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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
|
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -amdgpu-ir-lower-kernel-arguments=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN,HSA-VI,FUNC %s
; Repeat of some problematic tests in kernel-args.ll, with the IR
; argument lowering pass disabled. Struct padding needs to be
; accounted for, as well as legalization of types changing offsets.
; FUNC-LABEL: {{^}}i1_arg:
; GCN: s_load_dword s
; GCN: s_and_b32
; HSA-VI: .amdhsa_kernarg_size 12
define amdgpu_kernel void @i1_arg(ptr addrspace(1) %out, i1 %x) #0 {
store i1 %x, ptr addrspace(1) %out, align 1
ret void
}
; FUNC-LABEL: {{^}}v3i8_arg:
; HSA-VI: s_load_dword s{{[0-9]+}}, s[8:9], 0x8
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x0
; HSA-VI: .amdhsa_kernarg_size 12
define amdgpu_kernel void @v3i8_arg(ptr addrspace(1) nocapture %out, <3 x i8> %in) #0 {
entry:
store <3 x i8> %in, ptr addrspace(1) %out, align 4
ret void
}
; FUNC-LABEL: {{^}}v5i8_arg:
; GCN: s_load_dwordx2 s[0:1], s[8:9], 0x0
define amdgpu_kernel void @v5i8_arg(<5 x i8> %in) nounwind {
store <5 x i8> %in, ptr addrspace(1) null
ret void
}
; FUNC-LABEL: {{^}}v6i8_arg:
; GCN: s_load_dwordx2 s[0:1], s[8:9], 0x0
define amdgpu_kernel void @v6i8_arg(<6 x i8> %in) nounwind {
store <6 x i8> %in, ptr addrspace(1) null
ret void
}
; FUNC-LABEL: {{^}}v5i16_arg:
; GCN: s_load_dwordx4 s[0:3], s[8:9], 0x0
define amdgpu_kernel void @v5i16_arg(<5 x i16> %in) nounwind {
store <5 x i16> %in, ptr addrspace(1) null
ret void
}
; FUNC-LABEL: {{^}}v6i16_arg:
; GCN-DAG: s_load_dwordx4 s[0:3], s[8:9], 0x0
define amdgpu_kernel void @v6i16_arg(<6 x i16> %in) nounwind {
store <6 x i16> %in, ptr addrspace(1) null
ret void
}
; FUNC-LABEL: {{^}}v5i32_arg:
; GCN: s_load_dwordx4 s[0:3], s[8:9], 0x0
define amdgpu_kernel void @v5i32_arg(<5 x i32> %in) nounwind {
store <5 x i32> %in, ptr addrspace(1) null
ret void
}
; FUNC-LABEL: {{^}}v6i32_arg:
; GCN: s_load_dwordx4 s[0:3], s[8:9], 0x0
define amdgpu_kernel void @v6i32_arg(<6 x i32> %in) nounwind {
store <6 x i32> %in, ptr addrspace(1) null
ret void
}
; FUNC-LABEL: {{^}}i65_arg:
; HSA-VI: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x0
; HSA-VI: .amdhsa_kernarg_size 24
define amdgpu_kernel void @i65_arg(ptr addrspace(1) nocapture %out, i65 %in) #0 {
entry:
store i65 %in, ptr addrspace(1) %out, align 4
ret void
}
; FUNC-LABEL: {{^}}empty_struct_arg:
; HSA-VI: .amdhsa_kernarg_size 0
define amdgpu_kernel void @empty_struct_arg({} %in) #0 {
ret void
}
; The correct load offsets for these:
; load 4 from 0,
; load 8 from 8
; load 4 from 24
; load 8 from 32
; With the SelectionDAG argument lowering, the alignments for the
; struct members is not properly considered, making these wrong.
