[llvm] [PGO][NFC] Avoid floating-point block uniformity check (PR #206547)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-pgo

Author: Yaxun (Sam) Liu (yxsamliu)

<details>
<summary>Changes</summary>

Use an integer threshold when deciding whether a block is mostly uniform.

This keeps the 90% rule exact and avoids relying on floating-point arithmetic in profile merging.

---
Full diff: https://github.com/llvm/llvm-project/pull/206547.diff


1 Files Affected:

- (modified) llvm/lib/ProfileData/InstrProf.cpp (+2-2) 


``````````diff
diff --git a/llvm/lib/ProfileData/InstrProf.cpp b/llvm/lib/ProfileData/InstrProf.cpp
index 1124cbd95b185..1002c80af7801 100644
--- a/llvm/lib/ProfileData/InstrProf.cpp
+++ b/llvm/lib/ProfileData/InstrProf.cpp
@@ -962,8 +962,8 @@ void InstrProfRecord::computeBlockUniformity() {
   for (size_t I = 0, E = Counts.size(); I < E; ++I) {
     uint64_t TotalCount = Counts[I];
     uint64_t UniformCount = UniformCounts[I];
-    bool IsUniform =
-        TotalCount == 0 || (double)UniformCount / TotalCount >= 0.9;
+    uint64_t MinUniformCount = TotalCount - TotalCount / 10;
+    bool IsUniform = UniformCount >= MinUniformCount;
     if (!IsUniform)
       UniformityBits[I / 8] &= ~(1 << (I % 8));
   }

``````````

</details>


https://github.com/llvm/llvm-project/pull/206547