1 /* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
2
3 #include "test_sve_acle.h"
4
5 /*
6 ** abd_f32_m_tied1:
7 ** fabd z0\.s, p0/m, z0\.s, z1\.s
8 ** ret
9 */
10 TEST_UNIFORM_Z (abd_f32_m_tied1, svfloat32_t,
11 z0 = svabd_f32_m (p0, z0, z1),
12 z0 = svabd_m (p0, z0, z1))
13
14 /*
15 ** abd_f32_m_tied2:
16 ** mov (z[0-9]+)\.d, z0\.d
17 ** movprfx z0, z1
18 ** fabd z0\.s, p0/m, z0\.s, \1\.s
19 ** ret
20 */
21 TEST_UNIFORM_Z (abd_f32_m_tied2, svfloat32_t,
22 z0 = svabd_f32_m (p0, z1, z0),
23 z0 = svabd_m (p0, z1, z0))
24
25 /*
26 ** abd_f32_m_untied:
27 ** movprfx z0, z1
28 ** fabd z0\.s, p0/m, z0\.s, z2\.s
29 ** ret
30 */
31 TEST_UNIFORM_Z (abd_f32_m_untied, svfloat32_t,
32 z0 = svabd_f32_m (p0, z1, z2),
33 z0 = svabd_m (p0, z1, z2))
34
35 /*
36 ** abd_s4_f32_m_tied1:
37 ** mov (z[0-9]+\.s), s4
38 ** fabd z0\.s, p0/m, z0\.s, \1
39 ** ret
40 */
41 TEST_UNIFORM_ZD (abd_s4_f32_m_tied1, svfloat32_t, float,
42 z0 = svabd_n_f32_m (p0, z0, d4),
43 z0 = svabd_m (p0, z0, d4))
44
45 /*
46 ** abd_s4_f32_m_untied:
47 ** mov (z[0-9]+\.s), s4
48 ** movprfx z0, z1
49 ** fabd z0\.s, p0/m, z0\.s, \1
50 ** ret
51 */
52 TEST_UNIFORM_ZD (abd_s4_f32_m_untied, svfloat32_t, float,
53 z0 = svabd_n_f32_m (p0, z1, d4),
54 z0 = svabd_m (p0, z1, d4))
55
56 /*
57 ** abd_1_f32_m_tied1:
58 ** fmov (z[0-9]+\.s), #1\.0(?:e\+0)?
59 ** fabd z0\.s, p0/m, z0\.s, \1
60 ** ret
61 */
62 TEST_UNIFORM_Z (abd_1_f32_m_tied1, svfloat32_t,
63 z0 = svabd_n_f32_m (p0, z0, 1),
64 z0 = svabd_m (p0, z0, 1))
65
66 /*
67 ** abd_1_f32_m_untied: { xfail *-*-* }
68 ** fmov (z[0-9]+\.s), #1\.0(?:e\+0)?
