xref: /llvm-project/llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll (revision 11b040192640ef3b1f481124c440f464ed6ec86a)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -O0 -mtriple=amdgcn- -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9-O0 %s
3; RUN: llc -mtriple=amdgcn- -mcpu=gfx900 -amdgpu-dpp-combine=false -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9-O3 %s
4
5define amdgpu_gfx void @strict_wwm_no_cfg(ptr addrspace(8) inreg %tmp14) {
6; GFX9-O0-LABEL: strict_wwm_no_cfg:
7; GFX9-O0:       ; %bb.0:
8; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
9; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
10; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
11; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
12; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
13; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
14; GFX9-O0-NEXT:    s_mov_b32 s40, s6
15; GFX9-O0-NEXT:    s_mov_b32 s34, s4
16; GFX9-O0-NEXT:    ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41
17; GFX9-O0-NEXT:    s_mov_b32 s41, s7
18; GFX9-O0-NEXT:    s_mov_b32 s42, s41
19; GFX9-O0-NEXT:    s_mov_b32 s43, s40
20; GFX9-O0-NEXT:    ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35
21; GFX9-O0-NEXT:    s_mov_b32 s35, s5
22; GFX9-O0-NEXT:    s_mov_b32 s44, s35
23; GFX9-O0-NEXT:    s_mov_b32 s36, s34
24; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39
25; GFX9-O0-NEXT:    s_mov_b32 s37, s44
26; GFX9-O0-NEXT:    s_mov_b32 s38, s43
27; GFX9-O0-NEXT:    s_mov_b32 s39, s42
28; GFX9-O0-NEXT:    s_mov_b32 s34, 0
29; GFX9-O0-NEXT:    buffer_load_dwordx2 v[5:6], off, s[36:39], s34
30; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
31; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v5
32; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v6
33; GFX9-O0-NEXT:    ; implicit-def: $sgpr40_sgpr41
34; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
35; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s34
36; GFX9-O0-NEXT:    v_cndmask_b32_e64 v0, v0, v4, s[40:41]
37; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
38; GFX9-O0-NEXT:    ; implicit-def: $sgpr40_sgpr41
39; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
40; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s34
41; GFX9-O0-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[40:41]
42; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s34
43; GFX9-O0-NEXT:    s_nop 1
44; GFX9-O0-NEXT:    v_mov_b32_dpp v2, v0 row_bcast:31 row_mask:0xc bank_mask:0xf
45; GFX9-O0-NEXT:    v_add_u32_e64 v0, v0, v2
46; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
47; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v0
48; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
49; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s34
50; GFX9-O0-NEXT:    s_nop 1
51; GFX9-O0-NEXT:    v_mov_b32_dpp v0, v1 row_bcast:31 row_mask:0xc bank_mask:0xf
52; GFX9-O0-NEXT:    v_add_u32_e64 v0, v1, v0
53; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
54; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v0
55; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[40:41], v3, v4
56; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s[40:41]
57; GFX9-O0-NEXT:    s_mov_b32 s35, 1
58; GFX9-O0-NEXT:    v_lshlrev_b32_e64 v3, s35, v3
59; GFX9-O0-NEXT:    s_mov_b32 s35, 2
60; GFX9-O0-NEXT:    v_and_b32_e64 v3, v3, s35
61; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[36:39], s34 offset:4
62; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
63; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
64; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
65; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
66; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
67; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
68; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
69;
70; GFX9-O3-LABEL: strict_wwm_no_cfg:
71; GFX9-O3:       ; %bb.0:
72; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
73; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
74; GFX9-O3-NEXT:    buffer_store_dword v0, off, s[0:3], s32 ; 4-byte Folded Spill
75; GFX9-O3-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
76; GFX9-O3-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
77; GFX9-O3-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
78; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
79; GFX9-O3-NEXT:    buffer_load_dwordx2 v[4:5], off, s[4:7], 0
80; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
81; GFX9-O3-NEXT:    v_mov_b32_e32 v0, 0
82; GFX9-O3-NEXT:    v_mov_b32_e32 v1, 0
83; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
84; GFX9-O3-NEXT:    v_cndmask_b32_e64 v2, 0, v4, s[34:35]
85; GFX9-O3-NEXT:    v_cndmask_b32_e64 v3, 0, v5, s[34:35]
86; GFX9-O3-NEXT:    s_nop 0
87; GFX9-O3-NEXT:    v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
88; GFX9-O3-NEXT:    v_mov_b32_dpp v0, v3 row_bcast:31 row_mask:0xc bank_mask:0xf
89; GFX9-O3-NEXT:    v_add_u32_e32 v1, v2, v1
90; GFX9-O3-NEXT:    v_add_u32_e32 v0, v3, v0
91; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
92; GFX9-O3-NEXT:    v_mov_b32_e32 v4, v1
93; GFX9-O3-NEXT:    v_mov_b32_e32 v5, v0
94; GFX9-O3-NEXT:    v_cmp_eq_u32_e32 vcc, v4, v5
95; GFX9-O3-NEXT:    v_cndmask_b32_e64 v4, 0, 1, vcc
96; GFX9-O3-NEXT:    v_lshlrev_b32_e32 v4, 1, v4
97; GFX9-O3-NEXT:    v_and_b32_e32 v4, 2, v4
98; GFX9-O3-NEXT:    buffer_store_dword v4, off, s[4:7], 0 offset:4
99; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
100; GFX9-O3-NEXT:    buffer_load_dword v0, off, s[0:3], s32 ; 4-byte Folded Reload
101; GFX9-O3-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
102; GFX9-O3-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
103; GFX9-O3-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
104; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
105; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
106; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
107  %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)
108  %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
109  %tmp102 = extractelement <2 x i32> %tmp101, i32 0
110  %tmp103 = extractelement <2 x i32> %tmp101, i32 1
111  %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
112  %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
113
114
115  %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
116  %tmp121 = add i32 %tmp105, %tmp120
117  %tmp122 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp121)
118
119  %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
120  %tmp136 = add i32 %tmp107, %tmp135
121  %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
122
123  %tmp138 = icmp eq i32 %tmp122, %tmp137
124  %tmp139 = sext i1 %tmp138 to i32
125  %tmp140 = shl nsw i32 %tmp139, 1
126  %tmp141 = and i32 %tmp140, 2
127  %tmp145 = bitcast i32 %tmp141 to float
128  call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
129  ret void
130}
131
132define amdgpu_gfx void @strict_wwm_cfg(ptr addrspace(8) inreg %tmp14, i32 %arg) {
133; GFX9-O0-LABEL: strict_wwm_cfg:
134; GFX9-O0:       ; %bb.0: ; %entry
135; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
136; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
137; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
138; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
139; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
140; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
141; GFX9-O0-NEXT:    s_mov_b32 s40, s6
142; GFX9-O0-NEXT:    s_mov_b32 s34, s4
143; GFX9-O0-NEXT:    ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41
144; GFX9-O0-NEXT:    s_mov_b32 s41, s7
145; GFX9-O0-NEXT:    s_mov_b32 s42, s41
146; GFX9-O0-NEXT:    s_mov_b32 s43, s40
147; GFX9-O0-NEXT:    ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35
148; GFX9-O0-NEXT:    s_mov_b32 s35, s5
149; GFX9-O0-NEXT:    s_mov_b32 s44, s35
150; GFX9-O0-NEXT:    s_mov_b32 s36, s34
151; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39
152; GFX9-O0-NEXT:    s_mov_b32 s37, s44
153; GFX9-O0-NEXT:    s_mov_b32 s38, s43
154; GFX9-O0-NEXT:    s_mov_b32 s39, s42
155; GFX9-O0-NEXT:    ; implicit-def: $vgpr5 : SGPR spill to VGPR lane
156; GFX9-O0-NEXT:    v_writelane_b32 v5, s40, 0
157; GFX9-O0-NEXT:    v_writelane_b32 v5, s41, 1
158; GFX9-O0-NEXT:    v_writelane_b32 v5, s34, 2
159; GFX9-O0-NEXT:    v_writelane_b32 v5, s35, 3
160; GFX9-O0-NEXT:    s_mov_b32 s34, 0
161; GFX9-O0-NEXT:    s_nop 2
162; GFX9-O0-NEXT:    buffer_load_dwordx2 v[3:4], off, s[36:39], s34
163; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
164; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
165; GFX9-O0-NEXT:    s_nop 0
166; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
167; GFX9-O0-NEXT:    ; implicit-def: $sgpr36_sgpr37
168; GFX9-O0-NEXT:    ; implicit-def: $sgpr36_sgpr37
169; GFX9-O0-NEXT:    s_or_saveexec_b64 s[36:37], -1
170; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s34
171; GFX9-O0-NEXT:    v_cndmask_b32_e64 v1, v1, v3, s[36:37]
172; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s34
173; GFX9-O0-NEXT:    s_nop 1
174; GFX9-O0-NEXT:    v_mov_b32_dpp v2, v1 row_bcast:31 row_mask:0xc bank_mask:0xf
175; GFX9-O0-NEXT:    v_add_u32_e64 v1, v1, v2
176; GFX9-O0-NEXT:    s_mov_b64 exec, s[36:37]
177; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
178; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
179; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[36:37], v0, s34
180; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s34
181; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
182; GFX9-O0-NEXT:    s_mov_b64 s[34:35], exec
183; GFX9-O0-NEXT:    v_writelane_b32 v5, s34, 4
184; GFX9-O0-NEXT:    v_writelane_b32 v5, s35, 5
185; GFX9-O0-NEXT:    s_or_saveexec_b64 s[46:47], -1
186; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 ; 4-byte Folded Spill
187; GFX9-O0-NEXT:    s_mov_b64 exec, s[46:47]
188; GFX9-O0-NEXT:    s_and_b64 s[34:35], s[34:35], s[36:37]
189; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
190; GFX9-O0-NEXT:    s_cbranch_execz .LBB1_2
