In D127765#3814900, @fhahn wrote:

Hmm, aren't there other places in the code that rely on instructions being hoisted out by LICM? Unless I am mis-remembering, It's not clear to me if/why freeze should be treated in a special way.

Use opt to remove useless profiling data

@alexgatea Would you please also evaluate this patch on SPEC to see whether there are other regressions? Thanks

revert to previous revision that has been merged into main branch

Sorry for the mistake, I will keep this revision open and make it related to the update revision for the regression.

I will create a new revision for this patch

Avoid applying non-trivial unswitching in cold Function. Compared with previous PGO solution, this version will apply non-trivial unswitching on cold loops in hot functions.

@alexgatea please try whether the new revision could solve your problem. Thanks

As suggested by other users, we need to make the skip check more restrict to avoid performance degradation.

@alexgatea Thanks for the feedback.
isColdBlock should be global. One example is in (https://llvm.org/doxygen/ProfileSummaryInfo_8cpp_source.html#l00142), which uses isColdBlock to check whether a function is cold or not. So I assume the issue is the loop is non-cold in real case, but was concerned as cold in the profile data. I kindly suggest updating the profile data to see if there is still any degradation.
Please correct me if I am wrong. Thanks

This patch implements SimpleLoopUnswitch using FuncPass, which can be used to replace the LoopPass implementation currently used in the optimization pipeline.
This patch has been verified and can be re-opened once we find using FuncPass for SLU is necessary/beneficial.

[test] Update PGO SLU test

[test][SimpleLoopUnswitch] Precommit test for D129599

[llvm] Change test file for PGO on SLU

[llvm] Implement PGO for SLU
Using PGO for SLU to avoid non-trivial unswitch code on code loops, which will increase the binary code size without gain much benefit.

In D130126#3663981, @nikic wrote:

Please put the test in llvm/test/Transforms/CodeGenPrepare instead and use update_test_checks.py. Omit the target triple if possible, which it likely is.

Reformat the test file with update_test_checks.py

@bcahoon This revision is based on your previous implementation (https://reviews.llvm.org/D109762). In the parent revision(https://reviews.llvm.org/D127547) we use FuncPass to re-implement non-trivial LoopUnswitch, thus, we do not invoke DivergenceAnalysis for each loop. Would you please have are look for whether these modifications (implement non-trivial LoopUnswitch with FuncPass and integrate DivergenceAnalysis in the new FuncPass) work on your test cases? Thanks

Fix bug for hoisting freeze instruction