1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 --amdgpu-enable-flat-scratch < %s | FileCheck -check-prefixes=GCN %s
3
4declare void @extern_func() #0
5
6define amdgpu_kernel void @stack_object_addrspacecast_in_kernel_no_calls() {
7; GCN-LABEL: stack_object_addrspacecast_in_kernel_no_calls:
8; GCN:       ; %bb.0:
9; GCN-NEXT:    s_add_u32 s0, s0, s3
10; GCN-NEXT:    s_addc_u32 s1, s1, 0
11; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
12; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
13; GCN-NEXT:    v_mov_b32_e32 v0, 4
14; GCN-NEXT:    s_getreg_b32 s0, hwreg(HW_REG_SH_MEM_BASES, 0, 16)
15; GCN-NEXT:    v_mov_b32_e32 v2, 0
16; GCN-NEXT:    s_lshl_b32 s0, s0, 16
17; GCN-NEXT:    v_cmp_ne_u32_e32 vcc_lo, -1, v0
18; GCN-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc_lo
19; GCN-NEXT:    v_cndmask_b32_e64 v1, 0, s0, vcc_lo
20; GCN-NEXT:    flat_store_dword v[0:1], v2
21; GCN-NEXT:    s_waitcnt_vscnt null, 0x0
22; GCN-NEXT:    s_endpgm
23  %alloca = alloca i32, addrspace(5)
24  %cast = addrspacecast i32 addrspace(5)* %alloca to i32*
25  store volatile i32 0, i32* %cast
26  ret void
27}
28
29define amdgpu_kernel void @stack_object_in_kernel_no_calls() {
30; GCN-LABEL: stack_object_in_kernel_no_calls:
31; GCN:       ; %bb.0:
32; GCN-NEXT:    s_add_u32 s0, s0, s3
33; GCN-NEXT:    s_addc_u32 s1, s1, 0
34; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
35; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
36; GCN-NEXT:    v_mov_b32_e32 v0, 0
37; GCN-NEXT:    s_mov_b32 vcc_lo, 0
38; GCN-NEXT:    scratch_store_dword off, v0, vcc_lo offset:4
39; GCN-NEXT:    s_waitcnt_vscnt null, 0x0
40; GCN-NEXT:    s_endpgm
41  %alloca = alloca i32, addrspace(5)
42  store volatile i32 0, i32 addrspace(5)* %alloca
43  ret void
44}
45
46define amdgpu_kernel void @kernel_calls_no_stack() {
47; GCN-LABEL: kernel_calls_no_stack:
48; GCN:       ; %bb.0:
49; GCN-NEXT:    s_add_u32 s6, s6, s11
50; GCN-NEXT:    s_mov_b32 s32, 0
51; GCN-NEXT:    s_addc_u32 s7, s7, 0
52; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s6
53; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s7
54; GCN-NEXT:    s_getpc_b64 s[0:1]
55; GCN-NEXT:    s_add_u32 s0, s0, extern_func@gotpcrel32@lo+4
56; GCN-NEXT:    s_addc_u32 s1, s1, extern_func@gotpcrel32@hi+12
57; GCN-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
58; GCN-NEXT:    s_waitcnt lgkmcnt(0)
59; GCN-NEXT:    s_swappc_b64 s[30:31], s[0:1]
60; GCN-NEXT:    s_endpgm
61  call void @extern_func()
62  ret void
63}
64
65define amdgpu_kernel void @test(i32 addrspace(1)* %out, i32 %in) {
66; GCN-LABEL: test:
67; GCN:       ; %bb.0:
68; GCN-NEXT:    s_add_u32 s2, s2, s5
69; GCN-NEXT:    s_addc_u32 s3, s3, 0
70; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s2
71; GCN-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s3
72; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
73; GCN-NEXT:    s_mov_b32 s104, exec_lo
74; GCN-NEXT:    s_mov_b32 exec_lo, 3
75; GCN-NEXT:    s_mov_b32 s105, 0
76; GCN-NEXT:    scratch_store_dword off, v72, s105
77; GCN-NEXT:    s_waitcnt lgkmcnt(0)
78; GCN-NEXT:    v_writelane_b32 v72, s2, 0
79; GCN-NEXT:    s_mov_b32 s105, 4
80; GCN-NEXT:    v_writelane_b32 v72, s3, 1
81; GCN-NEXT:    scratch_store_dword off, v72, s105 ; 4-byte Folded Spill
