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,FLAT_SCR_OPT %s
3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 --mattr=+architected-flat-scratch < %s | FileCheck -check-prefixes=GCN,FLAT_SCR_ARCH %s
4
5declare void @extern_func() #0
6
7define amdgpu_kernel void @stack_object_addrspacecast_in_kernel_no_calls() {
8; FLAT_SCR_OPT-LABEL: stack_object_addrspacecast_in_kernel_no_calls:
9; FLAT_SCR_OPT:       ; %bb.0:
10; FLAT_SCR_OPT-NEXT:    s_add_u32 s0, s0, s3
11; FLAT_SCR_OPT-NEXT:    s_addc_u32 s1, s1, 0
12; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
13; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
14; FLAT_SCR_OPT-NEXT:    v_mov_b32_e32 v0, 4
15; FLAT_SCR_OPT-NEXT:    s_getreg_b32 s0, hwreg(HW_REG_SH_MEM_BASES, 0, 16)
16; FLAT_SCR_OPT-NEXT:    v_mov_b32_e32 v2, 0
17; FLAT_SCR_OPT-NEXT:    s_lshl_b32 s0, s0, 16
18; FLAT_SCR_OPT-NEXT:    v_cmp_ne_u32_e32 vcc_lo, -1, v0
19; FLAT_SCR_OPT-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc_lo
20; FLAT_SCR_OPT-NEXT:    v_cndmask_b32_e64 v1, 0, s0, vcc_lo
21; FLAT_SCR_OPT-NEXT:    flat_store_dword v[0:1], v2
22; FLAT_SCR_OPT-NEXT:    s_waitcnt_vscnt null, 0x0
23; FLAT_SCR_OPT-NEXT:    s_endpgm
24;
25; FLAT_SCR_ARCH-LABEL: stack_object_addrspacecast_in_kernel_no_calls:
26; FLAT_SCR_ARCH:       ; %bb.0:
27; FLAT_SCR_ARCH-NEXT:    v_mov_b32_e32 v0, 4
28; FLAT_SCR_ARCH-NEXT:    s_getreg_b32 s0, hwreg(HW_REG_SH_MEM_BASES, 0, 16)
29; FLAT_SCR_ARCH-NEXT:    v_mov_b32_e32 v2, 0
30; FLAT_SCR_ARCH-NEXT:    s_lshl_b32 s0, s0, 16
31; FLAT_SCR_ARCH-NEXT:    v_cmp_ne_u32_e32 vcc_lo, -1, v0
32; FLAT_SCR_ARCH-NEXT:    v_cndmask_b32_e32 v0, 0, v0, vcc_lo
33; FLAT_SCR_ARCH-NEXT:    v_cndmask_b32_e64 v1, 0, s0, vcc_lo
34; FLAT_SCR_ARCH-NEXT:    flat_store_dword v[0:1], v2
35; FLAT_SCR_ARCH-NEXT:    s_waitcnt_vscnt null, 0x0
36; FLAT_SCR_ARCH-NEXT:    s_endpgm
37  %alloca = alloca i32, addrspace(5)
38  %cast = addrspacecast i32 addrspace(5)* %alloca to i32*
39  store volatile i32 0, i32* %cast
40  ret void
41}
42
43define amdgpu_kernel void @stack_object_in_kernel_no_calls() {
44; FLAT_SCR_OPT-LABEL: stack_object_in_kernel_no_calls:
45; FLAT_SCR_OPT:       ; %bb.0:
46; FLAT_SCR_OPT-NEXT:    s_add_u32 s0, s0, s3
47; FLAT_SCR_OPT-NEXT:    s_addc_u32 s1, s1, 0
48; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s0
49; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s1
50; FLAT_SCR_OPT-NEXT:    v_mov_b32_e32 v0, 0
51; FLAT_SCR_OPT-NEXT:    s_mov_b32 vcc_lo, 0
52; FLAT_SCR_OPT-NEXT:    scratch_store_dword off, v0, vcc_lo offset:4
53; FLAT_SCR_OPT-NEXT:    s_waitcnt_vscnt null, 0x0
54; FLAT_SCR_OPT-NEXT:    s_endpgm
55;
56; FLAT_SCR_ARCH-LABEL: stack_object_in_kernel_no_calls:
57; FLAT_SCR_ARCH:       ; %bb.0:
58; FLAT_SCR_ARCH-NEXT:    v_mov_b32_e32 v0, 0
59; FLAT_SCR_ARCH-NEXT:    s_mov_b32 vcc_lo, 0
60; FLAT_SCR_ARCH-NEXT:    scratch_store_dword off, v0, vcc_lo offset:4
61; FLAT_SCR_ARCH-NEXT:    s_waitcnt_vscnt null, 0x0
62; FLAT_SCR_ARCH-NEXT:    s_endpgm
63  %alloca = alloca i32, addrspace(5)
64  store volatile i32 0, i32 addrspace(5)* %alloca
65  ret void
66}
67
68define amdgpu_kernel void @kernel_calls_no_stack() {
69; FLAT_SCR_OPT-LABEL: kernel_calls_no_stack:
70; FLAT_SCR_OPT:       ; %bb.0:
71; FLAT_SCR_OPT-NEXT:    s_add_u32 s6, s6, s11
72; FLAT_SCR_OPT-NEXT:    s_mov_b32 s32, 0
73; FLAT_SCR_OPT-NEXT:    s_addc_u32 s7, s7, 0
74; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s6
75; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s7
76; FLAT_SCR_OPT-NEXT:    s_getpc_b64 s[0:1]
77; FLAT_SCR_OPT-NEXT:    s_add_u32 s0, s0, extern_func@gotpcrel32@lo+4
78; FLAT_SCR_OPT-NEXT:    s_addc_u32 s1, s1, extern_func@gotpcrel32@hi+12
79; FLAT_SCR_OPT-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
80; FLAT_SCR_OPT-NEXT:    s_waitcnt lgkmcnt(0)
81; FLAT_SCR_OPT-NEXT:    s_swappc_b64 s[30:31], s[0:1]
82; FLAT_SCR_OPT-NEXT:    s_endpgm
83;
84; FLAT_SCR_ARCH-LABEL: kernel_calls_no_stack:
85; FLAT_SCR_ARCH:       ; %bb.0:
86; FLAT_SCR_ARCH-NEXT:    s_getpc_b64 s[0:1]
87; FLAT_SCR_ARCH-NEXT:    s_add_u32 s0, s0, extern_func@gotpcrel32@lo+4
88; FLAT_SCR_ARCH-NEXT:    s_addc_u32 s1, s1, extern_func@gotpcrel32@hi+12
89; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s32, 0
90; FLAT_SCR_ARCH-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
91; FLAT_SCR_ARCH-NEXT:    s_waitcnt lgkmcnt(0)
92; FLAT_SCR_ARCH-NEXT:    s_swappc_b64 s[30:31], s[0:1]
93; FLAT_SCR_ARCH-NEXT:    s_endpgm
94  call void @extern_func()
95  ret void
96}
97
