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

Generalize handleMaxMinNumReductions to handle any number of
F(Max|Min)Num reductions by collecting a vector of reductions to
convert.

We then add NaN checks for all of them, followed by adjusting the branch
controlling the vector loop region, and updating the resume phis.

Addresses a TODO from https://github.com/llvm/llvm-project/pull/148239

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

6 files changed, 259 insertions, 96 deletions

diff --git a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
index 1a66d2049a8d..f405c40611fc 100644
--- a/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
+++ b/llvm/lib/Transforms/Vectorize/VPlanConstruction.cpp
@@ -823,7 +823,8 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
   };
 
   VPRegionBlock *LoopRegion = Plan.getVectorLoopRegion();
-  VPReductionPHIRecipe *RedPhiR = nullptr;
+  SmallVector<std::pair<VPReductionPHIRecipe *, VPValue *>>
+      MinMaxNumReductionsToHandle;
   bool HasUnsupportedPhi = false;
   for (auto &R : LoopRegion->getEntryBasicBlock()->phis()) {
     if (isa<VPCanonicalIVPHIRecipe, VPWidenIntOrFpInductionRecipe>(&R))
@@ -834,19 +835,20 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
       HasUnsupportedPhi = true;
       continue;
     }
-    // For now, only a single reduction is supported.
-    // TODO: Support multiple MaxNum/MinNum reductions and other reductions.
-    if (RedPhiR)
-      return false;
     if (!RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(
             Cur->getRecurrenceKind())) {
       HasUnsupportedPhi = true;
       continue;
     }
-    RedPhiR = Cur;
+
+    VPValue *MinMaxOp = GetMinMaxCompareValue(Cur);
+    if (!MinMaxOp)
+      return false;
+
+    MinMaxNumReductionsToHandle.emplace_back(Cur, MinMaxOp);
   }
 
-  if (!RedPhiR)
+  if (MinMaxNumReductionsToHandle.empty())
     return true;
 
   // We won't be able to resume execution in the scalar tail, if there are
@@ -855,14 +857,6 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
   if (HasUnsupportedPhi || !Plan.hasScalarTail())
     return false;
 
-  VPValue *MinMaxOp = GetMinMaxCompareValue(RedPhiR);
-  if (!MinMaxOp)
-    return false;
-
-  assert(RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(
-             RedPhiR->getRecurrenceKind()) &&
-         "unsupported reduction");
-
   /// Check if the vector loop of \p Plan can early exit and restart
   /// execution of last vector iteration in the scalar loop. This requires all
   /// recipes up to early exit point be side-effect free as they are
@@ -879,52 +873,68 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
   }
 
   VPBasicBlock *LatchVPBB = LoopRegion->getExitingBasicBlock();
-  VPBuilder Builder(LatchVPBB->getTerminator());
-  auto *LatchExitingBranch = cast<VPInstruction>(LatchVPBB->getTerminator());
-  assert(LatchExitingBranch->getOpcode() == VPInstruction::BranchOnCount &&
+  VPBuilder LatchBuilder(LatchVPBB->getTerminator());
+  VPValue *AllNaNLanes = nullptr;
+  SmallPtrSet<VPValue *, 2> RdxResults;
+  for (const auto &[_, MinMaxOp] : MinMaxNumReductionsToHandle) {
+    VPValue *RedNaNLanes =
+        LatchBuilder.createFCmp(CmpInst::FCMP_UNO, MinMaxOp, MinMaxOp);
+    AllNaNLanes = AllNaNLanes ? LatchBuilder.createOr(AllNaNLanes, RedNaNLanes)
+                              : RedNaNLanes;
+  }
+
+  VPValue *AnyNaNLane =
+      LatchBuilder.createNaryOp(VPInstruction::AnyOf, {AllNaNLanes});
+  VPBasicBlock *MiddleVPBB = Plan.getMiddleBlock();
+  VPBuilder MiddleBuilder(MiddleVPBB, MiddleVPBB->begin());
+  for (const auto &[RedPhiR, _] : MinMaxNumReductionsToHandle) {
+    assert(RecurrenceDescriptor::isFPMinMaxNumRecurrenceKind(
+               RedPhiR->getRecurrenceKind()) &&
+           "unsupported reduction");
+
+    // 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->users(), [](VPUser *U, bool) -> VPSingleDefRecipe * {
+          auto *VPI = dyn_cast<VPInstruction>(U);
+          if (VPI && VPI->getOpcode() == VPInstruction::ComputeReductionResult)
+            return VPI;
+          return nullptr;
+        });
+
+    auto *NewSel = MiddleBuilder.createSelect(AnyNaNLane, RedPhiR,
+                                              RdxResult->getOperand(1));
