1 /* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
2
3 #include "test_sve_acle.h"
4
5 /*
6 ** subr_f16_m_tied1:
7 ** fsubr z0\.h, p0/m, z0\.h, z1\.h
8 ** ret
9 */
10 TEST_UNIFORM_Z (subr_f16_m_tied1, svfloat16_t,
11 z0 = svsubr_f16_m (p0, z0, z1),
12 z0 = svsubr_m (p0, z0, z1))
13
14 /*
15 ** subr_f16_m_tied2:
16 ** mov (z[0-9]+)\.d, z0\.d
17 ** movprfx z0, z1
18 ** fsubr z0\.h, p0/m, z0\.h, \1\.h
19 ** ret
20 */
21 TEST_UNIFORM_Z (subr_f16_m_tied2, svfloat16_t,
22 z0 = svsubr_f16_m (p0, z1, z0),
23 z0 = svsubr_m (p0, z1, z0))
24
25 /*
26 ** subr_f16_m_untied:
27 ** movprfx z0, z1
28 ** fsubr z0\.h, p0/m, z0\.h, z2\.h
29 ** ret
30 */
31 TEST_UNIFORM_Z (subr_f16_m_untied, svfloat16_t,
32 z0 = svsubr_f16_m (p0, z1, z2),
33 z0 = svsubr_m (p0, z1, z2))
34
35 /*
36 ** subr_h4_f16_m_tied1:
37 ** mov (z[0-9]+\.h), h4
38 ** fsubr z0\.h, p0/m, z0\.h, \1
39 ** ret
40 */
41 TEST_UNIFORM_ZD (subr_h4_f16_m_tied1, svfloat16_t, __fp16,
42 z0 = svsubr_n_f16_m (p0, z0, d4),
43 z0 = svsubr_m (p0, z0, d4))
44
45 /*
46 ** subr_h4_f16_m_untied:
47 ** mov (z[0-9]+\.h), h4
48 ** movprfx z0, z1
49 ** fsubr z0\.h, p0/m, z0\.h, \1
50 ** ret
51 */
52 TEST_UNIFORM_ZD (subr_h4_f16_m_untied, svfloat16_t, __fp16,
53 z0 = svsubr_n_f16_m (p0, z1, d4),
54 z0 = svsubr_m (p0, z1, d4))
55
56 /*
57 ** subr_1_f16_m_tied1:
58 ** fsubr z0\.h, p0/m, z0\.h, #1\.0
59 ** ret
60 */
61 TEST_UNIFORM_Z (subr_1_f16_m_tied1, svfloat16_t,
62 z0 = svsubr_n_f16_m (p0, z0, 1),
63 z0 = svsubr_m (p0, z0, 1))
64
65 /*
66 ** subr_1_f16_m_untied:
67 ** movprfx z0, z1
68 ** fsubr z0\.h, p0/m, z0\.h, #1\.0
69 ** ret
70 */
71 TEST_UNIFORM_Z (subr_1_f16_m_untied, svfloat16_t,
72 z0 = svsubr_n_f16_m (p0, z1, 1),
73 z0 = svsubr_m (p0, z1, 1))
74
75 /*
76 ** subr_0p5_f16_m_tied1:
77 ** fsubr z0\.h, p0/m, z0\.h, #0\.5
78 ** ret
79 */
80 TEST_UNIFORM_Z (subr_0p5_f16_m_tied1, svfloat16_t,
81 z0 = svsubr_n_f16_m (p0, z0, 0.5),
82 z0 = svsubr_m (p0, z0, 0.5))
83
84 /*
85 ** subr_0p5_f16_m_untied:
86 ** movprfx z0, z1
87 ** fsubr z0\.h, p0/m, z0\.h, #0\.5
88 ** ret
89 */
90 TEST_UNIFORM_Z (subr_0p5_f16_m_untied, svfloat16_t,
91 z0 = svsubr_n_f16_m (p0, z1, 0.5),
