[PATCH] D19191: AMDGPU/SI: add llvm.amdgcn.ps.live intrinsic

[Nicolai Hähnle via llvm-commits]

nhaehnle added a comment.

I don't think it should be convergent, because other threads in the wave do not affect the return value. It does have to be a source of divergence, though.

I didn't consider the possibility of reading from the EXEC mask, but it's tricky. The intrinsic must return the //initial// EXEC mask. In a pixel shader, the current EXEC mask can be a superset of the initial EXEC. Even if Mesa always emits the intrinsic call as the first thing in the shader, I don't think we can/should rely on the various scheduler steps in LLVM maintaining that ordering.


http://reviews.llvm.org/D19191