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

[Simon Pilgrim via llvm-commits]

================
@@ -158,6 +160,19 @@ 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: {
+      // We require "nnan" to use a shuffle reduction; "nsz" is implied by the
----------------
RKSimon wrote:

> So this pass provides incomplete legalization? 

Precisely - the expand-reductions pass only expands pow2 cases (except for sequential fadd cases which is does scalarize as expected) as a helper pass prior to DAG builder - the rest is done much later in DAG legalisation by TargetLowering::expandVecReduce

Neither do a good job of it imo :/

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