1; RUN: llc -march=amdgcn -mtriple=amdgcn-- -mcpu=tonga -mattr=-promote-alloca -verify-machineinstrs < %s | FileCheck -enable-var-scope --check-prefixes=GCN,GCN-ALLOCA %s
2; RUN: llc -march=amdgcn -mtriple=amdgcn-- -mcpu=tonga -mattr=+promote-alloca -verify-machineinstrs < %s | FileCheck -enable-var-scope --check-prefixes=GCN,GCN-PROMOTE %s
3; RUN: opt -S -mtriple=amdgcn-- -amdgpu-promote-alloca -sroa -instcombine < %s | FileCheck -check-prefix=OPT %s
4
5target datalayout = "A5"
6
7; OPT-LABEL: @vector_read_alloca_bitcast(
8; OPT-NOT:   alloca
9; OPT:       %0 = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index
10; OPT-NEXT:  store i32 %0, i32 addrspace(1)* %out, align 4
11
12; GCN-LABEL: {{^}}vector_read_alloca_bitcast:
13; GCN-ALLOCA-COUNT-4: buffer_store_dword
14; GCN-ALLOCA:         buffer_load_dword
15
16; GCN_PROMOTE: s_cmp_lg_u32 s{{[0-9]+}}, 2
17; GCN-PROMOTE: v_cmp_eq_u32_e64 [[CC1:[^,]+]], s{{[0-9]+}}, 1
18; GCN-PROMOTE: s_cselect_b64 vcc, -1, 0
19; GCN-PROMOTE: v_cndmask_b32_e{{32|64}} [[IND1:v[0-9]+]], 0, 1, [[CC1]]
20; GCN_PROMOTE: s_cmp_lg_u32 s{{[0-9]+}}, 3
21; GCN-PROMOTE: v_cndmask_b32_e{{32|64}} [[IND2:v[0-9]+]], 2, [[IND1]], vcc
22; GCN-PROMOTE: s_cselect_b64 vcc, -1, 0
23; GCN-PROMOTE: v_cndmask_b32_e{{32|64}} [[IND3:v[0-9]+]], 3, [[IND2]], vcc
24; GCN-PROMOTE: ScratchSize: 0
25
26define amdgpu_kernel void @vector_read_alloca_bitcast(i32 addrspace(1)* %out, i32 %index) {
27entry:
28  %tmp = alloca [4 x i32], addrspace(5)
29  %x = bitcast [4 x i32] addrspace(5)* %tmp to i32 addrspace(5)*
30  %y = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 1
31  %z = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 2
32  %w = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 3
33  store i32 0, i32 addrspace(5)* %x
34  store i32 1, i32 addrspace(5)* %y
35  store i32 2, i32 addrspace(5)* %z
36  store i32 3, i32 addrspace(5)* %w
37  %tmp1 = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 %index
38  %tmp2 = load i32, i32 addrspace(5)* %tmp1
39  store i32 %tmp2, i32 addrspace(1)* %out
40  ret void
41}
42
43; OPT-LABEL: @vector_write_alloca_bitcast(
44; OPT-NOT:   alloca
45; OPT:       %0 = insertelement <4 x i32> zeroinitializer, i32 1, i32 %w_index
46; OPT-NEXT:  %1 = extractelement <4 x i32> %0, i32 %r_index
47; OPT-NEXT:  store i32 %1, i32 addrspace(1)* %out, align
48
49; GCN-LABEL: {{^}}vector_write_alloca_bitcast:
50; GCN-ALLOCA-COUNT-5: buffer_store_dword
51; GCN-ALLOCA:         buffer_load_dword
52
53; GCN-PROMOTE-COUNT-7: v_cndmask
54
55; GCN-PROMOTE: ScratchSize: 0
56
57define amdgpu_kernel void @vector_write_alloca_bitcast(i32 addrspace(1)* %out, i32 %w_index, i32 %r_index) {
58entry:
59  %tmp = alloca [4 x i32], addrspace(5)
60  %x = bitcast [4 x i32] addrspace(5)* %tmp to i32 addrspace(5)*
61  %y = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 1
62  %z = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 2
