| blob | 52163982d1e2d81fa1193398187c77e938d3981e |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -global-isel -march=amdgcn -verify-machineinstrs -amdgpu-mem-intrinsic-expand-size=19 %s -o - | FileCheck -check-prefix=LOOP %s
3 ; RUN: llc -global-isel -march=amdgcn -verify-machineinstrs -amdgpu-mem-intrinsic-expand-size=21 %s -o - | FileCheck -check-prefix=UNROLL %s
5 declare void @llvm.memcpy.p1i8.p1i8.i32(i8 addrspace(1)*, i8 addrspace(1)*, i32, i1 immarg)
7 define amdgpu_cs void @memcpy_p1i8(i8 addrspace(1)* %dst, i8 addrspace(1)* %src) {
8 ; LOOP-LABEL: memcpy_p1i8:
9 ; LOOP: ; %bb.0:
10 ; LOOP-NEXT: s_mov_b32 s2, 0
11 ; LOOP-NEXT: s_mov_b32 s3, 0xf000
12 ; LOOP-NEXT: s_mov_b64 s[0:1], 0
13 ; LOOP-NEXT: v_mov_b32_e32 v5, v3
14 ; LOOP-NEXT: v_mov_b32_e32 v4, v2
15 ; LOOP-NEXT: v_mov_b32_e32 v7, v1
16 ; LOOP-NEXT: v_mov_b32_e32 v6, v0
17 ; LOOP-NEXT: v_mov_b32_e32 v8, s2
18 ; LOOP-NEXT: .LBB0_1: ; %load-store-loop
19 ; LOOP-NEXT: ; =>This Inner Loop Header: Depth=1
20 ; LOOP-NEXT: s_waitcnt expcnt(1)
21 ; LOOP-NEXT: buffer_load_ubyte v9, v[4:5], s[0:3], 0 addr64
22 ; LOOP-NEXT: s_waitcnt expcnt(0)
23 ; LOOP-NEXT: buffer_load_ubyte v10, v[4:5], s[0:3], 0 addr64 offset:1
24 ; LOOP-NEXT: s_waitcnt vmcnt(0)
25 ; LOOP-NEXT: v_lshlrev_b32_e32 v10, 8, v10
26 ; LOOP-NEXT: v_or_b32_e32 v9, v10, v9
27 ; LOOP-NEXT: buffer_load_ubyte v10, v[4:5], s[0:3], 0 addr64 offset:2
28 ; LOOP-NEXT: buffer_load_ubyte v11, v[4:5], s[0:3], 0 addr64 offset:3
29 ; LOOP-NEXT: s_waitcnt vmcnt(0)
30 ; LOOP-NEXT: v_lshlrev_b32_e32 v11, 24, v11
31 ; LOOP-NEXT: v_lshlrev_b32_e32 v10, 16, v10
32 ; LOOP-NEXT: v_or_b32_e32 v10, v11, v10
33 ; LOOP-NEXT: v_or_b32_e32 v9, v10, v9
34 ; LOOP-NEXT: buffer_load_ubyte v10, v[4:5], s[0:3], 0 addr64 offset:4
35 ; LOOP-NEXT: buffer_load_ubyte v11, v[4:5], s[0:3], 0 addr64 offset:5
36 ; LOOP-NEXT: s_waitcnt vmcnt(0)
37 ; LOOP-NEXT: v_lshlrev_b32_e32 v11, 8, v11
38 ; LOOP-NEXT: v_or_b32_e32 v10, v11, v10
39 ; LOOP-NEXT: buffer_load_ubyte v11, v[4:5], s[0:3], 0 addr64 offset:6
40 ; LOOP-NEXT: buffer_load_ubyte v12, v[4:5], s[0:3], 0 addr64 offset:7
41 ; LOOP-NEXT: s_waitcnt vmcnt(0)
42 ; LOOP-NEXT: v_lshlrev_b32_e32 v12, 24, v12
43 ; LOOP-NEXT: v_lshlrev_b32_e32 v11, 16, v11
44 ; LOOP-NEXT: v_or_b32_e32 v11, v12, v11
45 ; LOOP-NEXT: v_or_b32_e32 v10, v11, v10
46 ; LOOP-NEXT: buffer_load_ubyte v11, v[4:5], s[0:3], 0 addr64 offset:8
47 ; LOOP-NEXT: buffer_load_ubyte v12, v[4:5], s[0:3], 0 addr64 offset:9
48 ; LOOP-NEXT: s_waitcnt vmcnt(0)
49 ; LOOP-NEXT: v_lshlrev_b32_e32 v12, 8, v12
50 ; LOOP-NEXT: v_or_b32_e32 v11, v12, v11
51 ; LOOP-NEXT: buffer_load_ubyte v12, v[4:5], s[0:3], 0 addr64 offset:10
52 ; LOOP-NEXT: buffer_load_ubyte v13, v[4:5], s[0:3], 0 addr64 offset:11
53 ; LOOP-NEXT: s_waitcnt vmcnt(0)
54 ; LOOP-NEXT: v_lshlrev_b32_e32 v13, 24, v13
55 ; LOOP-NEXT: v_lshlrev_b32_e32 v12, 16, v12
56 ; LOOP-NEXT: v_or_b32_e32 v12, v13, v12
57 ; LOOP-NEXT: v_or_b32_e32 v11, v12, v11
58 ; LOOP-NEXT: buffer_load_ubyte v12, v[4:5], s[0:3], 0 addr64 offset:12