+    RdxResult->setOperand(1, NewSel);
+    assert(!RdxResults.contains(RdxResult) && "RdxResult already used");
+    RdxResults.insert(RdxResult);
+  }
+
+  auto *LatchExitingBranch = LatchVPBB->getTerminator();
+  assert(match(LatchExitingBranch, m_BranchOnCount(m_VPValue(), m_VPValue())) &&
          "Unexpected terminator");
-  auto *IsLatchExitTaken =
-      Builder.createICmp(CmpInst::ICMP_EQ, LatchExitingBranch->getOperand(0),
-                         LatchExitingBranch->getOperand(1));
-
-  VPValue *IsNaN = Builder.createFCmp(CmpInst::FCMP_UNO, MinMaxOp, MinMaxOp);
-  VPValue *AnyNaN = Builder.createNaryOp(VPInstruction::AnyOf, {IsNaN});
-  auto *AnyExitTaken =
-      Builder.createNaryOp(Instruction::Or, {AnyNaN, IsLatchExitTaken});
-  Builder.createNaryOp(VPInstruction::BranchOnCond, AnyExitTaken);
+  auto *IsLatchExitTaken = LatchBuilder.createICmp(
+      CmpInst::ICMP_EQ, LatchExitingBranch->getOperand(0),
+      LatchExitingBranch->getOperand(1));
+  auto *AnyExitTaken = LatchBuilder.createNaryOp(
+      Instruction::Or, {AnyNaNLane, IsLatchExitTaken});
+  LatchBuilder.createNaryOp(VPInstruction::BranchOnCond, AnyExitTaken);
   LatchExitingBranch->eraseFromParent();
 
-  // 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->users(), [](VPUser *U, bool) -> VPSingleDefRecipe * {
-        auto *VPI = dyn_cast<VPInstruction>(U);
-        if (VPI && VPI->getOpcode() == VPInstruction::ComputeReductionResult)
-          return VPI;
-        return nullptr;
-      });
-
-  auto *MiddleVPBB = Plan.getMiddleBlock();
-  Builder.setInsertPoint(MiddleVPBB, MiddleVPBB->begin());
-  auto *NewSel =
-      Builder.createSelect(AnyNaN, RedPhiR, RdxResult->getOperand(1));
-  RdxResult->setOperand(1, NewSel);
-
-  auto *ScalarPH = Plan.getScalarPreheader();
-  // Update resume phis for inductions in the scalar preheader. If AnyNaN is
+  // Update resume phis for inductions in the scalar preheader. If AnyNaNLane is
   // true, the resume from the start of the last vector iteration via the
   // canonical IV, otherwise from the original value.
-  for (auto &R : ScalarPH->phis()) {
+  for (auto &R : Plan.getScalarPreheader()->phis()) {
     auto *ResumeR = cast<VPPhi>(&R);
     VPValue *VecV = ResumeR->getOperand(0);
-    if (VecV == RdxResult)
+    if (RdxResults.contains(VecV))
       continue;
     if (auto *DerivedIV = dyn_cast<VPDerivedIVRecipe>(VecV)) {
       if (DerivedIV->getNumUsers() == 1 &&
           DerivedIV->getOperand(1) == &Plan.getVectorTripCount()) {
-        auto *NewSel = Builder.createSelect(
-            AnyNaN, LoopRegion->getCanonicalIV(), &Plan.getVectorTripCount());
-        DerivedIV->moveAfter(&*Builder.getInsertPoint());
+        auto *NewSel =
+            MiddleBuilder.createSelect(AnyNaNLane, LoopRegion->getCanonicalIV(),
+                                       &Plan.getVectorTripCount());
+        DerivedIV->moveAfter(&*MiddleBuilder.getInsertPoint());
         DerivedIV->setOperand(1, NewSel);
         continue;
       }
@@ -936,15 +946,16 @@ bool VPlanTransforms::handleMaxMinNumReductions(VPlan &Plan) {
                            "FMaxNum/FMinNum reduction.\n");
       return false;
     }
-    auto *NewSel =
-        Builder.createSelect(AnyNaN, LoopRegion->getCanonicalIV(), VecV);
+    auto *NewSel = MiddleBuilder.createSelect(
+        AnyNaNLane, LoopRegion->getCanonicalIV(), VecV);
     ResumeR->setOperand(0, NewSel);
   }
 
   auto *MiddleTerm = MiddleVPBB->getTerminator();
-  Builder.setInsertPoint(MiddleTerm);
+  MiddleBuilder.setInsertPoint(MiddleTerm);
   VPValue *MiddleCond = MiddleTerm->getOperand(0);
-  VPValue *NewCond = Builder.createAnd(MiddleCond, Builder.createNot(AnyNaN));
+  VPValue *NewCond =
+      MiddleBuilder.createAnd(MiddleCond, MiddleBuilder.createNot(AnyNaNLane));
   MiddleTerm->setOperand(0, NewCond);
   return true;
 }
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/fmax-without-fast-math-flags.ll b/llvm/test/Transforms/LoopVectorize/AArch64/fmax-without-fast-math-flags.ll
index 3c83c01929aa..7e58d9d6a8ec 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/fmax-without-fast-math-flags.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/fmax-without-fast-math-flags.ll