69 ** movprfx z0, z1
70 ** fabd z0\.s, p0/m, z0\.s, \1
71 ** ret
72 */
73 TEST_UNIFORM_Z (abd_1_f32_m_untied, svfloat32_t,
74 z0 = svabd_n_f32_m (p0, z1, 1),
75 z0 = svabd_m (p0, z1, 1))
76
77 /*
78 ** abd_f32_z_tied1:
79 ** movprfx z0\.s, p0/z, z0\.s
80 ** fabd z0\.s, p0/m, z0\.s, z1\.s
81 ** ret
82 */
83 TEST_UNIFORM_Z (abd_f32_z_tied1, svfloat32_t,
84 z0 = svabd_f32_z (p0, z0, z1),
85 z0 = svabd_z (p0, z0, z1))
86
87 /*
88 ** abd_f32_z_tied2:
89 ** movprfx z0\.s, p0/z, z0\.s
90 ** fabd z0\.s, p0/m, z0\.s, z1\.s
91 ** ret
92 */
93 TEST_UNIFORM_Z (abd_f32_z_tied2, svfloat32_t,
94 z0 = svabd_f32_z (p0, z1, z0),
95 z0 = svabd_z (p0, z1, z0))
96
97 /*
98 ** abd_f32_z_untied:
99 ** (
100 ** movprfx z0\.s, p0/z, z1\.s
101 ** fabd z0\.s, p0/m, z0\.s, z2\.s
102 ** |
103 ** movprfx z0\.s, p0/z, z2\.s
104 ** fabd z0\.s, p0/m, z0\.s, z1\.s
105 ** )
106 ** ret
107 */
108 TEST_UNIFORM_Z (abd_f32_z_untied, svfloat32_t,
109 z0 = svabd_f32_z (p0, z1, z2),
110 z0 = svabd_z (p0, z1, z2))
111
112 /*
113 ** abd_s4_f32_z_tied1:
114 ** mov (z[0-9]+\.s), s4
115 ** movprfx z0\.s, p0/z, z0\.s
116 ** fabd z0\.s, p0/m, z0\.s, \1
117 ** ret
118 */
119 TEST_UNIFORM_ZD (abd_s4_f32_z_tied1, svfloat32_t, float,
120 z0 = svabd_n_f32_z (p0, z0, d4),
121 z0 = svabd_z (p0, z0, d4))
122
123 /*
124 ** abd_s4_f32_z_untied:
125 ** mov (z[0-9]+\.s), s4
126 ** (
127 ** movprfx z0\.s, p0/z, z1\.s
128 ** fabd z0\.s, p0/m, z0\.s, \1
129 ** |
130 ** movprfx z0\.s, p0/z, \1
131 ** fabd z0\.s, p0/m, z0\.s, z1\.s
132 ** )
133 ** ret
134 */
135 TEST_UNIFORM_ZD (abd_s4_f32_z_untied, svfloat32_t, float,
136 z0 = svabd_n_f32_z (p0, z1, d4),
137 z0 = svabd_z (p0, z1, d4))
138
139 /*
140 ** abd_1_f32_z_tied1:
141 ** fmov (z[0-9]+\.s), #1\.0(?:e\+0)?
142 ** movprfx z0\.s, p0/z, z0\.s
143 ** fabd z0\.s, p0/m, z0\.s, \1
144 ** ret
145 */
146 TEST_UNIFORM_Z (abd_1_f32_z_tied1, svfloat32_t,
147 z0 = svabd_n_f32_z (p0, z0, 1),
148 z0 = svabd_z (p0, z0, 1))
149
150 /*
151 ** abd_1_f32_z_untied:
152 ** fmov (z[0-9]+\.s), #1\.0(?:e\+0)?
153 ** (
154 ** movprfx z0\.s, p0/z, z1\.s
155 ** fabd z0\.s, p0/m, z0\.s, \1
156 ** |
157 ** movprfx z0\.s, p0/z, \1
158 ** fabd z0\.s, p0/m, z0\.s, z1\.s
159 ** )
160 ** ret
161 */
162 TEST_UNIFORM_Z (abd_1_f32_z_untied, svfloat32_t,
163 z0 = svabd_n_f32_z (p0, z1, 1),
164 z0 = svabd_z (p0, z1, 1))
165
166 /*
167 ** abd_0p5_f32_z_tied1:
168 ** fmov (z[0-9]+\.s), #(?:0\.5|5\.0e-1)
169 ** movprfx z0\.s, p0/z, z0\.s
170 ** fabd z0\.s, p0/m, z0\.s, \1
171 ** ret
172 */
173 TEST_UNIFORM_Z (abd_0p5_f32_z_tied1, svfloat32_t,
174 z0 = svabd_n_f32_z (p0, z0, 0.5),
175 z0 = svabd_z (p0, z0, 0.5))
176
177 /*
178 ** abd_0p5_f32_z_untied:
179 ** fmov (z[0-9]+\.s), #(?:0\.5|5\.0e-1)
180 ** (
181 ** movprfx z0\.s, p0/z, z1\.s
182 ** fabd z0\.s, p0/m, z0\.s, \1
183 ** |
184 ** movprfx z0\.s, p0/z, \1
185 ** fabd z0\.s, p0/m, z0\.s, z1\.s
186 ** )
187 ** ret
188 */
189 TEST_UNIFORM_Z (abd_0p5_f32_z_untied, svfloat32_t,
190 z0 = svabd_n_f32_z (p0, z1, 0.5),
191 z0 = svabd_z (p0, z1, 0.5))
192
193 /*
194 ** abd_m1_f32_z_tied1:
195 ** fmov (z[0-9]+\.s), #-1\.0(?:e\+0)?