63  %w = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 3
64  store i32 0, i32 addrspace(5)* %x
65  store i32 0, i32 addrspace(5)* %y
66  store i32 0, i32 addrspace(5)* %z
67  store i32 0, i32 addrspace(5)* %w
68  %tmp1 = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 %w_index
69  store i32 1, i32 addrspace(5)* %tmp1
70  %tmp2 = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 %r_index
71  %tmp3 = load i32, i32 addrspace(5)* %tmp2
72  store i32 %tmp3, i32 addrspace(1)* %out
73  ret void
74}
75
76; OPT-LABEL: @vector_write_read_bitcast_to_float(
77; OPT-NOT:   alloca
78; OPT: bb2:
79; OPT:  %tmp.sroa.0.0 = phi <6 x float> [ undef, %bb ], [ %0, %bb2 ]
80; OPT:  %0 = insertelement <6 x float> %tmp.sroa.0.0, float %tmp73, i32 %tmp10
81; OPT: .preheader:
82; OPT:  %bc = bitcast <6 x float> %0 to <6 x i32>
83; OPT:  %1 = extractelement <6 x i32> %bc, i32 %tmp20
84
85; GCN-LABEL: {{^}}vector_write_read_bitcast_to_float:
86; GCN-ALLOCA: buffer_store_dword
87
88; GCN-PROMOTE-COUNT-6: v_cmp_eq_u16
89; GCN-PROMOTE-COUNT-6: v_cndmask
90
91; GCN: s_cbranch
92
93; GCN-ALLOCA: buffer_load_dword
94
95; GCN-PROMOTE: v_cmp_eq_u16
96; GCN-PROMOTE: v_cndmask
97; GCN-PROMOTE: v_cmp_eq_u16
98; GCN-PROMOTE: v_cndmask
99; GCN-PROMOTE: v_cmp_eq_u16
100; GCN-PROMOTE: v_cndmask
101; GCN-PROMOTE: v_cmp_eq_u16
102; GCN-PROMOTE: v_cndmask
103; GCN-PROMOTE: v_cmp_eq_u16
104; GCN-PROMOTE: v_cndmask
105
106; GCN-PROMOTE: ScratchSize: 0
107
108define amdgpu_kernel void @vector_write_read_bitcast_to_float(float addrspace(1)* %arg) {
109bb:
110  %tmp = alloca [6 x float], align 4, addrspace(5)
111  %tmp1 = bitcast [6 x float] addrspace(5)* %tmp to i8 addrspace(5)*
112  call void @llvm.lifetime.start.p5i8(i64 24, i8 addrspace(5)* %tmp1) #2
113  br label %bb2
114
115bb2:                                              ; preds = %bb2, %bb
116  %tmp3 = phi i32 [ 0, %bb ], [ %tmp13, %bb2 ]
117  %tmp4 = zext i32 %tmp3 to i64
118  %tmp5 = getelementptr inbounds float, float addrspace(1)* %arg, i64 %tmp4
119  %tmp6 = bitcast float addrspace(1)* %tmp5 to i32 addrspace(1)*
120  %tmp7 = load i32, i32 addrspace(1)* %tmp6, align 4
121  %tmp8 = trunc i32 %tmp3 to i16
122  %tmp9 = urem i16 %tmp8, 6
123  %tmp10 = zext i16 %tmp9 to i32
124  %tmp11 = getelementptr inbounds [6 x float], [6 x float] addrspace(5)* %tmp, i32 0, i32 %tmp10
125  %tmp12 = bitcast float addrspace(5)* %tmp11 to i32 addrspace(5)*
126  store i32 %tmp7, i32 addrspace(5)* %tmp12, align 4
127  %tmp13 = add nuw nsw i32 %tmp3, 1
128  %tmp14 = icmp eq i32 %tmp13, 1000
129  br i1 %tmp14, label %.preheader, label %bb2
130
131bb15:                                             ; preds = %.preheader
132  call void @llvm.lifetime.end.p5i8(i64 24, i8 addrspace(5)* %tmp1) #2
133  ret void
134
135.preheader:                                       ; preds = %.preheader, %bb2
136  %tmp16 = phi i32 [ %tmp27, %.preheader ], [ 0, %bb2 ]
137  %tmp17 = trunc i32 %tmp16 to i16
138  %tmp18 = urem i16 %tmp17, 6
