[llvm-profgen] Ignore stack samples before aggregation

[wlei]

Summary

With ignore-stack-samples, We can ignore the call stack before the samples aggregation which could reduce some redundant computations.

[hoy]

Good catch, thanks.

llvm/tools/llvm-profgen/PerfReader.cpp

650–656

make this an else case?

[wlei]

moved into if branch

[wenlei]

lgtm, thanks. do you see this making AutoFDO profile generation faster? and by how much?

[wlei]

In D111577#3058710, @wenlei wrote:

lgtm, thanks. do you see this making AutoFDO profile generation faster? and by how much?

Tested on clang-10 build, it can make profile generation a little bit faster, around 3%.

llvm/tools/llvm-profgen/PerfReader.cpp

if (!extractCallstack(TraceIt, Sample->CallStack)) {
     return;
   }
 
   warnIfMissingMMap();
 
   if (!TraceIt.isAtEoF() && TraceIt.getCurrentLine().startswith(" 0x")) {
     // Parsing LBR stack and populate into PerfSample.LBRStack
     if (extractLBRStack(TraceIt, Sample->LBRStack)) {
+      if (IgnoreStackSamples) {
+        Sample->CallStack.clear();
+      } else {
         // Canonicalize stack leaf to avoid 'random' IP from leaf frame skew LBR
         // ranges
         Sample->CallStack.front() = Sample->LBRStack[0].Target;
+      }
       // Record samples by aggregation
       AggregatedSamples[Hashable<PerfSample>(Sample)] += Count;
     }
   } else {
     // LBR sample is encoded in single line after stack sample
     exitWithError("'Hybrid perf sample is corrupted, No LBR sample line");
   }
 }