llvm/test/Transforms/LoopInterchange/fp-reductions.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt < %s -passes=loop-interchange -cache-line-size=64 -S | FileCheck %s

; Check that the loops aren't exchanged if there is a reduction of
; non-reassociative floating-point addition.
;
; float sum = 0;
; for (int i = 0; i < 2; i++)
;   for (int j = 0; j < 2; j++)
;     sum += A[j][i];

define void @reduction_fadd(ptr %A) {
; CHECK-LABEL: define void @reduction_fadd(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ]
; CHECK-NEXT:    [[SUM_I:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[SUM_I_LCSSA:%.*]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ 0, %[[FOR_I_HEADER]] ], [ [[J_INC:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[SUM_J:%.*]] = phi float [ [[SUM_I]], %[[FOR_I_HEADER]] ], [ [[SUM_J_NEXT:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[SUM_J_NEXT]] = fadd float [[SUM_J]], [[A]]
; CHECK-NEXT:    [[J_INC]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_J]], label %[[FOR_J]], label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[SUM_I_LCSSA]] = phi float [ [[SUM_J_NEXT]], %[[FOR_J]] ]
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[EXIT:.*]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %sum.i = phi float [ 0.0, %entry ], [ %sum.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %sum.j = phi float [ %sum.i, %for.i.header ], [ %sum.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %sum.j.next = fadd float %sum.j, %a
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %sum.i.lcssa = phi float [ %sum.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that the interchange is legal if the floating-point addition is marked
; as reassoc.
;
define void @reduction_reassoc_fadd(ptr %A) {
; CHECK-LABEL: define void @reduction_reassoc_fadd(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[SUM_J:%.*]] = phi float [ [[SUM_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[SUM_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[SUM_I]] = phi float [ [[SUM_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ 0.000000e+00, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[SUM_J_NEXT]] = fadd reassoc float [[SUM_J]], [[A]]
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[SUM_I_LCSSA]] = phi float [ [[SUM_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %sum.i = phi float [ 0.0, %entry ], [ %sum.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %sum.j = phi float [ %sum.i, %for.i.header ], [ %sum.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %sum.j.next = fadd reassoc float %sum.j, %a
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %sum.i.lcssa = phi float [ %sum.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; FIXME: Is it really legal to interchange the loops when
; both reassoc and ninf are set?
; Check that the interchange is legal if the floating-point addition is marked
; as reassoc.
;
define void @reduction_reassoc_ninf_fadd(ptr %A) {
; CHECK-LABEL: define void @reduction_reassoc_ninf_fadd(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[SUM_J:%.*]] = phi float [ [[SUM_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[SUM_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[SUM_I]] = phi float [ [[SUM_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ 0.000000e+00, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[SUM_J_NEXT]] = fadd reassoc float [[SUM_J]], [[A]]
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[SUM_I_LCSSA]] = phi float [ [[SUM_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %sum.i = phi float [ 0.0, %entry ], [ %sum.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %sum.j = phi float [ %sum.i, %for.i.header ], [ %sum.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %sum.j.next = fadd reassoc ninf float %sum.j, %a
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %sum.i.lcssa = phi float [ %sum.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that the loops aren't exchanged if there is a reduction of
; non-reassociative floating-point multiplication.
;
; float prod = 1;
; for (int i = 0; i < 2; i++)
;   for (int j = 0; j < 2; j++)
;     prod *= A[j][i];

