llvm/test/CodeGen/AMDGPU/amdgpu-simplify-libcall-sincos.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
; RUN: opt -S -O1 -mtriple=amdgcn-- -amdgpu-simplify-libcall=1 -amdgpu-prelink < %s | FileCheck %s

declare float @_Z3sinf(float) #0
declare float @_Z3cosf(float) #0
declare <2 x float> @_Z3sinDv2_f(<2 x float>) #0
declare <2 x float> @_Z3cosDv2_f(<2 x float>) #0
declare <3 x float> @_Z3sinDv3_f(<3 x float>) #0
declare <3 x float> @_Z3cosDv3_f(<3 x float>) #0
declare <4 x float> @_Z3sinDv4_f(<4 x float>) #0
declare <4 x float> @_Z3cosDv4_f(<4 x float>) #0
declare <8 x float> @_Z3sinDv8_f(<8 x float>) #0
declare <8 x float> @_Z3cosDv8_f(<8 x float>) #0
declare <16 x float> @_Z3sinDv16_f(<16 x float>) #0
declare <16 x float> @_Z3cosDv16_f(<16 x float>) #0


declare half @_Z3sinDh(half) #0
declare half @_Z3cosDh(half) #0
declare <2 x half> @_Z3sinDv2_Dh(<2 x half>) #0
declare <2 x half> @_Z3cosDv2_Dh(<2 x half>) #0
declare <3 x half> @_Z3sinDv3_Dh(<3 x half>) #0
declare <3 x half> @_Z3cosDv3_Dh(<3 x half>) #0
declare <4 x half> @_Z3sinDv4_Dh(<4 x half>) #0
declare <4 x half> @_Z3cosDv4_Dh(<4 x half>) #0
declare <8 x half> @_Z3sinDv8_Dh(<8 x half>) #0
declare <8 x half> @_Z3cosDv8_Dh(<8 x half>) #0
declare <16 x half> @_Z3sinDv16_Dh(<16 x half>) #0
declare <16 x half> @_Z3cosDv16_Dh(<16 x half>) #0


declare double @_Z3sind(double) #0
declare double @_Z3cosd(double) #0
declare <2 x double> @_Z3sinDv2_d(<2 x double>) #0
declare <2 x double> @_Z3cosDv2_d(<2 x double>) #0
declare <3 x double> @_Z3sinDv3_d(<3 x double>) #0
declare <3 x double> @_Z3cosDv3_d(<3 x double>) #0
declare <4 x double> @_Z3sinDv4_d(<4 x double>) #0
declare <4 x double> @_Z3cosDv4_d(<4 x double>) #0
declare <8 x double> @_Z3sinDv8_d(<8 x double>) #0
declare <8 x double> @_Z3cosDv8_d(<8 x double>) #0
declare <16 x double> @_Z3sinDv16_d(<16 x double>) #0
declare <16 x double> @_Z3cosDv16_d(<16 x double>) #0

declare float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) writeonly %ptr) #1
declare float @_Z6sincosfPU3AS0f(float %x, ptr writeonly %ptr) #1

define void @sincos_f16_nocontract(half %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f16_nocontract
; CHECK-SAME: (half [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 2)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 2)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2:[0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call half @_Z3sinDh(half [[X]])
; CHECK-NEXT:    store half [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 2
; CHECK-NEXT:    [[CALL1:%.*]] = tail call half @_Z3cosDh(half [[X]])
; CHECK-NEXT:    store half [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call half @_Z3sinDh(half %x)
  store half %call, ptr addrspace(1) %sin_out, align 2
  %call1 = tail call half @_Z3cosDh(half %x)
  store half %call1, ptr addrspace(1) %cos_out, align 2
  ret void
}


define void @sincos_v2f16_nocontract(<2 x half> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f16_nocontract
; CHECK-SAME: (<2 x half> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call <2 x half> @_Z3sinDv2_Dh(<2 x half> [[X]])
; CHECK-NEXT:    store <2 x half> [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call <2 x half> @_Z3cosDv2_Dh(<2 x half> [[X]])
; CHECK-NEXT:    store <2 x half> [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call <2 x half> @_Z3sinDv2_Dh(<2 x half> %x)
  store <2 x half> %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call <2 x half> @_Z3cosDv2_Dh(<2 x half> %x)
  store <2 x half> %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}


define void @sincos_f16(half %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f16
; CHECK-SAME: (half [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 2)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 2)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract half @_Z3sinDh(half [[X]])
; CHECK-NEXT:    store half [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 2
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract half @_Z3cosDh(half [[X]])
; CHECK-NEXT:    store half [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract half @_Z3sinDh(half %x)
  store half %call, ptr addrspace(1) %sin_out, align 2
  %call1 = tail call contract half @_Z3cosDh(half %x)
  store half %call1, ptr addrspace(1) %cos_out, align 2
  ret void
}

define void @sincos_f16_order1(half %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f16_order1
; CHECK-SAME: (half [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 2)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 2)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract half @_Z3cosDh(half [[X]])
; CHECK-NEXT:    store half [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 2
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract half @_Z3sinDh(half [[X]])
; CHECK-NEXT:    store half [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 2
; CHECK-NEXT:    ret void
;
entry:
  %call1 = tail call contract half @_Z3cosDh(half %x)
  store half %call1, ptr addrspace(1) %cos_out, align 2
  %call = tail call contract half @_Z3sinDh(half %x)
  store half %call, ptr addrspace(1) %sin_out, align 2
  ret void
}

