[PATCH] D22489: AMDGPU/SI: Implement readlane/readfirstlane intrinsics to expose the instructions.

Nicolai Hähnle via llvm-commits llvm-commits at lists.llvm.org
Wed Jul 27 06:48:16 PDT 2016


nhaehnle added a subscriber: nhaehnle.
nhaehnle added a comment.

Do you plan to add code to DivergenceAnalysis to recognize these intrinsics? We might eventually want to use a readfirstlane intrinsic from Mesa to hint when certain indices can be assumed to be uniform.

(Right now, LLVM ends up deciding where to insert the readfirstlane in those cases, since they are situations where buffer resource descriptors are loaded which must end up in SGPRs anyway. By giving a readfirstlane hint, we could perhaps generate more efficient code.)


https://reviews.llvm.org/D22489





More information about the llvm-commits mailing list