[llvm] [AMDGPU] Prevent folding of FMAC into FMAAK with two literals (PR #217438)

[Igor Wodiany via llvm-commits]

================
@@ -4001,6 +4001,10 @@ bool SIInstrInfo::foldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
       if (pseudoToMCOpcode(NewOpc) == -1)
         return false;
 
+      if ((Src0->isImm() && !isInlineConstant(*Src0)) ||
----------------
IgWod wrote:

So `isOperandLegal` already checks for this condition but it doesn't help in this case, I think. For this helper to work we need a valid instruction to check against, but we won't have that until we call `legalizeOperands` further down. So to re-use  `isOperandLegal`  we would have to construct the new instruction, check legality and then roll-back if that fails. Checking `isImm` and `isInlineConstant` at the point we do it is much simpler.

https://github.com/llvm/llvm-project/pull/217438