define void @sincos_v2f16(<2 x half> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f16
; CHECK-SAME: (<2 x half> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract <2 x half> @_Z3sinDv2_Dh(<2 x half> [[X]])
; CHECK-NEXT:    store <2 x half> [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract <2 x half> @_Z3cosDv2_Dh(<2 x half> [[X]])
; CHECK-NEXT:    store <2 x half> [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <2 x half> @_Z3sinDv2_Dh(<2 x half> %x)
  store <2 x half> %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract <2 x half> @_Z3cosDv2_Dh(<2 x half> %x)
  store <2 x half> %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_v3f16(<3 x half> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v3f16
; CHECK-SAME: (<3 x half> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract <3 x half> @_Z3sinDv3_Dh(<3 x half> [[X]])
; CHECK-NEXT:    [[EXTRACTVEC2:%.*]] = shufflevector <3 x half> [[CALL]], <3 x half> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
; CHECK-NEXT:    store <4 x half> [[EXTRACTVEC2]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    [[CALL5:%.*]] = tail call contract <3 x half> @_Z3cosDv3_Dh(<3 x half> [[X]])
; CHECK-NEXT:    [[EXTRACTVEC6:%.*]] = shufflevector <3 x half> [[CALL5]], <3 x half> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
; CHECK-NEXT:    store <4 x half> [[EXTRACTVEC6]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <3 x half> @_Z3sinDv3_Dh(<3 x half> %x)
  %extractVec2 = shufflevector <3 x half> %call, <3 x half> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
  store <4 x half> %extractVec2, ptr addrspace(1) %sin_out, align 8
  %call5 = tail call contract <3 x half> @_Z3cosDv3_Dh(<3 x half> %x)
  %extractVec6 = shufflevector <3 x half> %call5, <3 x half> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
  store <4 x half> %extractVec6, ptr addrspace(1) %cos_out, align 8
  ret void
}


define void @sincos_v4f16(<4 x half> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v4f16
; CHECK-SAME: (<4 x half> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract <4 x half> @_Z3sinDv4_Dh(<4 x half> [[X]])
; CHECK-NEXT:    store <4 x half> [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract <4 x half> @_Z3cosDv4_Dh(<4 x half> [[X]])
; CHECK-NEXT:    store <4 x half> [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <4 x half> @_Z3sinDv4_Dh(<4 x half> %x)
  store <4 x half> %call, ptr addrspace(1) %sin_out, align 8
  %call1 = tail call contract <4 x half> @_Z3cosDv4_Dh(<4 x half> %x)
  store <4 x half> %call1, ptr addrspace(1) %cos_out, align 8
  ret void
}

define void @sincos_v8f16(<8 x half> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v8f16
; CHECK-SAME: (<8 x half> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract <8 x half> @_Z3sinDv8_Dh(<8 x half> [[X]])
; CHECK-NEXT:    store <8 x half> [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 16
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract <8 x half> @_Z3cosDv8_Dh(<8 x half> [[X]])
; CHECK-NEXT:    store <8 x half> [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 16
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <8 x half> @_Z3sinDv8_Dh(<8 x half> %x)
  store <8 x half> %call, ptr addrspace(1) %sin_out, align 16
  %call1 = tail call contract <8 x half> @_Z3cosDv8_Dh(<8 x half> %x)
  store <8 x half> %call1, ptr addrspace(1) %cos_out, align 16
  ret void
}


define void @sincos_v16f16(<16 x half> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v16f16
; CHECK-SAME: (<16 x half> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract <16 x half> @_Z3sinDv16_Dh(<16 x half> [[X]])
; CHECK-NEXT:    store <16 x half> [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 32
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract <16 x half> @_Z3cosDv16_Dh(<16 x half> [[X]])
; CHECK-NEXT:    store <16 x half> [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 32
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <16 x half> @_Z3sinDv16_Dh(<16 x half> %x)
  store <16 x half> %call, ptr addrspace(1) %sin_out, align 32
  %call1 = tail call contract <16 x half> @_Z3cosDv16_Dh(<16 x half> %x)
  store <16 x half> %call1, ptr addrspace(1) %cos_out, align 32
  ret void
}


define void @sincos_f32_nocontract(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_nocontract
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3:[0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}


define void @sincos_v2f32_nocontract(<2 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f32_nocontract
; CHECK-SAME: (<2 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4:[0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <2 x float>, align 8, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call <2 x float> @_Z6sincosDv2_fPU3AS5S_(<2 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x float>, ptr addrspace(5) [[__SINCOS_]], align 8
; CHECK-NEXT:    store <2 x float> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    store <2 x float> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call <2 x float> @_Z3sinDv2_f(<2 x float> %x)
  store <2 x float> %call, ptr addrspace(1) %sin_out, align 8
  %call1 = tail call <2 x float> @_Z3cosDv2_f(<2 x float> %x)
  store <2 x float> %call1, ptr addrspace(1) %cos_out, align 8
  ret void
}


define void @sincos_f32(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_order1(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_order1
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call1 = tail call contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  ret void
}

define void @sincos_v2f32(<2 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f32
; CHECK-SAME: (<2 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <2 x float>, align 8, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <2 x float> @_Z6sincosDv2_fPU3AS5S_(<2 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x float>, ptr addrspace(5) [[__SINCOS_]], align 8
; CHECK-NEXT:    store <2 x float> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    store <2 x float> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <2 x float> @_Z3sinDv2_f(<2 x float> %x)
  store <2 x float> %call, ptr addrspace(1) %sin_out, align 8
  %call1 = tail call contract <2 x float> @_Z3cosDv2_f(<2 x float> %x)
  store <2 x float> %call1, ptr addrspace(1) %cos_out, align 8
  ret void
}

define void @sincos_v3f32(<3 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v3f32
; CHECK-SAME: (<3 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <3 x float>, align 16, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <3 x float> @_Z6sincosDv3_fPU3AS5S_(<3 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <3 x float>, ptr addrspace(5) [[__SINCOS_]], align 16
; CHECK-NEXT:    [[EXTRACTVEC2:%.*]] = shufflevector <3 x float> [[TMP0]], <3 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
; CHECK-NEXT:    store <4 x float> [[EXTRACTVEC2]], ptr addrspace(1) [[SIN_OUT]], align 16
; CHECK-NEXT:    [[EXTRACTVEC6:%.*]] = shufflevector <3 x float> [[TMP1]], <3 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
; CHECK-NEXT:    store <4 x float> [[EXTRACTVEC6]], ptr addrspace(1) [[COS_OUT]], align 16
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <3 x float> @_Z3sinDv3_f(<3 x float> %x)
  %extractVec2 = shufflevector <3 x float> %call, <3 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
  store <4 x float> %extractVec2, ptr addrspace(1) %sin_out, align 16
  %call5 = tail call contract <3 x float> @_Z3cosDv3_f(<3 x float> %x)
  %extractVec6 = shufflevector <3 x float> %call5, <3 x float> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
  store <4 x float> %extractVec6, ptr addrspace(1) %cos_out, align 16
  ret void
}

