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
225
226
227
228
229
230
231
232
|
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s --amdgpu-lower-module-lds-strategy=module | FileCheck -enable-var-scope -check-prefixes=CHECK,GFX9 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s --amdgpu-lower-module-lds-strategy=module | FileCheck -enable-var-scope -check-prefixes=CHECK,GFX10 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s --amdgpu-lower-module-lds-strategy=module | FileCheck -enable-var-scope -check-prefixes=CHECK,G_GFX9 %s
; RUN: llc -global-isel -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s --amdgpu-lower-module-lds-strategy=module | FileCheck -enable-var-scope -check-prefixes=CHECK,G_GFX10 %s
; Test case looks at the allocated offset of @used_by_both. It's at zero when
; allocated by itself, but at 8 when allocated in combination with the double.
; Redundantly also checks LDSByteSize.
@used_by_both = addrspace(3) global i32 poison
@used_by_kernel = addrspace(3) global i32 poison
@used_by_function = addrspace(3) global double poison
; kernel that calls no functions and uses an LDS variable allocates only that
; variable, so accesses at at offset 0 and LDSByteSize is 4
define amdgpu_kernel void @nocall_ideal() {
; CHECK-LABEL: nocall_ideal:
; CHECK: ; %bb.0:
; CHECK-NEXT: v_mov_b32_e32 v0, 0
; CHECK-NEXT: ds_write_b32 v0, v0
; CHECK-NEXT: s_endpgm
store i32 0, ptr addrspace(3) @used_by_kernel
ret void
}
; CHECK: ; LDSByteSize: 4 bytes
define void @nonkernel() {
; GFX9-LABEL: nonkernel:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX9-NEXT: v_mov_b32_e32 v0, 0
; GFX9-NEXT: v_mov_b32_e32 v1, v0
; GFX9-NEXT: ds_write_b32 v0, v0 offset:8
; GFX9-NEXT: ds_write_b64 v0, v[0:1]
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_setpc_b64 s[30:31]
;
; GFX10-LABEL: nonkernel:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; GFX10-NEXT: v_mov_b32_e32 v0, 0
; GFX10-NEXT: v_mov_b32_e32 v1, v0
; GFX10-NEXT: ds_write_b32 v0, v0 offset:8
; GFX10-NEXT: ds_write_b64 v0, v[0:1]
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_setpc_b64 s[30:31]
;
; G_GFX9-LABEL: nonkernel:
; G_GFX9: ; %bb.0:
; G_GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; G_GFX9-NEXT: v_mov_b32_e32 v2, 0
; G_GFX9-NEXT: v_mov_b32_e32 v3, 8
; G_GFX9-NEXT: v_mov_b32_e32 v0, 0
; G_GFX9-NEXT: v_mov_b32_e32 v1, 0
; G_GFX9-NEXT: ds_write_b32 v3, v2
; G_GFX9-NEXT: ds_write_b64 v2, v[0:1]
; G_GFX9-NEXT: s_waitcnt lgkmcnt(0)
; G_GFX9-NEXT: s_setpc_b64 s[30:31]
;
; G_GFX10-LABEL: nonkernel:
; G_GFX10: ; %bb.0:
; G_GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
; G_GFX10-NEXT: v_mov_b32_e32 v2, 0
; G_GFX10-NEXT: v_mov_b32_e32 v3, 8
; G_GFX10-NEXT: v_mov_b32_e32 v0, 0
; G_GFX10-NEXT: v_mov_b32_e32 v1, 0
; G_GFX10-NEXT: ds_write_b32 v3, v2
; G_GFX10-NEXT: ds_write_b64 v2, v[0:1]
; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
; G_GFX10-NEXT: s_setpc_b64 s[30:31]
store i32 0, ptr addrspace(3) @used_by_both