139  %tmp19 = sub nuw nsw i16 5, %tmp18
140  %tmp20 = zext i16 %tmp19 to i32
141  %tmp21 = getelementptr inbounds [6 x float], [6 x float] addrspace(5)* %tmp, i32 0, i32 %tmp20
142  %tmp22 = bitcast float addrspace(5)* %tmp21 to i32 addrspace(5)*
143  %tmp23 = load i32, i32 addrspace(5)* %tmp22, align 4
144  %tmp24 = zext i32 %tmp16 to i64
145  %tmp25 = getelementptr inbounds float, float addrspace(1)* %arg, i64 %tmp24
146  %tmp26 = bitcast float addrspace(1)* %tmp25 to i32 addrspace(1)*
147  store i32 %tmp23, i32 addrspace(1)* %tmp26, align 4
148  %tmp27 = add nuw nsw i32 %tmp16, 1
149  %tmp28 = icmp eq i32 %tmp27, 1000
150  br i1 %tmp28, label %bb15, label %.preheader
151}
152
153; OPT-LABEL: @vector_write_read_bitcast_to_double(
154; OPT-NOT:   alloca
155; OPT: bb2:
156; OPT:  %tmp.sroa.0.0 = phi <6 x double> [ undef, %bb ], [ %0, %bb2 ]
157; OPT:  %0 = insertelement <6 x double> %tmp.sroa.0.0, double %tmp73, i32 %tmp10
158; OPT: .preheader:
159; OPT:  %bc = bitcast <6 x double> %0 to <6 x i64>
160; OPT:  %1 = extractelement <6 x i64> %bc, i32 %tmp20
161
162; GCN-LABEL: {{^}}vector_write_read_bitcast_to_double:
163
164; GCN-ALLOCA-COUNT-2: buffer_store_dword
165; GCN-PROMOTE-COUNT-2: v_movreld_b32_e32
166
167; GCN: s_cbranch
168
169; GCN-ALLOCA-COUNT-2: buffer_load_dword
170; GCN-PROMOTE-COUNT-2: v_movrels_b32_e32
171
172; GCN-PROMOTE: ScratchSize: 0
173
174define amdgpu_kernel void @vector_write_read_bitcast_to_double(double addrspace(1)* %arg) {
175bb:
176  %tmp = alloca [6 x double], align 8, addrspace(5)
177  %tmp1 = bitcast [6 x double] addrspace(5)* %tmp to i8 addrspace(5)*
178  call void @llvm.lifetime.start.p5i8(i64 48, i8 addrspace(5)* %tmp1) #2
179  br label %bb2
180
181bb2:                                              ; preds = %bb2, %bb
182  %tmp3 = phi i32 [ 0, %bb ], [ %tmp13, %bb2 ]
183  %tmp4 = zext i32 %tmp3 to i64
184  %tmp5 = getelementptr inbounds double, double addrspace(1)* %arg, i64 %tmp4
185  %tmp6 = bitcast double addrspace(1)* %tmp5 to i64 addrspace(1)*
186  %tmp7 = load i64, i64 addrspace(1)* %tmp6, align 8
187  %tmp8 = trunc i32 %tmp3 to i16
188  %tmp9 = urem i16 %tmp8, 6
189  %tmp10 = zext i16 %tmp9 to i32
190  %tmp11 = getelementptr inbounds [6 x double], [6 x double] addrspace(5)* %tmp, i32 0, i32 %tmp10
191  %tmp12 = bitcast double addrspace(5)* %tmp11 to i64 addrspace(5)*
192  store i64 %tmp7, i64 addrspace(5)* %tmp12, align 8
193  %tmp13 = add nuw nsw i32 %tmp3, 1
194  %tmp14 = icmp eq i32 %tmp13, 1000
195  br i1 %tmp14, label %.preheader, label %bb2
196
197bb15:                                             ; preds = %.preheader
198  call void @llvm.lifetime.end.p5i8(i64 48, i8 addrspace(5)* %tmp1) #2
199  ret void
200
201.preheader:                                       ; preds = %.preheader, %bb2
202  %tmp16 = phi i32 [ %tmp27, %.preheader ], [ 0, %bb2 ]
203  %tmp17 = trunc i32 %tmp16 to i16
204  %tmp18 = urem i16 %tmp17, 6
