1; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,MOVREL,PREGFX9 %s
2; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,MOVREL,PREGFX9 %s
3; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -amdgpu-vgpr-index-mode -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,IDXMODE,PREGFX9 %s
4; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=gfx900 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,IDXMODE,GFX9 %s
5
6; Tests for indirect addressing on SI, which is implemented using dynamic
7; indexing of vectors.
8
9; GCN-LABEL: {{^}}extract_w_offset:
10; GCN-DAG: s_load_dword [[IN:s[0-9]+]]
11; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, 4.0
12; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, 0x40400000
13; GCN-DAG: v_mov_b32_e32 [[BASEREG:v[0-9]+]], 2.0
14; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, 1.0
15
16; MOVREL-DAG: s_mov_b32 m0, [[IN]]
17; MOVREL: v_movrels_b32_e32 v{{[0-9]+}}, [[BASEREG]]
18
19; IDXMODE: s_set_gpr_idx_on [[IN]], src0{{$}}
20; IDXMODE-NEXT: v_mov_b32_e32 v{{[0-9]+}}, [[BASEREG]]
21; IDXMODE-NEXT: s_set_gpr_idx_off
22define amdgpu_kernel void @extract_w_offset(float addrspace(1)* %out, i32 %in) {
23entry:
24  %idx = add i32 %in, 1
25  %elt = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %idx
26  store float %elt, float addrspace(1)* %out
27  ret void
28}
29
30; XXX: Could do v_or_b32 directly
31; GCN-LABEL: {{^}}extract_w_offset_salu_use_vector:
32; MOVREL: s_mov_b32 m0
33; GCN-DAG: s_or_b32
34; GCN-DAG: s_or_b32
35; GCN-DAG: s_or_b32
36; GCN-DAG: s_or_b32
37; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
38; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
39; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
40; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, s{{[0-9]+}}
41
42; MOVREL: v_movrels_b32_e32
43
44; IDXMODE: s_set_gpr_idx_on s{{[0-9]+}}, src0{{$}}
45; IDXMODE-NEXT: v_mov_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
46; IDXMODE-NEXT: s_set_gpr_idx_off
47define amdgpu_kernel void @extract_w_offset_salu_use_vector(i32 addrspace(1)* %out, i32 %in, <4 x i32> %or.val) {
48entry:
49  %idx = add i32 %in, 1
50  %vec = or <4 x i32> %or.val, <i32 1, i32 2, i32 3, i32 4>
51  %elt = extractelement <4 x i32> %vec, i32 %idx
52  store i32 %elt, i32 addrspace(1)* %out
53  ret void
54}
55
56; GCN-LABEL: {{^}}extract_wo_offset:
57; GCN-DAG: s_load_dword [[IN:s[0-9]+]]
58; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, 4.0
59; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, 0x40400000
60; GCN-DAG: v_mov_b32_e32 v{{[0-9]+}}, 2.0
61; GCN-DAG: v_mov_b32_e32 [[BASEREG:v[0-9]+]], 1.0
62
63; MOVREL-DAG: s_mov_b32 m0, [[IN]]
64; MOVREL: v_movrels_b32_e32 v{{[0-9]+}}, [[BASEREG]]
65
66; IDXMODE: s_set_gpr_idx_on [[IN]], src0{{$}}
67; IDXMODE-NEXT: v_mov_b32_e32 v{{[0-9]+}}, [[BASEREG]]
68; IDXMODE-NEXT: s_set_gpr_idx_off
69define amdgpu_kernel void @extract_wo_offset(float addrspace(1)* %out, i32 %in) {
70entry:
71  %elt = extractelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, i32 %in
72  store float %elt, float addrspace(1)* %out
73  ret void
74}
75
76; GCN-LABEL: {{^}}extract_neg_offset_sgpr:
77; The offset depends on the register that holds the first element of the vector.
78; MOVREL: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
79; MOVREL: v_movrels_b32_e32 v{{[0-9]}}, v0
80
81; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00{{$}}
82; IDXMODE: v_mov_b32_e32 v2, 2
83; IDXMODE: v_mov_b32_e32 v3, 3
84; IDXMODE-NEXT: s_set_gpr_idx_on [[ADD_IDX]], src0{{$}}
85; IDXMODE-NEXT: v_mov_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
86; IDXMODE-NEXT: s_set_gpr_idx_off
87define amdgpu_kernel void @extract_neg_offset_sgpr(i32 addrspace(1)* %out, i32 %offset) {
88entry:
89  %index = add i32 %offset, -512
90  %value = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index
91  store i32 %value, i32 addrspace(1)* %out
92  ret void
93}
94
95; GCN-LABEL: {{^}}extract_neg_offset_sgpr_loaded:
96; The offset depends on the register that holds the first element of the vector.
