| blob | ca160c091b229340cc7295a7773314e54712afec |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mattr=+simd128 | FileCheck %s
4 ;; Test that SIMD bitmask instruction can be selected
6 target triple = "wasm32-unknown-unknown"
8 define i16 @bitmask_v16i8(<16 x i8> %v) {
9 ; CHECK-LABEL: bitmask_v16i8:
10 ; CHECK: .functype bitmask_v16i8 (v128) -> (i32)
11 ; CHECK-NEXT: # %bb.0:
12 ; CHECK-NEXT: local.get 0
13 ; CHECK-NEXT: v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
14 ; CHECK-NEXT: i8x16.eq
15 ; CHECK-NEXT: i8x16.bitmask
16 ; CHECK-NEXT: # fallthrough-return
17 %cmp = icmp eq <16 x i8> %v, zeroinitializer
18 %bitmask = bitcast <16 x i1> %cmp to i16
19 ret i16 %bitmask
20 }
22 define i8 @bitmask_v8i16(<8 x i16> %v) {
23 ; CHECK-LABEL: bitmask_v8i16:
24 ; CHECK: .functype bitmask_v8i16 (v128) -> (i32)
25 ; CHECK-NEXT: # %bb.0:
26 ; CHECK-NEXT: local.get 0
27 ; CHECK-NEXT: v128.const 0, 0, 0, 0, 0, 0, 0, 0
28 ; CHECK-NEXT: i16x8.eq
29 ; CHECK-NEXT: i16x8.bitmask
30 ; CHECK-NEXT: # fallthrough-return
31 %cmp = icmp eq <8 x i16> %v, zeroinitializer
32 %bitmask = bitcast <8 x i1> %cmp to i8
33 ret i8 %bitmask
34 }
36 define i8 @bitmask_v4i32(<4 x i32> %v) {
37 ; CHECK-LABEL: bitmask_v4i32:
38 ; CHECK: .functype bitmask_v4i32 (v128) -> (i32)
39 ; CHECK-NEXT: # %bb.0:
40 ; CHECK-NEXT: local.get 0
41 ; CHECK-NEXT: v128.const 0, 0, 0, 0
42 ; CHECK-NEXT: i32x4.eq
43 ; CHECK-NEXT: i32x4.bitmask
44 ; CHECK-NEXT: # fallthrough-return
45 %cmp = icmp eq <4 x i32> %v, zeroinitializer
46 %bitmask = bitcast <4 x i1> %cmp to i4
47 %ext = zext i4 %bitmask to i8
48 ret i8 %ext
49 }
51 define i8 @bitmask_v2i64(<2 x i64> %v) {
52 ; CHECK-LABEL: bitmask_v2i64:
53 ; CHECK: .functype bitmask_v2i64 (v128) -> (i32)
54 ; CHECK-NEXT: # %bb.0:
55 ; CHECK-NEXT: local.get 0
56 ; CHECK-NEXT: v128.const 0, 0
57 ; CHECK-NEXT: i64x2.eq
58 ; CHECK-NEXT: i64x2.bitmask
59 ; CHECK-NEXT: # fallthrough-return
60 %cmp = icmp eq <2 x i64> %v, zeroinitializer
61 %bitmask = bitcast <2 x i1> %cmp to i2
62 %ext = zext i2 %bitmask to i8
63 ret i8 %ext
64 }
66 ;; Test unusual vectors
68 define i1 @bitmask_v1i8(<1 x i8> %v) {
69 ; CHECK-LABEL: bitmask_v1i8:
70 ; CHECK: .functype bitmask_v1i8 (v128) -> (i32)
71 ; CHECK-NEXT: # %bb.0:
72 ; CHECK-NEXT: local.get 0
73 ; CHECK-NEXT: i8x16.extract_lane_u 0
74 ; CHECK-NEXT: i32.eqz
75 ; CHECK-NEXT: # fallthrough-return
76 %cmp = icmp eq <1 x i8> %v, zeroinitializer
77 %bitmask = bitcast <1 x i1> %cmp to i1
78 ret i1 %bitmask
79 }
81 define i7 @bitmask_v7i8(<7 x i8> %v) {
82 ; CHECK-LABEL: bitmask_v7i8:
83 ; CHECK: .functype bitmask_v7i8 (i32, i32, i32, i32, i32, i32, i32) -> (i32)
84 ; CHECK-NEXT: .local v128
85 ; CHECK-NEXT: # %bb.0:
86 ; CHECK-NEXT: global.get __stack_pointer
87 ; CHECK-NEXT: i32.const 16
88 ; CHECK-NEXT: i32.sub
89 ; CHECK-NEXT: drop
90 ; CHECK-NEXT: local.get 0
91 ; CHECK-NEXT: i8x16.splat
92 ; CHECK-NEXT: local.get 1
93 ; CHECK-NEXT: i8x16.replace_lane 1
94 ; CHECK-NEXT: local.get 2
95 ; CHECK-NEXT: i8x16.replace_lane 2
96 ; CHECK-NEXT: local.get 3
97 ; CHECK-NEXT: i8x16.replace_lane 3
98 ; CHECK-NEXT: local.get 4
99 ; CHECK-NEXT: i8x16.replace_lane 4
100 ; CHECK-NEXT: local.get 5
101 ; CHECK-NEXT: i8x16.replace_lane 5
102 ; CHECK-NEXT: local.get 6
103 ; CHECK-NEXT: i8x16.replace_lane 6
104 ; CHECK-NEXT: v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
105 ; CHECK-NEXT: i8x16.eq
106 ; CHECK-NEXT: local.tee 7
107 ; CHECK-NEXT: i16x8.extract_lane_u 0
108 ; CHECK-NEXT: i32.const 1
109 ; CHECK-NEXT: i32.and
110 ; CHECK-NEXT: local.get 7
111 ; CHECK-NEXT: i16x8.extend_low_i8x16_s
112 ; CHECK-NEXT: local.tee 7
113 ; CHECK-NEXT: i16x8.extract_lane_u 1
114 ; CHECK-NEXT: i32.const 1
115 ; CHECK-NEXT: i32.and
116 ; CHECK-NEXT: i32.const 1
117 ; CHECK-NEXT: i32.shl
118 ; CHECK-NEXT: i32.or
119 ; CHECK-NEXT: local.get 7
120 ; CHECK-NEXT: i16x8.extract_lane_u 2
121 ; CHECK-NEXT: i32.const 1
122 ; CHECK-NEXT: i32.and
123 ; CHECK-NEXT: i32.const 2
124 ; CHECK-NEXT: i32.shl
125 ; CHECK-NEXT: i32.or
126 ; CHECK-NEXT: local.get 7
127 ; CHECK-NEXT: i16x8.extract_lane_u 3
128 ; CHECK-NEXT: i32.const 1
129 ; CHECK-NEXT: i32.and
130 ; CHECK-NEXT: i32.const 3
131 ; CHECK-NEXT: i32.shl
132 ; CHECK-NEXT: i32.or
133 ; CHECK-NEXT: local.get 7
134 ; CHECK-NEXT: i16x8.extract_lane_u 4
135 ; CHECK-NEXT: i32.const 1
136 ; CHECK-NEXT: i32.and
137 ; CHECK-NEXT: i32.const 4
138 ; CHECK-NEXT: i32.shl
139 ; CHECK-NEXT: i32.or
140 ; CHECK-NEXT: local.get 7
141 ; CHECK-NEXT: i16x8.extract_lane_u 5
142 ; CHECK-NEXT: i32.const 1
143 ; CHECK-NEXT: i32.and
144 ; CHECK-NEXT: i32.const 5
145 ; CHECK-NEXT: i32.shl
146 ; CHECK-NEXT: i32.or
147 ; CHECK-NEXT: local.get 7
148 ; CHECK-NEXT: i16x8.extract_lane_u 6
149 ; CHECK-NEXT: i32.const 6
150 ; CHECK-NEXT: i32.shl
151 ; CHECK-NEXT: i32.or
152 ; CHECK-NEXT: i32.const 127
153 ; CHECK-NEXT: i32.and
154 ; CHECK-NEXT: # fallthrough-return
155 %cmp = icmp eq <7 x i8> %v, zeroinitializer
156 %bitmask = bitcast <7 x i1> %cmp to i7
157 ret i7 %bitmask
158 }
160 define i8 @bitmask_v8i8(<8 x i8> %v) {
161 ; CHECK-LABEL: bitmask_v8i8:
162 ; CHECK: .functype bitmask_v8i8 (v128) -> (i32)
