiu $2, $zero, 0 st $2, 12($sp) addiu $2, $zero, 3 st $2, 8($sp) addiu $2, $sp, 8 st $2, 0($sp) addiu $sp, $sp, 16 ret $lr .set macro .set reorder .end main $tmp2: .size main, ($tmp2)-main .cfi_endproc

5.6 Operator mod, %

5.6.1 The DAG of %
Example input code ch4_6.cpp which contains the C operator % and its corresponding llvm IR, as follows,
// ch4_6.cpp int main() { int b = 11;

5.6. Operator mod, %

65

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

//

unsigned int b = 11;

b = (b+1)%12; return b; } ; ModuleID = ch4_6.bc target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128" target triple = "i386-apple-macosx10.8.0" define i32 @main() nounwind ssp { entry: %retval = alloca i32, align 4 %b = alloca i32, align 4 store i32 0, i32* %retval store i32 11, i32* %b, align 4 %0 = load i32* %b, align 4 %add = add nsw i32 %0, 1 %rem = srem i32 %add, 12 store i32 %rem, i32* %b, align 4 %1 = load i32* %b, align 4 ret i32 %1 }

LLVM srem is the IR corresponding %, reference sub-section srem instruction of 3 . Copy the reference as follows, Note: srem Instruction Syntax: <result> = srem <ty> <op1>, <op2> ; yields {ty}:result Overview: The srem instruction returns the remainder from the signed division of its two operands. This instruction can also take vector versions of the values in which case the elements must be integers. Arguments: The two arguments to the srem instruction must be integer or vector of integer values. Both arguments must have identical types. Semantics: This instruction returns the remainder of a division (where the result is either zero or has the same sign as the dividend, op1), not the modulo operator (where the result is either zero or has the same sign as the divisor, op2) of a value. For more information about the difference, see The Math Forum. For a table of how this is implemented in various languages, please see Wikipedia: modulo operation. Note that signed integer remainder and unsigned integer remainder are distinct operations; for unsigned integer remainder, use urem. Taking the remainder of a division by zero leads to undened behavior. Overow also leads to undened behavior; this is a rare case, but can occur, for example, by taking the remainder of a 32-bit division of -2147483648 by -1. (The remainder doesnt actually overow, but this rule lets srem be implemented using instructions that return both the result of the division and the remainder.) Example: <result> = srem i32 4, %var ; yields {i32}:result = 4 % %var Run 4/5/Cpu0 with input le ch4_6.bc and llc option view-isel-dags as follows, will get the error message as follows and the llvm DAG of ch4_6.bc DAG.
118-165-79-37:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -view-isel-dags -relocation-model=

66

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

pic -filetype=asm ch4_6.bc -o ch4_6.cpu0.s ... LLVM ERROR: Cannot select: 0x7fa73a02ea10: i32 = mulhs 0x7fa73a02c610, 0x7fa73a02e910 [ID=12] 0x7fa73a02c610: i32 = Constant<12> [ORD=5] [ID=7] 0x7fa73a02e910: i32 = Constant<715827883> [ID=9]

Figure 5.2: ch4_6.bc DAG LLVM replace srem divide operation with multiply operation in DAG optimization because DIV operation cost more in time than MUL. For example code int b = 11; b=(b+1)%12;, it translate into ch4_6.bc DAG. We verify the result and explain by calculate the value in each node. The 0xC*0x2AAAAAAB=0x200000004, (mulhs 0xC, 0x2AAAAAAAB) meaning get the Signed mul high word (32bits). Multiply with 2 operands of 1 word size generate

5.6. Operator mod, %

67

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

the 2 word size of result (0x2, 0xAAAAAAAB). The high word result, in this case is 0x2. The nal result (sub 12, 12) is 0 which match the statement (11+1)%12.

5.6.2 Arm solution

Lets run 4/6_1/Cpu0 with ch4_6.cpp as well as llc option -view-sched-dags to get Translate ch4_6.bc into cpu0 backend DAG. Similarly, SMMUL get the high word of multiply result. Follows is the result of run 4/6_1/Cpu0 with ch4_6.bc.
118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_6.bc -o ch4_6.cpu0.s 118-165-71-252:InputFiles Jonathan$ cat ch4_6.cpu0.s .section .mdebug.abi32 .previous .file "ch4_6.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .frame $sp,8,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: # %entry addiu $sp, $sp, -8 $2, $zero, 0 addiu st $2, 4($sp) addiu $2, $zero, 11 st $2, 0($sp) addiu $2, $zero, 10922 shl $2, $2, 16 addiu $3, $zero, 43691 or $3, $2, $3 addiu $2, $zero, 12 smmul $3, $2, $3 shr $4, $3, 31 sra $3, $3, 1 add $3, $3, $4 mul $3, $3, $2 sub $2, $2, $3 st $2, 0($sp) addiu $sp, $sp, 8 ret $lr .set macro .set reorder .end main $tmp1: .size main, ($tmp1)-main

The other instruction UMMUL and llvm IR mulhu are unsigned int type for operator %. You can check it by unmark the unsigned int b = 11; in ch4_6.cpp. Use SMMUL instruction to get the high word of multiplication result is adopted in ARM. The 4/6_1/Cpu0 use the ARM solution. With this solution, the following code is needed.

68

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 5.3: Translate ch4_6.bc into cpu0 backend DAG

5.6. Operator mod, %

69

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Cpu0InstrInfo.td ... // Transformation Function - get the lower 16 bits. def LO16 : SDNodeXForm<imm, [{ return getImm(N, N->getZExtValue() & 0xFFFF); }]>; // Transformation Function - get the higher 16 bits. def HI16 : SDNodeXForm<imm, [{ return getImm(N, (N->getZExtValue() >> 16) & 0xFFFF); }]>; ... def SMMUL : ArithLogicR<0x50, "smmul", mulhs, IIImul, CPURegs, 1>; def UMMUL : ArithLogicR<0x51, "ummul", mulhu, IIImul, CPURegs, 1>; ... // Arbitrary immediates def : Pat<(i32 imm:$imm), (OR (SHL (ADDiu ZERO, (HI16 imm:$imm)), 16), (ADDiu ZERO, (LO16 imm:$imm)))>;

5.6.3 Mips solution

Mips use MULT instruction and save the high & low part to register HI and LO. After that, use mfhi/mo to move register HI/LO to your general purpose register. ARM SMMUL is fast if you only need the HI part of result (it ignore the LO part of operation). Meanwhile Mips is fast if you need both the HI and LO result. If you need the LO part of result, you can use Cpu0 MUL instruction which only get the LO part of result. 4/6_2/Cpu0 is implemented with Mips MULT style. We choose it as the implementation of this book. For Mips style implementation, we add the following code in Cpu0RegisterInfo.td, Cpu0InstrInfo.td and Cpu0ISelDAGToDAG.cpp. And list the related DAG nodes mulhs and mulhu which are used in 4/6_2/Cpu0 from TargetSelectionDAG.td.
// Cpu0RegisterInfo.td ... // Hi/Lo registers def HI : Register<"hi">, DwarfRegNum<[18]>; def LO : Register<"lo">, DwarfRegNum<[19]>; // Cpu0Schedule.td ... def IIHiLo : InstrItinClass; ... def Cpu0GenericItineraries : ProcessorItineraries<[ALU, IMULDIV], [], [ ... InstrItinData<IIHiLo , [InstrStage<1, [IMULDIV]>]>, ... ]>; // Cpu0InstrInfo.td ... // Mul, Div class Mult<bits<8> op, string instr_asm, InstrItinClass itin, RegisterClass RC, list<Register> DefRegs>: FL<op, (outs), (ins RC:$ra, RC:$rb), !strconcat(instr_asm, "\t$ra, $rb"), [], itin> { let imm16 = 0; let isCommutable = 1; let Defs = DefRegs; let neverHasSideEffects = 1;

70

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

} class Mult32<bits<8> op, string instr_asm, InstrItinClass itin>: Mult<op, instr_asm, itin, CPURegs, [HI, LO]>; // Move from Hi/Lo class MoveFromLOHI<bits<8> op, string instr_asm, RegisterClass RC, list<Register> UseRegs>: FL<op, (outs RC:$ra), (ins), !strconcat(instr_asm, "\t$ra"), [], IIHiLo> { let rb = 0; let imm16 = 0; let Uses = UseRegs; let neverHasSideEffects = 1; } ... def MULT : Mult32<0x50, "mult", IIImul>; def MULTu : Mult32<0x51, "multu", IIImul>; def MFHI : MoveFromLOHI<0x40, "mfhi", CPURegs, [HI]>; def MFLO : MoveFromLOHI<0x41, "mflo", CPURegs, [LO]>; // Cpu0ISelDAGToDAG.cpp ... /// Select multiply instructions. std::pair<SDNode*, SDNode*> Cpu0DAGToDAGISel::SelectMULT(SDNode *N, unsigned Opc, DebugLoc dl, EVT Ty, bool HasLo, bool HasHi) { SDNode *Lo = 0, *Hi = 0; SDNode *Mul = CurDAG->getMachineNode(Opc, dl, MVT::Glue, N->getOperand(0), N->getOperand(1)); SDValue InFlag = SDValue(Mul, 0); if (HasLo) { Lo = CurDAG->getMachineNode(Cpu0::MFLO, dl, Ty, MVT::Glue, InFlag); InFlag = SDValue(Lo, 1); } if (HasHi) Hi = CurDAG->getMachineNode(Cpu0::MFHI, dl, Ty, InFlag); return std::make_pair(Lo, Hi); } /// Select instructions not customized! Used for /// expanded, promoted and normal instructions SDNode* Cpu0DAGToDAGISel::Select(SDNode *Node) { unsigned Opcode = Node->getOpcode(); ... switch(Opcode) { default: break; case ISD::MULHS: case ISD::MULHU: { MultOpc = (Opcode == ISD::MULHU ? Cpu0::MULTu : Cpu0::MULT); return SelectMULT(Node, MultOpc, dl, NodeTy, false, true).second; }

5.6. Operator mod, %

71

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

... } // TargetSelectionDAG.td ... def mulhs : SDNode<"ISD::MULHS" def mulhu : SDNode<"ISD::MULHU"

, SDTIntBinOp, [SDNPCommutative]>; , SDTIntBinOp, [SDNPCommutative]>;

Except the custom type, llvm IR operations of expand and promote type will call Cpu0DAGToDAGISel::Select() during instruction selection of DAG translation. In Select(), it return the HI part of multiplication result to HI register, for IR operations of mulhs or mulhu. After that, MFHI instruction move the HI register to cpu0 eld a register, $ra. MFHI instruction is FL format and only use cpu0 eld a register, we set the $rb and imm16 to 0. DAG for ch4_6.bc with Mips style MULT and ch4_6.cpu0.s are the result of compile ch4_6.bc.
118-165-66-82:InputFiles Jonathan$ cat ch4_6.cpu0.s .section .mdebug.abi32 .previous .file "ch4_6.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .cfi_startproc .frame $sp,8,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: addiu $sp, $sp, -8 $tmp1: .cfi_def_cfa_offset 8 addiu $2, $zero, 0 st $2, 4($sp) addiu $2, $zero, 11 st $2, 0($sp) addiu $2, $zero, 10922 shl $2, $2, 16 addiu $3, $zero, 43691 or $3, $2, $3 addiu $2, $zero, 12 mult $2, $3 mfhi $3 shr $4, $3, 31 sra $3, $3, 1 add $3, $3, $4 mul $3, $3, $2 sub $2, $2, $3 st $2, 0($sp) addiu $sp, $sp, 8 ret $lr .set macro .set reorder .end main $tmp2: .size main, ($tmp2)-main .cfi_endproc

72

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 5.4: DAG for ch4_6.bc with Mips style MULT

5.6. Operator mod, %

73

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

5.7 Full support %

The sensitive readers may nd the llvm using multiplication instead of div to get the % result just because our example use constant as divider, (b+1)%12 in our example. If programmer use variable as the divider like (b+1)%a, then what will happens in our code. The answer is our code will have error to take care this. In section Support arithmetic instructions, we use div a, b to hold the quotient part in register. The multiplication operator * need 64 bits of register to hold the result for two 32 bits of operands multiplication. We modify cpu0 to use the pair of registers LO and HI which just like Mips to solve this issue in last section. Now, its time to modify cpu0 for integer divide operator again. We use LO and HI registers to hold the quotient and remainder and use instructions mo and mfhi to get the result from LO or HI registers. With this solution, the c = a / b can be got by div a, b and mo c; the c = a % b can be got by div a, b and mfhi c. 4/6_4/Cpu0 support operator % and /. The code added in 4/6_4/Cpu0 as follows,
// Cpu0InstrInfo.cpp ... void Cpu0InstrInfo:: copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, DebugLoc DL, unsigned DestReg, unsigned SrcReg, bool KillSrc) const { unsigned Opc = 0, ZeroReg = 0; if (Cpu0::CPURegsRegClass.contains(DestReg)) { // Copy to CPU Reg. if (Cpu0::CPURegsRegClass.contains(SrcReg)) Opc = Cpu0::ADD, ZeroReg = Cpu0::ZERO; else if (SrcReg == Cpu0::HI) Opc = Cpu0::MFHI, SrcReg = 0; else if (SrcReg == Cpu0::LO) Opc = Cpu0::MFLO, SrcReg = 0; } else if (Cpu0::CPURegsRegClass.contains(SrcReg)) { // Copy from CPU Reg. if (DestReg == Cpu0::HI) Opc = Cpu0::MTHI, DestReg = 0; else if (DestReg == Cpu0::LO) Opc = Cpu0::MTLO, DestReg = 0; } assert(Opc && "Cannot copy registers"); MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(Opc)); if (DestReg) MIB.addReg(DestReg, RegState::Define); if (ZeroReg) MIB.addReg(ZeroReg); if (SrcReg) MIB.addReg(SrcReg, getKillRegState(KillSrc)); } // Cpu0InstrInfo.h ... virtual void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI, DebugLoc DL, unsigned DestReg, unsigned SrcReg, bool KillSrc) const;

74

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Cpu0InstrInfo.td ... def SDT_Cpu0DivRem

: SDTypeProfile<0, 2, [SDTCisInt<0>, SDTCisSameAs<0, 1>]>;

... // DivRem(u) nodes def Cpu0DivRem : SDNode<"Cpu0ISD::DivRem", SDT_Cpu0DivRem, [SDNPOutGlue]>; def Cpu0DivRemU : SDNode<"Cpu0ISD::DivRemU", SDT_Cpu0DivRem, [SDNPOutGlue]>; ... class Div<SDNode opNode, bits<8> op, string instr_asm, InstrItinClass itin, RegisterClass RC, list<Register> DefRegs>: FL<op, (outs), (ins RC:$rb, RC:$rc), !strconcat(instr_asm, "\t$$zero, $rb, $rc"), [(opNode RC:$rb, RC:$rc)], itin> { let imm16 = 0; let Defs = DefRegs; } class Div32<SDNode opNode, bits<8> op, string instr_asm, InstrItinClass itin>: Div<opNode, op, instr_asm, itin, CPURegs, [HI, LO]>; ... class MoveToLOHI<bits<8> op, string instr_asm, RegisterClass RC, list<Register> DefRegs>: FL<op, (outs), (ins RC:$ra), !strconcat(instr_asm, "\t$ra"), [], IIHiLo> { let rb = 0; let imm16 = 0; let Defs = DefRegs; let neverHasSideEffects = 1; } ... def SDIV : Div32<Cpu0DivRem, 0x16, "div", IIIdiv>; def UDIV : Div32<Cpu0DivRemU, 0x17, "divu", IIIdiv>; ... def MTHI : MoveToLOHI<0x42, "mthi", CPURegs, [HI]>; def MTLO : MoveToLOHI<0x43, "mtlo", CPURegs, [LO]>; // Cpu0ISelLowering.cpp ... Cpu0TargetLowering:: Cpu0TargetLowering(Cpu0TargetMachine &TM) : TargetLowering(TM, new TargetLoweringObjectFileELF()), Subtarget(&TM.getSubtarget<Cpu0Subtarget>()) { ... setOperationAction(ISD::SDIV, MVT::i32, Expand); setOperationAction(ISD::SREM, MVT::i32, Expand); setOperationAction(ISD::UDIV, MVT::i32, Expand); setOperationAction(ISD::UREM, MVT::i32, Expand); setTargetDAGCombine(ISD::SDIVREM); setTargetDAGCombine(ISD::UDIVREM); ... } ... static SDValue PerformDivRemCombine(SDNode *N, SelectionDAG& DAG,

5.7. Full support %

75

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

TargetLowering::DAGCombinerInfo &DCI, const Cpu0Subtarget* Subtarget) { if (DCI.isBeforeLegalizeOps()) return SDValue(); EVT Ty = unsigned unsigned unsigned N->getValueType(0); LO = Cpu0::LO; HI = Cpu0::HI; opc = N->getOpcode() == ISD::SDIVREM ? Cpu0ISD::DivRem : Cpu0ISD::DivRemU; DebugLoc dl = N->getDebugLoc(); SDValue DivRem = DAG.getNode(opc, dl, MVT::Glue, N->getOperand(0), N->getOperand(1)); SDValue InChain = DAG.getEntryNode(); SDValue InGlue = DivRem; // insert MFLO if (N->hasAnyUseOfValue(0)) { SDValue CopyFromLo = DAG.getCopyFromReg(InChain, dl, LO, Ty, InGlue); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), CopyFromLo); InChain = CopyFromLo.getValue(1); InGlue = CopyFromLo.getValue(2); } // insert MFHI if (N->hasAnyUseOfValue(1)) { SDValue CopyFromHi = DAG.getCopyFromReg(InChain, dl, HI, Ty, InGlue); DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), CopyFromHi); } return SDValue(); } SDValue Cpu0TargetLowering::PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { SelectionDAG &DAG = DCI.DAG; unsigned opc = N->getOpcode(); switch (opc) { default: break; case ISD::SDIVREM: case ISD::UDIVREM: return PerformDivRemCombine(N, DAG, DCI, Subtarget); } return SDValue(); } // Cpu0ISelLowering.h ... namespace llvm { namespace Cpu0ISD { enum NodeType { // Start the numbering from where ISD NodeType finishes. FIRST_NUMBER = ISD::BUILTIN_OP_END,

76

Chapter 5. Adding arithmetic and local pointer support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Ret, // DivRem(u) DivRem, DivRemU }; } ... // Cpu0RegisterInfo.td ... // Hi/Lo Registers def HILO : RegisterClass<"Cpu0", [i32], 32, (add HI, LO)>;

Run with ch4_1_2.cpp can get the result for operator / as below. But run with ch4_6_1.cpp as below, cannot get the div for operator %. It still use multiplication instead of div because llvm do Constant Propagation Optimization in this. The ch4_6_2.cpp can get the div for % result since it make the llvm Constant Propagation Optimization useless in this. Unfortunately, we cannot run it now since it need the function call support. We will verify % with ch4_6_2.cpp at the end of chapter Function Call. You can run with the end of Example Code of chapter Function Call, if you like to verify it now.
// ch4_1_2.cpp int main() { ... f = a / b; ... } 118-165-77-79:InputFiles Jonathan$ clang -c ch4_1_2.cpp -emit-llvm -o ch4_1_2.bc 118-165-77-79:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_1_2.bc -o ch4_1_2.cpu0.s 118-165-77-79:InputFiles Jonathan$ cat ch4_1_2.cpu0.s div $zero, $3, $2 mflo $2 ... // ch4_6_1.cpp int main() { int b = 11; int a = 12; b = (b+1)%a; return b; } // ch4_6_2.cpp #include <stdlib.h> int main() { int b = 11; // unsigned int b = 11; int c = rand(); b = (b+1)%c;

5.7. Full support %

77

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

return b; }

5.8 Summary
We support most of C operators in this chapter. Until now, we have around 3400 lines of source code with comments. With these 345 lines of source code added, it support the number of operators from three to over ten.

78

Chapter 5. Adding arithmetic and local pointer support

CHAPTER

SIX

GENERATING OBJECT FILES

The previous chapters only introduce the assembly code generated. This chapter will introduce you the obj support rst, and display the obj by objdump utility. With LLVM support, the cpu0 backend can generate both big endian and little endian obj les with only a few code added. The Target Registration mechanism and their structure will be introduced in this chapter.

6.1 Translate into obj le

Currently, we only support translate llvm IR code into assembly code. If you try to run 4/6_2/Cpu0 to translate obj code will get the error message as follows,
[Gamma@localhost 3]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/ llc -march=cpu0 -relocation-model=pic -filetype=obj ch4_1_2.bc -o ch4_1_2.cpu0.o /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc: target does not support generation of this file type!

The 5/Cpu0 support obj le generated. It can get result for big endian and little endian with command llc -march=cpu0 and llc -march=cpu0el. Run it will get the obj les as follows,
[Gamma@localhost InputFiles]$ cat ch4_1_2.cpu0.s ... .set nomacro # BB#0: addiu $sp, $sp, -72 addiu $2, $zero, 0 st $2, 68($sp) addiu $3, $zero, 5 st $3, 64($sp) ... [Gamma@localhost 3]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/ llc -march=cpu0 -relocation-model=pic -filetype=obj ch4_2.bc -o ch4_2.cpu0.o [Gamma@localhost InputFiles]$ objdump -s ch4_2.cpu0.o ch4_2.cpu0.o: file format elf32-big

Contents of section .text: 0000 09d0ffb8 09200000 012d0044 0010 013d0040 09300002 013d003c 0020 012d0034 012d0014 0930fffb 0030 012d000c 012d0008 002d003c 0040 13232000 012d0038 002d003c

09300005 012d0038 013d0010 003d0040 003d0040

..... ...-.D.0.. [email protected]...=.<.-.8 .-.4.-...0...=.. .-...-...-.<.=.@ .# ..-.8.-.<.=.@

79

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

0050 14232000 012d0034 002d003c 003d0040 .# ..-.4.-.<.=.@ 0060 15232000 012d0030 002d003c 003d0040 .# ..-.0.-.<.=.@ 0070 16232000 012d002c 002d003c 003d0040 .# ..-.,.-.<.=.@ 0080 18232000 012d0028 002d003c 003d0040 .# ..-.(.-.<.=.@ 0090 19232000 012d0024 002d003c 003d0040 .# ..-.$.-.<.=.@ 00a0 1a232000 012d0020 002d0040 1e220002 .# ..-. .-.@.".. 00b0 012d001c 002d0010 1e220002 012d0004 .-...-..."...-.. 00c0 002d0010 1f220002 012d000c 09d00048 .-..."...-.....H 00d0 2c00000e ,... Contents of section .eh_frame: 0000 00000010 00000000 017a5200 017c0e01 .........zR..|.. 0010 000c0d00 00000010 00000018 00000000 ................ 0020 000000d4 00440e48 .....D.H [Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/llc -march=cpu0el -relocation-model=pic -filetype=obj ch4_2.bc -o ch4_2.cpu0el.o [Gamma@localhost InputFiles]$ objdump -s ch4_2.cpu0el.o ch4_2.cpu0el.o: file format elf32-little

Contents of section .text: 0000 b8ffd009 00002009 44002d01 0010 40003d01 02003009 3c003d01 0020 34002d01 14002d01 fbff3009 0030 0c002d01 08002d01 3c002d00 0040 00202313 38002d01 3c002d00 0050 00202314 34002d01 3c002d00 0060 00202315 30002d01 3c002d00 0070 00202316 2c002d01 3c002d00 0080 00202318 28002d01 3c002d00 0090 00202319 24002d01 3c002d00 00a0 0020231a 20002d01 40002d00 00b0 1c002d01 10002d00 0200221e 00c0 10002d00 0200221f 0c002d01 00d0 0e00002c Contents of section .eh_frame: 0000 10000000 00000000 017a5200 0010 000c0d00 10000000 18000000 0020 d4000000 00440e48

05003009 38002d01 10003d01 40003d00 40003d00 40003d00 40003d00 40003d00 40003d00 40003d00 0200221e 04002d01 4800d009

...... .D.-...0. @.=...0.<.=.8.-. 4.-...-...0...=. ..-...-.<.-.@.=. . #.8.-.<.-.@.=. . #.4.-.<.-.@.=. . #.0.-.<.-.@.=. . #.,.-.<.-.@.=. . #.(.-.<.-.@.=. . #.$.-.<.-.@.=. . #. [email protected]...". ..-...-..."...-. ..-..."...-.H... ...,

017c0e01 00000000

.........zR..|.. ................ .....D.H

The rst instruction is addiu $sp, -72 and its corresponding obj is 0x09d0ffb8. The addiu opcode is 0x09, 8 bits, $sp register number is 13(0xd), 4bits, second register is useless, so assign it to 0x0, and the immediate is 16 bits -72(=0xffb8), so its correct. The third instruction st $2, 68($sp) and its corresponding obj is 0x012d0044. The st opcode is 0x0a, $2 is 0x2, $sp is 0xd and immediate is 68(0x0044). Thanks to cpu0 instruction format which opcode, register operand and offset(imediate value) size are multiple of 4 bits. The obj format is easy to check by eye. The big endian (B0, B1, B2, B3) = (09, d0, ff, b8), objdump from B0 to B3 as 0x09d0ffb8 and the little endian is (B3, B2, B1, B0) = (09, d0, ff, b8), objdump from B0 to B3 as 0xb8ffd009.