@@ -59,13 +59,13 @@ define float @fmaxnum(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[TMP7]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD]])
 ; CHECK-NEXT:    [[TMP8]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD2]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 8
-; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP4:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD2]]
 ; CHECK-NEXT:    [[TMP18:%.*]] = freeze <4 x i1> [[TMP3]]
 ; CHECK-NEXT:    [[TMP15:%.*]] = freeze <4 x i1> [[TMP4]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = or <4 x i1> [[TMP18]], [[TMP15]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP5]])
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = or i1 [[TMP6]], [[TMP9]]
 ; CHECK-NEXT:    br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -112,27 +112,84 @@ exit:
   ret float %max.next
 }
 
+; TODO: Could fold pairs of `fcmp uno` together.
 define float @test_fmax_and_fmin(ptr %src.0, ptr %src.1, i64 %n) {
 ; CHECK-LABEL: define float @test_fmax_and_fmin(
 ; CHECK-SAME: ptr [[SRC_0:%.*]], ptr [[SRC_1:%.*]], i64 [[N:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*]]:
-; CHECK-NEXT:    br label %[[LOOP:.*]]
-; CHECK:       [[LOOP]]:
-; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MIN:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MAX:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 8
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 8
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP6:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP7:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP4:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP5:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_0]], i64 [[IV]]
 ; CHECK-NEXT:    [[GEP_SRC_1:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_1]], i64 [[IV]]
-; CHECK-NEXT:    [[L_0:%.*]] = load float, ptr [[GEP_SRC_0]], align 4
-; CHECK-NEXT:    [[L_1:%.*]] = load float, ptr [[GEP_SRC_1]], align 4
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[GEP_SRC_0]], i32 4
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC_0]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD4:%.*]] = load <4 x float>, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds nuw float, ptr [[GEP_SRC_1]], i32 4
+; CHECK-NEXT:    [[WIDE_LOAD5:%.*]] = load <4 x float>, ptr [[GEP_SRC_1]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD6:%.*]] = load <4 x float>, ptr [[TMP3]], align 4
+; CHECK-NEXT:    [[TMP4]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI2]], <4 x float> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP5]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI3]], <4 x float> [[WIDE_LOAD4]])
+; CHECK-NEXT:    [[TMP6]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD5]])
+; CHECK-NEXT:    [[TMP7]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD6]])
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 8
+; CHECK-NEXT:    [[TMP8:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD5]], [[WIDE_LOAD5]]
+; CHECK-NEXT:    [[TMP9:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD6]], [[WIDE_LOAD6]]
+; CHECK-NEXT:    [[TMP14:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP15:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD4]], [[WIDE_LOAD4]]
+; CHECK-NEXT:    [[TMP12:%.*]] = or <4 x i1> [[TMP8]], [[TMP14]]
+; CHECK-NEXT:    [[TMP13:%.*]] = or <4 x i1> [[TMP9]], [[TMP15]]
+; CHECK-NEXT:    [[TMP16:%.*]] = freeze <4 x i1> [[TMP12]]
+; CHECK-NEXT:    [[TMP17:%.*]] = freeze <4 x i1> [[TMP13]]
+; CHECK-NEXT:    [[TMP18:%.*]] = or <4 x i1> [[TMP16]], [[TMP17]]
+; CHECK-NEXT:    [[TMP19:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP18]])
+; CHECK-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP20:%.*]] = or i1 [[TMP19]], [[TMP21]]
+; CHECK-NEXT:    br i1 [[TMP20]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP23:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI]], <4 x float> [[TMP6]]
+; CHECK-NEXT:    [[TMP24:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI1]], <4 x float> [[TMP7]]
+; CHECK-NEXT:    [[TMP25:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI2]], <4 x float> [[TMP4]]