97; MOVREL: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
98; MOVREL: v_movrels_b32_e32 v{{[0-9]}}, v0
99
100; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00{{$}}
101; IDXMODE: v_mov_b32_e32 v0,
102; IDXMODE: v_mov_b32_e32 v1,
103; IDXMODE: v_mov_b32_e32 v2,
104; IDXMODE: v_mov_b32_e32 v3,
105; IDXMODE-NEXT: s_set_gpr_idx_on [[ADD_IDX]], src0{{$}}
106; IDXMODE-NEXT: v_mov_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
107; IDXMODE-NEXT: s_set_gpr_idx_off
108define amdgpu_kernel void @extract_neg_offset_sgpr_loaded(i32 addrspace(1)* %out, <4 x i32> %vec0, <4 x i32> %vec1, i32 %offset) {
109entry:
110  %index = add i32 %offset, -512
111  %or = or <4 x i32> %vec0, %vec1
112  %value = extractelement <4 x i32> %or, i32 %index
113  store i32 %value, i32 addrspace(1)* %out
114  ret void
115}
116
117; GCN-LABEL: {{^}}extract_neg_offset_vgpr:
118; The offset depends on the register that holds the first element of the vector.
119
120; FIXME: The waitcnt for the argument load can go after the loop
121; GCN: s_mov_b64 s{{\[[0-9]+:[0-9]+\]}}, exec
122; GCN: [[LOOPBB:BB[0-9]+_[0-9]+]]:
123; GCN: v_readfirstlane_b32 [[READLANE:s[0-9]+]], v{{[0-9]+}}
124; GCN: s_and_saveexec_b64 vcc, vcc
125
126; MOVREL: s_add_i32 m0, [[READLANE]], 0xfffffe0
127; MOVREL: v_movrels_b32_e32 [[RESULT:v[0-9]+]], v1
128
129; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00
130; IDXMODE: s_set_gpr_idx_on [[ADD_IDX]], src0
131; IDXMODE: v_mov_b32_e32 [[RESULT:v[0-9]+]], v1
132; IDXMODE: s_set_gpr_idx_off
133
134; GCN: s_cbranch_execnz
135
136; GCN: buffer_store_dword [[RESULT]]
137define amdgpu_kernel void @extract_neg_offset_vgpr(i32 addrspace(1)* %out) {
138entry:
139  %id = call i32 @llvm.amdgcn.workitem.id.x() #1
140  %index = add i32 %id, -512
141  %value = extractelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 %index
142  store i32 %value, i32 addrspace(1)* %out
143  ret void
144}
145
146; GCN-LABEL: {{^}}extract_undef_offset_sgpr:
147; undefined behavior, but shouldn't crash compiler
148define amdgpu_kernel void @extract_undef_offset_sgpr(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
149entry:
150  %ld = load volatile <4 x i32>, <4  x i32> addrspace(1)* %in
151  %value = extractelement <4 x i32> %ld, i32 undef
152  store i32 %value, i32 addrspace(1)* %out
153  ret void
154}
155
156; GCN-LABEL: {{^}}insert_undef_offset_sgpr_vector_src:
157; undefined behavior, but shouldn't crash compiler
158define amdgpu_kernel void @insert_undef_offset_sgpr_vector_src(<4 x i32> addrspace(1)* %out, <4 x i32> addrspace(1)* %in) {
159entry:
160  %ld = load <4 x i32>, <4  x i32> addrspace(1)* %in
161  %value = insertelement <4 x i32> %ld, i32 5, i32 undef
162  store <4 x i32> %value, <4 x i32> addrspace(1)* %out
163  ret void
164}
165
166; GCN-LABEL: {{^}}insert_w_offset:
167; GCN-DAG: s_load_dword [[IN:s[0-9]+]]
168; MOVREL-DAG: s_mov_b32 m0, [[IN]]
169; GCN-DAG: v_mov_b32_e32 v[[ELT0:[0-9]+]], 1.0
170; GCN-DAG: v_mov_b32_e32 v[[ELT1:[0-9]+]], 2.0