define void @sincos_v4f32(<4 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v4f32
; CHECK-SAME: (<4 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <4 x float>, align 16, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <4 x float> @_Z6sincosDv4_fPU3AS5S_(<4 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <4 x float>, ptr addrspace(5) [[__SINCOS_]], align 16
; CHECK-NEXT:    store <4 x float> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 16
; CHECK-NEXT:    store <4 x float> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 16
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <4 x float> @_Z3sinDv4_f(<4 x float> %x)
  store <4 x float> %call, ptr addrspace(1) %sin_out, align 16
  %call1 = tail call contract <4 x float> @_Z3cosDv4_f(<4 x float> %x)
  store <4 x float> %call1, ptr addrspace(1) %cos_out, align 16
  ret void
}

define void @sincos_v8f32(<8 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v8f32
; CHECK-SAME: (<8 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <8 x float>, align 32, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <8 x float> @_Z6sincosDv8_fPU3AS5S_(<8 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <8 x float>, ptr addrspace(5) [[__SINCOS_]], align 32
; CHECK-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 32
; CHECK-NEXT:    store <8 x float> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 32
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <8 x float> @_Z3sinDv8_f(<8 x float> %x)
  store <8 x float> %call, ptr addrspace(1) %sin_out, align 32
  %call1 = tail call contract <8 x float> @_Z3cosDv8_f(<8 x float> %x)
  store <8 x float> %call1, ptr addrspace(1) %cos_out, align 32
  ret void
}

define void @sincos_v16f32(<16 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v16f32
; CHECK-SAME: (<16 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 64)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 64)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <16 x float>, align 64, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <16 x float> @_Z6sincosDv16_fPU3AS5S_(<16 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <16 x float>, ptr addrspace(5) [[__SINCOS_]], align 64
; CHECK-NEXT:    store <16 x float> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 64
; CHECK-NEXT:    store <16 x float> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 64
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <16 x float> @_Z3sinDv16_f(<16 x float> %x)
  store <16 x float> %call, ptr addrspace(1) %sin_out, align 64
  %call1 = tail call contract <16 x float> @_Z3cosDv16_f(<16 x float> %x)
  store <16 x float> %call1, ptr addrspace(1) %cos_out, align 64
  ret void
}

define void @sincos_f64_nocontract(double %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f64_nocontract
; CHECK-SAME: (double [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca double, align 8, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call double @_Z6sincosdPU3AS5d(double [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(5) [[__SINCOS_]], align 8
; CHECK-NEXT:    store double [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    store double [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call double @_Z3sind(double %x)
  store double %call, ptr addrspace(1) %sin_out, align 8
  %call1 = tail call double @_Z3cosd(double %x)
  store double %call1, ptr addrspace(1) %cos_out, align 8
  ret void
}


define void @sincos_v2f64_nocontract(<2 x double> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f64_nocontract
; CHECK-SAME: (<2 x double> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <2 x double>, align 16, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call <2 x double> @_Z6sincosDv2_dPU3AS5S_(<2 x double> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x double>, ptr addrspace(5) [[__SINCOS_]], align 16
; CHECK-NEXT:    store <2 x double> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 16
; CHECK-NEXT:    store <2 x double> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 16
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call <2 x double> @_Z3sinDv2_d(<2 x double> %x)
  store <2 x double> %call, ptr addrspace(1) %sin_out, align 16
  %call1 = tail call <2 x double> @_Z3cosDv2_d(<2 x double> %x)
  store <2 x double> %call1, ptr addrspace(1) %cos_out, align 16
  ret void
}

define void @sincos_f64(double %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f64
; CHECK-SAME: (double [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca double, align 8, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract double @_Z6sincosdPU3AS5d(double [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(5) [[__SINCOS_]], align 8
; CHECK-NEXT:    store double [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    store double [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract double @_Z3sind(double %x)
  store double %call, ptr addrspace(1) %sin_out, align 8
  %call1 = tail call contract double @_Z3cosd(double %x)
  store double %call1, ptr addrspace(1) %cos_out, align 8
  ret void
}

define void @sincos_f64_order1(double %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f64_order1
; CHECK-SAME: (double [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca double, align 8, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract double @_Z6sincosdPU3AS5d(double [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(5) [[__SINCOS_]], align 8
; CHECK-NEXT:    store double [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    store double [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call1 = tail call contract double @_Z3cosd(double %x)
  store double %call1, ptr addrspace(1) %cos_out, align 8
  %call = tail call contract double @_Z3sind(double %x)
  store double %call, ptr addrspace(1) %sin_out, align 8
  ret void
}

define void @sincos_v2f64(<2 x double> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f64
; CHECK-SAME: (<2 x double> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 16)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <2 x double>, align 16, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <2 x double> @_Z6sincosDv2_dPU3AS5S_(<2 x double> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x double>, ptr addrspace(5) [[__SINCOS_]], align 16
; CHECK-NEXT:    store <2 x double> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 16
; CHECK-NEXT:    store <2 x double> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 16
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <2 x double> @_Z3sinDv2_d(<2 x double> %x)
  store <2 x double> %call, ptr addrspace(1) %sin_out, align 16
  %call1 = tail call contract <2 x double> @_Z3cosDv2_d(<2 x double> %x)
  store <2 x double> %call1, ptr addrspace(1) %cos_out, align 16
  ret void
}