+; CHECK-NEXT:    [[TMP26:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI3]], <4 x float> [[TMP5]]
+; CHECK-NEXT:    [[TMP27:%.*]] = select i1 [[TMP19]], i64 [[IV]], i64 [[N_VEC]]
+; CHECK-NEXT:    [[RDX_MINMAX:%.*]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[TMP23]], <4 x float> [[TMP24]])
+; CHECK-NEXT:    [[TMP28:%.*]] = call float @llvm.vector.reduce.fmin.v4f32(<4 x float> [[RDX_MINMAX]])
+; CHECK-NEXT:    [[RDX_MINMAX9:%.*]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[TMP25]], <4 x float> [[TMP26]])
+; CHECK-NEXT:    [[TMP29:%.*]] = call float @llvm.vector.reduce.fmax.v4f32(<4 x float> [[RDX_MINMAX9]])
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP30:%.*]] = xor i1 [[TMP19]], true
+; CHECK-NEXT:    [[TMP31:%.*]] = and i1 [[CMP_N]], [[TMP30]]
+; CHECK-NEXT:    br i1 [[TMP31]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[TMP27]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP28]], %[[MIDDLE_BLOCK]] ], [ 0.000000e+00, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX8:%.*]] = phi float [ [[TMP29]], %[[MIDDLE_BLOCK]] ], [ 0.000000e+00, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV1:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MIN:%.*]] = phi float [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MAX:%.*]] = phi float [ [[BC_MERGE_RDX8]], %[[SCALAR_PH]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[GEP_SRC_2:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_0]], i64 [[IV1]]
+; CHECK-NEXT:    [[GEP_SRC_3:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_1]], i64 [[IV1]]
+; CHECK-NEXT:    [[L_0:%.*]] = load float, ptr [[GEP_SRC_2]], align 4
+; CHECK-NEXT:    [[L_1:%.*]] = load float, ptr [[GEP_SRC_3]], align 4
 ; CHECK-NEXT:    [[MAX_NEXT]] = tail call noundef float @llvm.maxnum.f32(float [[MAX]], float [[L_0]])
 ; CHECK-NEXT:    [[MIN_NEXT]] = tail call noundef float @llvm.minnum.f32(float [[MIN]], float [[L_1]])
-; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV1]], 1
 ; CHECK-NEXT:    [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
-; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
+; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[MAX_NEXT_LCSSA:%.*]] = phi float [ [[MAX_NEXT]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MIN_NEXT_LCSSA:%.*]] = phi float [ [[MIN_NEXT]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MAX_NEXT_LCSSA:%.*]] = phi float [ [[MAX_NEXT]], %[[LOOP]] ], [ [[TMP29]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[MIN_NEXT_LCSSA:%.*]] = phi float [ [[MIN_NEXT]], %[[LOOP]] ], [ [[TMP28]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[MAX_NEXT_LCSSA]], [[MIN_NEXT_LCSSA]]
 ; CHECK-NEXT:    ret float [[SUB]]
 ;
diff --git a/llvm/test/Transforms/LoopVectorize/AArch64/fmin-without-fast-math-flags.ll b/llvm/test/Transforms/LoopVectorize/AArch64/fmin-without-fast-math-flags.ll
index 711a9cd03ac1..1cc4c152649b 100644
--- a/llvm/test/Transforms/LoopVectorize/AArch64/fmin-without-fast-math-flags.ll
+++ b/llvm/test/Transforms/LoopVectorize/AArch64/fmin-without-fast-math-flags.ll
@@ -59,13 +59,13 @@ define float @fminnum(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[TMP7]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD]])
 ; CHECK-NEXT:    [[TMP8]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD2]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 8
-; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP4:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD2]]
 ; CHECK-NEXT:    [[TMP15:%.*]] = freeze <4 x i1> [[TMP3]]
 ; CHECK-NEXT:    [[TMP18:%.*]] = freeze <4 x i1> [[TMP4]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = or <4 x i1> [[TMP15]], [[TMP18]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP5]])
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = or i1 [[TMP6]], [[TMP9]]
 ; CHECK-NEXT:    br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
diff --git a/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags-interleave.ll b/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags-interleave.ll
index af648df9fc5c..01fab87209a3 100644
--- a/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags-interleave.ll
+++ b/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags-interleave.ll
