; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -passes=separate-const-offset-from-gep,gvn \
; RUN:     -reassociate-geps-verify-no-dead-code < %s \
; RUN:     | FileCheck --check-prefix=IR %s

; Verifies the SeparateConstOffsetFromGEP pass.
; The following code computes
; *output = array[x][y] + array[x][y+1] + array[x+1][y] + array[x+1][y+1]
;
; We expect SeparateConstOffsetFromGEP to transform it to
;
; ptr base = &a[x][y];
; *output = base[0] + base[1] + base[32] + base[33];
;
; so the backend can emit PTX that uses fewer virtual registers.

@array = internal addrspace(3) global [32 x [32 x float]] zeroinitializer, align 4

define void @sum_of_array(i32 %x, i32 %y, ptr nocapture %output) {
; IR-LABEL: define void @sum_of_array
; IR-SAME: (i32 [[X:%.*]], i32 [[Y:%.*]], ptr nocapture [[OUTPUT:%.*]]) #[[ATTR0:[0-9]+]] {
; IR-NEXT:  .preheader:
; IR-NEXT:    [[I:%.*]] = sext i32 [[Y]] to i64
; IR-NEXT:    [[I1:%.*]] = sext i32 [[X]] to i64
; IR-NEXT:    [[I2:%.*]] = getelementptr [32 x [32 x float]], ptr addrspace(3) @array, i32 0, i32 [[X]], i32 [[Y]]
; IR-NEXT:    [[I3:%.*]] = addrspacecast ptr addrspace(3) [[I2]] to ptr
; IR-NEXT:    [[I4:%.*]] = load float, ptr [[I3]], align 4
; IR-NEXT:    [[I5:%.*]] = fadd float [[I4]], 0.000000e+00
; IR-NEXT:    [[I87:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 4
; IR-NEXT:    [[I9:%.*]] = addrspacecast ptr addrspace(3) [[I87]] to ptr
; IR-NEXT:    [[I10:%.*]] = load float, ptr [[I9]], align 4
; IR-NEXT:    [[I11:%.*]] = fadd float [[I5]], [[I10]]
; IR-NEXT:    [[I1412:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 128
; IR-NEXT:    [[I15:%.*]] = addrspacecast ptr addrspace(3) [[I1412]] to ptr
; IR-NEXT:    [[I16:%.*]] = load float, ptr [[I15]], align 4
; IR-NEXT:    [[I17:%.*]] = fadd float [[I11]], [[I16]]
; IR-NEXT:    [[I1818:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 132
; IR-NEXT:    [[I19:%.*]] = addrspacecast ptr addrspace(3) [[I1818]] to ptr
; IR-NEXT:    [[I20:%.*]] = load float, ptr [[I19]], align 4
; IR-NEXT:    [[I21:%.*]] = fadd float [[I17]], [[I20]]
; IR-NEXT:    store float [[I21]], ptr [[OUTPUT]], align 4
; IR-NEXT:    ret void
;
.preheader:
  %i = sext i32 %y to i64
  %i1 = sext i32 %x to i64
  %i2 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i
  %i3 = addrspacecast ptr addrspace(3) %i2 to ptr
  %i4 = load float, ptr %i3, align 4
  %i5 = fadd float %i4, 0.000000e+00
  %i6 = add i32 %y, 1
  %i7 = sext i32 %i6 to i64
  %i8 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i7
  %i9 = addrspacecast ptr addrspace(3) %i8 to ptr
  %i10 = load float, ptr %i9, align 4
  %i11 = fadd float %i5, %i10
  %i12 = add i32 %x, 1
  %i13 = sext i32 %i12 to i64
  %i14 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i13, i64 %i
  %i15 = addrspacecast ptr addrspace(3) %i14 to ptr
  %i16 = load float, ptr %i15, align 4
  %i17 = fadd float %i11, %i16
  %i18 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i13, i64 %i7
  %i19 = addrspacecast ptr addrspace(3) %i18 to ptr
  %i20 = load float, ptr %i19, align 4
  %i21 = fadd float %i17, %i20
  store float %i21, ptr %output, align 4
  ret void
}


; TODO: GVN is unable to preserve the "inbounds" keyword on the first GEP. Need
; some infrastructure changes to enable such optimizations.