205  %tmp19 = sub nuw nsw i16 5, %tmp18
206  %tmp20 = zext i16 %tmp19 to i32
207  %tmp21 = getelementptr inbounds [6 x double], [6 x double] addrspace(5)* %tmp, i32 0, i32 %tmp20
208  %tmp22 = bitcast double addrspace(5)* %tmp21 to i64 addrspace(5)*
209  %tmp23 = load i64, i64 addrspace(5)* %tmp22, align 8
210  %tmp24 = zext i32 %tmp16 to i64
211  %tmp25 = getelementptr inbounds double, double addrspace(1)* %arg, i64 %tmp24
212  %tmp26 = bitcast double addrspace(1)* %tmp25 to i64 addrspace(1)*
213  store i64 %tmp23, i64 addrspace(1)* %tmp26, align 8
214  %tmp27 = add nuw nsw i32 %tmp16, 1
215  %tmp28 = icmp eq i32 %tmp27, 1000
216  br i1 %tmp28, label %bb15, label %.preheader
217}
218
219; OPT-LABEL: @vector_write_read_bitcast_to_i64(
220; OPT-NOT:   alloca
221; OPT: bb2:
222; OPT:  %tmp.sroa.0.0 = phi <6 x i64> [ undef, %bb ], [ %0, %bb2 ]
223; OPT:  %0 = insertelement <6 x i64> %tmp.sroa.0.0, i64 %tmp6, i32 %tmp9
224; OPT: .preheader:
225; OPT:  %1 = extractelement <6 x i64> %0, i32 %tmp18
226
227; GCN-LABEL: {{^}}vector_write_read_bitcast_to_i64:
228
229; GCN-ALLOCA-COUNT-2: buffer_store_dword
230; GCN-PROMOTE-COUNT-2: v_movreld_b32_e32
231
232; GCN: s_cbranch
233
234; GCN-ALLOCA-COUNT-2: buffer_load_dword
235; GCN-PROMOTE-COUNT-2: v_movrels_b32_e32
236
237; GCN-PROMOTE: ScratchSize: 0
238
239define amdgpu_kernel void @vector_write_read_bitcast_to_i64(i64 addrspace(1)* %arg) {
240bb:
241  %tmp = alloca [6 x i64], align 8, addrspace(5)
242  %tmp1 = bitcast [6 x i64] addrspace(5)* %tmp to i8 addrspace(5)*
243  call void @llvm.lifetime.start.p5i8(i64 48, i8 addrspace(5)* %tmp1) #2
244  br label %bb2
245
246bb2:                                              ; preds = %bb2, %bb
247  %tmp3 = phi i32 [ 0, %bb ], [ %tmp11, %bb2 ]
248  %tmp4 = zext i32 %tmp3 to i64
249  %tmp5 = getelementptr inbounds i64, i64 addrspace(1)* %arg, i64 %tmp4
250  %tmp6 = load i64, i64 addrspace(1)* %tmp5, align 8
251  %tmp7 = trunc i32 %tmp3 to i16
252  %tmp8 = urem i16 %tmp7, 6
253  %tmp9 = zext i16 %tmp8 to i32
254  %tmp10 = getelementptr inbounds [6 x i64], [6 x i64] addrspace(5)* %tmp, i32 0, i32 %tmp9
255  store i64 %tmp6, i64 addrspace(5)* %tmp10, align 8
256  %tmp11 = add nuw nsw i32 %tmp3, 1
257  %tmp12 = icmp eq i32 %tmp11, 1000
258  br i1 %tmp12, label %.preheader, label %bb2
259
260bb13:                                             ; preds = %.preheader
261  call void @llvm.lifetime.end.p5i8(i64 48, i8 addrspace(5)* %tmp1) #2
262  ret void
263
264.preheader:                                       ; preds = %.preheader, %bb2
265  %tmp14 = phi i32 [ %tmp23, %.preheader ], [ 0, %bb2 ]
266  %tmp15 = trunc i32 %tmp14 to i16
267  %tmp16 = urem i16 %tmp15, 6
268  %tmp17 = sub nuw nsw i16 5, %tmp16
269  %tmp18 = zext i16 %tmp17 to i32
270  %tmp19 = getelementptr inbounds [6 x i64], [6 x i64] addrspace(5)* %tmp, i32 0, i32 %tmp18
271  %tmp20 = load i64, i64 addrspace(5)* %tmp19, align 8
272  %tmp21 = zext i32 %tmp14 to i64
