xref: /llvm-project/llvm/test/CodeGen/AMDGPU/convergence-tokens.ll (revision 8a65ee8b2a05e276ab64682c3c9d3f93490cedcc)
1; RUN: llc -stop-after=amdgpu-isel -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs -o - %s | FileCheck --check-prefixes=CHECK,ISEL %s
2; RUN: llc -stop-after=dead-mi-elimination -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs -o - %s | FileCheck --check-prefixes=CHECK,DEADMI %s
3; RUN: llc -global-isel -stop-after=irtranslator -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs -o - %s | FileCheck %s --check-prefixes=CHECK,GISEL
4
5; CHECK-LABEL: name:            basic_call
6;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
7;        ISEL:    {{.*}} SI_CALL_ISEL {{.*}}, @foo, csr_amdgpu, {{.*}}, implicit [[TOKEN]]
8;      DEADMI:    {{.*}} SI_CALL {{.*}}, @foo, csr_amdgpu, {{.*}}, implicit [[TOKEN]]
9;       GISEL:    {{.*}} G_SI_CALL {{.*}}, @foo, csr_amdgpu, {{.*}}, implicit [[TOKEN]]
10define i32 @basic_call(i32 %src) #0 {
11  %t = call token @llvm.experimental.convergence.entry()
12  %r = call i32 @foo(i32 %src) [ "convergencectrl"(token %t) ]
13  ret i32 %r
14}
15
16; CHECK-LABEL: name:            basic_intrinsic
17;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
18;        ISEL:    CONVERGENCECTRL_GLUE [[TOKEN]]
19;  DEADMI-NOT:    CONVERGENCECTRL_GLUE
20;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[TOKEN]]
21;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[TOKEN]]
22define i32 @basic_intrinsic(i32 %src) #0 {
23  %t = call token @llvm.experimental.convergence.anchor()
24  %r = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t) ]
25  ret i32 %r
26}
27
28; There's nothing to check here. The test is just meant to catch any crashes
29; when a convergent call has no token.
30define i32 @uncontrolled_call(i32 %src) #0 {
31  %r = call i32 @foo(i32 %src)
32  ret i32 %r
33}
34
35; CHECK-LABEL: name:            basic_branch
36;       CHECK:  bb.[[#]].entry:
37;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
38;       CHECK:  bb.[[#]].then:
39;        ISEL:    CONVERGENCECTRL_GLUE [[TOKEN]]
40;  DEADMI-NOT:    CONVERGENCECTRL_GLUE
41;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[TOKEN]]
42;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[TOKEN]]
43define i32 @basic_branch(i32 %src, i1 %cond) #0 {
44entry:
45  %t = call token @llvm.experimental.convergence.anchor()
46  %x = add i32 %src, 1
47  br i1 %cond, label %then, label %else
48
49then:
50  %r = call i32 @llvm.amdgcn.readfirstlane(i32 %x) [ "convergencectrl"(token %t) ]
51  br label %else
52
53else:
54  %p = phi i32 [%r, %then], [%x, %entry]
55  ret i32 %p
56}
57
58; CHECK-LABEL: name:            basic_loop
59;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
60;       CHECK:  bb.[[#]].loop:
61;       CHECK:    [[LOOP:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_LOOP [[TOKEN]]
62;        ISEL:    CONVERGENCECTRL_GLUE [[LOOP]]
63;  DEADMI-NOT:    CONVERGENCECTRL_GLUE
64;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[LOOP]]
65;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[LOOP]]
66define i32 @basic_loop(i32 %src, i1 %cond) #0 {
67  %t1 = call token @llvm.experimental.convergence.anchor()
68  br label %loop
69
70loop:
71  %t2 = call token @llvm.experimental.convergence.loop() [ "convergencectrl"(token %t1) ]
72  %r = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t2) ]
73  br i1 %cond, label %loop, label %end
74
75end:
76  ret i32 %r
77}
78
79; CHECK-LABEL: name: nested
80;       CHECK:    [[ENTRY:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
81;       CHECK:    [[ANCHOR:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ANCHOR
82;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[ANCHOR]]
83;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[ANCHOR]]
84;        ISEL:    {{.*}} = V_READFIRSTLANE_B32 {{.*}}, implicit [[ENTRY]]
85;       GISEL:    {{.*}} = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane){{.*}}, implicit [[ENTRY]]
86define i32 @nested(i32 %src) #0 {
87  %t1 = call token @llvm.experimental.convergence.entry()
88  %t2 = call token @llvm.experimental.convergence.anchor()
89  %r2 = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t2) ]
90  %r1 = call i32 @llvm.amdgcn.readfirstlane(i32 %src) [ "convergencectrl"(token %t1) ]
91  %sum = add i32 %r1, %r2
92  ret i32 %sum
93}
94
95; CHECK-LABEL: name:            tail_call_void_func_void
96;       CHECK:    [[TOKEN:%[0-9]+]]{{[^ ]*}} = CONVERGENCECTRL_ENTRY
97;       CHECK:    {{.*}} SI_TCRETURN {{.*}}, @external_void_func_void, 0, csr_amdgpu, {{.*}}implicit [[TOKEN]]
98define void @tail_call_void_func_void() #0 {
99  %t1 = call token @llvm.experimental.convergence.entry()
100  tail call void @external_void_func_void() [ "convergencectrl"(token %t1) ]
101  ret void
102}
103
104declare hidden void @external_void_func_void() #0
105declare i32 @foo(i32 %x) #0
106
107declare i32 @llvm.amdgcn.readfirstlane(i32) #0
108
109declare token @llvm.experimental.convergence.entry()
110declare token @llvm.experimental.convergence.anchor()
111declare token @llvm.experimental.convergence.loop()
112
113attributes #0 = { nounwind readnone convergent }
114attributes #1 = { nounwind }
115