[llvm] [AMDGPU] In promote-alloca, if index is dynamic, sandwich load with bitcasts to reduce number of extractelements as they have large expansion in the backend. (PR #171253)

[Nicolai Hähnle via llvm-commits]

================
@@ -556,6 +557,35 @@ static Value *promoteAllocaUserToVector(
       auto *SubVecTy = FixedVectorType::get(VecEltTy, NumLoadedElts);
       assert(DL.getTypeStoreSize(SubVecTy) == DL.getTypeStoreSize(AccessTy));
 
+      // If idx is dynamic, then sandwich load with bitcasts.
+      // ie. <64 x i8> -> <16 x i8>  instead do
+      //     <64 x i8> -> <4 x i128> -> i128 -> <16 x i8>
+      // Extracting subvector with dynamic index has very large expansion in
+      // the amdgpu backend. Limit to pow2 for UDiv.
+      if (!isa<ConstantInt>(Index) && SubVecTy->isIntOrIntVectorTy() &&
+          llvm::isPowerOf2_32(VectorTy->getNumElements()) &&
+          llvm::isPowerOf2_32(SubVecTy->getNumElements())) {
----------------
nhaehnle wrote:

Check instead that the subvector is a power-of-two and the vector is a multiple of it? That covers more cases that a reasonable programmer might use and I would expect/hope that it still gives reasonable codegen.

https://github.com/llvm/llvm-project/pull/171253