171; GCN-DAG: v_mov_b32_e32 v[[ELT2:[0-9]+]], 0x40400000
172; GCN-DAG: v_mov_b32_e32 v[[ELT3:[0-9]+]], 4.0
173; GCN-DAG: v_mov_b32_e32 v[[INS:[0-9]+]], 0x40a00000
174
175; MOVREL: v_movreld_b32_e32 v[[ELT1]], v[[INS]]
176; MOVREL: buffer_store_dwordx4 v{{\[}}[[ELT0]]:[[ELT3]]{{\]}}
177define amdgpu_kernel void @insert_w_offset(<4 x float> addrspace(1)* %out, i32 %in) {
178entry:
179  %0 = add i32 %in, 1
180  %1 = insertelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, float 5.0, i32 %0
181  store <4 x float> %1, <4 x float> addrspace(1)* %out
182  ret void
183}
184
185; GCN-LABEL: {{^}}insert_wo_offset:
186; GCN: s_load_dword [[IN:s[0-9]+]]
187
188; MOVREL: s_mov_b32 m0, [[IN]]
189; MOVREL: v_movreld_b32_e32 v[[ELT0:[0-9]+]]
190
191; IDXMODE: s_set_gpr_idx_on [[IN]], dst
192; IDXMODE-NEXT: v_mov_b32_e32 v[[ELT0:[0-9]+]], v{{[0-9]+}}
193; IDXMODE-NEXT: s_set_gpr_idx_off
194
195; GCN: buffer_store_dwordx4 v{{\[}}[[ELT0]]:
196define amdgpu_kernel void @insert_wo_offset(<4 x float> addrspace(1)* %out, i32 %in) {
197entry:
198  %0 = insertelement <4 x float> <float 1.0, float 2.0, float 3.0, float 4.0>, float 5.0, i32 %in
199  store <4 x float> %0, <4 x float> addrspace(1)* %out
200  ret void
201}
202
203; GCN-LABEL: {{^}}insert_neg_offset_sgpr:
204; The offset depends on the register that holds the first element of the vector.
205; MOVREL: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
206; MOVREL: v_movreld_b32_e32 v0, 5
207
208; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00{{$}}
209; IDXMODE: s_set_gpr_idx_on [[ADD_IDX]], dst
210; IDXMODE-NEXT: v_mov_b32_e32 v0, 5
211; IDXMODE-NEXT: s_set_gpr_idx_off
212define amdgpu_kernel void @insert_neg_offset_sgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out, i32 %offset) {
213entry:
214  %index = add i32 %offset, -512
215  %value = insertelement <4 x i32> <i32 0, i32 1, i32 2, i32 3>, i32 5, i32 %index
216  store <4 x i32> %value, <4 x i32> addrspace(1)* %out
217  ret void
218}
219
220; The vector indexed into is originally loaded into an SGPR rather
221; than built with a reg_sequence
222
223; GCN-LABEL: {{^}}insert_neg_offset_sgpr_loadreg:
224; The offset depends on the register that holds the first element of the vector.
225; MOVREL: s_add_i32 m0, s{{[0-9]+}}, 0xfffffe{{[0-9a-z]+}}
226; MOVREL: v_movreld_b32_e32 v0, 5
227
228; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00{{$}}
229; IDXMODE: s_set_gpr_idx_on [[ADD_IDX]], dst
230; IDXMODE-NEXT: v_mov_b32_e32 v0, 5
231; IDXMODE-NEXT: s_set_gpr_idx_off
232define amdgpu_kernel void @insert_neg_offset_sgpr_loadreg(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out, <4 x i32> %vec, i32 %offset) {
233entry:
234  %index = add i32 %offset, -512
235  %value = insertelement <4 x i32> %vec, i32 5, i32 %index
236  store <4 x i32> %value, <4 x i32> addrspace(1)* %out
237  ret void
238}
239
240; GCN-LABEL: {{^}}insert_neg_offset_vgpr:
241; The offset depends on the register that holds the first element of the vector.