define void @reduction_fmul(ptr %A) {
; CHECK-LABEL: define void @reduction_fmul(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ]
; CHECK-NEXT:    [[PROD_I:%.*]] = phi float [ 1.000000e+00, %[[ENTRY]] ], [ [[PROD_I_LCSSA:%.*]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ 0, %[[FOR_I_HEADER]] ], [ [[J_INC:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[PROD_J:%.*]] = phi float [ [[PROD_I]], %[[FOR_I_HEADER]] ], [ [[PROD_J_NEXT:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[PROD_J_NEXT]] = fmul float [[PROD_J]], [[A]]
; CHECK-NEXT:    [[J_INC]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_J]], label %[[FOR_J]], label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[PROD_I_LCSSA]] = phi float [ [[PROD_J_NEXT]], %[[FOR_J]] ]
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[EXIT:.*]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %prod.i = phi float [ 1.0, %entry ], [ %prod.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %prod.j = phi float [ %prod.i, %for.i.header ], [ %prod.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %prod.j.next = fmul float %prod.j, %a
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %prod.i.lcssa = phi float [ %prod.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that the interchange is legal if the floating-point multiplication is
; marked as reassoc.
;
define void @reduction_reassoc_fmul(ptr %A) {
; CHECK-LABEL: define void @reduction_reassoc_fmul(
; CHECK-SAME: ptr [[A:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[PROD_J:%.*]] = phi float [ [[PROD_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[PROD_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[PROD_I]] = phi float [ [[PROD_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ 1.000000e+00, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[PROD_J_NEXT]] = fmul reassoc float [[PROD_J]], [[A]]
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[PROD_I_LCSSA]] = phi float [ [[PROD_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %prod.i = phi float [ 1.0, %entry ], [ %prod.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %prod.j = phi float [ %prod.i, %for.i.header ], [ %prod.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %prod.j.next = fmul reassoc float %prod.j, %a
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %prod.i.lcssa = phi float [ %prod.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that the loops aren't exchanged if there is a reduction of
; non-reassociative floating-point fmuladd.
;
; float fmuladd = 0;
; for (int i = 0; i < 2; i++)
;   for (int j = 0; j < 2; j++)
;     fmuladd += A[j][i] * B[j][i];

define void @reduction_fmuladd(ptr %A, ptr %B) {
; CHECK-LABEL: define void @reduction_fmuladd(
; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ]
; CHECK-NEXT:    [[FMULADD_I:%.*]] = phi float [ 1.000000e+00, %[[ENTRY]] ], [ [[FMULADD_I_LCSSA:%.*]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ 0, %[[FOR_I_HEADER]] ], [ [[J_INC:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[FMULADD_J:%.*]] = phi float [ [[FMULADD_I]], %[[FOR_I_HEADER]] ], [ [[FMULADD_J_NEXT:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[IDX_A:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[IDX_B:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[B]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX_A]], align 4
; CHECK-NEXT:    [[B:%.*]] = load float, ptr [[IDX_B]], align 4
; CHECK-NEXT:    [[FMULADD_J_NEXT]] = call float @llvm.fmuladd.f32(float [[A]], float [[B]], float [[FMULADD_J]])
; CHECK-NEXT:    [[J_INC]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_J]], label %[[FOR_J]], label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[FMULADD_I_LCSSA]] = phi float [ [[FMULADD_J_NEXT]], %[[FOR_J]] ]
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[EXIT:.*]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %fmuladd.i = phi float [ 1.0, %entry ], [ %fmuladd.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %fmuladd.j = phi float [ %fmuladd.i, %for.i.header ], [ %fmuladd.j.next, %for.j ]
  %idx.a = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %idx.b = getelementptr inbounds [2 x [2 x i32]], ptr %B, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx.a, align 4
  %b = load float, ptr %idx.b, align 4
  %fmuladd.j.next = call float @llvm.fmuladd.f32(float %a, float %b, float %fmuladd.j)
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %fmuladd.i.lcssa = phi float [ %fmuladd.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that the interchange is legal if the floating-point fmuladd is marked
; as reassoc.
;
define void @reduction_reassoc_fmuladd(ptr %A, ptr %B) {
; CHECK-LABEL: define void @reduction_reassoc_fmuladd(
; CHECK-SAME: ptr [[A:%.*]], ptr [[B:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[FMULADD_J:%.*]] = phi float [ [[FMULADD_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[FMULADD_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[FMULADD_I]] = phi float [ [[FMULADD_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ 1.000000e+00, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX_A:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[IDX_B:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[B]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX_A]], align 4
; CHECK-NEXT:    [[B:%.*]] = load float, ptr [[IDX_B]], align 4
; CHECK-NEXT:    [[FMULADD_J_NEXT]] = call reassoc float @llvm.fmuladd.f32(float [[A]], float [[B]], float [[FMULADD_J]])
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[FMULADD_I_LCSSA]] = phi float [ [[FMULADD_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %fmuladd.i = phi float [ 1.0, %entry ], [ %fmuladd.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %fmuladd.j = phi float [ %fmuladd.i, %for.i.header ], [ %fmuladd.j.next, %for.j ]
  %idx.a = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %idx.b = getelementptr inbounds [2 x [2 x i32]], ptr %B, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx.a, align 4
  %b = load float, ptr %idx.b, align 4
  %fmuladd.j.next = call reassoc float @llvm.fmuladd.f32(float %a, float %b, float %fmuladd.j)
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %fmuladd.i.lcssa = phi float [ %fmuladd.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that interchanging the loops is legal for the reassociative
; floating-point minimum.
;
; float fmin = init;
; for (int i = 0; i < 2; i++)
;   for (int j = 0; j < 2; j++)
;     fmin = (A[j][i] < fmin) ? A[j][i] : fmin;

