[llvm] [AMDGPU] Lower `llvm.amdgcn.queue.ptr` instrinsic to using implicit kernel argument if feasible (PR #103490)
Changpeng Fang via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 16 09:47:44 PDT 2024
changpeng wrote:
>
> But the thing is, the two SGPRs are never set in COV5. If you still want to use them, you basically have to revert those patches that @changpeng has done to disable allocation of the two SGPRs completely in COV5.
COV5 introduced a queue_ptr field in implicit kernarg, and there is no reason to allocate SGPRs. For gfx9+, the doc implies that queue_ptr field does not exist. However, the intrinsic has to return the queue_ptr. (My first impression is that the intrinsic is no longer needed, and should never be invoked). Suppose we have to keep this intrinsic, I would suggest we update the doc to say so that queue_ptr field is still needed when intrinsic call exists.
https://github.com/llvm/llvm-project/pull/103490
More information about the llvm-commits
mailing list