196 ** movprfx z0\.s, p0/z, z0\.s
197 ** fabd z0\.s, p0/m, z0\.s, \1
198 ** ret
199 */
200 TEST_UNIFORM_Z (abd_m1_f32_z_tied1, svfloat32_t,
201 z0 = svabd_n_f32_z (p0, z0, -1),
202 z0 = svabd_z (p0, z0, -1))
203
204 /*
205 ** abd_m1_f32_z_untied:
206 ** fmov (z[0-9]+\.s), #-1\.0(?:e\+0)?
207 ** (
208 ** movprfx z0\.s, p0/z, z1\.s
209 ** fabd z0\.s, p0/m, z0\.s, \1
210 ** |
211 ** movprfx z0\.s, p0/z, \1
212 ** fabd z0\.s, p0/m, z0\.s, z1\.s
213 ** )
214 ** ret
215 */
216 TEST_UNIFORM_Z (abd_m1_f32_z_untied, svfloat32_t,
217 z0 = svabd_n_f32_z (p0, z1, -1),
218 z0 = svabd_z (p0, z1, -1))
219
220 /*
221 ** abd_m0p5_f32_z_tied1:
222 ** fmov (z[0-9]+\.s), #-(?:0\.5|5\.0e-1)
223 ** movprfx z0\.s, p0/z, z0\.s
224 ** fabd z0\.s, p0/m, z0\.s, \1
225 ** ret
226 */
227 TEST_UNIFORM_Z (abd_m0p5_f32_z_tied1, svfloat32_t,
228 z0 = svabd_n_f32_z (p0, z0, -0.5),
229 z0 = svabd_z (p0, z0, -0.5))
230
231 /*
232 ** abd_m0p5_f32_z_untied:
233 ** fmov (z[0-9]+\.s), #-(?:0\.5|5\.0e-1)
234 ** (
235 ** movprfx z0\.s, p0/z, z1\.s
236 ** fabd z0\.s, p0/m, z0\.s, \1
237 ** |
238 ** movprfx z0\.s, p0/z, \1
239 ** fabd z0\.s, p0/m, z0\.s, z1\.s
240 ** )
241 ** ret
242 */
243 TEST_UNIFORM_Z (abd_m0p5_f32_z_untied, svfloat32_t,
244 z0 = svabd_n_f32_z (p0, z1, -0.5),
245 z0 = svabd_z (p0, z1, -0.5))
246
247 /*
248 ** abd_m2_f32_z:
249 ** fmov (z[0-9]+\.s), #-2\.0(?:e\+0)?
