[PATCH] D16603: AMDGPU/SI: Detect uniform branches and emit s_cbranch instructions

[Marek Olšák via llvm-commits]

mareko added a subscriber: mareko.
mareko added a comment.

Is it this hardware bug handled properly?

Description: VCCZ bit corrupted by SMRD
Scope: SI+CI
Solution: Anytime a Scalar Memory read is issued, the shader compiler must assume that VCCZ is no longer valid AND cannot be loaded with a new value until all of the scalar memory reads have completed and returned their data, as guaranteed by inserting an S_WAITCNT LGKM_CNT(0) before issuing an instruction which updates VCCZ.


http://reviews.llvm.org/D16603