6.2 Backend Target Registration Structure

Now, lets examine Cpu0MCTargetDesc.cpp.
// Cpu0MCTargetDesc.cpp ... extern "C" void LLVMInitializeCpu0TargetMC() { // Register the MC asm info.

80

Chapter 6. Generating object les

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

RegisterMCAsmInfoFn X(TheCpu0Target, createCpu0MCAsmInfo); RegisterMCAsmInfoFn Y(TheCpu0elTarget, createCpu0MCAsmInfo); // Register the MC codegen info. TargetRegistry::RegisterMCCodeGenInfo(TheCpu0Target, createCpu0MCCodeGenInfo); TargetRegistry::RegisterMCCodeGenInfo(TheCpu0elTarget, createCpu0MCCodeGenInfo); // Register the MC instruction info. TargetRegistry::RegisterMCInstrInfo(TheCpu0Target, createCpu0MCInstrInfo); TargetRegistry::RegisterMCInstrInfo(TheCpu0elTarget, createCpu0MCInstrInfo); // Register the MC register info. TargetRegistry::RegisterMCRegInfo(TheCpu0Target, createCpu0MCRegisterInfo); TargetRegistry::RegisterMCRegInfo(TheCpu0elTarget, createCpu0MCRegisterInfo); // Register the MC Code Emitter TargetRegistry::RegisterMCCodeEmitter(TheCpu0Target, createCpu0MCCodeEmitterEB); TargetRegistry::RegisterMCCodeEmitter(TheCpu0elTarget, createCpu0MCCodeEmitterEL); // Register the object streamer. TargetRegistry::RegisterMCObjectStreamer(TheCpu0Target, createMCStreamer); TargetRegistry::RegisterMCObjectStreamer(TheCpu0elTarget, createMCStreamer); // Register the asm backend. TargetRegistry::RegisterMCAsmBackend(TheCpu0Target, createCpu0AsmBackendEB32); TargetRegistry::RegisterMCAsmBackend(TheCpu0elTarget, createCpu0AsmBackendEL32); // Register the MC subtarget info. TargetRegistry::RegisterMCSubtargetInfo(TheCpu0Target, createCpu0MCSubtargetInfo); TargetRegistry::RegisterMCSubtargetInfo(TheCpu0elTarget, createCpu0MCSubtargetInfo); // Register the MCInstPrinter. TargetRegistry::RegisterMCInstPrinter(TheCpu0Target, createCpu0MCInstPrinter); TargetRegistry::RegisterMCInstPrinter(TheCpu0elTarget, createCpu0MCInstPrinter); }

Cpu0MCTargetDesc.cpp do the target registration as mentioned in section Target Registration 1 of the last chapter. Drawing the register function and those class it registered in Register Cpu0MCAsmInfo to Register Cpu0InstPrinter for explanation. In Register Cpu0MCAsmInfo, registering the object of class Cpu0AsmInfo for target TheCpu0Target and TheCpu0elTarget. TheCpu0Target is for big endian and TheCpu0elTarget is for little endian. Cpu0AsmInfo is derived from MCAsmInfo which is llvm built-in class. Most code is implemented in its parent, back end reuse those code by inherit. In Register MCCodeGenInfo, instancing MCCodeGenInfo, and initialize it by pass Roloc::PIC because we use command llc -relocation-model=pic to tell llc compile using position-independent code mode. Recall the addressing mode in system program book has two mode, one is PIC mode, the other is absolute addressing mode. MC stands for Machine Code. In Register MCInstrInfo, instancing MCInstrInfo object X, and initialize it by InitCpu0MCInstrInfo(X). Since InitCpu0MCInstrInfo(X) is dened in Cpu0GenInstrInfo.inc, it will add the information fromCpu0InstrInfo.td we
1

http://jonathan2251.github.com/lbd/llvmstructure.html#target-registration

6.2. Backend Target Registration Structure

81

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.1: Register Cpu0MCAsmInfo

82

Chapter 6. Generating object les

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.2: Register MCCodeGenInfo

Figure 6.3: Register MCInstrInfo

6.2. Backend Target Registration Structure

83

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.4: Register MCRegisterInfo

84

Chapter 6. Generating object les

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.5: Register Cpu0MCCodeEmitter

6.2. Backend Target Registration Structure

85

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.6: Register MCELFStreamer

86

Chapter 6. Generating object les

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.7: Register Cpu0AsmBackend

6.2. Backend Target Registration Structure

87

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.8: Register Cpu0MCSubtargetInfo

88

Chapter 6. Generating object les

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.9: Register Cpu0InstPrinter

6.2. Backend Target Registration Structure

89

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 6.10: MCELFStreamer inherit tree

90

Chapter 6. Generating object les

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

specied. Register MCRegisterInfo is similar to Register MCInstrInfo, but it initialize the register information specied in Cpu0RegisterInfo.td. They share a lot of code with instruction/register td description. Register Cpu0MCCodeEmitter, instancing two objects Cpu0MCCodeEmitter, one is for big endian and the other is for little endian. They take care the obj format generated. So, its not dened in 4/6_2/Cpu0 which support assembly code only. Register MCELFStreamer, MCELFStreamer take care the obj format also. Register Cpu0MCCodeEmitter Cpu0MCCodeEmitter take care code emitter while MCELFStreamer take care the obj output streamer. MCELFStreamer inherit tree is MCELFStreamer inherit tree. You can nd a lot of operations in that inherit tree. Reader maybe has the question for what are the actual arguments in createCpu0MCCodeEmitterEB(const MCInstrInfo &MCII, const MCSubtargetInfo &STI, MCContext &Ctx) and at when they are assigned. Yes, we didnt assign it, we register the createXXX() function by function pointer only (according C, TargetRegistry::RegisterXXX(TheCpu0Target, createXXX()) where createXXX is function pointer). LLVM keep a function pointer to createXXX() when we call target registry, and will call these createXXX() function back at proper time with arguments assigned during the target registration process, RegisterXXX(). Register Cpu0AsmBackend, Cpu0AsmBackend class is the bridge for asm to obj. Two objects take care big endian and little endian also. It derived from MCAsmBackend. Most of code for object le generated is implemented by MCELFStreamer and its parent, MCAsmBackend. Register Cpu0MCSubtargetInfo, instancing MCSubtargetInfo object and initialize with Cpu0.td information. Register Cpu0InstPrinter, instancing Cpu0InstPrinter to take care printing function for instructions. Like Register Cpu0MCAsmInfo to Register MCRegisterInfo, it has been dened in 4/6_2/Cpu0 code for assembly le generated support.

6.2. Backend Target Registration Structure

91

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

92

Chapter 6. Generating object les

CHAPTER

SEVEN

GLOBAL VARIABLES, STRUCTS AND ARRAYS

In the previous two chapters, we only access the local variables. This chapter will deal global variable access translation. After that, introducing the types of struct and array as well as their corresponding llvm IR statement, and how the cpu0 translate these llvm IR statements in section Array and struct support. The global variable DAG translation is different from the previous DAG translation we have now. It create DAG nodes at run time in our backend C++ code according the llc -relocation-model option while the others of DAG just do IR DAG to Machine DAG translation directly according the input le IR DAG.

7.1 Global variable

6/1/Cpu0 support the global variable, lets compile ch6_1.cpp with this version rst, and explain the code changes after that.
// // // // // // // // // // clang -c ch5_1.cpp -emit-llvm -o ch5_1.bc /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/llc -march=cpu0 -march=cpu0 -march=cpu0 -march=cpu0 -march=cpu0

-relocation-model=pic -filet -relocation-model=static -cp -relocation-model=static -fi -relocation-model=pic -filet -relocation-model=static -fi

/Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model /Applications/Xcode.app/Contents/Developer/usr/bin/lldb -- /Users/Jonathan/llvm/3.1.test/cpu0/1/cm

int gI = 100; int main() { int c = 0; c = gI; return c; } 118-165-66-82:InputFiles Jonathan$ llvm-dis ch6_1.bc -o ch6_1.ll 118-165-66-82:InputFiles Jonathan$ cat ch6_1.ll ; ModuleID = ch6_1.bc target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-

93

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16: 32:64-S128" target triple = "x86_64-apple-macosx10.8.0" @gI = global i32 100, align 4 define i32 @main() nounwind uwtable ssp { %1 = alloca i32, align 4 %c = alloca i32, align 4 store i32 0, i32* %1 store i32 0, i32* %c, align 4 %2 = load i32* @gI, align 4 store i32 %2, i32* %c, align 4 %3 = load i32* %c, align 4 ret i32 %3 } 118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch6_1.bc -o ch6_1.cpu0.s 118-165-66-82:InputFiles Jonathan$ cat ch6_1.cpu0.s .section .mdebug.abi32 .previous .file "ch6_1.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .cfi_startproc .frame $sp,8,$lr .mask 0x00000000,0 .set noreorder .cpload $t9 .set nomacro # BB#0: addiu $sp, $sp, -8 $tmp1: .cfi_def_cfa_offset 8 addiu $2, $zero, 0 st $2, 4($sp) st $2, 0($sp) ld $2, %got(gI)($gp) ld $2, 0($2) st $2, 0($sp) addiu $sp, $sp, 8 ret $lr .set macro .set reorder .end main $tmp2: .size main, ($tmp2)-main .cfi_endproc .type gI,@object .data .globl gI # @gI

94

Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.align 2 gI: .4byte 100 .size gI, 4

# 0x64

As above code, it translate load i32* @gI, align 4 into ld $2, %got(gI)($gp) for llc -march=cpu0 -relocation-model=pic, position-independent mode. More specically, it translate the global integer variable gI address into offset of register gp and load from $gp+(the offset) into register $2.

7.1.1 Static mode

We can also translate it with absolute address mode by following command,
118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=static -filetype=asm ch6_1.bc -o ch6_1.cpu0.static.s 118-165-66-82:InputFiles Jonathan$ cat ch6_1.cpu0.static.s ... addiu $2, $zero, %hi(gI) shl $2, $2, 16 addiu $2, $2, %lo(gI) ld $2, 0($2)

Above code, it load the high address part of gI absolute address (16 bits) to register $2 and shift 16 bits. Now, the register $2 got its high part of gI absolute address. Next, it load the low part of gI absolute address into register 3. Finally, add register $2 and $3 into $2, and load the content of address $2+offset 0 into register $2. The llc -relocationmodel=static is for static link mode which binding the address in static, compile/link time, not dynamic/run time. In this mode, you can also translate code with following command,
118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=static -cpu0-islinux-f ormat=false -filetype=asm ch6_1.bc -o ch6_1.cpu0.islinux-format-false.s 118-165-66-82:InputFiles Jonathan$ cat ch6_1.cpu0.islinux-format-false.s ... st $2, 0($sp) addiu $2, $gp, %gp_rel(gI) ld $2, 0($2) ... .section .sdata,"aw",@progbits .globl gI

As above, it translate code with llc -relocation-model=static -cpu0-islinux-format=false. The -cpu0-islinux-format default is true which will allocate global variables in data section. With false, it will allocate global variables in sdata section. Section data and sdata are areas for global variable with initial value, int gI = 100 in this example. Section bss and sbss are areas for global variables without initial value (for example, int gI;). Allocate variables in sdata or sbss sections is addressable by 16 bits + $gp. The static mode with -cpu0-islinux-format=false is still static mode (variable is binding in compile/link time) even its use $gp relative address. The $gp content is assigned in compile/link time, change only in program be loaded, and is xed during running the program; while the -relocation-model=pic the $gp can be changed during program running. For example, if $gp is assigned to start of .sdata like this example, then %gp_rel(gI) = (the relative address distance between gI and $gp) (is 0 in this case). When sdata is loaded into address x, then the gI variable can be got from address x+0 where x is the address stored in $gp, 0 is $gp_rel(gI). To support global variable, rst add IsLinuxOpt command variable to Cpu0Subtarget.cpp. After that, user can run llc with argument llc -cpu0-islinux-format=false to specify IsLinuxOpt to false. The IsLinuxOpt is default to true if without specify it. About the cl command, you can refer to http://llvm.org/docs/CommandLine.html further.

7.1. Global variable

95

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Cpu0Subtarget.cpp static cl::opt<bool> IsLinuxOpt("cpu0-islinux-format", cl::Hidden, cl::init(true), cl::desc("Always use linux format.")); Next add the following code to Cpu0ISelLowering.cpp. // Cpu0ISelLowering.cpp Cpu0TargetLowering:: Cpu0TargetLowering(Cpu0TargetMachine &TM) : TargetLowering(TM, new Cpu0TargetObjectFile()), Subtarget(&TM.getSubtarget<Cpu0Subtarget>()) { ... // Cpu0 Custom Operations setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); ... } SDValue Cpu0TargetLowering:: LowerOperation(SDValue Op, SelectionDAG &DAG) const { switch (Op.getOpcode()) { case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG); } return SDValue(); } //===----------------------------------------------------------------------===// // Lower helper functions //===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===// // Misc Lower Operation implementation //===----------------------------------------------------------------------===// SDValue Cpu0TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const { // FIXME there isnt actually debug info here DebugLoc dl = Op.getDebugLoc(); const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); if (getTargetMachine().getRelocationModel() != Reloc::PIC_) { SDVTList VTs = DAG.getVTList(MVT::i32); Cpu0TargetObjectFile &TLOF = (Cpu0TargetObjectFile&)getObjFileLowering(); // %gp_rel relocation if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) { SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_GPREL); SDValue GPRelNode = DAG.getNode(Cpu0ISD::GPRel, dl, VTs, &GA, 1); SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32); return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode); } // %hi/%lo relocation SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_ABS_HI); SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_ABS_LO);

96

Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

SDValue HiPart = DAG.getNode(Cpu0ISD::Hi, dl, VTs, &GAHi, 1); SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, MVT::i32, GALo); return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo); } EVT ValTy = Op.getValueType(); bool HasGotOfst = (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV))); unsigned GotFlag = (HasGotOfst ? Cpu0II::MO_GOT : Cpu0II::MO_GOT16); SDValue GA = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, GotFlag); GA = DAG.getNode(Cpu0ISD::Wrapper, dl, ValTy, GetGlobalReg(DAG, ValTy), GA); SDValue ResNode = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), GA, MachinePointerInfo(), false, false, false, 0); // On functions and global targets not internal linked only // a load from got/GP is necessary for PIC to work. if (!HasGotOfst) return ResNode; SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, Cpu0II::MO_ABS_LO); SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, ValTy, GALo); return DAG.getNode(ISD::ADD, dl, ValTy, ResNode, Lo); }

The setOperationAction(ISD::GlobalAddress, MVT::i32, Custom) tell llc that we implement global address operation in C++ function Cpu0TargetLowering::LowerOperation() and llvm will call this function only when llvm want to translate IR DAG of loading global variable into machine code. Since may have many Custom type of setOperationAction(ISD::XXX, MVT::XXX, Custom) in construction function Cpu0TargetLowering(), and llvm will call Cpu0TargetLowering::LowerOperation() for each ISD IR DAG node of Custom type translation. The global address access can be identied by check the DAG node of opcode is ISD::GlobalAddress. For static mode, LowerGlobalAddress() will check the translation is for IsGlobalInSmallSection() or not. When IsLinuxOpt is true and static mode, IsGlobalInSmallSection() always return false. LowerGlobalAddress() will translate global variable by create 2 DAG IR nodes ABS_HI and ABS_LO for high part and low part of address and one extra node ADD. List it again as follows.
// Cpu0ISelLowering.cpp ... // %hi/%lo relocation SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_ABS_HI); SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_ABS_LO); SDValue HiPart = DAG.getNode(Cpu0ISD::Hi, dl, VTs, &GAHi, 1); SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, MVT::i32, GALo); return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo);

The DAG list form for these three DAG nodes as above code created can be represented as (ADD (Hi(h1, h2), Lo (l1, l2)). Since some DAG nodes are not with two arguments, we will dene the list as (ADD (Hi (...), Lo (...)) or (ADD (Hi, Lo)) sometimes in this book. The corresponding machine instructions of these three IR nodes are dened in Cpu0InstrInfo.td as follows,
// Cpu0InstrInfo.td // Hi and Lo nodes are used to handle global addresses. Used on // Cpu0ISelLowering to lower stuff like GlobalAddress, ExternalSymbol // static model. (nothing to do with Cpu0 Registers Hi and Lo) def Cpu0Hi : SDNode<"Cpu0ISD::Hi", SDTIntUnaryOp>; def Cpu0Lo : SDNode<"Cpu0ISD::Lo", SDTIntUnaryOp>; ... // hi/lo relocs def : Pat<(Cpu0Hi tglobaladdr:$in), (SHL (ADDiu ZERO, tglobaladdr:$in), 16)>;

7.1. Global variable

97

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Expect cpu0 add LUi support, like Mips //def : Pat<(Cpu0Hi tglobaladdr:$in), (LUi tglobaladdr:$in)>; def : Pat<(Cpu0Lo tglobaladdr:$in), (ADDiu ZERO, tglobaladdr:$in)>; def : Pat<(add CPURegs:$hi, (Cpu0Lo tglobaladdr:$lo)), (ADDiu CPURegs:$hi, tglobaladdr:$lo)>; // gp_rel relocs def : Pat<(add CPURegs:$gp, (Cpu0GPRel tglobaladdr:$in)), (ADDiu CPURegs:$gp, tglobaladdr:$in)>;

Above code meaning translate ABS_HI into ADDiu and SHL two instructions. Remember the DAG and Instruction Selection introduced in chapter Back end structure, DAG list (SHL (ADDiu ...), 16) meaning DAG node ADDiu and its parent DAG node SHL two instructions nodes is for list IR DAG ABS_HI. The Pat<> has two list DAG representation. The left is IR DAG and the right is machine instruction DAG. So after Instruction Selection and Register Allocation, it translate ABS_HI to,
addiu $2, %hi(gI) shl $2, $2, 16

According above code, we know llvm allocate register $2 for the output operand of ADDiu instruction and $2 for SHL instruction in this example. Since (SHL (ADDiu), 16), the ADDiu output result will be the SHL rst register. The result is shl $2, 16. Above Pat<> also dene DAG list (add $hi, (ABS_LO)) will be translated into (ADD $hi, (ADDiu ZERO, ...)) where ADD is machine instruction add and ADDiu is machine instruction ldi which dened in Cpu0InstrInfo.td too. Remember (add $hi, (ABS_LO)) meaning add DAG has two operands, the rst is $hi and the second is the register which the ABS_LO output result register save to. So, the IR DAG pattern and its corresponding machine instruction node as follows,
addiu $3, %lo(gI) // def : Pat<(Cpu0Lo tglobaladdr:$in), (ADDiu ZERO, // tglobaladdr:$in)>;

// def : Pat<(add CPURegs:$hi, (Cpu0Lo tglobaladdr:$lo)), (ADD CPURegs:$hi, // (LDI ZERO, tglobaladdr:$lo))>; // So, the second register for add is the output register of ABS_LO IR DAG // translation result saved to; // Since LowerGlobalAddress() create list (ADD (Hi, Lo)) with 3 DAG nodes, // the Hi output register $2 will be the first input register for add. add $2, $2, $3

After translated as above, the register $2 is the global variable address, so get the global variable by IR DAG load will translate into machine instruction as follows,
%2 = load i32* @gI, align 4 => ld $2, 0($2)

When IsLinuxOpt is false and static mode, LowerGlobalAddress() will run the following code to create a DAG list (ADD GOT, GPRel).
// %gp_rel relocation if (TLOF.IsGlobalInSmallSection(GV, getTargetMachine())) { SDValue GA = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_GPREL); SDValue GPRelNode = DAG.getNode(Cpu0ISD::GPRel, dl, VTs, &GA, 1); SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(MVT::i32); return DAG.getNode(ISD::ADD, dl, MVT::i32, GOT, GPRelNode); }

As mentioned just before, all global variables allocated in sdata or sbss sections which is addressable by 16 bits + $gp in compile/link time (address binding in compile time). Its equal to offset+GOT where GOT is the base address for 98 Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

global variable and offset is 16 bits. Now, according the following Cpu0InstrInfo.td denition,
// Cpu0InstrInfo.td def Cpu0GPRel : SDNode<"Cpu0ISD::GPRel", SDTIntUnaryOp>; ... // gp_rel relocs def : Pat<(add CPURegs:$gp, (Cpu0GPRel tglobaladdr:$in)), (ADD CPURegs:$gp, (ADDiu ZERO, tglobaladdr:$in))>;

It translate global variable address of list (ADD GOT, GPRel) into machine instructions as follows,
addiu $2, $gp, %gp_rel(gI)

7.1.2 PIC mode

When PIC mode, LowerGlobalAddress() will create the DAG list (load DAG.getEntryNode(), (Wrapper GetGlobalReg(), GA)) by the following code and the code in Cpu0ISeleDAGToDAG.cpp as follows,
bool HasGotOfst = (GV->hasInternalLinkage() || (GV->hasLocalLinkage() && !isa<Function>(GV))); unsigned GotFlag = (HasGotOfst ? Cpu0II::MO_GOT : Cpu0II::MO_GOT16); SDValue GA = DAG.getTargetGlobalAddress(GV, dl, ValTy, 0, GotFlag); GA = DAG.getNode(Cpu0ISD::Wrapper, dl, ValTy, GetGlobalReg(DAG, ValTy), GA); SDValue ResNode = DAG.getLoad(ValTy, dl, DAG.getEntryNode(), GA, MachinePointerInfo(), false, false, false, 0); // On functions and global targets not internal linked only // a load from got/GP is necessary for PIC to work. if (!HasGotOfst) return ResNode; // Cpu0ISelDAGToDAG.cpp /// ComplexPattern used on Cpu0InstrInfo /// Used on Cpu0 Load/Store instructions bool Cpu0DAGToDAGISel:: SelectAddr(SDNode *Parent, SDValue Addr, SDValue &Base, SDValue &Offset) { ... // on PIC code Load GA if (Addr.getOpcode() == Cpu0ISD::Wrapper) { = Addr.getOperand(0); Base Offset = Addr.getOperand(1); return true; } ... }

Then it translate into the following code,

ld $2, %got(gI)($gp)

Where DAG.getEntryNode() is the register $2 which decide by Register Allocator, and (Wrapper GetGlobalReg(), GA) translate into Base=$gp and the 16 bits Offset for $gp. Apart from above code, add the following code to Cpu0AsmPrinter.cpp and it will emit .cpload asm pseudo instruction,
// Cpu0AsmPrinter.cpp /// EmitFunctionBodyStart - Targets can override this to emit stuff before /// the first basic block in the function. void Cpu0AsmPrinter::EmitFunctionBodyStart() {

7.1. Global variable

99

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

... // Emit .cpload directive if needed. if (EmitCPLoad) //- .cpload $t9 OutStreamer.EmitRawText(StringRef("\t.cpload\t$t9")); ... } // ch6_1.cpu0.s .cpload $t9 .set nomacro # BB#0: ldi $sp, -8

According Mips Application Binary Interface (ABI), $t9 ($25) is the register used in jalr $25 for long distance function pointer (far subroutine call). The jal %subroutine has 24 bits range of address offset relative to Program Counter (PC) while jalr has 32 bits address range in register size is 32 bits. One example of PIC mode is used in share library. Share library is re-entry code which can be loaded in different memory address decided on run time. The static mode (absolute address mode) is usually designed to load in specic memory address decided on compile time. Since share library can be loaded in different memory address, the global variable address cannot be decided in compile time. As above, the global variable address is translated into the relative address of $gp. In example code ch6_1.ll, .cpload is a asm pseudo instruction just before the rst instruction of main(), ldi. When the share library main() function be loaded, the loader will assign the $t9 value to $gp when it meet .cpload $t9. After that, the $gp value is $9 which point to main(), and the global variable address is the relative address to main().

7.1.3 Global variable print support

Above code is for global address DAG translation. Next, add the following code to Cpu0MCInstLower.cpp and Cpu0InstPrinter.cpp for global variable printing operand function.
// Cpu0MCInstLower.cpp MCOperand Cpu0MCInstLower::LowerSymbolOperand(const MachineOperand &MO, MachineOperandType MOTy, unsigned Offset) const { MCSymbolRefExpr::VariantKind Kind; const MCSymbol *Symbol; switch(MO.getTargetFlags()) { default: llvm_unreachable("Invalid target flag!"); // Cpu0_GPREL is for llc -march=cpu0 -relocation-model=static // -cpu0-islinux-format=false (global var in .sdata) case Cpu0II::MO_GPREL: Kind = MCSymbolRefExpr::VK_Cpu0_GPREL; break; case Cpu0II::MO_GOT16: Kind = MCSymbolRefExpr::VK_Cpu0_GOT16; break; case Cpu0II::MO_GOT: Kind = MCSymbolRefExpr::VK_Cpu0_GOT; break; // ABS_HI and ABS_LO is for llc -march=cpu0 -relocation-model=static // (global var in .data) case Cpu0II::MO_ABS_HI: Kind = MCSymbolRefExpr::VK_Cpu0_ABS_HI; break; case Cpu0II::MO_ABS_LO: Kind = MCSymbolRefExpr::VK_Cpu0_ABS_LO; break; } switch (MOTy) { case MachineOperand::MO_GlobalAddress: Symbol = Mang->getSymbol(MO.getGlobal()); break; default:

100

Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

llvm_unreachable("<unknown operand type>"); } ... } MCOperand Cpu0MCInstLower::LowerOperand(const MachineOperand& MO, unsigned offset) const { MachineOperandType MOTy = MO.getType(); switch (MOTy) { ... case MachineOperand::MO_GlobalAddress: return LowerSymbolOperand(MO, MOTy, offset); ... } // Cpu0InstPrinter.cpp ... static void printExpr(const MCExpr *Expr, raw_ostream &OS) { ... switch (Kind) { default: llvm_unreachable("Invalid kind!"); case MCSymbolRefExpr::VK_None: break; // Cpu0_GPREL is for llc -march=cpu0 -relocation-model=static case MCSymbolRefExpr::VK_Cpu0_GPREL: OS << "%gp_rel("; break; case MCSymbolRefExpr::VK_Cpu0_GOT16: OS << "%got("; break; case MCSymbolRefExpr::VK_Cpu0_GOT: OS << "%got("; break; case MCSymbolRefExpr::VK_Cpu0_ABS_HI: OS << "%hi("; break; case MCSymbolRefExpr::VK_Cpu0_ABS_LO: OS << "%lo("; break; } ... }

OS is the output stream which output to the assembly le.

