Tune insert/extract subvector cost

Change-Id: I5c18aa058ff272ca399687ef73eb7ead427f1e76
llvm · May 20, 2024 · bb0f618 · bb0f618
1 parent 128bd38
commit bb0f618
Show file tree

Hide file tree

Showing 3 changed files with 41 additions and 40 deletions.
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetTransformInfo.cpp
@@ -1159,9 +1159,10 @@ InstructionCost GCNTTIImpl::getShuffleCost(TTI::ShuffleKind Kind,
     }
     case TTI::SK_ExtractSubvector:
     case TTI::SK_InsertSubvector: {
-      if (HasVOP3P && NumVectorElts == 2)
+      // Even aligned accesses are free
+      if (!(Index % 2))
         return 0;
-      // Insert/extract subvectors require only shifts / extract code to get the
+      // Insert/extract subvectors only require shifts / extract code to get the
       // relevant bits
       return alignTo(RequestedElts, 2) / 2;
     }