[llvm-profgen] Clean up unnecessary memory reservations between phases.

[hoy]

Summary

Cleaning up data structures that are not used after a certain point. This further brings down peak memory usage by 15% for a large benchmark.

Before:

note: Before parsePerfTraces
note: VM: 40.73 GB   RSS: 39.18 GB
note: Before parseAndAggregateTrace
note: VM: 40.73 GB   RSS: 39.18 GB
note: After parseAndAggregateTrace
note: VM: 88.93 GB   RSS: 87.97 GB
note: Before generateUnsymbolizedProfile
note: VM: 88.95 GB   RSS: 87.99 GB
note: After generateUnsymbolizedProfile
note: VM: 93.50 GB   RSS: 92.53 GB
note: After computeSizeForProfiledFunctions
note: VM: 101.13 GB   RSS: 99.36 GB
note: After generateProbeBasedProfile
note: VM: 215.61 GB   RSS: 210.88 GB
note: After postProcessProfiles
note: VM: 237.48 GB   RSS: 212.50 GB

After:

note: Before parsePerfTraces
note: VM: 40.73 GB   RSS: 39.18 GB
note: Before parseAndAggregateTrace
note: VM: 40.73 GB   RSS: 39.18 GB
note: After parseAndAggregateTrace
note: VM: 88.93 GB   RSS: 87.96 GB
note: Before generateUnsymbolizedProfile
note: VM: 88.95 GB   RSS: 87.97 GB
note: After generateUnsymbolizedProfile
note: VM: 93.50 GB   RSS: 92.51 GB
note: After computeSizeForProfiledFunctions
note: VM: 93.50 GB   RSS: 92.53 GB
note: After generateProbeBasedProfile
note: VM: 164.87 GB   RSS: 163.55 GB
note: After postProcessProfiles
note: VM: 182.28 GB   RSS: 179.43 GB

[wlei]

LGTM, thanks for the cleanup!

[wenlei]

Good change!

llvm/tools/llvm-profgen/llvm-profgen.cpp

161

nit: using a lexical scope for this purpose might be better.

[hoy]

llvm/tools/llvm-profgen/llvm-profgen.cpp

161

Lexical scope is nice, but in this case Reader and Generator have overlapping lifetimes, i.e, we need Reader to be released before Generator->write(); but after Generator->generateProfile();. So one lexical scope including both objects may not work.

Actually maybe we can define Reader->getSampleCounters() as a unique_ptr too, so that when it is transferred to Generator, Generator will take its ownership and release it when appropriate.

[wenlei]

llvm/tools/llvm-profgen/llvm-profgen.cpp

161

Lexical scope is nice, but in this case Reader and Generator have overlapping lifetimes, i.e, we need Reader to be released before Generator->write(); but after Generator->generateProfile();. So one lexical scope including both objects may not work.

I was thinking about hoisting the unique_ptr for Generator above, initialized with null, then set to the real thing after reader is available. This way the lexical scope only applies to reader.

Actually maybe we can define Reader->getSampleCounters() as a unique_ptr too, so that when it is transferred to Generator, Generator will take its ownership and release it when appropriate.

I'm not sure if this is a good idea. With this, we can leave Reader in a corrupted state after calling the getter, which is probably not a good API design.

[hoy]

llvm/tools/llvm-profgen/llvm-profgen.cpp

161

So far Reader is released between two Generator invocations. This relies on the knowledge that Reader is only used in first invocation. I was thinking that Reader could be released immediately after the needed data is passed into Generator. Usually this would be done by passing SampleCounters by value from Reader into Generator. We are now passing it by reference which holds reader live until Generator doesn't use it. This seems a tangled definition without a clear boundary between the two objects. Maybe defining SampleCounters as a shared pointer is better?

I was thinking about hoisting the unique_ptr for Generator above, initialized with null, then set to the real thing after reader is available. This way the lexical scope only applies to reader.

