COMPUTER ORGANIZATION AND DESIGN

RISC-V
Edition
The Hardware/Software Interface

Module 9
The Processor Pipelining II

Instructor: Dr. Yujie Tang

The lecture slides are based on the slides by the publisher, but modified and
corrected by the instructor.
 §4.7 Data Hazards: Forwarding vs. Stalling
Data Hazards in ALU Instructions
 Consider this sequence:
sub x2, x1,x3
and x12,x2,x5
or x13,x6,x2
add x14,x2,x2
sd x15,100(x2)

 We can resolve hazards with forwarding

 How do we detect when to forward?

Chapter 4 — The Processor — 2

Dependencies & Forwarding

Chapter 4 — The Processor — 3

Detecting the Need to Forward
 Pass register numbers along pipeline
 e.g., ID/EX.RegisterRs1 = register number for Rs1
sitting in ID/EX pipeline register
 ALU operand register numbers in EX stage
are given by
 ID/EX.RegisterRs1, ID/EX.RegisterRs2
 Data hazards when
Fwd from
1a. EX/MEM.RegisterRd = ID/EX.RegisterRs1 EX/MEM
pipeline reg
1b. EX/MEM.RegisterRd = ID/EX.RegisterRs2
2a. MEM/WB.RegisterRd = ID/EX.RegisterRs1 Fwd from
MEM/WB
2b. MEM/WB.RegisterRd = ID/EX.RegisterRs2 pipeline reg

Chapter 4 — The Processor — 4

Detecting the Need to Forward
 But only if forwarding instruction will write
to a register!
 EX/MEM.RegWrite, MEM/WB.RegWrite

 And only if Rd for that instruction is not x0

 EX/MEM.RegisterRd ≠ 0,
MEM/WB.RegisterRd ≠ 0

Chapter 4 — The Processor — 5

Forwarding Paths

Chapter 4 — The Processor — 6

 Forwarding Conditions
Mux control Source Explanation
ForwardA = 00 ID/EX The first ALU operand comes from the register file.
ForwardA = 10 EX/MEM The first ALU operand is forwarded from the prior
ALU result.
ForwardA = 01 MEM/WB The first ALU operand is forwarded from data
memory or an earlier ALU result.

ForwardB = 00 ID/EX The second ALU operand comes from the register
file.
ForwardB = 10 EX/MEM The second ALU operand is forwarded from the prior
ALU result.
ForwardB = 01 MEM/WB The second ALU operand is forwarded from data
memory or an earlier ALU result.

Chapter 4 — The Processor — 7

Forwarding Conditions
 EX hazard
 if (EX/MEM.RegWrite and (EX/MEM.RegisterRd ≠ 0)
and (EX/MEM.RegisterRd = ID/EX.RegisterRs))
ForwardA = 10
 if (EX/MEM.RegWrite and (EX/MEM.RegisterRd ≠ 0)
and (EX/MEM.RegisterRd = ID/EX.RegisterRt))
ForwardB = 10
 MEM hazard
 if (MEM/WB.RegWrite and (MEM/WB.RegisterRd ≠ 0)
and (MEM/WB.RegisterRd = ID/EX.RegisterRs))
ForwardA = 01
 if (MEM/WB.RegWrite and (MEM/WB.RegisterRd ≠ 0)
and (MEM/WB.RegisterRd = ID/EX.RegisterRt))
ForwardB = 01

Chapter 4 — The Processor — 8

Double Data Hazard
 Consider the sequence:
add x1,x1,x2
add x1,x1,x3
add x1,x1,x4
 Both hazards occur
 Want to use the most recent
 Revise MEM hazard condition
 Only fwd if EX hazard condition isn’t true

Chapter 4 — The Processor — 9

Revised Forwarding Condition
 MEM hazard
 if (MEM/WB.RegWrite
and (MEM/WB.RegisterRd ≠ 0)
and not(EX/MEM.RegWrite and (EX/MEM.RegisterRd ≠ 0)
and (EX/MEM.RegisterRd ≠ ID/EX.RegisterRs1))
and (MEM/WB.RegisterRd = ID/EX.RegisterRs1)) ForwardA = 01
 if (MEM/WB.RegWrite
and (MEM/WB.RegisterRd ≠ 0)
and not(EX/MEM.RegWrite and (EX/MEM.RegisterRd ≠ 0)
and (EX/MEM.RegisterRd ≠ ID/EX.RegisterRs2))
and (MEM/WB.RegisterRd = ID/EX.RegisterRs2)) ForwardB = 01

