[llvm] [LoopUnroll] Introduce parallel accumulators when unrolling FP reductions. (PR #166630)

[Julian Nagele via llvm-commits]

================
@@ -1256,14 +1257,19 @@ llvm::canParallelizeReductionWhenUnrolling(PHINode &Phi, Loop *L,
     return std::nullopt;
   RecurKind RK = RdxDesc.getRecurrenceKind();
   // Skip unsupported reductions.
-  // TODO: Handle additional reductions, including FP and min-max
-  // reductions.
-  if (!RecurrenceDescriptor::isIntegerRecurrenceKind(RK) ||
+  // TODO: Handle additional reductions, including min-max reductions.
+  if (!(RecurrenceDescriptor::isIntegerRecurrenceKind(RK) ||
+        RecurrenceDescriptor::isFloatingPointRecurrenceKind(RK)) ||
       RecurrenceDescriptor::isAnyOfRecurrenceKind(RK) ||
       RecurrenceDescriptor::isFindIVRecurrenceKind(RK) ||
       RecurrenceDescriptor::isMinMaxRecurrenceKind(RK))
     return std::nullopt;
 
+  if (RecurrenceDescriptor::isFloatingPointRecurrenceKind(RK)) {
+    if (!RdxDesc.getFastMathFlags().allowReassoc())
+      return std::nullopt;
+  }
----------------
juliannagele wrote:

Hm, afaict `isRecurrenceInstr` doesn't return `None` but remembers the instruction that doesn't have reassoc
```
InstDesc(Kind == RecurKind::FAdd, I,
                    I->hasAllowReassoc() ? nullptr : I);
```
So I think we do need a check but we could simplify it to
```
  if (RdxDesc.hasExactFPMath())
      return std::nullopt;
```
where
```
  /// Returns true if the recurrence has floating-point math that requires
  /// precise (ordered) operations.
  bool hasExactFPMath() const { return ExactFPMathInst != nullptr; }
```

https://github.com/llvm/llvm-project/pull/166630