define void @sincos_v3f64(<3 x double> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v3f64
; CHECK-SAME: (<3 x double> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <3 x double>, align 32, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <3 x double> @_Z6sincosDv3_dPU3AS5S_(<3 x double> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <3 x double>, ptr addrspace(5) [[__SINCOS_]], align 32
; CHECK-NEXT:    [[EXTRACTVEC2:%.*]] = shufflevector <3 x double> [[TMP0]], <3 x double> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
; CHECK-NEXT:    store <4 x double> [[EXTRACTVEC2]], ptr addrspace(1) [[SIN_OUT]], align 32
; CHECK-NEXT:    [[EXTRACTVEC6:%.*]] = shufflevector <3 x double> [[TMP1]], <3 x double> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
; CHECK-NEXT:    store <4 x double> [[EXTRACTVEC6]], ptr addrspace(1) [[COS_OUT]], align 32
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <3 x double> @_Z3sinDv3_d(<3 x double> %x)
  %extractVec2 = shufflevector <3 x double> %call, <3 x double> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
  store <4 x double> %extractVec2, ptr addrspace(1) %sin_out, align 32
  %call5 = tail call contract <3 x double> @_Z3cosDv3_d(<3 x double> %x)
  %extractVec6 = shufflevector <3 x double> %call5, <3 x double> poison, <4 x i32> <i32 0, i32 1, i32 2, i32 poison>
  store <4 x double> %extractVec6, ptr addrspace(1) %cos_out, align 32
  ret void
}

define void @sincos_v4f64(<4 x double> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v4f64
; CHECK-SAME: (<4 x double> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 32)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <4 x double>, align 32, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <4 x double> @_Z6sincosDv4_dPU3AS5S_(<4 x double> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <4 x double>, ptr addrspace(5) [[__SINCOS_]], align 32
; CHECK-NEXT:    store <4 x double> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 32
; CHECK-NEXT:    store <4 x double> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 32
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <4 x double> @_Z3sinDv4_d(<4 x double> %x)
  store <4 x double> %call, ptr addrspace(1) %sin_out, align 32
  %call1 = tail call contract <4 x double> @_Z3cosDv4_d(<4 x double> %x)
  store <4 x double> %call1, ptr addrspace(1) %cos_out, align 32
  ret void
}

define void @sincos_v8f64(<8 x double> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v8f64
; CHECK-SAME: (<8 x double> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 64)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 64)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <8 x double>, align 64, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <8 x double> @_Z6sincosDv8_dPU3AS5S_(<8 x double> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <8 x double>, ptr addrspace(5) [[__SINCOS_]], align 64
; CHECK-NEXT:    store <8 x double> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 64
; CHECK-NEXT:    store <8 x double> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 64
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <8 x double> @_Z3sinDv8_d(<8 x double> %x)
  store <8 x double> %call, ptr addrspace(1) %sin_out, align 64
  %call1 = tail call contract <8 x double> @_Z3cosDv8_d(<8 x double> %x)
  store <8 x double> %call1, ptr addrspace(1) %cos_out, align 64
  ret void
}

define void @sincos_v16f64(<16 x double> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v16f64
; CHECK-SAME: (<16 x double> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 128)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 128)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <16 x double>, align 128, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract <16 x double> @_Z6sincosDv16_dPU3AS5S_(<16 x double> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <16 x double>, ptr addrspace(5) [[__SINCOS_]], align 128
; CHECK-NEXT:    store <16 x double> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 128
; CHECK-NEXT:    store <16 x double> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 128
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract <16 x double> @_Z3sinDv16_d(<16 x double> %x)
  store <16 x double> %call, ptr addrspace(1) %sin_out, align 128
  %call1 = tail call contract <16 x double> @_Z3cosDv16_d(<16 x double> %x)
  store <16 x double> %call1, ptr addrspace(1) %cos_out, align 128
  ret void
}



define void @sincos_f32_different_blocks(i1 %cond, float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_different_blocks
; CHECK-SAME: (i1 [[COND:%.*]], float [[X:%.*]], ptr addrspace(1) writeonly captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    br i1 [[COND]], label [[BB0:%.*]], label [[BB1:%.*]]
; CHECK:       common.ret:
; CHECK-NEXT:    ret void
; CHECK:       bb0:
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    br label [[COMMON_RET:%.*]]
; CHECK:       bb1:
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    br label [[COMMON_RET]]
;
entry:
  br i1 %cond, label %bb0, label %bb1

bb0:
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  ret void

bb1:
  %call1 = tail call contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

; The sin and cos are in different blocks but always execute
define void @sincos_f32_different_blocks_dominating_always_execute(i1 %cond, float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out, ptr addrspace(1) %other) {
; CHECK-LABEL: define void @sincos_f32_different_blocks_dominating_always_execute
; CHECK-SAME: (i1 [[COND:%.*]], float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]], ptr addrspace(1) writeonly captures(none) [[OTHER:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    br i1 [[COND]], label [[BB0:%.*]], label [[BB1:%.*]]
; CHECK:       bb0:
; CHECK-NEXT:    store i32 0, ptr addrspace(1) [[OTHER]], align 4
; CHECK-NEXT:    br label [[BB1]]
; CHECK:       bb1:
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  br i1 %cond, label %bb0, label %bb1

bb0:
  store i32 0, ptr addrspace(1) %other
  br label %bb1

bb1:
  %call1 = tail call contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