7.1.4 Summary
The global variable Instruction Selection for DAG translation is not like the ordinary IR node translation, it has static (absolute address) and PIC mode. Backend deal this translation by create DAG nodes in function LowerGlobalAddress() which called by LowerOperation(). Function LowerOperation() take care all Custom type of operation. Backend set global address as Custom operation by setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); in Cpu0TargetLowering() constructor. Different address mode has its corresponding DAG list be created. By set the pattern Pat<> in Cpu0InstrInfo.td, the llvm can apply the compiler mechanism, pattern match, in the Instruction Selection stage. There are three type for setXXXAction(), Promote, Expand and Custom. Except Custom, the other two usually no need to coding. The section Instruction Selector of http://llvm.org/docs/WritingAnLLVMBackend.html is the references.

7.2 Array and struct support

Shifting our work to iMac at this point. The Linux platform is ne. The reason we do the shift is for new platform using experience. LLVM use getelementptr to represent the array and struct type in C. Please reference section getelementptr of 1 . For
1

http://llvm.org/docs/CommandLine.html

7.2. Array and struct support

101

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

ch6_2.cpp, the llvm IR as follows,

// ch6_2.cpp struct Date { int year; int month; int day; }; Date date = {2012, 10, 12}; int a[3] = {2012, 10, 12}; int main() { int day = date.day; int i = a[1]; return 0; } // ch6_2.ll ; ModuleID = ch6_2.bc target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128" target triple = "i386-apple-macosx10.8.0" %struct.Date = type { i32, i32, i32 } @date = global %struct.Date { i32 2012, i32 10, i32 12 }, align 4 @a = global [3 x i32] [i32 2012, i32 10, i32 12], align 4 define i32 @main() nounwind ssp { entry: %retval = alloca i32, align 4 %day = alloca i32, align 4 %i = alloca i32, align 4 store i32 0, i32* %retval %0 = load i32* getelementptr inbounds (%struct.Date* @date, i32 0, i32 2), align 4 store i32 %0, i32* %day, align 4 %1 = load i32* getelementptr inbounds ([3 x i32]* @a, i32 0, i32 1), align 4 store i32 %1, i32* %i, align 4 ret i32 0 }

Run 6/1/Cpu0 with ch6_2.bc on static mode will get the incorrect asm le as follows,
118-165-66-82:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_ debug_build/bin/Debug/llc -march=cpu0 -relocation-model=static -filetype=asm ch6_2.bc -o ch6_2.cpu0.static.s 118-165-66-82:InputFiles Jonathan$ cat ch6_2.cpu0.static.s .section .mdebug.abi32 .previous .file "ch6_2.bc" .text .globl main .align 2 .type main,@function

102

Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.ent main # @main main: .cfi_startproc .frame $sp,16,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: addiu $sp, $sp, -16 $tmp1: .cfi_def_cfa_offset 16 addiu $2, $zero, 0 st $2, 12($sp) addiu $2, $zero, %hi(date) shl $2, $2, 16 addiu $2, $2, %lo(date) ld $2, 0($2) // the correct one is st $2, 8($sp) addiu $2, $zero, %hi(a) shl $2, $2, 16 addiu $2, $2, %lo(a) ld $2, 0($2) st $2, 4($sp) addiu $sp, $sp, 16 ret $lr .set macro .set reorder .end main $tmp2: .size main, ($tmp2)-main .cfi_endproc .type date,@object .data .globl date .align 2 date: .4byte 2012 .4byte 10 .4byte 12 .size date, 12 .type a,@object .globl a .align 2 a: .4byte 2012 .4byte 10 .4byte 12 .size a, 12 # @date

ld

$2, 8($2)

# 0x7dc # 0xa # 0xc

# @a

# 0x7dc # 0xa # 0xc

For day = date.day, the correct one is ld $2, 8($2), not ld $2, 0($2), since date.day is offset 8(date). Type int is 4 bytes in cpu0, and the date.day has elds year and month before it. Let use debug option in llc to see whats wrong,
jonathantekiimac:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -debug -relocation-model=static -filetype=asm ch6_2.bc -o ch6_2.cpu0.static.s ...

7.2. Array and struct support

103

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

=== main Initial selection DAG: BB#0 main:entry SelectionDAG has 20 nodes: 0x7f7f5b02d210: i32 = undef [ORD=1] 0x7f7f5ac10590: ch = EntryToken [ORD=1] 0x7f7f5b02d010: i32 = Constant<0> [ORD=1] 0x7f7f5b02d110: i32 = FrameIndex<0> [ORD=1] 0x7f7f5b02d210: <multiple use> 0x7f7f5b02d310: ch = store 0x7f7f5ac10590, 0x7f7f5b02d010, 0x7f7f5b02d110, 0x7f7f5b02d210<ST4[%retval]> [ORD=1] 0x7f7f5b02d410: i32 = GlobalAddress<%struct.Date* @date> 0 [ORD=2] 0x7f7f5b02d510: i32 = Constant<8> [ORD=2] 0x7f7f5b02d610: i32 = add 0x7f7f5b02d410, 0x7f7f5b02d510 [ORD=2] 0x7f7f5b02d210: <multiple use> 0x7f7f5b02d710: i32,ch = load 0x7f7f5b02d310, 0x7f7f5b02d610, 0x7f7f5b02d210 <LD4[getelementptr inbounds (%struct.Date* @date, i32 0, i32 2)]> [ORD=3] 0x7f7f5b02db10: i64 = Constant<4> 0x7f7f5b02d710: <multiple use> 0x7f7f5b02d710: <multiple use> 0x7f7f5b02d810: i32 = FrameIndex<1> [ORD=4] 0x7f7f5b02d210: <multiple use> 0x7f7f5b02d910: ch = store 0x7f7f5b02d710:1, 0x7f7f5b02d710, 0x7f7f5b02d810, 0x7f7f5b02d210<ST4[%day]> [ORD=4] 0x7f7f5b02da10: i32 = GlobalAddress<[3 x i32]* @a> 0 [ORD=5] 0x7f7f5b02dc10: i32 = Constant<4> [ORD=5] 0x7f7f5b02dd10: i32 = add 0x7f7f5b02da10, 0x7f7f5b02dc10 [ORD=5] 0x7f7f5b02d210: <multiple use> 0x7f7f5b02de10: i32,ch = load 0x7f7f5b02d910, 0x7f7f5b02dd10, 0x7f7f5b02d210 <LD4[getelementptr inbounds ([3 x i32]* @a, i32 0, i32 1)]> [ORD=6] ...

Replacing.3 0x7f7f5b02dd10: i32 = add 0x7f7f5b02da10, 0x7f7f5b02dc10 [ORD=5] With: 0x7f7f5b030010: i32 = GlobalAddress<[3 x i32]* @a> + 4

Replacing.3 0x7f7f5b02d610: i32 = add 0x7f7f5b02d410, 0x7f7f5b02d510 [ORD=2] With: 0x7f7f5b02db10: i32 = GlobalAddress<%struct.Date* @date> + 8 Optimized lowered selection DAG: BB#0 main:entry

104

Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

SelectionDAG has 15 nodes: 0x7f7f5b02d210: i32 = undef [ORD=1] 0x7f7f5ac10590: ch = EntryToken [ORD=1] 0x7f7f5b02d010: i32 = Constant<0> [ORD=1] 0x7f7f5b02d110: i32 = FrameIndex<0> [ORD=1] 0x7f7f5b02d210: <multiple use> 0x7f7f5b02d310: ch = store 0x7f7f5ac10590, 0x7f7f5b02d010, 0x7f7f5b02d110, 0x7f7f5b02d210<ST4[%retval]> [ORD=1] 0x7f7f5b02db10: i32 = GlobalAddress<%struct.Date* @date> + 8 0x7f7f5b02d210: <multiple use> 0x7f7f5b02d710: i32,ch = load 0x7f7f5b02d310, 0x7f7f5b02db10, 0x7f7f5b02d210 <LD4[getelementptr inbounds (%struct.Date* @date, i32 0, i32 2)]> [ORD=3] 0x7f7f5b02d710: <multiple use> 0x7f7f5b02d710: <multiple use> 0x7f7f5b02d810: i32 = FrameIndex<1> [ORD=4] 0x7f7f5b02d210: <multiple use> 0x7f7f5b02d910: ch = store 0x7f7f5b02d710:1, 0x7f7f5b02d710, 0x7f7f5b02d810, 0x7f7f5b02d210<ST4[%day]> [ORD=4] 0x7f7f5b030010: i32 = GlobalAddress<[3 x i32]* @a> + 4 0x7f7f5b02d210: <multiple use> 0x7f7f5b02de10: i32,ch = load 0x7f7f5b02d910, 0x7f7f5b030010, 0x7f7f5b02d210 <LD4[getelementptr inbounds ([3 x i32]* @a, i32 0, i32 1)]> [ORD=6] ...

By llc -debug, you can see the DAG translation process. As above, the DAG list for date.day (add GlobalAddress<[3 x i32]* @a> 0, Constant<8>) with 3 nodes is replaced by 1 node GlobalAddress<%struct.Date* @date> + 8. The DAG list for a[1] is same. The replacement occurs since TargetLowering.cpp::isOffsetFoldingLegal(...) return true in llc -static static addressing mode as below. In Cpu0 the ld instruction format is ld $r1, offset($r2) which meaning load $r2 address+offset to $r1. So, we just replace the isOffsetFoldingLegal(...) function by override mechanism as below.
// TargetLowering.cpp bool TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { // Assume that everything is safe in static mode. if (getTargetMachine().getRelocationModel() == Reloc::Static) return true; // In dynamic-no-pic mode, assume that known defined values are safe. if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && GA && !GA->getGlobal()->isDeclaration() && !GA->getGlobal()->isWeakForLinker()) return true; // Otherwise assume nothing is safe. return false;

7.2. Array and struct support

105

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

} // Cpu0TargetLowering.cpp bool Cpu0TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { // The Cpu0 target isnt yet aware of offsets. return false; }

Beyond that, we need to add the following code fragment to Cpu0ISelDAGToDAG.cpp,

// Cpu0ISelDAGToDAG.cpp /// ComplexPattern used on Cpu0InstrInfo /// Used on Cpu0 Load/Store instructions bool Cpu0DAGToDAGISel:: SelectAddr(SDNode *Parent, SDValue Addr, SDValue &Base, SDValue &Offset) { ... // Addresses of the form FI+const or FI|const if (CurDAG->isBaseWithConstantOffset(Addr)) { ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1)); if (isInt<16>(CN->getSExtValue())) { // If the first operand is a FI, get the TargetFI Node if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode> (Addr.getOperand(0))) Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), ValTy); else Base = Addr.getOperand(0); Offset = CurDAG->getTargetConstant(CN->getZExtValue(), ValTy); return true; } } }

Recall we have translated DAG list for date.day (add GlobalAddress<[3 x i32]* @a> 0, Constant<8>) into (add (add Cpu0ISD::Hi (Cpu0II::MO_ABS_HI), Cpu0ISD::Lo(Cpu0II::MO_ABS_LO)), Constant<8>) by the following code in Cpu0ISelLowering.cpp.
// Cpu0ISelLowering.cpp SDValue Cpu0TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const { ... // %hi/%lo relocation SDValue GAHi = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_ABS_HI); SDValue GALo = DAG.getTargetGlobalAddress(GV, dl, MVT::i32, 0, Cpu0II::MO_ABS_LO); SDValue HiPart = DAG.getNode(Cpu0ISD::Hi, dl, VTs, &GAHi, 1); SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, MVT::i32, GALo); return DAG.getNode(ISD::ADD, dl, MVT::i32, HiPart, Lo); ... }

So, when the SelectAddr(...) of Cpu0ISelDAGToDAG.cpp is called. The Addr SDValue in SelectAddr(..., Addr, ...) is DAG list for date.day (add (add Cpu0ISD::Hi (Cpu0II::MO_ABS_HI), Cpu0ISD::Lo(Cpu0II::MO_ABS_LO)), Constant<8>). Since Addr.getOpcode() = ISD:ADD, Addr.getOperand(0) = (add Cpu0ISD::Hi (Cpu0II::MO_ABS_HI), Cpu0ISD::Lo(Cpu0II::MO_ABS_LO)) and Addr.getOperand(1).getOpcode() = ISD::Constant, the Base = SDValue (add Cpu0ISD::Hi (Cpu0II::MO_ABS_HI), Cpu0ISD::Lo(Cpu0II::MO_ABS_LO)) and Offset = Constant<8>. After 106 Chapter 7. Global variables, structs and arrays

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

set Base and Offset, the load DAG will translate the global address date.day into machine instruction ld $r1, 8($r2) in Instruction Selection stage. 6/2/Cpu0 include these changes as above, you can run it with ch6_2.cpp to get the correct generated instruction ld $r1, 8($r2) for date.day access, as follows.
... ld $2, 8($2) st $2, 8($sp) addiu $2, $zero, %hi(a) shl $2, $2, 16 addiu $2, $2, %lo(a) ld $2, 4($2)

7.2. Array and struct support

107

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

108

Chapter 7. Global variables, structs and arrays

CHAPTER

EIGHT

CONTROL FLOW STATEMENTS

This chapter illustrates the corresponding IR for control ow statements, like if else, while and for loop statements in C, and how to translate these control ow statements of llvm IR into cpu0 instructions.

8.1 Control ow statement

Run ch7_1_1.cpp with clang will get result as follows,
// ch7_1_1.cpp int main() { unsigned int a = 0; int b = 1; int c = 2; int d = 3; int e = 4; int f = 5; int g = 6; int h = 7; int i = 8; if (a == 0) { a++; } if (b != 0) { b++; } if (c > 0) { c++; } if (d >= 0) { d++; } if (e < 0) { e++; } if (f <= 0) { f++; } if (g <= 1) { g++; }

109

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

if (h >= 1) { h++; } if (i < h) { i++; } if (a != b) { a++; } return a; } ; ModuleID = ch7_1_1.bc target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128" target triple = "i386-apple-macosx10.8.0" define i32 @main() nounwind ssp { entry: %retval = alloca i32, align 4 %a = alloca i32, align 4 %b = alloca i32, align 4 %c = alloca i32, align 4 %d = alloca i32, align 4 %e = alloca i32, align 4 %f = alloca i32, align 4 %g = alloca i32, align 4 %h = alloca i32, align 4 %i = alloca i32, align 4 store i32 0, i32* %retval store i32 0, i32* %a, align 4 store i32 1, i32* %b, align 4 store i32 2, i32* %c, align 4 store i32 3, i32* %d, align 4 store i32 4, i32* %e, align 4 store i32 5, i32* %f, align 4 store i32 6, i32* %g, align 4 store i32 7, i32* %h, align 4 store i32 8, i32* %i, align 4 %0 = load i32* %a, align 4 %cmp = icmp eq i32 %0, 0 br i1 %cmp, label %if.then, label %if.end if.then: %1 = load i32* %a, align 4 %inc = add i32 %1, 1 store i32 %inc, i32* %a, align 4 br label %if.end if.end: %2 = load i32* %b, align 4 %cmp1 = icmp ne i32 %2, 0 br i1 %cmp1, label %if.then2, label %if.end4 if.then2: %3 = load i32* %b, align 4 %inc3 = add nsw i32 %3, 1 ; preds = %entry

; preds = %if.then, %entry

; preds = %if.end

110

Chapter 8. Control ow statements

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

store i32 %inc3, i32* %b, align 4 br label %if.end4 if.end4: %4 = load i32* %c, align 4 %cmp5 = icmp sgt i32 %4, 0 br i1 %cmp5, label %if.then6, label %if.end8 if.then6: %5 = load i32* %c, align 4 %inc7 = add nsw i32 %5, 1 store i32 %inc7, i32* %c, align 4 br label %if.end8 ; preds = %if.then2, %if.end

; preds = %if.end4

if.end8: ; preds = %if.then6, %if.end4 %6 = load i32* %d, align 4 %cmp9 = icmp sge i32 %6, 0 br i1 %cmp9, label %if.then10, label %if.end12 if.then10: %7 = load i32* %d, align 4 %inc11 = add nsw i32 %7, 1 store i32 %inc11, i32* %d, align 4 br label %if.end12 ; preds = %if.end8

if.end12: ; preds = %if.then10, %if.end8 %8 = load i32* %e, align 4 %cmp13 = icmp slt i32 %8, 0 br i1 %cmp13, label %if.then14, label %if.end16 if.then14: %9 = load i32* %e, align 4 %inc15 = add nsw i32 %9, 1 store i32 %inc15, i32* %e, align 4 br label %if.end16 ; preds = %if.end12

if.end16: ; preds = %if.then14, %if.end12 %10 = load i32* %f, align 4 %cmp17 = icmp sle i32 %10, 0 br i1 %cmp17, label %if.then18, label %if.end20 if.then18: %11 = load i32* %f, align 4 %inc19 = add nsw i32 %11, 1 store i32 %inc19, i32* %f, align 4 br label %if.end20 ; preds = %if.end16

if.end20: ; preds = %if.then18, %if.end16 %12 = load i32* %g, align 4 %cmp21 = icmp sle i32 %12, 1 br i1 %cmp21, label %if.then22, label %if.end24 if.then22: %13 = load i32* %g, align 4 %inc23 = add nsw i32 %13, 1 store i32 %inc23, i32* %g, align 4 br label %if.end24 ; preds = %if.end20

8.1. Control ow statement

111

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

if.end24: ; preds = %if.then22, %if.end20 %14 = load i32* %h, align 4 %cmp25 = icmp sge i32 %14, 1 br i1 %cmp25, label %if.then26, label %if.end28 if.then26: %15 = load i32* %h, align 4 %inc27 = add nsw i32 %15, 1 store i32 %inc27, i32* %h, align 4 br label %if.end28 ; preds = %if.end24

if.end28: ; preds = %if.then26, %if.end24 %16 = load i32* %i, align 4 %17 = load i32* %h, align 4 %cmp29 = icmp slt i32 %16, %17 br i1 %cmp29, label %if.then30, label %if.end32 if.then30: %18 = load i32* %i, align 4 %inc31 = add nsw i32 %18, 1 store i32 %inc31, i32* %i, align 4 br label %if.end32 ; preds = %if.end28

if.end32: ; preds = %if.then30, %if.end28 %19 = load i32* %a, align 4 %20 = load i32* %b, align 4 %cmp33 = icmp ne i32 %19, %20 br i1 %cmp33, label %if.then34, label %if.end36 if.then34: %21 = load i32* %a, align 4 %inc35 = add i32 %21, 1 store i32 %inc35, i32* %a, align 4 br label %if.end36 if.end36: %22 = load i32* %a, align 4 ret i32 %22 } ; preds = %if.end32

; preds = %if.then34, %if.end32

The icmp ne stand for integer compare NotEqual, slt stand for Set Less Than, sle stand for Set Less Equal. Run version 6/2/Cpu0 with llc -view-isel-dags or -debug option, you can see it has translated if statement into (br (brcond (%1, setcc(%2, Constant<c>, setne)), BasicBlock_02), BasicBlock_01). Ignore %1, we get the form (br (brcond (setcc(%2, Constant<c>, setne)), BasicBlock_02), BasicBlock_01). For explanation, We list the IR DAG as follows,
%cond=setcc(%2, Constant<c>, setne) brcond %cond, BasicBlock_02 br BasicBlock_01 We want to translate them into cpu0 instructions DAG as follows, addiu %3, ZERO, Constant<c> cmp %2, %3 jne BasicBlock_02 jmp BasicBlock_01

For the rst addiu instruction as above which move Constant<c> into register, we have dened it before by the following code,

112

Chapter 8. Control ow statements

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Cpu0InstrInfo.td ... // Small immediates def : Pat<(i32 immSExt16:$in), (ADDiu ZERO, imm:$in)>; // Arbitrary immediates def : Pat<(i32 imm:$imm), (OR (SHL (ADDiu ZERO, (HI16 imm:$imm)), 16), (ADDiu ZERO, (LO16 imm:$imm)))>;

For the last IR br, we translate unconditional branch (br BasicBlock_01) into jmp BasicBlock_01 by the following pattern denition,
def brtarget : Operand<OtherVT> { let EncoderMethod = "getBranchTargetOpValue"; let OperandType = "OPERAND_PCREL"; let DecoderMethod = "DecodeBranchTarget"; } ... // Unconditional branch class UncondBranch<bits<8> op, string instr_asm>: BranchBase<op, (outs), (ins brtarget:$imm24), !strconcat(instr_asm, "\t$imm24"), [(br bb:$imm24)], IIBranch> { let isBranch = 1; let isTerminator = 1; let isBarrier = 1; let hasDelaySlot = 0; } ... def JMP : UncondBranch<0x26, "jmp">;

The pattern [(br bb:$imm24)] in class UncondBranch is translated into jmp machine instruction. The other two cpu0 instructions translation is more complicate than simple one-to-one IR to machine instruction translation we have experienced until now. To solve this chained IR to machine instructions translation, we dene the following pattern,
// brcond patterns multiclass BrcondPats<RegisterClass RC, Instruction JEQOp, Instruction JNEOp, Instruction JLTOp, Instruction JGTOp, Instruction JLEOp, Instruction JGEOp, Instruction CMPOp> { ... def : Pat<(brcond (i32 (setne RC:$lhs, RC:$rhs)), bb:$dst), (JNEOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; ... def : Pat<(brcond RC:$cond, bb:$dst), (JNEOp (CMPOp RC:$cond, ZEROReg), bb:$dst)>;

Above denition support (setne RC:$lhs, RC:$rhs) register to register compare. There are other compare pattern like, seteq, setlt, . . . . In addition to seteq, setne, . . . , we dene setueq, setune, . . . , by reference Mips code even though we didnt nd how setune came from. We have tried to dene unsigned int type, but clang still generate setne instead of setune. Pattern search order is according their appear order in context. The last pattern (brcond RC:$cond, bb:$dst) is meaning branch to $dst if $cond != 0, it is equal to (JNEOp (CMPOp RC:$cond, ZEROReg), bb:$dst) in cpu0 translation. The CMP instruction will set the result to register SW, and then JNE check the condition based on SW status. Since SW is a reserved register, it will be correct even an instruction is inserted between CMP and JNE as follows,

8.1. Control ow statement

113

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

cmp %2, %3 addiu $r1, $r2, 3 jne BasicBlock_02

// $r1 register never be allocated to $SW

The reserved registers setting by the following function code we dened before,
// Cpu0RegisterInfo.cpp ... // pure virtual method BitVector Cpu0RegisterInfo:: getReservedRegs(const MachineFunction &MF) const { static const uint16_t ReservedCPURegs[] = { Cpu0::ZERO, Cpu0::AT, Cpu0::GP, Cpu0::FP, Cpu0::SW, Cpu0::SP, Cpu0::LR, Cpu0::PC }; BitVector Reserved(getNumRegs()); typedef TargetRegisterClass::iterator RegIter; for (unsigned I = 0; I < array_lengthof(ReservedCPURegs); ++I) Reserved.set(ReservedCPURegs[I]); // If GP is dedicated as a global base register, reserve it. if (MF.getInfo<Cpu0FunctionInfo>()->globalBaseRegFixed()) { Reserved.set(Cpu0::GP); } return Reserved; }

Although the following denition in Cpu0RegisterInfo.td has no real effect in Reserved Registers, you should comment the Reserved Registers in it for readability.
// Cpu0RegisterInfo.td ... //===----------------------------------------------------------------------===// // Register Classes //===----------------------------------------------------------------------===// def CPURegs : RegisterClass<"Cpu0", [i32], 32, (add // Return Values and Arguments V0, V1, A0, A1, // Not preserved across procedure calls T9, // Callee save S0, S1, S2, // Reserved ZERO, AT, GP, FP, SW, SP, LR, PC)>;

By the following llc option, you can get the obj le and dump its content by hexdump as follows,
118-165-79-206:InputFiles Jonathan$ cat ch7_1_1.cpu0.s ... ld $3, 32($sp) cmp $3, $2 jne $BB0_2 jmp $BB0_1 $BB0_1: # %if.then ld $2, 32($sp) addiu $2, $2, 1

114

Chapter 8. Control ow statements

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

st $2, 32($sp) $BB0_2: ld $2, 28($sp) ...

