nhaehnle added a comment. In D124158#3486110 <https://reviews.llvm.org/D124158#3486110>, @jdoerfert wrote:
> I agree. As far as I can tell you have two options, both are specific to the > shuffle functions: > > 1. Do not set noundef for calls to them as they allow undef values for all > lanes we don't read the value. > 2. Freeze the inputs unconditionally. Right, with some nitpicks. Option #1 is semantically more accurate: __shfl_sync, subgroupShuffe, and all similar instructions across GPU programming languages are meant to be conceptually similar to `select` in that they select a value from a lane. The "data" argument and return value should allow undef and poison. The incoming value is simply returned as-is, and so poison is propagated instead of causing immediate UB, just as it is for a `select` instruction. However, the lane argument can (and arguably should) still be `noundef`. There's a separate curious issue in that apparently, reading from an uninitialized variable is **not** UB in CUDA/HIP/GLSL/HLSL/etc. If it was, a lot of code existing out there in the wild would be broken. But that's a matter for the relevant language standards to decide (for the subset of languages that //have// proper standards to begin with). Repository: rG LLVM Github Monorepo CHANGES SINCE LAST ACTION https://reviews.llvm.org/D124158/new/ https://reviews.llvm.org/D124158 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