250 ** movprfx z0\.s, p0/z, z0\.s
251 ** fabd z0\.s, p0/m, z0\.s, \1
252 ** ret
253 */
254 TEST_UNIFORM_Z (abd_m2_f32_z, svfloat32_t,
255 z0 = svabd_n_f32_z (p0, z0, -2),
256 z0 = svabd_z (p0, z0, -2))
257
258 /*
259 ** abd_f32_x_tied1:
260 ** fabd z0\.s, p0/m, z0\.s, z1\.s
261 ** ret
262 */
263 TEST_UNIFORM_Z (abd_f32_x_tied1, svfloat32_t,
264 z0 = svabd_f32_x (p0, z0, z1),
265 z0 = svabd_x (p0, z0, z1))
266
267 /*
268 ** abd_f32_x_tied2:
269 ** fabd z0\.s, p0/m, z0\.s, z1\.s
270 ** ret
271 */
272 TEST_UNIFORM_Z (abd_f32_x_tied2, svfloat32_t,
273 z0 = svabd_f32_x (p0, z1, z0),
274 z0 = svabd_x (p0, z1, z0))
275
276 /*
277 ** abd_f32_x_untied:
278 ** (
279 ** movprfx z0, z1
280 ** fabd z0\.s, p0/m, z0\.s, z2\.s
281 ** |
282 ** movprfx z0, z2
283 ** fabd z0\.s, p0/m, z0\.s, z1\.s
284 ** )
285 ** ret
286 */
287 TEST_UNIFORM_Z (abd_f32_x_untied, svfloat32_t,
288 z0 = svabd_f32_x (p0, z1, z2),
289 z0 = svabd_x (p0, z1, z2))
290
291 /*
292 ** abd_s4_f32_x_tied1:
293 ** mov (z[0-9]+\.s), s4
294 ** fabd z0\.s, p0/m, z0\.s, \1
295 ** ret
296 */
297 TEST_UNIFORM_ZD (abd_s4_f32_x_tied1, svfloat32_t, float,
298 z0 = svabd_n_f32_x (p0, z0, d4),
299 z0 = svabd_x (p0, z0, d4))
300
301 /*
302 ** abd_s4_f32_x_untied:
303 ** mov z0\.s, s4
304 ** fabd z0\.s, p0/m, z0\.s, z1\.s
305 ** ret
306 */
307 TEST_UNIFORM_ZD (abd_s4_f32_x_untied, svfloat32_t, float,
308 z0 = svabd_n_f32_x (p0, z1, d4),
309 z0 = svabd_x (p0, z1, d4))
310
311 /*
312 ** abd_1_f32_x_tied1:
313 ** fmov (z[0-9]+\.s), #1\.0(?:e\+0)?
314 ** fabd z0\.s, p0/m, z0\.s, \1
315 ** ret
316 */
317 TEST_UNIFORM_Z (abd_1_f32_x_tied1, svfloat32_t,
318 z0 = svabd_n_f32_x (p0, z0, 1),
319 z0 = svabd_x (p0, z0, 1))
320
321 /*
322 ** abd_1_f32_x_untied:
323 ** fmov z0\.s, #1\.0(?:e\+0)?
324 ** fabd z0\.s, p0/m, z0\.s, z1\.s
325 ** ret
326 */
327 TEST_UNIFORM_Z (abd_1_f32_x_untied, svfloat32_t,
328 z0 = svabd_n_f32_x (p0, z1, 1),
329 z0 = svabd_x (p0, z1, 1))
330
331 /*
332 ** abd_0p5_f32_x_tied1:
333 ** fmov (z[0-9]+\.s), #(?:0\.5|5\.0e-1)
334 ** fabd z0\.s, p0/m, z0\.s, \1
335 ** ret
336 */
337 TEST_UNIFORM_Z (abd_0p5_f32_x_tied1, svfloat32_t,
338 z0 = svabd_n_f32_x (p0, z0, 0.5),
339 z0 = svabd_x (p0, z0, 0.5))
340
341 /*
342 ** abd_0p5_f32_x_untied:
343 ** fmov z0\.s, #(?:0\.5|5\.0e-1)
344 ** fabd z0\.s, p0/m, z0\.s, z1\.s
345 ** ret
346 */
347 TEST_UNIFORM_Z (abd_0p5_f32_x_untied, svfloat32_t,
348 z0 = svabd_n_f32_x (p0, z1, 0.5),
349 z0 = svabd_x (p0, z1, 0.5))
350
351 /*
352 ** abd_m1_f32_x_tied1:
353 ** fmov (z[0-9]+\.s), #-1\.0(?:e\+0)?
354 ** fabd z0\.s, p0/m, z0\.s, \1
355 ** ret
356 */
357 TEST_UNIFORM_Z (abd_m1_f32_x_tied1, svfloat32_t,
358 z0 = svabd_n_f32_x (p0, z0, -1),
359 z0 = svabd_x (p0, z0, -1))
360
361 /*
362 ** abd_m1_f32_x_untied:
363 ** fmov z0\.s, #-1\.0(?:e\+0)?