define void @reduction_fmin(ptr %A, float %init) {
; CHECK-LABEL: define void @reduction_fmin(
; CHECK-SAME: ptr [[A:%.*]], float [[INIT:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[FMIN_J:%.*]] = phi float [ [[FMIN_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[FMIN_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[FMIN_I]] = phi float [ [[FMIN_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ [[INIT]], %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[CMP:%.*]] = fcmp nnan nsz olt float [[A]], [[FMIN_J]]
; CHECK-NEXT:    [[FMIN_J_NEXT]] = select nnan nsz i1 [[CMP]], float [[A]], float [[FMIN_J]]
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[FMIN_I_LCSSA]] = phi float [ [[FMIN_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %fmin.i = phi float [ %init, %entry ], [ %fmin.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %fmin.j = phi float [ %fmin.i, %for.i.header ], [ %fmin.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %cmp = fcmp nnan nsz olt float %a, %fmin.j
  %fmin.j.next = select nnan nsz i1 %cmp, float %a, float %fmin.j
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %fmin.i.lcssa = phi float [ %fmin.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that interchanging the loops is legal for the floating-point
; llvm.minimumnum.
;
define void @reduction_fmininumnum(ptr %A, float %init) {
; CHECK-LABEL: define void @reduction_fmininumnum(
; CHECK-SAME: ptr [[A:%.*]], float [[INIT:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[FMIN_J:%.*]] = phi float [ [[FMIN_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[FMIN_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[FMIN_I]] = phi float [ [[FMIN_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ [[INIT]], %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[FMIN_J_NEXT]] = call float @llvm.minimumnum.f32(float [[A]], float [[FMIN_J]])
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[FMIN_I_LCSSA]] = phi float [ [[FMIN_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %fmin.i = phi float [ %init, %entry ], [ %fmin.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %fmin.j = phi float [ %fmin.i, %for.i.header ], [ %fmin.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %fmin.j.next = call float @llvm.minimumnum.f32(float %a, float %fmin.j)
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %fmin.i.lcssa = phi float [ %fmin.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that interchanging the loops is legal for the reassociative
; floating-point maximum.
;
; float fmax = init;
; for (int i = 0; i < 2; i++)
;   for (int j = 0; j < 2; j++)
;     fmax = (A[j][i] > fmax) ? A[j][i] : fmax;

define void @reduction_fmax(ptr %A, float %init) {
; CHECK-LABEL: define void @reduction_fmax(
; CHECK-SAME: ptr [[A:%.*]], float [[INIT:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[FMAX_J:%.*]] = phi float [ [[FMAX_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[FMAX_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[FMAX_I]] = phi float [ [[FMAX_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ [[INIT]], %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[CMP:%.*]] = fcmp nnan nsz ogt float [[A]], [[FMAX_J]]
; CHECK-NEXT:    [[FMAX_J_NEXT]] = select nnan nsz i1 [[CMP]], float [[A]], float [[FMAX_J]]
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[FMAX_I_LCSSA]] = phi float [ [[FMAX_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %fmax.i = phi float [ %init, %entry ], [ %fmax.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %fmax.j = phi float [ %fmax.i, %for.i.header ], [ %fmax.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %cmp = fcmp nnan nsz ogt float %a, %fmax.j
  %fmax.j.next = select nnan nsz i1 %cmp, float %a, float %fmax.j
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %fmax.i.lcssa = phi float [ %fmax.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

; Check that interchanging the loops is legal for the floating-point
; llvm.maximumnum.

define void @reduction_fmaxinumnum(ptr %A, float %init) {
; CHECK-LABEL: define void @reduction_fmaxinumnum(
; CHECK-SAME: ptr [[A:%.*]], float [[INIT:%.*]]) {
; CHECK-NEXT:  [[ENTRY:.*:]]
; CHECK-NEXT:    br label %[[FOR_J_PREHEADER:.*]]
; CHECK:       [[FOR_I_HEADER_PREHEADER:.*]]:
; CHECK-NEXT:    br label %[[FOR_I_HEADER:.*]]
; CHECK:       [[FOR_I_HEADER]]:
; CHECK-NEXT:    [[I:%.*]] = phi i32 [ [[I_INC:%.*]], %[[FOR_I_LATCH:.*]] ], [ 0, %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    [[FMAX_J:%.*]] = phi float [ [[FMAX_J_NEXT:%.*]], %[[FOR_I_LATCH]] ], [ [[FMAX_I:%.*]], %[[FOR_I_HEADER_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_J_SPLIT1:.*]]
; CHECK:       [[FOR_J_PREHEADER]]:
; CHECK-NEXT:    br label %[[FOR_J:.*]]
; CHECK:       [[FOR_J]]:
; CHECK-NEXT:    [[J:%.*]] = phi i32 [ [[TMP0:%.*]], %[[FOR_J_SPLIT:.*]] ], [ 0, %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    [[FMAX_I]] = phi float [ [[FMAX_I_LCSSA:%.*]], %[[FOR_J_SPLIT]] ], [ [[INIT]], %[[FOR_J_PREHEADER]] ]
; CHECK-NEXT:    br label %[[FOR_I_HEADER_PREHEADER]]
; CHECK:       [[FOR_J_SPLIT1]]:
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x i32]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[FMAX_J_NEXT]] = call float @llvm.maximumnum.f32(float [[A]], float [[FMAX_J]])
; CHECK-NEXT:    [[J_INC:%.*]] = add i32 [[J]], 1
; CHECK-NEXT:    [[CMP_J:%.*]] = icmp slt i32 [[J_INC]], 2
; CHECK-NEXT:    br label %[[FOR_I_LATCH]]
; CHECK:       [[FOR_J_SPLIT]]:
; CHECK-NEXT:    [[FMAX_I_LCSSA]] = phi float [ [[FMAX_J_NEXT]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[TMP0]] = add i32 [[J]], 1
; CHECK-NEXT:    [[TMP1:%.*]] = icmp slt i32 [[TMP0]], 2
; CHECK-NEXT:    br i1 [[TMP1]], label %[[FOR_J]], label %[[EXIT:.*]]
; CHECK:       [[FOR_I_LATCH]]:
; CHECK-NEXT:    [[I_INC]] = add i32 [[I]], 1
; CHECK-NEXT:    [[CMP_I:%.*]] = icmp slt i32 [[I_INC]], 2
; CHECK-NEXT:    br i1 [[CMP_I]], label %[[FOR_I_HEADER]], label %[[FOR_J_SPLIT]]
; CHECK:       [[EXIT]]:
; CHECK-NEXT:    ret void
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %fmax.i = phi float [ %init, %entry ], [ %fmax.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %fmax.j = phi float [ %fmax.i, %for.i.header ], [ %fmax.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x i32]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %fmax.j.next = call float @llvm.maximumnum.f32(float %a, float %fmax.j)
  %j.inc = add i32 %j, 1
  %cmp.j = icmp slt i32 %j.inc, 2
  br i1 %cmp.j, label %for.j, label %for.i.latch

for.i.latch:
  %fmax.i.lcssa = phi float [ %fmax.j.next, %for.j ]
  %i.inc = add i32 %i, 1
  %cmp.i = icmp slt i32 %i.inc, 2
  br i1 %cmp.i, label %for.i.header, label %exit

exit:
  ret void
}

declare float @llvm.fmuladd.f32(float %a, float %b, float %c)
declare float @llvm.minimumnum.f32(float %a, float %b)
declare float @llvm.maximumnum.f32(float %a, float %b)