test/CodeGen/AMDGPU/promote-alloca-byte-ptr-cast.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -S -mtriple=amdgpu9.00-amd-amdhsa -passes=amdgpu-promote-alloca,verify < %s | FileCheck %s

; Promoting these allocas to a byte vector forces a conversion between a
; byte-vector and a pointer of the same size. Byte types are not integers, so
; before CreateBitPreservingCastChain learned to treat them like integers this
; produced a direct (and invalid) `bitcast <2 x b32> to ptr`, which later
; tripped the MachineVerifier ("bitcast cannot convert between pointers and
; other types"). The conversion must instead route through a pointer-sized
; integer via inttoptr/ptrtoint.

target datalayout = "e-p:64:64-p5:32:32-A5"

define amdgpu_kernel void @ptr_from_byte_vec(ptr %out, ptr %inptr) {
; CHECK-LABEL: @ptr_from_byte_vec(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[A:%.*]] = freeze <2 x b32> poison
; CHECK-NEXT:    [[PBITS:%.*]] = load b64, ptr [[INPTR:%.*]], align 8
; CHECK-NEXT:    [[TMP0:%.*]] = bitcast b64 [[PBITS]] to <2 x b32>
; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <2 x b32> [[TMP0]] to i64
; CHECK-NEXT:    [[TMP2:%.*]] = inttoptr i64 [[TMP1]] to ptr
; CHECK-NEXT:    store ptr [[TMP2]], ptr [[OUT:%.*]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %a = alloca [2 x b32], align 8, addrspace(5)
  %pbits = load b64, ptr %inptr, align 8
  store b64 %pbits, ptr addrspace(5) %a, align 8
  %p = load ptr, ptr addrspace(5) %a, align 8
  store ptr %p, ptr %out, align 8
  ret void
}

define amdgpu_kernel void @byte_vec_from_ptr(ptr %out, ptr %argp) {
; CHECK-LABEL: @byte_vec_from_ptr(
; CHECK-NEXT:  entry:
; CHECK-NEXT:    [[A:%.*]] = freeze <2 x b32> poison
; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[ARGP:%.*]] to i64
; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i64 [[TMP0]] to <2 x b32>
; CHECK-NEXT:    store <2 x b32> [[TMP1]], ptr [[OUT:%.*]], align 8
; CHECK-NEXT:    ret void
;
entry:
  %a = alloca [2 x b32], align 8, addrspace(5)
  store ptr %argp, ptr addrspace(5) %a, align 8
  %v = load <2 x b32>, ptr addrspace(5) %a, align 8
  store <2 x b32> %v, ptr %out, align 8
  ret void
}