[llvm] [AMDGPU] Ensure all WMMA instructions are marked as convergent (PR #178314)

[via llvm-commits]

================
@@ -0,0 +1,48 @@
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx9-generic    -run-pass=machine-sink -mattr=+wavefrontsize64 %s -stats 2>&1 | \
+# RUN:   not grep "Number of machine instructions sunk"
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx10-1-generic -run-pass=machine-sink -mattr=+wavefrontsize64 %s -stats 2>&1 | \
+# RUN:   not grep "Number of machine instructions sunk"
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx11-generic   -run-pass=machine-sink -mattr=+wavefrontsize64 %s -stats 2>&1 | \
+# RUN:   not grep "Number of machine instructions sunk"
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx12-generic   -run-pass=machine-sink -mattr=+wavefrontsize64 %s -stats 2>&1 | \
+# RUN:   not grep "Number of machine instructions sunk"
+# machine-sink must not sink WMMA* instructions.
----------------
LU-JOHN wrote:

Changed to only one target.

https://github.com/llvm/llvm-project/pull/178314