[PATCH] D15873: AMDGPU: Implement readcyclecounter

[Matt Arsenault via llvm-commits]

arsenm added a comment.

In http://reviews.llvm.org/D15873#319504, @nhaehnle wrote:

> I don't have a strong opinion s_memtime vs. s_memrealtime. Probably the least surprise is to mirror HSAIL, but I don't really care.
>
> I am against forcing there to be a waitcnt immediately after the s_memtime, because that might cause unnecessary stalls. However, we should prevent the s_memtime itself from getting reordered. GL_ARB_shader_clock explicitly states that the clock functions are code motion barriers. Isn't the intrinsic marked as having side effects?


Yes, the intrinsic already won't be reordered with other side effects / memory writes


http://reviews.llvm.org/D15873