[PATCH] D105248: [CSSPGO] Do not import pseudo probe desc in thinLTO

[Hongtao Yu via Phabricator via llvm-commits]

hoy added a comment.



In D105248#2855763 <https://reviews.llvm.org/D105248#2855763>, @wenlei wrote:

> In D105248#2853977 <https://reviews.llvm.org/D105248#2853977>, @hoy wrote:
>
>> Verified that for SPEC, the generated profiles w/ and w/o this changes are very close. The runtime performance of pass2 has no difference two. Regarding the pseudo probe sections, the .pseudo_probe section is exactly the same for all benchmarks. The .pseudo_probe_desc sections are mostly the same, except for a couple benchmarks, for which the probe decoding gives exact result. There might be a glitch in setting the comdat group name for duplicated functions, like the ones with `.llvm.` suffix, or an issue of our internal thinlto dedup work.
>
> Thanks for checking. For pseudo_probe_desc, do you know what is the actual difference, and why decoding leads to same output with different pseudo_probe_desc?

It appeared due to a glitch of our internal thinlto comdat deduplication change. Only part of the duplicated comdat groups were removed.  llvm-profgen can tolerate that by taking only the last duplicate.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D105248/new/

https://reviews.llvm.org/D105248