1; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -enable-ipra=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -enable-ipra=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-ipra=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
4
5declare void @external_void_func_void() #0
6
7; GCN-LABEL: {{^}}test_kernel_call_external_void_func_void_clobber_s30_s31_call_external_void_func_void:
8; GCN: s_mov_b32 s33, s7
9; GCN: s_getpc_b64 s[34:35]
10; GCN-NEXT: s_add_u32 s34, s34,
11; GCN-NEXT: s_addc_u32 s35, s35,
12; GCN-NEXT: s_mov_b32 s4, s33
13; GCN-NEXT: s_mov_b32 s32, s33
14; GCN: s_swappc_b64 s[30:31], s[34:35]
15
16; GCN-NEXT: s_mov_b32 s4, s33
17; GCN-NEXT: #ASMSTART
18; GCN-NEXT: #ASMEND
19; GCN-NEXT: s_swappc_b64 s[30:31], s[34:35]
20define amdgpu_kernel void @test_kernel_call_external_void_func_void_clobber_s30_s31_call_external_void_func_void() #0 {
21  call void @external_void_func_void()
22  call void asm sideeffect "", ""() #0
23  call void @external_void_func_void()
24  ret void
25}
26
27; GCN-LABEL: {{^}}test_func_call_external_void_func_void_clobber_s30_s31_call_external_void_func_void:
28; GCN: v_writelane_b32 v32, s33, 0
29; GCN: v_writelane_b32 v32, s34, 1
30; GCN: v_writelane_b32 v32, s35, 2
31; GCN: v_writelane_b32 v32, s36, 3
32; GCN: v_writelane_b32 v32, s37, 4
33
34; GCN: s_mov_b32 s33, s5
35; GCN-NEXT: s_swappc_b64
36; GCN-NEXT: s_mov_b32 s5, s33
37; GCN-NEXT: s_mov_b32 s33, s5
38; GCN-NEXT: ;;#ASMSTART
39; GCN-NEXT: ;;#ASMEND
40; GCN-NEXT: s_swappc_b64
41; GCN-NEXT: s_mov_b32 s5, s33
42; GCN: v_readlane_b32 s37, v32, 4
43; GCN: v_readlane_b32 s36, v32, 3
44; GCN: v_readlane_b32 s35, v32, 2
45; GCN: v_readlane_b32 s34, v32, 1
46; GCN: v_readlane_b32 s33, v32, 0
47; GCN: s_setpc_b64
48define void @test_func_call_external_void_func_void_clobber_s30_s31_call_external_void_func_void() #0 {
49  call void @external_void_func_void()
50  call void asm sideeffect "", ""() #0
51  call void @external_void_func_void()
52  ret void
53}
54
55; FIXME: Avoid extra restore of FP in between calls.
56; GCN-LABEL: {{^}}test_func_call_external_void_funcx2:
57; GCN: s_mov_b32 s33, s5
58; GCN-NEXT: s_swappc_b64
59; GCN-NEXT: s_mov_b32 s5, s33
60; GCN-NEXT: s_mov_b32 s33, s5
61; GCN-NEXT: s_swappc_b64
62; GCN-NEXT: s_mov_b32 s5, s33
63define void @test_func_call_external_void_funcx2() #0 {
64  call void @external_void_func_void()
65  call void @external_void_func_void()
66  ret void
67}
68
69; GCN-LABEL: {{^}}void_func_void_clobber_s30_s31:
70; GCN: s_waitcnt
71; GCN-NEXT: s_mov_b64 [[SAVEPC:s\[[0-9]+:[0-9]+\]]], s[30:31]
72; GCN-NEXT: #ASMSTART
73; GCN: ; clobber
74; GCN-NEXT: #ASMEND
75; GCN-NEXT: s_mov_b64 s[30:31], [[SAVEPC]]
76; GCN-NEXT: s_setpc_b64 s[30:31]
77define void @void_func_void_clobber_s30_s31() #2 {
78  call void asm sideeffect "; clobber", "~{s[30:31]}"() #0
79  ret void
80}
81
82; GCN-LABEL: {{^}}void_func_void_clobber_vcc:
83; GCN: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
84; GCN-NEXT: ;;#ASMSTART
85; GCN-NEXT: ;;#ASMEND
86; GCN-NEXT: s_setpc_b64 s[30:31]
87define void @void_func_void_clobber_vcc() #2 {
