1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
| ; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,LOOP %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=hawaii -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,LOOP %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=fiji -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,LOOP %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,NOLOOP %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,NOLOOP,GFX10 %s
; Make sure the op is emitted bundled with a waitcnt with and without the retry loop, and the bundle is not removed by ExpandPostRAPseudos.
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -stop-after=postrapseudos -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=MIR %s
; RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -stop-after=postrapseudos -o - -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=MIR %s
; Minimum offset
; GCN-LABEL: {{^}}gws_barrier_offset0:
; NOLOOP-DAG: s_load_dword [[BAR_NUM:s[0-9]+]]
; NOLOOP-DAG: s_mov_b32 m0, 0{{$}}
; NOLOOP: v_mov_b32_e32 v0, [[BAR_NUM]]
; NOLOOP: ds_gws_barrier v0 gds{{$}}
; LOOP: s_mov_b32 m0, 0{{$}}
; LOOP: [[LOOP:BB[0-9]+_[0-9]+]]:
; LOOP-NEXT: s_setreg_imm32_b32 hwreg(HW_REG_TRAPSTS, 8, 1), 0
; LOOP-NEXT: ds_gws_barrier v0 gds
; LOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; LOOP-NEXT: s_getreg_b32 [[GETREG:s[0-9]+]], hwreg(HW_REG_TRAPSTS, 8, 1)
; LOOP-NEXT: s_cmp_lg_u32 [[GETREG]], 0
; LOOP-NEXT: s_cbranch_scc1 [[LOOP]]
; MIR-LABEL: name: gws_barrier_offset0{{$}}
; MIR: BUNDLE implicit{{( killed)?( renamable)?}} $vgpr0, implicit $m0, implicit $exec {
; MIR-NEXT: DS_GWS_BARRIER renamable $vgpr0, 0, -1, implicit $m0, implicit $exec :: (load 4 from custom GWSResource)
; MIR-NEXT: S_WAITCNT 0
; MIR-NEXT: }
define amdgpu_kernel void @gws_barrier_offset0(i32 %val) #0 {
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 0)
ret void
}
; MIR-LABEL: name: gws_barrier_offset63{{$}}
; Maximum offset
; GCN-LABEL: {{^}}gws_barrier_offset63:
; NOLOOP-DAG: s_load_dword [[BAR_NUM:s[0-9]+]]
; NOLOOP-DAG: s_mov_b32 m0, 0{{$}}
; NOLOOP-DAG: v_mov_b32_e32 v0, [[BAR_NUM]]
; NOLOOP: ds_gws_barrier v0 offset:63 gds{{$}}
define amdgpu_kernel void @gws_barrier_offset63(i32 %val) #0 {
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 63)
ret void
}
; FIXME: Should be able to shift directly into m0
; GCN-LABEL: {{^}}gws_barrier_sgpr_offset:
; NOLOOP-DAG: s_load_dwordx2 s{{\[}}[[BAR_NUM:[0-9]+]]:[[OFFSET:[0-9]+]]{{\]}}
; NOLOOP-DAG: s_lshl_b32 [[SHL:s[0-9]+]], s[[OFFSET]], 16
; NOLOOP-DAG: s_mov_b32 m0, [[SHL]]{{$}}
; NOLOOP-DAG: v_mov_b32_e32 [[GWS_VAL:v[0-9]+]], s[[BAR_NUM]]
; NOLOOP: ds_gws_barrier [[GWS_VAL]] gds{{$}}
define amdgpu_kernel void @gws_barrier_sgpr_offset(i32 %val, i32 %offset) #0 {
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 %offset)
ret void
}
; Variable offset in SGPR with constant add