59 ; LOOP-NEXT: buffer_load_ubyte v13, v[4:5], s[0:3], 0 addr64 offset:13
60 ; LOOP-NEXT: s_waitcnt vmcnt(0)
61 ; LOOP-NEXT: v_lshlrev_b32_e32 v13, 8, v13
62 ; LOOP-NEXT: v_or_b32_e32 v12, v13, v12
63 ; LOOP-NEXT: buffer_load_ubyte v13, v[4:5], s[0:3], 0 addr64 offset:14
64 ; LOOP-NEXT: buffer_load_ubyte v14, v[4:5], s[0:3], 0 addr64 offset:15
65 ; LOOP-NEXT: s_waitcnt vmcnt(0)
66 ; LOOP-NEXT: v_lshlrev_b32_e32 v14, 24, v14
67 ; LOOP-NEXT: v_lshlrev_b32_e32 v13, 16, v13
68 ; LOOP-NEXT: v_or_b32_e32 v13, v14, v13
69 ; LOOP-NEXT: v_or_b32_e32 v12, v13, v12
70 ; LOOP-NEXT: v_lshrrev_b32_e32 v13, 16, v9
71 ; LOOP-NEXT: v_bfe_u32 v14, v9, 8, 8
72 ; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64
73 ; LOOP-NEXT: buffer_store_byte v14, v[6:7], s[0:3], 0 addr64 offset:1
74 ; LOOP-NEXT: s_waitcnt expcnt(1)
75 ; LOOP-NEXT: v_lshrrev_b32_e32 v9, 24, v9
76 ; LOOP-NEXT: buffer_store_byte v13, v[6:7], s[0:3], 0 addr64 offset:2
77 ; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64 offset:3
78 ; LOOP-NEXT: s_waitcnt expcnt(0)
79 ; LOOP-NEXT: v_lshrrev_b32_e32 v9, 16, v10
80 ; LOOP-NEXT: v_bfe_u32 v13, v10, 8, 8
81 ; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:4
82 ; LOOP-NEXT: buffer_store_byte v13, v[6:7], s[0:3], 0 addr64 offset:5
83 ; LOOP-NEXT: s_waitcnt expcnt(1)
84 ; LOOP-NEXT: v_lshrrev_b32_e32 v10, 24, v10
85 ; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64 offset:6
86 ; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:7
87 ; LOOP-NEXT: s_waitcnt expcnt(1)
88 ; LOOP-NEXT: v_lshrrev_b32_e32 v9, 16, v11
89 ; LOOP-NEXT: s_waitcnt expcnt(0)
90 ; LOOP-NEXT: v_bfe_u32 v10, v11, 8, 8
91 ; LOOP-NEXT: buffer_store_byte v11, v[6:7], s[0:3], 0 addr64 offset:8
92 ; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:9
93 ; LOOP-NEXT: s_waitcnt expcnt(0)
94 ; LOOP-NEXT: v_lshrrev_b32_e32 v10, 24, v11
95 ; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64 offset:10
96 ; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:11
97 ; LOOP-NEXT: s_waitcnt expcnt(1)
98 ; LOOP-NEXT: v_lshrrev_b32_e32 v9, 16, v12
99 ; LOOP-NEXT: s_waitcnt expcnt(0)
100 ; LOOP-NEXT: v_bfe_u32 v10, v12, 8, 8
101 ; LOOP-NEXT: buffer_store_byte v12, v[6:7], s[0:3], 0 addr64 offset:12
102 ; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:13
103 ; LOOP-NEXT: s_waitcnt expcnt(0)
104 ; LOOP-NEXT: v_lshrrev_b32_e32 v10, 24, v12
105 ; LOOP-NEXT: buffer_store_byte v9, v[6:7], s[0:3], 0 addr64 offset:14
106 ; LOOP-NEXT: buffer_store_byte v10, v[6:7], s[0:3], 0 addr64 offset:15
107 ; LOOP-NEXT: v_add_i32_e32 v8, vcc, 1, v8
108 ; LOOP-NEXT: s_xor_b64 s[4:5], vcc, -1
109 ; LOOP-NEXT: v_add_i32_e32 v6, vcc, 16, v6
110 ; LOOP-NEXT: v_addc_u32_e32 v7, vcc, 0, v7, vcc
111 ; LOOP-NEXT: v_add_i32_e32 v4, vcc, 16, v4
112 ; LOOP-NEXT: v_addc_u32_e32 v5, vcc, 0, v5, vcc
113 ; LOOP-NEXT: s_xor_b64 s[4:5], s[4:5], -1
114 ; LOOP-NEXT: s_and_b64 vcc, exec, s[4:5]
115 ; LOOP-NEXT: s_cbranch_vccnz .LBB0_1
