1; RUN: llc -march=amdgcn -mtriple=amdgcn-- -mcpu=verde -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SIVI -check-prefix=SI -check-prefix=FUNC %s
2; RUN: llc -march=amdgcn -mtriple=amdgcn-- -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SIVI -check-prefix=VI -check-prefix=FUNC %s
3; RUN: llc -march=amdgcn -mtriple=amdgcn-- -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=GFX9 -check-prefix=FUNC %s
4; RUN: llc -march=r600 -mtriple=r600-- -mcpu=redwood -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
5; RUN: llc -march=r600 -mtriple=r600-- -mcpu=cayman -verify-machineinstrs < %s | FileCheck -check-prefix=CM -check-prefix=FUNC %s
6
7; FUNC-LABEL: {{^}}store_i1:
8; EG: MEM_RAT MSKOR
9; EG-NOT: MEM_RAT MSKOR
10
11; CM: MEM_RAT MSKOR
12; CM-NOT: MEM_RAT MSKOR
13
14; SIVI: buffer_store_byte
15; GFX9: global_store_byte
16define amdgpu_kernel void @store_i1(i1 addrspace(1)* %out) {
17entry:
18  store i1 true, i1 addrspace(1)* %out
19  ret void
20}
21
22; i8 store
23; FUNC-LABEL: {{^}}store_i8:
24; EG: MEM_RAT MSKOR T[[RW_GPR:[0-9]]].XW, T{{[0-9]}}.X
25; EG-NOT: MEM_RAT MSKOR
26
27; EG: VTX_READ_8
28; EG: AND_INT
29; EG: AND_INT
30; EG: LSHL
31; EG: LSHL
32; EG: LSHL
33
34; SIVI: buffer_store_byte
35; GFX9: global_store_byte
36define amdgpu_kernel void @store_i8(i8 addrspace(1)* %out, i8 %in) {
37entry:
38  store i8 %in, i8 addrspace(1)* %out
39  ret void
40}
41
42; i16 store
43; FUNC-LABEL: {{^}}store_i16:
44; EG: MEM_RAT MSKOR T[[RW_GPR:[0-9]]].XW, T{{[0-9]}}.X
45; EG-NOT: MEM_RAT MSKOR
46
47; EG: VTX_READ_16
48; EG: AND_INT
49; EG: AND_INT
50; EG: LSHL
51; EG: LSHL
52; EG: LSHL
53
54
55; SIVI: buffer_store_short
56; GFX9: global_store_short
57define amdgpu_kernel void @store_i16(i16 addrspace(1)* %out, i16 %in) {
58entry:
59  store i16 %in, i16 addrspace(1)* %out
60  ret void
61}
62
63; FUNC-LABEL: {{^}}store_i24:
64; SIVI: s_lshr_b32 s{{[0-9]+}}, s{{[0-9]+}}, 16
65; SIVI-DAG: buffer_store_byte
66; SIVI-DAG: buffer_store_short
67
68; GFX9-DAG: global_store_byte_d16_hi v{{[0-9]+}}, v{{[0-9]+}}, s{{\[[0-9]+:[0-9]+\]}} offset:2
69; GFX9-DAG: global_store_short
70
71; EG: MEM_RAT MSKOR
72; EG: MEM_RAT MSKOR
73define amdgpu_kernel void @store_i24(i24 addrspace(1)* %out, i24 %in) {
74entry:
75  store i24 %in, i24 addrspace(1)* %out
76  ret void
77}
78
79; FUNC-LABEL: {{^}}store_i25:
80; GCN: s_and_b32 [[AND:s[0-9]+]], s{{[0-9]+}}, 0x1ffffff{{$}}
81; GCN: v_mov_b32_e32 [[VAND:v[0-9]+]], [[AND]]
82; SIVI: buffer_store_dword [[VAND]]
83; GFX9: global_store_dword v{{[0-9]+}}, [[VAND]], s
84
85; EG: MEM_RAT_CACHELESS STORE_RAW
86; EG-NOT: MEM_RAT
87
88; CM: MEM_RAT_CACHELESS STORE_DWORD
89; CM-NOT: MEM_RAT
90define amdgpu_kernel void @store_i25(i25 addrspace(1)* %out, i25 %in) {
91entry:
92  store i25 %in, i25 addrspace(1)* %out
93  ret void
94}
95
96; FUNC-LABEL: {{^}}store_v2i8:
97; v2i8 is naturally 2B aligned
98; EG: MEM_RAT MSKOR
99; EG-NOT: MEM_RAT MSKOR
100
101; CM: MEM_RAT MSKOR
102; CM-NOT: MEM_RAT MSKOR
103
104; SIVI: buffer_store_short
105; GFX9: global_store_short
106define amdgpu_kernel void @store_v2i8(<2 x i8> addrspace(1)* %out, <2 x i32> %in) {
107entry:
108  %0 = trunc <2 x i32> %in to <2 x i8>