82; GCN-NEXT:    s_waitcnt_depctr 0xffe3
83; GCN-NEXT:    s_mov_b32 s105, 0
84; GCN-NEXT:    scratch_load_dword v72, off, s105
85; GCN-NEXT:    s_waitcnt vmcnt(0)
86; GCN-NEXT:    s_waitcnt_depctr 0xffe3
87; GCN-NEXT:    s_mov_b32 exec_lo, s104
88; GCN-NEXT:    s_load_dword vcc_lo, s[0:1], 0x8
89; GCN-NEXT:    ; kill: killed $sgpr0_sgpr1
90; GCN-NEXT:    ;;#ASMSTART
91; GCN-NEXT:    ;;#ASMEND
92; GCN-NEXT:    ;;#ASMSTART
93; GCN-NEXT:    ;;#ASMEND
94; GCN-NEXT:    ;;#ASMSTART
95; GCN-NEXT:    ;;#ASMEND
96; GCN-NEXT:    ;;#ASMSTART
97; GCN-NEXT:    ;;#ASMEND
98; GCN-NEXT:    ;;#ASMSTART
99; GCN-NEXT:    ;;#ASMEND
100; GCN-NEXT:    ;;#ASMSTART
101; GCN-NEXT:    ;;#ASMEND
102; GCN-NEXT:    ;;#ASMSTART
103; GCN-NEXT:    ;;#ASMEND
104; GCN-NEXT:    ;;#ASMSTART
105; GCN-NEXT:    ;;#ASMEND
106; GCN-NEXT:    ;;#ASMSTART
107; GCN-NEXT:    ;;#ASMEND
108; GCN-NEXT:    ;;#ASMSTART
109; GCN-NEXT:    ;;#ASMEND
110; GCN-NEXT:    ;;#ASMSTART
111; GCN-NEXT:    ;;#ASMEND
112; GCN-NEXT:    ;;#ASMSTART
113; GCN-NEXT:    ;;#ASMEND
114; GCN-NEXT:    ;;#ASMSTART
115; GCN-NEXT:    ;;#ASMEND
116; GCN-NEXT:    ;;#ASMSTART
117; GCN-NEXT:    ;;#ASMEND
118; GCN-NEXT:    ;;#ASMSTART
119; GCN-NEXT:    ;;#ASMEND
120; GCN-NEXT:    ;;#ASMSTART
121; GCN-NEXT:    ;;#ASMEND
122; GCN-NEXT:    ;;#ASMSTART
123; GCN-NEXT:    ;;#ASMEND
124; GCN-NEXT:    ;;#ASMSTART
125; GCN-NEXT:    ;;#ASMEND
126; GCN-NEXT:    ;;#ASMSTART
127; GCN-NEXT:    ;;#ASMEND
128; GCN-NEXT:    ;;#ASMSTART
129; GCN-NEXT:    ;;#ASMEND
130; GCN-NEXT:    ;;#ASMSTART
131; GCN-NEXT:    ;;#ASMEND
132; GCN-NEXT:    ;;#ASMSTART
133; GCN-NEXT:    ;;#ASMEND
134; GCN-NEXT:    ;;#ASMSTART
135; GCN-NEXT:    ;;#ASMEND
136; GCN-NEXT:    ;;#ASMSTART
137; GCN-NEXT:    ;;#ASMEND
138; GCN-NEXT:    ;;#ASMSTART
139; GCN-NEXT:    ;;#ASMEND
140; GCN-NEXT:    ;;#ASMSTART
141; GCN-NEXT:    ;;#ASMEND
142; GCN-NEXT:    ;;#ASMSTART
143; GCN-NEXT:    ;;#ASMEND
144; GCN-NEXT:    ;;#ASMSTART
145; GCN-NEXT:    ;;#ASMEND
146; GCN-NEXT:    ;;#ASMSTART
147; GCN-NEXT:    ;;#ASMEND
148; GCN-NEXT:    ;;#ASMSTART
149; GCN-NEXT:    ;;#ASMEND
150; GCN-NEXT:    ;;#ASMSTART
151; GCN-NEXT:    ;;#ASMEND
152; GCN-NEXT:    ;;#ASMSTART
153; GCN-NEXT:    ;;#ASMEND
154; GCN-NEXT:    ;;#ASMSTART
155; GCN-NEXT:    ;;#ASMEND
156; GCN-NEXT:    ;;#ASMSTART
157; GCN-NEXT:    ;;#ASMEND
158; GCN-NEXT:    ;;#ASMSTART
159; GCN-NEXT:    ;;#ASMEND
160; GCN-NEXT:    ;;#ASMSTART
161; GCN-NEXT:    ;;#ASMEND
162; GCN-NEXT:    ;;#ASMSTART
163; GCN-NEXT:    ;;#ASMEND
164; GCN-NEXT:    ;;#ASMSTART
165; GCN-NEXT:    ;;#ASMEND
166; GCN-NEXT:    ;;#ASMSTART
167; GCN-NEXT:    ;;#ASMEND
168; GCN-NEXT:    ;;#ASMSTART
169; GCN-NEXT:    ;;#ASMEND
170; GCN-NEXT:    ;;#ASMSTART
171; GCN-NEXT:    ;;#ASMEND
172; GCN-NEXT:    ;;#ASMSTART
173; GCN-NEXT:    ;;#ASMEND
174; GCN-NEXT:    s_waitcnt lgkmcnt(0)
175; GCN-NEXT:    v_mov_b32_e32 v0, vcc_lo
176; GCN-NEXT:    ;;#ASMSTART
177; GCN-NEXT:    ;;#ASMEND
178; GCN-NEXT:    ;;#ASMSTART
179; GCN-NEXT:    ;;#ASMEND
180; GCN-NEXT:    ;;#ASMSTART
181; GCN-NEXT:    ;;#ASMEND
182; GCN-NEXT:    ;;#ASMSTART
183; GCN-NEXT:    ;;#ASMEND
184; GCN-NEXT:    s_mov_b32 s2, exec_lo
185; GCN-NEXT:    s_mov_b32 exec_lo, 3
186; GCN-NEXT:    s_mov_b32 s3, 0
187; GCN-NEXT:    scratch_store_dword off, v2, s3
188; GCN-NEXT:    s_waitcnt_depctr 0xffe3
