[llvm] [SLP] Support ordered fadd reduction via reduction intrinsics (PR #189451)

[Ryan Buchner via llvm-commits]

================
@@ -29046,6 +29138,239 @@ class HorizontalReduction {
     return VectorizedTree;
   }
 
+  /// Attempt to vectorize an ordered (linearized) reduction chain.
+  /// Reduced values from matchOrderedReduction() are in accumulation order.
+  /// Vectorized subsets are immediately reduced via ordered reduction
+  /// intrinsics; non-vectorized values are folded linearly.
+  Value *tryToReduceOrdered(BoUpSLP &V, const DataLayout &DL,
+                            TargetTransformInfo *TTI,
+                            const TargetLibraryInfo &TLI, AssumptionCache *AC,
+                            DominatorTree &DT) {
+    constexpr unsigned RegMaxNumber = 4;
+    constexpr unsigned RedValsMaxNumber = 128;
+
+    assert(RK == ReductionOrdering::Ordered && "Expected ordered reduction");
+    assert(ReducedVals.size() == 1 &&
+           "Expected single group from matchOrderedReduction");
+
+    IRBuilder<TargetFolder> Builder(ReductionRoot->getContext(),
+                                    TargetFolder(DL));
+    Instruction *RdxRootInst = cast<Instruction>(ReductionRoot);
+    Builder.SetInsertPoint(RdxRootInst);
+
+    SmallVector<Value *> Candidates(ReducedVals.back());
+
+    // Intersect the fast-math-flags from all reduction operations.
+    FastMathFlags RdxFMF;
+    RdxFMF.set();
+    for (Value *RdxVal : Candidates)
+      for (Instruction *Op : ReducedValsToOps.at(RdxVal))
+        if (auto *FPMO = dyn_cast<FPMathOperator>(Op))
+          RdxFMF &= FPMO->getFastMathFlags();
----------------
bababuck wrote:

In a test case where I set the `reassoc` flag, I noticed that the output is `reassoc llvm.vector.reduce.fadd`, but I was wondering if it would be better to output `llvm.vector.partial.reduce.fadd` in such cases
 e.g
```
  %add28 = fadd reassoc float %add6, %add11
  %add29 = fadd reassoc float %add28, %add16
  %add30 = fadd reassoc float %add29, %add21
```

https://github.com/llvm/llvm-project/pull/189451