191; GFX9-O0-NEXT:  ; %bb.1: ; %if
192; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
193; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
194; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
195; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v4
196; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
197; GFX9-O0-NEXT:    v_mov_b32_e32 v1, 0
198; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
199; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
200; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
201; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v1, v0, s[34:35]
202; GFX9-O0-NEXT:    s_nop 1
203; GFX9-O0-NEXT:    v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
204; GFX9-O0-NEXT:    v_add_u32_e64 v1, v2, v1
205; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
206; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v1
207; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
208; GFX9-O0-NEXT:  .LBB1_2: ; %merge
209; GFX9-O0-NEXT:    s_or_saveexec_b64 s[46:47], -1
210; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 ; 4-byte Folded Reload
211; GFX9-O0-NEXT:    s_mov_b64 exec, s[46:47]
212; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
213; GFX9-O0-NEXT:    v_readlane_b32 s36, v5, 4
214; GFX9-O0-NEXT:    v_readlane_b32 s37, v5, 5
215; GFX9-O0-NEXT:    s_or_b64 exec, exec, s[36:37]
216; GFX9-O0-NEXT:    v_readlane_b32 s38, v5, 0
217; GFX9-O0-NEXT:    v_readlane_b32 s39, v5, 1
218; GFX9-O0-NEXT:    v_readlane_b32 s34, v5, 2
219; GFX9-O0-NEXT:    v_readlane_b32 s35, v5, 3
220; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
221; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
222; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
223; GFX9-O0-NEXT:    v_cmp_eq_u32_e64 s[36:37], v0, v3
224; GFX9-O0-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[36:37]
225; GFX9-O0-NEXT:    s_mov_b32 s36, 1
226; GFX9-O0-NEXT:    v_lshlrev_b32_e64 v0, s36, v0
227; GFX9-O0-NEXT:    s_mov_b32 s36, 2
228; GFX9-O0-NEXT:    v_and_b32_e64 v0, v0, s36
229; GFX9-O0-NEXT:    s_mov_b32 s40, s35
230; GFX9-O0-NEXT:    s_mov_b32 s36, s34
231; GFX9-O0-NEXT:    s_mov_b32 s34, s39
232; GFX9-O0-NEXT:    s_mov_b32 s35, s38
233; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39
234; GFX9-O0-NEXT:    s_mov_b32 s37, s40
235; GFX9-O0-NEXT:    s_mov_b32 s38, s35
236; GFX9-O0-NEXT:    s_mov_b32 s39, s34
237; GFX9-O0-NEXT:    s_mov_b32 s34, 0
238; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[36:39], s34 offset:4
239; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
240; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
241; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
242; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
243; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
244; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
245; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
246;
247; GFX9-O3-LABEL: strict_wwm_cfg:
248; GFX9-O3:       ; %bb.0: ; %entry
249; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
250; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
251; GFX9-O3-NEXT:    buffer_store_dword v1, off, s[0:3], s32 ; 4-byte Folded Spill
252; GFX9-O3-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
253; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
254; GFX9-O3-NEXT:    buffer_load_dwordx2 v[3:4], off, s[4:7], 0
255; GFX9-O3-NEXT:    v_mov_b32_e32 v5, 0
256; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
257; GFX9-O3-NEXT:    v_mov_b32_e32 v1, 0
258; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
259; GFX9-O3-NEXT:    v_cndmask_b32_e64 v2, 0, v3, s[34:35]
260; GFX9-O3-NEXT:    s_nop 1
261; GFX9-O3-NEXT:    v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
262; GFX9-O3-NEXT:    v_add_u32_e32 v1, v2, v1
263; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
264; GFX9-O3-NEXT:    v_mov_b32_e32 v3, v1
265; GFX9-O3-NEXT:    v_cmp_eq_u32_e32 vcc, 0, v0
266; GFX9-O3-NEXT:    s_and_saveexec_b64 s[34:35], vcc
267; GFX9-O3-NEXT:  ; %bb.1: ; %if
268; GFX9-O3-NEXT:    s_or_saveexec_b64 s[36:37], -1
269; GFX9-O3-NEXT:    v_mov_b32_e32 v1, 0
270; GFX9-O3-NEXT:    v_cndmask_b32_e64 v2, 0, v4, s[36:37]
271; GFX9-O3-NEXT:    s_nop 1
272; GFX9-O3-NEXT:    v_mov_b32_dpp v1, v2 row_bcast:31 row_mask:0xc bank_mask:0xf
273; GFX9-O3-NEXT:    v_add_u32_e32 v1, v2, v1
274; GFX9-O3-NEXT:    s_mov_b64 exec, s[36:37]
275; GFX9-O3-NEXT:    v_mov_b32_e32 v5, v1
276; GFX9-O3-NEXT:  ; %bb.2: ; %merge
277; GFX9-O3-NEXT:    s_or_b64 exec, exec, s[34:35]
278; GFX9-O3-NEXT:    v_cmp_eq_u32_e32 vcc, v3, v5
279; GFX9-O3-NEXT:    v_cndmask_b32_e64 v0, 0, 1, vcc
280; GFX9-O3-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
281; GFX9-O3-NEXT:    v_and_b32_e32 v0, 2, v0
282; GFX9-O3-NEXT:    buffer_store_dword v0, off, s[4:7], 0 offset:4
283; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
284; GFX9-O3-NEXT:    buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
285; GFX9-O3-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
286; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
287; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
288; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
289entry:
290  %tmp100 = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8) %tmp14, i32 0, i32 0, i32 0)
291  %tmp101 = bitcast <2 x float> %tmp100 to <2 x i32>
292  %tmp102 = extractelement <2 x i32> %tmp101, i32 0
293  %tmp105 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp102, i32 0)
294
295  %tmp120 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp105, i32 323, i32 12, i32 15, i1 false)
296  %tmp121 = add i32 %tmp105, %tmp120
297  %tmp122 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp121)
298
299  %cond = icmp eq i32 %arg, 0
300  br i1 %cond, label %if, label %merge
301if:
302  %tmp103 = extractelement <2 x i32> %tmp101, i32 1
303  %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %tmp103, i32 0)
304
305  %tmp135 = tail call i32 @llvm.amdgcn.update.dpp.i32(i32 0, i32 %tmp107, i32 323, i32 12, i32 15, i1 false)
306  %tmp136 = add i32 %tmp107, %tmp135
307  %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
308  br label %merge
309
310merge:
311  %merge_value = phi i32 [ 0, %entry ], [%tmp137, %if ]
312  %tmp138 = icmp eq i32 %tmp122, %merge_value
313  %tmp139 = sext i1 %tmp138 to i32
314  %tmp140 = shl nsw i32 %tmp139, 1
315  %tmp141 = and i32 %tmp140, 2
316  %tmp145 = bitcast i32 %tmp141 to float
317  call void @llvm.amdgcn.raw.ptr.buffer.store.f32(float %tmp145, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
318  ret void
319}
320
321define hidden amdgpu_gfx i32 @strict_wwm_called(i32 %a) noinline {
322; GFX9-O0-LABEL: strict_wwm_called:
323; GFX9-O0:       ; %bb.0:
324; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
325; GFX9-O0-NEXT:    v_add_u32_e64 v1, v0, v0
326; GFX9-O0-NEXT:    v_mul_lo_u32 v0, v1, v0
327; GFX9-O0-NEXT:    v_sub_u32_e64 v0, v0, v1
328; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
329;
330; GFX9-O3-LABEL: strict_wwm_called:
331; GFX9-O3:       ; %bb.0:
332; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
333; GFX9-O3-NEXT:    v_add_u32_e32 v1, v0, v0
334; GFX9-O3-NEXT:    v_mul_lo_u32 v0, v1, v0
335; GFX9-O3-NEXT:    v_sub_u32_e32 v0, v0, v1
336; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
337  %add = add i32 %a, %a
338  %mul = mul i32 %add, %a
339  %sub = sub i32 %mul, %add
340  ret i32 %sub
341}
342
343define amdgpu_gfx void @strict_wwm_call(ptr addrspace(8) inreg %tmp14, i32 inreg %arg) {
344; GFX9-O0-LABEL: strict_wwm_call:
345; GFX9-O0:       ; %bb.0:
346; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
347; GFX9-O0-NEXT:    s_mov_b32 s48, s33
348; GFX9-O0-NEXT:    s_mov_b32 s33, s32
349; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
350; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s33 ; 4-byte Folded Spill
351; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
352; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
353; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
354; GFX9-O0-NEXT:    s_add_i32 s32, s32, 0x400
355; GFX9-O0-NEXT:    v_writelane_b32 v3, s30, 0
356; GFX9-O0-NEXT:    v_writelane_b32 v3, s31, 1
357; GFX9-O0-NEXT:    s_mov_b32 s40, s6
358; GFX9-O0-NEXT:    s_mov_b32 s34, s4
359; GFX9-O0-NEXT:    ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41
360; GFX9-O0-NEXT:    s_mov_b32 s41, s7
361; GFX9-O0-NEXT:    s_mov_b32 s42, s41
362; GFX9-O0-NEXT:    s_mov_b32 s43, s40
363; GFX9-O0-NEXT:    ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35
364; GFX9-O0-NEXT:    s_mov_b32 s35, s5
365; GFX9-O0-NEXT:    s_mov_b32 s44, s35
366; GFX9-O0-NEXT:    s_mov_b32 s36, s34
367; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39
368; GFX9-O0-NEXT:    s_mov_b32 s37, s44
369; GFX9-O0-NEXT:    s_mov_b32 s38, s43
370; GFX9-O0-NEXT:    s_mov_b32 s39, s42
371; GFX9-O0-NEXT:    s_mov_b32 s34, 0
372; GFX9-O0-NEXT:    ; implicit-def: $sgpr40_sgpr41
373; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s8
374; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
375; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s34
376; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v1, v0, s[40:41]
377; GFX9-O0-NEXT:    s_getpc_b64 s[42:43]
378; GFX9-O0-NEXT:    s_add_u32 s42, s42, strict_wwm_called@rel32@lo+4
379; GFX9-O0-NEXT:    s_addc_u32 s43, s43, strict_wwm_called@rel32@hi+12
380; GFX9-O0-NEXT:    s_mov_b64 s[46:47], s[2:3]
381; GFX9-O0-NEXT:    s_mov_b64 s[44:45], s[0:1]
382; GFX9-O0-NEXT:    s_mov_b64 s[0:1], s[44:45]
383; GFX9-O0-NEXT:    s_mov_b64 s[2:3], s[46:47]
384; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v2
385; GFX9-O0-NEXT:    s_swappc_b64 s[30:31], s[42:43]
386; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v0
387; GFX9-O0-NEXT:    v_add_u32_e64 v1, v1, v2
388; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
389; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v1
390; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[36:39], s34 offset:4
