[llvm] [VPlan] Support multiple F(Max|Min)Num reductions. (PR #161735)

[Mel Chen via llvm-commits]

================
@@ -880,52 +868,72 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
   }
 
   VPBasicBlock *LatchVPBB = LoopRegion->getExitingBasicBlock();
+  VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
+  VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->begin());
   VPBuilder Builder(LatchVPBB->getTerminator());
-  auto *LatchExitingBranch = cast<VPInstruction>(LatchVPBB->getTerminator());
-  assert(LatchExitingBranch->getOpcode() == VPInstruction::BranchOnCount &&
+  VPValue *AnyNaN = nullptr;
+  SmallPtrSet<VPValue *, 2> RdxResults;
+  for (VPReductionPHIRecipe *RedPhiR : ReductionsToConvert) {
+    assert(RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(
+               RedPhiR->getRecurrenceKind()) &&
+           "unsupported reduction");
+
+    VPValue *MinMaxOp = GetMinMaxCompareValue(RedPhiR);
+    if (!MinMaxOp)
+      return false;
+
+    VPValue *IsNaN = Builder.createFCmp(CmpInst::FCMP_UNO, MinMaxOp, MinMaxOp);
+    VPValue *HasNaN = Builder.createNaryOp(VPInstruction::AnyOf, {IsNaN});
+    if (AnyNaN)
+      AnyNaN = Builder.createOr(AnyNaN, HasNaN);
+    else
+      AnyNaN = HasNaN;
+
+    // If we exit early due to NaNs, compute the final reduction result based
+    // on the reduction phi at the beginning of the last vector iteration.
+    auto *RdxResult = find_singleton<VPSingleDefRecipe>(
+        RedPhiR->getBackedgeValue()->users(),
+        [RedPhiR](VPUser *U, bool) -> VPSingleDefRecipe * {
+          auto *VPI = dyn_cast<VPInstruction>(U);
+          if (VPI && VPI->getOpcode() == VPInstruction::ComputeReductionResult)
+            return VPI;
+          assert(U == RedPhiR &&
+                 "Backedge value must only be used by "
+                 "ComputeReductionResult and the reduction phi");
----------------
Mel-Chen wrote:

If CM.usePredicatedReductionSelect() is false, there might be a DU chain like backedge recipe --> select --> ComputeReductionResult. Would this still be able to find ComputeReductionResult correctly?

https://github.com/llvm/llvm-project/pull/161735