1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 < %s | FileCheck -check-prefixes=GCN,GFX9,MUBUF %s
3; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 --amdgpu-enable-flat-scratch < %s | FileCheck -check-prefixes=GCN,GFX9,FLATSCR %s
4
5; Make sure we use the correct frame offset is used with the local
6; frame area.
7;
8; %pin.low is allocated to offset 0.
9;
10; %local.area is assigned to the local frame offset by the
11; LocalStackSlotAllocation pass at offset 4096.
12;
13; The %load1 access to %gep.large.offset initially used the stack
14; pointer register and directly referenced the frame index. After
15; LocalStackSlotAllocation, it would no longer refer to a frame index
16; so eliminateFrameIndex would not adjust the access to use the
17; correct FP offset.
18
19define amdgpu_kernel void @local_stack_offset_uses_sp(i64 addrspace(1)* %out, i8 addrspace(1)* %in) {
20; MUBUF-LABEL: local_stack_offset_uses_sp:
21; MUBUF:       ; %bb.0: ; %entry
22; MUBUF-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x0
23; MUBUF-NEXT:    s_add_u32 flat_scratch_lo, s6, s9
24; MUBUF-NEXT:    s_addc_u32 flat_scratch_hi, s7, 0
25; MUBUF-NEXT:    s_add_u32 s0, s0, s9
26; MUBUF-NEXT:    v_mov_b32_e32 v1, 0x3000
27; MUBUF-NEXT:    s_addc_u32 s1, s1, 0
28; MUBUF-NEXT:    v_add_u32_e32 v0, 64, v1
29; MUBUF-NEXT:    v_mov_b32_e32 v2, 0
30; MUBUF-NEXT:    v_mov_b32_e32 v3, 0x2000
31; MUBUF-NEXT:    s_mov_b32 s6, 0
32; MUBUF-NEXT:    buffer_store_dword v2, v3, s[0:3], 0 offen
33; MUBUF-NEXT:  BB0_1: ; %loadstoreloop
34; MUBUF-NEXT:    ; =>This Inner Loop Header: Depth=1
35; MUBUF-NEXT:    v_add_u32_e32 v3, s6, v1
36; MUBUF-NEXT:    s_add_i32 s6, s6, 1
37; MUBUF-NEXT:    s_cmpk_lt_u32 s6, 0x2120
38; MUBUF-NEXT:    buffer_store_byte v2, v3, s[0:3], 0 offen
39; MUBUF-NEXT:    s_cbranch_scc1 BB0_1
40; MUBUF-NEXT:  ; %bb.2: ; %split
41; MUBUF-NEXT:    v_mov_b32_e32 v1, 0x3000
42; MUBUF-NEXT:    v_add_u32_e32 v1, 0x20d0, v1
43; MUBUF-NEXT:    buffer_load_dword v2, v1, s[0:3], 0 offen
44; MUBUF-NEXT:    buffer_load_dword v1, v1, s[0:3], 0 offen offset:4
45; MUBUF-NEXT:    buffer_load_dword v3, v0, s[0:3], 0 offen
46; MUBUF-NEXT:    buffer_load_dword v4, v0, s[0:3], 0 offen offset:4
47; MUBUF-NEXT:    s_waitcnt vmcnt(1)
48; MUBUF-NEXT:    v_add_co_u32_e32 v0, vcc, v2, v3
49; MUBUF-NEXT:    s_waitcnt vmcnt(0)
50; MUBUF-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v4, vcc
51; MUBUF-NEXT:    v_mov_b32_e32 v2, 0
52; MUBUF-NEXT:    s_waitcnt lgkmcnt(0)
53; MUBUF-NEXT:    global_store_dwordx2 v2, v[0:1], s[4:5]
54; MUBUF-NEXT:    s_endpgm
55;
56; FLATSCR-LABEL: local_stack_offset_uses_sp:
57; FLATSCR:       ; %bb.0: ; %entry
58; FLATSCR-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
59; FLATSCR-NEXT:    s_add_u32 flat_scratch_lo, s2, s5
60; FLATSCR-NEXT:    s_addc_u32 flat_scratch_hi, s3, 0
61; FLATSCR-NEXT:    v_mov_b32_e32 v0, 0
62; FLATSCR-NEXT:    s_movk_i32 vcc_hi, 0x2000
63; FLATSCR-NEXT:    s_mov_b32 s2, 0
64; FLATSCR-NEXT:    scratch_store_dword off, v0, vcc_hi
65; FLATSCR-NEXT:  BB0_1: ; %loadstoreloop
66; FLATSCR-NEXT:    ; =>This Inner Loop Header: Depth=1
67; FLATSCR-NEXT:    s_add_u32 s3, 0x3000, s2
