Attach maximum function count to Module when using PGO mode.

[eraman]

Summary

This sets the maximum entry count among all functions in the program to the module using module flags. This allows the optimizer to use this information.

[davidxl]

Should also add a test case as a followup. Example:

./projects/compiler-rt/test/profile/instrprof-basic.c

David

[davidxl]

LGTM. Wait a little longer in case other reviews want to chime in.

[rsmith]

lib/CodeGen/CodeGenModule.cpp

380–381 ↗

(On Diff #41683)

Can you add a test for this to clang's test/CodeGen?

[eraman]

Added a test case.

[eraman]

I've added a test case to check for the presence of MaxFunctionCount module flag. I'll check in this patch soon.

[eraman]

I have reverted the commit in r255416 because the test failed in many architectures. In many cases the linker is not able to find libclang_rt.profile-$ARCH.a file. There are also other errors. Here is one:

0. Program arguments: /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.obj/Release+Asserts/bin/clang -cc1 -triple hexagon-unknown--elf -emit-obj -disable-free -main-file-name pgo-max-function-count.c -mrelocation-model static -mthread-model posix -mdisable-fp-elim -fmath-errno -masm-verbose -mconstructor-aliases -target-cpu hexagonv4 -mqdsp6-compat -Wreturn-type -fshort-enums -mllvm -machine-sink-split=0 -target-linker-version 2.25 -dwarf-column-info -fprofile-instr-generate -resource-dir /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.obj/Release+Asserts/bin/../lib/clang/3.8.0 -internal-externc-isystem /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.obj/Release+Asserts/bin/../../gnu/lib/gcc/hexagon/0.0.0/include -internal-externc-isystem /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.obj/Release+Asserts/bin/../../gnu/lib/gcc/hexagon/0.0.0/include-fixed -internal-externc-isystem /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.obj/Release+Asserts/bin/../../gnu/hexagon/include -O2 -fdebug-compilation-dir /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.obj/tools/clang/test/CodeGen -ferror-limit 19 -fmessage-length 0 -fshort-enums -fno-signed-char -fno-use-cxa-atexit -fobjc-runtime=gcc -fdiagnostics-show-option -vectorize-loops -vectorize-slp -o /tmp/pgo-max-function-count-816195.o -x c /var/lib/buildbot/slaves/hexagon-build-03/clang-hexagon-elf/llvm.src/tools/clang/test/CodeGen/pgo-max-function-count.c

<snip>

Target: hexagon-unknown--elf

Seems like the command line is messed up (I assume there should be a space between hexagon-unknown and --elf).

Any suggestions on how to fix this test case? AFAICT, there are no PGO related tests in clang and I based this on PGO tests in compiler-rt/

[davidxl]

cfe/trunk/test/CodeGen/pgo-max-function-count.c

1

Use -fprofile-instr-generate option. -fprofile-generate is for GCC compatibility.

4

use -fprofile-instr-use=%t.profdata ..

[eraman]

Updated the patch addressing Justin's comments and a new test case.

(Should I open a new review thread since phabricator thinks this has been submitted?)

[eraman]

Fix the comment in the test case

[davidxl]

I prefer using the profile from the original test case where Max count is not 1.

[eraman]

Updated the test case.

[eraman]

Justin, does this patch look ok?

[davidxl]

LGTM

cfe/trunk/lib/CodeGen/CodeGenModule.cpp

if (Context.getLangOpts().CUDA && !Context.getLangOpts().CUDAIsDevice &&
       CUDARuntime) {
     if (llvm::Function *CudaCtorFunction = CUDARuntime->makeModuleCtorFunction())
       AddGlobalCtor(CudaCtorFunction);
     if (llvm::Function *CudaDtorFunction = CUDARuntime->makeModuleDtorFunction())
       AddGlobalDtor(CudaDtorFunction);
   }
   if (PGOReader && PGOStats.hasDiagnostics())
     PGOStats.reportDiagnostics(getDiags(), getCodeGenOpts().MainFileName);
+  // In PGO mode, attach maximum function count to the module.
+  if (PGOReader)
+    getModule().setMaximumFunctionCount(PGOReader->getMaximumFunctionCount());
   EmitCtorList(GlobalCtors, "llvm.global_ctors");
   EmitCtorList(GlobalDtors, "llvm.global_dtors");
   EmitGlobalAnnotations();
   EmitStaticExternCAliases();
   EmitDeferredUnusedCoverageMappings();
   if (CoverageMapping)
     CoverageMapping->emit();
   emitLLVMUsed();

cfe/trunk/test/CodeGen/pgo-max-function-count.c

+// RUN: %clang -fprofile-generate -o %t -O2 %s
+// RUN: env LLVM_PROFILE_FILE=%t.profraw  %t
+// RUN: llvm-profdata merge -o %t.profdata %t.profraw
+// RUN: %clang -fprofile-use=%t.profdata -o - -S -emit-llvm %s | FileCheck %s
+// Check
+int main() {
+  return 0;
+}
+// CHECK: !{{[0-9]+}} = !{i32 1, !"MaxFunctionCount", i32 1}