1; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s
2; RUN: %if ptxas %{ llc < %s -march=nvptx64 -mcpu=sm_20 | %ptxas-verify %}
3
4; CHECK-LABEL: plain
5define void @plain(i8* %a, i16* %b, i32* %c, i64* %d) local_unnamed_addr {
6  ; CHECK: ld.u8 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
7  %a.load = load i8, i8* %a
8  %a.add = add i8 %a.load, 1
9  ; CHECK: st.u8 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
10  store i8 %a.add, i8* %a
11
12  ; CHECK: ld.u16 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
13  %b.load = load i16, i16* %b
14  %b.add = add i16 %b.load, 1
15  ; CHECK: st.u16 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
16  store i16 %b.add, i16* %b
17
18  ; CHECK: ld.u32 %r{{[0-9]+}}, [%rd{{[0-9]+}}]
19  %c.load = load i32, i32* %c
20  %c.add = add i32 %c.load, 1
21  ; CHECK: st.u32 [%rd{{[0-9]+}}], %r{{[0-9]+}}
22  store i32 %c.add, i32* %c
23
24  ; CHECK: ld.u64 %rd{{[0-9]+}}, [%rd{{[0-9]+}}]
25  %d.load = load i64, i64* %d
26  %d.add = add i64 %d.load, 1
27  ; CHECK: st.u64 [%rd{{[0-9]+}}], %rd{{[0-9]+}}
28  store i64 %d.add, i64* %d
29
30  ret void
31}
32
33; CHECK-LABEL: volatile
34define void @volatile(i8* %a, i16* %b, i32* %c, i64* %d) local_unnamed_addr {
35  ; CHECK: ld.volatile.u8 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
36  %a.load = load volatile i8, i8* %a
37  %a.add = add i8 %a.load, 1
38  ; CHECK: st.volatile.u8 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
39  store volatile i8 %a.add, i8* %a
40
41  ; CHECK: ld.volatile.u16 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
42  %b.load = load volatile i16, i16* %b
43  %b.add = add i16 %b.load, 1
44  ; CHECK: st.volatile.u16 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
45  store volatile i16 %b.add, i16* %b
46
47  ; CHECK: ld.volatile.u32 %r{{[0-9]+}}, [%rd{{[0-9]+}}]
48  %c.load = load volatile i32, i32* %c
49  %c.add = add i32 %c.load, 1
50  ; CHECK: st.volatile.u32 [%rd{{[0-9]+}}], %r{{[0-9]+}}
51  store volatile i32 %c.add, i32* %c
52
53  ; CHECK: ld.volatile.u64 %rd{{[0-9]+}}, [%rd{{[0-9]+}}]
54  %d.load = load volatile i64, i64* %d
55  %d.add = add i64 %d.load, 1
56  ; CHECK: st.volatile.u64 [%rd{{[0-9]+}}], %rd{{[0-9]+}}
57  store volatile i64 %d.add, i64* %d
58
59  ret void
60}
61
62; CHECK-LABEL: monotonic
63define void @monotonic(i8* %a, i16* %b, i32* %c, i64* %d, float* %e) local_unnamed_addr {
64  ; CHECK: ld.volatile.u8 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
65  %a.load = load atomic i8, i8* %a monotonic, align 1
66  %a.add = add i8 %a.load, 1
67  ; CHECK: st.volatile.u8 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
68  store atomic i8 %a.add, i8* %a monotonic, align 1
69
70  ; CHECK: ld.volatile.u16 %rs{{[0-9]+}}, [%rd{{[0-9]+}}]
71  %b.load = load atomic i16, i16* %b monotonic, align 2
72  %b.add = add i16 %b.load, 1
73  ; CHECK: st.volatile.u16 [%rd{{[0-9]+}}], %rs{{[0-9]+}}
74  store atomic i16 %b.add, i16* %b monotonic, align 2
75
76  ; CHECK: ld.volatile.u32 %r{{[0-9]+}}, [%rd{{[0-9]+}}]
77  %c.load = load atomic i32, i32* %c monotonic, align 4
78  %c.add = add i32 %c.load, 1
79  ; CHECK: st.volatile.u32 [%rd{{[0-9]+}}], %r{{[0-9]+}}
80  store atomic i32 %c.add, i32* %c monotonic, align 4
81
82  ; CHECK: ld.volatile.u64 %rd{{[0-9]+}}, [%rd{{[0-9]+}}]
83  %d.load = load atomic i64, i64* %d monotonic, align 8
84  %d.add = add i64 %d.load, 1
85  ; CHECK: st.volatile.u64 [%rd{{[0-9]+}}], %rd{{[0-9]+}}
86  store atomic i64 %d.add, i64* %d monotonic, align 8
87
88  ; CHECK: ld.volatile.f32 %f{{[0-9]+}}, [%rd{{[0-9]+}}]
89  %e.load = load atomic float, float* %e monotonic, align 4
90  %e.add = fadd float %e.load, 1.0
91  ; CHECK: st.volatile.f32 [%rd{{[0-9]+}}], %f{{[0-9]+}}
92  store atomic float %e.add, float* %e monotonic, align 4
93
94  ret void
95}
96