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

[Shilei Tian via llvm-commits]

================
@@ -463,8 +463,14 @@ bool AMDGPUCombinerHelper::matchCombineFmulWithSelectToFldexp(
   LLT DestTy = MRI.getType(Dst);
   LLT ScalarDestTy = DestTy.getScalarType();
 
-  if ((ScalarDestTy != LLT::float64() && ScalarDestTy != LLT::float32() &&
-       ScalarDestTy != LLT::float16()) ||
+  if (MUI && !MUI->isDivergent(Dst) && this->STI.hasSALUFloatInsts() &&
----------------
shiltian wrote:

```suggestion
  if (MUI && !MUI->isDivergent(Dst) && STI.hasSALUFloatInsts() &&
```

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