116 ; LOOP-NEXT: ; %bb.2: ; %memcpy-split
117 ; LOOP-NEXT: s_mov_b32 s2, 0
118 ; LOOP-NEXT: s_mov_b32 s3, 0xf000
119 ; LOOP-NEXT: s_mov_b64 s[0:1], 0
120 ; LOOP-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:17
121 ; LOOP-NEXT: buffer_load_ubyte v5, v[2:3], s[0:3], 0 addr64 offset:19
122 ; LOOP-NEXT: buffer_load_ubyte v6, v[2:3], s[0:3], 0 addr64 offset:18
123 ; LOOP-NEXT: buffer_load_ubyte v2, v[2:3], s[0:3], 0 addr64 offset:16
124 ; LOOP-NEXT: s_waitcnt vmcnt(3)
125 ; LOOP-NEXT: v_lshlrev_b32_e32 v3, 8, v4
126 ; LOOP-NEXT: s_waitcnt vmcnt(2)
127 ; LOOP-NEXT: v_lshlrev_b32_e32 v4, 24, v5
128 ; LOOP-NEXT: s_waitcnt vmcnt(1)
129 ; LOOP-NEXT: v_lshlrev_b32_e32 v5, 16, v6
130 ; LOOP-NEXT: s_waitcnt vmcnt(0)
131 ; LOOP-NEXT: v_or_b32_e32 v2, v3, v2
132 ; LOOP-NEXT: v_or_b32_e32 v3, v4, v5
133 ; LOOP-NEXT: v_or_b32_e32 v2, v3, v2
134 ; LOOP-NEXT: v_lshrrev_b32_e32 v3, 16, v2
135 ; LOOP-NEXT: v_bfe_u32 v4, v2, 8, 8
136 ; LOOP-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64 offset:16
137 ; LOOP-NEXT: s_waitcnt expcnt(0)
138 ; LOOP-NEXT: v_lshrrev_b32_e32 v2, 24, v2
139 ; LOOP-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:17
140 ; LOOP-NEXT: buffer_store_byte v3, v[0:1], s[0:3], 0 addr64 offset:18
141 ; LOOP-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64 offset:19
142 ; LOOP-NEXT: s_endpgm
143 ;
144 ; UNROLL-LABEL: memcpy_p1i8:
145 ; UNROLL: ; %bb.0:
146 ; UNROLL-NEXT: s_mov_b32 s2, 0
147 ; UNROLL-NEXT: s_mov_b32 s3, 0xf000
148 ; UNROLL-NEXT: s_mov_b64 s[0:1], 0
149 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64
150 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
151 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64
152 ; UNROLL-NEXT: s_waitcnt expcnt(0)
153 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:1
154 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
155 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:1
156 ; UNROLL-NEXT: s_waitcnt expcnt(0)
157 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:2
158 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
159 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:2
160 ; UNROLL-NEXT: s_waitcnt expcnt(0)
161 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:3
162 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
163 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:3
164 ; UNROLL-NEXT: s_waitcnt expcnt(0)
165 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:4
166 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
167 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:4
168 ; UNROLL-NEXT: s_waitcnt expcnt(0)
169 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:5
170 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
171 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:5
172 ; UNROLL-NEXT: s_waitcnt expcnt(0)
173 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:6
174 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
175 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:6