68; FLATSCR-NEXT:    s_add_i32 s2, s2, 1
69; FLATSCR-NEXT:    s_cmpk_lt_u32 s2, 0x2120
70; FLATSCR-NEXT:    scratch_store_byte off, v0, s3
71; FLATSCR-NEXT:    s_cbranch_scc1 BB0_1
72; FLATSCR-NEXT:  ; %bb.2: ; %split
73; FLATSCR-NEXT:    s_movk_i32 s2, 0x2000
74; FLATSCR-NEXT:    s_add_u32 s2, 0x3000, s2
75; FLATSCR-NEXT:    scratch_load_dwordx2 v[0:1], off, s2 offset:208
76; FLATSCR-NEXT:    s_movk_i32 s2, 0x3000
77; FLATSCR-NEXT:    scratch_load_dwordx2 v[2:3], off, s2 offset:64
78; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
79; FLATSCR-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
80; FLATSCR-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
81; FLATSCR-NEXT:    v_mov_b32_e32 v2, 0
82; FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
83; FLATSCR-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
84; FLATSCR-NEXT:    s_endpgm
85entry:
86  %pin.low = alloca i32, align 8192, addrspace(5)
87  %local.area = alloca [1060 x i64], align 4096, addrspace(5)
88  store volatile i32 0, i32 addrspace(5)* %pin.low
89  %local.area.cast = bitcast [1060 x i64] addrspace(5)* %local.area to i8 addrspace(5)*
90  call void @llvm.memset.p5i8.i32(i8 addrspace(5)* align 4 %local.area.cast, i8 0, i32 8480, i1 true)
91  %gep.large.offset = getelementptr inbounds [1060 x i64], [1060 x i64] addrspace(5)* %local.area, i64 0, i64 1050
92  %gep.small.offset = getelementptr inbounds [1060 x i64], [1060 x i64] addrspace(5)* %local.area, i64 0, i64 8
93  %load0 = load volatile i64, i64 addrspace(5)* %gep.large.offset
94  %load1 = load volatile i64, i64 addrspace(5)* %gep.small.offset
95  %add0 = add i64 %load0, %load1
96  store volatile i64 %add0, i64 addrspace(1)* %out
97  ret void
98}
99
100define void @func_local_stack_offset_uses_sp(i64 addrspace(1)* %out, i8 addrspace(1)* %in) {
101; MUBUF-LABEL: func_local_stack_offset_uses_sp:
102; MUBUF:       ; %bb.0: ; %entry
103; MUBUF-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
104; MUBUF-NEXT:    s_add_u32 s4, s32, 0x7ffc0
105; MUBUF-NEXT:    s_mov_b32 s5, s33
106; MUBUF-NEXT:    s_and_b32 s33, s4, 0xfff80000
107; MUBUF-NEXT:    v_lshrrev_b32_e64 v3, 6, s33
108; MUBUF-NEXT:    v_add_u32_e32 v3, 0x1000, v3
109; MUBUF-NEXT:    v_mov_b32_e32 v4, 0
110; MUBUF-NEXT:    v_add_u32_e32 v2, 64, v3
111; MUBUF-NEXT:    s_mov_b32 s4, 0
112; MUBUF-NEXT:    s_add_u32 s32, s32, 0x180000
113; MUBUF-NEXT:    buffer_store_dword v4, off, s[0:3], s33
114; MUBUF-NEXT:  BB1_1: ; %loadstoreloop
115; MUBUF-NEXT:    ; =>This Inner Loop Header: Depth=1
116; MUBUF-NEXT:    v_add_u32_e32 v5, s4, v3
117; MUBUF-NEXT:    s_add_i32 s4, s4, 1
118; MUBUF-NEXT:    s_cmpk_lt_u32 s4, 0x2120
119; MUBUF-NEXT:    buffer_store_byte v4, v5, s[0:3], 0 offen
120; MUBUF-NEXT:    s_cbranch_scc1 BB1_1
121; MUBUF-NEXT:  ; %bb.2: ; %split
122; MUBUF-NEXT:    v_lshrrev_b32_e64 v3, 6, s33
123; MUBUF-NEXT:    v_add_u32_e32 v3, 0x1000, v3
124; MUBUF-NEXT:    v_add_u32_e32 v3, 0x20d0, v3
125; MUBUF-NEXT:    buffer_load_dword v4, v3, s[0:3], 0 offen
126; MUBUF-NEXT:    buffer_load_dword v3, v3, s[0:3], 0 offen offset:4
127; MUBUF-NEXT:    buffer_load_dword v5, v2, s[0:3], 0 offen
