[llvm] [AMDGPU] Hoist WMMA coexecution hazard V_NOPs from loops to preheaders (PR #176895)

[Dark Steve via llvm-commits]

================
@@ -49,6 +58,10 @@ static cl::opt<unsigned>
     NopPadding("amdgpu-snop-padding", cl::init(0), cl::Hidden,
                cl::desc("Insert a s_nop x before every instruction"));
 
+static cl::opt<bool> EnableWMMAVnopHoisting(
+    "amdgpu-wmma-vnop-hoisting", cl::init(true), cl::Hidden,
+    cl::desc("Hoist WMMA hazard V_NOPs from loops to preheaders"));
+
----------------
PrasoonMishra wrote:

Keeping for debugging and the corner case where hoisting hurts zero-trip loops (we pay V_NOP cost even when loop body isn't executed).

https://github.com/llvm/llvm-project/pull/176895