88  call void asm sideeffect "", "~{VCC}"() #0
89  ret void
90}
91
92; GCN-LABEL: {{^}}test_call_void_func_void_clobber_vcc:
93; GCN: s_getpc_b64
94; GCN-NEXT: s_add_u32
95; GCN-NEXT: s_addc_u32
96; GCN: s_mov_b64 s[34:35], vcc
97; GCN-NEXT: s_swappc_b64
98; GCN: s_mov_b64 vcc, s[34:35]
99define amdgpu_kernel void @test_call_void_func_void_clobber_vcc(i32 addrspace(1)* %out) #0 {
100  %vcc = call i64 asm sideeffect "; def $0", "={vcc}"()
101  call void @void_func_void_clobber_vcc()
102  %val0 = load volatile i32, i32 addrspace(1)* undef
103  %val1 = load volatile i32, i32 addrspace(1)* undef
104  call void asm sideeffect "; use $0", "{vcc}"(i64 %vcc)
105  ret void
106}
107
108; GCN-LABEL: {{^}}test_call_void_func_void_mayclobber_s31:
109; GCN: s_mov_b32 s33, s31
110; GCN-NEXT: s_swappc_b64
111; GCN-NEXT: s_mov_b32 s31, s33
112define amdgpu_kernel void @test_call_void_func_void_mayclobber_s31(i32 addrspace(1)* %out) #0 {
113  %s31 = call i32 asm sideeffect "; def $0", "={s31}"()
114  call void @external_void_func_void()
115  call void asm sideeffect "; use $0", "{s31}"(i32 %s31)
116  ret void
117}
118
119; GCN-LABEL: {{^}}test_call_void_func_void_mayclobber_v31:
120; GCN: v_mov_b32_e32 v32, v31
121; GCN-NEXT: s_swappc_b64
122; GCN-NEXT: v_mov_b32_e32 v31, v32
123define amdgpu_kernel void @test_call_void_func_void_mayclobber_v31(i32 addrspace(1)* %out) #0 {
124  %v31 = call i32 asm sideeffect "; def $0", "={v31}"()
125  call void @external_void_func_void()
126  call void asm sideeffect "; use $0", "{v31}"(i32 %v31)
127  ret void
128}
129
130; GCN-LABEL: {{^}}test_call_void_func_void_preserves_s33:
131; GCN: s_mov_b32 s34, s9
132; GCN: ; def s33
133; GCN-NEXT: #ASMEND
134; GCN: s_getpc_b64 s[6:7]
135; GCN-NEXT: s_add_u32 s6, s6, external_void_func_void@rel32@lo+4
136; GCN-NEXT: s_addc_u32 s7, s7, external_void_func_void@rel32@hi+4
137; GCN-NEXT: s_mov_b32 s4, s34
138; GCN-NEXT: s_mov_b32 s32, s34
139; GCN-NEXT: s_swappc_b64 s[30:31], s[6:7]
140; GCN-NEXT: ;;#ASMSTART
141; GCN-NEXT: ; use s33
142; GCN-NEXT: ;;#ASMEND
143; GCN-NEXT: s_endpgm
144define amdgpu_kernel void @test_call_void_func_void_preserves_s33(i32 addrspace(1)* %out) #0 {
145  %s33 = call i32 asm sideeffect "; def $0", "={s33}"()
146  call void @external_void_func_void()
147  call void asm sideeffect "; use $0", "{s33}"(i32 %s33)
148  ret void
149}
150
151; GCN-LABEL: {{^}}test_call_void_func_void_preserves_v32:
152; GCN: s_mov_b32 s33, s9
153; GCN: ; def v32
154; GCN-NEXT: #ASMEND
155; GCN: s_getpc_b64 s[6:7]
156; GCN-NEXT: s_add_u32 s6, s6, external_void_func_void@rel32@lo+4
157; GCN-NEXT: s_addc_u32 s7, s7, external_void_func_void@rel32@hi+4
158; GCN-NEXT: s_mov_b32 s4, s33
159; GCN-NEXT: s_mov_b32 s32, s33
160; GCN-NEXT: s_swappc_b64 s[30:31], s[6:7]
161; GCN-NEXT: ;;#ASMSTART
162; GCN-NEXT: ; use v32
163; GCN-NEXT: ;;#ASMEND
164; GCN-NEXT: s_endpgm
165define amdgpu_kernel void @test_call_void_func_void_preserves_v32(i32 addrspace(1)* %out) #0 {
166  %v32 = call i32 asm sideeffect "; def $0", "={v32}"()
167  call void @external_void_func_void()