; FIXME: Total argument size is computed wrong
; FUNC-LABEL: {{^}}struct_argument_alignment:
; HSA-VI: s_load_dword s{{[0-9]+}}, s[8:9], 0x0
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x8
; HSA-VI: s_load_dword s{{[0-9]+}}, s[8:9], 0x18
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x20
; HSA-VI: .amdhsa_kernarg_size 40
define amdgpu_kernel void @struct_argument_alignment({i32, i64} %arg0, i8, {i32, i64} %arg1) #0 {
%val0 = extractvalue {i32, i64} %arg0, 0
%val1 = extractvalue {i32, i64} %arg0, 1
%val2 = extractvalue {i32, i64} %arg1, 0
%val3 = extractvalue {i32, i64} %arg1, 1
store volatile i32 %val0, ptr addrspace(1) null
store volatile i64 %val1, ptr addrspace(1) null
store volatile i32 %val2, ptr addrspace(1) null
store volatile i64 %val3, ptr addrspace(1) null
ret void
}
; No padding between i8 and next struct, but round up at end to 4 byte
; multiple.
; FUNC-LABEL: {{^}}packed_struct_argument_alignment:
; HSA-VI-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
; HSA-VI: global_load_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}} offset:13
; HSA-VI: global_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}} offset:17
; HSA-VI: s_load_dword s{{[0-9]+}}, s[8:9], 0x0
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x4
; HSA-VI: .amdhsa_kernarg_size 28
define amdgpu_kernel void @packed_struct_argument_alignment(<{i32, i64}> %arg0, i8, <{i32, i64}> %arg1) #0 {
%val0 = extractvalue <{i32, i64}> %arg0, 0
%val1 = extractvalue <{i32, i64}> %arg0, 1
%val2 = extractvalue <{i32, i64}> %arg1, 0
%val3 = extractvalue <{i32, i64}> %arg1, 1
store volatile i32 %val0, ptr addrspace(1) null
store volatile i64 %val1, ptr addrspace(1) null
store volatile i32 %val2, ptr addrspace(1) null
store volatile i64 %val3, ptr addrspace(1) null
ret void
}
; GCN-LABEL: {{^}}struct_argument_alignment_after:
; HSA-VI: s_load_dword s{{[0-9]+}}, s[8:9], 0x0
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x8
; HSA-VI: s_load_dword s{{[0-9]+}}, s[8:9], 0x18
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x20
; HSA-VI: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x30
; HSA-VI: .amdhsa_kernarg_size 64
define amdgpu_kernel void @struct_argument_alignment_after({i32, i64} %arg0, i8, {i32, i64} %arg2, i8, <4 x i32> %arg4) #0 {
%val0 = extractvalue {i32, i64} %arg0, 0
%val1 = extractvalue {i32, i64} %arg0, 1
%val2 = extractvalue {i32, i64} %arg2, 0
%val3 = extractvalue {i32, i64} %arg2, 1
store volatile i32 %val0, ptr addrspace(1) null
store volatile i64 %val1, ptr addrspace(1) null
store volatile i32 %val2, ptr addrspace(1) null
store volatile i64 %val3, ptr addrspace(1) null
store volatile <4 x i32> %arg4, ptr addrspace(1) null
ret void
}
; GCN-LABEL: {{^}}array_3xi32:
; HSA-VI: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x0
define amdgpu_kernel void @array_3xi32(i16 %arg0, [3 x i32] %arg1) {
store volatile i16 %arg0, ptr addrspace(1) poison
store volatile [3 x i32] %arg1, ptr addrspace(1) poison
ret void
}
; GCN-LABEL: {{^}}array_3xi16:
; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x0
define amdgpu_kernel void @array_3xi16(i8 %arg0, [3 x i16] %arg1) {
store volatile i8 %arg0, ptr addrspace(1) poison
store volatile [3 x i16] %arg1, ptr addrspace(1) poison
ret void
}
; GCN-LABEL: {{^}}v2i15_arg:
; GCN: s_load_dword [[DWORD:s[0-9]+]]
; GCN-DAG: s_bfe_u32 [[BFE:s[0-9]+]], [[DWORD]], 0x100010{{$}}
; GCN-DAG: s_and_b32 [[AND:s[0-9]+]], [[DWORD]], 0x7fff{{$}}
define amdgpu_kernel void @v2i15_arg(ptr addrspace(1) nocapture %out, <2 x i15> %in) #0 {
entry:
store <2 x i15> %in, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}v3i15_arg:
; GCN: s_load_dwordx4 [[DWORDX4:s\[[0-9]+:[0-9]+\]]]
; GCN: s_lshl_b64
; GCN: s_and_b32
; GCN: s_and_b32
; GCN: s_or_b32
define amdgpu_kernel void @v3i15_arg(ptr addrspace(1) nocapture %out, <3 x i15> %in) #0 {
entry:
store <3 x i15> %in, ptr addrspace(1) %out, align 4
ret void
}
; Byref pointers should only be treated as offsets from kernarg
; GCN-LABEL: {{^}}byref_constant_i8_arg:
; GCN: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
; GCN: global_load_ubyte v{{[0-9]+}}, [[ZERO]], s[8:9] offset:8
; GCN: .amdhsa_kernarg_size 12
define amdgpu_kernel void @byref_constant_i8_arg(ptr addrspace(1) nocapture %out, ptr addrspace(4) byref(i8) %in.byref) #0 {
%in = load i8, ptr addrspace(4) %in.byref
%ext = zext i8 %in to i32
store i32 %ext, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_constant_i16_arg:
; GCN: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
; GCN: global_load_ushort v{{[0-9]+}}, [[ZERO]], s[8:9] offset:8
; GCN: .amdhsa_kernarg_size 12
define amdgpu_kernel void @byref_constant_i16_arg(ptr addrspace(1) nocapture %out, ptr addrspace(4) byref(i16) %in.byref) #0 {
%in = load i16, ptr addrspace(4) %in.byref
%ext = zext i16 %in to i32
store i32 %ext, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_constant_i32_arg:
; GCN: s_load_dwordx4 [[LOAD:s\[[0-9]+:[0-9]+\]]], s[8:9], 0x0{{$}}
; GCN: .amdhsa_kernarg_size 16
define amdgpu_kernel void @byref_constant_i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(4) byref(i32) %in.byref, i32 %after.offset) #0 {
%in = load i32, ptr addrspace(4) %in.byref
store volatile i32 %in, ptr addrspace(1) %out, align 4
store volatile i32 %after.offset, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_constant_v4i32_arg:
; GCN: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x10{{$}}
; GCN: s_load_dword s{{[0-9]+}}, s[8:9], 0x20{{$}}
; GCN: .amdhsa_kernarg_size 36
define amdgpu_kernel void @byref_constant_v4i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(4) byref(<4 x i32>) %in.byref, i32 %after.offset) #0 {
%in = load <4 x i32>, ptr addrspace(4) %in.byref
store volatile <4 x i32> %in, ptr addrspace(1) %out, align 4
store volatile i32 %after.offset, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_align_constant_i32_arg:
; GCN-DAG: s_load_dwordx2 s[[[IN:[0-9]+]]:[[AFTER_OFFSET:[0-9]+]]], s[8:9], 0x100{{$}}
; GCN-DAG: v_mov_b32_e32 [[V_IN:v[0-9]+]], s[[IN]]
; GCN-DAG: v_mov_b32_e32 [[V_AFTER_OFFSET:v[0-9]+]], s[[AFTER_OFFSET]]
; GCN: global_store_dword v{{[0-9]+}}, [[V_IN]], s
; GCN: global_store_dword v{{[0-9]+}}, [[V_AFTER_OFFSET]], s
; GCN: .amdhsa_kernarg_size 264
define amdgpu_kernel void @byref_align_constant_i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(4) byref(i32) align(256) %in.byref, i32 %after.offset) #0 {
%in = load i32, ptr addrspace(4) %in.byref
store volatile i32 %in, ptr addrspace(1) %out, align 4
store volatile i32 %after.offset, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_natural_align_constant_v16i32_arg:
; GCN-DAG: s_load_dword s{{[0-9]+}}, s[8:9], 0x80
; GCN-DAG: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[8:9], 0x40{{$}}
; GCN: .amdhsa_kernarg_size 132
define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(ptr addrspace(1) nocapture %out, i8, ptr addrspace(4) byref(<16 x i32>) align(64) %in.byref, i32 %after.offset) #0 {
%in = load <16 x i32>, ptr addrspace(4) %in.byref
store volatile <16 x i32> %in, ptr addrspace(1) %out, align 4
store volatile i32 %after.offset, ptr addrspace(1) %out, align 4
ret void
}
; Also accept byref kernel arguments with other global address spaces.