391; GFX9-O0-NEXT:    v_readlane_b32 s31, v3, 1
392; GFX9-O0-NEXT:    v_readlane_b32 s30, v3, 0
393; GFX9-O0-NEXT:    s_mov_b32 s32, s33
394; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
395; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s33 ; 4-byte Folded Reload
396; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
397; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
398; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
399; GFX9-O0-NEXT:    s_mov_b32 s33, s48
400; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
401; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
402;
403; GFX9-O3-LABEL: strict_wwm_call:
404; GFX9-O3:       ; %bb.0:
405; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
406; GFX9-O3-NEXT:    s_mov_b32 s38, s33
407; GFX9-O3-NEXT:    s_mov_b32 s33, s32
408; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
409; GFX9-O3-NEXT:    buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
410; GFX9-O3-NEXT:    buffer_store_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
411; GFX9-O3-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
412; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
413; GFX9-O3-NEXT:    v_writelane_b32 v3, s30, 0
414; GFX9-O3-NEXT:    s_addk_i32 s32, 0x400
415; GFX9-O3-NEXT:    v_writelane_b32 v3, s31, 1
416; GFX9-O3-NEXT:    v_mov_b32_e32 v0, s8
417; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
418; GFX9-O3-NEXT:    v_cndmask_b32_e64 v2, 0, v0, s[34:35]
419; GFX9-O3-NEXT:    v_mov_b32_e32 v0, v2
420; GFX9-O3-NEXT:    s_getpc_b64 s[36:37]
421; GFX9-O3-NEXT:    s_add_u32 s36, s36, strict_wwm_called@rel32@lo+4
422; GFX9-O3-NEXT:    s_addc_u32 s37, s37, strict_wwm_called@rel32@hi+12
423; GFX9-O3-NEXT:    s_swappc_b64 s[30:31], s[36:37]
424; GFX9-O3-NEXT:    v_mov_b32_e32 v1, v0
425; GFX9-O3-NEXT:    v_add_u32_e32 v1, v1, v2
426; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
427; GFX9-O3-NEXT:    v_mov_b32_e32 v0, v1
428; GFX9-O3-NEXT:    buffer_store_dword v0, off, s[4:7], 0 offset:4
429; GFX9-O3-NEXT:    v_readlane_b32 s31, v3, 1
430; GFX9-O3-NEXT:    v_readlane_b32 s30, v3, 0
431; GFX9-O3-NEXT:    s_mov_b32 s32, s33
432; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
433; GFX9-O3-NEXT:    buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
434; GFX9-O3-NEXT:    buffer_load_dword v1, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
435; GFX9-O3-NEXT:    buffer_load_dword v3, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
436; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
437; GFX9-O3-NEXT:    s_mov_b32 s33, s38
438; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
439; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
440  %tmp107 = tail call i32 @llvm.amdgcn.set.inactive.i32(i32 %arg, i32 0)
441  %tmp134 = call amdgpu_gfx i32 @strict_wwm_called(i32 %tmp107)
442  %tmp136 = add i32 %tmp134, %tmp107
443  %tmp137 = tail call i32 @llvm.amdgcn.strict.wwm.i32(i32 %tmp136)
444  call void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32 %tmp137, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
445  ret void
446}
447
448define amdgpu_gfx i64 @strict_wwm_called_i64(i64 %a) noinline {
449; GFX9-O0-LABEL: strict_wwm_called_i64:
450; GFX9-O0:       ; %bb.0:
451; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
452; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v0
453; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
454; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
455; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
456; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
457; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
458; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v2
459; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v2
460; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v3
461; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v3
462; GFX9-O0-NEXT:    v_add_co_u32_e64 v4, s[34:35], v4, v5
463; GFX9-O0-NEXT:    v_addc_co_u32_e64 v0, s[34:35], v0, v1, s[34:35]
464; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec
465; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v0
466; GFX9-O0-NEXT:    s_mov_b32 s34, 32
467; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v2
468; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v3
469; GFX9-O0-NEXT:    v_lshrrev_b64 v[0:1], s34, v[0:1]
470; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v0
471; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v4
472; GFX9-O0-NEXT:    v_mul_lo_u32 v1, v0, v1
473; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v2
474; GFX9-O0-NEXT:    v_lshrrev_b64 v[6:7], s34, v[4:5]
475; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v6
476; GFX9-O0-NEXT:    v_mul_lo_u32 v2, v2, v3
477; GFX9-O0-NEXT:    v_mad_u64_u32 v[6:7], s[36:37], v0, v3, 0
478; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v7
479; GFX9-O0-NEXT:    v_add3_u32 v0, v0, v1, v2
480; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
481; GFX9-O0-NEXT:    ; implicit-def: $sgpr36
482; GFX9-O0-NEXT:    ; implicit-def: $sgpr36
483; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s35
484; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1 killed $exec
485; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v2
486; GFX9-O0-NEXT:    v_lshlrev_b64 v[1:2], s34, v[0:1]
487; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v2
488; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 killed $vgpr6_vgpr7 killed $exec
489; GFX9-O0-NEXT:    s_mov_b32 s35, 0
490; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
491; GFX9-O0-NEXT:    v_mov_b32_e32 v0, 0
492; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
493; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v0
494; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v7
495; GFX9-O0-NEXT:    v_or_b32_e64 v0, v0, v3
496; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v1
497; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v6
498; GFX9-O0-NEXT:    v_or_b32_e64 v6, v1, v2
499; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
500; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v0
501; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v6
502; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v4
503; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v7
504; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v5
505; GFX9-O0-NEXT:    v_sub_co_u32_e64 v1, s[36:37], v1, v3
506; GFX9-O0-NEXT:    v_subb_co_u32_e64 v0, s[36:37], v0, v2, s[36:37]
507; GFX9-O0-NEXT:    ; kill: def $vgpr1 killed $vgpr1 def $vgpr1_vgpr2 killed $exec
508; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v0
509; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v1
510; GFX9-O0-NEXT:    v_lshrrev_b64 v[1:2], s34, v[1:2]
511; GFX9-O0-NEXT:    ; kill: def $vgpr1 killed $vgpr1 killed $vgpr1_vgpr2 killed $exec
512; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
513;
514; GFX9-O3-LABEL: strict_wwm_called_i64:
515; GFX9-O3:       ; %bb.0:
516; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
517; GFX9-O3-NEXT:    v_add_co_u32_e32 v2, vcc, v0, v0
518; GFX9-O3-NEXT:    v_addc_co_u32_e32 v3, vcc, v1, v1, vcc
519; GFX9-O3-NEXT:    v_mul_lo_u32 v4, v3, v0
520; GFX9-O3-NEXT:    v_mul_lo_u32 v5, v2, v1
521; GFX9-O3-NEXT:    v_mad_u64_u32 v[0:1], s[34:35], v2, v0, 0
522; GFX9-O3-NEXT:    v_add3_u32 v1, v1, v5, v4
523; GFX9-O3-NEXT:    v_sub_co_u32_e32 v0, vcc, v0, v2
524; GFX9-O3-NEXT:    v_subb_co_u32_e32 v1, vcc, v1, v3, vcc
525; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
526  %add = add i64 %a, %a
527  %mul = mul i64 %add, %a
528  %sub = sub i64 %mul, %add
529  ret i64 %sub
530}
531
532define amdgpu_gfx void @strict_wwm_call_i64(ptr addrspace(8) inreg %tmp14, i64 inreg %arg) {
533; GFX9-O0-LABEL: strict_wwm_call_i64:
534; GFX9-O0:       ; %bb.0:
535; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
536; GFX9-O0-NEXT:    s_mov_b32 s46, s33
537; GFX9-O0-NEXT:    s_mov_b32 s33, s32
538; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
539; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s33 ; 4-byte Folded Spill
540; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
541; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
542; GFX9-O0-NEXT:    s_nop 0
543; GFX9-O0-NEXT:    buffer_store_dword v9, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
544; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
545; GFX9-O0-NEXT:    s_nop 0
546; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
547; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
548; GFX9-O0-NEXT:    s_nop 0
549; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Spill
550; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Spill
551; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s33 offset:36 ; 4-byte Folded Spill
552; GFX9-O0-NEXT:    buffer_store_dword v10, off, s[0:3], s33 offset:40 ; 4-byte Folded Spill
553; GFX9-O0-NEXT:    buffer_store_dword v11, off, s[0:3], s33 offset:44 ; 4-byte Folded Spill
554; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
555; GFX9-O0-NEXT:    s_add_i32 s32, s32, 0x1000
556; GFX9-O0-NEXT:    v_writelane_b32 v10, s30, 0
557; GFX9-O0-NEXT:    v_writelane_b32 v10, s31, 1
558; GFX9-O0-NEXT:    s_mov_b32 s34, s8
559; GFX9-O0-NEXT:    s_mov_b32 s38, s6
560; GFX9-O0-NEXT:    s_mov_b32 s36, s4
561; GFX9-O0-NEXT:    ; kill: def $sgpr38 killed $sgpr38 def $sgpr38_sgpr39
562; GFX9-O0-NEXT:    s_mov_b32 s39, s7
563; GFX9-O0-NEXT:    s_mov_b32 s35, s39
564; GFX9-O0-NEXT:    s_mov_b32 s44, s38
565; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37
566; GFX9-O0-NEXT:    s_mov_b32 s37, s5
567; GFX9-O0-NEXT:    s_mov_b32 s45, s37
568; GFX9-O0-NEXT:    s_mov_b32 s40, s36
569; GFX9-O0-NEXT:    ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41_sgpr42_sgpr43
570; GFX9-O0-NEXT:    s_mov_b32 s41, s45
571; GFX9-O0-NEXT:    s_mov_b32 s42, s44
572; GFX9-O0-NEXT:    s_mov_b32 s43, s35
573; GFX9-O0-NEXT:    ; implicit-def: $vgpr11 : SGPR spill to VGPR lane
574; GFX9-O0-NEXT:    v_writelane_b32 v11, s40, 0
575; GFX9-O0-NEXT:    v_writelane_b32 v11, s41, 1