store double 0.0, ptr addrspace(3) @used_by_function
ret void
}
; Needs to allocate both variables, store to used_by_both is at sizeof(double)
define amdgpu_kernel void @withcall() {
; GFX9-LABEL: withcall:
; GFX9: ; %bb.0:
; GFX9-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
; GFX9-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
; GFX9-NEXT: s_mov_b32 s22, -1
; GFX9-NEXT: s_mov_b32 s23, 0xe00000
; GFX9-NEXT: s_add_u32 s20, s20, s11
; GFX9-NEXT: s_addc_u32 s21, s21, 0
; GFX9-NEXT: s_mov_b32 s12, s8
; GFX9-NEXT: s_add_u32 s8, s4, 36
; GFX9-NEXT: s_mov_b32 s13, s9
; GFX9-NEXT: s_addc_u32 s9, s5, 0
; GFX9-NEXT: s_getpc_b64 s[4:5]
; GFX9-NEXT: s_add_u32 s4, s4, nonkernel@gotpcrel32@lo+4
; GFX9-NEXT: s_addc_u32 s5, s5, nonkernel@gotpcrel32@hi+12
; GFX9-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX9-NEXT: s_mov_b32 s14, s10
; GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX9-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX9-NEXT: s_mov_b64 s[0:1], s[20:21]
; GFX9-NEXT: v_mov_b32_e32 v3, 0
; GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
; GFX9-NEXT: s_mov_b64 s[2:3], s[22:23]
; GFX9-NEXT: s_mov_b32 s32, 0
; GFX9-NEXT: ds_write_b32 v3, v3 offset:8
; GFX9-NEXT: s_waitcnt lgkmcnt(0)
; GFX9-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX9-NEXT: s_endpgm
;
; GFX10-LABEL: withcall:
; GFX10: ; %bb.0:
; GFX10-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
; GFX10-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
; GFX10-NEXT: s_mov_b32 s22, -1
; GFX10-NEXT: s_mov_b32 s23, 0x31c16000
; GFX10-NEXT: s_add_u32 s20, s20, s11
; GFX10-NEXT: s_addc_u32 s21, s21, 0
; GFX10-NEXT: s_mov_b32 s12, s8
; GFX10-NEXT: s_add_u32 s8, s4, 36
; GFX10-NEXT: s_mov_b32 s13, s9
; GFX10-NEXT: s_addc_u32 s9, s5, 0
; GFX10-NEXT: s_getpc_b64 s[4:5]
; GFX10-NEXT: s_add_u32 s4, s4, nonkernel@gotpcrel32@lo+4
; GFX10-NEXT: s_addc_u32 s5, s5, nonkernel@gotpcrel32@hi+12
; GFX10-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; GFX10-NEXT: s_load_dwordx2 s[16:17], s[4:5], 0x0
; GFX10-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; GFX10-NEXT: v_mov_b32_e32 v3, 0
; GFX10-NEXT: s_mov_b32 s14, s10
; GFX10-NEXT: s_mov_b64 s[10:11], s[6:7]
; GFX10-NEXT: s_mov_b64 s[4:5], s[0:1]
; GFX10-NEXT: v_or3_b32 v31, v0, v1, v2
; GFX10-NEXT: s_mov_b64 s[6:7], s[2:3]
; GFX10-NEXT: s_mov_b64 s[0:1], s[20:21]
; GFX10-NEXT: s_mov_b64 s[2:3], s[22:23]
; GFX10-NEXT: s_mov_b32 s32, 0
; GFX10-NEXT: ds_write_b32 v3, v3 offset:8
; GFX10-NEXT: s_waitcnt lgkmcnt(0)
; GFX10-NEXT: s_swappc_b64 s[30:31], s[16:17]
; GFX10-NEXT: s_endpgm
;
; G_GFX9-LABEL: withcall:
; G_GFX9: ; %bb.0:
; G_GFX9-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
; G_GFX9-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
; G_GFX9-NEXT: s_mov_b32 s22, -1
; G_GFX9-NEXT: s_mov_b32 s23, 0xe00000
; G_GFX9-NEXT: s_add_u32 s20, s20, s11
; G_GFX9-NEXT: s_addc_u32 s21, s21, 0
; G_GFX9-NEXT: s_mov_b32 s16, s8
; G_GFX9-NEXT: s_add_u32 s8, s4, 36
; G_GFX9-NEXT: s_mov_b32 s15, s9
