[SPGO] Use std::hash instead of MD5 to avoid run time regression in llvm-profgen

[HighW4y2H3ll]

#66164 changed the hashing in SampleContextFrame from std::hash to MD5 in a very hot function (ContextTrieNode::getOrCrateChildContext()) in llvm-profgen. This creates over 2x run time regression when running llvm-profgen with csspgo preinliner enabled, since the MD5 computation is tripled comparing to the Murmur hash in the std library. An llvm-profgen run time comparison shows follows:

$ time llvm-profgen -binary $BINARY--perfscript $SAMPLES --populate-profile-symbol-list --show-density --output=XXX

# MD5 hash
real    105m31.644s
user    104m51.334s
sys     0m35.033s

# std::hash
real    46m0.340s
user    45m17.998s
sys     0m38.420s

Can confirm that this patch recovers the run time regression in llvm-profgen, and the perf testing in our internal services shows neutral.

[llvmbot]

@llvm/pr-subscribers-pgo

Author: None (HighW4y2H3ll)

Changes

#66164 changed the hashing in SampleContextFrame from std::hash to MD5 in a very hot function (ContextTrieNode::getOrCrateChildContext()) in llvm-profgen. This creates over 2x run time regression when running llvm-profgen with csspgo preinliner enabled, since the MD5 computation is tripled comparing to the Murmur hash in the std library. An llvm-profgen run time comparison shows follows:

$ time llvm-profgen -binary $BINARY--perfscript $SAMPLES --populate-profile-symbol-list --show-density --output=XXX

# MD5 hash
real    105m31.644s
user    104m51.334s
sys     0m35.033s

# std::hash
real    46m0.340s
user    45m17.998s
sys     0m38.420s

Can confirm that this patch recovers the run time regression in llvm-profgen, and the perf testing in our internal services shows neutral.

---

Full diff: https://github.com/llvm/llvm-project/pull/180581.diff

1 Files Affected:

• (modified) llvm/include/llvm/ProfileData/SampleProf.h (+3-1)

diff --git a/llvm/include/llvm/ProfileData/SampleProf.h b/llvm/include/llvm/ProfileData/SampleProf.h
index b75dffaff19f7..8766ab23ac1da 100644
--- a/llvm/include/llvm/ProfileData/SampleProf.h
+++ b/llvm/include/llvm/ProfileData/SampleProf.h
@@ -522,7 +522,9 @@ struct SampleContextFrame {
   }
 
   uint64_t getHashCode() const {
-    uint64_t NameHash = Func.getHashCode();
+    // Context frame hash is heavily used in llvm-profgen context-sensitive
+    // pre-inliner. Use a lightweight hashing here to avoid speed regression.
+    uint64_t NameHash = std::hash<std::string>{}(Func.str());
     uint64_t LocId = Location.getHashCode();
     return NameHash + (LocId << 5) + LocId;
   }

[github-actions]

🐧 Linux x64 Test Results

• 189867 tests passed
• 5060 tests skipped

✅ The build succeeded and all tests passed.

[apolloww]

I think we only need to recompute hash when the FunctionId is a string.

[HighW4y2H3ll]

ah, good catch.. updated! thx

[apolloww]

LGTM