llvm/test/Transforms/LoadStoreVectorizer/AMDGPU/prefetch-intrinsics.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 6
; RUN: opt -mtriple=amdgpu12.50-amd-amdhsa -passes=load-store-vectorizer -S < %s | FileCheck %s

; A prefetch does not write the memory it prefetches, so loads on both sides
; of it can be vectorized.

define i32 @global_prefetch(ptr addrspace(1) %p) {
; CHECK-LABEL: define i32 @global_prefetch(
; CHECK-SAME: ptr addrspace(1) [[P:%.*]]) {
; CHECK-NEXT:    [[TMP1:%.*]] = load <2 x i32>, ptr addrspace(1) [[P]], align 8
; CHECK-NEXT:    [[A1:%.*]] = extractelement <2 x i32> [[TMP1]], i64 0
; CHECK-NEXT:    [[B2:%.*]] = extractelement <2 x i32> [[TMP1]], i64 1
; CHECK-NEXT:    call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) [[P]], i32 8)
; CHECK-NEXT:    [[S:%.*]] = add i32 [[A1]], [[B2]]
; CHECK-NEXT:    ret i32 [[S]]
;
  %p1 = getelementptr inbounds i32, ptr addrspace(1) %p, i64 1
  %a = load i32, ptr addrspace(1) %p, align 8
  call void @llvm.amdgcn.global.prefetch(ptr addrspace(1) %p, i32 8)
  %b = load i32, ptr addrspace(1) %p1, align 4
  %s = add i32 %a, %b
  ret i32 %s
}