92 z0 = svsubr_m (p0, z1, 0.5))
93
94 /*
95 ** subr_m1_f16_m_tied1:
96 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
97 ** fsubr z0\.h, p0/m, z0\.h, \1
98 ** ret
99 */
100 TEST_UNIFORM_Z (subr_m1_f16_m_tied1, svfloat16_t,
101 z0 = svsubr_n_f16_m (p0, z0, -1),
102 z0 = svsubr_m (p0, z0, -1))
103
104 /*
105 ** subr_m1_f16_m_untied: { xfail *-*-* }
106 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
107 ** movprfx z0, z1
108 ** fsubr z0\.h, p0/m, z0\.h, \1
109 ** ret
110 */
111 TEST_UNIFORM_Z (subr_m1_f16_m_untied, svfloat16_t,
112 z0 = svsubr_n_f16_m (p0, z1, -1),
113 z0 = svsubr_m (p0, z1, -1))
114
115 /*
116 ** subr_f16_z_tied1:
117 ** movprfx z0\.h, p0/z, z0\.h
118 ** fsubr z0\.h, p0/m, z0\.h, z1\.h
119 ** ret
120 */
121 TEST_UNIFORM_Z (subr_f16_z_tied1, svfloat16_t,
122 z0 = svsubr_f16_z (p0, z0, z1),
123 z0 = svsubr_z (p0, z0, z1))
124
125 /*
126 ** subr_f16_z_tied2:
127 ** movprfx z0\.h, p0/z, z0\.h
128 ** fsub z0\.h, p0/m, z0\.h, z1\.h
129 ** ret
130 */
131 TEST_UNIFORM_Z (subr_f16_z_tied2, svfloat16_t,
132 z0 = svsubr_f16_z (p0, z1, z0),
133 z0 = svsubr_z (p0, z1, z0))
134
135 /*
136 ** subr_f16_z_untied:
137 ** (
138 ** movprfx z0\.h, p0/z, z1\.h
139 ** fsubr z0\.h, p0/m, z0\.h, z2\.h
140 ** |
141 ** movprfx z0\.h, p0/z, z2\.h
142 ** fsub z0\.h, p0/m, z0\.h, z1\.h
143 ** )
144 ** ret
145 */
146 TEST_UNIFORM_Z (subr_f16_z_untied, svfloat16_t,
147 z0 = svsubr_f16_z (p0, z1, z2),
148 z0 = svsubr_z (p0, z1, z2))
149
150 /*
151 ** subr_h4_f16_z_tied1:
152 ** mov (z[0-9]+\.h), h4
153 ** movprfx z0\.h, p0/z, z0\.h
154 ** fsubr z0\.h, p0/m, z0\.h, \1
155 ** ret
156 */
157 TEST_UNIFORM_ZD (subr_h4_f16_z_tied1, svfloat16_t, __fp16,
158 z0 = svsubr_n_f16_z (p0, z0, d4),
159 z0 = svsubr_z (p0, z0, d4))
160
161 /*
162 ** subr_h4_f16_z_untied:
163 ** mov (z[0-9]+\.h), h4
164 ** (
165 ** movprfx z0\.h, p0/z, z1\.h
166 ** fsubr z0\.h, p0/m, z0\.h, \1
167 ** |
168 ** movprfx z0\.h, p0/z, \1
169 ** fsub z0\.h, p0/m, z0\.h, z1\.h
170 ** )
171 ** ret
172 */
173 TEST_UNIFORM_ZD (subr_h4_f16_z_untied, svfloat16_t, __fp16,
174 z0 = svsubr_n_f16_z (p0, z1, d4),
175 z0 = svsubr_z (p0, z1, d4))
