[LLVMdev] Duplicate records in instrumented profile data?
Justin Bogner
mail at justinbogner.com
Tue May 12 11:24:13 PDT 2015
Diego Novillo <dnovillo at google.com> writes:
> As a follow up to my function entry count
> (http://reviews.llvm.org/D9628), I modified codegen to emit the entry
> counters:
>
> --- a/lib/CodeGen/CodeGenPGO.cpp
> +++ b/lib/CodeGen/CodeGenPGO.cpp
> @@ -773,6 +773,8 @@
> CodeGenPGO::applyFunctionAttributes(llvm::IndexedInstrProfReader
> *PGOReader,
> // Turn on Cold attribute for cold functions.
> // FIXME: 1% is from preliminary tuning on SPEC, it may not be optimal.
> Fn->addFnAttr(llvm::Attribute::Cold);
> +
> + Fn->setEntryCount(FunctionCount);
> }
>
> void CodeGenPGO::emitCounterIncrement(CGBuilderTy &Builder, const Stmt *S) {
>
>
> This is causing a couple of failures in
> projects/compiler-rt/test/profile/instrprof-dynamic-*-shared.test.
> There is a comparison failure in the bytecode files between the shared
> and static link cases.
>
> The entry counts in the statically linked binary are exactly twice
> those of the share-linked library. At first I thought it was
> llvm-profdata somehow merging the wrong records, but the raw count
> file seems to be wrong.
>
> I modified llvm-profdata to show me the records it's merging. And it
> is, indeed, getting duplicate records for one of the template
> functions. For instrproof-dynamic-one-shared.test, I get the following
> for the statically linked profile:
>
> $ bin/llvm-profdata merge -o /dev/null
> projects/compiler-rt/test/profile/Output/instrprof-dynamic-one-shared.test.tmp-static.profraw
> Added record for: _Z1av - hash: 10 - counts: { 1 1 }
> Added record for: _Z3barIvEvv - hash: 10 - counts: { 3 3 }
> Added record for: _Z3barIcEvv - hash: 10 - counts: { 1 1 }
> Added record for: _Z1bv - hash: 10 - counts: { 1 1 }
> Added record for: _Z3barIvEvv - hash: 10 - counts: { 3 3 }
> Added record for: _Z3barIiEvv - hash: 10 - counts: { 1 1 }
> Added record for: _Z3fooi - hash: 10 - counts: { 1 1 }
> Added record for: main - hash: 0 - counts: { 1 }
>
> Notice how bar<void>() (_Z3barIvEvv) shows up twice. The entry counter
> for it ends up at 6 because of this.
>
> For the share-linked profile, I get:
>
> bin/llvm-profdata merge -o /dev/null
> projects/compiler-rt/test/profile/Output/instrprof-dynamic-one-shared.test.tmp-shared.profraw
> Added record for: _Z1bv - hash: 10 - counts: { 1 1 }
> Added record for: _Z3barIvEvv - hash: 10 - counts: { 3 3 }
> Added record for: _Z3barIiEvv - hash: 10 - counts: { 1 1 }
> Added record for: _Z3fooi - hash: 10 - counts: { 1 1 }
> Added record for: main - hash: 0 - counts: { 1 }
> Added record for: _Z1av - hash: 10 - counts: { 1 1 }
> Added record for: _Z3barIvEvv - hash: 10 - counts: { 0 0 }
> Added record for: _Z3barIcEvv - hash: 10 - counts: { 1 1 }
>
> In this case, bar<void>() is also mentioned twice, but the second set
> of counters is set to 0, so we managed to get it right.
>
> The problem seems to be somewhere in the instrumentation generated.
> Do you folks recognize this? To reproduce, you would need to apply
> http://reviews.llvm.org/D9628 (it is still under review) and the
> patchlet I included at the top of this message.
This is a bug in how we're dealing with the linkage of counters, and it
sounds kind of bad to me. Could you file a PR? If you want to
investigate I can probably point you in the right direction, otherwise
I'll look at this tomorrow.
More information about the llvm-dev
mailing list