; @sum_of_array2 is very similar to @sum_of_array. The only difference is in
; the order of "sext" and "add" when computing the array indices. @sum_of_array
; computes add before sext, e.g., array[sext(x + 1)][sext(y + 1)], while
; @sum_of_array2 computes sext before add,
; e.g., array[sext(x) + 1][sext(y) + 1]. SeparateConstOffsetFromGEP should be
; able to extract constant offsets from both forms.
define void @sum_of_array2(i32 %x, i32 %y, ptr nocapture %output) {
; IR-LABEL: define void @sum_of_array2
; IR-SAME: (i32 [[X:%.*]], i32 [[Y:%.*]], ptr nocapture [[OUTPUT:%.*]]) #[[ATTR0]] {
; IR-NEXT:  .preheader:
; IR-NEXT:    [[I:%.*]] = sext i32 [[Y]] to i64
; IR-NEXT:    [[I1:%.*]] = sext i32 [[X]] to i64
; IR-NEXT:    [[I2:%.*]] = getelementptr [32 x [32 x float]], ptr addrspace(3) @array, i32 0, i32 [[X]], i32 [[Y]]
; IR-NEXT:    [[I3:%.*]] = addrspacecast ptr addrspace(3) [[I2]] to ptr
; IR-NEXT:    [[I4:%.*]] = load float, ptr [[I3]], align 4
; IR-NEXT:    [[I5:%.*]] = fadd float [[I4]], 0.000000e+00
; IR-NEXT:    [[I77:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 4
; IR-NEXT:    [[I8:%.*]] = addrspacecast ptr addrspace(3) [[I77]] to ptr
; IR-NEXT:    [[I9:%.*]] = load float, ptr [[I8]], align 4
; IR-NEXT:    [[I10:%.*]] = fadd float [[I5]], [[I9]]
; IR-NEXT:    [[I1212:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 128
; IR-NEXT:    [[I13:%.*]] = addrspacecast ptr addrspace(3) [[I1212]] to ptr
; IR-NEXT:    [[I14:%.*]] = load float, ptr [[I13]], align 4
; IR-NEXT:    [[I15:%.*]] = fadd float [[I10]], [[I14]]
; IR-NEXT:    [[I1618:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 132
; IR-NEXT:    [[I17:%.*]] = addrspacecast ptr addrspace(3) [[I1618]] to ptr
; IR-NEXT:    [[I18:%.*]] = load float, ptr [[I17]], align 4
; IR-NEXT:    [[I19:%.*]] = fadd float [[I15]], [[I18]]
; IR-NEXT:    store float [[I19]], ptr [[OUTPUT]], align 4
; IR-NEXT:    ret void
;
.preheader:
  %i = sext i32 %y to i64
  %i1 = sext i32 %x to i64
  %i2 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i
  %i3 = addrspacecast ptr addrspace(3) %i2 to ptr
  %i4 = load float, ptr %i3, align 4
  %i5 = fadd float %i4, 0.000000e+00
  %i6 = add i64 %i, 1
  %i7 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i6
  %i8 = addrspacecast ptr addrspace(3) %i7 to ptr
  %i9 = load float, ptr %i8, align 4
  %i10 = fadd float %i5, %i9
  %i11 = add i64 %i1, 1
  %i12 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i11, i64 %i
  %i13 = addrspacecast ptr addrspace(3) %i12 to ptr
  %i14 = load float, ptr %i13, align 4
  %i15 = fadd float %i10, %i14
  %i16 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i11, i64 %i6
  %i17 = addrspacecast ptr addrspace(3) %i16 to ptr
  %i18 = load float, ptr %i17, align 4
  %i19 = fadd float %i15, %i18
  store float %i19, ptr %output, align 4
  ret void
}

; This function loads
;   array[zext(x)][zext(y)]
;   array[zext(x)][zext(y +nuw 1)]
;   array[zext(x +nuw 1)][zext(y)]
;   array[zext(x +nuw 1)][zext(y +nuw 1)].
;
; This function is similar to @sum_of_array, but it
; 1) extends array indices using zext instead of sext;
; 2) annotates the addition with "nuw"; otherwise, zext(x + 1) => zext(x) + 1
;    may be invalid.