176 ; UNROLL-NEXT: s_waitcnt expcnt(0)
177 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:7
178 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
179 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:7
180 ; UNROLL-NEXT: s_waitcnt expcnt(0)
181 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:8
182 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
183 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:8
184 ; UNROLL-NEXT: s_waitcnt expcnt(0)
185 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:9
186 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
187 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:9
188 ; UNROLL-NEXT: s_waitcnt expcnt(0)
189 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:10
190 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
191 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:10
192 ; UNROLL-NEXT: s_waitcnt expcnt(0)
193 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:11
194 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
195 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:11
196 ; UNROLL-NEXT: s_waitcnt expcnt(0)
197 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:12
198 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
199 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:12
200 ; UNROLL-NEXT: s_waitcnt expcnt(0)
201 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:13
202 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
203 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:13
204 ; UNROLL-NEXT: s_waitcnt expcnt(0)
205 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:14
206 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
207 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:14
208 ; UNROLL-NEXT: s_waitcnt expcnt(0)
209 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:15
210 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
211 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:15
212 ; UNROLL-NEXT: s_waitcnt expcnt(0)
213 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:16
214 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
215 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:16
216 ; UNROLL-NEXT: s_waitcnt expcnt(0)
217 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:17
218 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
219 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:17
220 ; UNROLL-NEXT: s_waitcnt expcnt(0)
221 ; UNROLL-NEXT: buffer_load_ubyte v4, v[2:3], s[0:3], 0 addr64 offset:18
222 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
223 ; UNROLL-NEXT: buffer_store_byte v4, v[0:1], s[0:3], 0 addr64 offset:18
224 ; UNROLL-NEXT: buffer_load_ubyte v2, v[2:3], s[0:3], 0 addr64 offset:19
225 ; UNROLL-NEXT: s_waitcnt vmcnt(0)
226 ; UNROLL-NEXT: buffer_store_byte v2, v[0:1], s[0:3], 0 addr64 offset:19
227 ; UNROLL-NEXT: s_endpgm
228 call void @llvm.memcpy.p1i8.p1i8.i32(i8 addrspace(1)* %dst, i8 addrspace(1)* %src, i32 20, i1 false)
229 ret void
230 }