Chapter 4 — The Processor — 10

Datapath with Forwarding

Chapter 4 — The Processor — 11

Load-Use Hazard Detection
 Check when using instruction is decoded
in ID stage
 ALU operand register numbers in ID stage
are given by
 IF/ID.RegisterRs1, IF/ID.RegisterRs2
 Load-use hazard when
 ID/EX.MemRead and
((ID/EX.RegisterRd = IF/ID.RegisterRs1) or
(ID/EX.RegisterRd = IF/ID.RegisterRs2))
 If detected, stall and insert bubble

Chapter 4 — The Processor — 12

How to Stall the Pipeline
 Force control values in ID/EX register
to 0
 EX, MEM and WB do nop (no-operation)
 Prevent update of PC and IF/ID register
 Using instruction is decoded again
 Following instruction is fetched again
 1-cycle stall allows MEM to read data for ld
 Can subsequently forward to EX stage

Chapter 4 — The Processor — 13

Load-Use Data Hazard

Stall inserted
here

Chapter 4 — The Processor — 14

Datapath with Hazard Detection

Chapter 4 — The Processor — 15

Stalls and Performance
The BIG Picture
 Stalls reduce performance
 But are required to get correct results
 Compiler can arrange code to avoid
hazards and stalls
 Requires knowledge of the pipeline structure

Chapter 4 — The Processor — 16

 §4.8 Control Hazards
Branch Hazards
 If branch outcome determined in MEM

Flush these
instructions
(Set control
values to 0)

PC

Chapter 4 — The Processor — 17

Reducing Branch Delay
 Move hardware to determine outcome to ID stage
 Target address adder
 Register comparator
 Example: branch taken
36: sub x10, x4, x8
40: beq x1, x3, 16 // PC-relative branch
// to 40+16*2=72
44: and x12, x2, x5
48: orr x13, x2, x6
52: add x14, x4, x2
56: sub x15, x6, x7
...
72: ld x4, 50(x7)

Chapter 4 — The Processor — 18

Example: Branch Taken

Chapter 4 — The Processor — 19

Example: Branch Taken

Chapter 4 — The Processor — 20

Dynamic Branch Prediction
 In deeper and superscalar pipelines, branch
penalty is more significant
 Use dynamic prediction
 Branch prediction buffer (aka branch history table)
 Indexed by recent branch instruction addresses
 Stores outcome (taken/not taken)
 To execute a branch
 Check table, expect the same outcome
 Start fetching from fall-through or target
 If wrong, flush pipeline and flip prediction

Chapter 4 — The Processor — 21

1-Bit Predictor: Shortcoming
 Inner loop branches mispredicted twice!
outer: …
…
inner: …
…
beq …, …, inner
…
beq …, …, outer

 Mispredict as taken on last iteration of inner

loop
 Then mispredict as not taken on first
iteration of inner loop next time around
Chapter 4 — The Processor — 22
2-Bit Predictor
 Only change prediction on two successive
mispredictions

Chapter 4 — The Processor — 23

Calculating the Branch Target
 Even with predictor, still need to calculate
the target address
 1-cycle penalty for a taken branch
 Branch target buffer
 Cache of target addresses
 Indexed by PC when instruction fetched
 If hit and instruction is branch predicted taken, can
fetch target immediately

Chapter 4 — The Processor — 24

 §4.9 Exceptions
Exceptions and Interrupts
 “Unexpected” events requiring change
in flow of control
 Different ISAs use the terms differently
 Exception
 Arises within the CPU
 e.g., undefined opcode, syscall, …
 Interrupt
 From an external I/O controller
 Dealing with them without sacrificing
performance is hard

Chapter 4 — The Processor — 25

Handling Exceptions
 Save PC of offending (or interrupted) instruction
 In RISC-V: Supervisor Exception Program Counter
(SEPC)

 Save indication of the problem

 In RISC-V: Supervisor Exception Cause Register
(SCAUSE)
 64 bits, but most bits unused
 Exception code field: 2 for undefined opcode, 12 for
hardware malfunction, …
 Jump to handler
 Assume at 0000 0000 1C09 0000hex