define void @sum_of_array3(i32 %x, i32 %y, ptr nocapture %output) {
; IR-LABEL: define void @sum_of_array3
; IR-SAME: (i32 [[X:%.*]], i32 [[Y:%.*]], ptr nocapture [[OUTPUT:%.*]]) #[[ATTR0]] {
; IR-NEXT:  .preheader:
; IR-NEXT:    [[I:%.*]] = zext i32 [[Y]] to i64
; IR-NEXT:    [[I1:%.*]] = zext i32 [[X]] to i64
; IR-NEXT:    [[I2:%.*]] = getelementptr [32 x [32 x float]], ptr addrspace(3) @array, i32 0, i32 [[X]], i32 [[Y]]
; IR-NEXT:    [[I3:%.*]] = addrspacecast ptr addrspace(3) [[I2]] to ptr
; IR-NEXT:    [[I4:%.*]] = load float, ptr [[I3]], align 4
; IR-NEXT:    [[I5:%.*]] = fadd float [[I4]], 0.000000e+00
; IR-NEXT:    [[I87:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 4
; IR-NEXT:    [[I9:%.*]] = addrspacecast ptr addrspace(3) [[I87]] to ptr
; IR-NEXT:    [[I10:%.*]] = load float, ptr [[I9]], align 4
; IR-NEXT:    [[I11:%.*]] = fadd float [[I5]], [[I10]]
; IR-NEXT:    [[I1412:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 128
; IR-NEXT:    [[I15:%.*]] = addrspacecast ptr addrspace(3) [[I1412]] to ptr
; IR-NEXT:    [[I16:%.*]] = load float, ptr [[I15]], align 4
; IR-NEXT:    [[I17:%.*]] = fadd float [[I11]], [[I16]]
; IR-NEXT:    [[I1818:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 132
; IR-NEXT:    [[I19:%.*]] = addrspacecast ptr addrspace(3) [[I1818]] to ptr
; IR-NEXT:    [[I20:%.*]] = load float, ptr [[I19]], align 4
; IR-NEXT:    [[I21:%.*]] = fadd float [[I17]], [[I20]]
; IR-NEXT:    store float [[I21]], ptr [[OUTPUT]], align 4
; IR-NEXT:    ret void
;
.preheader:
  %i = zext i32 %y to i64
  %i1 = zext i32 %x to i64
  %i2 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i
  %i3 = addrspacecast ptr addrspace(3) %i2 to ptr
  %i4 = load float, ptr %i3, align 4
  %i5 = fadd float %i4, 0.000000e+00
  %i6 = add nuw i32 %y, 1
  %i7 = zext i32 %i6 to i64
  %i8 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i7
  %i9 = addrspacecast ptr addrspace(3) %i8 to ptr
  %i10 = load float, ptr %i9, align 4
  %i11 = fadd float %i5, %i10
  %i12 = add nuw i32 %x, 1
  %i13 = zext i32 %i12 to i64
  %i14 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i13, i64 %i
  %i15 = addrspacecast ptr addrspace(3) %i14 to ptr
  %i16 = load float, ptr %i15, align 4
  %i17 = fadd float %i11, %i16
  %i18 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i13, i64 %i7
  %i19 = addrspacecast ptr addrspace(3) %i18 to ptr
  %i20 = load float, ptr %i19, align 4
  %i21 = fadd float %i17, %i20
  store float %i21, ptr %output, align 4
  ret void
}

