[llvm] [LV] Don't treat the pointer operand of histograms as scalar uses (PR #215769)

Benjamin Maxwell via llvm-commits llvm-commits at lists.llvm.org
Wed Aug 12 07:10:22 PDT 2026


================
@@ -2239,7 +2239,8 @@ void LoopVectorizationCostModel::collectLoopScalars(ElementCount VF) {
         return WideningDecision == CM_Scalarize;
     assert(Ptr == getLoadStorePointerOperand(MemAccess) &&
            "Ptr is neither a value or pointer operand");
-    return WideningDecision != CM_GatherScatter;
+    return WideningDecision != CM_GatherScatter &&
----------------
MacDue wrote:

It's scalarize. The cost most does not really seem to know about histograms, so it's costing it as if it's a normal gather/scatter and finds it's apparently cheaper to scalarize it. However, we map any `store` with histogram info to a vector histogram intrinsic. 

https://github.com/llvm/llvm-project/pull/215769


More information about the llvm-commits mailing list