1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -enable-var-scope -check-prefix=GCN %s
3
4; Load argument depends on waitcnt which should be skipped.
5define amdgpu_kernel void @call_memory_arg_load(i32 addrspace(3)* %ptr, i32) #0 {
6; GCN-LABEL: call_memory_arg_load:
7; GCN:       ; %bb.0:
8; GCN-NEXT:    s_load_dword s4, s[4:5], 0x0
9; GCN-NEXT:    s_add_u32 flat_scratch_lo, s6, s9
10; GCN-NEXT:    s_addc_u32 flat_scratch_hi, s7, 0
11; GCN-NEXT:    s_add_u32 s0, s0, s9
12; GCN-NEXT:    s_addc_u32 s1, s1, 0
13; GCN-NEXT:    s_waitcnt lgkmcnt(0)
14; GCN-NEXT:    v_mov_b32_e32 v0, s4
15; GCN-NEXT:    ds_read_b32 v0, v0
16; GCN-NEXT:    s_getpc_b64 s[4:5]
17; GCN-NEXT:    s_add_u32 s4, s4, func@rel32@lo+4
18; GCN-NEXT:    s_addc_u32 s5, s5, func@rel32@hi+4
19; GCN-NEXT:    s_mov_b32 s32, 0
20; GCN-NEXT:    s_swappc_b64 s[30:31], s[4:5]
21; GCN-NEXT:    s_endpgm
22  %vgpr = load volatile i32, i32 addrspace(3)* %ptr
23  call void @func(i32 %vgpr)
24  ret void
25}
26
27; Memory waitcnt with no register dependence on the call
28define amdgpu_kernel void @call_memory_no_dep(i32 addrspace(1)* %ptr, i32) #0 {
29; GCN-LABEL: call_memory_no_dep:
30; GCN:       ; %bb.0:
31; GCN-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
32; GCN-NEXT:    s_add_u32 flat_scratch_lo, s6, s9
33; GCN-NEXT:    s_addc_u32 flat_scratch_hi, s7, 0
34; GCN-NEXT:    s_add_u32 s0, s0, s9
35; GCN-NEXT:    v_mov_b32_e32 v2, 0
36; GCN-NEXT:    s_waitcnt lgkmcnt(0)
37; GCN-NEXT:    v_mov_b32_e32 v0, s4
38; GCN-NEXT:    v_mov_b32_e32 v1, s5
39; GCN-NEXT:    s_addc_u32 s1, s1, 0
40; GCN-NEXT:    global_store_dword v[0:1], v2, off
41; GCN-NEXT:    v_mov_b32_e32 v0, 0
42; GCN-NEXT:    s_getpc_b64 s[6:7]
43; GCN-NEXT:    s_add_u32 s6, s6, func@rel32@lo+4
44; GCN-NEXT:    s_addc_u32 s7, s7, func@rel32@hi+4
45; GCN-NEXT:    s_mov_b32 s32, 0
46; GCN-NEXT:    s_swappc_b64 s[30:31], s[6:7]
47; GCN-NEXT:    s_endpgm
48  store i32 0, i32 addrspace(1)* %ptr
49  call void @func(i32 0)
50  ret void
51}
52
53; Should not wait after the call before memory
54define amdgpu_kernel void @call_no_wait_after_call(i32 addrspace(1)* %ptr, i32) #0 {
55; GCN-LABEL: call_no_wait_after_call:
56; GCN:       ; %bb.0:
57; GCN-NEXT:    s_add_u32 flat_scratch_lo, s6, s9
58; GCN-NEXT:    s_load_dwordx2 s[34:35], s[4:5], 0x0
59; GCN-NEXT:    s_addc_u32 flat_scratch_hi, s7, 0
60; GCN-NEXT:    s_add_u32 s0, s0, s9
61; GCN-NEXT:    s_addc_u32 s1, s1, 0
62; GCN-NEXT:    v_mov_b32_e32 v0, 0
63; GCN-NEXT:    s_getpc_b64 s[4:5]
64; GCN-NEXT:    s_add_u32 s4, s4, func@rel32@lo+4
65; GCN-NEXT:    s_addc_u32 s5, s5, func@rel32@hi+4
66; GCN-NEXT:    s_mov_b32 s32, 0
67; GCN-NEXT:    v_mov_b32_e32 v32, 0
68; GCN-NEXT:    s_swappc_b64 s[30:31], s[4:5]
69; GCN-NEXT:    v_mov_b32_e32 v0, s34
70; GCN-NEXT:    v_mov_b32_e32 v1, s35
71; GCN-NEXT:    global_store_dword v[0:1], v32, off
72; GCN-NEXT:    s_endpgm
73  call void @func(i32 0)
74  store i32 0, i32 addrspace(1)* %ptr
75  ret void
76}
77
78define amdgpu_kernel void @call_no_wait_after_call_return_val(i32 addrspace(1)* %ptr, i32) #0 {