; G_GFX9-NEXT: s_addc_u32 s9, s5, 0
; G_GFX9-NEXT: s_mov_b64 s[12:13], s[0:1]
; G_GFX9-NEXT: s_getpc_b64 s[0:1]
; G_GFX9-NEXT: s_add_u32 s0, s0, nonkernel@gotpcrel32@lo+4
; G_GFX9-NEXT: s_addc_u32 s1, s1, nonkernel@gotpcrel32@hi+12
; G_GFX9-NEXT: s_load_dwordx2 s[18:19], s[0:1], 0x0
; G_GFX9-NEXT: s_mov_b32 s14, s10
; G_GFX9-NEXT: s_mov_b64 s[10:11], s[6:7]
; G_GFX9-NEXT: s_mov_b64 s[6:7], s[2:3]
; G_GFX9-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; G_GFX9-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; G_GFX9-NEXT: s_mov_b64 s[0:1], s[20:21]
; G_GFX9-NEXT: v_mov_b32_e32 v3, 0
; G_GFX9-NEXT: v_mov_b32_e32 v4, 8
; G_GFX9-NEXT: v_or3_b32 v31, v0, v1, v2
; G_GFX9-NEXT: s_mov_b64 s[2:3], s[22:23]
; G_GFX9-NEXT: s_mov_b64 s[4:5], s[12:13]
; G_GFX9-NEXT: s_mov_b32 s12, s16
; G_GFX9-NEXT: s_mov_b32 s13, s15
; G_GFX9-NEXT: s_mov_b32 s32, 0
; G_GFX9-NEXT: ds_write_b32 v4, v3
; G_GFX9-NEXT: s_waitcnt lgkmcnt(0)
; G_GFX9-NEXT: s_swappc_b64 s[30:31], s[18:19]
; G_GFX9-NEXT: s_endpgm
;
; G_GFX10-LABEL: withcall:
; G_GFX10: ; %bb.0:
; G_GFX10-NEXT: s_mov_b32 s20, SCRATCH_RSRC_DWORD0
; G_GFX10-NEXT: s_mov_b32 s21, SCRATCH_RSRC_DWORD1
; G_GFX10-NEXT: s_mov_b32 s22, -1
; G_GFX10-NEXT: s_mov_b32 s23, 0x31c16000
; G_GFX10-NEXT: s_add_u32 s20, s20, s11
; G_GFX10-NEXT: s_addc_u32 s21, s21, 0
; G_GFX10-NEXT: s_mov_b32 s16, s8
; G_GFX10-NEXT: s_add_u32 s8, s4, 36
; G_GFX10-NEXT: s_mov_b32 s15, s9
; G_GFX10-NEXT: s_addc_u32 s9, s5, 0
; G_GFX10-NEXT: s_mov_b64 s[12:13], s[0:1]
; G_GFX10-NEXT: s_getpc_b64 s[0:1]
; G_GFX10-NEXT: s_add_u32 s0, s0, nonkernel@gotpcrel32@lo+4
; G_GFX10-NEXT: s_addc_u32 s1, s1, nonkernel@gotpcrel32@hi+12
; G_GFX10-NEXT: v_lshlrev_b32_e32 v1, 10, v1
; G_GFX10-NEXT: s_load_dwordx2 s[18:19], s[0:1], 0x0
; G_GFX10-NEXT: v_lshlrev_b32_e32 v2, 20, v2
; G_GFX10-NEXT: v_mov_b32_e32 v3, 0
; G_GFX10-NEXT: v_mov_b32_e32 v4, 8
; G_GFX10-NEXT: s_mov_b32 s14, s10
; G_GFX10-NEXT: s_mov_b64 s[10:11], s[6:7]
; G_GFX10-NEXT: v_or3_b32 v31, v0, v1, v2
; G_GFX10-NEXT: s_mov_b64 s[6:7], s[2:3]
; G_GFX10-NEXT: s_mov_b64 s[0:1], s[20:21]
; G_GFX10-NEXT: s_mov_b64 s[2:3], s[22:23]
; G_GFX10-NEXT: s_mov_b64 s[4:5], s[12:13]
; G_GFX10-NEXT: s_mov_b32 s12, s16
; G_GFX10-NEXT: s_mov_b32 s13, s15
; G_GFX10-NEXT: s_mov_b32 s32, 0
; G_GFX10-NEXT: ds_write_b32 v4, v3
; G_GFX10-NEXT: s_waitcnt lgkmcnt(0)
; G_GFX10-NEXT: s_swappc_b64 s[30:31], s[18:19]
; G_GFX10-NEXT: s_endpgm
store i32 0, ptr addrspace(3) @used_by_both
call void @nonkernel()
ret void
}
; CHECK: ; LDSByteSize: 16 bytes
; Previous lowering was less efficient here than necessary as the i32 used
; by the kernel is also used by an unrelated non-kernel function. Codegen
; is now the same as nocall_ideal.
define amdgpu_kernel void @nocall_false_sharing() {
; CHECK-LABEL: nocall_false_sharing:
; CHECK: ; %bb.0:
; CHECK-NEXT: v_mov_b32_e32 v0, 0
; CHECK-NEXT: ds_write_b32 v0, v0
; CHECK-NEXT: s_endpgm
store i32 0, ptr addrspace(3) @used_by_both
ret void
}
; CHECK: ; LDSByteSize: 4 bytes
|
