ssahasra wrote: > SPIRV is a storage format and flavoring it as AMDGPU shouldn't introduce > unique optimization issues. If SPIRV is adding custom optimizations that is > harmful to the ultimate finalization, that's an issue to directly address in > the SPIRV optimizations.
This seems to me as the core issue. SPIRV being a target-neutral representation, how is it that target-neutral optimizations in LLVM are harmful to any one target? If hypothetically a HIP program were parsed into an identical LLVM IR, wouldn't the same optimizations make the same decisions for the AMDGCN target? https://github.com/llvm/llvm-project/pull/154765 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