; GCN-LABEL: {{^}}gws_barrier_sgpr_offset_add1:
; NOLOOP-DAG: s_load_dwordx2 s{{\[}}[[BAR_NUM:[0-9]+]]:[[OFFSET:[0-9]+]]{{\]}}
; NOLOOP-DAG: s_lshl_b32 [[SHL:s[0-9]+]], s[[OFFSET]], 16
; NOLOOP-DAG: s_mov_b32 m0, [[SHL]]{{$}}
; NOLOOP-DAG: v_mov_b32_e32 [[GWS_VAL:v[0-9]+]], s[[BAR_NUM]]
; NOLOOP: ds_gws_barrier [[GWS_VAL]] offset:1 gds{{$}}
define amdgpu_kernel void @gws_barrier_sgpr_offset_add1(i32 %val, i32 %offset.base) #0 {
%offset = add i32 %offset.base, 1
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 %offset)
ret void
}
; GCN-LABEL: {{^}}gws_barrier_vgpr_offset:
; NOLOOP-DAG: s_load_dword [[BAR_NUM:s[0-9]+]]
; NOLOOP-DAG: v_readfirstlane_b32 [[READLANE:s[0-9]+]], v0
; NOLOOP-DAG: s_lshl_b32 [[SHL:s[0-9]+]], [[READLANE]], 16
; NOLOOP-DAG: s_mov_b32 m0, [[SHL]]{{$}}
; NOLOOP-DAG: v_mov_b32_e32 [[GWS_VAL:v[0-9]+]], [[BAR_NUM]]
; NOLOOP: ds_gws_barrier [[GWS_VAL]] gds{{$}}
define amdgpu_kernel void @gws_barrier_vgpr_offset(i32 %val) #0 {
%vgpr.offset = call i32 @llvm.amdgcn.workitem.id.x()
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 %vgpr.offset)
ret void
}
; Variable offset in VGPR with constant add
; GCN-LABEL: {{^}}gws_barrier_vgpr_offset_add:
; NOLOOP-DAG: s_load_dword [[BAR_NUM:s[0-9]+]]
; NOLOOP-DAG: v_readfirstlane_b32 [[READLANE:s[0-9]+]], v0
; NOLOOP-DAG: s_lshl_b32 [[SHL:s[0-9]+]], [[READLANE]], 16
; NOLOOP-DAG: s_mov_b32 m0, [[SHL]]{{$}}
; NOLOOP-DAG: v_mov_b32_e32 [[GWS_VAL:v[0-9]+]], [[BAR_NUM]]
; NOLOOP: ds_gws_barrier [[GWS_VAL]] offset:3 gds{{$}}
define amdgpu_kernel void @gws_barrier_vgpr_offset_add(i32 %val) #0 {
%vgpr.offset.base = call i32 @llvm.amdgcn.workitem.id.x()
%vgpr.offset = add i32 %vgpr.offset.base, 3
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 %vgpr.offset)
ret void
}
@lds = internal unnamed_addr addrspace(3) global i32 undef
; Check if m0 initialization is shared
; GCN-LABEL: {{^}}gws_barrier_save_m0_barrier_constant_offset:
; NOLOOP: s_mov_b32 m0, 0
; NOLOOP: ds_gws_barrier v{{[0-9]+}} offset:10 gds
; LOOP: s_mov_b32 m0, -1
; LOOP: ds_write_b32
; LOOP: s_mov_b32 m0, 0
; LOOP: s_setreg_imm32_b32
; LOOP: ds_gws_barrier v{{[0-9]+}} offset:10 gds
; LOOP: s_cbranch_scc1
; LOOP: s_mov_b32 m0, -1
; LOOP: ds_write_b32
define amdgpu_kernel void @gws_barrier_save_m0_barrier_constant_offset(i32 %val) #0 {
store i32 1, i32 addrspace(3)* @lds
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 10)
store i32 2, i32 addrspace(3)* @lds
ret void
}
; Make sure this increments lgkmcnt
; GCN-LABEL: {{^}}gws_barrier_lgkmcnt:
; NOLOOP: s_mov_b32 m0, 0{{$}}
; NOLOOP: ds_gws_barrier v0 gds{{$}}
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; NOLOOP-NEXT: s_setpc_b64
define void @gws_barrier_lgkmcnt(i32 %val) {
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 0)
ret void
}
; Does not imply memory fence on its own
; GCN-LABEL: {{^}}gws_barrier_wait_before:
; NOLOOP: s_waitcnt
; NOLOOP-NOT: s_waitcnt{{$}}