# %if.end

118-165-79-206:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=obj ch7_1_1.bc -o ch7_1_1.cpu0.o 118-165-79-206:InputFiles Jonathan$ hexdump ch7_1_1.cpu0.o // jmp offset is 0x10=16 bytes which is correct 0000080 ............................ 10 20 20 02 21 00 00 10 0000090 26 00 00 00 ...............................................

The immediate value of jne (op 0x21) is 16; The offset between jne and $BB0_2 is 20 (5 words = 5*4 bytes). Suppose the jne address is X, then the label $BB0_2 is X+20. Cpu0 is a RISC cpu0 with 3 stages of pipeline which are fetch, decode and execution according to cpu0 web site information. The cpu0 do branch instruction execution at decode stage which like mips. After the jne instruction fetched, the PC (Program Counter) is X+4 since cpu0 update PC at fetch stage. The $BB0_2 address is equal to PC+16 for the jne branch instruction execute at decode stage. List and explain this again as follows,
// // // // // // Fetch instruction stage for jne instruction. The fetch stage can be divided into 2 cycles. First cycle fetch the instruction. Second cycle adjust PC = PC+4. Do jne compare in decode stage. PC = X+4 at this stage. When jne immediate value is 16, PC = PC+16. It will fetch X+20 which equal to label $BB0_2 instruction, ld $2, 28($sp). # %if.then

jne $BB0_2

jmp $BB0_1 $BB0_1: ld $2, 32($sp) addiu $2, $2, 1 st $2, 32($sp) $BB0_2: ld $2, 28($sp)

# %if.end

If cpu0 do jne compare in execution stage, then we should set PC=PC+12, offset of ($BB0_2, jn e $BB02) 8, in this example. Cpu0 is for teaching purpose and didnt consider the performance with design. In reality, the conditional branch is important in performance of CPU design. According bench mark information, every 7 instructions will meet 1 branch instruction in average. Cpu0 take 2 instructions for conditional branch, (jne(cmp...)), while Mips use one instruction (bne). Finally we list the code added for full support of control ow statement,
// Cpu0MCCodeEmitter.cpp /// getBranchTargetOpValue - Return binary encoding of the branch /// target operand. If the machine operand requires relocation, /// record the relocation and return zero. unsigned Cpu0MCCodeEmitter:: getBranchTargetOpValue(const MCInst &MI, unsigned OpNo, SmallVectorImpl<MCFixup> &Fixups) const { const MCOperand &MO = MI.getOperand(OpNo); assert(MO.isExpr() && "getBranchTargetOpValue expects only expressions"); const MCExpr *Expr = MO.getExpr(); Fixups.push_back(MCFixup::Create(0, Expr,

8.1. Control ow statement

115

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

MCFixupKind(Cpu0::fixup_Cpu0_PC24))); return 0; } // Cpu0MCInstLower.cpp MCOperand Cpu0MCInstLower::LowerSymbolOperand(const MachineOperand &MO, MachineOperandType MOTy, unsigned Offset) const { ... switch(MO.getTargetFlags()) { default: llvm_unreachable("Invalid target flag!"); case Cpu0II::MO_NO_FLAG: Kind = MCSymbolRefExpr::VK_None; break; ... } ... switch (MOTy) { case MachineOperand::MO_MachineBasicBlock: Symbol = MO.getMBB()->getSymbol(); break; ... } MCOperand Cpu0MCInstLower::LowerOperand(const MachineOperand& MO, unsigned offset) const { MachineOperandType MOTy = MO.getType(); switch (MOTy) { default: llvm_unreachable("unknown operand type"); case MachineOperand::MO_Register: ... case MachineOperand::MO_MachineBasicBlock: case MachineOperand::MO_GlobalAddress: case MachineOperand::MO_BlockAddress: ... } ... } // Cpu0ISelLowering.cpp Cpu0TargetLowering:: Cpu0TargetLowering(Cpu0TargetMachine &TM) : TargetLowering(TM, new Cpu0TargetObjectFile()), Subtarget(&TM.getSubtarget<Cpu0Subtarget>()) { ... // Used by legalize types to correctly generate the setcc result. // Without this, every float setcc comes with a AND/OR with the result, // we dont want this, since the fpcmp result goes to a flag register, // which is used implicitly by brcond and select operations. AddPromotedToType(ISD::SETCC, MVT::i1, MVT::i32); ... setOperationAction(ISD::BRCOND, MVT::Other, Custom); // Operations not directly supported by Cpu0. setOperationAction(ISD::BR_CC, MVT::Other, Expand); ... } // Cpu0InstrFormats.td

116

Chapter 8. Control ow statements

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

class BranchBase<bits<8> op, dag outs, dag ins, string asmstr, list<dag> pattern, InstrItinClass itin>: Cpu0Inst<outs, ins, asmstr, pattern, itin, FrmL> { bits<24> imm24; let Opcode = op; let Inst{23-0} } // Cpu0InstrInfo.td // Instruction operand types def brtarget : Operand<OtherVT> { let EncoderMethod = "getBranchTargetOpValue"; let OperandType = "OPERAND_PCREL"; let DecoderMethod = "DecodeBranchTarget"; } ... /// Conditional Branch class CBranch<bits<8> op, string instr_asm, RegisterClass RC>: BranchBase<op, (outs), (ins RC:$cond, brtarget:$imm24), !strconcat(instr_asm, "\t$imm24"), [], IIBranch> { let isBranch = 1; let isTerminator = 1; let hasDelaySlot = 0; } // Unconditional branch class UncondBranch<bits<8> op, string instr_asm>: BranchBase<op, (outs), (ins brtarget:$imm24), !strconcat(instr_asm, "\t$imm24"), [(br bb:$imm24)], IIBranch> { let isBranch = 1; let isTerminator = 1; let isBarrier = 1; let hasDelaySlot = 0; } ... /// Jump and Branch Instructions def JEQ : CBranch<0x20, "jeq", CPURegs>; def JNE : CBranch<0x21, "jne", CPURegs>; def JLT : CBranch<0x22, "jlt", CPURegs>; def JGT : CBranch<0x23, "jgt", CPURegs>; def JLE : CBranch<0x24, "jle", CPURegs>; def JGE : CBranch<0x25, "jge", CPURegs>; def JMP : UncondBranch<0x26, "jmp">; ... // brcond patterns multiclass BrcondPats<RegisterClass RC, Instruction JEQOp, Instruction JNEOp, Instruction JLTOp, Instruction JGTOp, Instruction JLEOp, Instruction JGEOp, Instruction CMPOp, Register ZEROReg> { def : Pat<(brcond (i32 (seteq RC:$lhs, RC:$rhs)), bb:$dst), (JEQOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setueq RC:$lhs, RC:$rhs)), bb:$dst), (JEQOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setne RC:$lhs, RC:$rhs)), bb:$dst), = imm24;

8.1. Control ow statement

117

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

(JNEOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setune RC:$lhs, RC:$rhs)), bb:$dst), (JNEOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setlt RC:$lhs, RC:$rhs)), bb:$dst), (JLTOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setult RC:$lhs, RC:$rhs)), bb:$dst), (JLTOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setgt RC:$lhs, RC:$rhs)), bb:$dst), (JGTOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setugt RC:$lhs, RC:$rhs)), bb:$dst), (JGTOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setle RC:$lhs, RC:$rhs)), bb:$dst), (JLEOp (CMPOp RC:$rhs, RC:$lhs), bb:$dst)>; def : Pat<(brcond (i32 (setule RC:$lhs, RC:$rhs)), bb:$dst), (JLEOp (CMPOp RC:$rhs, RC:$lhs), bb:$dst)>; def : Pat<(brcond (i32 (setge RC:$lhs, RC:$rhs)), bb:$dst), (JGEOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond (i32 (setuge RC:$lhs, RC:$rhs)), bb:$dst), (JGEOp (CMPOp RC:$lhs, RC:$rhs), bb:$dst)>; def : Pat<(brcond RC:$cond, bb:$dst), (JNEOp (CMPOp RC:$cond, ZEROReg), bb:$dst)>; } defm : BrcondPats<CPURegs, JEQ, JNE, JLT, JGT, JLE, JGE, CMP, ZERO>;

The ch7_1_2.cpp is for nest if test. The ch7_1_3.cpp is the for loop as well as while loop, continue, break, goto test. You can run with them if you like to test more. Finally, 7/1/Cpu0 support the local array denition by add the LowerCall() empty function in Cpu0ISelLowering.cpp as follows,
// Cpu0ISelLowering.cpp SDValue Cpu0TargetLowering::LowerCall(SDValue InChain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg, bool doesNotRet, bool &isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, const SmallVectorImpl<ISD::InputArg> &Ins, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { return InChain; }

With this LowerCall(), it can translate ch7_1_4.cpp, ch7_1_4.bc to ch7_1_4.cpu0.s as follows,

// ch7_1_4.cpp int main() { int a[3]={0, 1, 2}; return 0; } ; ModuleID = ch7_1_4 .bc target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128-n8:16:32-S128" target triple = "i386-apple-macosx10.8.0"

118

Chapter 8. Control ow statements

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

@_ZZ4mainE1a = private unnamed_addr constant [3 x i32] [i32 0, i32 1, i32 2], align 4 define i32 @main() nounwind ssp { entry: %retval = alloca i32, align 4 %a = alloca [3 x i32], align 4 store i32 0, i32* %retval %0 = bitcast [3 x i32]* %a to i8* call void @llvm.memcpy.p0i8.p0i8.i32(i8* %0, i8* bitcast ([3 x i32]* @_ZZ4mainE1a to i8*), i32 12, i32 4, i1 false) ret i32 0 } 118-165-79-206:InputFiles Jonathan$ cat ch7_1_4.cpu0.s .section .mdebug.abi32 .previous .file "ch7_1_4.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .frame $sp,24,$lr .mask 0x00000000,0 .set noreorder .cpload $t9 .set nomacro # BB#0: # %entry addiu $sp, $sp, -24 ld $2, %got(__stack_chk_guard)($gp) ld $3, 0($2) st $3, 20($sp) addiu $3, $zero, 0 st $3, 16($sp) ld $3, %got($_ZZ4mainE1a)($gp) addiu $3, $3, %lo($_ZZ4mainE1a) ld $4, 8($3) st $4, 12($sp) ld $4, 4($3) st $4, 8($sp) ld $3, 0($3) st $3, 4($sp) ld $2, 0($2) ld $3, 20($sp) cmp $2, $3 jne $BB0_2 jmp $BB0_1 $BB0_1: # %SP_return addiu $sp, $sp, 24 ret $lr $BB0_2: # %CallStackCheckFailBlk .set macro .set reorder .end main $tmp1: .size main, ($tmp1)-main

8.1. Control ow statement

119

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.type $_ZZ4mainE1a,@object # .section .rodata,"a",@progbits .align 2 $_ZZ4mainE1a: .4byte 0 # .4byte 1 # .4byte 2 # .size $_ZZ4mainE1a, 12

@_ZZ4mainE1a

0x0 0x1 0x2

8.2 RISC CPU knowledge

As mentioned in the previous section, cpu0 is a RISC (Reduced Instruction Set Computer) CPU with 3 stages of pipeline. RISC CPU is full in world. Even the X86 of CISC (Complex Instruction Set Computer) is RISC inside. (It translate CISC instruction into micro-instruction which do pipeline as RISC). Knowledge with RISC will make you satised in compiler design. List these two excellent books we have read which include the real RISC CPU knowledge needed for reference. Sure, there are many books in Computer Architecture, and some of them contain real RISC CPU knowledge needed, but these two are what we read. Computer Organization and Design: The Hardware/Software Interface (The Morgan Kaufmann Series in Computer Architecture and Design) Computer Architecture: A Quantitative Approach (The Morgan Kaufmann Series in Computer Architecture and Design) The book of Computer Organization and Design: The Hardware/Software Interface (there are 4 editions until the book is written) is for the introduction (simple). Computer Architecture: A Quantitative Approach (there are 5 editions until the book is written) is more complicate and deep in CPU architecture.

120

Chapter 8. Control ow statements

CHAPTER

NINE

FUNCTION CALL
The subroutine/function call of backend code translation is supported in this chapter. A lots of code needed in function call. We break it down according llvm supplied interface for easy to explanation. This chapter start from introducing the Mips stack frame structure since we borrow many part of ABI from it. Although each CPU has its own ABI, most of RISC CPUs ABI are similar. In addition to support xed number of arguments function call, cpu0 also upport variable number of arguments since C/C++ support this feature. Supply Mips ABI and assemble language manual on internet link in this chapter for your reference. The section 4.5 DAG Lowering of tricore_llvm.pdf contains some knowledge about Lowering process. Section 4.5.1 Calling Conventions of tricore_llvm.pdf is the related materials you can reference. This chapter is more complicate than any of the previous chapter. It include stack frame and the related ABI support. If you have problem in reading the stack frame illustrated in the rst three sections of this chapter, you can read the appendix B of Procedure Call Convention of book Computer Organization and Design which listed in section RISC CPU knowledge of chapter Control ow statement 1 , Run Time Memory of compiler book, or Function Call Sequence and Stack Frame of Mips ABI.

9.1 Mips stack frame

The rst thing for design the cpu0 function call is deciding how to pass arguments in function call. There are two options. The rst is pass arguments all in stack. Second is pass arguments in the registers which are reserved for function arguments, and put the other arguments in stack if it over the number of registers reserved for function call. For example, Mips pass the rst 4 arguments in register $a0, $a1, $a2, $a3, and the other arguments in stack if it over 4 arguments. Mips stack frame is the Mips stack frame.

Figure 9.1: Mips stack frame Run llc -march=mips for ch8_1.bc, you will get the following result.
1

http://jonathan2251.github.com/lbd/ctrlow.html#risc-cpu-knowledge

121

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// ch8_1.cpp int sum_i(int x1, int x2, int x3, int x4, int x5, int x6) { int sum = x1 + x2 + x3 + x4 + x5 + x6; return sum; } int main() { int a = sum_i(1, 2, 3, 4, 5, 6); return a; } 118-165-79-31:InputFiles Jonathan$ clang -c ch8_1.cpp -emit-llvm -o ch8_1.bc 118-165-79-31:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=mips -relocation-model=pic -filetype=asm ch8_1.bc -o ch8_1.mips.s 118-165-79-31:InputFiles Jonathan$ cat ch8_1.mips.s .section .mdebug.abi32 .previous .file "ch8_1.bc" .text .globl _Z5sum_iiiiiii .align 2 .type _Z5sum_iiiiiii,@function .ent _Z5sum_iiiiiii # @_Z5sum_iiiiiii _Z5sum_iiiiiii: .frame $sp,32,$ra .mask 0x00000000,0 .fmask 0x00000000,0 .set noreorder .set nomacro # BB#0: # %entry addiu $sp, $sp, -32 st $4, 28($sp) st $5, 24($sp) st $6, 20($sp) st $7, 16($sp) ld $2, 48($sp) // load argument 5 st $2, 12($sp) ld $2, 52($sp) // load argument 6 st $2, 8($sp) ld $3, 24($sp) ld $4, 28($sp) addu $3, $4, $3 ld $4, 20($sp) addu $3, $3, $4 ld $4, 16($sp) addu $3, $3, $4 ld $4, 12($sp) addu $3, $3, $4 addu $2, $3, $2 st $2, 4($sp) addiu $sp, $sp, 32 jr $ra nop

122

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.set .set .end $tmp1: .size

macro reorder _Z5sum_iiiiiii _Z5sum_iiiiiii, ($tmp1)-_Z5sum_iiiiiii

.globl main .align 2 .type main,@function .ent main # @main main: .frame $sp,48,$ra .mask 0x80000000,-4 .fmask 0x00000000,0 .set noreorder .cpload $25 .set nomacro # BB#0: # %entry addiu $sp, $sp, -48 st $ra, 44($sp) # 4-byte Folded Spill .cprestore 24 st $zero, 40($sp) addiu $2, $zero, 6 st $2, 20($sp) // Save argument 6 to 20($sp) addiu $2, $zero, 5 st $2, 16($sp) // Save argument 5 to 16($sp) ld $25, %call16(_Z5sum_iiiiiii)($gp) addiu $4, $zero, 1 // Pass argument 1 to $4 (=$a0) addiu $5, $zero, 2 // Pass argument 2 to $5 (=$a1) addiu $6, $zero, 3 addiu $7, $zero, 4 jalr $25 nop ld $gp, 24($sp) st $2, 36($sp) ld $ra, 44($sp) # 4-byte Folded Reload addiu $sp, $sp, 48 jr $ra nop .set macro .set reorder .end main $tmp4: .size main, ($tmp4)-main

From the mips assembly code generated as above, we know it save the rst 4 arguments to $a0..$a3 and last 2 arguments to 16($sp) and 20($sp). Mips arguments location in stack frame is the arguments location for example code ch8_1.cpp. It load argument 5 from 48($sp) in sum_i() since the argument 5 is saved to 16($sp) in main(). The stack size of sum_i() is 32, so 16+32($sp) is the location of incoming argument 5. The 007-2418-003.pdf in 2 is the Mips assembly language manual. include the Mips stack frame.
2 3

is Mips Application Binary Interface which

https://www.dropbox.com/sh/2pkh1fewlq2zag9/OHnrYn2nOs/doc/MIPSproAssemblyLanguageProgrammerGuide http://www.linux-mips.org/pub/linux/mips/doc/ABI/mipsabi.pdf

9.1. Mips stack frame

123

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 9.2: Mips arguments location in stack frame

124

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

9.2 Load incoming arguments from stack frame

From last section, to support function call, we need implementing the arguments pass mechanism with stack frame. Before do that, lets run the old version of code 7/1/Cpu0 with ch8_1.cpp and see what happen.
118-165-79-31:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch8_1.bc -o ch8_1.cpu0.s Assertion failed: (InVals.size() == Ins.size() && "LowerFormalArguments didnt emit the correct number of values!"), function LowerArguments, file /Users/ Jonathan/llvm/3.1.test/cpu0/1/src/lib/CodeGen/SelectionDAG/ SelectionDAGBuilder.cpp, line 6671. Stack dump: 0. Program arguments: /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch8_1.bc -o ch8_1.cpu0.s 1. Running pass Function Pass Manager on module ch8_1.bc. 2. Running pass CPU0 DAG->DAG Pattern Instruction Selection on function @_Z5sum_iiiiiii Illegal instruction: 4

Since 7/1/Cpu0 dene the LowerFormalArguments() with empty, we get the error message as above. Before dene LowerFormalArguments(), we have to choose how to pass arguments in function call. We choose pass arguments all in stack frame. We dont reserve any dedicated register for arguments passing since cpu0 has only 16 registers while Mips has 32 registers. Cpu0CallingConv.td is dened for cpu0 passing rule as follows,
// Cpu0CallingConv.td ... def RetCC_Cpu0EABI : CallingConv<[ // i32 are returned in registers V0, V1 CCIfType<[i32], CCAssignToReg<[V0, V1]>> ]>; //===----------------------------------------------------------------------===// // Cpu0 EABI Calling Convention //===----------------------------------------------------------------------===// def CC_Cpu0EABI : CallingConv<[ // Promote i8/i16 arguments to i32. CCIfType<[i8, i16], CCPromoteToType<i32>>, // Integer values get stored in stack slots that are 4 bytes in // size and 4-byte aligned. CCIfType<[i32], CCAssignToStack<4, 4>> ]>;

//===----------------------------------------------------------------------===// // Cpu0 Calling Convention Dispatch //===----------------------------------------------------------------------===// def CC_Cpu0 : CallingConv<[ CCDelegateTo<CC_Cpu0EABI> ]>;

def RetCC_Cpu0 : CallingConv<[ CCDelegateTo<RetCC_Cpu0EABI> ]>;

9.2. Load incoming arguments from stack frame

125

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

def CSR_O32 : CalleeSavedRegs<(add LR, FP, (sequence "S%u", 2, 0))>;

As above, CC_Cpu0 is the cpu0 Calling Convention which delegate to CC_Cpu0EABI and dene the CC_Cpu0EABI. The reason we dont dene the Calling Convention directly in CC_Cpu0 is that a real general CPU like Mips can have several Calling Convention. Combine with the mechanism of section Target Registration 4 which llvm supplied, we can use different Calling Convention in different target. Although cpu0 only have a Calling Convention right now, dene with a dedicate Call Convention name (CC_Cpu0EABI in this example) is a better solution for system expand, and naming your Calling Convention. CC_Cpu0EABI as above, say it pass arguments in stack frame. Function LowerFormalArguments() charge function incoming arguments creation. We dene it as follows,
// Cpu0ISelLowering.cpp ... /// LowerFormalArguments - transform physical registers into virtual registers /// and generate load operations for arguments places on the stack. SDValue Cpu0TargetLowering::LowerFormalArguments(SDValue Chain, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { MachineFunction &MF = DAG.getMachineFunction(); MachineFrameInfo *MFI = MF.getFrameInfo(); Cpu0FunctionInfo *Cpu0FI = MF.getInfo<Cpu0FunctionInfo>(); Cpu0FI->setVarArgsFrameIndex(0); // Used with vargs to acumulate store chains. std::vector<SDValue> OutChains; // Assign locations to all of the incoming arguments. SmallVector<CCValAssign, 16> ArgLocs; CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), getTargetMachine(), ArgLocs, *DAG.getContext()); CCInfo.AnalyzeFormalArguments(Ins, CC_Cpu0); Function::const_arg_iterator FuncArg = DAG.getMachineFunction().getFunction()->arg_begin(); int LastFI = 0;// Cpu0FI->LastInArgFI is 0 at the entry of this function. for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i, ++FuncArg) { CCValAssign &VA = ArgLocs[i]; EVT ValVT = VA.getValVT(); ISD::ArgFlagsTy Flags = Ins[i].Flags; bool IsRegLoc = VA.isRegLoc(); if (Flags.isByVal()) { assert(Flags.getByValSize() && "ByVal args of size 0 should have been ignored by front-end."); continue; } // sanity check
4

http://jonathan2251.github.com/lbd/llvmstructure.html#target-registration

126

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

assert(VA.isMemLoc()); // The stack pointer offset is relative to the caller stack frame. LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8, VA.getLocMemOffset(), true); // Create load nodes to retrieve arguments from the stack SDValue FIN = DAG.getFrameIndex(LastFI, getPointerTy()); InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN, MachinePointerInfo::getFixedStack(LastFI), false, false, false, 0)); } Cpu0FI->setLastInArgFI(LastFI); // All stores are grouped in one node to allow the matching between // the size of Ins and InVals. This only happens when on varg functions if (!OutChains.empty()) { OutChains.push_back(Chain); Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &OutChains[0], OutChains.size()); } return Chain; }

Refresh section Global variable 5 , we handled global variable translation by create the IR DAG in LowerGlobalAddress() rst, and then do the Instruction Selection by their corresponding machine instruction DAG in Cpu0InstrInfo.td. LowerGlobalAddress() is called when llc meet the global variable access. LowerFormalArguments() work with the same way. It is called when function is entered. It get incoming arguments information by CCInfo(CallConv,..., ArgLocs, . . . ) before enter for loop. In ch8_1.cpp, there are 6 arguments in sum_i(...) function call and we use the stack frame only for arguments passing without any arguments pass in registers. So ArgLocs.size() is 6, each argument information is in ArgLocs[i] and ArgLocs[i].isMemLoc() is true. In for loop, it create each frame index object by LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,VA.getLocMemOffset(), true) and FIN = DAG.getFrameIndex(LastFI, getPointerTy()). And then create IR DAG load node and put the load node into vector InVals by InVals.push_back(DAG.getLoad(ValVT, dl, Chain, FIN, MachinePointerInfo::getFixedStack(LastFI), false, false, false, 0)). Cpu0FI->setVarArgsFrameIndex(0) and Cpu0FI->setLastInArgFI(LastFI) are called when before and after above work. In ch8_1.cpp example, LowerFormalArguments() will be called twice. First time is for sum_i() which will create 6 load DAG for 6 incoming arguments passing into this function. Second time is for main() which didnt create any load DAG for no incoming argument passing into main(). In addition to LowerFormalArguments() which create the load DAG, we need to dene the loadRegFromStackSlot() to issue the machine instruction ld $r, offset($sp) to load incoming arguments from stack frame offset. GetMemOperand(..., FI, ...) return the Memory location of the frame index variable, which is the offset.
// Cpu0InstrInfo.cpp ... static MachineMemOperand* GetMemOperand(MachineBasicBlock &MBB, int FI, unsigned Flag) { MachineFunction &MF = *MBB.getParent(); MachineFrameInfo &MFI = *MF.getFrameInfo(); unsigned Align = MFI.getObjectAlignment(FI); return MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(FI), Flag, MFI.getObjectSize(FI), Align); } void Cpu0InstrInfo:: loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
5

http://jonathan2251.github.com/lbd/globalvar.html#global-variable

9.2. Load incoming arguments from stack frame

127

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

unsigned DestReg, int FI, const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const { DebugLoc DL; if (I != MBB.end()) DL = I->getDebugLoc(); MachineMemOperand *MMO = GetMemOperand(MBB, FI, MachineMemOperand::MOLoad); unsigned Opc = 0; if (RC == Cpu0::CPURegsRegisterClass) Opc = Cpu0::LD; assert(Opc && "Register class not handled!"); BuildMI(MBB, I, DL, get(Opc), DestReg).addFrameIndex(FI).addImm(0) .addMemOperand(MMO); }

