[llvm] [MemProf] Suppress duplicate clones in the LTO backend (PR #161551)

Kazu Hirata via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 1 14:06:16 PDT 2025


================
@@ -5186,19 +5191,129 @@ bool CallsiteContextGraph<DerivedCCG, FuncTy, CallTy>::assignFunctions() {
   return Changed;
 }
 
+// Compute a SHA1 hash of the callsite and alloc version information of clone I
+// in the summary, to use in detection of duplicate clones.
+std::string ComputeHash(StringMap<Function *> &HashToFunc, FunctionSummary *FS,
+                        unsigned I) {
+  SHA1 Hasher;
+  // Update hash with any callsites that call non-default (non-zero) callee
+  // versions.
+  for (auto &SN : FS->callsites()) {
+    // In theory all callsites and allocs in this function should have the same
+    // number of clone entries, but handle any discrepancies gracefully below
+    // for NDEBUG builds.
+    assert(
+        SN.Clones.size() > I &&
+        "Callsite summary has fewer entries than other summaries in function");
+    if (SN.Clones.size() <= I || !SN.Clones[I])
+      continue;
+    uint8_t Data[4];
+    support::endian::write32le(Data, SN.Clones[I]);
+    Hasher.update(Data);
+  }
+  // Update hash with any allocs that have non-default (non-None) hints.
+  for (auto &AN : FS->allocs()) {
+    // In theory all callsites and allocs in this function should have the same
+    // number of clone entries, but handle any discrepancies gracefully below
+    // for NDEBUG builds.
+    assert(AN.Versions.size() > I &&
+           "Alloc summary has fewer entries than other summaries in function");
+    if (AN.Versions.size() <= I ||
+        (AllocationType)AN.Versions[I] == AllocationType::None)
+      continue;
+    Hasher.update(ArrayRef<uint8_t>(&AN.Versions[I], 1));
+  }
+  return toHex(Hasher.result());
----------------
kazutakahirata wrote:

If you are OK with returning `uint64_t` instead, may I suggest the following?:

```suggestion
  return support::endian::read64le(Hasher.result().data());
```

https://github.com/llvm/llvm-project/pull/161551


More information about the llvm-commits mailing list