163 ; CHECK-NEXT: # %bb.0:
164 ; CHECK-NEXT: local.get 0
165 ; CHECK-NEXT: v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
166 ; CHECK-NEXT: i8x16.eq
167 ; CHECK-NEXT: i16x8.extend_low_i8x16_s
168 ; CHECK-NEXT: i16x8.bitmask
169 ; CHECK-NEXT: # fallthrough-return
170 %cmp = icmp eq <8 x i8> %v, zeroinitializer
171 %bitmask = bitcast <8 x i1> %cmp to i8
172 ret i8 %bitmask
173 }
175 define i32 @bitmask_v32i8(<32 x i8> %v) {
176 ; CHECK-LABEL: bitmask_v32i8:
177 ; CHECK: .functype bitmask_v32i8 (v128, v128) -> (i32)
178 ; CHECK-NEXT: .local v128
179 ; CHECK-NEXT: # %bb.0:
180 ; CHECK-NEXT: global.get __stack_pointer
181 ; CHECK-NEXT: i32.const 16
182 ; CHECK-NEXT: i32.sub
183 ; CHECK-NEXT: drop
184 ; CHECK-NEXT: local.get 0
185 ; CHECK-NEXT: v128.const 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
186 ; CHECK-NEXT: local.tee 2
187 ; CHECK-NEXT: i8x16.eq
188 ; CHECK-NEXT: local.tee 0
189 ; CHECK-NEXT: i8x16.extract_lane_u 0
190 ; CHECK-NEXT: i32.const 1
191 ; CHECK-NEXT: i32.and
192 ; CHECK-NEXT: local.get 0
193 ; CHECK-NEXT: i8x16.extract_lane_u 1
194 ; CHECK-NEXT: i32.const 1
195 ; CHECK-NEXT: i32.and
196 ; CHECK-NEXT: i32.const 1
197 ; CHECK-NEXT: i32.shl
198 ; CHECK-NEXT: i32.or
199 ; CHECK-NEXT: local.get 0
200 ; CHECK-NEXT: i8x16.extract_lane_u 2
201 ; CHECK-NEXT: i32.const 1
202 ; CHECK-NEXT: i32.and
203 ; CHECK-NEXT: i32.const 2
204 ; CHECK-NEXT: i32.shl
205 ; CHECK-NEXT: i32.or
206 ; CHECK-NEXT: local.get 0
207 ; CHECK-NEXT: i8x16.extract_lane_u 3
208 ; CHECK-NEXT: i32.const 1
209 ; CHECK-NEXT: i32.and
210 ; CHECK-NEXT: i32.const 3
211 ; CHECK-NEXT: i32.shl
212 ; CHECK-NEXT: i32.or
213 ; CHECK-NEXT: local.get 0
214 ; CHECK-NEXT: i8x16.extract_lane_u 4
215 ; CHECK-NEXT: i32.const 1
216 ; CHECK-NEXT: i32.and
217 ; CHECK-NEXT: i32.const 4
218 ; CHECK-NEXT: i32.shl
219 ; CHECK-NEXT: i32.or
220 ; CHECK-NEXT: local.get 0
221 ; CHECK-NEXT: i8x16.extract_lane_u 5
222 ; CHECK-NEXT: i32.const 1
223 ; CHECK-NEXT: i32.and
224 ; CHECK-NEXT: i32.const 5
225 ; CHECK-NEXT: i32.shl
226 ; CHECK-NEXT: i32.or
227 ; CHECK-NEXT: local.get 0
228 ; CHECK-NEXT: i8x16.extract_lane_u 6
229 ; CHECK-NEXT: i32.const 1
230 ; CHECK-NEXT: i32.and
231 ; CHECK-NEXT: i32.const 6
232 ; CHECK-NEXT: i32.shl
233 ; CHECK-NEXT: i32.or
234 ; CHECK-NEXT: local.get 0
235 ; CHECK-NEXT: i8x16.extract_lane_u 7
236 ; CHECK-NEXT: i32.const 1
237 ; CHECK-NEXT: i32.and
238 ; CHECK-NEXT: i32.const 7
239 ; CHECK-NEXT: i32.shl
240 ; CHECK-NEXT: i32.or
241 ; CHECK-NEXT: local.get 0
242 ; CHECK-NEXT: i8x16.extract_lane_u 8
243 ; CHECK-NEXT: i32.const 1