242
243; GCN-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], 1{{$}}
244; GCN-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], 2{{$}}
245; GCN-DAG: v_mov_b32_e32 [[VEC_ELT2:v[0-9]+]], 3{{$}}
246; GCN-DAG: v_mov_b32_e32 [[VEC_ELT3:v[0-9]+]], 4{{$}}
247
248; GCN: s_mov_b64 [[SAVEEXEC:s\[[0-9]+:[0-9]+\]]], exec
249; GCN: [[LOOPBB:BB[0-9]+_[0-9]+]]:
250; GCN: v_readfirstlane_b32 [[READLANE:s[0-9]+]]
251; GCN: s_and_saveexec_b64 vcc, vcc
252
253; MOVREL: s_add_i32 m0, [[READLANE]], 0xfffffe00
254; MOVREL: v_movreld_b32_e32 [[VEC_ELT0]], 5
255
256; IDXMODE: s_addk_i32 [[ADD_IDX:s[0-9]+]], 0xfe00{{$}}
257; IDXMODE: s_set_gpr_idx_on [[ADD_IDX]], dst
258; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, 5
259; IDXMODE: s_set_gpr_idx_off
260
261; GCN: s_cbranch_execnz [[LOOPBB]]
262; GCN: s_mov_b64 exec, [[SAVEEXEC]]
263
264; GCN: buffer_store_dword
265define amdgpu_kernel void @insert_neg_offset_vgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out) {
266entry:
267  %id = call i32 @llvm.amdgcn.workitem.id.x() #1
268  %index = add i32 %id, -512
269  %value = insertelement <4 x i32> <i32 1, i32 2, i32 3, i32 4>, i32 5, i32 %index
270  store <4 x i32> %value, <4 x i32> addrspace(1)* %out
271  ret void
272}
273
274; GCN-LABEL: {{^}}insert_neg_inline_offset_vgpr:
275
276; GCN-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], 1{{$}}
277; GCN-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], 2{{$}}
278; GCN-DAG: v_mov_b32_e32 [[VEC_ELT2:v[0-9]+]], 3{{$}}
279; GCN-DAG: v_mov_b32_e32 [[VEC_ELT3:v[0-9]+]], 4{{$}}
280; GCN-DAG: v_mov_b32_e32 [[VAL:v[0-9]+]], 0x1f4{{$}}
281
282; GCN: s_mov_b64 [[SAVEEXEC:s\[[0-9]+:[0-9]+\]]], exec
283
284; The offset depends on the register that holds the first element of the vector.
285; GCN: v_readfirstlane_b32 [[READLANE:s[0-9]+]]
286
287; MOVREL: s_add_i32 m0, [[READLANE]], -16
288; MOVREL: v_movreld_b32_e32 [[VEC_ELT0]], [[VAL]]
289
290; IDXMODE: s_add_i32 [[ADD_IDX:s[0-9]+]], [[READLANE]], -16
291; IDXMODE: s_set_gpr_idx_on [[ADD_IDX]], dst
292; IDXMODE: v_mov_b32_e32 [[VEC_ELT0]], [[VAL]]
293; IDXMODE: s_set_gpr_idx_off
294
295; GCN: s_cbranch_execnz
296define amdgpu_kernel void @insert_neg_inline_offset_vgpr(i32 addrspace(1)* %in, <4 x i32> addrspace(1)* %out) {
297entry:
298  %id = call i32 @llvm.amdgcn.workitem.id.x() #1
299  %index = add i32 %id, -16
300  %value = insertelement <4 x i32> <i32 1, i32 2, i32 3, i32 4>, i32 500, i32 %index
301  store <4 x i32> %value, <4 x i32> addrspace(1)* %out
302  ret void
303}
304
305; When the block is split to insert the loop, make sure any other
306; places that need to be expanded in the same block are also handled.
307
308; GCN-LABEL: {{^}}extract_vgpr_offset_multiple_in_block:
309
310; FIXME: Why is vector copied in between?
311
312; GCN-DAG: {{buffer|flat|global}}_load_dword [[IDX0:v[0-9]+]]
313; GCN-DAG: s_mov_b32 [[S_ELT1:s[0-9]+]], 9
314; GCN-DAG: s_mov_b32 [[S_ELT0:s[0-9]+]], 7
315; GCN-DAG: v_mov_b32_e32 [[VEC_ELT0:v[0-9]+]], [[S_ELT0]]
316; GCN-DAG: v_mov_b32_e32 [[VEC_ELT1:v[0-9]+]], [[S_ELT1]]
317
318; GCN: s_mov_b64 [[MASK:s\[[0-9]+:[0-9]+\]]], exec