273  %tmp22 = getelementptr inbounds i64, i64 addrspace(1)* %arg, i64 %tmp21
274  store i64 %tmp20, i64 addrspace(1)* %tmp22, align 8
275  %tmp23 = add nuw nsw i32 %tmp14, 1
276  %tmp24 = icmp eq i32 %tmp23, 1000
277  br i1 %tmp24, label %bb13, label %.preheader
278}
279
280; TODO: llvm.assume can be ingored
281
282; OPT-LABEL: @vector_read_alloca_bitcast_assume(
283; OPT:      %tmp = alloca <4 x i32>, align 16, addrspace(5)
284; OPT-NEXT: %x = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %tmp, i64 0, i64 0
285; OPT-NEXT: store i32 0, i32 addrspace(5)* %x, align 16
286; OPT-NEXT: %0 = load <4 x i32>, <4 x i32> addrspace(5)* %tmp, align 16
287; OPT-NEXT: %1 = shufflevector <4 x i32> %0, <4 x i32> <i32 undef, i32 1, i32 2, i32 3>, <4 x i32> <i32 0, i32 5, i32 6, i32 7>
288; OPT-NEXT: store <4 x i32> %1, <4 x i32> addrspace(5)* %tmp, align 16
289; OPT-NEXT: %2 = extractelement <4 x i32> %1, i32 %index
290; OPT-NEXT: store i32 %2, i32 addrspace(1)* %out, align 4
291
292; GCN-LABEL: {{^}}vector_read_alloca_bitcast_assume:
293; GCN-COUNT-4: buffer_store_dword
294
295define amdgpu_kernel void @vector_read_alloca_bitcast_assume(i32 addrspace(1)* %out, i32 %index) {
296entry:
297  %tmp = alloca [4 x i32], addrspace(5)
298  %x = bitcast [4 x i32] addrspace(5)* %tmp to i32 addrspace(5)*
299  %cmp = icmp ne i32 addrspace(5)* %x, null
300  call void @llvm.assume(i1 %cmp)
301  %y = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 1
302  %z = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 2
303  %w = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 3
304  store i32 0, i32 addrspace(5)* %x
305  store i32 1, i32 addrspace(5)* %y
306  store i32 2, i32 addrspace(5)* %z
307  store i32 3, i32 addrspace(5)* %w
308  %tmp1 = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 %index
309  %tmp2 = load i32, i32 addrspace(5)* %tmp1
310  store i32 %tmp2, i32 addrspace(1)* %out
311  ret void
312}
313
314; OPT-LABEL: @vector_read_alloca_multiuse(
315; OPT-NOT:   alloca
316; OPT:       %0 = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index
317; OPT-NEXT:  %add2 = add nuw nsw i32 %0, 1
318; OPT-NEXT:  store i32 %add2, i32 addrspace(1)* %out, align 4
319
320; GCN-LABEL: {{^}}vector_read_alloca_multiuse:
321; GCN-ALLOCA-COUNT-4: buffer_store_dword
322; GCN-ALLOCA:         buffer_load_dword
323
324; GCN_PROMOTE: s_cmp_lg_u32 s{{[0-9]+}}, 2
325; GCN-PROMOTE: v_cmp_eq_u32_e64 [[CC1:[^,]+]], s{{[0-9]+}}, 1
326; GCN-PROMOTE: s_cselect_b64 vcc, -1, 0
327; GCN-PROMOTE: v_cndmask_b32_e{{32|64}} [[IND1:v[0-9]+]], 0, 1, [[CC1]]
328; GCN_PROMOTE: s_cmp_lg_u32 s{{[0-9]+}}, 3
329; GCN-PROMOTE: v_cndmask_b32_e{{32|64}} [[IND2:v[0-9]+]], 2, [[IND1]], vcc
330; GCN-PROMOTE: s_cselect_b64 vcc, -1, 0
331; GCN-PROMOTE: v_cndmask_b32_e{{32|64}} [[IND3:v[0-9]+]], 3, [[IND2]], vcc
332
333; GCN-PROMOTE: ScratchSize: 0
334
335define amdgpu_kernel void @vector_read_alloca_multiuse(i32 addrspace(1)* %out, i32 %index) {