168  call void asm sideeffect "; use $0", "{v32}"(i32 %v32)
169  ret void
170}
171
172; GCN-LABEL: {{^}}void_func_void_clobber_s33:
173; GCN: v_writelane_b32 v0, s33, 0
174; GCN-NEXT: #ASMSTART
175; GCN-NEXT: ; clobber
176; GCN-NEXT: #ASMEND
177; GCN-NEXT:	v_readlane_b32 s33, v0, 0
178; GCN-NEXT: s_setpc_b64
179define void @void_func_void_clobber_s33() #2 {
180  call void asm sideeffect "; clobber", "~{s33}"() #0
181  ret void
182}
183
184; GCN-LABEL: {{^}}test_call_void_func_void_clobber_s33:
185; GCN: s_mov_b32 s33, s7
186; GCN: s_getpc_b64
187; GCN-NEXT: s_add_u32
188; GCN-NEXT: s_addc_u32
189; GCN-NEXT: s_mov_b32 s4, s33
190; GCN-NEXT: s_mov_b32 s32, s33
191; GCN: s_swappc_b64
192; GCN-NEXT: s_endpgm
193define amdgpu_kernel void @test_call_void_func_void_clobber_s33() #0 {
194  call void @void_func_void_clobber_s33()
195  ret void
196}
197
198; GCN-LABEL: {{^}}callee_saved_sgpr_func:
199; GCN-NOT: s40
200; GCN: v_writelane_b32 v32, s40
201; GCN: s_swappc_b64
202; GCN-NOT: s40
203; GCN: ; use s40
204; GCN-NOT: s40
205; GCN: v_readlane_b32 s40, v32
206; GCN-NOT: s40
207define void @callee_saved_sgpr_func() #2 {
208  %s40 = call i32 asm sideeffect "; def s40", "={s40}"() #0
209  call void @external_void_func_void()
210  call void asm sideeffect "; use $0", "s"(i32 %s40) #0
211  ret void
212}
213
214; GCN-LABEL: {{^}}callee_saved_sgpr_kernel:
215; GCN-NOT: s40
216; GCN: ; def s40
217; GCN-NOT: s40
218; GCN: s_swappc_b64
219; GCN-NOT: s40
220; GCN: ; use s40
221; GCN-NOT: s40
222define amdgpu_kernel void @callee_saved_sgpr_kernel() #2 {
223  %s40 = call i32 asm sideeffect "; def s40", "={s40}"() #0
224  call void @external_void_func_void()
225  call void asm sideeffect "; use $0", "s"(i32 %s40) #0
226  ret void
227}
228
229; First call preserved VGPR is used so it can't be used for SGPR spills.
230; GCN-LABEL: {{^}}callee_saved_sgpr_vgpr_func:
231; GCN-NOT: s40
232; GCN: v_writelane_b32 v33, s40
233; GCN: s_swappc_b64
234; GCN-NOT: s40
235; GCN: ; use s40
236; GCN-NOT: s40
237; GCN: v_readlane_b32 s40, v33
238; GCN-NOT: s40
239define void @callee_saved_sgpr_vgpr_func() #2 {
240  %s40 = call i32 asm sideeffect "; def s40", "={s40}"() #0
241  %v32 = call i32 asm sideeffect "; def v32", "={v32}"() #0
242  call void @external_void_func_void()
243  call void asm sideeffect "; use $0", "s"(i32 %s40) #0
244  call void asm sideeffect "; use $0", "v"(i32 %v32) #0
245  ret void
246}
247
248; GCN-LABEL: {{^}}callee_saved_sgpr_vgpr_kernel:
249; GCN-NOT: s40
250; GCN: ; def s40
251; GCN-NOT: s40
252; GCN: s_swappc_b64
253; GCN-NOT: s40
254; GCN: ; use s40
255; GCN-NOT: s40
256define amdgpu_kernel void @callee_saved_sgpr_vgpr_kernel() #2 {
257  %s40 = call i32 asm sideeffect "; def s40", "={s40}"() #0
258  %v32 = call i32 asm sideeffect "; def v32", "={v32}"() #0
259  call void @external_void_func_void()
260  call void asm sideeffect "; use $0", "s"(i32 %s40) #0
261  call void asm sideeffect "; use $0", "v"(i32 %v32) #0
262  ret void
263}
264
265attributes #0 = { nounwind }
266attributes #1 = { nounwind readnone }
267attributes #2 = { nounwind noinline }
268