319
320; GCN: [[LOOP0:BB[0-9]+_[0-9]+]]:
321; GCN-NEXT: s_waitcnt vmcnt(0)
322; GCN-NEXT: v_readfirstlane_b32 [[READLANE:s[0-9]+]], [[IDX0]]
323; GCN: v_cmp_eq_u32_e32 vcc, [[READLANE]], [[IDX0]]
324; GCN: s_and_saveexec_b64 vcc, vcc
325
326; MOVREL: s_mov_b32 m0, [[READLANE]]
327; MOVREL: v_movrels_b32_e32 [[MOVREL0:v[0-9]+]], [[VEC_ELT0]]
328
329; IDXMODE: s_set_gpr_idx_on [[READLANE]], src0
330; IDXMODE: v_mov_b32_e32 [[MOVREL0:v[0-9]+]], [[VEC_ELT0]]
331; IDXMODE: s_set_gpr_idx_off
332
333; GCN-NEXT: s_xor_b64 exec, exec, vcc
334; GCN-NEXT: s_cbranch_execnz [[LOOP0]]
335
336; FIXME: Redundant copy
337; GCN: s_mov_b64 exec, [[MASK]]
338
339; GCN: v_mov_b32_e32 [[VEC_ELT1_2:v[0-9]+]], [[S_ELT1]]
340
341; GCN: s_mov_b64 [[MASK2:s\[[0-9]+:[0-9]+\]]], exec
342
343; GCN: [[LOOP1:BB[0-9]+_[0-9]+]]:
344; GCN-NEXT: v_readfirstlane_b32 [[READLANE:s[0-9]+]], [[IDX0]]
345; GCN: v_cmp_eq_u32_e32 vcc, [[READLANE]], [[IDX0]]
346; GCN: s_and_saveexec_b64 vcc, vcc
347
348; MOVREL: s_mov_b32 m0, [[READLANE]]
349; MOVREL-NEXT: v_movrels_b32_e32 [[MOVREL1:v[0-9]+]], [[VEC_ELT1_2]]
350
351; IDXMODE: s_set_gpr_idx_on [[READLANE]], src0
352; IDXMODE-NEXT: v_mov_b32_e32 [[MOVREL1:v[0-9]+]], [[VEC_ELT1_2]]
353; IDXMODE: s_set_gpr_idx_off
354
355; GCN-NEXT: s_xor_b64 exec, exec, vcc
356; GCN: s_cbranch_execnz [[LOOP1]]
357
358; GCN: buffer_store_dword [[MOVREL0]]
359; GCN: buffer_store_dword [[MOVREL1]]
360define amdgpu_kernel void @extract_vgpr_offset_multiple_in_block(i32 addrspace(1)* %out0, i32 addrspace(1)* %out1, i32 addrspace(1)* %in) #0 {
361entry:
362  %id = call i32 @llvm.amdgcn.workitem.id.x() #1
363  %id.ext = zext i32 %id to i64
364  %gep = getelementptr inbounds i32, i32 addrspace(1)* %in, i64 %id.ext
365  %idx0 = load volatile i32, i32 addrspace(1)* %gep
366  %idx1 = add i32 %idx0, 1
367  %val0 = extractelement <4 x i32> <i32 7, i32 9, i32 11, i32 13>, i32 %idx0
368  %live.out.reg = call i32 asm sideeffect "s_mov_b32 $0, 17", "={s4}" ()
369  %val1 = extractelement <4 x i32> <i32 7, i32 9, i32 11, i32 13>, i32 %idx1
370  store volatile i32 %val0, i32 addrspace(1)* %out0
371  store volatile i32 %val1, i32 addrspace(1)* %out0
372  %cmp = icmp eq i32 %id, 0
373  br i1 %cmp, label %bb1, label %bb2
374
375bb1:
376  store volatile i32 %live.out.reg, i32 addrspace(1)* undef
377  br label %bb2
378
379bb2:
380  ret void
381}
382
383; GCN-LABEL: {{^}}insert_vgpr_offset_multiple_in_block:
384; GCN-DAG: s_load_dwordx4 s{{\[}}[[S_ELT0:[0-9]+]]:[[S_ELT3:[0-9]+]]{{\]}}
385; GCN-DAG: {{buffer|flat|global}}_load_dword [[IDX0:v[0-9]+]]
386; GCN-DAG: v_mov_b32 [[INS0:v[0-9]+]], 62
387
388; GCN-DAG: v_mov_b32_e32 v[[VEC_ELT3:[0-9]+]], s[[S_ELT3]]
389; GCN: v_mov_b32_e32 v[[VEC_ELT2:[0-9]+]], s{{[0-9]+}}
390; GCN: v_mov_b32_e32 v[[VEC_ELT1:[0-9]+]], s{{[0-9]+}}
391; GCN: v_mov_b32_e32 v[[VEC_ELT0:[0-9]+]], s[[S_ELT0]]
392
393; GCN: [[LOOP0:BB[0-9]+_[0-9]+]]:
394; GCN-NEXT: s_waitcnt vmcnt(0)
395; GCN-NEXT: v_readfirstlane_b32 [[READLANE:s[0-9]+]], [[IDX0]]
396; GCN: v_cmp_eq_u32_e32 vcc, [[READLANE]], [[IDX0]]
397; GCN: s_and_saveexec_b64 vcc, vcc
398
399; MOVREL: s_mov_b32 m0, [[READLANE]]