; sin dominates cos but cos doesn't always execute.
define void @sincos_f32_different_blocks_dominating_conditional_execute(i1 %cond, float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out, ptr addrspace(1) %other) {
; CHECK-LABEL: define void @sincos_f32_different_blocks_dominating_conditional_execute
; CHECK-SAME: (i1 [[COND:%.*]], float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]], ptr addrspace(1) readnone captures(none) [[OTHER:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[SPEC_SELECT:%.*]] = select i1 [[COND]], float [[TMP1]], float 0.000000e+00
; CHECK-NEXT:    store float [[SPEC_SELECT]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  br i1 %cond, label %bb0, label %bb1

bb0:
  %call1 = tail call contract float @_Z3cosf(float %x)
  br label %bb1

bb1:
  %phi = phi float [%call1, %bb0], [0.0, %entry]
  store float %phi, ptr addrspace(1) %cos_out, align 4
  ret void
}

define float @select_sin_or_cos_f32(i1 %cond, float %x) {
; CHECK-LABEL: define float @select_sin_or_cos_f32
; CHECK-SAME: (i1 [[COND:%.*]], float [[X:%.*]]) local_unnamed_addr #[[ATTR5:[0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    [[SELECT:%.*]] = select i1 [[COND]], float [[TMP0]], float [[TMP1]]
; CHECK-NEXT:    ret float [[SELECT]]
;
entry:
  %sin = tail call contract float @_Z3sinf(float %x)
  %cos = tail call contract float @_Z3cosf(float %x)
  %select = select i1 %cond, float %sin, float %cos
  ret float %select
}

declare void @func(ptr addrspace(1))

define void @sincos_f32_value_is_instr(ptr addrspace(1) %value.ptr, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_value_is_instr
; CHECK-SAME: (ptr addrspace(1) [[VALUE_PTR:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    tail call void @func(ptr addrspace(1) [[VALUE_PTR]])
; CHECK-NEXT:    [[X:%.*]] = load float, ptr addrspace(1) [[VALUE_PTR]], align 4
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  call void @func(ptr addrspace(1) %value.ptr)
  %x = load float, ptr addrspace(1) %value.ptr
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_value_is_same_constexpr(ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_value_is_same_constexpr
; CHECK-SAME: (ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float bitcast (i32 ptrtoint (ptr @func to i32) to float), ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float bitcast (i32 ptrtoint (ptr @func to i32) to float))
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float bitcast (i32 ptrtoint (ptr @func to i32) to float))
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_value_is_different_constexpr(ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_value_is_different_constexpr
; CHECK-SAME: (ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract float @_Z3sinf(float bitcast (i32 ptrtoint (ptr @func to i32) to float))
; CHECK-NEXT:    store float [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract float @_Z3cosf(float bitcast (i32 ptrtoint (ptr @sincos_f32_value_is_different_constexpr to i32) to float))
; CHECK-NEXT:    store float [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float bitcast (i32 ptrtoint (ptr @func to i32) to float))
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float bitcast (i32 ptrtoint (ptr @sincos_f32_value_is_different_constexpr to i32) to float))
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_value_is_same_constantfp(ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_value_is_same_constantfp
; CHECK-SAME: (ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract float @_Z3sinf(float 4.200000e+01)
; CHECK-NEXT:    store float [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract float @_Z3cosf(float 4.200000e+01)
; CHECK-NEXT:    store float [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float 42.0)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float 42.0)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_value_is_different_constantfp(ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_value_is_different_constantfp
; CHECK-SAME: (ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract float @_Z3sinf(float 4.200000e+01)
; CHECK-NEXT:    store float [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract float @_Z3cosf(float 4.250000e+01)
; CHECK-NEXT:    store float [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float 42.0)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float 42.5)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_different_args(float %x, float %y, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_different_args
; CHECK-SAME: (float [[X:%.*]], float [[Y:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR2]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract float @_Z3sinf(float [[X]])
; CHECK-NEXT:    store float [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract float @_Z3cosf(float [[Y]])
; CHECK-NEXT:    store float [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float %y)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_flag_intersect0(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_flag_intersect0
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call nnan contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call ninf contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_flag_intersect1(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_flag_intersect1
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call nnan contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call nnan nsz contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call nnan ninf contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_v2f32_flag_intersect1(<2 x float> %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_v2f32_flag_intersect1
; CHECK-SAME: (<2 x float> [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 8)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca <2 x float>, align 8, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call nnan contract <2 x float> @_Z6sincosDv2_fPU3AS5S_(<2 x float> [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x float>, ptr addrspace(5) [[__SINCOS_]], align 8
; CHECK-NEXT:    store <2 x float> [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 8
; CHECK-NEXT:    store <2 x float> [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract ninf nnan <2 x float> @_Z3sinDv2_f(<2 x float> %x)
  store <2 x float> %call, ptr addrspace(1) %sin_out, align 8
  %call1 = tail call contract nsz nnan <2 x float> @_Z3cosDv2_f(<2 x float> %x)
  store <2 x float> %call1, ptr addrspace(1) %cos_out, align 8
  ret void
}

declare void @use_stack_ptrs(ptr addrspace(5), ptr addrspace(5))

define void @sincos_f32_alloca_insertpt(float %x) {
; CHECK-LABEL: define void @sincos_f32_alloca_insertpt
; CHECK-SAME: (float [[X:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[ALLOCA0:%.*]] = alloca i32, align 4, addrspace(5)
; CHECK-NEXT:    [[ALLOCA1:%.*]] = alloca i32, align 4, addrspace(5)
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(5) [[ALLOCA0]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(5) [[ALLOCA1]], align 4
; CHECK-NEXT:    call void @use_stack_ptrs(ptr addrspace(5) [[ALLOCA0]], ptr addrspace(5) [[ALLOCA1]])
; CHECK-NEXT:    ret void
;
entry:
  %alloca0 = alloca i32, addrspace(5)
  %alloca1 = alloca i32, addrspace(5)
  %sin = tail call contract float @_Z3sinf(float %x)
  %cos = tail call contract float @_Z3cosf(float %x)
  store float %sin, ptr addrspace(5) %alloca0
  store float %cos, ptr addrspace(5) %alloca1
  call void @use_stack_ptrs(ptr addrspace(5) %alloca0, ptr addrspace(5) %alloca1)
  ret void
}