79; GCN-LABEL: call_no_wait_after_call_return_val:
80; GCN:       ; %bb.0:
81; GCN-NEXT:    s_add_u32 flat_scratch_lo, s6, s9
82; GCN-NEXT:    s_load_dwordx2 s[34:35], s[4:5], 0x0
83; GCN-NEXT:    s_addc_u32 flat_scratch_hi, s7, 0
84; GCN-NEXT:    s_add_u32 s0, s0, s9
85; GCN-NEXT:    s_addc_u32 s1, s1, 0
86; GCN-NEXT:    v_mov_b32_e32 v0, 0
87; GCN-NEXT:    s_getpc_b64 s[4:5]
88; GCN-NEXT:    s_add_u32 s4, s4, func.return@rel32@lo+4
89; GCN-NEXT:    s_addc_u32 s5, s5, func.return@rel32@hi+4
90; GCN-NEXT:    s_mov_b32 s32, 0
91; GCN-NEXT:    s_swappc_b64 s[30:31], s[4:5]
92; GCN-NEXT:    v_mov_b32_e32 v1, s34
93; GCN-NEXT:    v_mov_b32_e32 v2, s35
94; GCN-NEXT:    global_store_dword v[1:2], v0, off
95; GCN-NEXT:    s_endpgm
96  %rv = call i32 @func.return(i32 0)
97  store i32 %rv, i32 addrspace(1)* %ptr
98  ret void
99}
100
101; Need to wait for the address dependency
102define amdgpu_kernel void @call_got_load(i32 addrspace(1)* %ptr, i32) #0 {
103; GCN-LABEL: call_got_load:
104; GCN:       ; %bb.0:
105; GCN-NEXT:    s_add_u32 flat_scratch_lo, s6, s9
106; GCN-NEXT:    s_addc_u32 flat_scratch_hi, s7, 0
107; GCN-NEXT:    s_add_u32 s0, s0, s9
108; GCN-NEXT:    s_addc_u32 s1, s1, 0
109; GCN-NEXT:    s_getpc_b64 s[4:5]
110; GCN-NEXT:    s_add_u32 s4, s4, got.func@gotpcrel32@lo+4
111; GCN-NEXT:    s_addc_u32 s5, s5, got.func@gotpcrel32@hi+4
112; GCN-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
113; GCN-NEXT:    v_mov_b32_e32 v0, 0
114; GCN-NEXT:    s_mov_b32 s32, 0
115; GCN-NEXT:    s_waitcnt lgkmcnt(0)
116; GCN-NEXT:    s_swappc_b64 s[30:31], s[4:5]
117; GCN-NEXT:    s_endpgm
118  call void @got.func(i32 0)
119  ret void
120}
121
122; Need to wait for the address dependency
123define void @tailcall_got_load(i32 addrspace(1)* %ptr, i32) #0 {
124; GCN-LABEL: tailcall_got_load:
125; GCN:       ; %bb.0:
126; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
127; GCN-NEXT:    s_getpc_b64 s[4:5]
128; GCN-NEXT:    s_add_u32 s4, s4, got.func@gotpcrel32@lo+4
129; GCN-NEXT:    s_addc_u32 s5, s5, got.func@gotpcrel32@hi+4
130; GCN-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
131; GCN-NEXT:    v_mov_b32_e32 v0, 0
132; GCN-NEXT:    s_waitcnt lgkmcnt(0)
133; GCN-NEXT:    s_setpc_b64 s[4:5]
134  tail call void @got.func(i32 0)
135  ret void
136}
137
138; No need to wait for the load.
139define void @tail_call_memory_arg_load(i32 addrspace(3)* %ptr, i32) #0 {
140; GCN-LABEL: tail_call_memory_arg_load:
141; GCN:       ; %bb.0:
142; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
143; GCN-NEXT:    ds_read_b32 v0, v0
144; GCN-NEXT:    s_getpc_b64 s[4:5]
145; GCN-NEXT:    s_add_u32 s4, s4, func@rel32@lo+4
146; GCN-NEXT:    s_addc_u32 s5, s5, func@rel32@hi+4
147; GCN-NEXT:    s_setpc_b64 s[4:5]
148  %vgpr = load volatile i32, i32 addrspace(3)* %ptr
149  tail call void @func(i32 %vgpr)
150  ret void
151}
152
153declare hidden void @func(i32) #0
154declare hidden i32 @func.return(i32) #0
155declare void @got.func(i32) #0
156
157attributes #0 = { nounwind }
158