400; MOVREL-NEXT: v_movreld_b32_e32 v[[VEC_ELT0]], [[INS0]]
401
402; IDXMODE: s_set_gpr_idx_on [[READLANE]], dst
403; IDXMODE-NEXT: v_mov_b32_e32 v[[VEC_ELT0]], [[INS0]]
404; IDXMODE: s_set_gpr_idx_off
405
406; GCN-NEXT: s_xor_b64 exec, exec, vcc
407; GCN: s_cbranch_execnz [[LOOP0]]
408
409; FIXME: Redundant copy
410; GCN: s_mov_b64 exec, [[MASK:s\[[0-9]+:[0-9]+\]]]
411
412; GCN: s_mov_b64 [[MASK]], exec
413
414; GCN: [[LOOP1:BB[0-9]+_[0-9]+]]:
415; GCN-NEXT: v_readfirstlane_b32 [[READLANE:s[0-9]+]], [[IDX0]]
416; GCN: v_cmp_eq_u32_e32 vcc, [[READLANE]], [[IDX0]]
417; GCN: s_and_saveexec_b64 vcc, vcc
418
419; MOVREL: s_mov_b32 m0, [[READLANE]]
420; MOVREL-NEXT: v_movreld_b32_e32 v[[VEC_ELT1]], 63
421
422; IDXMODE: s_set_gpr_idx_on [[READLANE]], dst
423; IDXMODE-NEXT: v_mov_b32_e32 v[[VEC_ELT1]], 63
424; IDXMODE: s_set_gpr_idx_off
425
426; GCN-NEXT: s_xor_b64 exec, exec, vcc
427; GCN: s_cbranch_execnz [[LOOP1]]
428
429; GCN: buffer_store_dwordx4 v{{\[}}[[VEC_ELT0]]:
430
431; GCN: buffer_store_dword [[INS0]]
432define amdgpu_kernel void @insert_vgpr_offset_multiple_in_block(<4 x i32> addrspace(1)* %out0, <4 x i32> addrspace(1)* %out1, i32 addrspace(1)* %in, <4 x i32> %vec0) #0 {
433entry:
434  %id = call i32 @llvm.amdgcn.workitem.id.x() #1
435  %id.ext = zext i32 %id to i64
436  %gep = getelementptr inbounds i32, i32 addrspace(1)* %in, i64 %id.ext
437  %idx0 = load volatile i32, i32 addrspace(1)* %gep
438  %idx1 = add i32 %idx0, 1
439  %live.out.val = call i32 asm sideeffect "v_mov_b32 $0, 62", "=v"()
440  %vec1 = insertelement <4 x i32> %vec0, i32 %live.out.val, i32 %idx0
441  %vec2 = insertelement <4 x i32> %vec1, i32 63, i32 %idx1
442  store volatile <4 x i32> %vec2, <4 x i32> addrspace(1)* %out0
443  %cmp = icmp eq i32 %id, 0
444  br i1 %cmp, label %bb1, label %bb2
445
446bb1:
447  store volatile i32 %live.out.val, i32 addrspace(1)* undef
448  br label %bb2
449
450bb2:
451  ret void
452}
453
454
455; GCN-LABEL: {{^}}insert_adjacent_blocks:
456define amdgpu_kernel void @insert_adjacent_blocks(i32 %arg, float %val0) #0 {
457bb:
458  %tmp = icmp eq i32 %arg, 0
459  br i1 %tmp, label %bb1, label %bb4
460
461bb1:                                              ; preds = %bb
462  %tmp2 = load volatile <4 x float>, <4 x float> addrspace(1)* undef
463  %tmp3 = insertelement <4 x float> %tmp2, float %val0, i32 undef
464  call void asm sideeffect "; reg use $0", "v"(<4 x float> %tmp3) #0 ; Prevent block optimize out
465  br label %bb7
466
467bb4:                                              ; preds = %bb
468  %tmp5 = load volatile <4 x float>, <4 x float> addrspace(1)* undef
469  %tmp6 = insertelement <4 x float> %tmp5, float %val0, i32 undef
470  call void asm sideeffect "; reg use $0", "v"(<4 x float> %tmp6) #0 ; Prevent block optimize out
471  br label %bb7
472
473bb7:                                              ; preds = %bb4, %bb1
474  %tmp8 = phi <4 x float> [ %tmp3, %bb1 ], [ %tmp6, %bb4 ]
475  store volatile <4 x float> %tmp8, <4 x float> addrspace(1)* undef
476  ret void
477}
478
479; FIXME: Should be able to fold zero input to movreld to inline imm?
