[llvm] [AMDGPU] Skip s_delay_alu for gfx12+ WMMA C-reuse chains (PR #214101)

[Matt Arsenault via llvm-commits]

================
@@ -390,6 +394,10 @@ class AMDGPUInsertDelayAlu {
         State = DelayState();
       } else if (Type != OTHER) {
         DelayInfo Delay;
+        // GFX12+ C-reuse: back-to-back WMMAs into the same C register forward
+        // the accumulator in place, so the tied srcC read has no dependency.
+        bool IsWMMACReuse = ST->hasGFX12Insts() && PrevWMMAVDst.isValid() &&
----------------
arsenm wrote:

Don't need the gfx12insts check?

https://github.com/llvm/llvm-project/pull/214101