[llvm] AMDGPU/GlobalISel: RegBankLegalize rules for mfma_scale (PR #203279)

[Petar Avramovic via llvm-commits]

================
@@ -2532,6 +2532,13 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
             {{VgprOrAgprAnyTy},
              {IntrId, VgprAnyTy, VgprAnyTy, VgprOrAgprAnyTy, VgprAnyTy}}});
 
+  addRulesForIOpcs({amdgcn_mfma_scale_f32_32x32x64_f8f6f4,
+                    amdgcn_mfma_scale_f32_16x16x128_f8f6f4})
+      .Any({{DivAnyTy},
+            {{VgprOrAgprAnyTy},
+             {IntrId, VgprAnyTy, VgprAnyTy, VgprAnyTy, Imm, Imm, Imm, Vgpr32,
----------------
petar-avramovic wrote:

added more VgprOrAgprAnyTy for operands that accept it, now we force more agprs then sdag, sdag does not force agprs on src0/1. Is this what we want ?

https://github.com/llvm/llvm-project/pull/203279