109  store <2 x i8> %0, <2 x i8> addrspace(1)* %out
110  ret void
111}
112
113; FUNC-LABEL: {{^}}store_v2i8_unaligned:
114; EG: MEM_RAT MSKOR
115; EG: MEM_RAT MSKOR
116; EG-NOT: MEM_RAT MSKOR
117
118; CM: MEM_RAT MSKOR
119; CM: MEM_RAT MSKOR
120; CM-NOT: MEM_RAT MSKOR
121
122; SI: buffer_store_byte
123define amdgpu_kernel void @store_v2i8_unaligned(<2 x i8> addrspace(1)* %out, <2 x i32> %in) {
124entry:
125  %0 = trunc <2 x i32> %in to <2 x i8>
126  store <2 x i8> %0, <2 x i8> addrspace(1)* %out, align 1
127  ret void
128}
129
130
131; FUNC-LABEL: {{^}}store_v2i16:
132; EG: MEM_RAT_CACHELESS STORE_RAW
133
134; CM: MEM_RAT_CACHELESS STORE_DWORD
135
136; SIVI: buffer_store_dword
137; GFX9: global_store_dword
138define amdgpu_kernel void @store_v2i16(<2 x i16> addrspace(1)* %out, <2 x i32> %in) {
139entry:
140  %0 = trunc <2 x i32> %in to <2 x i16>
141  store <2 x i16> %0, <2 x i16> addrspace(1)* %out
142  ret void
143}
144
145; FUNC-LABEL: {{^}}store_v2i16_unaligned:
146; EG: MEM_RAT MSKOR
147; EG: MEM_RAT MSKOR
148; EG-NOT: MEM_RAT MSKOR
149; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
150
151; CM: MEM_RAT MSKOR
152; CM: MEM_RAT MSKOR
153; CM-NOT: MEM_RAT MSKOR
154; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
155
156; SIVI: buffer_store_short
157; SIVI: buffer_store_short
158
159; GFX9: global_store_short
160; GFX9: global_store_short
161define amdgpu_kernel void @store_v2i16_unaligned(<2 x i16> addrspace(1)* %out, <2 x i32> %in) {
162entry:
163  %0 = trunc <2 x i32> %in to <2 x i16>
164  store <2 x i16> %0, <2 x i16> addrspace(1)* %out, align 2
165  ret void
166}
167
168; FUNC-LABEL: {{^}}store_v4i8:
169; EG: MEM_RAT_CACHELESS STORE_RAW
170
171; CM: MEM_RAT_CACHELESS STORE_DWORD
172
173; SIVI: buffer_store_dword
174; GFX9: global_store_dword
175define amdgpu_kernel void @store_v4i8(<4 x i8> addrspace(1)* %out, <4 x i32> %in) {
176entry:
177  %0 = trunc <4 x i32> %in to <4 x i8>
178  store <4 x i8> %0, <4 x i8> addrspace(1)* %out
179  ret void
180}
181
182; FUNC-LABEL: {{^}}store_v4i8_unaligned:
183; EG: MEM_RAT MSKOR
184; EG: MEM_RAT MSKOR
185; EG: MEM_RAT MSKOR
186; EG: MEM_RAT MSKOR
187; EG-NOT: MEM_RAT MSKOR
188; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
189
190; CM: MEM_RAT MSKOR
191; CM: MEM_RAT MSKOR
192; CM: MEM_RAT MSKOR
193; CM: MEM_RAT MSKOR
194; CM-NOT: MEM_RAT MSKOR
195; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
196
197; SI: buffer_store_byte
198; SI: buffer_store_byte
199; SI: buffer_store_byte
200; SI: buffer_store_byte
201; SI-NOT: buffer_store_dword
202define amdgpu_kernel void @store_v4i8_unaligned(<4 x i8> addrspace(1)* %out, <4 x i32> %in) {
203entry:
204  %0 = trunc <4 x i32> %in to <4 x i8>
205  store <4 x i8> %0, <4 x i8> addrspace(1)* %out, align 1
206  ret void
207}
208
209; FUNC-LABEL: {{^}}store_v4i8_halfaligned:
210; EG: MEM_RAT MSKOR
211; EG: MEM_RAT MSKOR
212; EG-NOT: MEM_RAT MSKOR
213; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
214
215; CM: MEM_RAT MSKOR
216; CM: MEM_RAT MSKOR
217; CM-NOT: MEM_RAT MSKOR
218; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
219
220; SI: buffer_store_short
221; SI: buffer_store_short
222; SI-NOT: buffer_store_dword
