1 ; RUN: llc -march=r600 -mcpu=redwood < %s | FileCheck %s -check-prefix=R600 -check-prefix=FUNC
2 ; RUN: llc -show-mc-encoding -mattr=+promote-alloca -verify-machineinstrs -march=amdgcn -mcpu=SI < %s | FileCheck %s -check-prefix=SI-PROMOTE -check-prefix=SI -check-prefix=FUNC
3 ; RUN: llc -show-mc-encoding -mattr=-promote-alloca -verify-machineinstrs -march=amdgcn -mcpu=SI < %s | FileCheck %s -check-prefix=SI-ALLOCA -check-prefix=SI -check-prefix=FUNC
4 ; RUN: llc -show-mc-encoding -mattr=+promote-alloca -verify-machineinstrs -march=amdgcn -mcpu=tonga < %s | FileCheck %s -check-prefix=SI-PROMOTE -check-prefix=SI -check-prefix=FUNC
5 ; RUN: llc -show-mc-encoding -mattr=-promote-alloca -verify-machineinstrs -march=amdgcn -mcpu=tonga < %s | FileCheck %s -check-prefix=SI-ALLOCA -check-prefix=SI -check-prefix=FUNC
7 declare i32 @llvm.r600.read.tidig.x() nounwind readnone
9 ; FUNC-LABEL: {{^}}mova_same_clause:
16 ; SI-PROMOTE: ds_write_b32
17 ; SI-PROMOTE: ds_write_b32
18 ; SI-PROMOTE: ds_read_b32
19 ; SI-PROMOTE: ds_read_b32
21 ; SI-ALLOCA: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ; encoding: [0x00,0x10,0x70,0xe0
22 ; SI-ALLOCA: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ; encoding: [0x00,0x10,0x70,0xe0
23 define void @mova_same_clause(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
25 %stack = alloca [5 x i32], align 4
26 %0 = load i32, i32 addrspace(1)* %in, align 4
27 %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %0
28 store i32 4, i32* %arrayidx1, align 4
29 %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
30 %1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
31 %arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %1
32 store i32 5, i32* %arrayidx3, align 4
33 %arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 0
34 %2 = load i32, i32* %arrayidx10, align 4
35 store i32 %2, i32 addrspace(1)* %out, align 4
36 %arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 1
37 %3 = load i32, i32* %arrayidx12
38 %arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
39 store i32 %3, i32 addrspace(1)* %arrayidx13
43 ; This test checks that the stack offset is calculated correctly for structs.
44 ; All register loads/stores should be optimized away, so there shouldn't be
45 ; any MOVA instructions.
47 ; XXX: This generated code has unnecessary MOVs, we should be able to optimize
50 ; FUNC-LABEL: {{^}}multiple_structs:
54 %struct.point = type { i32, i32 }
56 define void @multiple_structs(i32 addrspace(1)* %out) {
58 %a = alloca %struct.point
59 %b = alloca %struct.point
60 %a.x.ptr = getelementptr %struct.point, %struct.point* %a, i32 0, i32 0
61 %a.y.ptr = getelementptr %struct.point, %struct.point* %a, i32 0, i32 1
62 %b.x.ptr = getelementptr %struct.point, %struct.point* %b, i32 0, i32 0
63 %b.y.ptr = getelementptr %struct.point, %struct.point* %b, i32 0, i32 1
64 store i32 0, i32* %a.x.ptr
65 store i32 1, i32* %a.y.ptr
66 store i32 2, i32* %b.x.ptr
67 store i32 3, i32* %b.y.ptr
68 %a.indirect.ptr = getelementptr %struct.point, %struct.point* %a, i32 0, i32 0
69 %b.indirect.ptr = getelementptr %struct.point, %struct.point* %b, i32 0, i32 0
70 %a.indirect = load i32, i32* %a.indirect.ptr
71 %b.indirect = load i32, i32* %b.indirect.ptr
72 %0 = add i32 %a.indirect, %b.indirect
73 store i32 %0, i32 addrspace(1)* %out
77 ; Test direct access of a private array inside a loop. The private array
78 ; loads and stores should be lowered to copies, so there shouldn't be any
81 ; FUNC-LABEL: {{^}}direct_loop:
85 define void @direct_loop(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
87 %prv_array_const = alloca [2 x i32]
88 %prv_array = alloca [2 x i32]
89 %a = load i32, i32 addrspace(1)* %in
90 %b_src_ptr = getelementptr i32, i32 addrspace(1)* %in, i32 1
91 %b = load i32, i32 addrspace(1)* %b_src_ptr
92 %a_dst_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array_const, i32 0, i32 0