98define amdgpu_kernel void @test(i32 addrspace(1)* %out, i32 %in) {
99; FLAT_SCR_OPT-LABEL: test:
100; FLAT_SCR_OPT:       ; %bb.0:
101; FLAT_SCR_OPT-NEXT:    s_add_u32 s2, s2, s5
102; FLAT_SCR_OPT-NEXT:    s_addc_u32 s3, s3, 0
103; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_LO), s2
104; FLAT_SCR_OPT-NEXT:    s_setreg_b32 hwreg(HW_REG_FLAT_SCR_HI), s3
105; FLAT_SCR_OPT-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
106; FLAT_SCR_OPT-NEXT:    s_mov_b32 s104, exec_lo
107; FLAT_SCR_OPT-NEXT:    s_mov_b32 exec_lo, 3
108; FLAT_SCR_OPT-NEXT:    s_mov_b32 s105, 0
109; FLAT_SCR_OPT-NEXT:    scratch_store_dword off, v72, s105
110; FLAT_SCR_OPT-NEXT:    s_waitcnt lgkmcnt(0)
111; FLAT_SCR_OPT-NEXT:    v_writelane_b32 v72, s2, 0
112; FLAT_SCR_OPT-NEXT:    s_mov_b32 s105, 4
113; FLAT_SCR_OPT-NEXT:    v_writelane_b32 v72, s3, 1
114; FLAT_SCR_OPT-NEXT:    scratch_store_dword off, v72, s105 ; 4-byte Folded Spill
115; FLAT_SCR_OPT-NEXT:    s_waitcnt_depctr 0xffe3
116; FLAT_SCR_OPT-NEXT:    s_mov_b32 s105, 0
117; FLAT_SCR_OPT-NEXT:    scratch_load_dword v72, off, s105
118; FLAT_SCR_OPT-NEXT:    s_waitcnt vmcnt(0)
119; FLAT_SCR_OPT-NEXT:    s_waitcnt_depctr 0xffe3
120; FLAT_SCR_OPT-NEXT:    s_mov_b32 exec_lo, s104
121; FLAT_SCR_OPT-NEXT:    s_load_dword vcc_lo, s[0:1], 0x8
122; FLAT_SCR_OPT-NEXT:    ; kill: killed $sgpr0_sgpr1
123; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
124; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
125; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
126; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
127; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
128; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
129; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
130; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
131; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
132; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
133; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
134; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
135; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
136; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
137; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
138; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
139; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
140; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
141; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
142; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
143; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
144; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
145; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
146; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
147; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
148; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
149; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
150; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
151; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
152; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
153; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
154; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
155; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
156; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
157; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
158; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
159; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
160; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
161; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
162; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
163; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
164; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
165; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
166; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
167; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
168; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
169; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
170; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
171; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
172; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
173; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
174; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
175; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
176; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
177; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
178; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
179; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
180; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
181; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
182; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
183; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
184; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
185; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
186; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
187; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
188; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
189; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
190; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
191; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
192; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
193; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
194; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
195; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
196; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
197; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
198; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
199; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
200; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
201; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
202; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
203; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
204; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
205; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
206; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
207; FLAT_SCR_OPT-NEXT:    s_waitcnt lgkmcnt(0)
208; FLAT_SCR_OPT-NEXT:    v_mov_b32_e32 v0, vcc_lo
209; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
210; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
211; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
212; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
213; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
214; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
215; FLAT_SCR_OPT-NEXT:    ;;#ASMSTART
216; FLAT_SCR_OPT-NEXT:    ;;#ASMEND
217; FLAT_SCR_OPT-NEXT:    s_mov_b32 s2, exec_lo
218; FLAT_SCR_OPT-NEXT:    s_mov_b32 exec_lo, 3
219; FLAT_SCR_OPT-NEXT:    s_mov_b32 s3, 0
220; FLAT_SCR_OPT-NEXT:    scratch_store_dword off, v2, s3
221; FLAT_SCR_OPT-NEXT:    s_waitcnt_depctr 0xffe3
222; FLAT_SCR_OPT-NEXT:    s_mov_b32 s3, 4
223; FLAT_SCR_OPT-NEXT:    scratch_load_dword v2, off, s3 ; 4-byte Folded Reload
224; FLAT_SCR_OPT-NEXT:    s_waitcnt_depctr 0xffe3
225; FLAT_SCR_OPT-NEXT:    s_mov_b32 s3, 0
226; FLAT_SCR_OPT-NEXT:    s_waitcnt vmcnt(0)
227; FLAT_SCR_OPT-NEXT:    v_readlane_b32 s0, v2, 0
228; FLAT_SCR_OPT-NEXT:    v_readlane_b32 s1, v2, 1
229; FLAT_SCR_OPT-NEXT:    scratch_load_dword v2, off, s3
230; FLAT_SCR_OPT-NEXT:    s_waitcnt vmcnt(0)
231; FLAT_SCR_OPT-NEXT:    s_waitcnt_depctr 0xffe3
232; FLAT_SCR_OPT-NEXT:    s_mov_b32 exec_lo, s2
233; FLAT_SCR_OPT-NEXT:    v_mov_b32_e32 v1, 0
234; FLAT_SCR_OPT-NEXT:    global_store_dword v1, v0, s[0:1]
235; FLAT_SCR_OPT-NEXT:    s_endpgm
236;
237; FLAT_SCR_ARCH-LABEL: test:
238; FLAT_SCR_ARCH:       ; %bb.0:
239; FLAT_SCR_ARCH-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
240; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s104, exec_lo
241; FLAT_SCR_ARCH-NEXT:    s_mov_b32 exec_lo, 3
242; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s105, 0
243; FLAT_SCR_ARCH-NEXT:    scratch_store_dword off, v72, s105
244; FLAT_SCR_ARCH-NEXT:    s_waitcnt lgkmcnt(0)
245; FLAT_SCR_ARCH-NEXT:    v_writelane_b32 v72, s2, 0
246; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s105, 4
247; FLAT_SCR_ARCH-NEXT:    v_writelane_b32 v72, s3, 1
248; FLAT_SCR_ARCH-NEXT:    scratch_store_dword off, v72, s105 ; 4-byte Folded Spill
249; FLAT_SCR_ARCH-NEXT:    s_waitcnt_depctr 0xffe3
250; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s105, 0
251; FLAT_SCR_ARCH-NEXT:    scratch_load_dword v72, off, s105
252; FLAT_SCR_ARCH-NEXT:    s_waitcnt vmcnt(0)
253; FLAT_SCR_ARCH-NEXT:    s_waitcnt_depctr 0xffe3
254; FLAT_SCR_ARCH-NEXT:    s_mov_b32 exec_lo, s104
255; FLAT_SCR_ARCH-NEXT:    s_load_dword vcc_lo, s[0:1], 0x8
256; FLAT_SCR_ARCH-NEXT:    ; kill: killed $sgpr0_sgpr1
257; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
258; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
259; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
260; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
261; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
262; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
263; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
264; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
265; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
266; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
267; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
268; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
269; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
270; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
271; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
272; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
273; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
274; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
275; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
276; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
277; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
278; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
279; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
280; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
281; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
282; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
283; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
284; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
285; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
286; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
287; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
288; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
289; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
290; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
291; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
292; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
293; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
294; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
295; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
296; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
297; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
298; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
299; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
300; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
301; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
302; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
303; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
304; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
305; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
306; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
307; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
308; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
309; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
310; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
311; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
312; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
313; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
314; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
315; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
316; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
317; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
318; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
319; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
320; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
321; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
322; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
323; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
324; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
325; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
326; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
327; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
328; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
329; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
330; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
331; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
332; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
333; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
334; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
335; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
336; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
337; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
338; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
339; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
340; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
341; FLAT_SCR_ARCH-NEXT:    s_waitcnt lgkmcnt(0)
342; FLAT_SCR_ARCH-NEXT:    v_mov_b32_e32 v0, vcc_lo
343; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
344; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
345; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
346; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
347; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
348; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
349; FLAT_SCR_ARCH-NEXT:    ;;#ASMSTART
350; FLAT_SCR_ARCH-NEXT:    ;;#ASMEND
351; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s2, exec_lo
352; FLAT_SCR_ARCH-NEXT:    s_mov_b32 exec_lo, 3
353; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s3, 0
354; FLAT_SCR_ARCH-NEXT:    scratch_store_dword off, v2, s3
355; FLAT_SCR_ARCH-NEXT:    s_waitcnt_depctr 0xffe3
356; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s3, 4
357; FLAT_SCR_ARCH-NEXT:    scratch_load_dword v2, off, s3 ; 4-byte Folded Reload
358; FLAT_SCR_ARCH-NEXT:    s_waitcnt_depctr 0xffe3
359; FLAT_SCR_ARCH-NEXT:    s_mov_b32 s3, 0
360; FLAT_SCR_ARCH-NEXT:    s_waitcnt vmcnt(0)
361; FLAT_SCR_ARCH-NEXT:    v_readlane_b32 s0, v2, 0
362; FLAT_SCR_ARCH-NEXT:    v_readlane_b32 s1, v2, 1
363; FLAT_SCR_ARCH-NEXT:    scratch_load_dword v2, off, s3
364; FLAT_SCR_ARCH-NEXT:    s_waitcnt vmcnt(0)
365; FLAT_SCR_ARCH-NEXT:    s_waitcnt_depctr 0xffe3
366; FLAT_SCR_ARCH-NEXT:    s_mov_b32 exec_lo, s2
367; FLAT_SCR_ARCH-NEXT:    v_mov_b32_e32 v1, 0
368; FLAT_SCR_ARCH-NEXT:    global_store_dword v1, v0, s[0:1]
369; FLAT_SCR_ARCH-NEXT:    s_endpgm
370  call void asm sideeffect "", "~{s[0:7]}" ()
371  call void asm sideeffect "", "~{s[8:15]}" ()
372  call void asm sideeffect "", "~{s[16:23]}" ()
373  call void asm sideeffect "", "~{s[24:31]}" ()
374  call void asm sideeffect "", "~{s[32:39]}" ()
375  call void asm sideeffect "", "~{s[40:47]}" ()
376  call void asm sideeffect "", "~{s[48:55]}" ()
377  call void asm sideeffect "", "~{s[56:63]}" ()
378  call void asm sideeffect "", "~{s[64:71]}" ()
379  call void asm sideeffect "", "~{s[72:79]}" ()
380  call void asm sideeffect "", "~{s[80:87]}" ()
381  call void asm sideeffect "", "~{s[88:95]}" ()
382  call void asm sideeffect "", "~{s[96:103]}" ()
383  call void asm sideeffect "", "~{s[104:105]}" ()
384  call void asm sideeffect "", "~{v[0:7]}" ()
385  call void asm sideeffect "", "~{v[8:15]}" ()
386  call void asm sideeffect "", "~{v[16:23]}" ()
387  call void asm sideeffect "", "~{v[24:31]}" ()
388  call void asm sideeffect "", "~{v[32:39]}" ()
389  call void asm sideeffect "", "~{v[40:47]}" ()
390  call void asm sideeffect "", "~{v[48:55]}" ()
391  call void asm sideeffect "", "~{v[56:63]}" ()
392  call void asm sideeffect "", "~{v[64:71]}" ()
393  call void asm sideeffect "", "~{v[72:79]}" ()
394  call void asm sideeffect "", "~{v[80:87]}" ()
395  call void asm sideeffect "", "~{v[88:95]}" ()
396  call void asm sideeffect "", "~{v[96:103]}" ()
397  call void asm sideeffect "", "~{v[104:111]}" ()
398  call void asm sideeffect "", "~{v[112:119]}" ()
399  call void asm sideeffect "", "~{v[120:127]}" ()
400  call void asm sideeffect "", "~{v[128:135]}" ()
401  call void asm sideeffect "", "~{v[136:143]}" ()
402  call void asm sideeffect "", "~{v[144:151]}" ()
403  call void asm sideeffect "", "~{v[152:159]}" ()
404  call void asm sideeffect "", "~{v[160:167]}" ()
405  call void asm sideeffect "", "~{v[168:175]}" ()
406  call void asm sideeffect "", "~{v[176:183]}" ()
407  call void asm sideeffect "", "~{v[184:191]}" ()
408  call void asm sideeffect "", "~{v[192:199]}" ()
409  call void asm sideeffect "", "~{v[200:207]}" ()
410  call void asm sideeffect "", "~{v[208:215]}" ()
411  call void asm sideeffect "", "~{v[216:223]}" ()
412  call void asm sideeffect "", "~{v[224:231]}" ()
413  call void asm sideeffect "", "~{v[232:239]}" ()
414  call void asm sideeffect "", "~{v[240:247]}" ()
415  call void asm sideeffect "", "~{v[248:255]}" ()
416
417  store i32 %in, i32 addrspace(1)* %out
418  ret void
419}
420
421define amdgpu_kernel void @kernel_no_calls_no_stack() {
422; GCN-LABEL: kernel_no_calls_no_stack:
423; GCN:       ; %bb.0:
424; GCN-NEXT:    s_endpgm
425  ret void
426}
427
428attributes #0 = { nounwind }
429