Chapter 4 — The Processor — 26

An Alternate Mechanism
 Vectored Interrupts
 Handler address determined by the cause
 Exception vector address to be added to a
vector table base register:
 Undefined opcode 00 0100 0000two
 Hardware malfunction: 01 1000 0000two
 …: …
 Instructions either
 Deal with the interrupt, or
 Jump to real handler
Chapter 4 — The Processor — 27
Handler Actions
 Read cause, and transfer to relevant
handler
 Determine action required
 If restartable
 Take corrective action
 use SEPC to return to program
 Otherwise
 Terminate program
 Report error using SEPC, SCAUSE, …

Chapter 4 — The Processor — 28

Exceptions in a Pipeline
 Another form of control hazard
 Consider malfunction on add in EX stage
add x1, x2, x1
 Prevent x1 from being clobbered

 Complete previous instructions

 Flush add and subsequent instructions

 Set SEPC and SCAUSE register values

 Transfer control to handler

 Similar to mispredicted branch

 Use much of the same hardware

Chapter 4 — The Processor — 29

Pipeline with Exceptions

Chapter 4 — The Processor — 30

Exception Properties
 Restartable exceptions
 Pipeline can flush the instruction
 Handler executes, then returns to the
instruction
 Refetched and executed from scratch
 PC saved in SEPC register
 Identifies causing instruction

Chapter 4 — The Processor — 31

Exception Example
 Exception on add in
40 sub x11, x2, x4
44 and x12, x2, x5
48 orr x13, x2, x6
4c add x1, x2, x1
50 sub x15, x6, x7
54 ld x16, 100(x7)
…
 Handler
1C090000 sd x26, 1000(x10)
1c090004 sd x27, 1008(x10)
…

Chapter 4 — The Processor — 32

Exception Example

Chapter 4 — The Processor — 33

Exception Example

Chapter 4 — The Processor — 34

Multiple Exceptions
 Pipelining overlaps multiple instructions
 Could have multiple exceptions at once
 Simple approach: deal with exception from
earliest instruction
 Flush subsequent instructions
 “Precise” exceptions
 In complex pipelines
 Multiple instructions issued per cycle
 Out-of-order completion
 Maintaining precise exceptions is difficult!

Chapter 4 — The Processor — 35

Imprecise Exceptions
 Just stop pipeline and save state
 Including exception cause(s)
 Let the handler work out
 Which instruction(s) had exceptions
 Which to complete or flush
 May require “manual” completion
 Simplifies hardware, but more complex handler
software
 Not feasible for complex multiple-issue
out-of-order pipelines

Chapter 4 — The Processor — 36

 §4.10 Parallelism via Instructions
Instruction-Level Parallelism (ILP)
 Pipelining: executing multiple instructions in
parallel
 To increase ILP
 Deeper pipeline
 Less work per stage  shorter clock cycle
 Multiple issue
 Replicate pipeline stages  multiple pipelines
 Start multiple instructions per clock cycle
 CPI < 1, so use Instructions Per Cycle (IPC)
 E.g., 4GHz 4-way multiple-issue
 16 BIPS, peak CPI = 0.25, peak IPC = 4
 But dependencies reduce this in practice

Chapter 4 — The Processor — 37

Multiple Issue
 Static multiple issue
 Compiler groups instructions to be issued together
 Packages them into “issue slots”
 Compiler detects and avoids hazards
 Dynamic multiple issue
 CPU examines instruction stream and chooses
instructions to issue each cycle
 Compiler can help by reordering instructions
 CPU resolves hazards using advanced techniques at
runtime

Chapter 4 — The Processor — 38

Speculation
 “Guess” what to do with an instruction
 Start operation as soon as possible
 Check whether guess was right
 If so, complete the operation
 If not, roll-back and do the right thing
 Common to static and dynamic multiple issue
 Examples
 Speculate on branch outcome
 Roll back if path taken is different
 Speculate on load
 Roll back if location is updated

Chapter 4 — The Processor — 39

Compiler/Hardware Speculation
 Compiler can reorder instructions
 e.g., move load before branch
 Can include “fix-up” instructions to recover
from incorrect guess
 Hardware can look ahead for instructions
to execute
 Buffer results until it determines they are
actually needed
 Flush buffers on incorrect speculation

