llvm-project

ac00a114 - [AMDGPU] Ensure v_mfma_scale_f32_{16x16x128|32x32x64}_f8f6f4 instructions are convergent (#178627)

[AMDGPU] Ensure v_mfma_scale_f32_{16x16x128|32x32x64}_f8f6f4 instructions are convergent (#178627) The scaled variants of mfma instructions are not properly marked as "convergent" and hence the machine-sink pass sinks them which is incorrect. This patch ensures that the instructions get marked as "convergent". The new test also covers other mfma variants, but only the scale variants are mistreated without the changes from this patch.