1; RUN: llc < %s -march=nvptx -mcpu=sm_20 | FileCheck %s
2
3define void @foo(<2 x i32>* %a) {
4; CHECK: .func foo
5; CHECK: ld.v2.u32
6; CHECK-NEXT: ld.v2.u32
7; CHECK-NEXT: ld.v2.u32
8; CHECK-NEXT: ld.v2.u32
9; CHECK-NEXT: add.s32
10; CHECK-NEXT: add.s32
11; CHECK-NEXT: add.s32
12; CHECK-NEXT: add.s32
13; CHECK-NEXT: add.s32
14; CHECK-NEXT: add.s32
15  %ptr0 = getelementptr <2 x i32>, <2 x i32>* %a, i32 0
16  %val0 = load <2 x i32>, <2 x i32>* %ptr0
17  %ptr1 = getelementptr <2 x i32>, <2 x i32>* %a, i32 1
18  %val1 = load <2 x i32>, <2 x i32>* %ptr1
19  %ptr2 = getelementptr <2 x i32>, <2 x i32>* %a, i32 2
20  %val2 = load <2 x i32>, <2 x i32>* %ptr2
21  %ptr3 = getelementptr <2 x i32>, <2 x i32>* %a, i32 3
22  %val3 = load <2 x i32>, <2 x i32>* %ptr3
23
24  %t0 = add <2 x i32> %val0, %val1
25  %t1 = add <2 x i32> %t0, %val2
26  %t2 = add <2 x i32> %t1, %val3
27
28  store <2 x i32> %t2, <2 x i32>* %a
29
30  ret void
31}
32
33