Open
Description
We currently failed to make use of s_or_b64/s_and_b64/s_xor_b64 in the uniform cases, and scalarize into pairs of the 32-bit binary operators. Making them legal is simple (starter patch attached), but avoiding all the regressions is more difficult.
0001-AMDGPU-Make-v2i32-and-or-xor-legal.patch
We need additional combiner work to avoid the regressions