[llvm] [CodeGen] Expand fmaximum/fminimum reductions (PR #195169)

[Matt Arsenault via llvm-commits]

================
@@ -158,6 +163,16 @@ bool expandReductions(Function &F, const TargetTransformInfo *TTI,
       Rdx = getShuffleReduction(Builder, Vec, RdxOpcode, RS, RK);
       break;
     }
+    case Intrinsic::vector_reduce_fmaximum:
+    case Intrinsic::vector_reduce_fminimum: {
+      Value *Vec = II->getArgOperand(0);
+      if (!isPowerOf2_32(
+              cast<FixedVectorType>(Vec->getType())->getNumElements()))
+        continue;
----------------
arsenm wrote:

```suggestion
```

I think this is completely illegitimate, regardless of whatever existing code is doing which should also be removed. Lowering passes should be comprehensive. This also needs tests 

https://github.com/llvm/llvm-project/pull/195169