define amdgpu_kernel void @gws_barrier_wait_before(i32 %val, i32 addrspace(1)* %ptr) #0 {
store i32 0, i32 addrspace(1)* %ptr
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 7)
ret void
}
; GCN-LABEL: {{^}}gws_barrier_wait_after:
; NOLOOP: s_mov_b32 m0, 0{{$}}
; NOLOOP: ds_gws_barrier v{{[0-9]+}} offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; NOLOOP-NEXT: load_dword
define amdgpu_kernel void @gws_barrier_wait_after(i32 %val, i32 addrspace(1)* %ptr) #0 {
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 7)
%load = load volatile i32, i32 addrspace(1)* %ptr
ret void
}
; Does not imply memory fence on its own
; GCN-LABEL: {{^}}gws_barrier_fence_before:
; NOLOOP: s_mov_b32 m0, 0{{$}}
; NOLOOP: store_dword
; NOLOOP: s_waitcnt vmcnt(0) lgkmcnt(0)
; NOLOOP: ds_gws_barrier v{{[0-9]+}} offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
define amdgpu_kernel void @gws_barrier_fence_before(i32 %val, i32 addrspace(1)* %ptr) #0 {
store i32 0, i32 addrspace(1)* %ptr
fence release
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 7)
ret void
}
; FIXME: Extra waitcnt
; GCN-LABEL: {{^}}gws_barrier_fence_after:
; NOLOOP: s_mov_b32 m0, 0{{$}}
; NOLOOP: ds_gws_barrier v{{[0-9]+}} offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; NOLOOP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; NOLOOP-NEXT: load_dword
define amdgpu_kernel void @gws_barrier_fence_after(i32 %val, i32 addrspace(1)* %ptr) #0 {
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 7)
fence release
%load = load volatile i32, i32 addrspace(1)* %ptr
ret void
}
; FIXME: Should a wait be inserted here, or is an explicit fence needed?
; GCN-LABEL: {{^}}gws_init_barrier:
; NOLOOP: s_mov_b32 m0, 0
; NOLOOP: ds_gws_init v0 offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; NOLOOP-NEXT: ds_gws_barrier v0 offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
define amdgpu_kernel void @gws_init_barrier(i32 %val) #0 {
call void @llvm.amdgcn.ds.gws.init(i32 %val, i32 7)
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 7)
ret void
}
; FIXME: Why vmcnt, not expcnt?
; GCN-LABEL: {{^}}gws_init_fence_barrier:
; NOLOOP: s_mov_b32 m0, 0
; NOLOOP: ds_gws_init v0 offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; NOLOOP-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
; NOLOOP-NEXT: ds_gws_barrier v0 offset:7 gds
; NOLOOP-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
define amdgpu_kernel void @gws_init_fence_barrier(i32 %val) #0 {
call void @llvm.amdgcn.ds.gws.init(i32 %val, i32 7)
fence release
call void @llvm.amdgcn.ds.gws.barrier(i32 %val, i32 7)
ret void
}
declare void @llvm.amdgcn.ds.gws.barrier(i32, i32) #1
declare void @llvm.amdgcn.ds.gws.init(i32, i32) #2
declare i32 @llvm.amdgcn.workitem.id.x() #3
attributes #0 = { nounwind }
attributes #1 = { convergent inaccessiblememonly nounwind }
attributes #2 = { convergent inaccessiblememonly nounwind writeonly }
attributes #3 = { nounwind readnone speculatable }
|