176
177 /*
178 ** subr_1_f16_z_tied1:
179 ** movprfx z0\.h, p0/z, z0\.h
180 ** fsubr z0\.h, p0/m, z0\.h, #1\.0
181 ** ret
182 */
183 TEST_UNIFORM_Z (subr_1_f16_z_tied1, svfloat16_t,
184 z0 = svsubr_n_f16_z (p0, z0, 1),
185 z0 = svsubr_z (p0, z0, 1))
186
187 /*
188 ** subr_1_f16_z_untied:
189 ** movprfx z0\.h, p0/z, z1\.h
190 ** fsubr z0\.h, p0/m, z0\.h, #1\.0
191 ** ret
192 */
193 TEST_UNIFORM_Z (subr_1_f16_z_untied, svfloat16_t,
194 z0 = svsubr_n_f16_z (p0, z1, 1),
195 z0 = svsubr_z (p0, z1, 1))
196
197 /*
198 ** subr_0p5_f16_z_tied1:
199 ** movprfx z0\.h, p0/z, z0\.h
200 ** fsubr z0\.h, p0/m, z0\.h, #0\.5
201 ** ret
202 */
203 TEST_UNIFORM_Z (subr_0p5_f16_z_tied1, svfloat16_t,
204 z0 = svsubr_n_f16_z (p0, z0, 0.5),
205 z0 = svsubr_z (p0, z0, 0.5))
206
207 /*
208 ** subr_0p5_f16_z_untied:
209 ** movprfx z0\.h, p0/z, z1\.h
210 ** fsubr z0\.h, p0/m, z0\.h, #0\.5
211 ** ret
212 */
213 TEST_UNIFORM_Z (subr_0p5_f16_z_untied, svfloat16_t,
214 z0 = svsubr_n_f16_z (p0, z1, 0.5),
215 z0 = svsubr_z (p0, z1, 0.5))
216
217 /*
218 ** subr_m1_f16_z_tied1:
219 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
220 ** movprfx z0\.h, p0/z, z0\.h
221 ** fsubr z0\.h, p0/m, z0\.h, \1
222 ** ret
223 */
224 TEST_UNIFORM_Z (subr_m1_f16_z_tied1, svfloat16_t,
225 z0 = svsubr_n_f16_z (p0, z0, -1),
226 z0 = svsubr_z (p0, z0, -1))
227
228 /*
229 ** subr_m1_f16_z_untied:
230 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
231 ** (
232 ** movprfx z0\.h, p0/z, z1\.h
233 ** fsubr z0\.h, p0/m, z0\.h, \1
234 ** |
235 ** movprfx z0\.h, p0/z, \1
236 ** fsub z0\.h, p0/m, z0\.h, z1\.h
237 ** )
238 ** ret
239 */
240 TEST_UNIFORM_Z (subr_m1_f16_z_untied, svfloat16_t,
241 z0 = svsubr_n_f16_z (p0, z1, -1),
242 z0 = svsubr_z (p0, z1, -1))
243
244 /*
245 ** subr_f16_x_tied1:
246 ** fsubr z0\.h, p0/m, z0\.h, z1\.h
247 ** ret
248 */
249 TEST_UNIFORM_Z (subr_f16_x_tied1, svfloat16_t,
250 z0 = svsubr_f16_x (p0, z0, z1),
251 z0 = svsubr_x (p0, z0, z1))
252
253 /*
254 ** subr_f16_x_tied2:
255 ** fsub z0\.h, p0/m, z0\.h, z1\.h
256 ** ret
257 */
258 TEST_UNIFORM_Z (subr_f16_x_tied2, svfloat16_t,
259 z0 = svsubr_f16_x (p0, z1, z0),