Yeah, this way should work. The adjusted code look like below. Still feels a bit weird. Maybe just make the release call explicitly if we don't want to change the api?

PerfInputFile PerfFile = getPerfInputFile();
std::unique_ptr<ProfileGeneratorBase> Generator;

{
    std::unique_ptr<PerfReaderBase> Reader =
        PerfReaderBase::create(Binary.get(), PerfFile);
    // Parse perf events and samples
    Reader->parsePerfTraces();

    if (SkipSymbolization)
      return EXIT_SUCCESS;

    Generator =
      ProfileGeneratorBase::create(Binary.get(), Reader->getSampleCounters(),
                                   Reader->profileIsCSFlat());
    Generator->generateProfile();
  }

  Generator->write();

[wenlei]

llvm/tools/llvm-profgen/llvm-profgen.cpp

161

I don't have strong opinion. It's ok to leave it as is, but add a comment to explain why the explicit release/scope is needed.

Still feels a bit weird.

It is just as weird as a release in the middle. :)

[hoy]

llvm/tools/llvm-profgen/llvm-profgen.cpp

161

It is just as weird as a release in the middle. :)

Agreed. I hope we'll have some time to readdress the api design in the future.

[hoy]

Updating D118677: [llvm-profgen] Clean up unnecessary memory reservations between phases.

[wenlei]

lgtm, thanks.

[hoy]

llvm/tools/llvm-profgen/llvm-profgen.cpp

163

Should call reset instead of release.

[hoy]

Updating D118677: [llvm-profgen] Clean up unnecessary memory reservations between phases.

[hoy]

llvm/tools/llvm-profgen/llvm-profgen.cpp

163

OK, looks like Reader is needed by Generator->write(), especially for UnsymbolizedProfileReader, where the context strings are built inside the reader. I'll remove the release here. This is not on the critical path. The critical path is Generator->generateProfile();.

[hoy]

Updating D118677: [llvm-profgen] Clean up unnecessary memory reservations between phases.

[wenlei]

Is the memory reduction in the description still accurate?

[hoy]

In D118677#3288970, @wenlei wrote:

Is the memory reduction in the description still accurate?

Yes, since that was measured based on using unique_ptr release which didn't really free anything.

[vitalybuka]

New leaks https://lab.llvm.org/buildbot/#/builders/5/builds/18370/steps/13/logs/stdio

[hoy]

In D118677#3288981, @vitalybuka wrote:

New leaks https://lab.llvm.org/buildbot/#/builders/5/builds/18370/steps/13/logs/stdio

Thanks for letting me know. The commit has been reverted. I'll recheck it in with a fix.

llvm/tools/llvm-profgen/PerfReader.cpp

emitWarningSummary(NumLeafExternalFrame, NumTotalSample,
                      "of samples have leaf external frame in call stack.");
   emitWarningSummary(NumLeadingOutgoingLBR, NumTotalSample,
                      "of samples have leading external LBR.");
 
   // Generate unsymbolized profile.
   warnTruncatedStack();
   warnInvalidRange();
   generateUnsymbolizedProfile();
+  AggregatedSamples.clear();
 
   if (SkipSymbolization)
     writeUnsymbolizedProfile(OutputFilename);
 }
 
 } // end namespace sampleprof
 } // end namespace llvm

llvm/tools/llvm-profgen/llvm-profgen.cpp

int main(int argc, const char *argv[]) {
 
   if (SkipSymbolization)
     return EXIT_SUCCESS;
 
   std::unique_ptr<ProfileGeneratorBase> Generator =
       ProfileGeneratorBase::create(Binary.get(), Reader->getSampleCounters(),
                                    Reader->profileIsCSFlat());
   Generator->generateProfile();
+  // The Reader object, espcially its SampleCounters field, is not needed at
+  // this point, so releasing it to reduce peak memory usage.
+  Reader.reset();
   Generator->write();
 
   return EXIT_SUCCESS;
 }