In addition to Calling Convention and LowerFormalArguments(), 8/2/Cpu0 add the following code for cpu0 instructions swi (Software Interrupt), jsub and jalr (function call) denition and printing.
// Cpu0InstrFormats.td ... // Cpu0 Pseudo Instructions Format class Cpu0Pseudo<dag outs, dag ins, string asmstr, list<dag> pattern>: Cpu0Inst<outs, ins, asmstr, pattern, IIPseudo, Pseudo> { let isCodeGenOnly = 1; let isPseudo = 1; } // Cpu0InstrInfo.td ... def SDT_Cpu0JmpLink : SDTypeProfile<0, 1, [SDTCisVT<0, iPTR>]>; ... // Call def Cpu0JmpLink : SDNode<"Cpu0ISD::JmpLink",SDT_Cpu0JmpLink, [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue, SDNPVariadic]>; ... def jmptarget : Operand<OtherVT> { let EncoderMethod = "getJumpTargetOpValue"; } ... def calltarget : Operand<iPTR> { let EncoderMethod = "getJumpTargetOpValue"; } ... // Jump and Link (Call) let isCall=1, hasDelaySlot=0 in { class JumpLink<bits<8> op, string instr_asm>: FJ<op, (outs), (ins calltarget:$target, variable_ops), !strconcat(instr_asm, "\t$target"), [(Cpu0JmpLink imm:$target)], IIBranch> { let DecoderMethod = "DecodeJumpTarget"; } class JumpLinkReg<bits<8> op, string instr_asm, RegisterClass RC>: FA<op, (outs), (ins RC:$rb, variable_ops), !strconcat(instr_asm, "\t$rb"), [(Cpu0JmpLink RC:$rb)], IIBranch> {

128

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

let rc = 0; let ra = 14; let shamt = 0; } } ... /// def def ... def ... def ... // Cpu0InstPrinter.cpp ... static void printExpr(const MCExpr *Expr, raw_ostream &OS) { switch (Kind) { ... case MCSymbolRefExpr::VK_Cpu0_GOT_CALL: OS << "%call24("; break; ... } ... } // Cpu0MCCodeEmitter.cpp ... unsigned Cpu0MCCodeEmitter:: getMachineOpValue(const MCInst &MI, const MCOperand &MO, SmallVectorImpl<MCFixup> &Fixups) const { ... switch(cast<MCSymbolRefExpr>(Expr)->getKind()) { ... case MCSymbolRefExpr::VK_Cpu0_GOT_CALL: FixupKind = Cpu0::fixup_Cpu0_CALL24; break; ... } ... } // Cpu0MachineFucntion.h class Cpu0FunctionInfo : public MachineFunctionInfo { ... /// VarArgsFrameIndex - FrameIndex for start of varargs area. int VarArgsFrameIndex; // Range of frame object indices. // InArgFIRange: Range of indices of all frame objects created during call to // LowerFormalArguments. // OutArgFIRange: Range of indices of all frame objects created during call to // LowerCall except for the frame object for restoring $gp. std::pair<int, int> InArgFIRange, OutArgFIRange; int GPFI; // Index of the frame object for restoring $gp mutable int DynAllocFI; // Frame index of dynamically allocated stack area. unsigned MaxCallFrameSize;

Jump and Branch Instructions SWI : JumpLink<0x2A, "swi">; JSUB : JumpLink<0x2B, "jsub">; JALR : JumpLinkReg<0x2D, "jalr", CPURegs>; : Pat<(Cpu0JmpLink (i32 tglobaladdr:$dst)), (JSUB tglobaladdr:$dst)>;

9.2. Load incoming arguments from stack frame

129

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

public: Cpu0FunctionInfo(MachineFunction& MF) : MF(MF), GlobalBaseReg(0), VarArgsFrameIndex(0), InArgFIRange(std::make_pair(-1, 0)), OutArgFIRange(std::make_pair(-1, 0)), GPFI(0), DynAllocFI(0), MaxCallFrameSize(0) {} bool isInArgFI(int FI) const { return FI <= InArgFIRange.first && FI >= InArgFIRange.second; } void setLastInArgFI(int FI) { InArgFIRange.second = FI; } void extendOutArgFIRange(int FirstFI, int LastFI) { if (!OutArgFIRange.second) // this must be the first time this function was called. OutArgFIRange.first = FirstFI; OutArgFIRange.second = LastFI; } int getGPFI() const { return GPFI; } void setGPFI(int FI) { GPFI = FI; } bool needGPSaveRestore() const { return getGPFI(); } bool isGPFI(int FI) const { return GPFI && GPFI == FI; } // The first call to this function creates a frame object for dynamically // allocated stack area. int getDynAllocFI() const { if (!DynAllocFI) DynAllocFI = MF.getFrameInfo()->CreateFixedObject(4, 0, true); return DynAllocFI; } bool isDynAllocFI(int FI) const { return DynAllocFI && DynAllocFI == FI; } ... int getVarArgsFrameIndex() const { return VarArgsFrameIndex; } void setVarArgsFrameIndex(int Index) { VarArgsFrameIndex = Index; } unsigned getMaxCallFrameSize() const { return MaxCallFrameSize; } void setMaxCallFrameSize(unsigned S) { MaxCallFrameSize = S; } };

After above changes, you can run 8/2/Cpu0 with ch8_1.cpp and see what happens in the following,
118-165-79-83:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch8_1.bc -o ch8_1.cpu0.s Assertion failed: ((isTailCall || InVals.size() == Ins.size()) && "LowerCall didnt emit the correct number of values!"), function LowerCallTo, file /Users/Jonathan/llvm/3.1.test/cpu0/1/src/lib/CodeGen/SelectionDAG/ SelectionDAGBuilder.cpp, line 6482. Stack dump: 0. Program arguments: /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch8_1.bc -o ch8_1.cpu0.s 1. Running pass Function Pass Manager on module ch8_1.bc. 2. Running pass CPU0 DAG->DAG Pattern Instruction Selection on function @main

130

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Illegal instruction: 4

9.3 Store outgoing arguments to stack frame

Mips arguments location in stack frame depicted two steps to take care arguments passing. One is store outgoing arguments in caller function, and the other is load incoming arguments in callee function. We dened LowerFormalArguments() for load incoming arguments in callee function last section. Now, we will nish store outgoing arguments in caller function. LowerCall() is responsible to do this. The implementation as follows,
// Cpu0ISelLowering.cpp ... SDValue Cpu0TargetLowering::LowerCall(SDValue InChain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg, bool doesNotRet, bool &isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, const SmallVectorImpl<ISD::InputArg> &Ins, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { // Cpu0 target does not yet support tail call optimization. isTailCall = false; MachineFunction &MF = DAG.getMachineFunction(); MachineFrameInfo *MFI = MF.getFrameInfo(); const TargetFrameLowering *TFL = MF.getTarget().getFrameLowering(); bool IsPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_; Cpu0FunctionInfo *Cpu0FI = MF.getInfo<Cpu0FunctionInfo>(); // Analyze operands of the call, assigning locations to each operand. SmallVector<CCValAssign, 16> ArgLocs; CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), getTargetMachine(), ArgLocs, *DAG.getContext()); CCInfo.AnalyzeCallOperands(Outs, CC_Cpu0); // Get a count of how many bytes are to be pushed on the stack. unsigned NextStackOffset = CCInfo.getNextStackOffset(); // If this is the first call, create a stack frame object that points to // a location to which .cprestore saves $gp. if (IsPIC && Cpu0FI->globalBaseRegFixed() && !Cpu0FI->getGPFI()) Cpu0FI->setGPFI(MFI->CreateFixedObject(4, 0, true)); // Get the frame index of the stack frame object that points to the location // of dynamically allocated area on the stack. int DynAllocFI = Cpu0FI->getDynAllocFI(); unsigned MaxCallFrameSize = Cpu0FI->getMaxCallFrameSize(); if (MaxCallFrameSize < NextStackOffset) { Cpu0FI->setMaxCallFrameSize(NextStackOffset); // Set the offsets relative to $sp of the $gp restore slot and dynamically // allocated stack space. These offsets must be aligned to a boundary // determined by the stack alignment of the ABI. unsigned StackAlignment = TFL->getStackAlignment(); NextStackOffset = (NextStackOffset + StackAlignment - 1) /

9.3. Store outgoing arguments to stack frame

131

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

StackAlignment * StackAlignment; MFI->setObjectOffset(DynAllocFI, NextStackOffset); } // Chain is the output chain of the last Load/Store or CopyToReg node. // ByValChain is the output chain of the last Memcpy node created for copying // byval arguments to the stack. SDValue Chain, CallSeqStart, ByValChain; SDValue NextStackOffsetVal = DAG.getIntPtrConstant(NextStackOffset, true); Chain = CallSeqStart = DAG.getCALLSEQ_START(InChain, NextStackOffsetVal); ByValChain = InChain; // With EABI is it possible to have 16 args on registers. SmallVector<std::pair<unsigned, SDValue>, 16> RegsToPass; SmallVector<SDValue, 8> MemOpChains; int FirstFI = -MFI->getNumFixedObjects() - 1, LastFI = 0; // Walk the register/memloc assignments, inserting copies/loads. for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { SDValue Arg = OutVals[i]; CCValAssign &VA = ArgLocs[i]; MVT ValVT = VA.getValVT(), LocVT = VA.getLocVT(); ISD::ArgFlagsTy Flags = Outs[i].Flags; // ByVal Arg. if (Flags.isByVal()) { assert("!!!Error!!!, Flags.isByVal()==true"); assert(Flags.getByValSize() && "ByVal args of size 0 should have been ignored by front-end."); continue; } // Register cant get to this point... assert(VA.isMemLoc()); // Create the frame index object for this incoming parameter LastFI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8, VA.getLocMemOffset(), true); SDValue PtrOff = DAG.getFrameIndex(LastFI, getPointerTy()); // emit ISD::STORE whichs stores the // parameter value to a stack Location MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff, MachinePointerInfo(), false, false, 0)); } // // // if Extend range of indices of frame objects for outgoing arguments that were created during this function call. Skip this step if no such objects were created. (LastFI) Cpu0FI->extendOutArgFIRange(FirstFI, LastFI);

// If a memcpy has been created to copy a byval arg to a stack, replace the // chain input of CallSeqStart with ByValChain. if (InChain != ByValChain) DAG.UpdateNodeOperands(CallSeqStart.getNode(), ByValChain, NextStackOffsetVal);

132

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Transform all store nodes into one single node because all store // nodes are independent of each other. if (!MemOpChains.empty()) Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &MemOpChains[0], MemOpChains.size()); // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol // node so that legalize doesnt hack it. unsigned char OpFlag; bool IsPICCall = IsPIC; // true if calls are translated to jalr $25 bool GlobalOrExternal = false; SDValue CalleeLo; if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { OpFlag = IsPICCall ? Cpu0II::MO_GOT_CALL : Cpu0II::MO_NO_FLAG; Callee = DAG.getTargetGlobalAddress(G->getGlobal(), dl, getPointerTy(), 0, OpFlag); GlobalOrExternal = true; } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { if (!IsPIC) // static OpFlag = Cpu0II::MO_NO_FLAG; else // O32 & PIC OpFlag = Cpu0II::MO_GOT_CALL; Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(), OpFlag); GlobalOrExternal = true; } SDValue InFlag; // Create nodes that load address of callee and copy it to T9 if (IsPICCall) { if (GlobalOrExternal) { // Load callee address Callee = DAG.getNode(Cpu0ISD::Wrapper, dl, getPointerTy(), GetGlobalReg(DAG, getPointerTy()), Callee); SDValue LoadValue = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee, MachinePointerInfo::getGOT(), false, false, false, 0); // Use GOT+LO if callee has internal linkage. if (CalleeLo.getNode()) { SDValue Lo = DAG.getNode(Cpu0ISD::Lo, dl, getPointerTy(), CalleeLo); Callee = DAG.getNode(ISD::ADD, dl, getPointerTy(), LoadValue, Lo); } else Callee = LoadValue; } } // T9 should contain the address of the callee function if // -reloction-model=pic or it is an indirect call. if (IsPICCall || !GlobalOrExternal) { // copy to T9 unsigned T9Reg = Cpu0::T9; Chain = DAG.getCopyToReg(Chain, dl, T9Reg, Callee, SDValue(0, 0)); InFlag = Chain.getValue(1);

9.3. Store outgoing arguments to stack frame

133

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Callee = DAG.getRegister(T9Reg, getPointerTy()); } // Cpu0JmpLink = #chain, #target_address, #opt_in_flags... // = Chain, Callee, Reg#1, Reg#2, ... // // Returns a chain & a flag for retval copy to use. SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); SmallVector<SDValue, 8> Ops; Ops.push_back(Chain); Ops.push_back(Callee); // Add argument registers to the end of the list so that they are // known live into the call. for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) Ops.push_back(DAG.getRegister(RegsToPass[i].first, RegsToPass[i].second.getValueType())); // Add a register mask operand representing the call-preserved registers. const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); const uint32_t *Mask = TRI->getCallPreservedMask(CallConv); assert(Mask && "Missing call preserved mask for calling convention"); Ops.push_back(DAG.getRegisterMask(Mask)); if (InFlag.getNode()) Ops.push_back(InFlag); Chain = DAG.getNode(Cpu0ISD::JmpLink, dl, NodeTys, &Ops[0], Ops.size()); InFlag = Chain.getValue(1); // Create the CALLSEQ_END node. Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NextStackOffset, true), DAG.getIntPtrConstant(0, true), InFlag); InFlag = Chain.getValue(1); // Handle result values, copying them out of physregs into vregs that we // return. return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG, InVals); } /// LowerCallResult - Lower the result values of a call into the /// appropriate copies out of appropriate physical registers. SDValue Cpu0TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg, const SmallVectorImpl<ISD::InputArg> &Ins, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { // Assign locations to each value returned by this call. SmallVector<CCValAssign, 16> RVLocs; CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), getTargetMachine(), RVLocs, *DAG.getContext()); CCInfo.AnalyzeCallResult(Ins, RetCC_Cpu0); // Copy all of the result registers out of their specified physreg.

134

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

for (unsigned i = 0; i != RVLocs.size(); ++i) { Chain = DAG.getCopyFromReg(Chain, dl, RVLocs[i].getLocReg(), RVLocs[i].getValVT(), InFlag).getValue(1); InFlag = Chain.getValue(2); InVals.push_back(Chain.getValue(0)); } return Chain; }

Just like load incoming arguments from stack frame, we call CCInfo(CallConv,..., ArgLocs, . . . ) to get outgoing arguments information before enter for loop and set stack alignment with 8 bytes. Theyre almost same in for loop with LowerFormalArguments(), except LowerCall() create store DAG vector instead of load DAG vector. After the for loop, it create ld $6, %call24(_Z5sum_iiiiiii)($gp) and jalr $6 for calling subroutine (the $6 is $t9) in PIC mode. DAG.getCALLSEQ_START() and DAG.getCALLSEQ_END() are set before the for loop and after call subroutine, they insert CALLSEQ_START, CALLSEQ_END, and translate into pseudo machine instructions !ADJCALLSTACKDOWN, !ADJCALLSTACKUP later according Cpu0InstrInfo.td denition as follows.
// Cpu0InstrInfo.td ... def SDT_Cpu0CallSeqStart : SDCallSeqStart<[SDTCisVT<0, i32>]>; def SDT_Cpu0CallSeqEnd : SDCallSeqEnd<[SDTCisVT<0, i32>, SDTCisVT<1, i32>]>; ... // These are target-independent nodes, but have target-specific formats. def callseq_start : SDNode<"ISD::CALLSEQ_START", SDT_Cpu0CallSeqStart, [SDNPHasChain, SDNPOutGlue]>; def callseq_end : SDNode<"ISD::CALLSEQ_END", SDT_Cpu0CallSeqEnd, [SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>; //===----------------------------------------------------------------------===// // Pseudo instructions //===----------------------------------------------------------------------===// // As stack alignment is always done with addiu, we need a 16-bit immediate let Defs = [SP], Uses = [SP] in { def ADJCALLSTACKDOWN : Cpu0Pseudo<(outs), (ins uimm16:$amt), "!ADJCALLSTACKDOWN $amt", [(callseq_start timm:$amt)]>; def ADJCALLSTACKUP : Cpu0Pseudo<(outs), (ins uimm16:$amt1, uimm16:$amt2), "!ADJCALLSTACKUP $amt1", [(callseq_end timm:$amt1, timm:$amt2)]>; }

Like load incoming arguments, we need to implement storeRegToStackSlot() for store outgoing arguments to stack frame offset. .. code-block:: c++ // Cpu0InstrInfo.cpp ... //- st SrcReg, MMO(FI) void Cpu0InstrInfo:: storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I, unsigned SrcReg, bool isKill, int FI, const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const { DebugLoc DL;

9.3. Store outgoing arguments to stack frame

135

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

if (I != MBB.end()) DL = I->getDebugLoc(); MachineMemOperand *MMO = GetMemOperand(MBB, FI, MachineMemOperand::MOStore); unsigned Opc = 0; if (RC == Cpu0::CPURegsRegisterClass) Opc = Cpu0::ST; assert(Opc && "Register class not handled!"); BuildMI(MBB, I, DL, get(Opc)).addReg(SrcReg, getKillRegState(isKill)) .addFrameIndex(FI).addImm(0).addMemOperand(MMO); }

Now, lets run 8/3/Cpu0 with ch8_1.cpp to get result as follows,

118-165-79-83:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch8_1.bc -o ch8_1.cpu0.s 118-165-79-83:InputFiles Jonathan$ cat ch8_1.cpu0.s .section .mdebug.abi32 .previous .file "ch8_1.bc" .text .globl _Z5sum_iiiiiii .align 2 .type _Z5sum_iiiiiii,@function .ent _Z5sum_iiiiiii # @_Z5sum_iiiiiii _Z5sum_iiiiiii: .frame $sp,32,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: # %entry addiu $sp, $sp, -32 ld $2, 32($sp) st $2, 28($sp) ld $2, 36($sp) st $2, 24($sp) ld $2, 40($sp) st $2, 20($sp) ld $2, 44($sp) st $2, 16($sp) ld $2, 48($sp) st $2, 12($sp) ld $2, 52($sp) st $2, 8($sp) ld $3, 24($sp) ld $4, 28($sp) add $3, $4, $3 ld $4, 20($sp) add $3, $3, $4 ld $4, 16($sp) add $3, $3, $4 ld $4, 12($sp) add $3, $3, $4 add $2, $3, $2 st $2, 4($sp) addiu $sp, $sp, 32 ret $lr

136

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.set .set .end $tmp1: .size

macro reorder _Z5sum_iiiiiii _Z5sum_iiiiiii, ($tmp1)-_Z5sum_iiiiiii

.globl main .align 2 .type main,@function .ent main # @main main: .frame $sp,40,$lr .mask 0x00004000,-4 .set noreorder .cpload $t9 .set nomacro # BB#0: # %entry addiu $sp, $sp, -40 st $lr, 36($sp) # 4-byte Folded Spill addiu $2, $zero, 0 st $2, 32($sp) !ADJCALLSTACKDOWN 24 $2, $zero, 6 addiu st $2, 60($sp) // wrong offset $2, $zero, 5 addiu st $2, 56($sp) addiu $2, $zero, 4 st $2, 52($sp) addiu $2, $zero, 3 st $2, 48($sp) addiu $2, $zero, 2 st $2, 44($sp) addiu $2, $zero, 1 st $2, 40($sp) ld $6, %call24(_Z5sum_iiiiiii)($gp) jalr $6 !ADJCALLSTACKUP 24 st $2, 28($sp) ld $lr, 36($sp) # 4-byte Folded Reload addiu $sp, $sp, 40 ret $lr .set macro .set reorder .end main $tmp4: .size main, ($tmp4)-main

It store the arguments to wrong offset. We will x this issue and take care !ADJCALLSTACKUP and !ADJCALLSTACKDOWN in next two sections.

9.4 Fix the wrong offset in storing arguments to stack frame

To x the wrong offset in storing arguments, we modify the following code in eliminateFrameIndex() as follows. The bold text as below is added in 8/4/Cpu0 to set the caller outgoing arguments into spOffset($sp) (8/3/Cpu0 set them to pOffset+stackSize($sp).

9.4. Fix the wrong offset in storing arguments to stack frame

137

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Cpu0RegisterInfo.cpp ... void Cpu0RegisterInfo:: eliminateFrameIndex(MachineBasicBlock::iterator II, int SPAdj, RegScavenger *RS) const { ... Cpu0FunctionInfo *Cpu0FI = MF.getInfo<Cpu0FunctionInfo>(); ... if (Cpu0FI->isOutArgFI(FrameIndex) || Cpu0FI->isDynAllocFI(FrameIndex) || (FrameIndex >= MinCSFI && FrameIndex <= MaxCSFI)) FrameReg = Cpu0::SP; else FrameReg = getFrameRegister(MF); ... // Calculate final offset. // - There is no need to change the offset if the frame object is one of the // following: an outgoing argument, pointer to a dynamically allocated // stack space or a $gp restore location, // - If the frame object is any of the following, its offset must be adjusted // by adding the size of the stack: // incoming argument, callee-saved register location or local variable. if (Cpu0FI->isOutArgFI(FrameIndex) || Cpu0FI->isGPFI(FrameIndex) || Cpu0FI->isDynAllocFI(FrameIndex)) Offset = spOffset; else Offset = spOffset + (int64_t)stackSize; Offset += MI.getOperand(i+1).getImm(); DEBUG(errs() << "Offset : " << Offset << "\n" << "<--------->\n");

// If MI is not a debug value, make sure Offset fits in the 16-bit immediate // field. if (!MI.isDebugValue() && !isInt<16>(Offset)) { assert("(!MI.isDebugValue() && !isInt<16>(Offset))"); } MI.getOperand(i).ChangeToRegister(FrameReg, false); MI.getOperand(i+1).ChangeToImmediate(Offset); } // Cpu0MachineFunction.h ... bool isOutArgFI(int FI) const { return FI <= OutArgFIRange.first && FI >= OutArgFIRange.second; }

Run 8/4/Cpu0 with ch8_1.cpp will get the following result. It correct arguements offset im main() from (0+40)$sp, (8+40)$sp, ..., to (0)$sp, (8)$sp, ..., where the stack size is 40 in main().
118-165-76-131:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype= asm ch8_1.bc -o ch8_1.cpu0.s 118-165-76-131:InputFiles Jonathan$ cat ch8_1.cpu0.s .section .mdebug.abi32 .previous .file "ch8_1.bc" .text

138

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.globl _Z5sum_iiiiiii .align 2 .type _Z5sum_iiiiiii,@function .ent _Z5sum_iiiiiii # @_Z5sum_iiiiiii _Z5sum_iiiiiii: .cfi_startproc .frame $sp,32,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: addiu $sp, $sp, -32 $tmp1: .cfi_def_cfa_offset 32 ld $2, 32($sp) st $2, 28($sp) ld $2, 36($sp) st $2, 24($sp) ld $2, 40($sp) st $2, 20($sp) ld $2, 44($sp) st $2, 16($sp) ld $2, 48($sp) st $2, 12($sp) ld $2, 52($sp) st $2, 8($sp) ld $3, 24($sp) ld $4, 28($sp) add $3, $4, $3 ld $4, 20($sp) add $3, $3, $4 ld $4, 16($sp) add $3, $3, $4 ld $4, 12($sp) add $3, $3, $4 add $2, $3, $2 st $2, 4($sp) addiu $sp, $sp, 32 ret $lr .set macro .set reorder .end _Z5sum_iiiiiii $tmp2: .size _Z5sum_iiiiiii, ($tmp2)-_Z5sum_iiiiiii .cfi_endproc .globl main .align 2 .type main,@function .ent main main: .cfi_startproc .frame $sp,40,$lr .mask 0x00004000,-4 .set noreorder .cpload $t9 .set nomacro # BB#0:

# @main

9.4. Fix the wrong offset in storing arguments to stack frame

139

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

addiu $sp, $sp, -40 $tmp5: .cfi_def_cfa_offset 40 st $lr, 36($sp) # 4-byte Folded Spill $tmp6: .cfi_offset 14, -4 addiu $2, $zero, 0 st $2, 32($sp) !ADJCALLSTACKDOWN 24 addiu $2, $zero, 6 st $2, 20($sp) addiu $2, $zero, 5 st $2, 16($sp) addiu $2, $zero, 4 st $2, 12($sp) addiu $2, $zero, 3 st $2, 8($sp) addiu $2, $zero, 2 st $2, 4($sp) addiu $2, $zero, 1 st $2, 0($sp) ld $6, %call24(_Z5sum_iiiiiii)($gp) jalr $6 !ADJCALLSTACKUP 24 st $2, 28($sp) ld $lr, 36($sp) # 4-byte Folded Reload addiu $sp, $sp, 40 ret $lr .set macro .set reorder .end main $tmp7: .size main, ($tmp7)-main .cfi_endproc

The incoming arguments is the formal arguments dened in compiler and program language books. The outgoing arguments is the actual arguments. Summary callee incoming arguments and caller outgoing arguments as Callee incoming arguments and caller outgoing arguments.