480
481; GCN-LABEL: {{^}}multi_same_block:
482
483; GCN: s_load_dword [[ARG:s[0-9]+]]
484
485; MOVREL: v_mov_b32_e32 v{{[0-9]+}}, 0x41900000
486; MOVREL: v_mov_b32_e32 v{{[0-9]+}}, 0x41b0cccd
487; MOVREL: s_waitcnt
488; MOVREL: s_add_i32 m0, [[ARG]], -16
489; MOVREL: v_movreld_b32_e32 v{{[0-9]+}}, 4.0
490; MOVREL: v_movreld_b32_e32 v{{[0-9]+}}, -4.0
491; MOVREL: s_mov_b32 m0, -1
492
493
494; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, 0x41900000
495; IDXMODE: s_waitcnt
496; IDXMODE: s_add_i32 [[ARG]], [[ARG]], -16
497; IDXMODE: s_set_gpr_idx_on [[ARG]], dst
498; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, 4.0
499; IDXMODE: s_set_gpr_idx_off
500; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, 0x41b0cccd
501; IDXMODE: s_set_gpr_idx_on [[ARG]], dst
502; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, -4.0
503; IDXMODE: s_set_gpr_idx_off
504
505; GCN: ds_write_b32
506; GCN: ds_write_b32
507; GCN: s_endpgm
508define amdgpu_kernel void @multi_same_block(i32 %arg) #0 {
509bb:
510  %tmp1 = add i32 %arg, -16
511  %tmp2 = insertelement <6 x float> <float 1.700000e+01, float 1.800000e+01, float 1.900000e+01, float 2.000000e+01, float 2.100000e+01, float 2.200000e+01>, float 4.000000e+00, i32 %tmp1
512  %tmp3 = add i32 %arg, -16
513  %tmp4 = insertelement <6 x float> <float 0x40311999A0000000, float 0x40321999A0000000, float 0x40331999A0000000, float 0x40341999A0000000, float 0x40351999A0000000, float 0x40361999A0000000>, float -4.0, i32 %tmp3
514  %tmp5 = bitcast <6 x float> %tmp2 to <6 x i32>
515  %tmp6 = extractelement <6 x i32> %tmp5, i32 1
516  %tmp7 = bitcast <6 x float> %tmp4 to <6 x i32>
517  %tmp8 = extractelement <6 x i32> %tmp7, i32 5
518  store volatile i32 %tmp6, i32 addrspace(3)* undef, align 4
519  store volatile i32 %tmp8, i32 addrspace(3)* undef, align 4
520  ret void
521}
522
523; offset puts outside of superegister bounaries, so clamp to 1st element.
524; GCN-LABEL: {{^}}extract_largest_inbounds_offset:
525; GCN-DAG: buffer_load_dwordx4 v{{\[}}[[LO_ELT:[0-9]+]]:[[HI_ELT:[0-9]+]]{{\]}}
526; GCN-DAG: s_load_dword [[IDX:s[0-9]+]]
527; MOVREL: s_mov_b32 m0, [[IDX]]
528; MOVREL: v_movrels_b32_e32 [[EXTRACT:v[0-9]+]], v[[HI_ELT]]
529
530; IDXMODE: s_set_gpr_idx_on [[IDX]], src0
531; IDXMODE: v_mov_b32_e32 [[EXTRACT:v[0-9]+]], v[[HI_ELT]]
532; IDXMODE: s_set_gpr_idx_off
533
534; GCN: buffer_store_dword [[EXTRACT]]
535define amdgpu_kernel void @extract_largest_inbounds_offset(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %idx) {
536entry:
537  %ld = load volatile <4 x i32>, <4  x i32> addrspace(1)* %in
538  %offset = add i32 %idx, 3
539  %value = extractelement <4 x i32> %ld, i32 %offset
540  store i32 %value, i32 addrspace(1)* %out
541  ret void
542}
543
544; GCN-LABEL: {{^}}extract_out_of_bounds_offset:
545; GCN-DAG: buffer_load_dwordx4 v{{\[}}[[LO_ELT:[0-9]+]]:[[HI_ELT:[0-9]+]]{{\]}}
546; GCN-DAG: s_load_dword [[IDX:s[0-9]+]]
547; MOVREL: s_add_i32 m0, [[IDX]], 4
548; MOVREL: v_movrels_b32_e32 [[EXTRACT:v[0-9]+]], v[[LO_ELT]]
549
550; IDXMODE: s_add_i32 [[ADD_IDX:s[0-9]+]], [[IDX]], 4
551; IDXMODE: s_set_gpr_idx_on [[ADD_IDX]], src0
552; IDXMODE: v_mov_b32_e32 [[EXTRACT:v[0-9]+]], v[[LO_ELT]]
553; IDXMODE: s_set_gpr_idx_off
554
555; GCN: buffer_store_dword [[EXTRACT]]
556define amdgpu_kernel void @extract_out_of_bounds_offset(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %idx) {