define float @sincos_f32_unused_result_cos(float %x) {
; CHECK-LABEL: define float @sincos_f32_unused_result_cos
; CHECK-SAME: (float [[X:%.*]]) local_unnamed_addr #[[ATTR6:[0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[SIN:%.*]] = tail call contract float @_Z3sinf(float [[X]])
; CHECK-NEXT:    ret float [[SIN]]
;
entry:
  %alloca0 = alloca i32, addrspace(5)
  %alloca1 = alloca i32, addrspace(5)
  %sin = tail call contract float @_Z3sinf(float %x)
  %dead = tail call contract float @_Z3cosf(float %x)
  ret float %sin
}

define float @sincos_f32_unused_result_sin(float %x) {
; CHECK-LABEL: define float @sincos_f32_unused_result_sin
; CHECK-SAME: (float [[X:%.*]]) local_unnamed_addr #[[ATTR6]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS:%.*]] = tail call contract float @_Z3cosf(float [[X]])
; CHECK-NEXT:    ret float [[COS]]
;
entry:
  %alloca0 = alloca i32, addrspace(5)
  %alloca1 = alloca i32, addrspace(5)
  %dead = tail call contract float @_Z3sinf(float %x)
  %cos = tail call contract float @_Z3cosf(float %x)
  ret float %cos
}

define void @sincos_f32_repeated_uses(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define void @sincos_f32_repeated_uses
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) [[SIN_OUT:%.*]], ptr addrspace(1) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR7:[0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store volatile float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store volatile float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store volatile float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    store volatile float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    store volatile float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %sin0 = tail call contract float @_Z3sinf(float %x)
  store volatile float %sin0, ptr addrspace(1) %sin_out, align 4

  %sin1 = tail call contract float @_Z3sinf(float %x)
  store volatile float %sin1, ptr addrspace(1) %sin_out, align 4

  %cos0 = tail call contract float @_Z3cosf(float %x)
  store volatile float %cos0, ptr addrspace(1) %cos_out, align 4

  %cos1 = tail call contract float @_Z3cosf(float %x)
  store volatile float %cos1, ptr addrspace(1) %cos_out, align 4

  %sin2 = tail call contract float @_Z3sinf(float %x)
  store volatile float %sin2, ptr addrspace(1) %sin_out, align 4

  ret void
}

define void @sin_f32_indirect_call_user(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out, ptr %func.ptr) {
; CHECK-LABEL: define void @sin_f32_indirect_call_user
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]], ptr readonly captures(none) [[FUNC_PTR:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract float @_Z3sinf(float [[X]])
; CHECK-NEXT:    store float [[CALL]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract float [[FUNC_PTR]](float [[X]])
; CHECK-NEXT:    store float [[CALL1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x)
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float %func.ptr(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @cos_f32_indirect_call_user(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out, ptr %func.ptr) {
; CHECK-LABEL: define void @cos_f32_indirect_call_user
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]], ptr readonly captures(none) [[FUNC_PTR:%.*]]) local_unnamed_addr #[[ATTR4]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[CALL:%.*]] = tail call contract float @_Z3cosf(float [[X]])
; CHECK-NEXT:    store float [[CALL]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[CALL1:%.*]] = tail call contract float [[FUNC_PTR]](float [[X]])
; CHECK-NEXT:    store float [[CALL1]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3cosf(float %x)
  store float %call, ptr addrspace(1) %cos_out, align 4
  %call1 = tail call contract float %func.ptr(float %x)
  store float %call1, ptr addrspace(1) %sin_out, align 4
  ret void
}

define void @sincos_f32_preserve_fpmath_0(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_preserve_fpmath_0
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]]), !fpmath [[META5:![0-9]+]]
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x), !fpmath !5
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float %x), !fpmath !5
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_preserve_fpmath_1(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_preserve_fpmath_1
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]]), !fpmath [[META6:![0-9]+]]
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x), !fpmath !5
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float %x), !fpmath !6
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

; Should drop the metadata
define void @sincos_f32_drop_fpmath(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) {
; CHECK-LABEL: define void @sincos_f32_drop_fpmath
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret void
;
entry:
  %call = tail call contract float @_Z3sinf(float %x), !fpmath !5
  store float %call, ptr addrspace(1) %sin_out, align 4
  %call1 = tail call contract float @_Z3cosf(float %x)
  store float %call1, ptr addrspace(1) %cos_out, align 4
  ret void
}

define void @sincos_f32_debuginfo(float %x, ptr addrspace(1) nocapture writeonly %sin_out, ptr addrspace(1) nocapture writeonly %cos_out) !dbg !15 {
; CHECK-LABEL: define void @sincos_f32_debuginfo
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] !dbg [[DBG7:![0-9]+]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5), !dbg [[DBG14:![0-9]+]]
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]]), !dbg [[DBG14]]
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4, !dbg [[DBG14]]
; CHECK-NEXT:      #dbg_value(float [[TMP0]], [[META11:![0-9]+]], !DIExpression(), [[META15:![0-9]+]])
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4, !dbg [[DBG16:![0-9]+]]
; CHECK-NEXT:      #dbg_value(float [[TMP1]], [[META13:![0-9]+]], !DIExpression(), [[META17:![0-9]+]])
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4, !dbg [[DBG18:![0-9]+]]
; CHECK-NEXT:    ret void, !dbg [[DBG19:![0-9]+]]
;
entry:
  %call = tail call contract float @_Z3sinf(float %x), !dbg !19
  call void @llvm.dbg.value(metadata float %call, metadata !17, metadata !DIExpression()), !dbg !19
  store float %call, ptr addrspace(1) %sin_out, align 4, !dbg !20
  %call1 = tail call contract float @_Z3cosf(float %x), !dbg !21
  call void @llvm.dbg.value(metadata float %call1, metadata !18, metadata !DIExpression()), !dbg !21
  store float %call1, ptr addrspace(1) %cos_out, align 4, !dbg !22
  ret void, !dbg !23
}