Figure 9.3: Callee incoming arguments and caller outgoing arguments

140

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

9.5 Pseudo hook instruction ADJCALLSTACKDOWN and ADJCALLSTACKUP

To x the !ADJSTACKDOWN and !1ADJSTACKUP, we call Cpu0GenInstrInfo(Cpu0:: ADJCALLSTACKDOWN, Cpu0::ADJCALLSTACKUP) in Cpu0InstrInfo() constructor function and dene eliminateCallFramePseudoInstr() as follows,
// Cpu0InstrInfo.cpp ... Cpu0InstrInfo::Cpu0InstrInfo(Cpu0TargetMachine &tm) : Cpu0GenInstrInfo(Cpu0::ADJCALLSTACKDOWN, Cpu0::ADJCALLSTACKUP), ... // Cpu0RegisterInfo.cpp ... // Cpu0 // This function eliminate ADJCALLSTACKDOWN, // ADJCALLSTACKUP pseudo instructions void Cpu0RegisterInfo:: eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB, MachineBasicBlock::iterator I) const { // Simply discard ADJCALLSTACKDOWN, ADJCALLSTACKUP instructions. MBB.erase(I); }

With above denition, eliminateCallFramePseudoInstr() will be called when llvm meet pseudo instructions ADJCALLSTACKDOWN and ADJCALLSTACKUP. We just discard these 2 pseudo instructions. Run 8/5/Cpu0 with ch8_1.cpp will get the following result.
118-165-76-131:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype =asm ch8_1.bc -o ch8_1.cpu0.s 118-165-76-131:InputFiles Jonathan$ cat ch8_1.cpu0.s .section .mdebug.abi32 .previous .file "ch8_1.bc" .text .globl _Z5sum_iiiiiii .align 2 .type _Z5sum_iiiiiii,@function .ent _Z5sum_iiiiiii # @_Z5sum_iiiiiii _Z5sum_iiiiiii: .cfi_startproc .frame $sp,32,$lr .mask 0x00000000,0 .set noreorder .set nomacro # BB#0: addiu $sp, $sp, -32 $tmp1: .cfi_def_cfa_offset 32 ld $2, 32($sp) st $2, 28($sp) ld $2, 36($sp) st $2, 24($sp) ld $2, 40($sp) st $2, 20($sp)

9.5. Pseudo hook instruction ADJCALLSTACKDOWN and ADJCALLSTACKUP

141

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

ld $2, 44($sp) st $2, 16($sp) ld $2, 48($sp) st $2, 12($sp) ld $2, 52($sp) st $2, 8($sp) ld $3, 24($sp) ld $4, 28($sp) add $3, $4, $3 ld $4, 20($sp) add $3, $3, $4 ld $4, 16($sp) add $3, $3, $4 ld $4, 12($sp) add $3, $3, $4 add $2, $3, $2 st $2, 4($sp) addiu $sp, $sp, 32 ret $lr .set macro .set reorder .end _Z5sum_iiiiiii $tmp2: .size _Z5sum_iiiiiii, ($tmp2)-_Z5sum_iiiiiii .cfi_endproc .globl main .align 2 .type main,@function .ent main main: .cfi_startproc .frame $sp,64,$lr .mask 0x00004000,-4 .set noreorder .cpload $t9 .set nomacro # BB#0: addiu $sp, $sp, -64 $tmp5: .cfi_def_cfa_offset 64 st $lr, 60($sp) $tmp6: .cfi_offset 14, -4 addiu $2, $zero, 0 st $2, 56($sp) addiu $2, $zero, 6 st $2, 20($sp) addiu $2, $zero, 5 st $2, 16($sp) addiu $2, $zero, 4 st $2, 12($sp) addiu $2, $zero, 3 st $2, 8($sp) addiu $2, $zero, 2 st $2, 4($sp) addiu $2, $zero, 1 st $2, 0($sp)

# @main

# 4-byte Folded Spill

142

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

ld $6, %call24(_Z5sum_iiiiiii)($gp) jalr $6 st $2, 52($sp) ld $lr, 60($sp) # 4-byte Folded Reload addiu $sp, $sp, 64 ret $lr .set macro .set reorder .end main $tmp7: .size main, ($tmp7)-main .cfi_endproc

9.6 Handle $gp register in PIC addressing mode

In section Global variable 5 , we mentioned two addressing mode, the static address mode and PIC (positionindependent code) mode. We also mentioned, one example of PIC mode is used in share library. Share library usually can be loaded in different memory address decided at run time. The static mode (absolute address mode) is usually designed to load in specic memory address decided at compile time. Since share library can be loaded in different memory address, the global variable address cannot be decided at compile time. But, we can caculate the distance between the global variable address and shared library function if they will be loaded to the contiguous memory space together. Lets run 8/6/Cpu0 with ch8_2.cpp to get the following result of we putting the comment in it for explanation.
118-165-67-25:InputFiles Jonathan$ cat ch8_2.cpu0.s .section .mdebug.abi32 .previous .file "ch8_2.bc" .text .globl _Z5sum_iiiiiii .align 2 .type _Z5sum_iiiiiii,@function .ent _Z5sum_iiiiiii # @_Z5sum_iiiiiii _Z5sum_iiiiiii: ... .cpload $t9 // assign $gp = $t9 by loader when loader load re-entry // function (shared library) of _Z5sum_iiiiiii .set nomacro # BB#0: addiu $sp, $sp, -32 $tmp1: .cfi_def_cfa_offset 32 ... ld $3, %got(gI)($gp) // %got(gI) is offset of (gI - _Z5sum_iiiiiii) ... ret $lr .set macro .set reorder .end _Z5sum_iiiiiii ... .ent main # @main main: .cfi_startproc .frame $sp,72,$lr .mask 0x00004000,-4

9.6. Handle $gp register in PIC addressing mode

143

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.set noreorder .cpload $t9 .set nomacro # BB#0: addiu $sp, $sp, -72 $tmp5: .cfi_def_cfa_offset 72 st $lr, 68($sp) # 4-byte Folded Spill $tmp6: .cfi_offset 14, -4 .cprestore 24 // save $gp to 24($sp) addiu $2, $zero, 0 ... ld $6, %call24(_Z5sum_iiiiiii)($gp) jalr $6 // $t9 register number is 6, meaning $6 and %t9 are the // same register ld $gp, 24($sp) // restore $gp from 24($sp) ... addiu $sp, $sp, 72 ret $lr .set macro .set reorder .end main $tmp7: .size main, ($tmp7)-main .cfi_endproc .type .data .globl .align gI: .4byte .size 100 gI, 4 # 0x64 gI,@object gI 2 # @gI

As above code comment, .cprestore 24 is a pseudo instruction for saving $gp to 24($sp); Instruction ld $gp, 24($sp) will restore the $gp. In other word, $gp is caller saved register, so main() need to save/restore $gp before/after call the shared library _Z5sum_iiiiiii() function. In _Z5sum_iiiiiii() function, we translate global variable gI address by ld $3, %got(gI)($gp) where %got(gI) is offset of (gI - _Z5sum_iiiiiii) (we can write our cpu0 compiler to produce obj code by calculate the offset value). According the original cpu0 web site information, it only support jsub 24 bits address range access. We add jalr to cpu0 and expand it to 32 bit address. We did this change for two reason. One is cpu0 can be expand to 32 bit address space by only add this instruction. The other is cpu0 is designed for teaching purpose, this book has the same purpose for llvm backend design. We reserve jalr as PIC mode for shared library or dynamic loading code to demonstrate the caller how to handle the caller saved register $gp in calling the shared library and the shared library how to use $gp to access global variable address. This solution is popular in reality and deserve change cpu0 ofcial design as a compiler book. Mips use the same solution in 32 bits Mips32 CPU. Now, as the following code added in 8/6/Cpu0, we can issue .cprestore in emitPrologue() and emit ld $gp, ($gp save slot on stack) after jalr by create le Cpu0EmitGPRestore.cpp which run as a function pass.
// # CMakeLists.txt ... add_llvm_target(Cpu0CodeGen ... Cpu0EmitGPRestore.cpp ...

144

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// Cpu0TargetMachine.cpp ... bool Cpu0PassConfig::addPreRegAlloc() { // Do not restore $gp if target is Cpu064. // In N32/64, $gp is a callee-saved register. PM->add(createCpu0EmitGPRestorePass(getCpu0TargetMachine())); return true; } // Cpu0.h ... FunctionPass *createCpu0EmitGPRestorePass(Cpu0TargetMachine &TM); // Cpu0FrameLowering.cpp ... void Cpu0FrameLowering::emitPrologue(MachineFunction &MF) const { ... unsigned RegSize = 4; unsigned LocalVarAreaOffset = Cpu0FI->needGPSaveRestore() ? (MFI->getObjectOffset(Cpu0FI->getGPFI()) + RegSize) : Cpu0FI->getMaxCallFrameSize(); .... // Restore GP from the saved stack location if (Cpu0FI->needGPSaveRestore()) { unsigned Offset = MFI->getObjectOffset(Cpu0FI->getGPFI()); BuildMI(MBB, MBBI, dl, TII.get(Cpu0::CPRESTORE)).addImm(Offset) .addReg(Cpu0::GP); } } // Cpu0InstrInfo.td ... // When handling PIC code the assembler // directives. If the real instructions // are used, we have the same behavior, // from the assembler. let neverHasSideEffects = 1 in def CPRESTORE : Cpu0Pseudo<(outs), (ins ".cprestore\t$loc", []>;

needs .cpload and .cprestore corresponding these directives but get also a bunch of warnings

i32imm:$loc, CPURegs:$gp),

// Cpu0SelLowering.cpp ... SDValue Cpu0TargetLowering::LowerCall(SDValue InChain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg, bool doesNotRet, bool &isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, const SmallVectorImpl<ISD::InputArg> &Ins, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { ... // If this is the first call, create a stack frame object that points to // a location to which .cprestore saves $gp. if (IsPIC && Cpu0FI->globalBaseRegFixed() && !Cpu0FI->getGPFI())

9.6. Handle $gp register in PIC addressing mode

145

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

... if (MaxCallFrameSize < NextStackOffset) { if (Cpu0FI->needGPSaveRestore()) MFI->setObjectOffset(Cpu0FI->getGPFI(), NextStackOffset); ... } // Cpu0EmitGPRestore.cpp //===-- Cpu0EmitGPRestore.cpp - Emit GP Restore Instruction ---------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // // This pass emits instructions that restore $gp right // after jalr instructions. // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "emit-gp-restore" #include #include #include #include #include #include #include "Cpu0.h" "Cpu0TargetMachine.h" "Cpu0MachineFunction.h" "llvm/CodeGen/MachineFunctionPass.h" "llvm/CodeGen/MachineInstrBuilder.h" "llvm/Target/TargetInstrInfo.h" "llvm/ADT/Statistic.h"

using namespace llvm; namespace { struct Inserter : public MachineFunctionPass { TargetMachine &TM; const TargetInstrInfo *TII; static char ID; Inserter(TargetMachine &tm) : MachineFunctionPass(ID), TM(tm), TII(tm.getInstrInfo()) { } virtual const char *getPassName() const { return "Cpu0 Emit GP Restore"; } bool runOnMachineFunction(MachineFunction &F); }; char Inserter::ID = 0; } // end of anonymous namespace bool Inserter::runOnMachineFunction(MachineFunction &F) { Cpu0FunctionInfo *Cpu0FI = F.getInfo<Cpu0FunctionInfo>(); if ((TM.getRelocationModel() != Reloc::PIC_) || (!Cpu0FI->globalBaseRegFixed())) return false;

146

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

bool Changed = false; int FI = Cpu0FI->getGPFI(); for (MachineFunction::iterator MFI = F.begin(), MFE = F.end(); MFI != MFE; ++MFI) { MachineBasicBlock& MBB = *MFI; MachineBasicBlock::iterator I = MFI->begin(); /// IsLandingPad - Indicate that this basic block is entered via an /// exception handler. // If MBB is a landing pad, insert instruction that restores $gp after // EH_LABEL. if (MBB.isLandingPad()) { // Find EH_LABEL first. for (; I->getOpcode() != TargetOpcode::EH_LABEL; ++I) ; // Insert ld. ++I; DebugLoc dl = I != MBB.end() ? I->getDebugLoc() : DebugLoc(); BuildMI(MBB, I, dl, TII->get(Cpu0::LD), Cpu0::GP).addFrameIndex(FI) .addImm(0); Changed = true; } while (I != MFI->end()) { if (I->getOpcode() != Cpu0::JALR) { ++I; continue; } DebugLoc dl = I->getDebugLoc(); // emit ld $gp, ($gp save slot on stack) after jalr BuildMI(MBB, ++I, dl, TII->get(Cpu0::LD), Cpu0::GP).addFrameIndex(FI) .addImm(0); Changed = true; } } return Changed; } /// createCpu0EmitGPRestorePass - Returns a pass that emits instructions that /// restores $gp clobbered by jalr instructions. FunctionPass *llvm::createCpu0EmitGPRestorePass(Cpu0TargetMachine &tm) { return new Inserter(tm); } //===-- Cpu0MachineFunctionInfo.h - Private data used for Cpu0 ----*- C++ -*-=// ... class Cpu0FunctionInfo : public MachineFunctionInfo { ... bool EmitNOAT; public: Cpu0FunctionInfo(MachineFunction& MF) : ... MaxCallFrameSize(0), EmitNOAT(false) ...

9.6. Handle $gp register in PIC addressing mode

147

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

bool getEmitNOAT() const { return EmitNOAT; } void setEmitNOAT() { EmitNOAT = true; } }; } // end of namespace llvm #endif // CPU0_MACHINE_FUNCTION_INFO_H // Cpu0AsmPrinter.cpp ... void Cpu0AsmPrinter::EmitInstrWithMacroNoAT(const MachineInstr *MI) { MCInst TmpInst; MCInstLowering.Lower(MI, TmpInst); OutStreamer.EmitRawText(StringRef("\t.set\tmacro")); if (Cpu0FI->getEmitNOAT()) OutStreamer.EmitRawText(StringRef("\t.set\tat")); OutStreamer.EmitInstruction(TmpInst); if (Cpu0FI->getEmitNOAT()) OutStreamer.EmitRawText(StringRef("\t.set\tnoat")); OutStreamer.EmitRawText(StringRef("\t.set\tnomacro")); } void Cpu0AsmPrinter::EmitInstruction(const MachineInstr *MI) { ... unsigned Opc = MI->getOpcode(); MCInst TmpInst0; SmallVector<MCInst, 4> MCInsts; switch (Opc) { case Cpu0::CPRESTORE: { const MachineOperand &MO = MI->getOperand(0); assert(MO.isImm() && "CPRESTOREs operand must be an immediate."); int64_t Offset = MO.getImm(); if (OutStreamer.hasRawTextSupport()) { if (!isInt<16>(Offset)) { EmitInstrWithMacroNoAT(MI); return; } } else { MCInstLowering.LowerCPRESTORE(Offset, MCInsts); for (SmallVector<MCInst, 4>::iterator I = MCInsts.begin(); I != MCInsts.end(); ++I) OutStreamer.EmitInstruction(*I); return; } break; } default: break; } MCInstLowering.Lower(MI, TmpInst0);

148

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

OutStreamer.EmitInstruction(TmpInst0); } void Cpu0AsmPrinter::EmitFunctionBodyStart() { ... if (OutStreamer.hasRawTextSupport()) { ... if (Cpu0FI->getEmitNOAT()) OutStreamer.EmitRawText(StringRef("\t.set\tnoat")); } else if (EmitCPLoad) { SmallVector<MCInst, 4> MCInsts; MCInstLowering.LowerCPLOAD(MCInsts); for (SmallVector<MCInst, 4>::iterator I = MCInsts.begin(); I != MCInsts.end(); ++I) OutStreamer.EmitInstruction(*I); } } // Cpu0MCInstLower.cpp ... sstatic void CreateMCInst(MCInst& Inst, unsigned Opc, const MCOperand& Opnd0, const MCOperand& Opnd1, const MCOperand& Opnd2 = MCOperand()) { Inst.setOpcode(Opc); Inst.addOperand(Opnd0); Inst.addOperand(Opnd1); if (Opnd2.isValid()) Inst.addOperand(Opnd2); } // Lower ".cpload $reg" to // "addiu $gp, $zero, %hi(_gp_disp)" // "shl $gp, $gp, 16" // "addiu $gp, $gp, %lo(_gp_disp)" // "addu $gp, $gp, $t9" void Cpu0MCInstLower::LowerCPLOAD(SmallVector<MCInst, 4>& MCInsts) { MCOperand GPReg = MCOperand::CreateReg(Cpu0::GP); MCOperand T9Reg = MCOperand::CreateReg(Cpu0::T9); MCOperand ZEROReg = MCOperand::CreateReg(Cpu0::ZERO); StringRef SymName("_gp_disp"); const MCSymbol *Sym = Ctx->GetOrCreateSymbol(SymName); const MCSymbolRefExpr *MCSym; MCSym = MCSymbolRefExpr::Create(Sym, MCSymbolRefExpr::VK_Cpu0_ABS_HI, *Ctx); MCOperand SymHi = MCOperand::CreateExpr(MCSym); MCSym = MCSymbolRefExpr::Create(Sym, MCSymbolRefExpr::VK_Cpu0_ABS_LO, *Ctx); MCOperand SymLo = MCOperand::CreateExpr(MCSym); MCInsts.resize(4); CreateMCInst(MCInsts[0], CreateMCInst(MCInsts[1], CreateMCInst(MCInsts[2], CreateMCInst(MCInsts[3], } // Lower ".cprestore offset" to "st $gp, offset($sp)". void Cpu0MCInstLower::LowerCPRESTORE(int64_t Offset, Cpu0::ADDiu, GPReg, ZEROReg, SymHi); Cpu0::SHL, GPReg, GPReg, MCOperand::CreateImm(16)); Cpu0::ADDiu, GPReg, GPReg, SymLo); Cpu0::ADD, GPReg, GPReg, T9Reg);

9.6. Handle $gp register in PIC addressing mode

149

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

SmallVector<MCInst, 4>& MCInsts) { assert(isInt<32>(Offset) && (Offset >= 0) && "Imm operand of .cprestore must be a non-negative 32-bit value."); MCOperand SPReg = MCOperand::CreateReg(Cpu0::SP), BaseReg = SPReg; MCOperand GPReg = MCOperand::CreateReg(Cpu0::GP); MCOperand ZEROReg = MCOperand::CreateReg(Cpu0::ZERO); if (!isInt<16>(Offset)) { unsigned Hi = ((Offset + 0x8000) >> 16) & 0xffff; Offset &= 0xffff; MCOperand ATReg = MCOperand::CreateReg(Cpu0::AT); BaseReg = ATReg; // addiu at,zero,hi // shl at,at,16 // add at,at,sp MCInsts.resize(3); CreateMCInst(MCInsts[0], Cpu0::ADDiu, ATReg, ZEROReg, MCOperand::CreateImm(Hi)); CreateMCInst(MCInsts[1], Cpu0::SHL, ATReg, ATReg, MCOperand::CreateImm(16)); CreateMCInst(MCInsts[2], Cpu0::ADD, ATReg, ATReg, SPReg); } MCInst St; CreateMCInst(St, Cpu0::ST, GPReg, BaseReg, MCOperand::CreateImm(Offset)); MCInsts.push_back(St); }

The above added code of Cpu0AsmPrinter.cpp will call the LowerCPLOAD() and LowerCPRESTORE() when user run with llc -filetype=obj. The above added code of Cpu0MCInstLower.cpp take care the .cpload and .cprestore machine instructions. It translate pseudo asm .cpload into four machine instructions, and .cprestore into one machine instruction as below. As mentioned in section Global variable 5 . When the share library main() function be loaded, the loader will set the $t9 value to $gp when meet .cpload $t9. After that, the $gp value is $t9 which point to main(), and the global variable address is the relative address to main(). The _gp_disp is zero as the following reason from Mips ABI.
// Lower ".cpload $reg" to // "addiu $gp, $zero, %hi(_gp_disp)" // "shl $gp, $gp, 16" // "addiu $gp, $gp, %lo(_gp_disp)" // "addu $gp, $gp, $t9" // Lower ".cprestore offset" to "st $gp, offset($sp)".

Note: // Mips ABI: _gp_disp After calculating the gp, a function allocates the local stack space and saves the gp on the stack, so it can be restored after subsequent function calls. In other words, the gp is a caller saved register. ... _gp_disp represents the offset between the beginning of the function and the global offset table. Various optimizations are possible in this code example and the others that follow. For example, the calculation of gp need not be done for a position-independent function that is strictly local to an object module. By run with llc -filetype=obj, the .cpload and .cprestore are translated into machine code as follows,
118-165-76-131:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=

150

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

obj ch8_2.bc -o ch8_2.cpu0.o 118-165-76-131:InputFiles Jonathan$ hexdump ch8_2.cpu0.o ... // .cpload machine instructions "09 a0 00 00 to 13 aa 60 00" 0000030 00 0a 00 07 09 a0 00 00 1e aa 00 10 09 aa 00 00 0000040 13 aa 60 00 09 dd ff e0 00 2d 00 20 01 2d 00 1c ... // .cpload machine instructions "09 a0 00 00 to 13 aa 60 00" 00000b0 09 dd 00 20 2c 00 00 00 09 a0 00 00 1e aa 00 10 00000c0 09 aa 00 00 13 aa 60 00 09 dd ff b8 01 ed 00 44 // .cprestore machine instruction 01 ad 00 18 00000d0 01 ad 00 18 09 20 00 00 01 2d 00 40 09 20 00 06 ... 118-165-67-25:InputFiles Jonathan$ cat ch8_2.cpu0.s ... .ent _Z5sum_iiiiiii # @_Z5sum_iiiiiii _Z5sum_iiiiiii: ... .cpload $t9 // assign $gp = $t9 by loader when loader load re-entry function // (shared library) of _Z5sum_iiiiiii .set nomacro # BB#0: ... .ent main # @main ... .cpload $t9 .set nomacro ... .cprestore 24 // save $gp to 24($sp) ...

Run llc -static will call jsub instruction instead of jalr as follows,
118-165-76-131:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=static -filetype= asm ch8_2.bc -o ch8_2.cpu0.s 118-165-76-131:InputFiles Jonathan$ cat ch8_2.cpu0.s ... jsub _Z5sum_iiiiiii ...