336entry:
337  %tmp = alloca [4 x i32], addrspace(5)
338  %b = bitcast [4 x i32] addrspace(5)* %tmp to float addrspace(5)*
339  %x = bitcast float addrspace(5)* %b to i32 addrspace(5)*
340  %y = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 1
341  %z = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 2
342  %w = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 3
343  store i32 0, i32 addrspace(5)* %x
344  store i32 1, i32 addrspace(5)* %y
345  store i32 2, i32 addrspace(5)* %z
346  store i32 3, i32 addrspace(5)* %w
347  %tmp1 = getelementptr [4 x i32], [4 x i32] addrspace(5)* %tmp, i32 0, i32 %index
348  %tmp2 = load i32, i32 addrspace(5)* %tmp1
349  %tmp3 = load i32, i32 addrspace(5)* %x
350  %tmp4 = load i32, i32 addrspace(5)* %y
351  %add1 = add i32 %tmp2, %tmp3
352  %add2 = add i32 %add1, %tmp4
353  store i32 %add2, i32 addrspace(1)* %out
354  ret void
355}
356
357; OPT-LABEL: @bitcast_vector_to_vector(
358; OPT-NOT:   alloca
359; OPT:       store <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(1)* %out, align 16
360
361; GCN-LABEL: {{^}}bitcast_vector_to_vector:
362; GCN: v_mov_b32_e32 v0, 1
363; GCN: v_mov_b32_e32 v1, 2
364; GCN: v_mov_b32_e32 v2, 3
365; GCN: v_mov_b32_e32 v3, 4
366
367; GCN: ScratchSize: 0
368
369define amdgpu_kernel void @bitcast_vector_to_vector(<4 x i32> addrspace(1)* %out)  {
370.entry:
371  %alloca = alloca <4 x float>, align 16, addrspace(5)
372  %cast = bitcast <4 x float> addrspace(5)* %alloca to <4 x i32> addrspace(5)*
373  store <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(5)* %cast
374  %load = load <4 x i32>, <4 x i32> addrspace(5)* %cast, align 16
375  store <4 x i32> %load, <4 x i32> addrspace(1)* %out
376  ret void
377}
378
379; OPT-LABEL: @vector_bitcast_from_alloca_array(
380; OPT-NOT:   alloca
381; OPT:       store <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(1)* %out, align 16
382
383; GCN-LABEL: {{^}}vector_bitcast_from_alloca_array:
384; GCN: v_mov_b32_e32 v0, 1
385; GCN: v_mov_b32_e32 v1, 2
386; GCN: v_mov_b32_e32 v2, 3
387; GCN: v_mov_b32_e32 v3, 4
388
389; GCN: ScratchSize: 0
390
391define amdgpu_kernel void @vector_bitcast_from_alloca_array(<4 x i32> addrspace(1)* %out)  {
392.entry:
393  %alloca = alloca [4 x float], align 16, addrspace(5)
394  %cast = bitcast [4 x float] addrspace(5)* %alloca to <4 x i32> addrspace(5)*
395  store <4 x i32> <i32 1, i32 2, i32 3, i32 4>, <4 x i32> addrspace(5)* %cast
396  %load = load <4 x i32>, <4 x i32> addrspace(5)* %cast, align 16
397  store <4 x i32> %load, <4 x i32> addrspace(1)* %out
398  ret void
399}
400
401; OPT-LABEL: @vector_bitcast_to_array_from_alloca_array(
402; OPT-NOT:   alloca
403; OPT:      %out.repack = getelementptr inbounds [4 x i32], [4 x i32] addrspace(1)* %out, i64 0, i64 0
404; OPT-NEXT: store i32 1, i32 addrspace(1)* %out.repack, align 4
405; OPT-NEXT: %out.repack1 = getelementptr inbounds [4 x i32], [4 x i32] addrspace(1)* %out, i64 0, i64 1
