llvm-project

fc4fed4d - [HLSL] Codegen column-major matrix initializer lists without a vector shuffle (#186228)

[HLSL] Codegen column-major matrix initializer lists without a vector shuffle (#186228) Fixes #185518 The SPIR-V backend does not handle the lowering of `shufflevector` instructions on vectors with more than 4 elements. This PR changes the codegen of matrix init lists to directly emit vectors with elements in column-major order when the default matrix memory layout is column-major, as opposed to in linear/row-major order followed by a vector shuffle. While an alternative fix could be to change the default depth of [`canEvaluateShuffled`](https://github.com/llvm/llvm-project/blob/main/llvm/lib/Transforms/InstCombine/InstCombineVectorOps.cpp#L1865-L1866) to 16 in `InstCombineVectorOps.cpp` to eliminate the vector shuffle for vectors of up to 16 elements in size (to handle 4x4 matrices), this change would have broader impacts than just HLSL, which does not seem necessary for the scope of this issue (which regards only matrix initializer list codegen). Another alternative fix would be to extend the `shufflevector` lowering in the SPIR-V backend to support vectors of more than 4 elements. However, again, this goes beyond the scope of just matrix initializer list codegen which is so far the only case where a vector shuffle of a vector more than 4 elements appeared. Assisted-by: claude-opus-4.6