244 ; CHECK-NEXT: i32.and
245 ; CHECK-NEXT: i32.const 8
246 ; CHECK-NEXT: i32.shl
247 ; CHECK-NEXT: i32.or
248 ; CHECK-NEXT: local.get 0
249 ; CHECK-NEXT: i8x16.extract_lane_u 9
250 ; CHECK-NEXT: i32.const 1
251 ; CHECK-NEXT: i32.and
252 ; CHECK-NEXT: i32.const 9
253 ; CHECK-NEXT: i32.shl
254 ; CHECK-NEXT: i32.or
255 ; CHECK-NEXT: local.get 0
256 ; CHECK-NEXT: i8x16.extract_lane_u 10
257 ; CHECK-NEXT: i32.const 1
258 ; CHECK-NEXT: i32.and
259 ; CHECK-NEXT: i32.const 10
260 ; CHECK-NEXT: i32.shl
261 ; CHECK-NEXT: i32.or
262 ; CHECK-NEXT: local.get 0
263 ; CHECK-NEXT: i8x16.extract_lane_u 11
264 ; CHECK-NEXT: i32.const 1
265 ; CHECK-NEXT: i32.and
266 ; CHECK-NEXT: i32.const 11
267 ; CHECK-NEXT: i32.shl
268 ; CHECK-NEXT: i32.or
269 ; CHECK-NEXT: local.get 0
270 ; CHECK-NEXT: i8x16.extract_lane_u 12
271 ; CHECK-NEXT: i32.const 1
272 ; CHECK-NEXT: i32.and
273 ; CHECK-NEXT: i32.const 12
274 ; CHECK-NEXT: i32.shl
275 ; CHECK-NEXT: i32.or
276 ; CHECK-NEXT: local.get 0
277 ; CHECK-NEXT: i8x16.extract_lane_u 13
278 ; CHECK-NEXT: i32.const 1
279 ; CHECK-NEXT: i32.and
280 ; CHECK-NEXT: i32.const 13
281 ; CHECK-NEXT: i32.shl
282 ; CHECK-NEXT: i32.or
283 ; CHECK-NEXT: local.get 0
284 ; CHECK-NEXT: i8x16.extract_lane_u 14
285 ; CHECK-NEXT: i32.const 1
286 ; CHECK-NEXT: i32.and
287 ; CHECK-NEXT: i32.const 14
288 ; CHECK-NEXT: i32.shl
289 ; CHECK-NEXT: i32.or
290 ; CHECK-NEXT: local.get 0
291 ; CHECK-NEXT: i8x16.extract_lane_u 15
292 ; CHECK-NEXT: i32.const 15
293 ; CHECK-NEXT: i32.shl
294 ; CHECK-NEXT: i32.or
295 ; CHECK-NEXT: i32.const 65535
296 ; CHECK-NEXT: i32.and
297 ; CHECK-NEXT: local.get 1
298 ; CHECK-NEXT: local.get 2
299 ; CHECK-NEXT: i8x16.eq
300 ; CHECK-NEXT: local.tee 0
301 ; CHECK-NEXT: i8x16.extract_lane_u 15
302 ; CHECK-NEXT: i32.const 31
303 ; CHECK-NEXT: i32.shl
304 ; CHECK-NEXT: local.get 0
305 ; CHECK-NEXT: i8x16.extract_lane_u 14
306 ; CHECK-NEXT: i32.const 1
307 ; CHECK-NEXT: i32.and
308 ; CHECK-NEXT: i32.const 30
309 ; CHECK-NEXT: i32.shl
310 ; CHECK-NEXT: local.get 0
311 ; CHECK-NEXT: i8x16.extract_lane_u 13
312 ; CHECK-NEXT: i32.const 1
313 ; CHECK-NEXT: i32.and
314 ; CHECK-NEXT: i32.const 29
315 ; CHECK-NEXT: i32.shl
316 ; CHECK-NEXT: local.get 0
317 ; CHECK-NEXT: i8x16.extract_lane_u 12
318 ; CHECK-NEXT: i32.const 1
319 ; CHECK-NEXT: i32.and
320 ; CHECK-NEXT: i32.const 28
321 ; CHECK-NEXT: i32.shl
322 ; CHECK-NEXT: local.get 0
323 ; CHECK-NEXT: i8x16.extract_lane_u 11
324 ; CHECK-NEXT: i32.const 1
325 ; CHECK-NEXT: i32.and
326 ; CHECK-NEXT: i32.const 27
327 ; CHECK-NEXT: i32.shl
328 ; CHECK-NEXT: local.get 0