223define amdgpu_kernel void @store_v4i8_halfaligned(<4 x i8> addrspace(1)* %out, <4 x i32> %in) {
224entry:
225  %0 = trunc <4 x i32> %in to <4 x i8>
226  store <4 x i8> %0, <4 x i8> addrspace(1)* %out, align 2
227  ret void
228}
229
230; floating-point store
231; FUNC-LABEL: {{^}}store_f32:
232; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+\.X, T[0-9]+\.X}}, 1
233
234; CM: MEM_RAT_CACHELESS STORE_DWORD T{{[0-9]+\.X, T[0-9]+\.X}}
235
236; SIVI: buffer_store_dword
237; GFX9: global_store_dword
238
239define amdgpu_kernel void @store_f32(float addrspace(1)* %out, float %in) {
240  store float %in, float addrspace(1)* %out
241  ret void
242}
243
244; FUNC-LABEL: {{^}}store_v4i16:
245; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+}}.XY
246
247; CM: MEM_RAT_CACHELESS STORE_DWORD T{{[0-9]+}}
248
249; SIVI: buffer_store_dwordx2
250; GFX9: global_store_dwordx2
251define amdgpu_kernel void @store_v4i16(<4 x i16> addrspace(1)* %out, <4 x i32> %in) {
252entry:
253  %0 = trunc <4 x i32> %in to <4 x i16>
254  store <4 x i16> %0, <4 x i16> addrspace(1)* %out
255  ret void
256}
257
258; vec2 floating-point stores
259; FUNC-LABEL: {{^}}store_v2f32:
260; EG: MEM_RAT_CACHELESS STORE_RAW
261
262; CM: MEM_RAT_CACHELESS STORE_DWORD
263
264; SIVI: buffer_store_dwordx2
265; GFX9: global_store_dwordx2
266
267define amdgpu_kernel void @store_v2f32(<2 x float> addrspace(1)* %out, float %a, float %b) {
268entry:
269  %0 = insertelement <2 x float> <float 0.0, float 0.0>, float %a, i32 0
270  %1 = insertelement <2 x float> %0, float %b, i32 1
271  store <2 x float> %1, <2 x float> addrspace(1)* %out
272  ret void
273}
274
275; FUNC-LABEL: {{^}}store_v3i32:
276; SI-DAG: buffer_store_dword v
277; SI-DAG: buffer_store_dwordx2
278
279; VI: buffer_store_dwordx3
280
281; GFX9: global_store_dwordx3
282
283; EG-DAG: MEM_RAT_CACHELESS STORE_RAW {{T[0-9]+\.[XYZW]}}, {{T[0-9]+\.[XYZW]}},
284; EG-DAG: MEM_RAT_CACHELESS STORE_RAW {{T[0-9]+\.XY}}, {{T[0-9]+\.[XYZW]}},
285define amdgpu_kernel void @store_v3i32(<3 x i32> addrspace(1)* %out, <3 x i32> %a) nounwind {
286  store <3 x i32> %a, <3 x i32> addrspace(1)* %out, align 16
287  ret void
288}
289
290; FUNC-LABEL: {{^}}store_v4i32:
291; EG: MEM_RAT_CACHELESS STORE_RAW {{T[0-9]+\.XYZW}}
292; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
293
294; CM: MEM_RAT_CACHELESS STORE_DWORD
295; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
296
297; SIVI: buffer_store_dwordx4
298; GFX9: global_store_dwordx4
299define amdgpu_kernel void @store_v4i32(<4 x i32> addrspace(1)* %out, <4 x i32> %in) {
300entry:
301  store <4 x i32> %in, <4 x i32> addrspace(1)* %out
302  ret void
303}
304
305; FUNC-LABEL: {{^}}store_v4i32_unaligned:
306; EG: MEM_RAT_CACHELESS STORE_RAW {{T[0-9]+\.XYZW}}
307; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
308
309; CM: MEM_RAT_CACHELESS STORE_DWORD
310; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
311
312; SIVI: buffer_store_dwordx4
313; GFX9: global_store_dwordx4
314define amdgpu_kernel void @store_v4i32_unaligned(<4 x i32> addrspace(1)* %out, <4 x i32> %in) {
315entry:
316  store <4 x i32> %in, <4 x i32> addrspace(1)* %out, align 4
317  ret void
318}
319
320; v4f32 store