Chapter 4 — The Processor — 40

Speculation and Exceptions
 What if exception occurs on a
speculatively executed instruction?
 e.g., speculative load before null-pointer
check
 Static speculation
 Can add ISA support for deferring exceptions
 Dynamic speculation
 Can buffer exceptions until instruction
completion (which may not occur)

Chapter 4 — The Processor — 41

Static Multiple Issue
 Compiler groups instructions into “issue
packets”
 Group of instructions that can be issued on a
single cycle
 Determined by pipeline resources required
 Think of an issue packet as a very long
instruction
 Specifies multiple concurrent operations
  Very Long Instruction Word (VLIW)

Chapter 4 — The Processor — 42

Scheduling Static Multiple Issue
 Compiler must remove some/all hazards
 Reorder instructions into issue packets
 No dependencies with a packet
 Possibly some dependencies between
packets
 Varies between ISAs; compiler must know!
 Pad with nop if necessary

Chapter 4 — The Processor — 43

RISC-V with Static Dual Issue
 Two-issue packets
 One ALU/branch instruction
 One load/store instruction
 64-bit aligned
 ALU/branch, then load/store
 Pad an unused instruction with nop

Address Instruction type Pipeline Stages

n ALU/branch IF ID EX MEM WB
n+4 Load/store IF ID EX MEM WB
n+8 ALU/branch IF ID EX MEM WB
n + 12 Load/store IF ID EX MEM WB
n + 16 ALU/branch IF ID EX MEM WB
n + 20 Load/store IF ID EX MEM WB

Chapter 4 — The Processor — 44

RISC-V with Static Dual Issue

Chapter 4 — The Processor — 45

Hazards in the Dual-Issue RISC-V
 More instructions executing in parallel
 EX data hazard
 Forwarding avoided stalls with single-issue
 Now can’t use ALU result in load/store in same packet
 add x10, x0, x1
ld x2, 0(x10)
 Split into two packets, effectively a stall
 Load-use hazard
 Still one cycle use latency, but now two instructions
 More aggressive scheduling required

Chapter 4 — The Processor — 46

Scheduling Example
 Schedule this for dual-issue RISC-V
Loop: ld x31,0(x20) // x31=array element
add x31,x31,x21 // add scalar in x21
sd x31,0(x20) // store result
addi x20,x20,-8 // decrement pointer
blt x22,x20,Loop // branch if x22 < x20

ALU/branch Load/store cycle

Loop: nop ld x31,0(x20) 1
addi x20,x20,-8 nop 2
add x31,x31,x21 nop 3
blt x22,x20,Loop sd x31,8(x20) 4

 IPC = 5/4 = 1.25 (c.f. peak IPC = 2)

Chapter 4 — The Processor — 47
Loop Unrolling
 Replicate loop body to expose more
parallelism
 Reduces loop-control overhead
 Use different registers per replication
 Called “register renaming”
 Avoid loop-carried “anti-dependencies”
 Store followed by a load of the same register
 Aka “name dependence”
 Reuse of a register name

Chapter 4 — The Processor — 48

Loop Unrolling Example
ALU/branch Load/store cycle
Loop: addi x20,x20,-32 ld x28, 0(x20) 1
nop ld x29, 24(x20) 2
add x28,x28,x21 ld x30, 16(x20) 3
add x29,x29,x21 ld x31, 8(x20) 4
add x30,x30,x21 sd x28, 32(x20) 5
add x31,x31,x21 sd x29, 24(x20) 6
nop sd x30, 16(x20) 7
blt x22,x20,Loop sd x31, 8(x20) 8

 IPC = 14/8 = 1.75

 Closer to 2, but at cost of registers and code size

Chapter 4 — The Processor — 49

Dynamic Multiple Issue
 “Superscalar” processors
 CPU decides whether to issue 0, 1, 2, …
each cycle
 Avoiding structural and data hazards
 Avoids the need for compiler scheduling
 Though it may still help
 Code semantics ensured by the CPU

Chapter 4 — The Processor — 50

Dynamic Pipeline Scheduling
 Allow the CPU to execute instructions out
of order to avoid stalls
 But commit result to registers in order
 Example
ld x31,20(x21)
add x1,x31,x2
sub x23,x23,x3
andi x5,x23,20
 Can start sub while add is waiting for ld