93 store i32 %a, i32* %a_dst_ptr
94 %b_dst_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array_const, i32 0, i32 1
95 store i32 %b, i32* %b_dst_ptr
99 %inc = phi i32 [0, %entry], [%count, %for.body]
100 %x_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array_const, i32 0, i32 0
101 %x = load i32, i32* %x_ptr
102 %y_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array, i32 0, i32 0
103 %y = load i32, i32* %y_ptr
105 store i32 %xy, i32* %y_ptr
106 %count = add i32 %inc, 1
107 %done = icmp eq i32 %count, 4095
108 br i1 %done, label %for.end, label %for.body
111 %value_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array, i32 0, i32 0
112 %value = load i32, i32* %value_ptr
113 store i32 %value, i32 addrspace(1)* %out
117 ; FUNC-LABEL: {{^}}short_array:
121 ; SI-PROMOTE-DAG: buffer_store_short v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ; encoding: [0x00,0x10,0x68,0xe0
122 ; SI-PROMOTE-DAG: buffer_store_short v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen offset:2 ; encoding: [0x02,0x10,0x68,0xe0
123 ; SI-PROMOTE: buffer_load_sshort v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
124 define void @short_array(i32 addrspace(1)* %out, i32 %index) {
126 %0 = alloca [2 x i16]
127 %1 = getelementptr [2 x i16], [2 x i16]* %0, i32 0, i32 0
128 %2 = getelementptr [2 x i16], [2 x i16]* %0, i32 0, i32 1
131 %3 = getelementptr [2 x i16], [2 x i16]* %0, i32 0, i32 %index
132 %4 = load i16, i16* %3
133 %5 = sext i16 %4 to i32
134 store i32 %5, i32 addrspace(1)* %out
138 ; FUNC-LABEL: {{^}}char_array:
142 ; SI-DAG: buffer_store_byte v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ; encoding: [0x00,0x10,0x60,0xe0
143 ; SI-DAG: buffer_store_byte v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen offset:1 ; encoding: [0x01,0x10,0x60,0xe0
144 define void @char_array(i32 addrspace(1)* %out, i32 %index) {
147 %1 = getelementptr [2 x i8], [2 x i8]* %0, i32 0, i32 0
148 %2 = getelementptr [2 x i8], [2 x i8]* %0, i32 0, i32 1
151 %3 = getelementptr [2 x i8], [2 x i8]* %0, i32 0, i32 %index
153 %5 = sext i8 %4 to i32
154 store i32 %5, i32 addrspace(1)* %out
159 ; Make sure we don't overwrite workitem information with private memory
161 ; FUNC-LABEL: {{^}}work_item_info:
163 ; Additional check in case the move ends up in the last slot
164 ; R600-NOT: MOV * TO.X
166 ; SI-NOT: v_mov_b32_e{{(32|64)}} v0
167 define void @work_item_info(i32 addrspace(1)* %out, i32 %in) {
169 %0 = alloca [2 x i32]
170 %1 = getelementptr [2 x i32], [2 x i32]* %0, i32 0, i32 0
171 %2 = getelementptr [2 x i32], [2 x i32]* %0, i32 0, i32 1
174 %3 = getelementptr [2 x i32], [2 x i32]* %0, i32 0, i32 %in
175 %4 = load i32, i32* %3
176 %5 = call i32 @llvm.r600.read.tidig.x()
178 store i32 %6, i32 addrspace(1)* %out
182 ; Test that two stack objects are not stored in the same register
183 ; The second stack object should be in T3.X
184 ; FUNC-LABEL: {{^}}no_overlap:
186 ; R600_CHECK: [[CHAN:[XYZW]]]+
187 ; R600-NOT: [[CHAN]]+
188 ; SI: v_mov_b32_e32 v3
189 define void @no_overlap(i32 addrspace(1)* %out, i32 %in) {
191 %0 = alloca [3 x i8], align 1
192 %1 = alloca [2 x i8], align 1
193 %2 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 0
194 %3 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 1
195 %4 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 2
196 %5 = getelementptr [2 x i8], [2 x i8]* %1, i32 0, i32 0
197 %6 = getelementptr [2 x i8], [2 x i8]* %1, i32 0, i32 1
203 %7 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 %in
204 %8 = getelementptr [2 x i8], [2 x i8]* %1, i32 0, i32 %in
206 %10 = load i8, i8* %8
208 %12 = sext i8 %11 to i32
209 store i32 %12, i32 addrspace(1)* %out
213 define void @char_array_array(i32 addrspace(1)* %out, i32 %index) {
215 %alloca = alloca [2 x [2 x i8]]
216 %gep0 = getelementptr [2 x [2 x i8]], [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 0
