| blob | ff7eb92fca067008c50effb8d298bae4ea1115f7 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=-unaligned-scratch-access < %s | FileCheck --check-prefix=GFX7-ALIGNED %s
3 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=hawaii -mattr=+unaligned-scratch-access < %s | FileCheck --check-prefix=GFX7-UNALIGNED %s
4 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+unaligned-scratch-access < %s | FileCheck --check-prefix=GFX9 %s
5 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -mattr=+unaligned-scratch-access -amdgpu-enable-flat-scratch < %s | FileCheck --check-prefix=GFX9-FLASTSCR %s
6 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+unaligned-scratch-access < %s | FileCheck --check-prefix=GFX10 %s
7 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1010 -mattr=+unaligned-scratch-access -amdgpu-enable-flat-scratch < %s | FileCheck --check-prefix=GFX10-FLASTSCR %s
9 ; Should not merge this to a dword load
10 define i32 @private_load_2xi16_align2(i16 addrspace(5)* %p) #0 {
11 ; GFX7-ALIGNED-LABEL: private_load_2xi16_align2:
12 ; GFX7-ALIGNED: ; %bb.0:
13 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
14 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v1, vcc, 2, v0
15 ; GFX7-ALIGNED-NEXT: buffer_load_ushort v0, v0, s[0:3], 0 offen
16 ; GFX7-ALIGNED-NEXT: buffer_load_ushort v1, v1, s[0:3], 0 offen
17 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0)
18 ; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1
19 ; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v0, v1
20 ; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31]
21 ;
22 ; GFX7-UNALIGNED-LABEL: private_load_2xi16_align2:
23 ; GFX7-UNALIGNED: ; %bb.0:
24 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
25 ; GFX7-UNALIGNED-NEXT: v_add_i32_e32 v1, vcc, 2, v0
26 ; GFX7-UNALIGNED-NEXT: buffer_load_ushort v0, v0, s[0:3], 0 offen
27 ; GFX7-UNALIGNED-NEXT: buffer_load_ushort v1, v1, s[0:3], 0 offen
28 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0)
29 ; GFX7-UNALIGNED-NEXT: v_lshlrev_b32_e32 v1, 16, v1
30 ; GFX7-UNALIGNED-NEXT: v_or_b32_e32 v0, v0, v1
31 ; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31]
32 ;
33 ; GFX9-LABEL: private_load_2xi16_align2:
34 ; GFX9: ; %bb.0:
35 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
36 ; GFX9-NEXT: buffer_load_ushort v1, v0, s[0:3], 0 offen
37 ; GFX9-NEXT: buffer_load_ushort v2, v0, s[0:3], 0 offen offset:2
38 ; GFX9-NEXT: s_waitcnt vmcnt(0)
39 ; GFX9-NEXT: v_lshl_or_b32 v0, v2, 16, v1
40 ; GFX9-NEXT: s_setpc_b64 s[30:31]
41 ;
42 ; GFX9-FLASTSCR-LABEL: private_load_2xi16_align2:
43 ; GFX9-FLASTSCR: ; %bb.0:
44 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
45 ; GFX9-FLASTSCR-NEXT: scratch_load_ushort v1, v0, off
46 ; GFX9-FLASTSCR-NEXT: scratch_load_ushort v2, v0, off offset:2
47 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
48 ; GFX9-FLASTSCR-NEXT: v_lshl_or_b32 v0, v2, 16, v1
49 ; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
50 ;
51 ; GFX10-LABEL: private_load_2xi16_align2:
52 ; GFX10: ; %bb.0:
53 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
54 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
55 ; GFX10-NEXT: s_clause 0x1
56 ; GFX10-NEXT: buffer_load_ushort v1, v0, s[0:3], 0 offen
57 ; GFX10-NEXT: buffer_load_ushort v2, v0, s[0:3], 0 offen offset:2
58 ; GFX10-NEXT: s_waitcnt vmcnt(0)
59 ; GFX10-NEXT: v_lshl_or_b32 v0, v2, 16, v1
60 ; GFX10-NEXT: s_setpc_b64 s[30:31]
61 ;
62 ; GFX10-FLASTSCR-LABEL: private_load_2xi16_align2:
63 ; GFX10-FLASTSCR: ; %bb.0:
64 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
65 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
66 ; GFX10-FLASTSCR-NEXT: s_clause 0x1
67 ; GFX10-FLASTSCR-NEXT: scratch_load_ushort v1, v0, off
68 ; GFX10-FLASTSCR-NEXT: scratch_load_ushort v2, v0, off offset:2
69 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
70 ; GFX10-FLASTSCR-NEXT: v_lshl_or_b32 v0, v2, 16, v1
71 ; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
72 %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1
73 %p.0 = load i16, i16 addrspace(5)* %p, align 2
74 %p.1 = load i16, i16 addrspace(5)* %gep.p, align 2
75 %zext.0 = zext i16 %p.0 to i32
76 %zext.1 = zext i16 %p.1 to i32
77 %shl.1 = shl i32 %zext.1, 16
78 %or = or i32 %zext.0, %shl.1
79 ret i32 %or
80 }
82 ; Should not merge this to a dword store
83 define void @private_store_2xi16_align2(i16 addrspace(5)* %p, i16 addrspace(5)* %r) #0 {
84 ; GFX7-ALIGNED-LABEL: private_store_2xi16_align2:
85 ; GFX7-ALIGNED: ; %bb.0:
86 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
87 ; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v3, 1
88 ; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, 2
89 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v1
90 ; GFX7-ALIGNED-NEXT: buffer_store_short v3, v1, s[0:3], 0 offen
91 ; GFX7-ALIGNED-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
92 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0)
93 ; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31]
94 ;
95 ; GFX7-UNALIGNED-LABEL: private_store_2xi16_align2:
96 ; GFX7-UNALIGNED: ; %bb.0:
97 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
98 ; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v3, 1
99 ; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, 2
100 ; GFX7-UNALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v1
101 ; GFX7-UNALIGNED-NEXT: buffer_store_short v3, v1, s[0:3], 0 offen
102 ; GFX7-UNALIGNED-NEXT: buffer_store_short v0, v2, s[0:3], 0 offen
103 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0)
104 ; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31]
105 ;
106 ; GFX9-LABEL: private_store_2xi16_align2:
107 ; GFX9: ; %bb.0:
108 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
109 ; GFX9-NEXT: v_mov_b32_e32 v0, 1
110 ; GFX9-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
111 ; GFX9-NEXT: v_mov_b32_e32 v0, 2
112 ; GFX9-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen offset:2
113 ; GFX9-NEXT: s_waitcnt vmcnt(0)
114 ; GFX9-NEXT: s_setpc_b64 s[30:31]
115 ;
116 ; GFX9-FLASTSCR-LABEL: private_store_2xi16_align2:
117 ; GFX9-FLASTSCR: ; %bb.0:
118 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
119 ; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 1
120 ; GFX9-FLASTSCR-NEXT: scratch_store_short v1, v0, off
121 ; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 2
122 ; GFX9-FLASTSCR-NEXT: scratch_store_short v1, v0, off offset:2
123 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
124 ; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
125 ;
126 ; GFX10-LABEL: private_store_2xi16_align2:
127 ; GFX10: ; %bb.0:
128 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
129 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
130 ; GFX10-NEXT: v_mov_b32_e32 v0, 1
131 ; GFX10-NEXT: v_mov_b32_e32 v2, 2
132 ; GFX10-NEXT: buffer_store_short v0, v1, s[0:3], 0 offen
133 ; GFX10-NEXT: buffer_store_short v2, v1, s[0:3], 0 offen offset:2
134 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
135 ; GFX10-NEXT: s_setpc_b64 s[30:31]
136 ;
