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 -mtriple=amdgcn--amdhsa -mcpu=kaveri < %s | FileCheck %s -check-prefix=SI-PROMOTE -check-prefix=SI -check-prefix=FUNC -check-prefix=HSA-PROMOTE
4 ; 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
5 ; RUN: llc -show-mc-encoding -mattr=-promote-alloca -verify-machineinstrs -mtriple=amdgcn-amdhsa -mcpu=kaveri < %s | FileCheck %s -check-prefix=SI-ALLOCA -check-prefix=SI -check-prefix=FUNC -check-prefix=HSA-ALLOCA
6 ; 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
7 ; 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
9 declare i32 @llvm.r600.read.tidig.x() nounwind readnone
11 ; FUNC-LABEL: {{^}}mova_same_clause:
18 ; HSA-PROMOTE: .amd_kernel_code_t
19 ; HSA-PROMOTE: workgroup_group_segment_byte_size = 5120
20 ; HSA-PROMOTE: .end_amd_kernel_code_t
22 ; SI-PROMOTE: ds_write_b32
23 ; SI-PROMOTE: ds_write_b32
24 ; SI-PROMOTE: ds_read_b32
25 ; SI-PROMOTE: ds_read_b32
27 ; HSA-ALLOCA: .amd_kernel_code_t
28 ; FIXME: Creating the emergency stack slots causes us to over-estimate scratch
30 ; HSA-ALLOCA: workitem_private_segment_byte_size = 24
31 ; HSA-ALLOCA: .end_amd_kernel_code_t
33 ; 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
34 ; 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
35 define void @mova_same_clause(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* nocapture %in) {
37 %stack = alloca [5 x i32], align 4
38 %0 = load i32, i32 addrspace(1)* %in, align 4
39 %arrayidx1 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %0
40 store i32 4, i32* %arrayidx1, align 4
41 %arrayidx2 = getelementptr inbounds i32, i32 addrspace(1)* %in, i32 1
42 %1 = load i32, i32 addrspace(1)* %arrayidx2, align 4
43 %arrayidx3 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 %1
44 store i32 5, i32* %arrayidx3, align 4
45 %arrayidx10 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 0
46 %2 = load i32, i32* %arrayidx10, align 4
47 store i32 %2, i32 addrspace(1)* %out, align 4
48 %arrayidx12 = getelementptr inbounds [5 x i32], [5 x i32]* %stack, i32 0, i32 1
49 %3 = load i32, i32* %arrayidx12
50 %arrayidx13 = getelementptr inbounds i32, i32 addrspace(1)* %out, i32 1
51 store i32 %3, i32 addrspace(1)* %arrayidx13
55 ; This test checks that the stack offset is calculated correctly for structs.
56 ; All register loads/stores should be optimized away, so there shouldn't be
57 ; any MOVA instructions.
59 ; XXX: This generated code has unnecessary MOVs, we should be able to optimize
62 ; FUNC-LABEL: {{^}}multiple_structs:
66 %struct.point = type { i32, i32 }
68 define void @multiple_structs(i32 addrspace(1)* %out) {
70 %a = alloca %struct.point
71 %b = alloca %struct.point
72 %a.x.ptr = getelementptr %struct.point, %struct.point* %a, i32 0, i32 0
73 %a.y.ptr = getelementptr %struct.point, %struct.point* %a, i32 0, i32 1
74 %b.x.ptr = getelementptr %struct.point, %struct.point* %b, i32 0, i32 0
75 %b.y.ptr = getelementptr %struct.point, %struct.point* %b, i32 0, i32 1
76 store i32 0, i32* %a.x.ptr
77 store i32 1, i32* %a.y.ptr
78 store i32 2, i32* %b.x.ptr
79 store i32 3, i32* %b.y.ptr
80 %a.indirect.ptr = getelementptr %struct.point, %struct.point* %a, i32 0, i32 0
81 %b.indirect.ptr = getelementptr %struct.point, %struct.point* %b, i32 0, i32 0
82 %a.indirect = load i32, i32* %a.indirect.ptr
83 %b.indirect = load i32, i32* %b.indirect.ptr
84 %0 = add i32 %a.indirect, %b.indirect
85 store i32 %0, i32 addrspace(1)* %out
89 ; Test direct access of a private array inside a loop. The private array
90 ; loads and stores should be lowered to copies, so there shouldn't be any
93 ; FUNC-LABEL: {{^}}direct_loop:
97 define void @direct_loop(i32 addrspace(1)* %out, i32 addrspace(1)* %in) {
99 %prv_array_const = alloca [2 x i32]
100 %prv_array = alloca [2 x i32]
101 %a = load i32, i32 addrspace(1)* %in