; This function loads
;   array[zext(x)][zext(y)]
;   array[zext(x)][zext(y)]
;   array[zext(x) + 1][zext(y) + 1]
;   array[zext(x) + 1][zext(y) + 1].
;
; We expect the generated code to reuse the computation of
; &array[zext(x)][zext(y)]. See the expected IR and PTX for details.
define void @sum_of_array4(i32 %x, i32 %y, ptr nocapture %output) {
; IR-LABEL: define void @sum_of_array4
; IR-SAME: (i32 [[X:%.*]], i32 [[Y:%.*]], ptr nocapture [[OUTPUT:%.*]]) #[[ATTR0]] {
; IR-NEXT:  .preheader:
; IR-NEXT:    [[I:%.*]] = zext i32 [[Y]] to i64
; IR-NEXT:    [[I1:%.*]] = zext i32 [[X]] to i64
; IR-NEXT:    [[I2:%.*]] = getelementptr [32 x [32 x float]], ptr addrspace(3) @array, i32 0, i32 [[X]], i32 [[Y]]
; IR-NEXT:    [[I3:%.*]] = addrspacecast ptr addrspace(3) [[I2]] to ptr
; IR-NEXT:    [[I4:%.*]] = load float, ptr [[I3]], align 4
; IR-NEXT:    [[I5:%.*]] = fadd float [[I4]], 0.000000e+00
; IR-NEXT:    [[I77:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 4
; IR-NEXT:    [[I8:%.*]] = addrspacecast ptr addrspace(3) [[I77]] to ptr
; IR-NEXT:    [[I9:%.*]] = load float, ptr [[I8]], align 4
; IR-NEXT:    [[I10:%.*]] = fadd float [[I5]], [[I9]]
; IR-NEXT:    [[I1212:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 128
; IR-NEXT:    [[I13:%.*]] = addrspacecast ptr addrspace(3) [[I1212]] to ptr
; IR-NEXT:    [[I14:%.*]] = load float, ptr [[I13]], align 4
; IR-NEXT:    [[I15:%.*]] = fadd float [[I10]], [[I14]]
; IR-NEXT:    [[I1618:%.*]] = getelementptr inbounds i8, ptr addrspace(3) [[I2]], i32 132
; IR-NEXT:    [[I17:%.*]] = addrspacecast ptr addrspace(3) [[I1618]] to ptr
; IR-NEXT:    [[I18:%.*]] = load float, ptr [[I17]], align 4
; IR-NEXT:    [[I19:%.*]] = fadd float [[I15]], [[I18]]
; IR-NEXT:    store float [[I19]], ptr [[OUTPUT]], align 4
; IR-NEXT:    ret void
;
.preheader:
  %i = zext i32 %y to i64
  %i1 = zext i32 %x to i64
  %i2 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i
  %i3 = addrspacecast ptr addrspace(3) %i2 to ptr
  %i4 = load float, ptr %i3, align 4
  %i5 = fadd float %i4, 0.000000e+00
  %i6 = add i64 %i, 1
  %i7 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i1, i64 %i6
  %i8 = addrspacecast ptr addrspace(3) %i7 to ptr
  %i9 = load float, ptr %i8, align 4
  %i10 = fadd float %i5, %i9
  %i11 = add i64 %i1, 1
  %i12 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i11, i64 %i
  %i13 = addrspacecast ptr addrspace(3) %i12 to ptr
  %i14 = load float, ptr %i13, align 4
  %i15 = fadd float %i10, %i14
  %i16 = getelementptr inbounds [32 x [32 x float]], ptr addrspace(3) @array, i64 0, i64 %i11, i64 %i6
  %i17 = addrspacecast ptr addrspace(3) %i16 to ptr
  %i18 = load float, ptr %i17, align 4
  %i19 = fadd float %i15, %i18
  store float %i19, ptr %output, align 4
  ret void
}

; The source code is:
;   p0 = &input[sext(x + y)];
;   p1 = &input[sext(x + (y + 5))];
;
; Without reuniting extensions, SeparateConstOffsetFromGEP would emit
;   p0 = &input[sext(x + y)];
;   t1 = &input[sext(x) + sext(y)];
;   p1 = &t1[5];
;
; With reuniting extensions, it merges p0 and t1 and thus emits
;   p0 = &input[sext(x + y)];
;   p1 = &p0[5];
define void @reunion(i32 %x, i32 %y, ptr %input) {
; IR-LABEL: define void @reunion
; IR-SAME: (i32 [[X:%.*]], i32 [[Y:%.*]], ptr [[INPUT:%.*]]) #[[ATTR0]] {
; IR-NEXT:  entry:
; IR-NEXT:    [[XY:%.*]] = add nsw i32 [[X]], [[Y]]
; IR-NEXT:    [[I:%.*]] = sext i32 [[XY]] to i64
; IR-NEXT:    [[P0:%.*]] = getelementptr float, ptr [[INPUT]], i64 [[I]]
; IR-NEXT:    [[V0:%.*]] = load float, ptr [[P0]], align 4
; IR-NEXT:    call void @use(float [[V0]])
; IR-NEXT:    [[P13:%.*]] = getelementptr inbounds i8, ptr [[P0]], i64 20
; IR-NEXT:    [[V1:%.*]] = load float, ptr [[P13]], align 4
; IR-NEXT:    call void @use(float [[V1]])
; IR-NEXT:    ret void
;
entry:
  %xy = add nsw i32 %x, %y
  %i = sext i32 %xy to i64
  %p0 = getelementptr inbounds float, ptr %input, i64 %i
  %v0 = load float, ptr %p0, align 4
  call void @use(float %v0)
  %y5 = add nsw i32 %y, 5
  %xy5 = add nsw i32 %x, %y5
  %i1 = sext i32 %xy5 to i64
  %p1 = getelementptr inbounds float, ptr %input, i64 %i1
  %v1 = load float, ptr %p1, align 4
  call void @use(float %v1)
  ret void
}

declare void @use(float)