[llvm] [AMDGPU] Do not treat LDSDMA as VALU in several passes (PR #212866)

[Matt Arsenault via llvm-commits]

================
@@ -0,0 +1,30 @@
+# LDSDMA must not be treated as VALU by amdgpu-wait-sgpr-hazards. A tensor/async
+# load between a VALU SGPR write and its consumer must not clear the hazard.
+# RUN: llc -mtriple=amdgpu11.00 -mattr=+wavefrontsize64 -run-pass post-RA-hazard-rec,amdgpu-wait-sgpr-hazards -o - %s | FileCheck %s
+
+--- |
+  define amdgpu_gs void @ldsdma_not_valu_for_sgpr_hazard() { ret void }
+...
+
----------------
arsenm wrote:

The only benefit the IR section is providing here is the amdgpu_gs, does this really depend on using a shader CC 

https://github.com/llvm/llvm-project/pull/212866