Ronald Barnes

1
Combining Predication and Software Speculation
with Dynamic Speculation

• Ronald Barnes
• George Mason University

2
Motivation

• Architectural features empower compilers to
  express instruction-level parallelism (ILP)
• Large architecturally visible register files
• Predicated instruction sets
• Explicit speculation
• Compiler schedulers meticulously group
  instructions for parallel execution
• Makes static assumptions about control-flow based
  on profile information
• Assumes hit latency for memory operations

3
Compiler Expressed Parallelism

• Compiler can find a significant number of
  instructions for parallel execution on 6-issue
  processor

4
Motivation cont.

• Run-time stalls retard compilers carefully
  planned execution
• Instruction cache misses
• Branch mispredictions
• Long, fixed latency instructions
• Variable memory-access latency

5
Compiler Expressed Parallelism

• Dynamic stalls (of which cache misses are most
  important) drastically reduce observed performance

6
Experimental configuration

• Benchmarks compiled and (unless otherwise stated)
  aggressively optimized with the IMPACT C compiler

7
Microarchitectural solutions to tolerating memory
latency

• Out-of-order EPIC executionWang01
• In-order runaheadDundas97
• Multipass Barnes05

8
Dynamic Speculation in EPIC

• Many options for tolerating memory latency have
  characteristics of dynamic speculation/scheduling
• An instruction stream chosen by branch prediction
  is speculatively executed
• This execution (even if it is only prefetching)
  is speculated before control flow instructions
• In particular speculation during cache miss that
  otherwise would cause the processor to stall
• EPIC features are envisioned to statically expose
  ILPare they helpful w/ dynamic speculation?

9
EPIC Features

• Predication
• Allows compiler toeffectively manage parallel
  execution in the presence of control-dependences
• Converts control-flow dependence to data-flow
  dependence
• Software control speculation
• Enables compiler to schedule instructions around
  (otherwise serializing) branches

10
Predication

• Wang01 discussed some of the problems with
  predication and out-of-order execution
• Multiple definitions
• Predicate define on dependence chain

11
Register renaming and predication

• Register renaming requires complexity in RAT and
  select_ops

12
Compare instruction dependences

• Predication adds a data-flow dependence that
  hinders instruction dispatch from issue queue
• Predicate slip Wang01 allows dispatch of
  predicated instruction before its predicate is
  computed
• Instruction cannot bypass to other instructions
  until its predicate is known

13
Dependence with distance from miss
14
Dynamic speculation performance
15
Branch prediction

• Biggest difference twolf lt 5 performance
  improvement

16
Using aggressive promotion to reduce dependences
17
Performance w/ aggressive promotion
18
Speculation

• Fill up reservation tables with instructions off
  path

19
Fullness of issue queues
20
Effect of speculation on dynamic scheduling
21
Conclusions

• Look for more EPIC-friendly ways to achieve
  benefits of dynamic speculation
• Runahead execution/Multipass pipelines
• Prefetching helper threads
• Researchers should consider characteristics of
  compilation
• Dynamic systems should be specifically targeted
• ex. Less predication, more aggressive promotion,
  less aggressive static speculation)

22
Runahead execution performance