364 ** fabd z0\.s, p0/m, z0\.s, z1\.s
365 ** ret
366 */
367 TEST_UNIFORM_Z (abd_m1_f32_x_untied, svfloat32_t,
368 z0 = svabd_n_f32_x (p0, z1, -1),
369 z0 = svabd_x (p0, z1, -1))
370
371 /*
372 ** abd_m0p5_f32_x_tied1:
373 ** fmov (z[0-9]+\.s), #-(?:0\.5|5\.0e-1)
374 ** fabd z0\.s, p0/m, z0\.s, \1
375 ** ret
376 */
377 TEST_UNIFORM_Z (abd_m0p5_f32_x_tied1, svfloat32_t,
378 z0 = svabd_n_f32_x (p0, z0, -0.5),
379 z0 = svabd_x (p0, z0, -0.5))
380
381 /*
382 ** abd_m0p5_f32_x_untied:
383 ** fmov z0\.s, #-(?:0\.5|5\.0e-1)
384 ** fabd z0\.s, p0/m, z0\.s, z1\.s
385 ** ret
386 */
387 TEST_UNIFORM_Z (abd_m0p5_f32_x_untied, svfloat32_t,
388 z0 = svabd_n_f32_x (p0, z1, -0.5),
389 z0 = svabd_x (p0, z1, -0.5))
390
391 /*
392 ** abd_2_f32_x_tied1:
393 ** fmov (z[0-9]+\.s), #2\.0(?:e\+0)?
394 ** fabd z0\.s, p0/m, z0\.s, \1
395 ** ret
396 */
397 TEST_UNIFORM_Z (abd_2_f32_x_tied1, svfloat32_t,
398 z0 = svabd_n_f32_x (p0, z0, 2),
399 z0 = svabd_x (p0, z0, 2))
400
401 /*
402 ** abd_2_f32_x_untied:
403 ** fmov z0\.s, #2\.0(?:e\+0)?
404 ** fabd z0\.s, p0/m, z0\.s, z1\.s
405 ** ret
406 */
407 TEST_UNIFORM_Z (abd_2_f32_x_untied, svfloat32_t,
408 z0 = svabd_n_f32_x (p0, z1, 2),
409 z0 = svabd_x (p0, z1, 2))
410
411 /*
412 ** ptrue_abd_f32_x_tied1:
413 ** ...
414 ** ptrue p[0-9]+\.b[^\n]*
415 ** ...
416 ** ret
417 */
418 TEST_UNIFORM_Z (ptrue_abd_f32_x_tied1, svfloat32_t,
419 z0 = svabd_f32_x (svptrue_b32 (), z0, z1),
420 z0 = svabd_x (svptrue_b32 (), z0, z1))
421
422 /*
423 ** ptrue_abd_f32_x_tied2:
424 ** ...
425 ** ptrue p[0-9]+\.b[^\n]*
426 ** ...
427 ** ret
428 */
429 TEST_UNIFORM_Z (ptrue_abd_f32_x_tied2, svfloat32_t,
430 z0 = svabd_f32_x (svptrue_b32 (), z1, z0),
431 z0 = svabd_x (svptrue_b32 (), z1, z0))
432
433 /*
434 ** ptrue_abd_f32_x_untied:
435 ** ...
436 ** ptrue p[0-9]+\.b[^\n]*
437 ** ...
438 ** ret
439 */
440 TEST_UNIFORM_Z (ptrue_abd_f32_x_untied, svfloat32_t,
441 z0 = svabd_f32_x (svptrue_b32 (), z1, z2),
442 z0 = svabd_x (svptrue_b32 (), z1, z2))
443
444 /*
445 ** ptrue_abd_1_f32_x_tied1:
446 ** ...
447 ** ptrue p[0-9]+\.b[^\n]*
448 ** ...
449 ** ret
450 */
451 TEST_UNIFORM_Z (ptrue_abd_1_f32_x_tied1, svfloat32_t,
452 z0 = svabd_n_f32_x (svptrue_b32 (), z0, 1),
453 z0 = svabd_x (svptrue_b32 (), z0, 1))
454
455 /*
456 ** ptrue_abd_1_f32_x_untied:
457 ** ...
458 ** ptrue p[0-9]+\.b[^\n]*
459 ** ...