Chapter 4 — The Processor — 51

 Dynamically Scheduled CPU
Preserves
dependencies

Hold pending
operands

Results also sent to

any waiting
reservation stations

Reorders buffer for

register writes
Can supply
operands for
issued instructions

Chapter 4 — The Processor — 52

Register Renaming
 Reservation stations and reorder buffer
effectively provide register renaming
 On instruction issue to reservation station
 If operand is available in register file or reorder
buffer
 Copied to reservation station
 No longer required in the register; can be
overwritten
 If operand is not yet available
 It will be provided to the reservation station by a
function unit
 Register update may not be required
Chapter 4 — The Processor — 53
Speculation
 Predict branch and continue issuing
 Don’t commit until branch outcome
determined
 Load speculation
 Avoid load and cache miss delay
 Predict the effective address
 Predict loaded value
 Load before completing outstanding stores
 Bypass stored values to load unit
 Don’t commit load until speculation cleared

Chapter 4 — The Processor — 54

Why Do Dynamic Scheduling?
 Why not just let the compiler schedule
code?
 Not all stalls are predicable
 e.g., cache misses
 Can’t always schedule around branches
 Branch outcome is dynamically determined
 Different implementations of an ISA have
different latencies and hazards

Chapter 4 — The Processor — 55

Does Multiple Issue Work?
The BIG Picture
 Yes, but not as much as we’d like
 Programs have real dependencies that limit ILP
 Some dependencies are hard to eliminate
 e.g., pointer aliasing
 Some parallelism is hard to expose
 Limited window size during instruction issue
 Memory delays and limited bandwidth
 Hard to keep pipelines full
 Speculation can help if done well
Chapter 4 — The Processor — 56
Power Efficiency
 Complexity of dynamic scheduling and
speculations requires power
 Multiple simpler cores may be better
Microprocessor Year Clock Rate Pipeline Issue Out-of-order/ Cores Power
Stages width Speculation
i486 1989 25MHz 5 1 No 1 5W
Pentium 1993 66MHz 5 2 No 1 10W
Pentium Pro 1997 200MHz 10 3 Yes 1 29W
P4 Willamette 2001 2000MHz 22 3 Yes 1 75W
P4 Prescott 2004 3600MHz 31 3 Yes 1 103W
Core 2006 2930MHz 14 4 Yes 2 75W
UltraSparc III 2003 1950MHz 14 4 No 1 90W
UltraSparc T1 2005 1200MHz 6 1 No 8 70W

Chapter 4 — The Processor — 57

 §4.11 Real Stuff: The ARM Cortex-A53 and Intel Core i7 Pipelines
Cortex A53 and Intel i7
Processor ARM A53 Intel Core i7 920
Market Personal Mobile Device Server, cloud
Thermal design power 100 milliWatts 130 Watts
(1 core @ 1 GHz)
Clock rate 1.5 GHz 2.66 GHz
Cores/Chip 4 (configurable) 4
Floating point? Yes Yes
Multiple issue? Dynamic Dynamic
Peak instructions/clock cycle 2 4
Pipeline stages 8 14
Pipeline schedule Static in-order Dynamic out-of-order
with speculation
Branch prediction Hybrid 2-level
1st level caches/core 16-64 KiB I, 16-64 KiB D 32 KiB I, 32 KiB D
2nd level caches/core 128-2048 KiB 256 KiB (per core)
3rd level caches (shared) (platform dependent) 2-8 MB