define float @sin_sincos_private_f32(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @sin_sincos_private_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[SIN0:%.*]] = tail call nnan ninf nsz contract float @_Z3sinf(float [[X]]), !fpmath [[META5]]
; CHECK-NEXT:    store float [[SIN0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[SIN1:%.*]] = call nnan contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[COS_TMP]]), !fpmath [[META6]]
; CHECK-NEXT:    [[COS1:%.*]] = load float, ptr addrspace(5) [[COS_TMP]], align 4
; CHECK-NEXT:    store float [[COS1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp = alloca float, addrspace(5)
  %sin0 = tail call nsz ninf nnan contract float @_Z3sinf(float %x), !fpmath !5
  store float %sin0, ptr addrspace(1) %sin_out, align 4
  %sin1 = call contract nnan float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) %cos.tmp), !fpmath !6
  %cos1 = load float, ptr addrspace(5) %cos.tmp
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @sin_sincos_generic_f32(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @sin_sincos_generic_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[SIN0:%.*]] = tail call nsz contract float @_Z3sinf(float [[X]]), !fpmath [[META5]]
; CHECK-NEXT:    store float [[SIN0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    [[COS_TMP_CAST:%.*]] = addrspacecast ptr addrspace(5) [[COS_TMP]] to ptr
; CHECK-NEXT:    [[SIN1:%.*]] = call ninf nsz contract float @_Z6sincosfPU3AS0f(float [[X]], ptr [[COS_TMP_CAST]]), !fpmath [[META6]]
; CHECK-NEXT:    [[COS1:%.*]] = load float, ptr addrspace(5) [[COS_TMP]], align 4
; CHECK-NEXT:    store float [[COS1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp = alloca float, addrspace(5)
  %sin0 = tail call nsz contract float @_Z3sinf(float %x), !fpmath !5
  store float %sin0, ptr addrspace(1) %sin_out, align 4
  %cos.tmp.cast = addrspacecast ptr addrspace(5) %cos.tmp to ptr
  %sin1 = call contract nsz ninf float @_Z6sincosfPU3AS0f(float %x, ptr %cos.tmp.cast), !fpmath !6
  %cos1 = load float, ptr addrspace(5) %cos.tmp
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @cos_sincos_private_f32(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @cos_sincos_private_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) readnone captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[COS0:%.*]] = tail call contract float @_Z3cosf(float [[X]])
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[SIN1:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[COS_TMP]])
; CHECK-NEXT:    [[COS1:%.*]] = load float, ptr addrspace(5) [[COS_TMP]], align 4
; CHECK-NEXT:    store float [[COS1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp = alloca float, addrspace(5)
  %cos0 = tail call contract float @_Z3cosf(float %x)
  store float %cos0, ptr addrspace(1) %cos_out, align 4
  %sin1 = call contract float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) %cos.tmp)
  %cos1 = load float, ptr addrspace(5) %cos.tmp
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @cos_sincos_generic_f32(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @cos_sincos_generic_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) readnone captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[COS0:%.*]] = tail call contract float @_Z3cosf(float [[X]])
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[COS_TMP_CAST:%.*]] = addrspacecast ptr addrspace(5) [[COS_TMP]] to ptr
; CHECK-NEXT:    [[SIN1:%.*]] = call contract float @_Z6sincosfPU3AS0f(float [[X]], ptr [[COS_TMP_CAST]])
; CHECK-NEXT:    [[COS1:%.*]] = load float, ptr addrspace(5) [[COS_TMP]], align 4
; CHECK-NEXT:    store float [[COS1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp = alloca float, addrspace(5)
  %cos0 = tail call contract float @_Z3cosf(float %x)
  store float %cos0, ptr addrspace(1) %cos_out, align 4
  %cos.tmp.cast = addrspacecast ptr addrspace(5) %cos.tmp to ptr
  %sin1 = call contract float @_Z6sincosfPU3AS0f(float %x, ptr %cos.tmp.cast)
  %cos1 = load float, ptr addrspace(5) %cos.tmp
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @sincos_private_f32_x2(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @sincos_private_f32_x2
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) readnone captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP0:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[COS_TMP1:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[SIN0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[COS_TMP0]])
; CHECK-NEXT:    [[COS0:%.*]] = load float, ptr addrspace(5) [[COS_TMP0]], align 4
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[SIN1:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[COS_TMP1]])
; CHECK-NEXT:    [[COS1:%.*]] = load float, ptr addrspace(5) [[COS_TMP1]], align 4
; CHECK-NEXT:    store float [[COS1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp0 = alloca float, addrspace(5)
  %cos.tmp1 = alloca float, addrspace(5)

  %sin0 = call contract float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) %cos.tmp0)
  %cos0 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos0, ptr addrspace(1) %cos_out, align 4

  %sin1 = call contract float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) %cos.tmp1)
  %cos1 = load float, ptr addrspace(5) %cos.tmp1
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @sincos_generic_f32_x2(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @sincos_generic_f32_x2
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) readnone captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP0:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[COS_TMP1:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[COS_TMP0_CAST:%.*]] = addrspacecast ptr addrspace(5) [[COS_TMP0]] to ptr
; CHECK-NEXT:    [[SIN0:%.*]] = call contract float @_Z6sincosfPU3AS0f(float [[X]], ptr [[COS_TMP0_CAST]])
; CHECK-NEXT:    [[COS0:%.*]] = load float, ptr addrspace(5) [[COS_TMP0]], align 4
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[COS_TMP1_CAST:%.*]] = addrspacecast ptr addrspace(5) [[COS_TMP1]] to ptr
; CHECK-NEXT:    [[SIN1:%.*]] = call contract float @_Z6sincosfPU3AS0f(float [[X]], ptr [[COS_TMP1_CAST]])
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp0 = alloca float, addrspace(5)
  %cos.tmp1 = alloca float, addrspace(5)
  %cos.tmp0.cast = addrspacecast ptr addrspace(5) %cos.tmp0 to ptr
  %sin0 = call contract float @_Z6sincosfPU3AS0f(float %x, ptr %cos.tmp0.cast)
  %cos0 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos0, ptr addrspace(1) %cos_out, align 4
  %cos.tmp1.cast = addrspacecast ptr addrspace(5) %cos.tmp1 to ptr
  %sin1 = call contract float @_Z6sincosfPU3AS0f(float %x, ptr %cos.tmp1.cast)
  %cos1 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @sincos_generic_private_f32(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @sincos_generic_private_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) readnone captures(none) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP0:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[COS_TMP1:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[SIN0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[COS_TMP0]])
; CHECK-NEXT:    [[COS0:%.*]] = load float, ptr addrspace(5) [[COS_TMP0]], align 4
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[COS_TMP1_CAST:%.*]] = addrspacecast ptr addrspace(5) [[COS_TMP1]] to ptr
; CHECK-NEXT:    [[SIN1:%.*]] = call contract float @_Z6sincosfPU3AS0f(float [[X]], ptr [[COS_TMP1_CAST]])
; CHECK-NEXT:    store float [[COS0]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[SIN1]]
;
entry:
  %cos.tmp0 = alloca float, addrspace(5)
  %cos.tmp1 = alloca float, addrspace(5)
  %sin0 = call contract float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) %cos.tmp0)
  %cos0 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos0, ptr addrspace(1) %cos_out, align 4
  %cos.tmp1.cast = addrspacecast ptr addrspace(5) %cos.tmp1 to ptr
  %sin1 = call contract float @_Z6sincosfPU3AS0f(float %x, ptr %cos.tmp1.cast)
  %cos1 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  ret float %sin1
}