576; GFX9-O0-NEXT:    v_writelane_b32 v11, s42, 2
577; GFX9-O0-NEXT:    v_writelane_b32 v11, s43, 3
578; GFX9-O0-NEXT:    ; kill: def $sgpr34 killed $sgpr34 def $sgpr34_sgpr35
579; GFX9-O0-NEXT:    s_mov_b32 s35, s9
580; GFX9-O0-NEXT:    ; kill: def $sgpr36_sgpr37 killed $sgpr34_sgpr35
581; GFX9-O0-NEXT:    s_mov_b32 s38, s35
582; GFX9-O0-NEXT:    s_mov_b64 s[36:37], 0
583; GFX9-O0-NEXT:    s_mov_b32 s40, s37
584; GFX9-O0-NEXT:    ; implicit-def: $sgpr42_sgpr43
585; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s38
586; GFX9-O0-NEXT:    s_or_saveexec_b64 s[38:39], -1
587; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s40
588; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v0, s[38:39]
589; GFX9-O0-NEXT:    s_mov_b64 exec, s[38:39]
590; GFX9-O0-NEXT:    ; kill: def $sgpr34 killed $sgpr34 killed $sgpr34_sgpr35
591; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 killed $sgpr36_sgpr37
592; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
593; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s34
594; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
595; GFX9-O0-NEXT:    v_writelane_b32 v11, s34, 4
596; GFX9-O0-NEXT:    v_writelane_b32 v11, s35, 5
597; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s36
598; GFX9-O0-NEXT:    v_cndmask_b32_e64 v3, v3, v0, s[34:35]
599; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
600; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
601; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v3
602; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v2
603; GFX9-O0-NEXT:    s_mov_b32 s34, 32
604; GFX9-O0-NEXT:    v_lshrrev_b64 v[4:5], s34, v[8:9]
605; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v4
606; GFX9-O0-NEXT:    s_getpc_b64 s[34:35]
607; GFX9-O0-NEXT:    s_add_u32 s34, s34, strict_wwm_called_i64@gotpcrel32@lo+4
608; GFX9-O0-NEXT:    s_addc_u32 s35, s35, strict_wwm_called_i64@gotpcrel32@hi+12
609; GFX9-O0-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
610; GFX9-O0-NEXT:    s_mov_b64 s[38:39], s[2:3]
611; GFX9-O0-NEXT:    s_mov_b64 s[36:37], s[0:1]
612; GFX9-O0-NEXT:    s_mov_b64 s[0:1], s[36:37]
613; GFX9-O0-NEXT:    s_mov_b64 s[2:3], s[38:39]
614; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v3
615; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v2
616; GFX9-O0-NEXT:    s_waitcnt lgkmcnt(0)
617; GFX9-O0-NEXT:    s_swappc_b64 s[30:31], s[34:35]
618; GFX9-O0-NEXT:    v_readlane_b32 s34, v11, 4
619; GFX9-O0-NEXT:    v_readlane_b32 s35, v11, 5
620; GFX9-O0-NEXT:    v_readlane_b32 s36, v11, 0
621; GFX9-O0-NEXT:    v_readlane_b32 s37, v11, 1
622; GFX9-O0-NEXT:    v_readlane_b32 s38, v11, 2
623; GFX9-O0-NEXT:    v_readlane_b32 s39, v11, 3
624; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v0
625; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
626; GFX9-O0-NEXT:    ; implicit-def: $sgpr40
627; GFX9-O0-NEXT:    ; implicit-def: $sgpr40
628; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v8
629; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v9
630; GFX9-O0-NEXT:    v_add_co_u32_e64 v2, s[40:41], v2, v4
631; GFX9-O0-NEXT:    v_addc_co_u32_e64 v3, s[40:41], v3, v5, s[40:41]
632; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
633; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v2
634; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v3
635; GFX9-O0-NEXT:    s_mov_b32 s34, 0
636; GFX9-O0-NEXT:    buffer_store_dwordx2 v[0:1], off, s[36:39], s34 offset:4
637; GFX9-O0-NEXT:    v_readlane_b32 s31, v10, 1
638; GFX9-O0-NEXT:    v_readlane_b32 s30, v10, 0
639; GFX9-O0-NEXT:    s_mov_b32 s32, s33
640; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
641; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s33 ; 4-byte Folded Reload
642; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
643; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
644; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
645; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
646; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
647; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
648; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s33 offset:28 ; 4-byte Folded Reload
649; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s33 offset:32 ; 4-byte Folded Reload
650; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s33 offset:36 ; 4-byte Folded Reload
651; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s33 offset:40 ; 4-byte Folded Reload
652; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s33 offset:44 ; 4-byte Folded Reload
653; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
654; GFX9-O0-NEXT:    s_mov_b32 s33, s46
655; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
656; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
657;
658; GFX9-O3-LABEL: strict_wwm_call_i64:
659; GFX9-O3:       ; %bb.0:
660; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
661; GFX9-O3-NEXT:    s_mov_b32 s38, s33
662; GFX9-O3-NEXT:    s_mov_b32 s33, s32
663; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
664; GFX9-O3-NEXT:    buffer_store_dword v6, off, s[0:3], s33 ; 4-byte Folded Spill
665; GFX9-O3-NEXT:    buffer_store_dword v7, off, s[0:3], s33 offset:4 ; 4-byte Folded Spill
666; GFX9-O3-NEXT:    buffer_store_dword v2, off, s[0:3], s33 offset:8 ; 4-byte Folded Spill
667; GFX9-O3-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:12 ; 4-byte Folded Spill
668; GFX9-O3-NEXT:    buffer_store_dword v2, off, s[0:3], s33 offset:16 ; 4-byte Folded Spill
669; GFX9-O3-NEXT:    s_nop 0
670; GFX9-O3-NEXT:    buffer_store_dword v3, off, s[0:3], s33 offset:20 ; 4-byte Folded Spill
671; GFX9-O3-NEXT:    buffer_store_dword v8, off, s[0:3], s33 offset:24 ; 4-byte Folded Spill
672; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
673; GFX9-O3-NEXT:    v_writelane_b32 v8, s30, 0
674; GFX9-O3-NEXT:    s_addk_i32 s32, 0x800
675; GFX9-O3-NEXT:    v_writelane_b32 v8, s31, 1
676; GFX9-O3-NEXT:    v_mov_b32_e32 v0, s9
677; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
678; GFX9-O3-NEXT:    v_cndmask_b32_e64 v6, 0, v0, s[34:35]
679; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
680; GFX9-O3-NEXT:    v_mov_b32_e32 v0, s8
681; GFX9-O3-NEXT:    s_or_saveexec_b64 s[36:37], -1
682; GFX9-O3-NEXT:    s_getpc_b64 s[34:35]
683; GFX9-O3-NEXT:    s_add_u32 s34, s34, strict_wwm_called_i64@gotpcrel32@lo+4
684; GFX9-O3-NEXT:    s_addc_u32 s35, s35, strict_wwm_called_i64@gotpcrel32@hi+12
685; GFX9-O3-NEXT:    s_load_dwordx2 s[34:35], s[34:35], 0x0
686; GFX9-O3-NEXT:    v_cndmask_b32_e64 v7, 0, v0, s[36:37]
687; GFX9-O3-NEXT:    v_mov_b32_e32 v0, v7
688; GFX9-O3-NEXT:    v_mov_b32_e32 v1, v6
689; GFX9-O3-NEXT:    s_waitcnt lgkmcnt(0)
690; GFX9-O3-NEXT:    s_swappc_b64 s[30:31], s[34:35]
691; GFX9-O3-NEXT:    v_mov_b32_e32 v2, v0
692; GFX9-O3-NEXT:    v_mov_b32_e32 v3, v1
693; GFX9-O3-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v7
694; GFX9-O3-NEXT:    v_addc_co_u32_e32 v3, vcc, v3, v6, vcc
695; GFX9-O3-NEXT:    s_mov_b64 exec, s[36:37]
696; GFX9-O3-NEXT:    v_mov_b32_e32 v0, v2
697; GFX9-O3-NEXT:    v_mov_b32_e32 v1, v3
698; GFX9-O3-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0 offset:4
699; GFX9-O3-NEXT:    v_readlane_b32 s31, v8, 1
700; GFX9-O3-NEXT:    v_readlane_b32 s30, v8, 0
701; GFX9-O3-NEXT:    s_mov_b32 s32, s33
702; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
703; GFX9-O3-NEXT:    buffer_load_dword v6, off, s[0:3], s33 ; 4-byte Folded Reload
704; GFX9-O3-NEXT:    buffer_load_dword v7, off, s[0:3], s33 offset:4 ; 4-byte Folded Reload
705; GFX9-O3-NEXT:    buffer_load_dword v2, off, s[0:3], s33 offset:8 ; 4-byte Folded Reload
706; GFX9-O3-NEXT:    buffer_load_dword v3, off, s[0:3], s33 offset:12 ; 4-byte Folded Reload
707; GFX9-O3-NEXT:    buffer_load_dword v2, off, s[0:3], s33 offset:16 ; 4-byte Folded Reload
708; GFX9-O3-NEXT:    buffer_load_dword v3, off, s[0:3], s33 offset:20 ; 4-byte Folded Reload
709; GFX9-O3-NEXT:    buffer_load_dword v8, off, s[0:3], s33 offset:24 ; 4-byte Folded Reload
710; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
711; GFX9-O3-NEXT:    s_mov_b32 s33, s38
712; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
713; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
714  %tmp107 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %arg, i64 0)
715  %tmp134 = call amdgpu_gfx i64 @strict_wwm_called_i64(i64 %tmp107)
716  %tmp136 = add i64 %tmp134, %tmp107
717  %tmp137 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp136)
718  %tmp138 = bitcast i64 %tmp137 to <2 x i32>
719  call void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32> %tmp138, ptr addrspace(8) %tmp14, i32 4, i32 0, i32 0)
720  ret void
721}
722
723define amdgpu_gfx void @strict_wwm_amdgpu_cs_main(<4 x i32> inreg %desc, i32 %index) {
724; GFX9-O0-LABEL: strict_wwm_amdgpu_cs_main:
725; GFX9-O0:       ; %bb.0:
726; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
727; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
728; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 ; 4-byte Folded Spill
729; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
730; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
731; GFX9-O0-NEXT:    s_nop 0
732; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
733; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
734; GFX9-O0-NEXT:    s_mov_b32 s36, s4
735; GFX9-O0-NEXT:    ; kill: def $sgpr36 killed $sgpr36 def $sgpr36_sgpr37_sgpr38_sgpr39
736; GFX9-O0-NEXT:    s_mov_b32 s37, s5
737; GFX9-O0-NEXT:    s_mov_b32 s38, s6
738; GFX9-O0-NEXT:    s_mov_b32 s39, s7
739; GFX9-O0-NEXT:    ; kill: def $sgpr40_sgpr41_sgpr42_sgpr43 killed $sgpr36_sgpr37_sgpr38_sgpr39
740; GFX9-O0-NEXT:    s_mov_b32 s34, 5
741; GFX9-O0-NEXT:    v_lshlrev_b32_e64 v0, s34, v0
742; GFX9-O0-NEXT:    s_mov_b32 s34, 0
743; GFX9-O0-NEXT:    buffer_load_dwordx4 v[11:14], v0, s[36:39], s34 offen
744; GFX9-O0-NEXT:    buffer_load_dwordx2 v[4:5], v0, s[36:39], s34 offen offset:16
745; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)
746; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v11
747; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v12
748; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
749; GFX9-O0-NEXT:    s_mov_b32 s35, 0x7fffffff
750; GFX9-O0-NEXT:    s_mov_b32 s44, -1
751; GFX9-O0-NEXT:    ; kill: def $sgpr44 killed $sgpr44 def $sgpr44_sgpr45
752; GFX9-O0-NEXT:    s_mov_b32 s45, s35
753; GFX9-O0-NEXT:    s_mov_b32 s42, s45
754; GFX9-O0-NEXT:    ; implicit-def: $sgpr46_sgpr47
755; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s42
756; GFX9-O0-NEXT:    v_cndmask_b32_e64 v1, v1, v7, s[40:41]
757; GFX9-O0-NEXT:    s_mov_b32 s35, s44
758; GFX9-O0-NEXT:    ; implicit-def: $sgpr44_sgpr45
759; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s35
760; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s[40:41]
761; GFX9-O0-NEXT:    ; implicit-def: $sgpr43
762; GFX9-O0-NEXT:    ; implicit-def: $sgpr43
763; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
764; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
765; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v2
766; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v3
767; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v13
768; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v14
769; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
770; GFX9-O0-NEXT:    ; implicit-def: $sgpr44_sgpr45
771; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s42
772; GFX9-O0-NEXT:    v_cndmask_b32_e64 v1, v1, v7, s[40:41]
773; GFX9-O0-NEXT:    ; implicit-def: $sgpr44_sgpr45
774; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s35
775; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v6, s[40:41]
776; GFX9-O0-NEXT:    ; implicit-def: $sgpr43
777; GFX9-O0-NEXT:    ; implicit-def: $sgpr43
778; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
779; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
780; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v2
781; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v3
782; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
783; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v5
784; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
785; GFX9-O0-NEXT:    ; implicit-def: $sgpr44_sgpr45
786; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s42
787; GFX9-O0-NEXT:    v_cndmask_b32_e64 v1, v1, v6, s[40:41]
788; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
789; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 killed $vgpr4_vgpr5 killed $exec
790; GFX9-O0-NEXT:    s_or_saveexec_b64 s[40:41], -1
791; GFX9-O0-NEXT:    ; implicit-def: $sgpr42_sgpr43
792; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s35
793; GFX9-O0-NEXT:    v_cndmask_b32_e64 v2, v2, v4, s[40:41]
794; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
795; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
796; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
797; GFX9-O0-NEXT:    s_mov_b64 exec, s[40:41]
798; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v2
799; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v3
800; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v10
801; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v9
802; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v8
803; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v7
804; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
805; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
806; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
807; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
808; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7_vgpr8_vgpr9 killed $exec
809; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v12
810; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v11
811; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v10
812; GFX9-O0-NEXT:    buffer_store_dwordx4 v[6:9], v0, s[36:39], s34 offen
813; GFX9-O0-NEXT:    buffer_store_dwordx2 v[4:5], v0, s[36:39], s34 offen offset:16
814; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
815; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
816; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
817; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
818; GFX9-O0-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
819; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
820; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
821; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
822;
823; GFX9-O3-LABEL: strict_wwm_amdgpu_cs_main:
824; GFX9-O3:       ; %bb.0:
825; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
826; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
827; GFX9-O3-NEXT:    buffer_store_dword v1, off, s[0:3], s32 ; 4-byte Folded Spill
828; GFX9-O3-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Spill
829; GFX9-O3-NEXT:    s_nop 0
830; GFX9-O3-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Spill
831; GFX9-O3-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Spill
832; GFX9-O3-NEXT:    s_nop 0
833; GFX9-O3-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Spill
834; GFX9-O3-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:20 ; 4-byte Folded Spill
835; GFX9-O3-NEXT:    s_nop 0
836; GFX9-O3-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
837; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
838; GFX9-O3-NEXT:    v_lshlrev_b32_e32 v0, 5, v0
839; GFX9-O3-NEXT:    buffer_load_dwordx4 v[8:11], v0, s[4:7], 0 offen
840; GFX9-O3-NEXT:    buffer_load_dwordx2 v[12:13], v0, s[4:7], 0 offen offset:16
841; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
842; GFX9-O3-NEXT:    v_bfrev_b32_e32 v1, -2
843; GFX9-O3-NEXT:    s_waitcnt vmcnt(1)
844; GFX9-O3-NEXT:    v_cndmask_b32_e64 v3, v1, v9, s[34:35]
845; GFX9-O3-NEXT:    v_cndmask_b32_e64 v2, -1, v8, s[34:35]
846; GFX9-O3-NEXT:    v_cndmask_b32_e64 v5, v1, v11, s[34:35]
847; GFX9-O3-NEXT:    v_cndmask_b32_e64 v4, -1, v10, s[34:35]
848; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
849; GFX9-O3-NEXT:    v_cndmask_b32_e64 v7, v1, v13, s[34:35]
850; GFX9-O3-NEXT:    v_cndmask_b32_e64 v6, -1, v12, s[34:35]
851; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
852; GFX9-O3-NEXT:    v_mov_b32_e32 v8, v2
853; GFX9-O3-NEXT:    v_mov_b32_e32 v10, v4
854; GFX9-O3-NEXT:    v_mov_b32_e32 v9, v3
855; GFX9-O3-NEXT:    v_mov_b32_e32 v11, v5
856; GFX9-O3-NEXT:    v_mov_b32_e32 v12, v6
857; GFX9-O3-NEXT:    v_mov_b32_e32 v13, v7
858; GFX9-O3-NEXT:    buffer_store_dwordx4 v[8:11], v0, s[4:7], 0 offen
859; GFX9-O3-NEXT:    buffer_store_dwordx2 v[12:13], v0, s[4:7], 0 offen offset:16
860; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
861; GFX9-O3-NEXT:    buffer_load_dword v1, off, s[0:3], s32 ; 4-byte Folded Reload
862; GFX9-O3-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:4 ; 4-byte Folded Reload
863; GFX9-O3-NEXT:    buffer_load_dword v3, off, s[0:3], s32 offset:8 ; 4-byte Folded Reload
864; GFX9-O3-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:12 ; 4-byte Folded Reload
865; GFX9-O3-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:16 ; 4-byte Folded Reload
866; GFX9-O3-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:20 ; 4-byte Folded Reload
867; GFX9-O3-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
868; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
869; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
870; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
871  %tmp17 = shl i32 %index, 5
872  %tmp18 = tail call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %desc, i32 %tmp17, i32 0)
873  %.i0.upto1.bc = bitcast <4 x i32> %tmp18 to <2 x i64>
874  %tmp19 = or i32 %tmp17, 16
875  %tmp20 = tail call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %desc, i32 %tmp19, i32 0)
876  %.i0.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 0
877  %tmp22 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i0.upto1.extract, i64 9223372036854775807)
878  %tmp97 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp22)
879  %.i1.upto1.extract = extractelement <2 x i64> %.i0.upto1.bc, i32 1
880  %tmp99 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i1.upto1.extract, i64 9223372036854775807)
881  %tmp174 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp99)
882  %.i25 = bitcast <2 x i32> %tmp20 to i64
883  %tmp176 = tail call i64 @llvm.amdgcn.set.inactive.i64(i64 %.i25, i64 9223372036854775807)
884  %tmp251 = tail call i64 @llvm.amdgcn.strict.wwm.i64(i64 %tmp176)
885  %.cast = bitcast i64 %tmp97 to <2 x float>
886  %.cast6 = bitcast i64 %tmp174 to <2 x float>
887  %.cast7 = bitcast i64 %tmp251 to <2 x float>
888  %tmp254 = shufflevector <2 x float> %.cast, <2 x float> %.cast6, <4 x i32> <i32 0, i32 1, i32 2, i32 3>
889  %desc.int = bitcast <4 x i32> %desc to i128
890  %desc.ptr = inttoptr i128 %desc.int to ptr addrspace(8)
891  tail call void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float> %tmp254, ptr addrspace(8) %desc.ptr, i32 %tmp17, i32 0, i32 0)
892  tail call void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float> %.cast7, ptr addrspace(8) %desc.ptr, i32 %tmp19, i32 0, i32 0)
893  ret void
894}
895
896; FIXME: This spills v40 and v41 twice, once in whole-wave-mode and once for the active lanes
897define amdgpu_gfx <32 x i32> @strict_wwm_callee_saves(<32 x i32> inreg %keep, ptr addrspace(5) %ptr, i64 %a, i64 %b, i64 %c, i64 %d, i64 %e) #0 {
898; GFX9-O0-LABEL: strict_wwm_callee_saves:
899; GFX9-O0:       ; %bb.0:
900; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
901; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
902; GFX9-O0-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Spill
903; GFX9-O0-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Spill
904; GFX9-O0-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Spill
905; GFX9-O0-NEXT:    s_nop 0
906; GFX9-O0-NEXT:    buffer_store_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Spill
907; GFX9-O0-NEXT:    s_mov_b64 exec, -1
908; GFX9-O0-NEXT:    buffer_store_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Spill
909; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
910; GFX9-O0-NEXT:    buffer_store_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Spill
911; GFX9-O0-NEXT:    buffer_store_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Spill
912; GFX9-O0-NEXT:    buffer_store_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Spill
913; GFX9-O0-NEXT:    buffer_store_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Spill
914; GFX9-O0-NEXT:    buffer_store_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Spill
915; GFX9-O0-NEXT:    buffer_store_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
916; GFX9-O0-NEXT:    buffer_store_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
917; GFX9-O0-NEXT:    v_writelane_b32 v47, s64, 0
918; GFX9-O0-NEXT:    v_writelane_b32 v47, s65, 1
919; GFX9-O0-NEXT:    v_writelane_b32 v47, s66, 2
920; GFX9-O0-NEXT:    v_writelane_b32 v47, s67, 3
921; GFX9-O0-NEXT:    buffer_store_dword v10, off, s[0:3], s32 offset:68 ; 4-byte Folded Spill
922; GFX9-O0-NEXT:    buffer_store_dword v9, off, s[0:3], s32 offset:60 ; 4-byte Folded Spill
923; GFX9-O0-NEXT:    buffer_store_dword v8, off, s[0:3], s32 offset:76 ; 4-byte Folded Spill
924; GFX9-O0-NEXT:    buffer_store_dword v7, off, s[0:3], s32 offset:64 ; 4-byte Folded Spill
925; GFX9-O0-NEXT:    buffer_store_dword v6, off, s[0:3], s32 offset:84 ; 4-byte Folded Spill
926; GFX9-O0-NEXT:    buffer_store_dword v5, off, s[0:3], s32 offset:72 ; 4-byte Folded Spill
927; GFX9-O0-NEXT:    buffer_store_dword v4, off, s[0:3], s32 offset:92 ; 4-byte Folded Spill
928; GFX9-O0-NEXT:    buffer_store_dword v3, off, s[0:3], s32 offset:80 ; 4-byte Folded Spill
929; GFX9-O0-NEXT:    buffer_store_dword v2, off, s[0:3], s32 offset:52 ; 4-byte Folded Spill
930; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:88 ; 4-byte Folded Spill
931; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:56 ; 4-byte Folded Spill
932; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:20
933; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
934; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:116 ; 4-byte Folded Spill
935; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:16
936; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
937; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:112 ; 4-byte Folded Spill
938; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:12
939; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
940; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:108 ; 4-byte Folded Spill
941; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:8
942; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
943; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:104 ; 4-byte Folded Spill
944; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:4
945; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
946; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:100 ; 4-byte Folded Spill
947; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32
948; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
949; GFX9-O0-NEXT:    buffer_store_dword v0, off, s[0:3], s32 offset:96 ; 4-byte Folded Spill
950; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
951; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
952; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
953; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
954; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
955; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
956; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s4
957; GFX9-O0-NEXT:    v_mov_b32_e32 v35, s5
958; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s6
959; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:168 ; 4-byte Folded Spill
960; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s7
961; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:164 ; 4-byte Folded Spill
962; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s8
963; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:160 ; 4-byte Folded Spill
964; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s9
965; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:156 ; 4-byte Folded Spill
966; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s10
967; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:152 ; 4-byte Folded Spill
968; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s11
969; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:148 ; 4-byte Folded Spill
970; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s12
971; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:144 ; 4-byte Folded Spill
972; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s13
973; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:140 ; 4-byte Folded Spill
974; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s14
975; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:136 ; 4-byte Folded Spill
976; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s15
977; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:132 ; 4-byte Folded Spill
978; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s16
979; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:128 ; 4-byte Folded Spill
980; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s17
981; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:124 ; 4-byte Folded Spill
982; GFX9-O0-NEXT:    v_mov_b32_e32 v40, s18
983; GFX9-O0-NEXT:    v_mov_b32_e32 v39, s19
984; GFX9-O0-NEXT:    v_mov_b32_e32 v38, s20
985; GFX9-O0-NEXT:    v_mov_b32_e32 v37, s21
986; GFX9-O0-NEXT:    v_mov_b32_e32 v36, s22
987; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s23
988; GFX9-O0-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:120 ; 4-byte Folded Spill
989; GFX9-O0-NEXT:    v_mov_b32_e32 v46, s24
990; GFX9-O0-NEXT:    v_mov_b32_e32 v45, s25
991; GFX9-O0-NEXT:    v_mov_b32_e32 v44, s26
992; GFX9-O0-NEXT:    v_mov_b32_e32 v43, s27
993; GFX9-O0-NEXT:    v_mov_b32_e32 v42, s28
994; GFX9-O0-NEXT:    v_mov_b32_e32 v41, s29
995; GFX9-O0-NEXT:    ; kill: def $vgpr0 killed $vgpr0 def $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 killed $exec
996; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v35
997; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:168 ; 4-byte Folded Reload
998; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
999; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v35
1000; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:164 ; 4-byte Folded Reload
1001; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1002; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v35
1003; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:160 ; 4-byte Folded Reload
1004; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1005; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v35
1006; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:156 ; 4-byte Folded Reload
1007; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1008; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v35
1009; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:152 ; 4-byte Folded Reload
1010; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1011; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v35
1012; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:148 ; 4-byte Folded Reload
1013; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1014; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v35
1015; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:144 ; 4-byte Folded Reload
1016; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1017; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v35
1018; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:140 ; 4-byte Folded Reload
1019; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1020; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v35
1021; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:136 ; 4-byte Folded Reload
1022; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1023; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v35
1024; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:132 ; 4-byte Folded Reload
1025; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1026; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v35
1027; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:128 ; 4-byte Folded Reload
1028; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1029; GFX9-O0-NEXT:    v_mov_b32_e32 v12, v35
1030; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:124 ; 4-byte Folded Reload
1031; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1032; GFX9-O0-NEXT:    v_mov_b32_e32 v13, v35
1033; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:120 ; 4-byte Folded Reload
1034; GFX9-O0-NEXT:    v_mov_b32_e32 v14, v40
1035; GFX9-O0-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
1036; GFX9-O0-NEXT:    v_mov_b32_e32 v15, v39
1037; GFX9-O0-NEXT:    buffer_load_dword v39, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
1038; GFX9-O0-NEXT:    v_mov_b32_e32 v16, v38
1039; GFX9-O0-NEXT:    buffer_load_dword v38, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
1040; GFX9-O0-NEXT:    v_mov_b32_e32 v17, v37
1041; GFX9-O0-NEXT:    buffer_load_dword v37, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
1042; GFX9-O0-NEXT:    v_mov_b32_e32 v18, v36
1043; GFX9-O0-NEXT:    buffer_load_dword v36, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
1044; GFX9-O0-NEXT:    s_waitcnt vmcnt(5)
1045; GFX9-O0-NEXT:    v_mov_b32_e32 v19, v35
1046; GFX9-O0-NEXT:    buffer_load_dword v35, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
1047; GFX9-O0-NEXT:    v_mov_b32_e32 v20, v46
1048; GFX9-O0-NEXT:    v_mov_b32_e32 v21, v45
1049; GFX9-O0-NEXT:    v_mov_b32_e32 v22, v44
1050; GFX9-O0-NEXT:    v_mov_b32_e32 v23, v43
1051; GFX9-O0-NEXT:    v_mov_b32_e32 v24, v42
1052; GFX9-O0-NEXT:    v_mov_b32_e32 v25, v41
1053; GFX9-O0-NEXT:    s_waitcnt vmcnt(5)
1054; GFX9-O0-NEXT:    v_mov_b32_e32 v26, v40
1055; GFX9-O0-NEXT:    s_waitcnt vmcnt(4)
1056; GFX9-O0-NEXT:    v_mov_b32_e32 v27, v39
1057; GFX9-O0-NEXT:    s_waitcnt vmcnt(3)