@@ -59,13 +59,13 @@ define float @fmaxnum(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[TMP7]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD]])
 ; CHECK-NEXT:    [[TMP8]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD2]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 8
-; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP4:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD2]]
 ; CHECK-NEXT:    [[TMP15:%.*]] = freeze <4 x i1> [[TMP3]]
 ; CHECK-NEXT:    [[TMP18:%.*]] = freeze <4 x i1> [[TMP4]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = or <4 x i1> [[TMP15]], [[TMP18]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP5]])
+; CHECK-NEXT:    [[TMP9:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = or i1 [[TMP6]], [[TMP9]]
 ; CHECK-NEXT:    br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -116,23 +116,79 @@ define float @test_fmax_and_fmin(ptr %src.0, ptr %src.1, i64 %n) {
 ; CHECK-LABEL: define float @test_fmax_and_fmin(
 ; CHECK-SAME: ptr [[SRC_0:%.*]], ptr [[SRC_1:%.*]], i64 [[N:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*]]:
-; CHECK-NEXT:    br label %[[LOOP:.*]]
-; CHECK:       [[LOOP]]:
-; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MIN:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MAX:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 8
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 8
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP6:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP7:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI2:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP4:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI3:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP5:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_0]], i64 [[IV]]
 ; CHECK-NEXT:    [[GEP_SRC_1:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_1]], i64 [[IV]]
-; CHECK-NEXT:    [[L_0:%.*]] = load float, ptr [[GEP_SRC_0]], align 4
-; CHECK-NEXT:    [[L_1:%.*]] = load float, ptr [[GEP_SRC_1]], align 4
+; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds nuw float, ptr [[GEP_SRC_0]], i32 4
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC_0]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD4:%.*]] = load <4 x float>, ptr [[TMP2]], align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds nuw float, ptr [[GEP_SRC_1]], i32 4
+; CHECK-NEXT:    [[WIDE_LOAD5:%.*]] = load <4 x float>, ptr [[GEP_SRC_1]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD6:%.*]] = load <4 x float>, ptr [[TMP3]], align 4
+; CHECK-NEXT:    [[TMP4]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI2]], <4 x float> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP5]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI3]], <4 x float> [[WIDE_LOAD4]])
+; CHECK-NEXT:    [[TMP6]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD5]])
+; CHECK-NEXT:    [[TMP7]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD6]])
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 8
+; CHECK-NEXT:    [[TMP8:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD5]], [[WIDE_LOAD5]]
+; CHECK-NEXT:    [[TMP9:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD6]], [[WIDE_LOAD6]]
+; CHECK-NEXT:    [[TMP14:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP15:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD4]], [[WIDE_LOAD4]]
+; CHECK-NEXT:    [[TMP12:%.*]] = or <4 x i1> [[TMP8]], [[TMP14]]
+; CHECK-NEXT:    [[TMP13:%.*]] = or <4 x i1> [[TMP9]], [[TMP15]]
+; CHECK-NEXT:    [[TMP16:%.*]] = freeze <4 x i1> [[TMP12]]
+; CHECK-NEXT:    [[TMP17:%.*]] = freeze <4 x i1> [[TMP13]]
+; CHECK-NEXT:    [[TMP18:%.*]] = or <4 x i1> [[TMP16]], [[TMP17]]
+; CHECK-NEXT:    [[TMP19:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP18]])
+; CHECK-NEXT:    [[TMP21:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP20:%.*]] = or i1 [[TMP19]], [[TMP21]]
+; CHECK-NEXT:    br i1 [[TMP20]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP23:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI]], <4 x float> [[TMP6]]
+; CHECK-NEXT:    [[TMP24:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI1]], <4 x float> [[TMP7]]
+; CHECK-NEXT:    [[TMP25:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI2]], <4 x float> [[TMP4]]
