[PATCH] D147740: [NFC][llvm-profdata] Refactoring Sample Profile Reader to increase FDO build speed

[Lei Wang via Phabricator via llvm-commits]

wlei added a comment.

Thanks for the improvement. I'm a bit concern if this results in any hash collision issue for CS profile. MD5 might be perfect for one function name, but for CS profile, this uses the "hash_combine_range", the context size can be very large, the probability of hash collision might be higher than non-CS profile.



================
Comment at: llvm/lib/ProfileData/SampleProfReader.cpp:661
+    if (FProfile.getContext() != FContext)
+      LLVM_DEBUG(errs() << "Hash collision detected: "
+                        << FProfile.getContext().toString() << " and "
----------------
Here it checked hash collision during reading the context, it also could happen during writing time(profile generation time or merging profile time..).


================
Comment at: llvm/tools/llvm-profgen/ProfileGenerator.cpp:971
     SampleContextFrames NewContext = *Contexts.insert(Context).first;
-    auto Ret = ProfileMap.emplace(NewContext, std::move(*FProfile));
+    auto Ret = ProfileMap.insert(std::make_pair(
+        SampleContext(NewContext).getHashCode(), std::move(*FProfile)));
----------------
Here for profile generation time, if hash collision happens, the sample is lost.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D147740/new/

https://reviews.llvm.org/D147740