xref: /llvm-project/llvm/test/CodeGen/NVPTX/match.ll (revision b279f6b098d3849f7f1c1f539b108307d5f8ae2d) 
1; RUN: llc < %s -mtriple=nvptx64 -mcpu=sm_70 -mattr=+ptx60 | FileCheck %s
2; RUN: %if ptxas %{ llc < %s -mtriple=nvptx64 -mcpu=sm_70 -mattr=+ptx60 | %ptxas-verify -arch=sm_70 %}
3
4declare i32 @llvm.nvvm.match.any.sync.i32(i32, i32)
5declare i32 @llvm.nvvm.match.any.sync.i64(i32, i64)
6
7; CHECK-LABEL: .func{{.*}}match_any_sync_i32
8define i32 @match_any_sync_i32(i32 %mask, i32 %value) {
9  ; CHECK: ld.param.u32 	[[MASK:%r[0-9]+]], [match_any_sync_i32_param_0];
10  ; CHECK: ld.param.u32 	[[VALUE:%r[0-9]+]], [match_any_sync_i32_param_1];
11
12  ; CHECK:  match.any.sync.b32  [[V0:%r[0-9]+]], [[VALUE]], [[MASK]];
13  %v0 = call i32 @llvm.nvvm.match.any.sync.i32(i32 %mask, i32 %value)
14  ; CHECK:  match.any.sync.b32  [[V1:%r[0-9]+]], [[VALUE]], 1;
15  %v1 = call i32 @llvm.nvvm.match.any.sync.i32(i32 1, i32 %value)
16  ; CHECK:  match.any.sync.b32  [[V2:%r[0-9]+]], 2, [[MASK]];
17  %v2 = call i32 @llvm.nvvm.match.any.sync.i32(i32 %mask, i32 2)
18  ; CHECK:  match.any.sync.b32  [[V3:%r[0-9]+]], 4, 3;
19  %v3 = call i32 @llvm.nvvm.match.any.sync.i32(i32 3, i32 4)
20  %sum1 = add i32 %v0, %v1
21  %sum2 = add i32 %v2, %v3
22  %sum3 = add i32 %sum1, %sum2
23  ret i32 %sum3;
24}
25
26; CHECK-LABEL: .func{{.*}}match_any_sync_i64
27define i32 @match_any_sync_i64(i32 %mask, i64 %value) {
28  ; CHECK: ld.param.u32 	[[MASK:%r[0-9]+]], [match_any_sync_i64_param_0];
29  ; CHECK: ld.param.u64 	[[VALUE:%rd[0-9]+]], [match_any_sync_i64_param_1];
30
31  ; CHECK:  match.any.sync.b64  [[V0:%r[0-9]+]], [[VALUE]], [[MASK]];
32  %v0 = call i32 @llvm.nvvm.match.any.sync.i64(i32 %mask, i64 %value)
33  ; CHECK:  match.any.sync.b64  [[V1:%r[0-9]+]], [[VALUE]], 1;
34  %v1 = call i32 @llvm.nvvm.match.any.sync.i64(i32 1, i64 %value)
35  ; CHECK:  match.any.sync.b64  [[V2:%r[0-9]+]], 2, [[MASK]];
36  %v2 = call i32 @llvm.nvvm.match.any.sync.i64(i32 %mask, i64 2)
37  ; CHECK:  match.any.sync.b64  [[V3:%r[0-9]+]], 4, 3;
38  %v3 = call i32 @llvm.nvvm.match.any.sync.i64(i32 3, i64 4)
39  %sum1 = add i32 %v0, %v1
40  %sum2 = add i32 %v2, %v3
41  %sum3 = add i32 %sum1, %sum2
42  ret i32 %sum3;
43}
44
45declare {i32, i1} @llvm.nvvm.match.all.sync.i32p(i32, i32)
46declare {i32, i1} @llvm.nvvm.match.all.sync.i64p(i32, i64)
47
48; CHECK-LABEL: .func{{.*}}match_all_sync_i32p(
49define {i32,i1} @match_all_sync_i32p(i32 %mask, i32 %value) {
50  ; CHECK: ld.param.u32 	[[MASK:%r[0-9]+]], [match_all_sync_i32p_param_0];
51  ; CHECK: ld.param.u32 	[[VALUE:%r[0-9]+]], [match_all_sync_i32p_param_1];
52
53  ; CHECK:  match.all.sync.b32 {{%r[0-9]+\|%p[0-9]+}}, [[VALUE]], [[MASK]];