460 ** ret
461 */
462 TEST_UNIFORM_Z (ptrue_abd_1_f32_x_untied, svfloat32_t,
463 z0 = svabd_n_f32_x (svptrue_b32 (), z1, 1),
464 z0 = svabd_x (svptrue_b32 (), z1, 1))
465
466 /*
467 ** ptrue_abd_0p5_f32_x_tied1:
468 ** ...
469 ** ptrue p[0-9]+\.b[^\n]*
470 ** ...
471 ** ret
472 */
473 TEST_UNIFORM_Z (ptrue_abd_0p5_f32_x_tied1, svfloat32_t,
474 z0 = svabd_n_f32_x (svptrue_b32 (), z0, 0.5),
475 z0 = svabd_x (svptrue_b32 (), z0, 0.5))
476
477 /*
478 ** ptrue_abd_0p5_f32_x_untied:
479 ** ...
480 ** ptrue p[0-9]+\.b[^\n]*
481 ** ...
482 ** ret
483 */
484 TEST_UNIFORM_Z (ptrue_abd_0p5_f32_x_untied, svfloat32_t,
485 z0 = svabd_n_f32_x (svptrue_b32 (), z1, 0.5),
486 z0 = svabd_x (svptrue_b32 (), z1, 0.5))
487
488 /*
489 ** ptrue_abd_m1_f32_x_tied1:
490 ** ...
491 ** ptrue p[0-9]+\.b[^\n]*
492 ** ...
493 ** ret
494 */
495 TEST_UNIFORM_Z (ptrue_abd_m1_f32_x_tied1, svfloat32_t,
496 z0 = svabd_n_f32_x (svptrue_b32 (), z0, -1),
497 z0 = svabd_x (svptrue_b32 (), z0, -1))
498
499 /*
500 ** ptrue_abd_m1_f32_x_untied:
501 ** ...
502 ** ptrue p[0-9]+\.b[^\n]*
503 ** ...
504 ** ret
505 */
506 TEST_UNIFORM_Z (ptrue_abd_m1_f32_x_untied, svfloat32_t,
507 z0 = svabd_n_f32_x (svptrue_b32 (), z1, -1),
508 z0 = svabd_x (svptrue_b32 (), z1, -1))
509
510 /*
511 ** ptrue_abd_m0p5_f32_x_tied1:
512 ** ...
513 ** ptrue p[0-9]+\.b[^\n]*
514 ** ...
515 ** ret
516 */
517 TEST_UNIFORM_Z (ptrue_abd_m0p5_f32_x_tied1, svfloat32_t,
518 z0 = svabd_n_f32_x (svptrue_b32 (), z0, -0.5),
519 z0 = svabd_x (svptrue_b32 (), z0, -0.5))
520
521 /*
522 ** ptrue_abd_m0p5_f32_x_untied:
523 ** ...
524 ** ptrue p[0-9]+\.b[^\n]*
525 ** ...
526 ** ret
527 */
528 TEST_UNIFORM_Z (ptrue_abd_m0p5_f32_x_untied, svfloat32_t,
529 z0 = svabd_n_f32_x (svptrue_b32 (), z1, -0.5),
530 z0 = svabd_x (svptrue_b32 (), z1, -0.5))
531
532 /*
533 ** ptrue_abd_2_f32_x_tied1:
534 ** ...
535 ** ptrue p[0-9]+\.b[^\n]*
536 ** ...
537 ** ret
538 */
539 TEST_UNIFORM_Z (ptrue_abd_2_f32_x_tied1, svfloat32_t,
540 z0 = svabd_n_f32_x (svptrue_b32 (), z0, 2),
541 z0 = svabd_x (svptrue_b32 (), z0, 2))
542
543 /*
544 ** ptrue_abd_2_f32_x_untied:
545 ** ...
546 ** ptrue p[0-9]+\.b[^\n]*
547 ** ...
548 ** ret
549 */
550 TEST_UNIFORM_Z (ptrue_abd_2_f32_x_untied, svfloat32_t,
551 z0 = svabd_n_f32_x (svptrue_b32 (), z1, 2),
552 z0 = svabd_x (svptrue_b32 (), z1, 2))