189; GCN-NEXT:    s_mov_b32 s3, 4
190; GCN-NEXT:    scratch_load_dword v2, off, s3 ; 4-byte Folded Reload
191; GCN-NEXT:    s_waitcnt_depctr 0xffe3
192; GCN-NEXT:    s_mov_b32 s3, 0
193; GCN-NEXT:    s_waitcnt vmcnt(0)
194; GCN-NEXT:    v_readlane_b32 s0, v2, 0
195; GCN-NEXT:    v_readlane_b32 s1, v2, 1
196; GCN-NEXT:    scratch_load_dword v2, off, s3
197; GCN-NEXT:    s_waitcnt vmcnt(0)
198; GCN-NEXT:    s_waitcnt_depctr 0xffe3
199; GCN-NEXT:    s_mov_b32 exec_lo, s2
200; GCN-NEXT:    v_mov_b32_e32 v1, 0
201; GCN-NEXT:    global_store_dword v1, v0, s[0:1]
202; GCN-NEXT:    s_endpgm
203  call void asm sideeffect "", "~{s[0:7]}" ()
204  call void asm sideeffect "", "~{s[8:15]}" ()
205  call void asm sideeffect "", "~{s[16:23]}" ()
206  call void asm sideeffect "", "~{s[24:31]}" ()
207  call void asm sideeffect "", "~{s[32:39]}" ()
208  call void asm sideeffect "", "~{s[40:47]}" ()
209  call void asm sideeffect "", "~{s[48:55]}" ()
210  call void asm sideeffect "", "~{s[56:63]}" ()
211  call void asm sideeffect "", "~{s[64:71]}" ()
212  call void asm sideeffect "", "~{s[72:79]}" ()
213  call void asm sideeffect "", "~{s[80:87]}" ()
214  call void asm sideeffect "", "~{s[88:95]}" ()
215  call void asm sideeffect "", "~{s[96:103]}" ()
216  call void asm sideeffect "", "~{s[104:105]}" ()
217  call void asm sideeffect "", "~{v[0:7]}" ()
218  call void asm sideeffect "", "~{v[8:15]}" ()
219  call void asm sideeffect "", "~{v[16:23]}" ()
220  call void asm sideeffect "", "~{v[24:31]}" ()
221  call void asm sideeffect "", "~{v[32:39]}" ()
222  call void asm sideeffect "", "~{v[40:47]}" ()
223  call void asm sideeffect "", "~{v[48:55]}" ()
224  call void asm sideeffect "", "~{v[56:63]}" ()
225  call void asm sideeffect "", "~{v[64:71]}" ()
226  call void asm sideeffect "", "~{v[72:79]}" ()
227  call void asm sideeffect "", "~{v[80:87]}" ()
228  call void asm sideeffect "", "~{v[88:95]}" ()
229  call void asm sideeffect "", "~{v[96:103]}" ()
230  call void asm sideeffect "", "~{v[104:111]}" ()
231  call void asm sideeffect "", "~{v[112:119]}" ()
232  call void asm sideeffect "", "~{v[120:127]}" ()
233  call void asm sideeffect "", "~{v[128:135]}" ()
234  call void asm sideeffect "", "~{v[136:143]}" ()
235  call void asm sideeffect "", "~{v[144:151]}" ()
236  call void asm sideeffect "", "~{v[152:159]}" ()
237  call void asm sideeffect "", "~{v[160:167]}" ()
238  call void asm sideeffect "", "~{v[168:175]}" ()
239  call void asm sideeffect "", "~{v[176:183]}" ()
240  call void asm sideeffect "", "~{v[184:191]}" ()
241  call void asm sideeffect "", "~{v[192:199]}" ()
242  call void asm sideeffect "", "~{v[200:207]}" ()
243  call void asm sideeffect "", "~{v[208:215]}" ()
244  call void asm sideeffect "", "~{v[216:223]}" ()
245  call void asm sideeffect "", "~{v[224:231]}" ()
246  call void asm sideeffect "", "~{v[232:239]}" ()
247  call void asm sideeffect "", "~{v[240:247]}" ()
248  call void asm sideeffect "", "~{v[248:255]}" ()
249
250  store i32 %in, i32 addrspace(1)* %out
251  ret void
252}
253
254define amdgpu_kernel void @kernel_no_calls_no_stack() {
255; GCN-LABEL: kernel_no_calls_no_stack:
256; GCN:       ; %bb.0:
257; GCN-NEXT:    s_endpgm
258  ret void
259}
260
261attributes #0 = { nounwind }
262