137 ; GFX10-FLASTSCR-LABEL: private_store_2xi16_align2:
138 ; GFX10-FLASTSCR: ; %bb.0:
139 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
140 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
141 ; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v0, 1
142 ; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v2, 2
143 ; GFX10-FLASTSCR-NEXT: scratch_store_short v1, v0, off
144 ; GFX10-FLASTSCR-NEXT: scratch_store_short v1, v2, off offset:2
145 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
146 ; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
147 %gep.r = getelementptr i16, i16 addrspace(5)* %r, i64 1
148 store i16 1, i16 addrspace(5)* %r, align 2
149 store i16 2, i16 addrspace(5)* %gep.r, align 2
150 ret void
151 }
153 ; Should produce align 1 dword when legal
154 define i32 @private_load_2xi16_align1(i16 addrspace(5)* %p) #0 {
155 ; GFX7-ALIGNED-LABEL: private_load_2xi16_align1:
156 ; GFX7-ALIGNED: ; %bb.0:
157 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
158 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v1, vcc, 2, v0
159 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 1, v0
160 ; GFX7-ALIGNED-NEXT: buffer_load_ubyte v3, v0, s[0:3], 0 offen
161 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v0, vcc, 3, v0
162 ; GFX7-ALIGNED-NEXT: buffer_load_ubyte v0, v0, s[0:3], 0 offen
163 ; GFX7-ALIGNED-NEXT: buffer_load_ubyte v2, v2, s[0:3], 0 offen
164 ; GFX7-ALIGNED-NEXT: buffer_load_ubyte v1, v1, s[0:3], 0 offen
165 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(2)
166 ; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v0, 8, v0
167 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(1)
168 ; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v2, 8, v2
169 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0)
170 ; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v0, v1
171 ; GFX7-ALIGNED-NEXT: v_or_b32_e32 v2, v2, v3
172 ; GFX7-ALIGNED-NEXT: v_lshlrev_b32_e32 v0, 16, v0
173 ; GFX7-ALIGNED-NEXT: v_or_b32_e32 v0, v2, v0
174 ; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31]
175 ;
176 ; GFX7-UNALIGNED-LABEL: private_load_2xi16_align1:
177 ; GFX7-UNALIGNED: ; %bb.0:
178 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
179 ; GFX7-UNALIGNED-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
180 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0)
181 ; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31]
182 ;
183 ; GFX9-LABEL: private_load_2xi16_align1:
184 ; GFX9: ; %bb.0:
185 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
186 ; GFX9-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
187 ; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff
188 ; GFX9-NEXT: s_mov_b32 s4, 0xffff
189 ; GFX9-NEXT: s_waitcnt vmcnt(0)
190 ; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0
191 ; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1
192 ; GFX9-NEXT: s_setpc_b64 s[30:31]
193 ;
194 ; GFX9-FLASTSCR-LABEL: private_load_2xi16_align1:
195 ; GFX9-FLASTSCR: ; %bb.0:
196 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
197 ; GFX9-FLASTSCR-NEXT: scratch_load_dword v0, v0, off
198 ; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v1, 0xffff
199 ; GFX9-FLASTSCR-NEXT: s_mov_b32 s0, 0xffff
200 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
201 ; GFX9-FLASTSCR-NEXT: v_bfi_b32 v1, v1, 0, v0
202 ; GFX9-FLASTSCR-NEXT: v_and_or_b32 v0, v0, s0, v1
203 ; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
204 ;
205 ; GFX10-LABEL: private_load_2xi16_align1:
206 ; GFX10: ; %bb.0:
207 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
208 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
209 ; GFX10-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
210 ; GFX10-NEXT: s_waitcnt vmcnt(0)
211 ; GFX10-NEXT: v_bfi_b32 v1, 0xffff, 0, v0
212 ; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v1
213 ; GFX10-NEXT: s_setpc_b64 s[30:31]
214 ;
215 ; GFX10-FLASTSCR-LABEL: private_load_2xi16_align1:
216 ; GFX10-FLASTSCR: ; %bb.0:
217 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
218 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
219 ; GFX10-FLASTSCR-NEXT: scratch_load_dword v0, v0, off
220 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
221 ; GFX10-FLASTSCR-NEXT: v_bfi_b32 v1, 0xffff, 0, v0
222 ; GFX10-FLASTSCR-NEXT: v_and_or_b32 v0, 0xffff, v0, v1
223 ; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
224 %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1
225 %p.0 = load i16, i16 addrspace(5)* %p, align 1
226 %p.1 = load i16, i16 addrspace(5)* %gep.p, align 1
227 %zext.0 = zext i16 %p.0 to i32
228 %zext.1 = zext i16 %p.1 to i32
229 %shl.1 = shl i32 %zext.1, 16
230 %or = or i32 %zext.0, %shl.1
231 ret i32 %or
232 }
234 ; Should produce align 1 dword when legal
235 define void @private_store_2xi16_align1(i16 addrspace(5)* %p, i16 addrspace(5)* %r) #0 {
236 ; GFX7-ALIGNED-LABEL: private_store_2xi16_align1:
237 ; GFX7-ALIGNED: ; %bb.0:
238 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
239 ; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v3, 1
240 ; GFX7-ALIGNED-NEXT: buffer_store_byte v3, v1, s[0:3], 0 offen
241 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v2, vcc, 2, v1
242 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v3, vcc, 1, v1
243 ; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v4, 0
244 ; GFX7-ALIGNED-NEXT: v_add_i32_e32 v1, vcc, 3, v1
245 ; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, 2
246 ; GFX7-ALIGNED-NEXT: buffer_store_byte v4, v3, s[0:3], 0 offen
247 ; GFX7-ALIGNED-NEXT: buffer_store_byte v4, v1, s[0:3], 0 offen
248 ; GFX7-ALIGNED-NEXT: buffer_store_byte v0, v2, s[0:3], 0 offen
249 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0)
250 ; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31]
251 ;
252 ; GFX7-UNALIGNED-LABEL: private_store_2xi16_align1:
253 ; GFX7-UNALIGNED: ; %bb.0:
254 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
255 ; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, 0x20001
256 ; GFX7-UNALIGNED-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
257 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0)
258 ; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31]
259 ;
260 ; GFX9-LABEL: private_store_2xi16_align1:
261 ; GFX9: ; %bb.0:
262 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
263 ; GFX9-NEXT: v_mov_b32_e32 v0, 0x20001
264 ; GFX9-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
265 ; GFX9-NEXT: s_waitcnt vmcnt(0)
266 ; GFX9-NEXT: s_setpc_b64 s[30:31]
267 ;
268 ; GFX9-FLASTSCR-LABEL: private_store_2xi16_align1:
269 ; GFX9-FLASTSCR: ; %bb.0:
270 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
271 ; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001
272 ; GFX9-FLASTSCR-NEXT: scratch_store_dword v1, v0, off
273 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
274 ; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
275 ;
276 ; GFX10-LABEL: private_store_2xi16_align1:
277 ; GFX10: ; %bb.0:
278 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
