Add loop distribution to the LTO pipeline

[sanwou01]

Summary

The LoopDistribute pass is missing from the LTO pipeline, so
-enable-loop-distribute has no effect during post-link. The pre-link
loop distribution doesn't seem to survive the LTO pipeline either.

With this patch (and -flto -mllvm -enable-loop-distribute) we see a 43%
uplift on SPEC 2006 hmmer for AArch64. The rest of SPECINT 2006 is
unaffected.

[dmgreen]

Is it possible to make a test, like llvm/test/Other/opt-O3-pipeline.ll but using -std-link-opts to print the structure of the passes? I don't think there is anything like that already for LTO, but would be useful.

[SjoerdMeijer]

FWIW, I really dislike these pipeline tests because some of them are actually very tricky to update and I doubt they provide any useful information. But agreed with Dave that for consistency such a test would probably be best (in case someone finds it useful).

[sanwou01]

Added LTO pipeline test

[sanwou01]

@SjoerdMeijer yeaahhh these pipeline tests are a bit of a pain, but nothing some big-brain sed scripting can't solve.

This is the diff after would-be-pre-committing the new pipeline test.

[SjoerdMeijer]

This looks consistent with the other pipeline, it's opt-in, and gives a nice uplift, so LGTM.

[dmgreen]

It's a shame that this adds so many other passes for something that isn't usually used, but I guess a lot of them should be free.

Is the new pass manager effected by this? It seems to go through different codepaths for LTO.

Otherwise yeah, LGTM.

[nikic]

In D89896#2346549, @dmgreen wrote:

It's a shame that this adds so many other passes for something that isn't usually used, but I guess a lot of them should be free.

Right, I noticed that as well... If this turns out to have a compile-time impact, I would expect at least some parts of this can be preserved. For example, a trivial way to "preserve" SCEV without worrying about the details would be to "forget all" if loop distribution actually happens, which it does not by default. I believe BasicAA and AAResults can be preserved pretty much always, though it also doesn't have much effect.

[sanwou01]

Reverted due to new test failing on a bunch of buildbots. I'll try again tomorrow, looks like the other pipeline tests manage to work around it.

[rsanthir.quic]

I am trying to enable LoopDistribute, to observe the 43% uplift on hmmer2006, however "flto -mllvm -enable-loop-distribute" does not seem to improve performance. Could you share what other flags you used?

[sanwou01]

Hi, our flags are "-Ofast -flto -mcpu=native -fomit-frame-pointer", where "native" is a Neoverse-N1 system. Let me know if that helps.

[pzheng]

In D89896#2454388, @sanwou01 wrote:

Hi, our flags are "-Ofast -flto -mcpu=native -fomit-frame-pointer", where "native" is a Neoverse-N1 system. Let me know if that helps.

Hi @sanwou01, did you observe a similar uplift on cortex-a57, or the improvement is only on neoverse-n1?

[sanwou01]

@pzheng I don't have perf data for Cortex-A57, but I would certainly expect to see an uplift there too.

llvm/lib/Transforms/IPO/PassManagerBuilder.cpp

void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
   if (EnableLoopInterchange)
     PM.add(createLoopInterchangePass());
   if (EnableLoopFlatten)
     PM.add(createLoopFlattenPass());
 
   // Unroll small loops
   PM.add(createSimpleLoopUnrollPass(OptLevel, DisableUnrollLoops,
                                     ForgetAllSCEVInLoopUnroll));
+  PM.add(createLoopDistributePass());
   PM.add(createLoopVectorizePass(true, !LoopVectorize));
   // The vectorizer may have significantly shortened a loop body; unroll again.
   PM.add(createLoopUnrollPass(OptLevel, DisableUnrollLoops,
                               ForgetAllSCEVInLoopUnroll));
 
   PM.add(createWarnMissedTransformationsPass());
 
   // Now that we've optimized loops (in particular loop induction variables),

llvm/test/Other/opt-LTO-pipeline.ll

 ; CHECK-NEXT:       LCSSA Verifier
 ; CHECK-NEXT:       Loop-Closed SSA Form Pass
 ; CHECK-NEXT:       Function Alias Analysis Results
 ; CHECK-NEXT:       Scalar Evolution Analysis
 ; CHECK-NEXT:       Loop Pass Manager
 ; CHECK-NEXT:         Induction Variable Simplification
 ; CHECK-NEXT:         Delete dead loops
 ; CHECK-NEXT:         Unroll loops
+; CHECK-NEXT:       Loop Access Analysis
+; CHECK-NEXT:       Lazy Branch Probability Analysis
+; CHECK-NEXT:       Lazy Block Frequency Analysis
+; CHECK-NEXT:       Optimization Remark Emitter
+; CHECK-NEXT:       Loop Distribution
 ; CHECK-NEXT:       Post-Dominator Tree Construction
 ; CHECK-NEXT:       Branch Probability Analysis
 ; CHECK-NEXT:       Block Frequency Analysis
+; CHECK-NEXT:       Scalar Evolution Analysis
+; CHECK-NEXT:       Basic Alias Analysis (stateless AA impl)
+; CHECK-NEXT:       Function Alias Analysis Results
 ; CHECK-NEXT:       Loop Access Analysis
 ; CHECK-NEXT:       Demanded bits analysis
 ; CHECK-NEXT:       Lazy Branch Probability Analysis
 ; CHECK-NEXT:       Lazy Block Frequency Analysis
 ; CHECK-NEXT:       Optimization Remark Emitter
 ; CHECK-NEXT:       Inject TLI Mappings
 ; CHECK-NEXT:       Loop Vectorization
 ; CHECK-NEXT:       Canonicalize natural loops