260 z0 = svsubr_x (p0, z1, z0))
261
262 /*
263 ** subr_f16_x_untied:
264 ** (
265 ** movprfx z0, z1
266 ** fsubr z0\.h, p0/m, z0\.h, z2\.h
267 ** |
268 ** movprfx z0, z2
269 ** fsub z0\.h, p0/m, z0\.h, z1\.h
270 ** )
271 ** ret
272 */
273 TEST_UNIFORM_Z (subr_f16_x_untied, svfloat16_t,
274 z0 = svsubr_f16_x (p0, z1, z2),
275 z0 = svsubr_x (p0, z1, z2))
276
277 /*
278 ** subr_h4_f16_x_tied1:
279 ** mov (z[0-9]+\.h), h4
280 ** fsubr z0\.h, p0/m, z0\.h, \1
281 ** ret
282 */
283 TEST_UNIFORM_ZD (subr_h4_f16_x_tied1, svfloat16_t, __fp16,
284 z0 = svsubr_n_f16_x (p0, z0, d4),
285 z0 = svsubr_x (p0, z0, d4))
286
287 /*
288 ** subr_h4_f16_x_untied:
289 ** mov z0\.h, h4
290 ** fsub z0\.h, p0/m, z0\.h, z1\.h
291 ** ret
292 */
293 TEST_UNIFORM_ZD (subr_h4_f16_x_untied, svfloat16_t, __fp16,
294 z0 = svsubr_n_f16_x (p0, z1, d4),
295 z0 = svsubr_x (p0, z1, d4))
296
297 /*
298 ** subr_1_f16_x_tied1:
299 ** fsubr z0\.h, p0/m, z0\.h, #1\.0
300 ** ret
301 */
302 TEST_UNIFORM_Z (subr_1_f16_x_tied1, svfloat16_t,
303 z0 = svsubr_n_f16_x (p0, z0, 1),
304 z0 = svsubr_x (p0, z0, 1))
305
306 /*
307 ** subr_1_f16_x_untied:
308 ** movprfx z0, z1
309 ** fsubr z0\.h, p0/m, z0\.h, #1\.0
310 ** ret
311 */
312 TEST_UNIFORM_Z (subr_1_f16_x_untied, svfloat16_t,
313 z0 = svsubr_n_f16_x (p0, z1, 1),
314 z0 = svsubr_x (p0, z1, 1))
315
316 /*
317 ** subr_0p5_f16_x_tied1:
318 ** fsubr z0\.h, p0/m, z0\.h, #0\.5
319 ** ret
320 */
321 TEST_UNIFORM_Z (subr_0p5_f16_x_tied1, svfloat16_t,
322 z0 = svsubr_n_f16_x (p0, z0, 0.5),
323 z0 = svsubr_x (p0, z0, 0.5))
324
325 /*
326 ** subr_0p5_f16_x_untied:
327 ** movprfx z0, z1
328 ** fsubr z0\.h, p0/m, z0\.h, #0\.5
329 ** ret
330 */
331 TEST_UNIFORM_Z (subr_0p5_f16_x_untied, svfloat16_t,
332 z0 = svsubr_n_f16_x (p0, z1, 0.5),
333 z0 = svsubr_x (p0, z1, 0.5))
334
335 /*
336 ** subr_m1_f16_x_tied1:
337 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
338 ** fsubr z0\.h, p0/m, z0\.h, \1
339 ** ret
340 */
341 TEST_UNIFORM_Z (subr_m1_f16_x_tied1, svfloat16_t,
342 z0 = svsubr_n_f16_x (p0, z0, -1),
343 z0 = svsubr_x (p0, z0, -1))
344
345 /*
346 ** subr_m1_f16_x_untied:
347 ** fmov z0\.h, #-1\.0(?:e\+0)?