557entry:
558  %ld = load volatile <4 x i32>, <4  x i32> addrspace(1)* %in
559  %offset = add i32 %idx, 4
560  %value = extractelement <4 x i32> %ld, i32 %offset
561  store i32 %value, i32 addrspace(1)* %out
562  ret void
563}
564
565; Test that the or is folded into the base address register instead of
566; added to m0
567
568; GCN-LABEL: {{^}}extractelement_v4i32_or_index:
569; GCN: s_load_dword [[IDX_IN:s[0-9]+]]
570; GCN: s_lshl_b32 [[IDX_SHL:s[0-9]+]], [[IDX_IN]]
571; GCN-NOT: [[IDX_SHL]]
572
573; MOVREL: s_mov_b32 m0, [[IDX_SHL]]
574; MOVREL: v_movrels_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
575
576; IDXMODE: s_set_gpr_idx_on [[IDX_SHL]], src0
577; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
578; IDXMODE: s_set_gpr_idx_off
579define amdgpu_kernel void @extractelement_v4i32_or_index(i32 addrspace(1)* %out, <4 x i32> addrspace(1)* %in, i32 %idx.in) {
580entry:
581  %ld = load volatile <4 x i32>, <4  x i32> addrspace(1)* %in
582  %idx.shl = shl i32 %idx.in, 2
583  %idx = or i32 %idx.shl, 1
584  %value = extractelement <4 x i32> %ld, i32 %idx
585  store i32 %value, i32 addrspace(1)* %out
586  ret void
587}
588
589; GCN-LABEL: {{^}}insertelement_v4f32_or_index:
590; GCN: s_load_dword [[IDX_IN:s[0-9]+]]
591; GCN: s_lshl_b32 [[IDX_SHL:s[0-9]+]], [[IDX_IN]]
592; GCN-NOT: [[IDX_SHL]]
593
594; MOVREL: s_mov_b32 m0, [[IDX_SHL]]
595; MOVREL: v_movreld_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
596
597; IDXMODE: s_set_gpr_idx_on [[IDX_SHL]], dst
598; IDXMODE: v_mov_b32_e32 v{{[0-9]+}}, v{{[0-9]+}}
599; IDXMODE: s_set_gpr_idx_off
600define amdgpu_kernel void @insertelement_v4f32_or_index(<4 x float> addrspace(1)* %out, <4 x float> %a, i32 %idx.in) nounwind {
601  %idx.shl = shl i32 %idx.in, 2
602  %idx = or i32 %idx.shl, 1
603  %vecins = insertelement <4 x float> %a, float 5.000000e+00, i32 %idx
604  store <4 x float> %vecins, <4 x float> addrspace(1)* %out, align 16
605  ret void
606}
607
608; GCN-LABEL: {{^}}broken_phi_bb:
609; GCN: v_mov_b32_e32 [[PHIREG:v[0-9]+]], 8
610
611; GCN: s_branch [[BB2:BB[0-9]+_[0-9]+]]
612
613; GCN: {{^BB[0-9]+_[0-9]+}}:
614; GCN: s_mov_b64 exec,
615
616; GCN: [[BB2]]:
617; GCN: v_cmp_le_i32_e32 vcc, s{{[0-9]+}}, [[PHIREG]]
618; GCN: buffer_load_dword
619
620; GCN: [[REGLOOP:BB[0-9]+_[0-9]+]]:
621; MOVREL: v_movreld_b32_e32
622
623; IDXMODE: s_set_gpr_idx_on
624; IDXMODE: v_mov_b32_e32
625; IDXMODE: s_set_gpr_idx_off
626
627; GCN: s_cbranch_execnz [[REGLOOP]]
628define amdgpu_kernel void @broken_phi_bb(i32 %arg, i32 %arg1) #0 {
629bb:
630  br label %bb2
631
632bb2:                                              ; preds = %bb4, %bb
633  %tmp = phi i32 [ 8, %bb ], [ %tmp7, %bb4 ]
634  %tmp3 = icmp slt i32 %tmp, %arg
635  br i1 %tmp3, label %bb4, label %bb8
636
637bb4:                                              ; preds = %bb2
638  %vgpr = load volatile i32, i32 addrspace(1)* undef
639  %tmp5 = insertelement <8 x i32> undef, i32 undef, i32 %vgpr
640  %tmp6 = insertelement <8 x i32> %tmp5, i32 %arg1, i32 %vgpr
641  %tmp7 = extractelement <8 x i32> %tmp6, i32 0
642  br label %bb2
643
644bb8:                                              ; preds = %bb2
645  ret void
646}
647
648declare i32 @llvm.amdgcn.workitem.id.x() #1
649declare void @llvm.amdgcn.s.barrier() #2
650
651attributes #0 = { nounwind }
652attributes #1 = { nounwind readnone }
653attributes #2 = { nounwind convergent }
654