279 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
280 ; GFX10-NEXT: v_mov_b32_e32 v0, 0x20001
281 ; GFX10-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
282 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
283 ; GFX10-NEXT: s_setpc_b64 s[30:31]
284 ;
285 ; GFX10-FLASTSCR-LABEL: private_store_2xi16_align1:
286 ; GFX10-FLASTSCR: ; %bb.0:
287 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
288 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
289 ; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001
290 ; GFX10-FLASTSCR-NEXT: scratch_store_dword v1, v0, off
291 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
292 ; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
293 %gep.r = getelementptr i16, i16 addrspace(5)* %r, i64 1
294 store i16 1, i16 addrspace(5)* %r, align 1
295 store i16 2, i16 addrspace(5)* %gep.r, align 1
296 ret void
297 }
299 ; Should merge this to a dword load
300 define i32 @private_load_2xi16_align4(i16 addrspace(5)* %p) #0 {
301 ; GFX7-LABEL: load_2xi16_align4:
302 ; GFX7: ; %bb.0:
303 ; GFX7-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
304 ; GFX7-NEXT: flat_load_dword v0, v[0:1]
305 ; GFX7-NEXT: s_waitcnt vmcnt(0) lgkmcnt(0)
306 ; GFX7-NEXT: s_setpc_b64 s[30:31]
307 ;
308 ; GFX7-ALIGNED-LABEL: private_load_2xi16_align4:
309 ; GFX7-ALIGNED: ; %bb.0:
310 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
311 ; GFX7-ALIGNED-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
312 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0)
313 ; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31]
314 ;
315 ; GFX7-UNALIGNED-LABEL: private_load_2xi16_align4:
316 ; GFX7-UNALIGNED: ; %bb.0:
317 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
318 ; GFX7-UNALIGNED-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
319 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0)
320 ; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31]
321 ;
322 ; GFX9-LABEL: private_load_2xi16_align4:
323 ; GFX9: ; %bb.0:
324 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
325 ; GFX9-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
326 ; GFX9-NEXT: v_mov_b32_e32 v1, 0xffff
327 ; GFX9-NEXT: s_mov_b32 s4, 0xffff
328 ; GFX9-NEXT: s_waitcnt vmcnt(0)
329 ; GFX9-NEXT: v_bfi_b32 v1, v1, 0, v0
330 ; GFX9-NEXT: v_and_or_b32 v0, v0, s4, v1
331 ; GFX9-NEXT: s_setpc_b64 s[30:31]
332 ;
333 ; GFX9-FLASTSCR-LABEL: private_load_2xi16_align4:
334 ; GFX9-FLASTSCR: ; %bb.0:
335 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
336 ; GFX9-FLASTSCR-NEXT: scratch_load_dword v0, v0, off
337 ; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v1, 0xffff
338 ; GFX9-FLASTSCR-NEXT: s_mov_b32 s0, 0xffff
339 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
340 ; GFX9-FLASTSCR-NEXT: v_bfi_b32 v1, v1, 0, v0
341 ; GFX9-FLASTSCR-NEXT: v_and_or_b32 v0, v0, s0, v1
342 ; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
343 ;
344 ; GFX10-LABEL: private_load_2xi16_align4:
345 ; GFX10: ; %bb.0:
346 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
347 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
348 ; GFX10-NEXT: buffer_load_dword v0, v0, s[0:3], 0 offen
349 ; GFX10-NEXT: s_waitcnt vmcnt(0)
350 ; GFX10-NEXT: v_bfi_b32 v1, 0xffff, 0, v0
351 ; GFX10-NEXT: v_and_or_b32 v0, 0xffff, v0, v1
352 ; GFX10-NEXT: s_setpc_b64 s[30:31]
353 ;
354 ; GFX10-FLASTSCR-LABEL: private_load_2xi16_align4:
355 ; GFX10-FLASTSCR: ; %bb.0:
356 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
357 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