; GCN-LABEL: {{^}}byref_global_i32_arg:
; GCN: s_load_dword [[IN:s[0-9]+]], s[8:9], 0x8{{$}}
; GCN: .amdhsa_kernarg_size 12
define amdgpu_kernel void @byref_global_i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(1) byref(i32) %in.byref) #0 {
%in = load i32, ptr addrspace(1) %in.byref
store i32 %in, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_flat_i32_arg:
; GCN: flat_load_dword [[IN:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}} offset:8{{$}}
define amdgpu_kernel void @byref_flat_i32_arg(ptr addrspace(1) nocapture %out, ptr byref(i32) %in.byref) #0 {
%in = load i32, ptr %in.byref
store i32 %in, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_constant_32bit_i32_arg:
; GCN: s_add_i32 s[[PTR_LO:[0-9]+]], s8, 8
; GCN: s_mov_b32 s[[PTR_HI:[0-9]+]], 0{{$}}
; GCN: s_load_dword s{{[0-9]+}}, s[[[PTR_LO]]:[[PTR_HI]]], 0x0{{$}}
define amdgpu_kernel void @byref_constant_32bit_i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(6) byref(i32) %in.byref) #0 {
%in = load i32, ptr addrspace(6) %in.byref
store i32 %in, ptr addrspace(1) %out, align 4
ret void
}
; define amdgpu_kernel void @byref_unknown_as_i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(999) byref %in.byref) {
; %in = load i32, ptr addrspace(999) %in.byref
; store i32 %in, ptr addrspace(1) %out, align 4
; ret void
; }
; GCN-LABEL: {{^}}multi_byref_constant_i32_arg:
; GCN: s_load_dwordx4 {{s\[[0-9]+:[0-9]+\]}}, s[8:9], 0x0
; GCN: .amdhsa_kernarg_size 20
define amdgpu_kernel void @multi_byref_constant_i32_arg(ptr addrspace(1) nocapture %out, ptr addrspace(4) byref(i32) %in0.byref, ptr addrspace(4) byref(i32) %in1.byref, i32 %after.offset) #0 {
%in0 = load i32, ptr addrspace(4) %in0.byref
%in1 = load i32, ptr addrspace(4) %in1.byref
store volatile i32 %in0, ptr addrspace(1) %out, align 4
store volatile i32 %in1, ptr addrspace(1) %out, align 4
store volatile i32 %after.offset, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}byref_constant_i32_arg_offset0:
; GCN-NOT: s4
; GCN-NOT: s5
; GCN: s_load_dword {{s[0-9]+}}, s[8:9], 0x0{{$}}
; GCN: .amdhsa_kernarg_size 4
define amdgpu_kernel void @byref_constant_i32_arg_offset0(ptr addrspace(4) byref(i32) %in.byref) #0 {
%in = load i32, ptr addrspace(4) %in.byref
store i32 %in, ptr addrspace(1) poison, align 4
ret void
}
attributes #0 = { "amdgpu-no-implicitarg-ptr" }
!llvm.module.flags = !{!0}
!0 = !{i32 1, !"amdhsa_code_object_version", i32 400}
|