102 %b_src_ptr = getelementptr i32, i32 addrspace(1)* %in, i32 1
103 %b = load i32, i32 addrspace(1)* %b_src_ptr
104 %a_dst_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array_const, i32 0, i32 0
105 store i32 %a, i32* %a_dst_ptr
106 %b_dst_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array_const, i32 0, i32 1
107 store i32 %b, i32* %b_dst_ptr
111 %inc = phi i32 [0, %entry], [%count, %for.body]
112 %x_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array_const, i32 0, i32 0
113 %x = load i32, i32* %x_ptr
114 %y_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array, i32 0, i32 0
115 %y = load i32, i32* %y_ptr
117 store i32 %xy, i32* %y_ptr
118 %count = add i32 %inc, 1
119 %done = icmp eq i32 %count, 4095
120 br i1 %done, label %for.end, label %for.body
123 %value_ptr = getelementptr [2 x i32], [2 x i32]* %prv_array, i32 0, i32 0
124 %value = load i32, i32* %value_ptr
125 store i32 %value, i32 addrspace(1)* %out
129 ; FUNC-LABEL: {{^}}short_array:
133 ; 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
134 ; 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
135 ; SI-PROMOTE: buffer_load_sshort v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
136 define void @short_array(i32 addrspace(1)* %out, i32 %index) {
138 %0 = alloca [2 x i16]
139 %1 = getelementptr [2 x i16], [2 x i16]* %0, i32 0, i32 0
140 %2 = getelementptr [2 x i16], [2 x i16]* %0, i32 0, i32 1
143 %3 = getelementptr [2 x i16], [2 x i16]* %0, i32 0, i32 %index
144 %4 = load i16, i16* %3
145 %5 = sext i16 %4 to i32
146 store i32 %5, i32 addrspace(1)* %out
150 ; FUNC-LABEL: {{^}}char_array:
154 ; 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
155 ; 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
156 define void @char_array(i32 addrspace(1)* %out, i32 %index) {
159 %1 = getelementptr [2 x i8], [2 x i8]* %0, i32 0, i32 0
160 %2 = getelementptr [2 x i8], [2 x i8]* %0, i32 0, i32 1
163 %3 = getelementptr [2 x i8], [2 x i8]* %0, i32 0, i32 %index
165 %5 = sext i8 %4 to i32
166 store i32 %5, i32 addrspace(1)* %out
171 ; Make sure we don't overwrite workitem information with private memory
173 ; FUNC-LABEL: {{^}}work_item_info:
175 ; Additional check in case the move ends up in the last slot
176 ; R600-NOT: MOV * TO.X
178 ; SI-NOT: v_mov_b32_e{{(32|64)}} v0
179 define void @work_item_info(i32 addrspace(1)* %out, i32 %in) {
181 %0 = alloca [2 x i32]
182 %1 = getelementptr [2 x i32], [2 x i32]* %0, i32 0, i32 0
183 %2 = getelementptr [2 x i32], [2 x i32]* %0, i32 0, i32 1
186 %3 = getelementptr [2 x i32], [2 x i32]* %0, i32 0, i32 %in
187 %4 = load i32, i32* %3
188 %5 = call i32 @llvm.r600.read.tidig.x()
190 store i32 %6, i32 addrspace(1)* %out
194 ; Test that two stack objects are not stored in the same register
195 ; The second stack object should be in T3.X
196 ; FUNC-LABEL: {{^}}no_overlap:
198 ; R600_CHECK: [[CHAN:[XYZW]]]+
199 ; R600-NOT: [[CHAN]]+
200 ; SI: v_mov_b32_e32 v3
201 define void @no_overlap(i32 addrspace(1)* %out, i32 %in) {
203 %0 = alloca [3 x i8], align 1
204 %1 = alloca [2 x i8], align 1
205 %2 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 0
206 %3 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 1
207 %4 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 2
208 %5 = getelementptr [2 x i8], [2 x i8]* %1, i32 0, i32 0
209 %6 = getelementptr [2 x i8], [2 x i8]* %1, i32 0, i32 1
215 %7 = getelementptr [3 x i8], [3 x i8]* %0, i32 0, i32 %in
216 %8 = getelementptr [2 x i8], [2 x i8]* %1, i32 0, i32 %in
218 %10 = load i8, i8* %8
220 %12 = sext i8 %11 to i32
221 store i32 %12, i32 addrspace(1)* %out
225 define void @char_array_array(i32 addrspace(1)* %out, i32 %index) {
227 %alloca = alloca [2 x [2 x i8]]
228 %gep0 = getelementptr [2 x [2 x i8]], [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 0
