[llvm] [SLP]Initial support for ordered reductions (PR #182644)

[Ryan Buchner via llvm-commits]

================
@@ -25349,31 +25354,38 @@ class HorizontalReduction {
   }
 
   /// Checks if instruction is associative and can be vectorized.
-  static bool isVectorizable(RecurKind Kind, Instruction *I,
-                             bool TwoElementReduction = false) {
+  enum class ReductionOrdering { Unordered, Ordered, None };
+  ReductionOrdering RK = ReductionOrdering::None;
+  static ReductionOrdering isVectorizable(RecurKind Kind, Instruction *I,
+                                          bool TwoElementReduction = false) {
     if (Kind == RecurKind::None)
-      return false;
+      return ReductionOrdering::None;
 
     // Integer ops that map to select instructions or intrinsics are fine.
     if (RecurrenceDescriptor::isIntMinMaxRecurrenceKind(Kind) ||
         isBoolLogicOp(I))
-      return true;
+      return ReductionOrdering::Unordered;
 
     // No need to check for associativity, if 2 reduced values.
     if (TwoElementReduction)
-      return true;
+      return ReductionOrdering::Unordered;
 
     if (Kind == RecurKind::FMax || Kind == RecurKind::FMin) {
       // FP min/max are associative except for NaN and -0.0. We do not
       // have to rule out -0.0 here because the intrinsic semantics do not
       // specify a fixed result for it.
-      return I->getFastMathFlags().noNaNs();
+      return I->getFastMathFlags().noNaNs() ? ReductionOrdering::Unordered
+                                            : ReductionOrdering::Ordered;
     }
 
     if (Kind == RecurKind::FMaximum || Kind == RecurKind::FMinimum)
-      return true;
+      return ReductionOrdering::Unordered;
 
-    return I->isAssociative();
+    if (I->isAssociative())
+      return ReductionOrdering::Unordered;
+
+    return ::isCommutative(I) ? ReductionOrdering::Ordered
+                              : ReductionOrdering::None;
----------------
bababuck wrote:

Would you mind if I have a go at it? (I have a couple things on my plate ahead of it so might be a couple weeks before I get to it).

https://github.com/llvm/llvm-project/pull/182644