+; CHECK-NEXT:    [[TMP26:%.*]] = select i1 [[TMP19]], <4 x float> [[VEC_PHI3]], <4 x float> [[TMP5]]
+; CHECK-NEXT:    [[TMP27:%.*]] = select i1 [[TMP19]], i64 [[IV]], i64 [[N_VEC]]
+; CHECK-NEXT:    [[RDX_MINMAX:%.*]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[TMP23]], <4 x float> [[TMP24]])
+; CHECK-NEXT:    [[TMP28:%.*]] = call float @llvm.vector.reduce.fmin.v4f32(<4 x float> [[RDX_MINMAX]])
+; CHECK-NEXT:    [[RDX_MINMAX9:%.*]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[TMP25]], <4 x float> [[TMP26]])
+; CHECK-NEXT:    [[TMP29:%.*]] = call float @llvm.vector.reduce.fmax.v4f32(<4 x float> [[RDX_MINMAX9]])
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP30:%.*]] = xor i1 [[TMP19]], true
+; CHECK-NEXT:    [[TMP31:%.*]] = and i1 [[CMP_N]], [[TMP30]]
+; CHECK-NEXT:    br i1 [[TMP31]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[TMP27]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP28]], %[[MIDDLE_BLOCK]] ], [ 0.000000e+00, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX8:%.*]] = phi float [ [[TMP29]], %[[MIDDLE_BLOCK]] ], [ 0.000000e+00, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV1:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MIN:%.*]] = phi float [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MAX:%.*]] = phi float [ [[BC_MERGE_RDX8]], %[[SCALAR_PH]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[GEP_SRC_2:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_0]], i64 [[IV1]]
+; CHECK-NEXT:    [[GEP_SRC_3:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_1]], i64 [[IV1]]
+; CHECK-NEXT:    [[L_0:%.*]] = load float, ptr [[GEP_SRC_2]], align 4
+; CHECK-NEXT:    [[L_1:%.*]] = load float, ptr [[GEP_SRC_3]], align 4
 ; CHECK-NEXT:    [[MAX_NEXT]] = tail call noundef float @llvm.maxnum.f32(float [[MAX]], float [[L_0]])
 ; CHECK-NEXT:    [[MIN_NEXT]] = tail call noundef float @llvm.minnum.f32(float [[MIN]], float [[L_1]])
-; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV1]], 1
 ; CHECK-NEXT:    [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
-; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
+; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP5:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[MAX_NEXT_LCSSA:%.*]] = phi float [ [[MAX_NEXT]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MIN_NEXT_LCSSA:%.*]] = phi float [ [[MIN_NEXT]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MAX_NEXT_LCSSA:%.*]] = phi float [ [[MAX_NEXT]], %[[LOOP]] ], [ [[TMP29]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[MIN_NEXT_LCSSA:%.*]] = phi float [ [[MIN_NEXT]], %[[LOOP]] ], [ [[TMP28]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[MAX_NEXT_LCSSA]], [[MIN_NEXT_LCSSA]]
 ; CHECK-NEXT:    ret float [[SUB]]
 ;
diff --git a/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags.ll b/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags.ll
index 242df1fcf761..e028bec138fa 100644
--- a/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags.ll
+++ b/llvm/test/Transforms/LoopVectorize/fmax-without-fast-math-flags.ll
@@ -205,10 +205,10 @@ define float @fmaxnum_1(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC]], align 4
 ; CHECK-NEXT:    [[TMP4]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[WIDE_LOAD]], <4 x float> [[VEC_PHI]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP2:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = freeze <4 x i1> [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP10]])
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = or i1 [[TMP3]], [[TMP5]]
 ; CHECK-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -270,10 +270,10 @@ define float @fmaxnum_2(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC]], align 4
 ; CHECK-NEXT:    [[TMP4]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP2:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = freeze <4 x i1> [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP10]])
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = or i1 [[TMP3]], [[TMP5]]
 ; CHECK-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -337,10 +337,10 @@ define float @fmaxnum_induction_starts_at_10(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC]], align 4
