[llvm] [LV] Vectorize FMax via OrderedFCmpSelect w/o fast-math flags. (PR #146711)

[via llvm-commits]

================
@@ -654,7 +654,105 @@ void VPlanTransforms::attachCheckBlock(VPlan &Plan, Value *Cond,
   }
 }
 
-bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
+static bool handleOrderedFCmpSelect(VPlan &Plan,
+                                    VPReductionPHIRecipe *RedPhiR) {
+  VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
+  VPWidenIntOrFpInductionRecipe *WideIV = nullptr;
+
+  // MaxOp feeding the reduction phi must be a select (either wide or a
+  // replicate recipe), where the phi is the last operand, and the compare
+  // predicate is strict. This ensures NaNs won't get propagated unless the
+  // initial value is NaN
+  auto *MaxOp = dyn_cast<VPRecipeWithIRFlags>(
+      RedPhiR->getBackedgeValue()->getDefiningRecipe());
+  if (!MaxOp)
+    return false;
+  auto *RepR = dyn_cast<VPReplicateRecipe>(MaxOp);
+  if (!isa<VPWidenSelectRecipe>(MaxOp) &&
+      !(RepR && (isa<SelectInst>(RepR->getUnderlyingInstr()))))
+    return false;
+
+  auto *Cmp = cast<VPRecipeWithIRFlags>(MaxOp->getOperand(0));
+  if (MaxOp->getOperand(1) == RedPhiR ||
+      !CmpInst::isStrictPredicate(Cmp->getPredicate()))
+    return false;
+
+  for (auto &R : LoopRegion->getEntryBasicBlock()->phis()) {
+    // We need a wide canonical IV
+    if (auto *CurIV = dyn_cast<VPWidenIntOrFpInductionRecipe>(&R)) {
+      if (CurIV->isCanonical()) {
+        WideIV = CurIV;
+        break;
+      }
+    }
+  }
+
+  // A wide canonical IV is currently required.
+  // TODO: Create an induction if no suitable existing one is available.
+  if (!WideIV)
+    return false;
+
+  // Create a reduction that tracks the first indices where the latest maximum
----------------
ayalz wrote:

```suggestion
  // Create a reduction that tracks the first indices where the running maximum
```

https://github.com/llvm/llvm-project/pull/146711