217 %gep1 = getelementptr [2 x [2 x i8]], [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 1
218 store i8 0, i8* %gep0
219 store i8 1, i8* %gep1
220 %gep2 = getelementptr [2 x [2 x i8]], [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 %index
221 %load = load i8, i8* %gep2
222 %sext = sext i8 %load to i32
223 store i32 %sext, i32 addrspace(1)* %out
227 define void @i32_array_array(i32 addrspace(1)* %out, i32 %index) {
229 %alloca = alloca [2 x [2 x i32]]
230 %gep0 = getelementptr [2 x [2 x i32]], [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 0
231 %gep1 = getelementptr [2 x [2 x i32]], [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 1
232 store i32 0, i32* %gep0
233 store i32 1, i32* %gep1
234 %gep2 = getelementptr [2 x [2 x i32]], [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 %index
235 %load = load i32, i32* %gep2
236 store i32 %load, i32 addrspace(1)* %out
240 define void @i64_array_array(i64 addrspace(1)* %out, i32 %index) {
242 %alloca = alloca [2 x [2 x i64]]
243 %gep0 = getelementptr [2 x [2 x i64]], [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 0
244 %gep1 = getelementptr [2 x [2 x i64]], [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 1
245 store i64 0, i64* %gep0
246 store i64 1, i64* %gep1
247 %gep2 = getelementptr [2 x [2 x i64]], [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 %index
248 %load = load i64, i64* %gep2
249 store i64 %load, i64 addrspace(1)* %out
253 %struct.pair32 = type { i32, i32 }
255 define void @struct_array_array(i32 addrspace(1)* %out, i32 %index) {
257 %alloca = alloca [2 x [2 x %struct.pair32]]
258 %gep0 = getelementptr [2 x [2 x %struct.pair32]], [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 0, i32 1
259 %gep1 = getelementptr [2 x [2 x %struct.pair32]], [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 1, i32 1
260 store i32 0, i32* %gep0
261 store i32 1, i32* %gep1
262 %gep2 = getelementptr [2 x [2 x %struct.pair32]], [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 %index, i32 0
263 %load = load i32, i32* %gep2
264 store i32 %load, i32 addrspace(1)* %out
268 define void @struct_pair32_array(i32 addrspace(1)* %out, i32 %index) {
270 %alloca = alloca [2 x %struct.pair32]
271 %gep0 = getelementptr [2 x %struct.pair32], [2 x %struct.pair32]* %alloca, i32 0, i32 0, i32 1
272 %gep1 = getelementptr [2 x %struct.pair32], [2 x %struct.pair32]* %alloca, i32 0, i32 1, i32 0
273 store i32 0, i32* %gep0
274 store i32 1, i32* %gep1
275 %gep2 = getelementptr [2 x %struct.pair32], [2 x %struct.pair32]* %alloca, i32 0, i32 %index, i32 0
276 %load = load i32, i32* %gep2
277 store i32 %load, i32 addrspace(1)* %out
281 define void @select_private(i32 addrspace(1)* %out, i32 %in) nounwind {
283 %tmp = alloca [2 x i32]
284 %tmp1 = getelementptr [2 x i32], [2 x i32]* %tmp, i32 0, i32 0
285 %tmp2 = getelementptr [2 x i32], [2 x i32]* %tmp, i32 0, i32 1
286 store i32 0, i32* %tmp1
287 store i32 1, i32* %tmp2
288 %cmp = icmp eq i32 %in, 0
289 %sel = select i1 %cmp, i32* %tmp1, i32* %tmp2
290 %load = load i32, i32* %sel
291 store i32 %load, i32 addrspace(1)* %out
295 ; AMDGPUPromoteAlloca does not know how to handle ptrtoint. When it
296 ; finds one, it should stop trying to promote.
298 ; FUNC-LABEL: ptrtoint:
300 ; SI: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen
301 ; SI: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ;
302 define void @ptrtoint(i32 addrspace(1)* %out, i32 %a, i32 %b) {
303 %alloca = alloca [16 x i32]
304 %tmp0 = getelementptr [16 x i32], [16 x i32]* %alloca, i32 0, i32 %a
305 store i32 5, i32* %tmp0
306 %tmp1 = ptrtoint [16 x i32]* %alloca to i32
307 %tmp2 = add i32 %tmp1, 5
308 %tmp3 = inttoptr i32 %tmp2 to i32*
309 %tmp4 = getelementptr i32, i32* %tmp3, i32 %b
310 %tmp5 = load i32, i32* %tmp4
311 store i32 %tmp5, i32 addrspace(1)* %out