define float @sincos_mixed_sin_cos_generic_private_f32(float %x, ptr addrspace(1) %sin_out, ptr addrspace(1) %cos_out) {
; CHECK-LABEL: define float @sincos_mixed_sin_cos_generic_private_f32
; CHECK-SAME: (float [[X:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[SIN_OUT:%.*]], ptr addrspace(1) writeonly captures(none) initializes((0, 4)) [[COS_OUT:%.*]]) local_unnamed_addr #[[ATTR3]] {
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[COS_TMP0:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[__SINCOS_:%.*]] = alloca float, align 4, addrspace(5)
; CHECK-NEXT:    [[TMP0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[__SINCOS_]])
; CHECK-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(5) [[__SINCOS_]], align 4
; CHECK-NEXT:    [[SIN0:%.*]] = call contract float @_Z6sincosfPU3AS5f(float [[X]], ptr addrspace(5) [[COS_TMP0]])
; CHECK-NEXT:    store float [[TMP0]], ptr addrspace(1) [[SIN_OUT]], align 4
; CHECK-NEXT:    store float [[TMP1]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    [[COS2:%.*]] = load float, ptr addrspace(5) [[COS_TMP0]], align 4
; CHECK-NEXT:    store float [[COS2]], ptr addrspace(1) [[COS_OUT]], align 4
; CHECK-NEXT:    ret float [[TMP0]]
;
entry:
  %cos.tmp0 = alloca float, addrspace(5)
  %cos.tmp1 = alloca float, addrspace(5)
  %sin0 = call contract float @_Z6sincosfPU3AS5f(float %x, ptr addrspace(5) %cos.tmp0)
  %sin1 = call contract float @_Z3sinf(float %x)
  store float %sin1, ptr addrspace(1) %sin_out
  %cos0 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos0, ptr addrspace(1) %cos_out, align 4
  %cos.tmp1.cast = addrspacecast ptr addrspace(5) %cos.tmp1 to ptr
  %cos1 = call contract float @_Z3cosf(float %x)
  store float %cos1, ptr addrspace(1) %cos_out, align 4
  %sin2 = call contract float @_Z6sincosfPU3AS0f(float %x, ptr %cos.tmp1.cast)
  %cos2 = load float, ptr addrspace(5) %cos.tmp0
  store float %cos2, ptr addrspace(1) %cos_out, align 4
  ret float %sin2
}

declare void @llvm.dbg.value(metadata, metadata, metadata) #0

attributes #0 = { nocallback nofree nosync nounwind speculatable willreturn memory(none) }
attributes #1 = { argmemonly nounwind willreturn }

!llvm.dbg.cu = !{!0}
!llvm.debugify = !{!2, !3}
!llvm.module.flags = !{!4}

!0 = distinct !DICompileUnit(language: DW_LANG_C, file: !1, producer: "debugify", isOptimized: true, runtimeVersion: 0, emissionKind: FullDebug)
!1 = !DIFile(filename: "/tmp/arst.ll", directory: "/")
!2 = !{i32 10}
!3 = !{i32 4}
!4 = !{i32 2, !"Debug Info Version", i32 3}
!5 = !{float 4.000000e+00}
!6 = !{float 2.000000e+00}
!7 = !DILocation(line: 5, column: 1, scope: !8)
!8 = distinct !DISubprogram(name: "sincos_f32_preserve_fpmath_1", linkageName: "sincos_f32_preserve_fpmath_1", scope: null, file: !1, line: 1, type: !9, scopeLine: 1, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !11)
!9 = !DISubroutineType(types: !10)
!10 = !{}
!11 = !{!12, !14}
!12 = !DILocalVariable(name: "1", scope: !8, file: !1, line: 1, type: !13)
!13 = !DIBasicType(name: "ty32", size: 32, encoding: DW_ATE_unsigned)
!14 = !DILocalVariable(name: "2", scope: !8, file: !1, line: 3, type: !13)
!15 = distinct !DISubprogram(name: "sincos_f32_debuginfo", linkageName: "sincos_f32_debuginfo", scope: null, file: !1, line: 6, type: !9, scopeLine: 6, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !16)
!16 = !{!17, !18}
!17 = !DILocalVariable(name: "3", scope: !15, file: !1, line: 6, type: !13)
!18 = !DILocalVariable(name: "4", scope: !15, file: !1, line: 8, type: !13)
!19 = !DILocation(line: 6, column: 1, scope: !15)
!20 = !DILocation(line: 7, column: 1, scope: !15)
!21 = !DILocation(line: 8, column: 1, scope: !15)
!22 = !DILocation(line: 9, column: 1, scope: !15)
!23 = !DILocation(line: 10, column: 1, scope: !15)