[llvm] [AMDGPU] Provide control to force VGPR MFMA form (PR #148079)

[Matt Arsenault via llvm-commits]

================
@@ -29,6 +29,16 @@ enum { MAX_LANES = 64 };
 
 using namespace llvm;
 
+// TODO -- delete this flag once we have more robust mechanisms to allocate the
+// optimal RC for Opc and Dest of MFMA. In particular, there are high RP cases
+// where it is better to produce the VGPR form (e.g. if there are VGPR users
+// of the MFMA result).
+cl::opt<bool> MFMAVGPRForm(
----------------
arsenm wrote:

Missing static 

https://github.com/llvm/llvm-project/pull/148079