[llvm] [memprof] Use BLAKE for FrameId (PR #109501)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-pgo

Author: Kazu Hirata (kazutakahirata)

<details>
<summary>Changes</summary>

This patch uses a stronger hash function for FrameId.

Without this patch, I've observed hash collisions in fairly common
scenarios.  Specifically, for a given GUID, I see three pairs of hash
collisions in the two-dimensional range 1 <= LineOffset <= 70 and
1 <= Column <= 50, which may be a bit too frequent.  With the new
function, I don't see collisions at all even for a large profile.

Impact on serialization/deserialization should be as follows:

- Up to indexed memprof format Version 2, inclusive: The FrameIds
  computed with the new hash function will show up as part of the
  profile file.  However, the deserializer only treats FrameIds as
  keys (but not hash values) to retrieve Frames from the on-disk hash
  table, so a change of the hash function shouldn't matter.

- For indexed memprof format Version 3, FrameIds do not show up at all
  in the resulting profile file.  FrameIds are only used to break ties
  when we sort Frames in writeMemProfFrameArray.


---
Full diff: https://github.com/llvm/llvm-project/pull/109501.diff


1 Files Affected:

- (modified) llvm/include/llvm/ProfileData/MemProf.h (+9-14) 


``````````diff
diff --git a/llvm/include/llvm/ProfileData/MemProf.h b/llvm/include/llvm/ProfileData/MemProf.h
index 892865a0a26fea..25e183ee4c1327 100644
--- a/llvm/include/llvm/ProfileData/MemProf.h
+++ b/llvm/include/llvm/ProfileData/MemProf.h
@@ -7,8 +7,10 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/IR/GlobalValue.h"
 #include "llvm/ProfileData/MemProfData.inc"
+#include "llvm/Support/BLAKE3.h"
 #include "llvm/Support/Endian.h"
 #include "llvm/Support/EndianStream.h"
+#include "llvm/Support/HashBuilder.h"
 #include "llvm/Support/raw_ostream.h"
 
 #include <bitset>
@@ -312,20 +314,13 @@ struct Frame {
   // hashing from llvm ADT since we are going to persist the hash id, the hash
   // combine algorithm in ADT uses a new randomized seed each time.
   inline FrameId hash() const {
-    auto HashCombine = [](auto Value, size_t Seed) {
-      std::hash<decltype(Value)> Hasher;
-      // The constant used below is the 64 bit representation of the fractional
-      // part of the golden ratio. Used here for the randomness in their bit
-      // pattern.
-      return Hasher(Value) + 0x9e3779b97f4a7c15 + (Seed << 6) + (Seed >> 2);
-    };
-
-    size_t Result = 0;
-    Result ^= HashCombine(Function, Result);
-    Result ^= HashCombine(LineOffset, Result);
-    Result ^= HashCombine(Column, Result);
-    Result ^= HashCombine(IsInlineFrame, Result);
-    return static_cast<FrameId>(Result);
+    llvm::HashBuilder<llvm::TruncatedBLAKE3<8>, llvm::endianness::little>
+        HashBuilder;
+    HashBuilder.add(Function, LineOffset, Column, IsInlineFrame);
+    llvm::BLAKE3Result<8> Hash = HashBuilder.final();
+    FrameId Id;
+    std::memcpy(&Id, Hash.data(), sizeof(Hash));
+    return Id;
   }
 };
 

``````````

</details>


https://github.com/llvm/llvm-project/pull/109501