1058; GFX9-O0-NEXT:    v_mov_b32_e32 v28, v38
1059; GFX9-O0-NEXT:    s_waitcnt vmcnt(2)
1060; GFX9-O0-NEXT:    v_mov_b32_e32 v29, v37
1061; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)
1062; GFX9-O0-NEXT:    v_mov_b32_e32 v30, v36
1063; GFX9-O0-NEXT:    ; kill: def $vgpr31 killed $vgpr35 killed $exec
1064; GFX9-O0-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:116 ; 4-byte Folded Reload
1065; GFX9-O0-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:112 ; 4-byte Folded Reload
1066; GFX9-O0-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:108 ; 4-byte Folded Reload
1067; GFX9-O0-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:104 ; 4-byte Folded Reload
1068; GFX9-O0-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:100 ; 4-byte Folded Reload
1069; GFX9-O0-NEXT:    buffer_load_dword v26, off, s[0:3], s32 offset:96 ; 4-byte Folded Reload
1070; GFX9-O0-NEXT:    buffer_load_dword v11, off, s[0:3], s32 offset:92 ; 4-byte Folded Reload
1071; GFX9-O0-NEXT:    buffer_load_dword v10, off, s[0:3], s32 offset:88 ; 4-byte Folded Reload
1072; GFX9-O0-NEXT:    buffer_load_dword v9, off, s[0:3], s32 offset:84 ; 4-byte Folded Reload
1073; GFX9-O0-NEXT:    buffer_load_dword v8, off, s[0:3], s32 offset:80 ; 4-byte Folded Reload
1074; GFX9-O0-NEXT:    buffer_load_dword v7, off, s[0:3], s32 offset:76 ; 4-byte Folded Reload
1075; GFX9-O0-NEXT:    buffer_load_dword v6, off, s[0:3], s32 offset:72 ; 4-byte Folded Reload
1076; GFX9-O0-NEXT:    buffer_load_dword v5, off, s[0:3], s32 offset:68 ; 4-byte Folded Reload
1077; GFX9-O0-NEXT:    buffer_load_dword v4, off, s[0:3], s32 offset:64 ; 4-byte Folded Reload
1078; GFX9-O0-NEXT:    buffer_load_dword v2, off, s[0:3], s32 offset:60 ; 4-byte Folded Reload
1079; GFX9-O0-NEXT:    buffer_load_dword v1, off, s[0:3], s32 offset:56 ; 4-byte Folded Reload
1080; GFX9-O0-NEXT:    buffer_load_dword v0, off, s[0:3], s32 offset:52 ; 4-byte Folded Reload
1081; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1082; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1083; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
1084; GFX9-O0-NEXT:    s_waitcnt vmcnt(4)
1085; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v5
1086; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1087; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1088; GFX9-O0-NEXT:    ; kill: def $vgpr4 killed $vgpr4 def $vgpr4_vgpr5 killed $exec
1089; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v7
1090; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1091; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1092; GFX9-O0-NEXT:    ; kill: def $vgpr6 killed $vgpr6 def $vgpr6_vgpr7 killed $exec
1093; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v9
1094; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1095; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1096; GFX9-O0-NEXT:    ; kill: def $vgpr8 killed $vgpr8 def $vgpr8_vgpr9 killed $exec
1097; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v11
1098; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1099; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
1100; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 def $vgpr10_vgpr11 killed $exec
1101; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1102; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v0
1103; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
1104; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
1105; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
1106; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
1107; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
1108; GFX9-O0-NEXT:    ; implicit-def: $sgpr36_sgpr37_sgpr38_sgpr39_sgpr40_sgpr41_sgpr42_sgpr43_sgpr44_sgpr45_sgpr46_sgpr47_sgpr48_sgpr49_sgpr50_sgpr51_sgpr52_sgpr53_sgpr54_sgpr55_sgpr56_sgpr57_sgpr58_sgpr59_sgpr60_sgpr61_sgpr62_sgpr63_sgpr64_sgpr65_sgpr66_sgpr67
1109; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v11
1110; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1111; GFX9-O0-NEXT:    s_mov_b64 s[38:39], 0
1112; GFX9-O0-NEXT:    s_mov_b32 s37, s39
1113; GFX9-O0-NEXT:    ; implicit-def: $sgpr40_sgpr41
1114; GFX9-O0-NEXT:    v_mov_b32_e32 v32, s37
1115; GFX9-O0-NEXT:    v_cndmask_b32_e64 v32, v32, v0, s[34:35]
1116; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1117; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v10
1118; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1119; GFX9-O0-NEXT:    s_mov_b32 s36, s38
1120; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1121; GFX9-O0-NEXT:    v_mov_b32_e32 v33, s36
1122; GFX9-O0-NEXT:    v_cndmask_b32_e64 v33, v33, v0, s[34:35]
1123; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1124; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1125; GFX9-O0-NEXT:    v_mov_b32_e32 v34, v32
1126; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1127; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v33
1128; GFX9-O0-NEXT:    v_mov_b32_e32 v11, v34
1129; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v9
1130; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1131; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1132; GFX9-O0-NEXT:    v_mov_b32_e32 v32, s37
1133; GFX9-O0-NEXT:    v_cndmask_b32_e64 v32, v32, v0, s[34:35]
1134; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1135; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v8
1136; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1137; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1138; GFX9-O0-NEXT:    v_mov_b32_e32 v33, s36
1139; GFX9-O0-NEXT:    v_cndmask_b32_e64 v33, v33, v0, s[34:35]
1140; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1141; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1142; GFX9-O0-NEXT:    v_mov_b32_e32 v34, v32
1143; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1144; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v33
1145; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v34
1146; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v7
1147; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1148; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1149; GFX9-O0-NEXT:    v_mov_b32_e32 v32, s37
1150; GFX9-O0-NEXT:    v_cndmask_b32_e64 v32, v32, v0, s[34:35]
1151; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1152; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v6
1153; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1154; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1155; GFX9-O0-NEXT:    v_mov_b32_e32 v33, s36
1156; GFX9-O0-NEXT:    v_cndmask_b32_e64 v33, v33, v0, s[34:35]
1157; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1158; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1159; GFX9-O0-NEXT:    v_mov_b32_e32 v34, v32
1160; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1161; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v33
1162; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v34
1163; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v5
1164; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1165; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1166; GFX9-O0-NEXT:    v_mov_b32_e32 v32, s37
1167; GFX9-O0-NEXT:    v_cndmask_b32_e64 v32, v32, v0, s[34:35]
1168; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1169; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v4
1170; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1171; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1172; GFX9-O0-NEXT:    v_mov_b32_e32 v33, s36
1173; GFX9-O0-NEXT:    v_cndmask_b32_e64 v33, v33, v0, s[34:35]
1174; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1175; GFX9-O0-NEXT:    ; implicit-def: $sgpr38
1176; GFX9-O0-NEXT:    v_mov_b32_e32 v34, v32
1177; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1178; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v33
1179; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v34
1180; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v3
1181; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1182; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1183; GFX9-O0-NEXT:    v_mov_b32_e32 v32, s37
1184; GFX9-O0-NEXT:    v_cndmask_b32_e64 v32, v32, v0, s[34:35]
1185; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1186; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v2
1187; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1
1188; GFX9-O0-NEXT:    ; implicit-def: $sgpr38_sgpr39
1189; GFX9-O0-NEXT:    v_mov_b32_e32 v33, s36
1190; GFX9-O0-NEXT:    v_cndmask_b32_e64 v33, v33, v0, s[34:35]
1191; GFX9-O0-NEXT:    ; implicit-def: $sgpr36
1192; GFX9-O0-NEXT:    ; implicit-def: $sgpr36
1193; GFX9-O0-NEXT:    v_mov_b32_e32 v34, v32
1194; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1195; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v33
1196; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v34
1197; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v11
1198; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:4
1199; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v10
1200; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen
1201; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v9
1202; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:12
1203; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v8
1204; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:8
1205; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v7
1206; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:20
1207; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v6
1208; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:16
1209; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v5
1210; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:28
1211; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v4
1212; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:24
1213; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v3
1214; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:36
1215; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v2
