[clang] [llvm] [Clang][AMDGPU] Deprecate `amdgpu-num-vgpr` and `amdgpu-num-sgpr` (PR #193023)

Jannik Silvanus via cfe-commits cfe-commits at lists.llvm.org
Tue Nov 3 08:00:14 PST 2026


jasilvanus wrote:

Dynamic VGPRs are only supported in compute shaders, and are used for non-inlined ray tracing pipelines where indirect function calls are lowered to `amdgpu_cs_chain` calls. 

For these, the compiler backend inherently doesn't have a full view of the pipeline, and both guaranteeing forward progress as well as reasonable performance requires a high-level plan/protocol to follow controlled by the frontend/pipeline compiler/pipeline linker that the compiler backend must support in the form of respecting required usage limits.

https://github.com/llvm/llvm-project/pull/193023


More information about the cfe-commits mailing list