Run with llc -obj, you can nd the Cx of jsub Cx is 0 since the Cx is calculated by linker as below. Mips has the same 0 in its jal instruction. The ch8_1_2.cpp, ch8_1_3.cpp and ch8_1_4.cpp are example code more for test.
// jsub _Z5sum_iiiiiii translate into 2B 00 00 00 00F0: 2B 00 00 00 01 2D 00 34 00 ED 00 3C 09 DD 00 40

9.7 Variable number of arguments

Until now, we support xed number of arguments in formal function denition (Incoming Arguments). This section support variable number of arguments since C language support this feature. Run 8/7/Cpu0 with ch8_3.cpp to get the following,

9.7. Variable number of arguments

151

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

// ch8_3.cpp //#include <stdio.h> #include <stdarg.h> int sum_i(int amount, ...) { int i = 0; int val = 0; int sum = 0; va_list vl; va_start(vl, amount); for (i = 0; i < amount; i++) { val = va_arg(vl, int); sum += val; } va_end(vl); return sum; } int main() { int a = sum_i(6, 1, 2, 3, 4, 5, 6); // printf("a = %d\n", a); return a; } 118-165-76-131:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch8_3.bc -o ch8_3.cpu0.s 118-165-76-131:InputFiles Jonathan$ cat ch8_3.cpu0.s .section .mdebug.abi32 .previous .file "ch8_3.bc" .text .globl _Z5sum_iiz .align 2 .type _Z5sum_iiz,@function .ent _Z5sum_iiz # @_Z5sum_iiz _Z5sum_iiz: .cfi_startproc .frame $sp,56,$lr .mask 0x00004000,-4 .set noreorder .cpload $t9 .set nomacro # BB#0: addiu $sp, $sp, -56 $tmp2: .cfi_def_cfa_offset 56 st $lr, 52($sp) # 4-byte Folded Spill $tmp3: .cfi_offset 14, -4 .cprestore 0 ld $2, %got(__stack_chk_guard)($gp)

152

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

ld $2, 0($2) st $2, 48($sp) // 48($sp) = 0 ld $2, 56($sp) // amount st $2, 44($sp) // amount addiu $2, $zero, 0 st $2, 40($sp) // i = 0 st $2, 36($sp) // val = 0 st $2, 32($sp) // sum = 0 addiu $3, $sp, 48 // $3 = 48($sp) st $3, 8($sp) // 8($sp) = 48($sp) = arg_ptr st $2, 40($sp) // i = 0 addiu $2, $zero, 40 // $2 = 40 $BB0_1: # =>This Inner Loop Header: Depth=1 ld $3, 44($sp) // $3 = amount ld $4, 40($sp) // $4 = i cmp $4, $3 jge $BB0_7 // i >= amount jmp $BB0_2 $BB0_2: # in Loop: Header=BB0_1 Depth=1 // i < amount ld $3, 8($sp) // $3 = arg_ptr cmp $3, $2 jgt $BB0_4 // arg_ptr > 40 jmp $BB0_3 $BB0_3: # in Loop: Header=BB0_1 Depth=1 // arg_ptr <= 40 addiu $4, $3, 8 ld $5, 20($sp) // *(20($sp)) = arg_offset = 12 st $4, 8($sp) // arg_ptr += 8 add $3, $5, $3 // $3 = (arg_ptr + arg_offset) jmp $BB0_5 $BB0_4: # in Loop: Header=BB0_1 Depth=1 ld $3, 16($sp) addiu $4, $3, 8 st $4, 16($sp) $BB0_5: # in Loop: Header=BB0_1 Depth=1 ld $3, 0($3) // $3 = val = *(arg_ptr + arg_offset) st $3, 36($sp) ld $4, 32($sp) // $4 = sum add $3, $4, $3 st $3, 32($sp) // sum += val # BB#6: # in Loop: Header=BB0_1 Depth=1 ld $3, 40($sp) // $3 = i addiu $3, $3, 1 st $3, 40($sp) // i = i + 1 jmp $BB0_1 $BB0_7: ld $2, %got(__stack_chk_guard)($gp) ld $2, 0($2) ld $3, 48($sp) cmp $2, $3 jne $BB0_9 jmp $BB0_8 $BB0_8: # %SP_return ld $lr, 52($sp) # 4-byte Folded Reload addiu $sp, $sp, 56 ret $lr $BB0_9: # %CallStackCheckFailBlk

9.7. Variable number of arguments

153

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

ld $6, %call24(__stack_chk_fail)($gp) jalr $6 ld $gp, 0($sp) .set macro .set reorder .end _Z5sum_iiz $tmp4: .size _Z5sum_iiz, ($tmp4)-_Z5sum_iiz .cfi_endproc .globl main .align 2 .type main,@function # @main .ent main main: .cfi_startproc .frame $sp,88,$lr .mask 0x00004000,-4 .set noreorder .cpload $t9 .set nomacro # BB#0: addiu $sp, $sp, -88 $tmp7: .cfi_def_cfa_offset 88 st $lr, 84($sp) # 4-byte Folded Spill $tmp8: .cfi_offset 14, -4 .cprestore 32 addiu $2, $zero, 0 st $2, 80($sp) addiu $2, $zero, 5 st $2, 20($sp) addiu $2, $zero, 4 st $2, 16($sp) addiu $2, $zero, 3 st $2, 12($sp) addiu $2, $zero, 2 st $2, 8($sp) addiu $2, $zero, 1 st $2, 4($sp) addiu $2, $zero, 6 st $2, 24($sp) st $2, 0($sp) ld $6, %call24(_Z5sum_iiz)($gp) jalr $6 ld $gp, 32($sp) st $2, 76($sp) ld $lr, 84($sp) # 4-byte Folded Reload addiu $sp, $sp, 88 ret $lr .set macro .set reorder .end main $tmp9: .size main, ($tmp9)-main .cfi_endproc

154

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

We have problem in analysis of the output ch8_3.cpu0.s. We guess and try to analysis as follows. As above code, we get the rst argument amount from ld $2, 56($sp) since the stack size of the callee function _Z5sum_iiz() is 56. Next, check i < amount in block $BB0_1. If i < amount, than enter into $BB0_2. We assume arg_ptr < 40 and the content of address 8($sp) is the arg_ptr. When it exits $BB0_2 and enter into $BB0_3, the register ($3 + $5) = (arg_ptr + arg_offset=12) is point to the second argument and it do the sum += val in $BB0_5. It do i += 1 in $BB0_6 and jumb to $BB0_1 enter into second round. The second round do as above again, it will get the third argument and add to sum in $BB0_5 since the ptr_arg (16($sp)) is added 8 in the previous run. We assume the arg_prt < 40 but actually according the analysis the arg_prt is 48($sp) which > 40, so the above analysis is not satised. The compare arg_prt with 40 is exist in llvm IR, and mips has the same translated output. So, we dont know whats wrong. We believe the arg < 40 is satised because the native Intel CPU has the arg_ptr < 40 in its assembly code and the Intel CPU native execution le can print correct result. You will see it soon in the bellow code. If the arg_ptr < 40 is satised and *(20($sp)) = arg_offset = 12, then the assembly output is correct. The llvm IR and mips assembly output as follows,
118-165-78-221:InputFiles Jonathan$ llvm-dis ch8_3.bc -o ch8_3.ll 118-165-78-221:InputFiles Jonathan$ cat ch8_3.ll ; ModuleID = ch8_3.bc target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16: 32:64-S128" target triple = "x86_64-apple-macosx10.8.0" %struct.__va_list_tag = type { i32, i32, i8*, i8* } define i32 @_Z5sum_iiz(i32 %amount, ...) nounwind uwtable ssp { ... ; <label>:8 ; preds = %4 ... %12 = icmp ule i32 %11, 40 br i1 %12, label %13, label %19 118-165-67-185:InputFiles Jonathan$ cat ch8_3.mips.s .section .mdebug.abi32 .previous .file "ch8_3.bc" .text .globl _Z5sum_iiz .align 2 .type _Z5sum_iiz,@function .ent _Z5sum_iiz # @_Z5sum_iiz _Z5sum_iiz: .cfi_startproc .frame $sp,72,$ra .mask 0x80000000,-4 .fmask 0x00000000,0 .set noreorder .cpload $25 .set nomacro # BB#0: addiu $sp, $sp, -72 $tmp2: .cfi_def_cfa_offset 72 sw $ra, 68($sp) # 4-byte Folded Spill $tmp3: .cfi_offset 31, -4 .cprestore 16 sw $7, 84($sp) sw $6, 80($sp) sw $5, 76($sp) // 76($sp) = arg[1]

9.7. Variable number of arguments

155

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

lw $2, %got(__stack_chk_guard)($gp) lw $2, 0($2) sw $2, 64($sp) sw $4, 60($sp) // 60($sp) = amount = arg[0] sw $zero, 56($sp) // i sw $zero, 52($sp) // val sw $zero, 48($sp) // sum addiu $2, $sp, 76 sw $2, 24($sp) // 24($sp) = arg_ptr sw $zero, 56($sp) addiu $2, $zero, 40 // $2 = 40 b $BB0_1 nop $BB0_5: # in Loop: Header=BB0_1 Depth=1 lw $3, 0($3) // $3 = *arg_ptr sw $3, 52($sp) // val lw $4, 48($sp) // sum addu $3, $4, $3 // sw $3, 48($sp) // sum += val lw $3, 56($sp) addiu $3, $3, 1 sw $3, 56($sp) // i += 1 $BB0_1: # =>This Inner Loop Header: Depth=1 lw $3, 60($sp) lw $4, 56($sp) slt $3, $4, $3 // set if i < amount beq $3, $zero, $BB0_6 // i >= amount nop # BB#2: # in Loop: Header=BB0_1 Depth=1 lw $3, 24($sp) // $3 = arg_ptr sltu $4, $2, $3 // set if 40 < arg_ptr bne $4, $zero, $BB0_4 nop # BB#3: # in Loop: Header=BB0_1 Depth=1 // arg_ptr <= 40 addiu $4, $3, 8 lw $5, 36($sp) // 36($sp) = 0, assume even though we didnt find the // 36($sp) is 0 sw $4, 24($sp) // arg_ptr += 8 addu $3, $5, $3 // arg_ptr + 0 b $BB0_5 nop $BB0_4: # in Loop: Header=BB0_1 Depth=1 // 40 < arg_ptr lw $3, 32($sp) addiu $4, $3, 8 sw $4, 32($sp) b $BB0_5 nop $BB0_6: lw $2, %got(__stack_chk_guard)($gp) lw $2, 0($2) lw $3, 64($sp) bne $2, $3, $BB0_8 nop # BB#7: # %SP_return lw $2, 48($sp) lw $ra, 68($sp) # 4-byte Folded Reload

156

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

addiu $sp, $sp, 72 jr $ra nop $BB0_8: # %CallStackCheckFailBlk lw $25, %call16(__stack_chk_fail)($gp) jalr $25 nop lw $gp, 16($sp) .set macro .set reorder .end _Z5sum_iiz $tmp4: .size _Z5sum_iiz, ($tmp4)-_Z5sum_iiz .cfi_endproc .globl main .align 2 .type main,@function .ent main # @main main: .cfi_startproc .frame $sp,64,$ra .mask 0x80000000,-4 .fmask 0x00000000,0 .set noreorder .cpload $25 .set nomacro # BB#0: addiu $sp, $sp, -64 $tmp7: .cfi_def_cfa_offset 64 sw $ra, 60($sp) # 4-byte Folded Spill $tmp8: .cfi_offset 31, -4 .cprestore 40 sw $zero, 56($sp) addiu $2, $zero, 7 sw $2, 28($sp) addiu $2, $zero, 6 sw $2, 24($sp) addiu $2, $zero, 5 sw $2, 20($sp) addiu $2, $zero, 4 sw $2, 16($sp) addiu $4, $zero, 8 sw $4, 32($sp) lw $25, %call16(_Z5sum_iiz)($gp) addiu $5, $zero, 1 addiu $6, $zero, 2 addiu $7, $zero, 3 jalr $25 nop lw $gp, 40($sp) sw $2, 52($sp) addu $2, $zero, $zero lw $ra, 60($sp) # 4-byte Folded Reload addiu $sp, $sp, 64 jr $ra

9.7. Variable number of arguments

157

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

nop .set macro .set reorder .end main $tmp9: .size main, ($tmp9)-main .cfi_endproc

We have veried the translation of ch8_3.cpp is correct by add printf in ch8_3.cpp to get ch8_3_3.cpp and run with lli llvm interpreter. We also translate it into native Intel CPU code and get the correct print result. Following are the ch8_3_3.cpp, and lli, Intel native code run result.
// ch8_3_3.cpp // clang -c ch8_3_3.cpp -emit-llvm -I/Applications/Xcode.app/Contents/ Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX10.8.sdk/usr/ include/ -o ch8_3_3.bc // /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake_debug_build/bin/Debug/llc ch8_3_3.bc -o ch8_3_3.s // clang++ ch8_3_3.s -o ch8_3_3.native // ./ch8_3_3.native // lldb -- ch8_3_3.native // b main // s // ... // print $rsp ; print %rsp, choose $ instead of % in assembly code // // // // mips-linux-gnu-g++ -g ch8_3_3.cpp -o ch8_3_3 -static qemu-mips ch8_3_3 mips-linux-gnu-g++ -S ch8_3_3.cpp cat ch8_3_3.s

#include <stdio.h> #include <stdarg.h> int sum_i(int amount, ...) { int i = 0; int val = 0; int sum = 0; va_list vl; va_start(vl, amount); for (i = 0; i < amount; i++) { val = va_arg(vl, int); sum += val; } va_end(vl); return sum; } int main() { int a = sum_i(6, 1, 2, 3, 4, 5, 6); printf("a = %d\n", a); return a;

158

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

} 118-165-78-221:InputFiles Jonathan$ lli ch8_3_3.bc a = 21 118-165-67-185:InputFiles Jonathan$ clang -c ch8_3_3.cpp -emit-llvm -I /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/ Developer/SDKs/MacOSX10.8.sdk/usr/include/ -o ch8_3_3.bc 118-165-67-185:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/ cmake_debug_build/bin/Debug/llc ch8_3_3.bc -o ch8_3_3.s 118-165-67-185:InputFiles Jonathan$ clang++ ch8_3_3.s -o ch8_3_3.native 118-165-67-185:InputFiles Jonathan$ ./ch8_3_3.native a = 21 118-165-67-185:InputFiles Jonathan$ cat ch8_3_3.s ... LBB0_3: ## =>This Inner Loop Header: Depth=1 movl 216(%rsp), %eax cmpl 220(%rsp), %eax jge LBB0_8 // i >= amount ## BB#4: ## in Loop: Header=BB0_3 Depth=1 movl 176(%rsp), %eax // i < amount cmpl $40, %eax // arg_ptr < 40 ja LBB0_6

We have run mips qemu on Linux by gcc. It get the correct print result, and the mips code has no analysis problem since it hasnt the arg_ptr < 40 in assembly output. The qemu mips gcc result as follows,
[Gamma@localhost InputFiles]$ qemu-mips ch8_3_3 a = 21 [Gamma@localhost InputFiles]$ mips-linux-gnu-g++ -g ch8_3_3.cpp -o ch8_3_3 -static [Gamma@localhost InputFiles]$ qemu-mips ch8_3_3 a = 21 [Gamma@localhost InputFiles]$ mips-linux-gnu-g++ -S ch8_3_3.cpp [Gamma@localhost InputFiles]$ cat ch8_3_3.s .file 1 "ch8_3_3.cpp" .section .mdebug.abi32 .previous .gnu_attribute 4, 1 .abicalls .option pic0 .text .align 2 .globl _Z5sum_iiz $LFB0 = . .set nomips16 .ent _Z5sum_iiz .type _Z5sum_iiz, @function _Z5sum_iiz: .frame $fp,32,$31 # vars= 16, regs= 1/0, args= 0, gp= 8 .mask 0x40000000,-4 .fmask 0x00000000,0 .set noreorder .set nomacro addiu $sp,$sp,-32 $LCFI0: sw $fp,28($sp)

9.7. Variable number of arguments

159

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

$LCFI1: move $fp,$sp $LCFI2: sw $5,36($fp) sw $6,40($fp) sw $7,44($fp) sw $4,32($fp) sw $0,16($fp) sw $0,12($fp) sw $0,8($fp) addiu $2,$fp,36 sw $2,20($fp) sw $0,16($fp) j $L2 nop

// arg[1]

// // // //

amount = arg[0] i = 0 val = 0 sum = 0

// arg_ptr = &arg[1]

$L3: // i < amount lw $2,20($fp) // arg_ptr addiu $3,$2,4 sw $3,20($fp) // arg_ptr += 4 lw $2,0($2) // $2 = *arg_ptr sw $2,12($fp) // val = *arg_ptr lw $3,8($fp) lw $2,12($fp) addu $2,$3,$2 sw $2,8($fp) // sum += val lw $2,16($fp) addiu $2,$2,1 sw $2,16($fp) // i += 1 $L2: lw $3,16($fp) lw $2,32($fp) slt $2,$3,$2 // set if i < amount andi $2,$2,0x00ff bne $2,$0,$L3 nop lw $2,8($fp) // i >= amount move $sp,$fp lw $fp,28($sp) addiu $sp,$sp,32 j $31 nop .set macro .set reorder .end _Z5sum_iiz $LFE0: .size _Z5sum_iiz, .-_Z5sum_iiz .rdata .align 2 $LC0: .ascii "a = %d\012\000" .text .align 2 .globl main $LFB1 = . .set nomips16

160

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.ent main .type main, @function main: .frame $fp,56,$31 # vars= 8, regs= 2/0, args= 32, gp= 8 .mask 0xc0000000,-4 .fmask 0x00000000,0 .set noreorder .set nomacro addiu $sp,$sp,-56 $LCFI3: sw $31,52($sp) $LCFI4: sw $fp,48($sp) $LCFI5: move $fp,$sp $LCFI6: li $2,4 # 0x4 sw $2,16($sp) li $2,5 # 0x5 sw $2,20($sp) li $2,6 # 0x6 sw $2,24($sp) li $4,6 # 0x6 li $5,1 # 0x1 li $6,2 # 0x2 li $7,3 # 0x3 jal _Z5sum_iiz nop sw $2,40($fp) lui $2,%hi($LC0) addiu $4,$2,%lo($LC0) lw $5,40($fp) jal printf nop lw $2,40($fp) move $sp,$fp lw $31,52($sp) lw $fp,48($sp) addiu $sp,$sp,56 j $31 nop .set macro .set reorder .end main $LFE1: .size main, .-main .section .eh_frame,"a",@progbits $Lframe1: .4byte $LECIE1-$LSCIE1 $LSCIE1: .4byte 0x0 .byte 0x1 .globl __gxx_personality_v0 .ascii "zP\000"

9.7. Variable number of arguments

161

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.uleb128 0x1 .sleb128 -4 .byte 0x1f .uleb128 0x5 .byte 0x0 .4byte __gxx_personality_v0 .byte 0xc .uleb128 0x1d .uleb128 0x0 .align 2 $LECIE1: $LSFDE3: .4byte $LEFDE3-$LASFDE3 $LASFDE3: .4byte $LASFDE3-$Lframe1 .4byte $LFB1 .4byte $LFE1-$LFB1 .uleb128 0x0 .byte 0x4 .4byte $LCFI3-$LFB1 .byte 0xe .uleb128 0x38 .byte 0x4 .4byte $LCFI5-$LCFI3 .byte 0x11 .uleb128 0x1e .sleb128 2 .byte 0x11 .uleb128 0x1f .sleb128 1 .byte 0x4 .4byte $LCFI6-$LCFI5 .byte 0xd .uleb128 0x1e .align 2 $LEFDE3: .ident "GCC: (GNU) 4.4.6" [Gamma@localhost InputFiles]$

To support variable number of arguments, the following code needed to add in 8/7/Cpu0. The ch8_3_2.cpp is C++ template example code, it can be translated into cpu0 backend code too.
// Cpu0TargetLowering.cpp ... Cpu0TargetLowering:: Cpu0TargetLowering(Cpu0TargetMachine &TM) : TargetLowering(TM, new Cpu0TargetObjectFile()), Subtarget(&TM.getSubtarget<Cpu0Subtarget>()) { ... setOperationAction(ISD::VASTART, MVT::Other, Custom); ... // Support va_arg(): variable numbers (not fixed numbers) of arguments // (parameters) for function all setOperationAction(ISD::VAARG, MVT::Other, Expand); setOperationAction(ISD::VACOPY, MVT::Other, Expand); setOperationAction(ISD::VAEND, MVT::Other, Expand); ... }

162

Chapter 9. Function call

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

... SDValue Cpu0TargetLowering:: LowerOperation(SDValue Op, SelectionDAG &DAG) const { switch (Op.getOpcode()) { ... case ISD::VASTART: return LowerVASTART(Op, DAG); } return SDValue(); } ... SDValue Cpu0TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { MachineFunction &MF = DAG.getMachineFunction(); Cpu0FunctionInfo *FuncInfo = MF.getInfo<Cpu0FunctionInfo>(); DebugLoc dl = Op.getDebugLoc(); SDValue FI = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), getPointerTy()); // vastart just stores the address of the VarArgsFrameIndex slot into the // memory location argument. const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); return DAG.getStore(Op.getOperand(0), dl, FI, Op.getOperand(1), MachinePointerInfo(SV), false, false, 0); } // ch8_3_2.cpp ... //#include <stdio.h> #include <stdarg.h> template<class T> T sum(T amount, ...) { T i = 0; T val = 0; T sum = 0; va_list vl; va_start(vl, amount); for (i = 0; i < amount; i++) { val = va_arg(vl, T); sum += val; } va_end(vl); return sum; } int main() { int a = sum<int>(6, 1, 2, 3, 4, 5, 6); // printf("a = %d\n", a); return a;

9.7. Variable number of arguments

163

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Mips qemu reference .

9.8 Verify DIV for operator %

Now, lets run 8/7/Cpu0 with 4_6_2.cpp to get the result as below. It translate (b+1)%c into div $zero, $3, $2 and mfhi $2.
// ch4_6_2.cpp #include <stdlib.h> int main() { int b = 11; // unsigned int b = 11; int c = rand(); b = (b+1)%c; return b; } 118-165-70-242:InputFiles Jonathan$ clang -c ch4_6_2.cpp -I/Applications/ Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/ MacOSX10.8.sdk/usr/include/ -emit-llvm -o ch4_6_2.bc 118-165-70-242:InputFiles Jonathan$ /Users/Jonathan/llvm/3.1.test/cpu0/1/cmake _debug_build/bin/Debug/llc -march=cpu0 -relocation-model=pic -filetype=asm ch4_6_2.bc -o ch4_6_2.cpu0.s 118-165-70-242:InputFiles Jonathan$ cat ch4_6_2.cpu0.s ... div $zero, $3, $2 mfhi $2 ...

9.9 Summary of this chapter

Until now, we have 5,700 lines of source code around in 8/7/Cpu0. The cpu0 backend code now can take care the integer function call and control statement just like the llvm front end tutorial example code. Look back the chapter of Back end structure, there are 3,000 lines of source code with taking three instructions only. With this 87% more of code, it can translate tens of instructions, global variable, control ow statement and function call.

164

Chapter 9. Function call

CHAPTER

TEN

ELF SUPPORT
Cpu0 backend generated the ELF format of obj. The ELF (Executable and Linkable Format) is a common standard le format for executables, object code, shared libraries and core dumps. First published in the System V Application Binary Interface specication, and later in the Tool Interface Standard, it was quickly accepted among different vendors of Unixsystems. In 1999 it was chosen as the standard binary le format for Unix and Unix-like systems on x86 by the x86open project. Please reference 1 . The binary encode of cpu0 instruction set in obj has been checked in the previous chapters. But we didnt dig into the ELF le format like elf header and relocation record at that time. This chapter will use the binutils which has been installed in sub-section Install other tools on iMac of Appendix A: Installing LLVM 2 to analysis cpu0 ELF le. You will learn the objdump, readelf, . . . , tools and understand the ELF le format itself through using these tools to analyze the cpu0 generated obj in this chapter. LLVM has the llvm-objdump tool which like objdump but its only support the native CPU. The binutils support other CPU ELF dump as a cross compiler tool chains. Linux platform has binutils already and no need to install it further. We use Linux binutils in this chapter just because iMac will display Chinese text. The iMac corresponding binutils have no problem except it use add g in command, for example, use gobjdump instead of objdump, and display your area language instead of pure English. The binutils tool we use is not a part of llvm tools, but its a powerful tool in ELF analysis. This chapter introduce the tool to readers since we think it is a valuable knowledge in this popular ELF format and the ELF binutils analysis tool. An LLVM compiler engineer has the responsibility to analyze the ELF since the obj is need to be handled by linker or loader later. With this tool, you can verify your generated ELF format. The cpu0 author has published a System Software book which introduce the topics of assembler, linker, loader, compiler and OS in concept, and at same time demonstrate how to use binutils and gcc to analysis ELF through the example code in his book. Its a Chinese book of System Software in concept and practice. This book does the real analysis through binutils. The System Software3 written by Beck is a famous book in concept of telling readers what is the compiler output, what is the linker output, what is the loader output, and how they work together. But it covers the concept only. You can reference it to understand how the Relocation Record works if you need to refresh or learning this knowledge for this chapter.
4 5 6

, , are the Chinese documents available from the cpu0 author on web site.

10.1 ELF format

ELF is a format used both in obj and executable le. So, there are two views in it as ELF le format overview.
1 2 3 4 5 6

http://en.wikipedia.org/wiki/Executable_and_Linkable_Format http://jonathan2251.github.com/lbd/install.html#install-other-tools-on-imac Leland Beck, System Software: An Introduction to Systems Programming. http://ccckmit.wikidot.com/lk:aout http://ccckmit.wikidot.com/lk:objle http://ccckmit.wikidot.com/lk:elf

165

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Figure 10.1: ELF le format overview

166

Chapter 10. ELF Support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

As ELF le format overview, the Section header table include sections .text, .rodata, . . . , .data which are sections layout for code, read only data, . . . , and read/write data. Program header table include segments include run time code and data. The denition of segments is run time layout for code and data, and sections is link time layout for code and data.