 ; CHECK-NEXT:    [[TMP3]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[WIDE_LOAD]], <4 x float> [[VEC_PHI]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP12:%.*]] = freeze <4 x i1> [[TMP5]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP12]])
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP7:%.*]] = or i1 [[TMP6]], [[TMP4]]
 ; CHECK-NEXT:    br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP6:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -405,10 +405,10 @@ define float @fmaxnum_induction_starts_at_value(ptr %src, i64 %start, i64 %n) {
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC]], align 4
 ; CHECK-NEXT:    [[TMP3]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[WIDE_LOAD]], <4 x float> [[VEC_PHI]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
-; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP5:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP12:%.*]] = freeze <4 x i1> [[TMP5]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP12]])
+; CHECK-NEXT:    [[TMP4:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP7:%.*]] = or i1 [[TMP6]], [[TMP4]]
 ; CHECK-NEXT:    br i1 [[TMP7]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP8:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -680,23 +680,62 @@ define float @test_fmax_and_fmax(ptr %src.0, ptr %src.1, i64 %n) {
 ; CHECK-LABEL: define float @test_fmax_and_fmax(
 ; CHECK-SAME: ptr [[SRC_0:%.*]], ptr [[SRC_1:%.*]], i64 [[N:%.*]]) {
 ; CHECK-NEXT:  [[ENTRY:.*]]:
-; CHECK-NEXT:    br label %[[LOOP:.*]]
-; CHECK:       [[LOOP]]:
-; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[ENTRY]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MIN:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MAX:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[N]], 4
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label %[[SCALAR_PH:.*]], label %[[VECTOR_PH:.*]]
+; CHECK:       [[VECTOR_PH]]:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[N]], 4
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[N]], [[N_MOD_VF]]
+; CHECK-NEXT:    br label %[[VECTOR_BODY:.*]]
+; CHECK:       [[VECTOR_BODY]]:
+; CHECK-NEXT:    [[IV:%.*]] = phi i64 [ 0, %[[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP3:%.*]], %[[VECTOR_BODY]] ]
+; CHECK-NEXT:    [[VEC_PHI1:%.*]] = phi <4 x float> [ zeroinitializer, %[[VECTOR_PH]] ], [ [[TMP2:%.*]], %[[VECTOR_BODY]] ]
 ; CHECK-NEXT:    [[GEP_SRC_0:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_0]], i64 [[IV]]
 ; CHECK-NEXT:    [[GEP_SRC_1:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_1]], i64 [[IV]]
-; CHECK-NEXT:    [[L_0:%.*]] = load float, ptr [[GEP_SRC_0]], align 4
-; CHECK-NEXT:    [[L_1:%.*]] = load float, ptr [[GEP_SRC_1]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC_0]], align 4
+; CHECK-NEXT:    [[WIDE_LOAD2:%.*]] = load <4 x float>, ptr [[GEP_SRC_1]], align 4
+; CHECK-NEXT:    [[TMP2]] = call <4 x float> @llvm.maxnum.v4f32(<4 x float> [[VEC_PHI1]], <4 x float> [[WIDE_LOAD]])
+; CHECK-NEXT:    [[TMP3]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD2]])
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
+; CHECK-NEXT:    [[TMP4:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD2]], [[WIDE_LOAD2]]
+; CHECK-NEXT:    [[TMP7:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
+; CHECK-NEXT:    [[TMP6:%.*]] = or <4 x i1> [[TMP4]], [[TMP7]]
+; CHECK-NEXT:    [[TMP8:%.*]] = freeze <4 x i1> [[TMP6]]
+; CHECK-NEXT:    [[TMP9:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP8]])
+; CHECK-NEXT:    [[TMP11:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP10:%.*]] = or i1 [[TMP9]], [[TMP11]]
+; CHECK-NEXT:    br i1 [[TMP10]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP10:![0-9]+]]
+; CHECK:       [[MIDDLE_BLOCK]]:
+; CHECK-NEXT:    [[TMP13:%.*]] = select i1 [[TMP9]], <4 x float> [[VEC_PHI]], <4 x float> [[TMP3]]
+; CHECK-NEXT:    [[TMP14:%.*]] = select i1 [[TMP9]], <4 x float> [[VEC_PHI1]], <4 x float> [[TMP2]]
+; CHECK-NEXT:    [[TMP15:%.*]] = select i1 [[TMP9]], i64 [[IV]], i64 [[N_VEC]]
+; CHECK-NEXT:    [[TMP16:%.*]] = call float @llvm.vector.reduce.fmin.v4f32(<4 x float> [[TMP13]])
+; CHECK-NEXT:    [[TMP17:%.*]] = call float @llvm.vector.reduce.fmax.v4f32(<4 x float> [[TMP14]])