Chapter 4 — The Processor — 58

ARM Cortex-A53 Pipeline

Chapter 4 — The Processor — 59

ARM Cortex-A53 Performance

Chapter 4 — The Processor — 60

Core i7 Pipeline

Chapter 4 — The Processor — 61

Core i7 Performance

Chapter 4 — The Processor — 62

 §4.12 Instruction-Level Parallelism and Matrix Multiply
Matrix Multiply
 Unrolled C code
1 #include <x86intrin.h>
2 #define UNROLL (4)
3
4 void dgemm (int n, double* A, double* B, double* C)
5 {
6 for ( int i = 0; i < n; i+=UNROLL*4 )
7 for ( int j = 0; j < n; j++ ) {
8 __m256d c[4];
9 for ( int x = 0; x < UNROLL; x++ )
10 c[x] = _mm256_load_pd(C+i+x*4+j*n);
11
12 for( int k = 0; k < n; k++ )
13 {
14 __m256d b = _mm256_broadcast_sd(B+k+j*n);
15 for (int x = 0; x < UNROLL; x++)
16 c[x] = _mm256_add_pd(c[x],
17 _mm256_mul_pd(_mm256_load_pd(A+n*k+x*4+i), b));
18 }
19
20 for ( int x = 0; x < UNROLL; x++ )
21 _mm256_store_pd(C+i+x*4+j*n, c[x]);
22 }
23 }

Chapter 4 — The Processor — 63

Matrix Multiply
 Assembly code:
1 vmovapd (%r11),%ymm4 # Load 4 elements of C into %ymm4
2 mov %rbx,%rax # register %rax = %rbx
3 xor %ecx,%ecx # register %ecx = 0
4 vmovapd 0x20(%r11),%ymm3 # Load 4 elements of C into %ymm3
5 vmovapd 0x40(%r11),%ymm2 # Load 4 elements of C into %ymm2
6 vmovapd 0x60(%r11),%ymm1 # Load 4 elements of C into %ymm1
7 vbroadcastsd (%rcx,%r9,1),%ymm0 # Make 4 copies of B element
8 add $0x8,%rcx # register %rcx = %rcx + 8
9 vmulpd (%rax),%ymm0,%ymm5 # Parallel mul %ymm1,4 A elements
10 vaddpd %ymm5,%ymm4,%ymm4 # Parallel add %ymm5, %ymm4
11 vmulpd 0x20(%rax),%ymm0,%ymm5 # Parallel mul %ymm1,4 A elements
12 vaddpd %ymm5,%ymm3,%ymm3 # Parallel add %ymm5, %ymm3
13 vmulpd 0x40(%rax),%ymm0,%ymm5 # Parallel mul %ymm1,4 A elements
14 vmulpd 0x60(%rax),%ymm0,%ymm0 # Parallel mul %ymm1,4 A elements
15 add %r8,%rax # register %rax = %rax + %r8
16 cmp %r10,%rcx # compare %r8 to %rax
17 vaddpd %ymm5,%ymm2,%ymm2 # Parallel add %ymm5, %ymm2
18 vaddpd %ymm0,%ymm1,%ymm1 # Parallel add %ymm0, %ymm1
19 jne 68 <dgemm+0x68> # jump if not %r8 != %rax
20 add $0x1,%esi # register % esi = % esi + 1
21 vmovapd %ymm4,(%r11) # Store %ymm4 into 4 C elements
22 vmovapd %ymm3,0x20(%r11) # Store %ymm3 into 4 C elements
23 vmovapd %ymm2,0x40(%r11) # Store %ymm2 into 4 C elements
24 vmovapd %ymm1,0x60(%r11) # Store %ymm1 into 4 C elements

Chapter 4 — The Processor — 64

Performance Impact

Chapter 4 — The Processor — 65

 §4.14 Fallacies and Pitfalls
Fallacies
 Pipelining is easy (!)
 The basic idea is easy
 The devil is in the details
 e.g., detecting data hazards
 Pipelining is independent of technology
 So why haven’t we always done pipelining?
 More transistors make more advanced techniques
feasible
 Pipeline-related ISA design needs to take account of
technology trends
 e.g., predicated instructions

Chapter 4 — The Processor — 66

Pitfalls
 Poor ISA design can make pipelining
harder
 e.g., complex instruction sets (VAX, IA-32)
 Significant overhead to make pipelining work
 IA-32 micro-op approach
 e.g., complex addressing modes
 Register update side effects, memory indirection
 e.g., delayed branches
 Advanced pipelines have long delay slots

Chapter 4 — The Processor — 67

 §4.14 Concluding Remarks
Concluding Remarks
 ISA influences design of datapath and control
 Datapath and control influence design of ISA
 Pipelining improves instruction throughput
using parallelism
 More instructions completed per second
 Latency for each instruction not reduced
 Hazards: structural, data, control
 Multiple issue and dynamic scheduling (ILP)
 Dependencies limit achievable parallelism
 Complexity leads to the power wall

Chapter 4 — The Processor — 68