10.2 ELF header and Section header table

Lets run 7/7/Cpu0 with ch6_1.cpp, and dump ELF header information by readelf -h to see what information the ELF header contains.
[Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=obj ch6_1.bc -o ch6_1.cpu0.o [Gamma@localhost InputFiles]$ readelf -h ch6_1.cpu0.o ELF Header: Magic: 7f 45 4c 46 01 02 01 08 00 00 00 00 00 00 00 00 Class: ELF32 Data: 2s complement, big endian Version: 1 (current) OS/ABI: UNIX - IRIX ABI Version: 0 Type: REL (Relocatable file) Machine: <unknown>: 0xc9 Version: 0x1 Entry point address: 0x0 Start of program headers: 0 (bytes into file) Start of section headers: 212 (bytes into file) Flags: 0x70000001 Size of this header: 52 (bytes) Size of program headers: 0 (bytes) Number of program headers: 0 Size of section headers: 40 (bytes) Number of section headers: 10 Section header string table index: 7 [Gamma@localhost InputFiles]$ [Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=mips -relocation-model=pic -filetype=obj ch6_1.bc -o ch6_1.mips.o [Gamma@localhost InputFiles]$ readelf -h ch6_1.mips.o ELF Header: Magic: 7f 45 4c 46 01 02 01 08 00 00 00 00 00 00 00 00 Class: ELF32 Data: 2s complement, big endian Version: 1 (current) OS/ABI: UNIX - IRIX ABI Version: 0 Type: REL (Relocatable file) Machine: MIPS R3000 Version: 0x1 Entry point address: 0x0 Start of program headers: 0 (bytes into file) Start of section headers: 212 (bytes into file) Flags: 0x70000001 Size of this header: 52 (bytes) Size of program headers: 0 (bytes)

10.2. ELF header and Section header table

167

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Number of program headers: Size of section headers: Number of section headers: Section header string table index: [Gamma@localhost InputFiles]$

0 40 (bytes) 11 8

As above ELF header display, it contains information of magic number, version, ABI, . . . , . Lets check ELF segments information as follows,
[Gamma@localhost InputFiles]$ readelf -l ch6_1.cpu0.o There are no program headers in this file. [Gamma@localhost InputFiles]$

The result is in expectation because cpu0 obj is for link only, not for execute. So, the segments is empty. Check ELF sections information as follows. It contains offset and size information for every section.
[Gamma@localhost InputFiles]$ readelf -S ch6_1.cpu0.o There are 10 section headers, starting at offset 0xd4: Section Headers: [Nr] Name Type Addr Off Size ES Flg Lk Inf Al [ 0] NULL 00000000 000000 000000 00 0 0 0 [ 1] .text PROGBITS 00000000 000034 000034 00 AX 0 0 4 [ 2] .rel.text REL 00000000 000310 000018 08 8 1 4 [ 3] .data PROGBITS 00000000 000068 000004 00 WA 0 0 4 [ 4] .bss NOBITS 00000000 00006c 000000 00 WA 0 0 4 [ 5] .eh_frame PROGBITS 00000000 00006c 000028 00 A 0 0 4 [ 6] .rel.eh_frame REL 00000000 000328 000008 08 8 5 4 [ 7] .shstrtab STRTAB 00000000 000094 00003e 00 0 0 1 [ 8] .symtab SYMTAB 00000000 000264 000090 10 9 6 4 [ 9] .strtab STRTAB 00000000 0002f4 00001b 00 0 0 1 Key to Flags: W (write), A (alloc), X (execute), M (merge), S (strings) I (info), L (link order), G (group), T (TLS), E (exclude), x (unknown) O (extra OS processing required) o (OS specific), p (processor specific) [Gamma@localhost InputFiles]$

10.3 Relocation Record

The cpu0 backend translate global variable as follows,
[Gamma@localhost InputFiles]$ clang -c ch6_1.cpp -emit-llvm -o ch6_1.bc [Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=asm ch6_1.bc -o ch6_1.cpu0.s [Gamma@localhost InputFiles]$ cat ch6_1.cpu0.s .section .mdebug.abi32 .previous .file "ch6_1.bc" .text .globl main .align 2 .type main,@function .ent main # @main main: .cfi_startproc

168

Chapter 10. ELF Support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

.frame $sp,8,$lr .mask 0x00000000,0 .set noreorder .cpload $t9 ... ld $2, %got(gI)($gp) ... .type gI,@object .data .globl gI .align 2 gI: .4byte 100 .size gI, 4

# @gI

# 0x64

[Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=obj ch6_1.bc -o ch6_1.cpu0.o [Gamma@localhost InputFiles]$ objdump -s ch6_1.cpu0.o ch6_1.cpu0.o: file format elf32-big

Contents of section .text: // .cpload machine instruction 0000 09a00000 1eaa0010 09aa0000 13aa6000 ... 0020 002a0000 00220000 012d0000 09dd0008 ... [Gamma@localhost InputFiles]$ Jonathan$

............... .*..."...-......

[Gamma@localhost InputFiles]$ readelf -tr ch6_1.cpu0.o There are 10 section headers, starting at offset 0xd4: Section Headers: [Nr] Name Type Addr Flags [ 0] NULL 00000000 [00000000]: [ 1] .text PROGBITS 00000000 [00000006]: ALLOC, EXEC [ 2] .rel.text REL 00000000 [00000000]: [ 3] .data PROGBITS 00000000 [00000003]: WRITE, ALLOC [ 4] .bss NOBITS 00000000 [00000003]: WRITE, ALLOC [ 5] .eh_frame PROGBITS 00000000 [00000002]: ALLOC [ 6] .rel.eh_frame REL 00000000 [00000000]:

Off

Size

ES

Lk Inf Al

000000 000000 00

000034 000034 00

000310 000018 08

000068 000004 00

00006c 000000 00

00006c 000028 00

000328 000008 08

10.3. Relocation Record

169

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

[ 7] .shstrtab STRTAB [00000000]: [ 8] .symtab SYMTAB [00000000]: [ 9] .strtab STRTAB [00000000]: Relocation section Offset Info 00000000 00000805 00000008 00000806 00000020 00000609

00000000 000094 00003e 00

00000000 000264 000090 10

00000000 0002f4 00001b 00

.rel.text at offset 0x310 contains 3 entries: Type Sym.Value Sym. Name unrecognized: 5 00000000 _gp_disp unrecognized: 6 00000000 _gp_disp unrecognized: 9 00000000 gI

Relocation section .rel.eh_frame at offset 0x328 contains 1 entries: Offset Info Type Sym.Value Sym. Name 0000001c 00000202 unrecognized: 2 00000000 .text [Gamma@localhost InputFiles]$ readelf -tr ch6_1.mips.o There are 10 section headers, starting at offset 0xd0: Section Headers: [Nr] Name Type Addr Flags [ 0] NULL 00000000 [00000000]: [ 1] .text PROGBITS 00000000 [00000006]: ALLOC, EXEC [ 2] .rel.text REL 00000000 [00000000]: [ 3] .data PROGBITS 00000000 [00000003]: WRITE, ALLOC [ 4] .bss NOBITS 00000000 [00000003]: WRITE, ALLOC [ 5] .eh_frame PROGBITS 00000000 [00000002]: ALLOC [ 6] .rel.eh_frame REL 00000000 [00000000]: [ 7] .shstrtab STRTAB 00000000 [00000000]: [ 8] .symtab SYMTAB 00000000 [00000000]: [ 9] .strtab STRTAB 00000000 [00000000]:

Off

Size

ES

Lk Inf Al

000000 000000 00

000034 000030 00

00030c 000018 08

000064 000004 00

000068 000000 00

000068 000028 00

000324 000008 08

000090 00003e 00

000260 000090 10

0002f0 00001b 00

Relocation section .rel.text at offset 0x30c contains 3 entries:

170

Chapter 10. ELF Support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

Offset 00000000 00000004 00000018

Info Type 00000805 R_MIPS_HI16 00000806 R_MIPS_LO16 00000609 R_MIPS_GOT16

Sym.Value 00000000 00000000 00000000

Sym. Name _gp_disp _gp_disp gI

Relocation section .rel.eh_frame at offset 0x324 contains 1 entries: Offset Info Type Sym.Value Sym. Name 0000001c 00000202 R_MIPS_32 00000000 .text

As depicted in section Handle $gp register in PIC addressing mode, it translate .cpload %reg into the following.
// Lower ".cpload $reg" to // "addiu $gp, $zero, %hi(_gp_disp)" // "shl $gp, $gp, 16" // "addiu $gp, $gp, %lo(_gp_disp)" // "addu $gp, $gp, $t9"

The _gp_disp value is determined by loader. So, its undened in obj. You can nd the Relocation Records for offset 0 and 8 of .text section referred to _gp_disp value. The offset 0 and 8 of .text section are instructions addiu $gp, $zero, %hi(_gp_disp) and addiu $gp, $gp, %lo(_gp_disp) and their corresponding obj encode are 09a00000 and 09aa0000. The obj translate the %hi(_gp_disp) and %lo(_gp_disp) into 0 since when loader load this obj into memory, loader will know the _gp_disp value at run time and will update these two offset relocation records into the correct offset value. You can check the cpu0 of %hi(_gp_disp) and %lo(_gp_disp) are correct by above mips Relocation Records of R_MIPS_HI(_gp_disp) and R_MIPS_LO(_gp_disp) even though the cpu0 is not a CPU recognized by greadelf utilitly. The instruction ld $2, %got(gI)($gp) is same since we dont know what the address of .data section variable will load to. So, translate the address to 0 and made a relocation record on 0x00000020 of .text section. Loader will change this address too. Run with ch8_3_3.cpp will get the unknown result in _Z5sum_iiz and other symbol reference as below. Loader or linker will take care them according the relocation records compiler generated.
[Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=cpu0 -relocation-model=pic -filetype=obj ch8_3_3.bc -o ch8_3__3. cpu0.o [Gamma@localhost InputFiles]$ readelf -tr ch8_3_3.cpu0.o There are 11 section headers, starting at offset 0x248: Section Headers: [Nr] Name Type Addr Flags [ 0] NULL 00000000 [00000000]: [ 1] .text PROGBITS 00000000 [00000006]: ALLOC, EXEC [ 2] .rel.text REL 00000000 [00000000]: [ 3] .data PROGBITS 00000000 [00000003]: WRITE, ALLOC [ 4] .bss NOBITS 00000000 [00000003]: WRITE, ALLOC [ 5] .rodata.str1.1 PROGBITS 00000000

Off

Size

ES

Lk Inf Al

000000 000000 00

000034 000178 00

000538 000058 08

0001ac 000000 00

0001ac 000000 00

0001ac 000008 01

10.3. Relocation Record

171

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

[00000032]: ALLOC, MERGE, STRINGS [ 6] .eh_frame PROGBITS 00000000 0001b4 000044 [00000002]: ALLOC [ 7] .rel.eh_frame REL 00000000 000590 000010 [00000000]: [ 8] .shstrtab STRTAB 00000000 0001f8 00004d [00000000]: [ 9] .symtab SYMTAB 00000000 000400 0000e0 [00000000]: [10] .strtab STRTAB 00000000 0004e0 000055 [00000000]: Relocation section Offset Info 00000000 00000c05 00000008 00000c06 0000001c 00000b09 000000b8 00000b09 000000dc 00000a0b 000000e8 00000c05 000000f0 00000c06 00000140 0000080b 00000154 00000209 00000158 00000206 00000160 00000d0b

00

08

00

10

10

00

.rel.text at offset 0x538 contains 11 entries: Type Sym.Value Sym. Name unrecognized: 5 00000000 _gp_disp unrecognized: 6 00000000 _gp_disp unrecognized: 9 00000000 __stack_chk_guard unrecognized: 9 00000000 __stack_chk_guard unrecognized: b 00000000 __stack_chk_fail unrecognized: 5 00000000 _gp_disp unrecognized: 6 00000000 _gp_disp unrecognized: b 00000000 _Z5sum_iiz unrecognized: 9 00000000 $.str unrecognized: 6 00000000 $.str unrecognized: b 00000000 printf

Relocation section .rel.eh_frame at offset 0x590 contains 2 entries: Offset Info Type Sym.Value Sym. Name 0000001c 00000302 unrecognized: 2 00000000 .text 00000034 00000302 unrecognized: 2 00000000 .text [Gamma@localhost InputFiles]$ /usr/local/llvm/3.1.test/cpu0/1/cmake_debug_build/ bin/llc -march=mips -relocation-model=pic -filetype=obj ch8_3_3.bc -o ch8_3__3. mips.o [Gamma@localhost InputFiles]$ readelf -tr ch8_3_3.mips.o There are 11 section headers, starting at offset 0x254: Section Headers: [Nr] Name Type Addr Flags [ 0] NULL 00000000 [00000000]: [ 1] .text PROGBITS 00000000 [00000006]: ALLOC, EXEC [ 2] .rel.text REL 00000000 [00000000]: [ 3] .data PROGBITS 00000000 [00000003]: WRITE, ALLOC [ 4] .bss

Off

Size

ES

Lk Inf Al

000000 000000 00

000034 000184 00

000544 000058 08

0001b8 000000 00

172

Chapter 10. ELF Support

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

NOBITS 00000000 0001b8 000000 [00000003]: WRITE, ALLOC [ 5] .rodata.str1.1 PROGBITS 00000000 0001b8 000008 [00000032]: ALLOC, MERGE, STRINGS [ 6] .eh_frame PROGBITS 00000000 0001c0 000044 [00000002]: ALLOC [ 7] .rel.eh_frame REL 00000000 00059c 000010 [00000000]: [ 8] .shstrtab STRTAB 00000000 000204 00004d [00000000]: [ 9] .symtab SYMTAB 00000000 00040c 0000e0 [00000000]: [10] .strtab STRTAB 00000000 0004ec 000055 [00000000]: Relocation section Offset Info 00000000 00000c05 00000004 00000c06 00000024 00000b09 000000c8 00000b09 000000f0 00000a0b 00000100 00000c05 00000104 00000c06 00000134 0000080b 00000154 00000209 00000158 00000206 0000015c 00000d0b

00

01

00

08

00

10

10

00

.rel.text at offset 0x544 contains 11 entries: Type Sym.Value Sym. Name R_MIPS_HI16 00000000 _gp_disp R_MIPS_LO16 00000000 _gp_disp R_MIPS_GOT16 00000000 __stack_chk_guard R_MIPS_GOT16 00000000 __stack_chk_guard R_MIPS_CALL16 00000000 __stack_chk_fail R_MIPS_HI16 00000000 _gp_disp R_MIPS_LO16 00000000 _gp_disp R_MIPS_CALL16 00000000 _Z5sum_iiz R_MIPS_GOT16 00000000 $.str R_MIPS_LO16 00000000 $.str R_MIPS_CALL16 00000000 printf

Relocation section .rel.eh_frame at offset 0x59c contains 2 entries: Offset Info Type Sym.Value Sym. Name 0000001c 00000302 R_MIPS_32 00000000 .text 00000034 00000302 R_MIPS_32 00000000 .text [Gamma@localhost InputFiles]$

10.4 Cpu0 ELF related les

Files Cpu0ELFObjectWrite.cpp and Cpu0MC*.cpp are the les take care the obj format. Most obj code translation are dened by Cpu0InstrInfo.td and Cpu0RegisterInfo.td. With these td description, LLVM translate the instruction into obj format automatically.

10.5 lld
The lld is a project of LLVM linker. Its under development and we cannot nish the installation by following the web site direction. Even with this, its really make sense to develop a new linker according its web site information. Please visit the web site 7 .
7

http://lld.llvm.org/

10.4. Cpu0 ELF related les

173

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

174

Chapter 10. ELF Support

CHAPTER

ELEVEN

PORTING TO LLVM 3.2

LLVM version 3.2 has released in December, 2012. Just a few API which Cpu0 used is changes. 9/1/Cpu0 which support 3.2 reect these changes. You can check these changes by compare tools. In addition to highlight these changes, this chapter also explains the Mips backend changes in 3.2. Todo The changes made to port from LLVM 3.1->3.2 should be reected throughout the rest of the book. We should assume the reader is using the current release of LLVM when working with this book.

11.1 API changes in LLVM 3.2

List API changes as follows, 1. The parameters of call back function for Target Registration is changed by add parameter MCRegisterInfo in the callback function for RegisterMCCodeEmitter() and StringRef in the callback function for RegisterMCAsmBackend. In other word, you can get more information of registers and CPU (type of StringRef) for your backend after this registration. Of course, these information came from TabGen which source is the Target Description .td you write.
extern "C" void LLVMInitializeCpu0TargetMC() { ... // Register the MC Code Emitter TargetRegistry::RegisterMCCodeEmitter(TheCpu0Target, createCpu0MCCodeEmitterEB); TargetRegistry::RegisterMCCodeEmitter(TheCpu0elTarget, createCpu0MCCodeEmitterEL); ... // Register the asm backend. TargetRegistry::RegisterMCAsmBackend(TheCpu0Target, createCpu0AsmBackendEB32); TargetRegistry::RegisterMCAsmBackend(TheCpu0elTarget, createCpu0AsmBackendEL32); ... }

Version 3.1 as follows,

MCCodeEmitter *createCpu0MCCodeEmitterEB(const MCInstrInfo &MCII, const MCSubtargetInfo &STI, MCContext &Ctx); MCCodeEmitter *createCpu0MCCodeEmitterEL(const MCInstrInfo &MCII,

175

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

const MCSubtargetInfo &STI, MCContext &Ctx); MCAsmBackend *createCpu0AsmBackendEB32(const Target &T, StringRef TT); MCAsmBackend *createCpu0AsmBackendEL32(const Target &T, StringRef TT);

Version 3.2 as follows,

MCCodeEmitter *createCpu0MCCodeEmitterEB(const MCInstrInfo &MCII, const MCRegisterInfo &MRI, const MCSubtargetInfo &STI, MCContext &Ctx); MCCodeEmitter *createCpu0MCCodeEmitterEL(const MCInstrInfo &MCII, const MCRegisterInfo &MRI, const MCSubtargetInfo &STI, MCContext &Ctx); MCAsmBackend *createCpu0AsmBackendEB32(const Target &T, StringRef TT, StringRef CPU); MCAsmBackend *createCpu0AsmBackendEL32(const Target &T, StringRef TT, StringRef CPU);

2. Change the name from CPURegsRegisterClass to CPURegsRegClass. The source of register class information came from your backend <register>.td. The new name CPURegsRegClass is call by reference type in C++ while the old CPURegsRegisterClass is pointer type. The reference type use . while pointer type use -> as follows,
unsigned CPURegSize = Cpu0::CPURegsRegClass.getSize(); unsigned CPURegSize = Cpu0::CPURegsRegisterClass->getSize();

3. Change LowerCall() parameters as follows, Version 3.1 as follows,

SDValue LowerCall(SDValue Chain, SDValue Callee, CallingConv::ID CallConv, bool isVarArg, bool doesNotRet, bool &isTailCall, const SmallVectorImpl<ISD::OutputArg> &Outs, const SmallVectorImpl<SDValue> &OutVals, const SmallVectorImpl<ISD::InputArg> &Ins, DebugLoc dl, SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const;

Version 3.2 as follows,

LowerCall(TargetLowering::CallLoweringInfo &CLI, SmallVectorImpl<SDValue> &InVals) const;

The TargetLowering::CallLoweringInfo is type of structure/class which contains the old version 3.1 parameters. You can get the 3.1 same information by,
SDValue Cpu0TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI, SmallVectorImpl<SDValue> &InVals) const { SelectionDAG &DAG = CLI.DAG; DebugLoc &dl = CLI.DL; SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs; SmallVector<SDValue, 32> &OutVals = CLI.OutVals; SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;

176

Chapter 11. Porting to LLVM 3.2

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

SDValue InChain SDValue Callee bool &isTailCall CallingConv::ID CallConv bool isVarArg ... }

= = = = =

CLI.Chain; CLI.Callee; CLI.IsTailCall; CLI.CallConv; CLI.IsVarArg;

As chapter function call, LowerCall() handle the outgoing arguments passing in function call. 4. Above changes is mandatory. There are some changes are adviced to follow. Like the below. We comment the Change Reason in the following code. You can get the Change Reason by internet searching.
MCObjectWriter *createObjectWriter(raw_ostream &OS) const { // Change Reason: // Reduce the exposure of Triple::OSType in the ELF object writer. This will // avoid including ADT/Triple.h in many places when the target specific bits // are moved. return createCpu0ELFObjectWriter(OS, MCELFObjectTargetWriter::getOSABI(OSType), IsLittle); // Even though, the old function still work on LLVM version 3.2 // return createCpu0ELFObjectWriter(OS, OSType, IsLittle); } class Cpu0MCCodeEmitter : public MCCodeEmitter { // #define LLVM_DELETED_FUNCTION // LLVM_DELETED_FUNCTION - Expands to = delete if the compiler supports it. // Use to mark functions as uncallable. Member functions with this should be // declared private so that some behavior is kept in C++03 mode. // class DontCopy { private: DontCopy(const DontCopy&) LLVM_DELETED_FUNCTION; // DontCopy &operator =(const DontCopy&) LLVM_DELETED_FUNCTION; public: ... }; // Definition at line 79 of file Compiler.h. Cpu0MCCodeEmitter(const Cpu0MCCodeEmitter &) LLVM_DELETED_FUNCTION; void operator=(const Cpu0MCCodeEmitter &) LLVM_DELETED_FUNCTION; // Even though, the old function still work on LLVM version 3.2 // Cpu0MCCodeEmitter(const Cpu0MCCodeEmitter &); // DO NOT IMPLEMENT // void operator=(const Cpu0MCCodeEmitter &); // DO NOT IMPLEMENT ...

11.2 Verify the Cpu0 for 3.2

3.2_src_les_modify include the LLVM 3.2 original les and those les modied for Cpu0 backend support. Please update the LLVM 3.2 as Appendix A indicate (just copy 3.2_src_les_modify/src_les_modify/src to your LLVM 3.2 source directory). 9/1/Cpu0 is the code for LLVM version 3.2. File ch_all.cpp include the all C/C++ operators, global variable, struct, array, control statement and function call test. Run 9/1/Cpu0 with ch_all.cpp will get the assembly code as below. By compare it with the output of 3.1 result, you can verify the correction as below. The difference is came from 3.2 correcting the label number for order.
//#include <stdio.h> #include <stdarg.h> #include <stdlib.h> int test_operators() { int a = 5;

11.2. Verify the Cpu0 for 3.2

177

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

int b = 2; int c = 0; int d = 0; int e, f, g, h, i, j, k, l = 0; unsigned int a1 = -5, k1 = 0, f1 = 0; c = a + b; d = a - b; e = a * b; f = a / b; f1 = a1 / b; g = (a & b); h = (a | b); i = (a ^ b); j = (a << 2); int j1 = (a1 << 2); k = (a >> 2); k1 = (a1 >> 2); b = !a; int* p = &b; b = (b+1)%a; c = rand(); b = (b+1)%c; return c; } int gI = 100; int test_globalvar() { int c = 0; c = gI; return c; } struct Date { int year; int month; int day; }; Date date = {2012, 10, 12}; int a[3] = {2012, 10, 12}; int test_struct() { int day = date.day; int i = a[1]; return 0; } template<class T>

178

Chapter 11. Porting to LLVM 3.2

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

T sum(T { T i = T val T sum

amount, ...) 0; = 0; = 0;

va_list vl; va_start(vl, amount); for (i = 0; i < amount; i++) { val = va_arg(vl, T); sum += val; } va_end(vl); return sum; } int main() { test_operators(); int a = sum<int>(6, 1, 2, 3, 4, 5, 6); // printf("a = %d\n", a); return a; } 118-165-78-60:InputFiles Jonathan$ diff ch_all.3.1.cpu0.s ch_all.3.2.cpu0.s 262c262 < jge $BB4_7 --> jge $BB4_6 285d284 < # BB#6: # in Loop: Header=BB4_1 Depth=1 290c289 < $BB4_7: --> $BB4_6: 295,297c294,296 < jne $BB4_9 < jmp $BB4_8 < $BB4_8: # %SP_return --> jne $BB4_8 > jmp $BB4_7 > $BB4_7: # %SP_return 301c300 < $BB4_9: # %CallStackCheckFailBlk --# %CallStackCheckFailBlk > $BB4_8: // ch_all.3.2.cpu0.s ... $BB4_5: ld $3, 0($3) st $3, 36($sp) ld $4, 32($sp) add $3, $4, $3

in Loop: Header=BB4_1 Depth=1

11.2. Verify the Cpu0 for 3.2

179

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

st $3, 32($sp) ld $3, 40($sp) addiu $3, $3, 1 st $3, 40($sp) jmp $BB4_1 $BB4_6: ld $2, %got(__stack_chk_guard)($gp) ld $2, 0($2) ld $3, 48($sp) cmp $2, $3 jne $BB4_8 jmp $BB4_7 $BB4_7: ...

# %SP_return

// ch_all.3.2.cpu0.s ... $BB4_5: ld $3, 0($3) st $3, 36($sp) ld $4, 32($sp) add $3, $4, $3 st $3, 32($sp) # BB#6: ld $3, 40($sp) addiu $3, $3, 1 st $3, 40($sp) jmp $BB4_1 $BB4_7: ld $2, %got(__stack_chk_guard)($gp) ld $2, 0($2) ld $3, 48($sp) cmp $2, $3 jne $BB4_9 jmp $BB4_8 $BB4_8: ...

in Loop: Header=BB4_1 Depth=1

in Loop: Header=BB4_1 Depth=1

# %SP_return

11.3 Mips backend changes in 3.2

In 3.1, Mips use .cpload and .cprestore pseudo assembly code. It removes these pseudo assembly code in 3.2. This change is good for spim (mips assembly code simulator) which run for Mips assembly code. According the theory of System Software, some pseudo assembly code (especially for those not in standard) cannot be translated by assembler. It will break out assembly code simulator. To be continue...

180

Chapter 11. Porting to LLVM 3.2

CHAPTER

TWELVE

TODO LIST
Todo Add info about LLVM documentation licensing. (The original entry is located in /z/tmp/github/lbd/source/about.rst, line 53.) Todo The changes made to port from LLVM 3.1->3.2 should be reected throughout the rest of the book. We should assume the reader is using the current release of LLVM when working with this book. (The original entry is located in /z/tmp/github/lbd/source/port3.2.rst, line 13.) Todo Fix centering for gure captions. (The original entry is located in /z/tmp/github/lbd/source/appendix.rst, line 11.) Todo Should we just write out commands in a terminal for people to execute? (The original entry is located in /z/tmp/github/lbd/source/appendix.rst, line 20.) Todo The html will follow the appear order in *.rst source context but latexpdf didnt. For example, the Create LLVM.xcodeproj by cmake Set option to generate Xcode project Figure 2.4 and Create LLVM.xcodeproj by cmake Before Adjust CMAKE_INSTALL_NAME_TOOL Figure 2.5 appear after the below text Click OK from ... in pdf. If nd the NoReorder or newpage directive, maybe can solve this problem. (The original entry is located in /z/tmp/github/lbd/source/appendix.rst, line 89.)

181

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

182

Chapter 12. Todo List

CHAPTER

THIRTEEN

BOOK EXAMPLE CODE

The example code is available in: http://jonathan2251.github.com/lbd/LLVMBackendTutorialExampleCode.tar.gz

183

Write An LLVM Backend Tutorial For Cpu0, Release 3.2.3

184

Chapter 13. Book example code

CHAPTER

FOURTEEN

ALTERNATE FORMATS
The book is also available in the following formats:

185