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[N]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP18:%.*]] = xor i1 [[TMP9]], true
+; CHECK-NEXT:    [[TMP19:%.*]] = and i1 [[CMP_N]], [[TMP18]]
+; CHECK-NEXT:    br i1 [[TMP19]], label %[[EXIT:.*]], label %[[SCALAR_PH]]
+; CHECK:       [[SCALAR_PH]]:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[TMP15]], %[[MIDDLE_BLOCK]] ], [ 0, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi float [ [[TMP16]], %[[MIDDLE_BLOCK]] ], [ 0.000000e+00, %[[ENTRY]] ]
+; CHECK-NEXT:    [[BC_MERGE_RDX3:%.*]] = phi float [ [[TMP17]], %[[MIDDLE_BLOCK]] ], [ 0.000000e+00, %[[ENTRY]] ]
+; CHECK-NEXT:    br label %[[LOOP:.*]]
+; CHECK:       [[LOOP]]:
+; CHECK-NEXT:    [[IV1:%.*]] = phi i64 [ [[BC_RESUME_VAL]], %[[SCALAR_PH]] ], [ [[IV_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MIN:%.*]] = phi float [ [[BC_MERGE_RDX]], %[[SCALAR_PH]] ], [ [[MIN_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MAX:%.*]] = phi float [ [[BC_MERGE_RDX3]], %[[SCALAR_PH]] ], [ [[MAX_NEXT:%.*]], %[[LOOP]] ]
+; CHECK-NEXT:    [[GEP_SRC_2:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_0]], i64 [[IV1]]
+; CHECK-NEXT:    [[GEP_SRC_3:%.*]] = getelementptr inbounds nuw float, ptr [[SRC_1]], i64 [[IV1]]
+; CHECK-NEXT:    [[L_0:%.*]] = load float, ptr [[GEP_SRC_2]], align 4
+; CHECK-NEXT:    [[L_1:%.*]] = load float, ptr [[GEP_SRC_3]], align 4
 ; CHECK-NEXT:    [[MAX_NEXT]] = tail call noundef float @llvm.maxnum.f32(float [[MAX]], float [[L_0]])
 ; CHECK-NEXT:    [[MIN_NEXT]] = tail call noundef float @llvm.minnum.f32(float [[MIN]], float [[L_1]])
-; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
+; CHECK-NEXT:    [[IV_NEXT]] = add nuw nsw i64 [[IV1]], 1
 ; CHECK-NEXT:    [[EC:%.*]] = icmp eq i64 [[IV_NEXT]], [[N]]
-; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT:.*]], label %[[LOOP]]
+; CHECK-NEXT:    br i1 [[EC]], label %[[EXIT]], label %[[LOOP]], !llvm.loop [[LOOP11:![0-9]+]]
 ; CHECK:       [[EXIT]]:
-; CHECK-NEXT:    [[MAX_NEXT_LCSSA:%.*]] = phi float [ [[MAX_NEXT]], %[[LOOP]] ]
-; CHECK-NEXT:    [[MIN_NEXT_LCSSA:%.*]] = phi float [ [[MIN_NEXT]], %[[LOOP]] ]
+; CHECK-NEXT:    [[MAX_NEXT_LCSSA:%.*]] = phi float [ [[MAX_NEXT]], %[[LOOP]] ], [ [[TMP17]], %[[MIDDLE_BLOCK]] ]
+; CHECK-NEXT:    [[MIN_NEXT_LCSSA:%.*]] = phi float [ [[MIN_NEXT]], %[[LOOP]] ], [ [[TMP16]], %[[MIDDLE_BLOCK]] ]
 ; CHECK-NEXT:    [[SUB:%.*]] = fsub float [[MAX_NEXT_LCSSA]], [[MIN_NEXT_LCSSA]]
 ; CHECK-NEXT:    ret float [[SUB]]
 ;
diff --git a/llvm/test/Transforms/LoopVectorize/fmin-without-fast-math-flags.ll b/llvm/test/Transforms/LoopVectorize/fmin-without-fast-math-flags.ll
index 7f65306bcbe5..368553dc2a7d 100644
--- a/llvm/test/Transforms/LoopVectorize/fmin-without-fast-math-flags.ll
+++ b/llvm/test/Transforms/LoopVectorize/fmin-without-fast-math-flags.ll
@@ -205,10 +205,10 @@ define float @fminnum_1(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC]], align 4
 ; CHECK-NEXT:    [[TMP4]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[WIDE_LOAD]], <4 x float> [[VEC_PHI]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP2:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = freeze <4 x i1> [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP10]])
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = or i1 [[TMP3]], [[TMP5]]
 ; CHECK-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]:
@@ -270,10 +270,10 @@ define float @fminnum_2(ptr %src, i64 %n) {
 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x float>, ptr [[GEP_SRC]], align 4
 ; CHECK-NEXT:    [[TMP4]] = call <4 x float> @llvm.minnum.v4f32(<4 x float> [[VEC_PHI]], <4 x float> [[WIDE_LOAD]])
 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[IV]], 4
-; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP2:%.*]] = fcmp uno <4 x float> [[WIDE_LOAD]], [[WIDE_LOAD]]
 ; CHECK-NEXT:    [[TMP10:%.*]] = freeze <4 x i1> [[TMP2]]
 ; CHECK-NEXT:    [[TMP3:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP10]])
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = or i1 [[TMP3]], [[TMP5]]
 ; CHECK-NEXT:    br i1 [[TMP6]], label %[[MIDDLE_BLOCK:.*]], label %[[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
 ; CHECK:       [[MIDDLE_BLOCK]]: