test/Transforms/LoopInterchange/reduction-minmax-multi-use.ll

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

; FP maximum reduction with an extra use outside the reduction chain.
define float @fmaximum_with_extra_use_of_reduction_phi(ptr %A) {
; CHECK-LABEL: define float @fmaximum_with_extra_use_of_reduction_phi(
; 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:    [[MAX_I:%.*]] = phi float [ 0.000000e+00, %[[ENTRY]] ], [ [[MAX_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:    [[MAX_J:%.*]] = phi float [ [[MAX_I]], %[[FOR_I_HEADER]] ], [ [[MAX_J_NEXT:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[IDX:%.*]] = getelementptr inbounds [2 x [2 x float]], ptr [[A]], i32 0, i32 [[J]], i32 [[I]]
; CHECK-NEXT:    [[A:%.*]] = load float, ptr [[IDX]], align 4
; CHECK-NEXT:    [[EXTRA:%.*]] = fadd float [[MAX_J]], 1.000000e+00
; CHECK-NEXT:    [[MAX_J_NEXT]] = call float @llvm.maximum.f32(float [[MAX_J]], float [[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:    [[MAX_I_LCSSA]] = phi float [ [[MAX_I]], %[[FOR_J]] ]
; CHECK-NEXT:    [[EXTRA_LCSSA:%.*]] = phi float [ [[EXTRA]], %[[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:    [[RES:%.*]] = phi float [ [[EXTRA_LCSSA]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    ret float [[RES]]
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %max.i = phi float [ 0.000000e+00, %entry ], [ %max.i.lcssa, %for.i.latch ]
  br label %for.j

for.j:
  %j = phi i32 [ 0, %for.i.header ], [ %j.inc, %for.j ]
  %max.j = phi float [ %max.i, %for.i.header ], [ %max.j.next, %for.j ]
  %idx = getelementptr inbounds [2 x [2 x float]], ptr %A, i32 0, i32 %j, i32 %i
  %a = load float, ptr %idx, align 4
  %extra = fadd float %max.j, 1.000000e+00
  %max.j.next = call float @llvm.maximum.f32(float %max.j, float %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:
  %max.i.lcssa = phi float [ %max.i, %for.j ]
  %extra.lcssa = phi float [ %extra, %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:
  %res = phi float [ %extra.lcssa, %for.i.latch ]
  ret float %res
}

; sum accumulates the running maximum, so the result depends on the order in
; which the elements are visited.
define i32 @smax_with_extra_use_of_reduction_phi(ptr %A) {
; CHECK-LABEL: define i32 @smax_with_extra_use_of_reduction_phi(
; 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:    [[MAX_I:%.*]] = phi i32 [ 0, %[[ENTRY]] ], [ [[MAX_I_LCSSA:%.*]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    [[SUM_I:%.*]] = phi i32 [ 0, %[[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:    [[MAX_J:%.*]] = phi i32 [ [[MAX_I]], %[[FOR_I_HEADER]] ], [ [[MAX_J_NEXT:%.*]], %[[FOR_J]] ]
; CHECK-NEXT:    [[SUM_J:%.*]] = phi i32 [ [[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 i32, ptr [[IDX]], align 4
; CHECK-NEXT:    [[SUM_J_NEXT]] = add i32 [[SUM_J]], [[MAX_J]]
; CHECK-NEXT:    [[MAX_J_NEXT]] = tail call i32 @llvm.smax.i32(i32 [[MAX_J]], i32 [[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:    [[MAX_I_LCSSA]] = phi i32 [ [[MAX_J_NEXT]], %[[FOR_J]] ]
; CHECK-NEXT:    [[SUM_I_LCSSA]] = phi i32 [ [[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:    [[RES:%.*]] = phi i32 [ [[SUM_I_LCSSA]], %[[FOR_I_LATCH]] ]
; CHECK-NEXT:    ret i32 [[RES]]
;
entry:
  br label %for.i.header

for.i.header:
  %i = phi i32 [ 0, %entry ], [ %i.inc, %for.i.latch ]
  %max.i = phi i32 [ 0, %entry ], [ %max.i.lcssa, %for.i.latch ]
  %sum.i = phi i32 [ 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 ]
  %max.j = phi i32 [ %max.i, %for.i.header ], [ %max.j.next, %for.j ]
  %sum.j = phi i32 [ %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 i32, ptr %idx, align 4
  %sum.j.next = add i32 %sum.j, %max.j
  %max.j.next = tail call i32 @llvm.smax.i32(i32 %max.j, i32 %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:
  %max.i.lcssa = phi i32 [ %max.j.next, %for.j ]
  %sum.i.lcssa = phi i32 [ %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:
  %res = phi i32 [ %sum.i.lcssa, %for.i.latch ]
  ret i32 %res
}