[clang] [lld] [llvm] Adding Matching and Inference Functionality to Propeller (PR #139008)
Sriraman Tallam via cfe-commits
cfe-commits at lists.llvm.org
Fri May 23 11:33:10 PDT 2025
================
@@ -1688,6 +1688,11 @@ def fprofile_sample_accurate : Flag<["-"], "fprofile-sample-accurate">,
as cold. Otherwise, treat callsites without profile samples as if
we have no profile}]>,
MarshallingInfoFlag<CodeGenOpts<"ProfileSampleAccurate">>;
+def fpropeller_profile_use_EQ : Joined<["-"], "fpropeller-profile-use=">,
+ Group<f_Group>, Visibility<[ClangOption, CC1Option, CLOption]>,
+ MetaVarName<"<pathname>">,
----------------
tmsri wrote:
So, we want :
1. Another option to expose raw propeller profiles, MBB hashes and counts to the compiler.
2. It would be mutually exclusive to -fbasic-block-sections=list options
I want to have a discussion on how much of a maintenance burden this would cause. This is because Propeller will now do code layout in two different ways. With stale profiles, the compiler will do it. By default, the offline tool would do it. For future optimizations, we need to figure out how to support two paths. @rlavaee @shenhanc78
Can we expose raw profiles using -fbasic-blocks-sections=list= itself? Do we need a new option? I like the stale profile matching ideas but I am not sure adding a new flag and a new file is the way to go forward.
https://github.com/llvm/llvm-project/pull/139008
More information about the cfe-commits
mailing list