321; FUNC-LABEL: {{^}}store_v4f32:
322; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+\.XYZW, T[0-9]+\.X}}, 1
323; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
324
325; CM: MEM_RAT_CACHELESS STORE_DWORD
326; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
327
328; SIVI: buffer_store_dwordx4
329; GFX9: global_store_dwordx4
330define amdgpu_kernel void @store_v4f32(<4 x float> addrspace(1)* %out, <4 x float> addrspace(1)* %in) {
331  %1 = load <4 x float>, <4 x float> addrspace(1) * %in
332  store <4 x float> %1, <4 x float> addrspace(1)* %out
333  ret void
334}
335
336; FUNC-LABEL: {{^}}store_i64_i8:
337; EG: MEM_RAT MSKOR
338
339; CM: MEM_RAT MSKOR
340
341; SIVI: buffer_store_byte
342; GFX9: global_store_byte
343define amdgpu_kernel void @store_i64_i8(i8 addrspace(1)* %out, i64 %in) {
344entry:
345  %0 = trunc i64 %in to i8
346  store i8 %0, i8 addrspace(1)* %out
347  ret void
348}
349
350; FUNC-LABEL: {{^}}store_i64_i16:
351; EG: MEM_RAT MSKOR
352; SIVI: buffer_store_short
353; GFX9: global_store_short
354define amdgpu_kernel void @store_i64_i16(i16 addrspace(1)* %out, i64 %in) {
355entry:
356  %0 = trunc i64 %in to i16
357  store i16 %0, i16 addrspace(1)* %out
358  ret void
359}
360
361; The stores in this function are combined by the optimizer to create a
362; 64-bit store with 32-bit alignment.  This is legal and the legalizer
363; should not try to split the 64-bit store back into 2 32-bit stores.
364
365; FUNC-LABEL: {{^}}vecload2:
366; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+\.XY, T[0-9]+\.X}}, 1
367; EG-NOT: MEM_RAT_CACHELESS STORE_RAW
368
369; CM: MEM_RAT_CACHELESS STORE_DWORD
370; CM-NOT: MEM_RAT_CACHELESS STORE_DWORD
371
372; SIVI: buffer_store_dwordx2
373; GFX9: global_store_dwordx2
374define amdgpu_kernel void @vecload2(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* nocapture %mem) #0 {
375entry:
376  %0 = load i32, i32 addrspace(4)* %mem, align 4
377  %arrayidx1.i = getelementptr inbounds i32, i32 addrspace(4)* %mem, i64 1
378  %1 = load i32, i32 addrspace(4)* %arrayidx1.i, align 4
379  store i32 %0, i32 addrspace(1)* %out, align 4
380  %arrayidx1 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 1
381  store i32 %1, i32 addrspace(1)* %arrayidx1, align 4
382  ret void
383}
384
385; When i128 was a legal type this program generated cannot select errors:
386
387; FUNC-LABEL: {{^}}"i128-const-store":
388; EG: MEM_RAT_CACHELESS STORE_RAW T{{[0-9]+}}.XYZW, T{{[0-9]+}}.X, 1
389
390; CM: MEM_RAT_CACHELESS STORE_DWORD T{{[0-9]+}}, T{{[0-9]+}}.X
391
392; SIVI: buffer_store_dwordx4
393; GFX9: global_store_dwordx4
394define amdgpu_kernel void @i128-const-store(i32 addrspace(1)* %out) {
395entry:
396  store i32 1, i32 addrspace(1)* %out, align 4
397  %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 1
398  store i32 1, i32 addrspace(1)* %arrayidx2, align 4
399  %arrayidx4 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 2
400  store i32 2, i32 addrspace(1)* %arrayidx4, align 4
401  %arrayidx6 = getelementptr inbounds i32, i32 addrspace(1)* %out, i64 3
402  store i32 2, i32 addrspace(1)* %arrayidx6, align 4
403  ret void
404}
405
406attributes #0 = { nounwind }
407