406; OPT-NEXT: store i32 2, i32 addrspace(1)* %out.repack1, align 4
407; OPT-NEXT: %out.repack2 = getelementptr inbounds [4 x i32], [4 x i32] addrspace(1)* %out, i64 0, i64 2
408; OPT-NEXT: store i32 3, i32 addrspace(1)* %out.repack2, align 4
409; OPT-NEXT: %out.repack3 = getelementptr inbounds [4 x i32], [4 x i32] addrspace(1)* %out, i64 0, i64 3
410; OPT-NEXT: store i32 4, i32 addrspace(1)* %out.repack3, align 4
411
412; GCN-LABEL: {{^}}vector_bitcast_to_array_from_alloca_array:
413; GCN: v_mov_b32_e32 v0, 1
414; GCN: v_mov_b32_e32 v1, 2
415; GCN: v_mov_b32_e32 v2, 3
416; GCN: v_mov_b32_e32 v3, 4
417
418; GCN: ScratchSize: 0
419
420define amdgpu_kernel void @vector_bitcast_to_array_from_alloca_array([4 x i32] addrspace(1)* %out)  {
421.entry:
422  %alloca = alloca [4 x float], align 16, addrspace(5)
423  %cast = bitcast [4 x float] addrspace(5)* %alloca to [4 x i32] addrspace(5)*
424  store [4 x i32] [i32 1, i32 2, i32 3, i32 4], [4 x i32] addrspace(5)* %cast
425  %load = load [4 x i32], [4 x i32] addrspace(5)* %cast, align 16
426  store [4 x i32] %load, [4 x i32] addrspace(1)* %out
427  ret void
428}
429
430; OPT-LABEL: @vector_bitcast_to_struct_from_alloca_array(
431; OPT-NOT:   alloca
432; OPT:      %out.repack = getelementptr inbounds %struct.v4, %struct.v4 addrspace(1)* %out, i64 0, i32 0
433; OPT-NEXT: store i32 1, i32 addrspace(1)* %out.repack, align 4
434; OPT-NEXT: %out.repack1 = getelementptr inbounds %struct.v4, %struct.v4 addrspace(1)* %out, i64 0, i32 1
435; OPT-NEXT: store i32 2, i32 addrspace(1)* %out.repack1, align 4
436; OPT-NEXT: %out.repack2 = getelementptr inbounds %struct.v4, %struct.v4 addrspace(1)* %out, i64 0, i32 2
437; OPT-NEXT: store i32 3, i32 addrspace(1)* %out.repack2, align 4
438; OPT-NEXT: %out.repack3 = getelementptr inbounds %struct.v4, %struct.v4 addrspace(1)* %out, i64 0, i32 3
439; OPT-NEXT: store i32 4, i32 addrspace(1)* %out.repack3, align 4
440
441; GCN-LABEL: {{^}}vector_bitcast_to_struct_from_alloca_array:
442; GCN: v_mov_b32_e32 v0, 1
443; GCN: v_mov_b32_e32 v1, 2
444; GCN: v_mov_b32_e32 v2, 3
445; GCN: v_mov_b32_e32 v3, 4
446
447; GCN: ScratchSize: 0
448
449%struct.v4 = type { i32, i32, i32, i32 }
450
451define amdgpu_kernel void @vector_bitcast_to_struct_from_alloca_array(%struct.v4 addrspace(1)* %out)  {
452.entry:
453  %alloca = alloca [4 x float], align 16, addrspace(5)
454  %cast = bitcast [4 x float] addrspace(5)* %alloca to %struct.v4 addrspace(5)*
455  store %struct.v4 { i32 1, i32 2, i32 3, i32 4 }, %struct.v4 addrspace(5)* %cast
456  %load = load %struct.v4, %struct.v4 addrspace(5)* %cast, align 16
457  store %struct.v4 %load, %struct.v4 addrspace(1)* %out
458  ret void
459}
460
461declare void @llvm.lifetime.start.p5i8(i64 immarg, i8 addrspace(5)* nocapture)
462
463declare void @llvm.lifetime.end.p5i8(i64 immarg, i8 addrspace(5)* nocapture)
464
465declare void @llvm.assume(i1)
466