[llvm] [AMDGPU] Skip s_delay_alu for gfx1250 WMMA C-reuse chains (PR #214101)

[via llvm-commits]

================
@@ -398,6 +406,10 @@ class AMDGPUInsertDelayAlu {
             // ignore this operand.
             if (MI.getOpcode() == AMDGPU::V_WRITELANE_B32 && Op.isTied())
               continue;
+            // Skip the tied srcC of a GFX1250 C-reuse edge.
+            if (IsWMMACReuse && Op.isTied() &&
+                TRI->regsOverlap(Op.getReg(), PrevWMMAVDst))
----------------
sstipano wrote:

I think we need full overlap here, in other words: `Op.getReg() == PrevWMMAVDst`.

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