128; MUBUF-NEXT:    buffer_load_dword v6, v2, s[0:3], 0 offen offset:4
129; MUBUF-NEXT:    s_sub_u32 s32, s32, 0x180000
130; MUBUF-NEXT:    s_mov_b32 s33, s5
131; MUBUF-NEXT:    s_waitcnt vmcnt(1)
132; MUBUF-NEXT:    v_add_co_u32_e32 v2, vcc, v4, v5
133; MUBUF-NEXT:    s_waitcnt vmcnt(0)
134; MUBUF-NEXT:    v_addc_co_u32_e32 v3, vcc, v3, v6, vcc
135; MUBUF-NEXT:    global_store_dwordx2 v[0:1], v[2:3], off
136; MUBUF-NEXT:    s_waitcnt vmcnt(0)
137; MUBUF-NEXT:    s_setpc_b64 s[30:31]
138;
139; FLATSCR-LABEL: func_local_stack_offset_uses_sp:
140; FLATSCR:       ; %bb.0: ; %entry
141; FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
142; FLATSCR-NEXT:    s_add_u32 s0, s32, 0x1fff
143; FLATSCR-NEXT:    s_mov_b32 s2, s33
144; FLATSCR-NEXT:    s_and_b32 s33, s0, 0xffffe000
145; FLATSCR-NEXT:    v_mov_b32_e32 v2, 0
146; FLATSCR-NEXT:    s_mov_b32 s0, 0
147; FLATSCR-NEXT:    s_add_u32 s32, s32, 0x6000
148; FLATSCR-NEXT:    scratch_store_dword off, v2, s33
149; FLATSCR-NEXT:  BB1_1: ; %loadstoreloop
150; FLATSCR-NEXT:    ; =>This Inner Loop Header: Depth=1
151; FLATSCR-NEXT:    s_add_u32 vcc_hi, s33, 0x1000
152; FLATSCR-NEXT:    s_add_u32 s1, vcc_hi, s0
153; FLATSCR-NEXT:    s_add_i32 s0, s0, 1
154; FLATSCR-NEXT:    s_cmpk_lt_u32 s0, 0x2120
155; FLATSCR-NEXT:    scratch_store_byte off, v2, s1
156; FLATSCR-NEXT:    s_cbranch_scc1 BB1_1
157; FLATSCR-NEXT:  ; %bb.2: ; %split
158; FLATSCR-NEXT:    s_movk_i32 s0, 0x2000
159; FLATSCR-NEXT:    s_add_u32 s1, s33, 0x1000
160; FLATSCR-NEXT:    s_add_u32 s0, s1, s0
161; FLATSCR-NEXT:    scratch_load_dwordx2 v[2:3], off, s0 offset:208
162; FLATSCR-NEXT:    s_add_u32 s0, s33, 0x1000
163; FLATSCR-NEXT:    scratch_load_dwordx2 v[4:5], off, s0 offset:64
164; FLATSCR-NEXT:    s_sub_u32 s32, s32, 0x6000
165; FLATSCR-NEXT:    s_mov_b32 s33, s2
166; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
167; FLATSCR-NEXT:    v_add_co_u32_e32 v2, vcc, v2, v4
168; FLATSCR-NEXT:    v_addc_co_u32_e32 v3, vcc, v3, v5, vcc
169; FLATSCR-NEXT:    global_store_dwordx2 v[0:1], v[2:3], off
170; FLATSCR-NEXT:    s_waitcnt vmcnt(0)
171; FLATSCR-NEXT:    s_setpc_b64 s[30:31]
172entry:
173  %pin.low = alloca i32, align 8192, addrspace(5)
174  %local.area = alloca [1060 x i64], align 4096, addrspace(5)
175  store volatile i32 0, i32 addrspace(5)* %pin.low
176  %local.area.cast = bitcast [1060 x i64] addrspace(5)* %local.area to i8 addrspace(5)*
177  call void @llvm.memset.p5i8.i32(i8 addrspace(5)* align 4 %local.area.cast, i8 0, i32 8480, i1 true)
178  %gep.large.offset = getelementptr inbounds [1060 x i64], [1060 x i64] addrspace(5)* %local.area, i64 0, i64 1050
179  %gep.small.offset = getelementptr inbounds [1060 x i64], [1060 x i64] addrspace(5)* %local.area, i64 0, i64 8
180  %load0 = load volatile i64, i64 addrspace(5)* %gep.large.offset
181  %load1 = load volatile i64, i64 addrspace(5)* %gep.small.offset
182  %add0 = add i64 %load0, %load1
183  store volatile i64 %add0, i64 addrspace(1)* %out
184  ret void
185}
186
187declare void @llvm.memset.p5i8.i32(i8 addrspace(5)* nocapture writeonly, i8, i32, i1 immarg) #0
188
189attributes #0 = { argmemonly nounwind willreturn writeonly }
190