1216; GFX9-O0-NEXT:    buffer_store_dword v0, v1, s[0:3], 0 offen offset:32
1217; GFX9-O0-NEXT:    v_mov_b32_e32 v0, s4
1218; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s5
1219; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s6
1220; GFX9-O0-NEXT:    v_mov_b32_e32 v3, s7
1221; GFX9-O0-NEXT:    v_mov_b32_e32 v4, s8
1222; GFX9-O0-NEXT:    v_mov_b32_e32 v5, s9
1223; GFX9-O0-NEXT:    v_mov_b32_e32 v6, s10
1224; GFX9-O0-NEXT:    v_mov_b32_e32 v7, s11
1225; GFX9-O0-NEXT:    v_mov_b32_e32 v8, s12
1226; GFX9-O0-NEXT:    v_mov_b32_e32 v9, s13
1227; GFX9-O0-NEXT:    v_mov_b32_e32 v10, s14
1228; GFX9-O0-NEXT:    v_mov_b32_e32 v11, s15
1229; GFX9-O0-NEXT:    v_mov_b32_e32 v12, s16
1230; GFX9-O0-NEXT:    v_mov_b32_e32 v13, s17
1231; GFX9-O0-NEXT:    v_mov_b32_e32 v14, s18
1232; GFX9-O0-NEXT:    v_mov_b32_e32 v15, s19
1233; GFX9-O0-NEXT:    v_mov_b32_e32 v16, s20
1234; GFX9-O0-NEXT:    v_mov_b32_e32 v17, s21
1235; GFX9-O0-NEXT:    v_mov_b32_e32 v18, s22
1236; GFX9-O0-NEXT:    v_mov_b32_e32 v19, s23
1237; GFX9-O0-NEXT:    v_mov_b32_e32 v20, s24
1238; GFX9-O0-NEXT:    v_mov_b32_e32 v21, s25
1239; GFX9-O0-NEXT:    v_mov_b32_e32 v22, s26
1240; GFX9-O0-NEXT:    v_mov_b32_e32 v23, s27
1241; GFX9-O0-NEXT:    v_mov_b32_e32 v24, s28
1242; GFX9-O0-NEXT:    v_mov_b32_e32 v25, s29
1243; GFX9-O0-NEXT:    v_readlane_b32 s67, v47, 3
1244; GFX9-O0-NEXT:    v_readlane_b32 s66, v47, 2
1245; GFX9-O0-NEXT:    v_readlane_b32 s65, v47, 1
1246; GFX9-O0-NEXT:    v_readlane_b32 s64, v47, 0
1247; GFX9-O0-NEXT:    buffer_load_dword v46, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
1248; GFX9-O0-NEXT:    buffer_load_dword v45, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
1249; GFX9-O0-NEXT:    buffer_load_dword v44, off, s[0:3], s32 offset:32 ; 4-byte Folded Reload
1250; GFX9-O0-NEXT:    buffer_load_dword v43, off, s[0:3], s32 offset:36 ; 4-byte Folded Reload
1251; GFX9-O0-NEXT:    buffer_load_dword v42, off, s[0:3], s32 offset:40 ; 4-byte Folded Reload
1252; GFX9-O0-NEXT:    buffer_load_dword v41, off, s[0:3], s32 offset:44 ; 4-byte Folded Reload
1253; GFX9-O0-NEXT:    buffer_load_dword v40, off, s[0:3], s32 offset:48 ; 4-byte Folded Reload
1254; GFX9-O0-NEXT:    s_xor_saveexec_b64 s[34:35], -1
1255; GFX9-O0-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:176 ; 4-byte Folded Reload
1256; GFX9-O0-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:180 ; 4-byte Folded Reload
1257; GFX9-O0-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:184 ; 4-byte Folded Reload
1258; GFX9-O0-NEXT:    buffer_load_dword v34, off, s[0:3], s32 offset:188 ; 4-byte Folded Reload
1259; GFX9-O0-NEXT:    s_mov_b64 exec, -1
1260; GFX9-O0-NEXT:    buffer_load_dword v47, off, s[0:3], s32 offset:172 ; 4-byte Folded Reload
1261; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
1262; GFX9-O0-NEXT:    s_waitcnt vmcnt(0)
1263; GFX9-O0-NEXT:    s_setpc_b64 s[30:31]
1264;
1265; GFX9-O3-LABEL: strict_wwm_callee_saves:
1266; GFX9-O3:       ; %bb.0:
1267; GFX9-O3-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
1268; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
1269; GFX9-O3-NEXT:    buffer_store_dword v32, off, s[0:3], s32 offset:24 ; 4-byte Folded Spill
1270; GFX9-O3-NEXT:    s_nop 0
1271; GFX9-O3-NEXT:    buffer_store_dword v33, off, s[0:3], s32 offset:28 ; 4-byte Folded Spill
1272; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1273; GFX9-O3-NEXT:    buffer_load_dword v26, off, s[0:3], s32
1274; GFX9-O3-NEXT:    buffer_load_dword v27, off, s[0:3], s32 offset:4
1275; GFX9-O3-NEXT:    buffer_load_dword v28, off, s[0:3], s32 offset:8
1276; GFX9-O3-NEXT:    buffer_load_dword v29, off, s[0:3], s32 offset:12
1277; GFX9-O3-NEXT:    buffer_load_dword v30, off, s[0:3], s32 offset:16
1278; GFX9-O3-NEXT:    buffer_load_dword v31, off, s[0:3], s32 offset:20
1279; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
1280; GFX9-O3-NEXT:    v_cndmask_b32_e64 v33, 0, v2, s[34:35]
1281; GFX9-O3-NEXT:    v_cndmask_b32_e64 v32, 0, v1, s[34:35]
1282; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1283; GFX9-O3-NEXT:    v_mov_b32_e32 v1, v32
1284; GFX9-O3-NEXT:    v_mov_b32_e32 v2, v33
1285; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
1286; GFX9-O3-NEXT:    v_cndmask_b32_e64 v33, 0, v4, s[34:35]
1287; GFX9-O3-NEXT:    v_cndmask_b32_e64 v32, 0, v3, s[34:35]
1288; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1289; GFX9-O3-NEXT:    v_mov_b32_e32 v3, v32
1290; GFX9-O3-NEXT:    v_mov_b32_e32 v4, v33
1291; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
1292; GFX9-O3-NEXT:    v_cndmask_b32_e64 v33, 0, v6, s[34:35]
1293; GFX9-O3-NEXT:    v_cndmask_b32_e64 v32, 0, v5, s[34:35]
1294; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1295; GFX9-O3-NEXT:    v_mov_b32_e32 v5, v32
1296; GFX9-O3-NEXT:    v_mov_b32_e32 v6, v33
1297; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
1298; GFX9-O3-NEXT:    v_cndmask_b32_e64 v33, 0, v8, s[34:35]
1299; GFX9-O3-NEXT:    v_cndmask_b32_e64 v32, 0, v7, s[34:35]
1300; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1301; GFX9-O3-NEXT:    v_mov_b32_e32 v7, v32
1302; GFX9-O3-NEXT:    v_mov_b32_e32 v8, v33
1303; GFX9-O3-NEXT:    s_or_saveexec_b64 s[34:35], -1
1304; GFX9-O3-NEXT:    v_cndmask_b32_e64 v33, 0, v10, s[34:35]
1305; GFX9-O3-NEXT:    v_cndmask_b32_e64 v32, 0, v9, s[34:35]
1306; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1307; GFX9-O3-NEXT:    v_mov_b32_e32 v9, v32
1308; GFX9-O3-NEXT:    v_mov_b32_e32 v10, v33
1309; GFX9-O3-NEXT:    buffer_store_dword v2, v0, s[0:3], 0 offen offset:4
1310; GFX9-O3-NEXT:    buffer_store_dword v1, v0, s[0:3], 0 offen
1311; GFX9-O3-NEXT:    buffer_store_dword v4, v0, s[0:3], 0 offen offset:12
1312; GFX9-O3-NEXT:    buffer_store_dword v3, v0, s[0:3], 0 offen offset:8
1313; GFX9-O3-NEXT:    buffer_store_dword v6, v0, s[0:3], 0 offen offset:20
1314; GFX9-O3-NEXT:    buffer_store_dword v5, v0, s[0:3], 0 offen offset:16
1315; GFX9-O3-NEXT:    buffer_store_dword v8, v0, s[0:3], 0 offen offset:28
1316; GFX9-O3-NEXT:    buffer_store_dword v7, v0, s[0:3], 0 offen offset:24
1317; GFX9-O3-NEXT:    buffer_store_dword v10, v0, s[0:3], 0 offen offset:36
1318; GFX9-O3-NEXT:    buffer_store_dword v9, v0, s[0:3], 0 offen offset:32
1319; GFX9-O3-NEXT:    v_mov_b32_e32 v0, s4
1320; GFX9-O3-NEXT:    v_mov_b32_e32 v1, s5
1321; GFX9-O3-NEXT:    v_mov_b32_e32 v2, s6
1322; GFX9-O3-NEXT:    v_mov_b32_e32 v3, s7
1323; GFX9-O3-NEXT:    v_mov_b32_e32 v4, s8
1324; GFX9-O3-NEXT:    v_mov_b32_e32 v5, s9
1325; GFX9-O3-NEXT:    v_mov_b32_e32 v6, s10
1326; GFX9-O3-NEXT:    v_mov_b32_e32 v7, s11
1327; GFX9-O3-NEXT:    v_mov_b32_e32 v8, s12
1328; GFX9-O3-NEXT:    v_mov_b32_e32 v9, s13
1329; GFX9-O3-NEXT:    v_mov_b32_e32 v10, s14
1330; GFX9-O3-NEXT:    v_mov_b32_e32 v11, s15
1331; GFX9-O3-NEXT:    v_mov_b32_e32 v12, s16
1332; GFX9-O3-NEXT:    v_mov_b32_e32 v13, s17
1333; GFX9-O3-NEXT:    v_mov_b32_e32 v14, s18
1334; GFX9-O3-NEXT:    v_mov_b32_e32 v15, s19
1335; GFX9-O3-NEXT:    v_mov_b32_e32 v16, s20
1336; GFX9-O3-NEXT:    v_mov_b32_e32 v17, s21
1337; GFX9-O3-NEXT:    v_mov_b32_e32 v18, s22
1338; GFX9-O3-NEXT:    v_mov_b32_e32 v19, s23
1339; GFX9-O3-NEXT:    v_mov_b32_e32 v20, s24
1340; GFX9-O3-NEXT:    v_mov_b32_e32 v21, s25
1341; GFX9-O3-NEXT:    v_mov_b32_e32 v22, s26
1342; GFX9-O3-NEXT:    v_mov_b32_e32 v23, s27
1343; GFX9-O3-NEXT:    v_mov_b32_e32 v24, s28
1344; GFX9-O3-NEXT:    v_mov_b32_e32 v25, s29
1345; GFX9-O3-NEXT:    s_xor_saveexec_b64 s[34:35], -1
1346; GFX9-O3-NEXT:    buffer_load_dword v32, off, s[0:3], s32 offset:24 ; 4-byte Folded Reload
1347; GFX9-O3-NEXT:    buffer_load_dword v33, off, s[0:3], s32 offset:28 ; 4-byte Folded Reload
1348; GFX9-O3-NEXT:    s_mov_b64 exec, s[34:35]
1349; GFX9-O3-NEXT:    s_waitcnt vmcnt(0)
1350; GFX9-O3-NEXT:    s_setpc_b64 s[30:31]
1351  %a2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %a, i64 0)
1352  %a2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %a2.i)
1353  %b2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %b, i64 0)
1354  %b2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %b2.i)
1355  %c2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %c, i64 0)
1356  %c2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %c2.i)
1357  %d2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %d, i64 0)
1358  %d2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %d2.i)
1359  %e2.i = call i64 @llvm.amdgcn.set.inactive.i64(i64 %e, i64 0)
1360  %e2 = call i64 @llvm.amdgcn.strict.wwm.i64(i64 %e2.i)
1361  store i64 %a2, ptr addrspace(5) %ptr
1362  %ptr_b = getelementptr i64, ptr addrspace(5) %ptr, i32 1
1363  store i64 %b2, ptr addrspace(5) %ptr_b
1364  %ptr_c = getelementptr i64, ptr addrspace(5) %ptr, i32 2
1365  store i64 %c2, ptr addrspace(5) %ptr_c
1366  %ptr_d = getelementptr i64, ptr addrspace(5) %ptr, i32 3
1367  store i64 %d2, ptr addrspace(5) %ptr_d
1368  %ptr_e = getelementptr i64, ptr addrspace(5) %ptr, i32 4
1369  store i64 %e2, ptr addrspace(5) %ptr_e
1370  ret <32 x i32> %keep
1371}
1372
1373declare i32 @llvm.amdgcn.strict.wwm.i32(i32)
1374declare i64 @llvm.amdgcn.strict.wwm.i64(i64)
1375declare i32 @llvm.amdgcn.set.inactive.i32(i32, i32)
1376declare i64 @llvm.amdgcn.set.inactive.i64(i64, i64)
1377declare i32 @llvm.amdgcn.update.dpp.i32(i32, i32, i32, i32, i32, i1)
1378declare <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.v2f32(ptr addrspace(8), i32, i32, i32)
1379declare void @llvm.amdgcn.raw.ptr.buffer.store.f32(float, ptr addrspace(8), i32, i32, i32)
1380declare void @llvm.amdgcn.raw.ptr.buffer.store.i32(i32, ptr addrspace(8), i32, i32, i32)
1381declare void @llvm.amdgcn.raw.ptr.buffer.store.v2i32(<2 x i32>, ptr addrspace(8), i32, i32, i32)
1382declare void @llvm.amdgcn.raw.ptr.buffer.store.v2f32(<2 x float>, ptr addrspace(8), i32, i32, i32)
1383declare void @llvm.amdgcn.raw.ptr.buffer.store.v4f32(<4 x float>, ptr addrspace(8), i32, i32, i32)
1384declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)
1385declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)
1386
1387attributes #0 = { "amdgpu-waves-per-eu"="5,5" }
1388