229 %gep1 = getelementptr [2 x [2 x i8]], [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 1
230 store i8 0, i8* %gep0
231 store i8 1, i8* %gep1
232 %gep2 = getelementptr [2 x [2 x i8]], [2 x [2 x i8]]* %alloca, i32 0, i32 0, i32 %index
233 %load = load i8, i8* %gep2
234 %sext = sext i8 %load to i32
235 store i32 %sext, i32 addrspace(1)* %out
239 define void @i32_array_array(i32 addrspace(1)* %out, i32 %index) {
241 %alloca = alloca [2 x [2 x i32]]
242 %gep0 = getelementptr [2 x [2 x i32]], [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 0
243 %gep1 = getelementptr [2 x [2 x i32]], [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 1
244 store i32 0, i32* %gep0
245 store i32 1, i32* %gep1
246 %gep2 = getelementptr [2 x [2 x i32]], [2 x [2 x i32]]* %alloca, i32 0, i32 0, i32 %index
247 %load = load i32, i32* %gep2
248 store i32 %load, i32 addrspace(1)* %out
252 define void @i64_array_array(i64 addrspace(1)* %out, i32 %index) {
254 %alloca = alloca [2 x [2 x i64]]
255 %gep0 = getelementptr [2 x [2 x i64]], [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 0
256 %gep1 = getelementptr [2 x [2 x i64]], [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 1
257 store i64 0, i64* %gep0
258 store i64 1, i64* %gep1
259 %gep2 = getelementptr [2 x [2 x i64]], [2 x [2 x i64]]* %alloca, i32 0, i32 0, i32 %index
260 %load = load i64, i64* %gep2
261 store i64 %load, i64 addrspace(1)* %out
265 %struct.pair32 = type { i32, i32 }
267 define void @struct_array_array(i32 addrspace(1)* %out, i32 %index) {
269 %alloca = alloca [2 x [2 x %struct.pair32]]
270 %gep0 = getelementptr [2 x [2 x %struct.pair32]], [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 0, i32 1
271 %gep1 = getelementptr [2 x [2 x %struct.pair32]], [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 1, i32 1
272 store i32 0, i32* %gep0
273 store i32 1, i32* %gep1
274 %gep2 = getelementptr [2 x [2 x %struct.pair32]], [2 x [2 x %struct.pair32]]* %alloca, i32 0, i32 0, i32 %index, i32 0
275 %load = load i32, i32* %gep2
276 store i32 %load, i32 addrspace(1)* %out
280 define void @struct_pair32_array(i32 addrspace(1)* %out, i32 %index) {
282 %alloca = alloca [2 x %struct.pair32]
283 %gep0 = getelementptr [2 x %struct.pair32], [2 x %struct.pair32]* %alloca, i32 0, i32 0, i32 1
284 %gep1 = getelementptr [2 x %struct.pair32], [2 x %struct.pair32]* %alloca, i32 0, i32 1, i32 0
285 store i32 0, i32* %gep0
286 store i32 1, i32* %gep1
287 %gep2 = getelementptr [2 x %struct.pair32], [2 x %struct.pair32]* %alloca, i32 0, i32 %index, i32 0
288 %load = load i32, i32* %gep2
289 store i32 %load, i32 addrspace(1)* %out
293 define void @select_private(i32 addrspace(1)* %out, i32 %in) nounwind {
295 %tmp = alloca [2 x i32]
296 %tmp1 = getelementptr [2 x i32], [2 x i32]* %tmp, i32 0, i32 0
297 %tmp2 = getelementptr [2 x i32], [2 x i32]* %tmp, i32 0, i32 1
298 store i32 0, i32* %tmp1
299 store i32 1, i32* %tmp2
300 %cmp = icmp eq i32 %in, 0
301 %sel = select i1 %cmp, i32* %tmp1, i32* %tmp2
302 %load = load i32, i32* %sel
303 store i32 %load, i32 addrspace(1)* %out
307 ; AMDGPUPromoteAlloca does not know how to handle ptrtoint. When it
308 ; finds one, it should stop trying to promote.
310 ; FUNC-LABEL: ptrtoint:
312 ; SI: buffer_store_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen
313 ; SI: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offen ;
314 define void @ptrtoint(i32 addrspace(1)* %out, i32 %a, i32 %b) {
315 %alloca = alloca [16 x i32]
316 %tmp0 = getelementptr [16 x i32], [16 x i32]* %alloca, i32 0, i32 %a
317 store i32 5, i32* %tmp0
318 %tmp1 = ptrtoint [16 x i32]* %alloca to i32
319 %tmp2 = add i32 %tmp1, 5
320 %tmp3 = inttoptr i32 %tmp2 to i32*
321 %tmp4 = getelementptr i32, i32* %tmp3, i32 %b
322 %tmp5 = load i32, i32* %tmp4
323 store i32 %tmp5, i32 addrspace(1)* %out