329 ; CHECK-NEXT: i8x16.extract_lane_u 10
330 ; CHECK-NEXT: i32.const 1
331 ; CHECK-NEXT: i32.and
332 ; CHECK-NEXT: i32.const 26
333 ; CHECK-NEXT: i32.shl
334 ; CHECK-NEXT: local.get 0
335 ; CHECK-NEXT: i8x16.extract_lane_u 9
336 ; CHECK-NEXT: i32.const 1
337 ; CHECK-NEXT: i32.and
338 ; CHECK-NEXT: i32.const 25
339 ; CHECK-NEXT: i32.shl
340 ; CHECK-NEXT: local.get 0
341 ; CHECK-NEXT: i8x16.extract_lane_u 8
342 ; CHECK-NEXT: i32.const 1
343 ; CHECK-NEXT: i32.and
344 ; CHECK-NEXT: i32.const 24
345 ; CHECK-NEXT: i32.shl
346 ; CHECK-NEXT: local.get 0
347 ; CHECK-NEXT: i8x16.extract_lane_u 7
348 ; CHECK-NEXT: i32.const 1
349 ; CHECK-NEXT: i32.and
350 ; CHECK-NEXT: i32.const 23
351 ; CHECK-NEXT: i32.shl
352 ; CHECK-NEXT: local.get 0
353 ; CHECK-NEXT: i8x16.extract_lane_u 6
354 ; CHECK-NEXT: i32.const 1
355 ; CHECK-NEXT: i32.and
356 ; CHECK-NEXT: i32.const 22
357 ; CHECK-NEXT: i32.shl
358 ; CHECK-NEXT: local.get 0
359 ; CHECK-NEXT: i8x16.extract_lane_u 5
360 ; CHECK-NEXT: i32.const 1
361 ; CHECK-NEXT: i32.and
362 ; CHECK-NEXT: i32.const 21
363 ; CHECK-NEXT: i32.shl
364 ; CHECK-NEXT: local.get 0
365 ; CHECK-NEXT: i8x16.extract_lane_u 4
366 ; CHECK-NEXT: i32.const 1
367 ; CHECK-NEXT: i32.and
368 ; CHECK-NEXT: i32.const 20
369 ; CHECK-NEXT: i32.shl
370 ; CHECK-NEXT: local.get 0
371 ; CHECK-NEXT: i8x16.extract_lane_u 3
372 ; CHECK-NEXT: i32.const 1
373 ; CHECK-NEXT: i32.and
374 ; CHECK-NEXT: i32.const 19
375 ; CHECK-NEXT: i32.shl
376 ; CHECK-NEXT: local.get 0
377 ; CHECK-NEXT: i8x16.extract_lane_u 2
378 ; CHECK-NEXT: i32.const 1
379 ; CHECK-NEXT: i32.and
380 ; CHECK-NEXT: i32.const 18
381 ; CHECK-NEXT: i32.shl
382 ; CHECK-NEXT: local.get 0
383 ; CHECK-NEXT: i8x16.extract_lane_u 1
384 ; CHECK-NEXT: i32.const 1
385 ; CHECK-NEXT: i32.and
386 ; CHECK-NEXT: i32.const 17
387 ; CHECK-NEXT: i32.shl
388 ; CHECK-NEXT: local.get 0
389 ; CHECK-NEXT: i8x16.extract_lane_u 0
390 ; CHECK-NEXT: i32.const 1
391 ; CHECK-NEXT: i32.and
392 ; CHECK-NEXT: i32.const 16
393 ; CHECK-NEXT: i32.shl
394 ; CHECK-NEXT: i32.or
395 ; CHECK-NEXT: i32.or
396 ; CHECK-NEXT: i32.or
397 ; CHECK-NEXT: i32.or
398 ; CHECK-NEXT: i32.or
399 ; CHECK-NEXT: i32.or
400 ; CHECK-NEXT: i32.or
401 ; CHECK-NEXT: i32.or
402 ; CHECK-NEXT: i32.or
403 ; CHECK-NEXT: i32.or
404 ; CHECK-NEXT: i32.or
405 ; CHECK-NEXT: i32.or
406 ; CHECK-NEXT: i32.or
407 ; CHECK-NEXT: i32.or
408 ; CHECK-NEXT: i32.or
409 ; CHECK-NEXT: i32.or
410 ; CHECK-NEXT: # fallthrough-return
411 %cmp = icmp eq <32 x i8> %v, zeroinitializer
412 %bitmask = bitcast <32 x i1> %cmp to i32
413 ret i32 %bitmask
414 }