[llvm] [AMDGPU][GlobalISel] Improve combining on v_ldexp (PR #190236)

[Jasmine Tang via llvm-commits]

================
@@ -518,6 +522,23 @@ bool AMDGPUCombinerHelper::matchCombineFmulWithSelectToFldexp(
   return true;
 }
 
+bool AMDGPUCombinerHelper::matchCombineFmulWithSelectToFldexp(
+    MachineInstr &MI, MachineInstr &Sel,
+    std::function<void(MachineIRBuilder &)> &MatchInfo) const {
+  Register Dst = MI.getOperand(0).getReg();
+  LLT ScalarDestTy = MRI.getType(Dst).getScalarType();
+
+  // fldexp has no SALU form. On targets with SALU float, defer this combine
+  // to the RegBankCombiner where register banks are known and we can limit it
+  // to VGPR (divergent) values only.
+  if (STI.hasSALUFloatInsts() &&
+      (ScalarDestTy == LLT::scalar(64) || ScalarDestTy == LLT::scalar(32) ||
+       ScalarDestTy == LLT::scalar(16)))
+    return false;
----------------
badumbatish wrote:

would it be simpler to gate this to
```
  if (STI.hasSALUFloatInsts() &&ScalarDestTy == LLT::scalar(32))
      return false;
```

this addresses the original issue and doesn't complicate the if-else case too much, leaving 16 and 64 case to early combine

https://github.com/llvm/llvm-project/pull/190236