1; RUN: llc < %s -march=nvptx64 -mcpu=sm_35 | FileCheck %s
2
3declare <2 x float> @barv(<2 x float> %input)
4declare <3 x float> @barv3(<3 x float> %input)
5declare [2 x float] @bara([2 x float] %input)
6declare {float, float} @bars({float, float} %input)
7
8define void @test_v2f32(<2 x float> %input, <2 x float>* %output) {
9; CHECK-LABEL: @test_v2f32
10  %call = tail call <2 x float> @barv(<2 x float> %input)
11; CHECK: .param .align 8 .b8 retval0[8];
12; CHECK: ld.param.v2.f32 {[[E0:%f[0-9]+]], [[E1:%f[0-9]+]]}, [retval0+0];
13  store <2 x float> %call, <2 x float>* %output, align 8
14; CHECK: st.v2.f32 [{{%rd[0-9]+}}], {[[E0]], [[E1]]}
15  ret void
16}
17
18define void @test_v3f32(<3 x float> %input, <3 x float>* %output) {
19; CHECK-LABEL: @test_v3f32
20;
21  %call = tail call <3 x float> @barv3(<3 x float> %input)
22; CHECK: .param .align 16 .b8 retval0[16];
23; CHECK-DAG: ld.param.v2.f32 {[[E0:%f[0-9]+]], [[E1:%f[0-9]+]]}, [retval0+0];
24; CHECK-DAG: ld.param.f32 [[E2:%f[0-9]+]], [retval0+8];
25; Make sure we don't load more values than than we need to.
26; CHECK-NOT: ld.param.f32 [[E3:%f[0-9]+]], [retval0+12];
27  store <3 x float> %call, <3 x float>* %output, align 8
28; CHECK-DAG: st.f32 [{{%rd[0-9]}}+8],
29; -- This is suboptimal. We should do st.v2.f32 instead
30;    of combining 2xf32 info i64.
31; CHECK-DAG: st.u64 [{{%rd[0-9]}}],
32; CHECK: ret;
33  ret void
34}
35
36define void @test_a2f32([2 x float] %input, [2 x float]* %output) {
37; CHECK-LABEL: @test_a2f32
38  %call = tail call [2 x float] @bara([2 x float] %input)
39; CHECK: .param .align 4 .b8 retval0[8];
40; CHECK-DAG: ld.param.f32 [[ELEMA1:%f[0-9]+]], [retval0+0];
41; CHECK-DAG: ld.param.f32 [[ELEMA2:%f[0-9]+]], [retval0+4];
42  store [2 x float] %call, [2 x float]* %output, align 4
43; CHECK: }
44; CHECK-DAG: st.f32 [{{%rd[0-9]+}}], [[ELEMA1]]
45; CHECK-DAG: st.f32 [{{%rd[0-9]+}}+4], [[ELEMA2]]
46  ret void
47; CHECK: ret
48}
49
50define void @test_s2f32({float, float} %input, {float, float}* %output) {
51; CHECK-LABEL: @test_s2f32
52  %call = tail call {float, float} @bars({float, float} %input)
53; CHECK: .param .align 4 .b8 retval0[8];
54; CHECK-DAG: ld.param.f32 [[ELEMS1:%f[0-9]+]], [retval0+0];
55; CHECK-DAG: ld.param.f32 [[ELEMS2:%f[0-9]+]], [retval0+4];
56  store {float, float} %call, {float, float}* %output, align 4
57; CHECK: }
58; CHECK-DAG: st.f32 [{{%rd[0-9]+}}], [[ELEMS1]]
59; CHECK-DAG: st.f32 [{{%rd[0-9]+}}+4], [[ELEMS2]]
60  ret void
61; CHECK: ret
62}
63