; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
; RUN: opt < %s -passes=nvptx-set-byval-param-align -S | FileCheck %s

target datalayout = "e-i64:64-i128:128-v16:16-v32:32-n16:32:64"
target triple = "nvptx64-nvidia-cuda"

%struct.S = type { i32, i32 }

; Kernel byval with no alignment: should get ABI align of i32 (4).
define ptx_kernel void @kernel_no_align(ptr byval(i32) %s) {
; CHECK-LABEL: define ptx_kernel void @kernel_no_align(
; CHECK-SAME: ptr byval(i32) align 4 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Kernel byval with low alignment: should be bumped to ABI align (4).
define ptx_kernel void @kernel_low_align(ptr byval(%struct.S) align 1 %s) {
; CHECK-LABEL: define ptx_kernel void @kernel_low_align(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Kernel byval with adequate alignment: should stay unchanged.
define ptx_kernel void @kernel_high_align(ptr byval(%struct.S) align 8 %s) {
; CHECK-LABEL: define ptx_kernel void @kernel_high_align(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 8 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; External device function: gets ABI alignment (4).
define void @device_extern_no_align(ptr byval(%struct.S) %s) {
; CHECK-LABEL: define void @device_extern_no_align(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Internal device function: gets max(16, ABI) = 16.
define internal void @device_internal_no_align(ptr byval(%struct.S) %s) {
; CHECK-LABEL: define internal void @device_internal_no_align(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 16 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Internal device function with adequate alignment: stays unchanged.
define internal void @device_internal_high_align(ptr byval(%struct.S) align 32 %s) {
; CHECK-LABEL: define internal void @device_internal_high_align(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 32 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Large vector type: ABI alignment exceeds 128 bytes, capped to PTX maximum.
define ptx_kernel void @kernel_align_cap(ptr byval(<256 x i32>) %s) {
; CHECK-LABEL: define ptx_kernel void @kernel_align_cap(
; CHECK-SAME: ptr byval(<256 x i32>) align 128 [[S:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Non-byval args should be untouched.
define ptx_kernel void @no_byval(ptr %p, i32 %x) {
; CHECK-LABEL: define ptx_kernel void @no_byval(
; CHECK-SAME: ptr [[P:%.*]], i32 [[X:%.*]]) {
; CHECK-NEXT:    ret void
;
  ret void
}

; Load propagation: loads at known offsets get alignment upgraded.
define ptx_kernel void @load_propagation(ptr byval(%struct.S) %s) {
; CHECK-LABEL: define ptx_kernel void @load_propagation(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[S:%.*]]) {
; CHECK-NEXT:    [[V1:%.*]] = load i32, ptr [[S]], align 4
; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[S]], i64 0, i32 1
; CHECK-NEXT:    [[V2:%.*]] = load i32, ptr [[GEP]], align 4
; CHECK-NEXT:    ret void
;
  %v1 = load i32, ptr %s, align 1
  %gep = getelementptr inbounds %struct.S, ptr %s, i64 0, i32 1
  %v2 = load i32, ptr %gep, align 1
  ret void
}

; Load propagation through bitcast.
define ptx_kernel void @load_through_bitcast(ptr byval(%struct.S) %s) {
; CHECK-LABEL: define ptx_kernel void @load_through_bitcast(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 4 [[S:%.*]]) {
; CHECK-NEXT:    [[V1:%.*]] = load i64, ptr [[S]], align 4
; CHECK-NEXT:    ret void
;
  %v1 = load i64, ptr %s, align 1
  ret void
}

; Internal function gets higher alignment, which propagates to loads.
define internal void @internal_load_propagation(ptr byval(%struct.S) %s) {
; CHECK-LABEL: define internal void @internal_load_propagation(
; CHECK-SAME: ptr byval([[STRUCT_S:%.*]]) align 16 [[S:%.*]]) {
; CHECK-NEXT:    [[V1:%.*]] = load i32, ptr [[S]], align 16
; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds [[STRUCT_S]], ptr [[S]], i64 0, i32 1
; CHECK-NEXT:    [[V2:%.*]] = load i32, ptr [[GEP]], align 4
; CHECK-NEXT:    ret void
;
  %v1 = load i32, ptr %s, align 1
  %gep = getelementptr inbounds %struct.S, ptr %s, i64 0, i32 1
  %v2 = load i32, ptr %gep, align 1
  ret void
}