[llvm] [AMDGPU] Lower `llvm.amdgcn.queue.ptr` instrinsic to using implicit kernel argument if feasible (PR #103490)

Matt Arsenault via llvm-commits llvm-commits at lists.llvm.org
Mon Aug 19 09:44:00 PDT 2024


arsenm wrote:

My expectation is for cov5, llvm.amdgcn.queue.ptr would not be used in the first place. If it were used, I would expect it to enable the same user SGPR input as it always did. Lowering it to the load from implicit argument is a strange midpoint that should not occur in real code. 

> IMHO, I don't see the win of 2

 Multiple representations of the same thing are bad. When we detect uses of implicit argument pointers, now we have an extra special case to consider. 

So I would just revert the changes that broke the queue pointer intrinsic.

https://github.com/llvm/llvm-project/pull/103490


More information about the llvm-commits mailing list