54  %r1 = call {i32, i1} @llvm.nvvm.match.all.sync.i32p(i32 %mask, i32 %value)
55  %v1 = extractvalue {i32, i1} %r1, 0
56  %p1 = extractvalue {i32, i1} %r1, 1
57
58  ; CHECK:  match.all.sync.b32 {{%r[0-9]+\|%p[0-9]+}}, 1, [[MASK]];
59  %r2 = call {i32, i1} @llvm.nvvm.match.all.sync.i32p(i32 %mask, i32 1)
60  %v2 = extractvalue {i32, i1} %r2, 0
61  %p2 = extractvalue {i32, i1} %r2, 1
62
63  ; CHECK:  match.all.sync.b32 {{%r[0-9]+\|%p[0-9]+}}, [[VALUE]], 2;
64  %r3 = call {i32, i1} @llvm.nvvm.match.all.sync.i32p(i32 2, i32 %value)
65  %v3 = extractvalue {i32, i1} %r3, 0
66  %p3 = extractvalue {i32, i1} %r3, 1
67
68  ; CHECK:  match.all.sync.b32 {{%r[0-9]+\|%p[0-9]+}}, 4, 3;
69  %r4 = call {i32, i1} @llvm.nvvm.match.all.sync.i32p(i32 3, i32 4)
70  %v4 = extractvalue {i32, i1} %r4, 0
71  %p4 = extractvalue {i32, i1} %r4, 1
72
73  %vsum1 = add i32 %v1, %v2
74  %vsum2 = add i32 %v3, %v4
75  %vsum3 = add i32 %vsum1, %vsum2
76  %psum1 = add i1 %p1, %p2
77  %psum2 = add i1 %p3, %p4
78  %psum3 = add i1 %psum1, %psum2
79  %ret0 = insertvalue {i32, i1} undef, i32 %vsum3, 0
80  %ret1 = insertvalue {i32, i1} %ret0, i1 %psum3, 1
81  ret {i32, i1} %ret1;
82}
83
84; CHECK-LABEL: .func{{.*}}match_all_sync_i64p(
85define {i32,i1} @match_all_sync_i64p(i32 %mask, i64 %value) {
86  ; CHECK: ld.param.u32 	[[MASK:%r[0-9]+]], [match_all_sync_i64p_param_0];
87  ; CHECK: ld.param.u64 	[[VALUE:%rd[0-9]+]], [match_all_sync_i64p_param_1];
88
89  ; CHECK:  match.all.sync.b64 {{%r[0-9]+\|%p[0-9]+}}, [[VALUE]], [[MASK]];
90  %r1 = call {i32, i1} @llvm.nvvm.match.all.sync.i64p(i32 %mask, i64 %value)
91  %v1 = extractvalue {i32, i1} %r1, 0
92  %p1 = extractvalue {i32, i1} %r1, 1
93
94  ; CHECK:  match.all.sync.b64 {{%r[0-9]+\|%p[0-9]+}}, 1, [[MASK]];
95  %r2 = call {i32, i1} @llvm.nvvm.match.all.sync.i64p(i32 %mask, i64 1)
96  %v2 = extractvalue {i32, i1} %r2, 0
97  %p2 = extractvalue {i32, i1} %r2, 1
98
99  ; CHECK:  match.all.sync.b64 {{%r[0-9]+\|%p[0-9]+}}, [[VALUE]], 2;
100  %r3 = call {i32, i1} @llvm.nvvm.match.all.sync.i64p(i32 2, i64 %value)
101  %v3 = extractvalue {i32, i1} %r3, 0
102  %p3 = extractvalue {i32, i1} %r3, 1
103
104  ; CHECK:  match.all.sync.b64 {{%r[0-9]+\|%p[0-9]+}}, 4, 3;
105  %r4 = call {i32, i1} @llvm.nvvm.match.all.sync.i64p(i32 3, i64 4)
106  %v4 = extractvalue {i32, i1} %r4, 0
107  %p4 = extractvalue {i32, i1} %r4, 1
108
109  %vsum1 = add i32 %v1, %v2
110  %vsum2 = add i32 %v3, %v4
111  %vsum3 = add i32 %vsum1, %vsum2
112  %psum1 = add i1 %p1, %p2
113  %psum2 = add i1 %p3, %p4
114  %psum3 = add i1 %psum1, %psum2
115  %ret0 = insertvalue {i32, i1} undef, i32 %vsum3, 0
116  %ret1 = insertvalue {i32, i1} %ret0, i1 %psum3, 1
117  ret {i32, i1} %ret1;
118}
119