348 ** fsub z0\.h, p0/m, z0\.h, z1\.h
349 ** ret
350 */
351 TEST_UNIFORM_Z (subr_m1_f16_x_untied, svfloat16_t,
352 z0 = svsubr_n_f16_x (p0, z1, -1),
353 z0 = svsubr_x (p0, z1, -1))
354
355 /*
356 ** ptrue_subr_f16_x_tied1:
357 ** fsub z0\.h, z1\.h, z0\.h
358 ** ret
359 */
360 TEST_UNIFORM_Z (ptrue_subr_f16_x_tied1, svfloat16_t,
361 z0 = svsubr_f16_x (svptrue_b16 (), z0, z1),
362 z0 = svsubr_x (svptrue_b16 (), z0, z1))
363
364 /*
365 ** ptrue_subr_f16_x_tied2:
366 ** fsub z0\.h, z0\.h, z1\.h
367 ** ret
368 */
369 TEST_UNIFORM_Z (ptrue_subr_f16_x_tied2, svfloat16_t,
370 z0 = svsubr_f16_x (svptrue_b16 (), z1, z0),
371 z0 = svsubr_x (svptrue_b16 (), z1, z0))
372
373 /*
374 ** ptrue_subr_f16_x_untied:
375 ** fsub z0\.h, z2\.h, z1\.h
376 ** ret
377 */
378 TEST_UNIFORM_Z (ptrue_subr_f16_x_untied, svfloat16_t,
379 z0 = svsubr_f16_x (svptrue_b16 (), z1, z2),
380 z0 = svsubr_x (svptrue_b16 (), z1, z2))
381
382 /*
383 ** ptrue_subr_1_f16_x_tied1:
384 ** ...
385 ** ptrue p[0-9]+\.b[^\n]*
386 ** ...
387 ** ret
388 */
389 TEST_UNIFORM_Z (ptrue_subr_1_f16_x_tied1, svfloat16_t,
390 z0 = svsubr_n_f16_x (svptrue_b16 (), z0, 1),
391 z0 = svsubr_x (svptrue_b16 (), z0, 1))
392
393 /*
394 ** ptrue_subr_1_f16_x_untied:
395 ** ...
396 ** ptrue p[0-9]+\.b[^\n]*
397 ** ...
398 ** ret
399 */
400 TEST_UNIFORM_Z (ptrue_subr_1_f16_x_untied, svfloat16_t,
401 z0 = svsubr_n_f16_x (svptrue_b16 (), z1, 1),
402 z0 = svsubr_x (svptrue_b16 (), z1, 1))
403
404 /*
405 ** ptrue_subr_0p5_f16_x_tied1:
406 ** ...
407 ** ptrue p[0-9]+\.b[^\n]*
408 ** ...
409 ** ret
410 */
411 TEST_UNIFORM_Z (ptrue_subr_0p5_f16_x_tied1, svfloat16_t,
412 z0 = svsubr_n_f16_x (svptrue_b16 (), z0, 0.5),
413 z0 = svsubr_x (svptrue_b16 (), z0, 0.5))
414
415 /*
416 ** ptrue_subr_0p5_f16_x_untied:
417 ** ...
418 ** ptrue p[0-9]+\.b[^\n]*
419 ** ...
420 ** ret
421 */
422 TEST_UNIFORM_Z (ptrue_subr_0p5_f16_x_untied, svfloat16_t,
423 z0 = svsubr_n_f16_x (svptrue_b16 (), z1, 0.5),
424 z0 = svsubr_x (svptrue_b16 (), z1, 0.5))
425
426 /*
427 ** ptrue_subr_m1_f16_x_tied1:
428 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
429 ** fsub z0\.h, \1, z0\.h
430 ** ret
431 */
432 TEST_UNIFORM_Z (ptrue_subr_m1_f16_x_tied1, svfloat16_t,
433 z0 = svsubr_n_f16_x (svptrue_b16 (), z0, -1),
434 z0 = svsubr_x (svptrue_b16 (), z0, -1))
435
436 /*
437 ** ptrue_subr_m1_f16_x_untied:
438 ** fmov (z[0-9]+\.h), #-1\.0(?:e\+0)?
439 ** fsub z0\.h, \1, z1\.h
440 ** ret
441 */
442 TEST_UNIFORM_Z (ptrue_subr_m1_f16_x_untied, svfloat16_t,
443 z0 = svsubr_n_f16_x (svptrue_b16 (), z1, -1),
444 z0 = svsubr_x (svptrue_b16 (), z1, -1))