358 ; GFX10-FLASTSCR-NEXT: scratch_load_dword v0, v0, off
359 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
360 ; GFX10-FLASTSCR-NEXT: v_bfi_b32 v1, 0xffff, 0, v0
361 ; GFX10-FLASTSCR-NEXT: v_and_or_b32 v0, 0xffff, v0, v1
362 ; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
363 %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1
364 %p.0 = load i16, i16 addrspace(5)* %p, align 4
365 %p.1 = load i16, i16 addrspace(5)* %gep.p, align 2
366 %zext.0 = zext i16 %p.0 to i32
367 %zext.1 = zext i16 %p.1 to i32
368 %shl.1 = shl i32 %zext.1, 16
369 %or = or i32 %zext.0, %shl.1
370 ret i32 %or
371 }
373 ; Should merge this to a dword store
374 define void @private_store_2xi16_align4(i16 addrspace(5)* %p, i16 addrspace(5)* %r) #0 {
375 ; GFX7-LABEL: private_store_2xi16_align4:
376 ; GFX7: ; %bb.0:
377 ; GFX7-NEXT: s_load_dwordx2 s[0:1], s[4:5], 0x2
378 ; GFX7-NEXT: v_mov_b32_e32 v2, 0x20001
379 ; GFX7-NEXT: s_waitcnt lgkmcnt(0)
380 ; GFX7-NEXT: v_mov_b32_e32 v0, s0
381 ; GFX7-NEXT: v_mov_b32_e32 v1, s1
382 ; GFX7-NEXT: flat_store_dword v[0:1], v2
383 ; GFX7-NEXT: s_endpgm
384 ;
385 ; GFX7-ALIGNED-LABEL: private_store_2xi16_align4:
386 ; GFX7-ALIGNED: ; %bb.0:
387 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
388 ; GFX7-ALIGNED-NEXT: v_mov_b32_e32 v0, 0x20001
389 ; GFX7-ALIGNED-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
390 ; GFX7-ALIGNED-NEXT: s_waitcnt vmcnt(0)
391 ; GFX7-ALIGNED-NEXT: s_setpc_b64 s[30:31]
392 ;
393 ; GFX7-UNALIGNED-LABEL: private_store_2xi16_align4:
394 ; GFX7-UNALIGNED: ; %bb.0:
395 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
396 ; GFX7-UNALIGNED-NEXT: v_mov_b32_e32 v0, 0x20001
397 ; GFX7-UNALIGNED-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
398 ; GFX7-UNALIGNED-NEXT: s_waitcnt vmcnt(0)
399 ; GFX7-UNALIGNED-NEXT: s_setpc_b64 s[30:31]
400 ;
401 ; GFX9-LABEL: private_store_2xi16_align4:
402 ; GFX9: ; %bb.0:
403 ; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
404 ; GFX9-NEXT: v_mov_b32_e32 v0, 0x20001
405 ; GFX9-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
406 ; GFX9-NEXT: s_waitcnt vmcnt(0)
407 ; GFX9-NEXT: s_setpc_b64 s[30:31]
408 ;
409 ; GFX9-FLASTSCR-LABEL: private_store_2xi16_align4:
410 ; GFX9-FLASTSCR: ; %bb.0:
411 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
412 ; GFX9-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001
413 ; GFX9-FLASTSCR-NEXT: scratch_store_dword v1, v0, off
414 ; GFX9-FLASTSCR-NEXT: s_waitcnt vmcnt(0)
415 ; GFX9-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
416 ;
417 ; GFX10-LABEL: private_store_2xi16_align4:
418 ; GFX10: ; %bb.0:
419 ; GFX10-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
420 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
421 ; GFX10-NEXT: v_mov_b32_e32 v0, 0x20001
422 ; GFX10-NEXT: buffer_store_dword v0, v1, s[0:3], 0 offen
423 ; GFX10-NEXT: s_waitcnt_vscnt null, 0x0
424 ; GFX10-NEXT: s_setpc_b64 s[30:31]
425 ;
426 ; GFX10-FLASTSCR-LABEL: private_store_2xi16_align4:
427 ; GFX10-FLASTSCR: ; %bb.0:
428 ; GFX10-FLASTSCR-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
429 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
430 ; GFX10-FLASTSCR-NEXT: v_mov_b32_e32 v0, 0x20001
431 ; GFX10-FLASTSCR-NEXT: scratch_store_dword v1, v0, off
432 ; GFX10-FLASTSCR-NEXT: s_waitcnt_vscnt null, 0x0
433 ; GFX10-FLASTSCR-NEXT: s_setpc_b64 s[30:31]
434 %gep.r = getelementptr i16, i16 addrspace(5)* %r, i64 1
435 store i16 1, i16 addrspace(5)* %r, align 4
436 store i16 2, i16 addrspace(5)* %gep.r, align 2
437 ret void
438 }