1 /* { dg-final { check-function-bodies "**" "" "-DCHECK_ASM" } } */
2
3 #include "test_sve_acle.h"
4
5 /*
6 ** max_f32_m_tied1:
7 ** fmax z0\.s, p0/m, z0\.s, z1\.s
8 ** ret
9 */
10 TEST_UNIFORM_Z (max_f32_m_tied1, svfloat32_t,
11 z0 = svmax_f32_m (p0, z0, z1),
12 z0 = svmax_m (p0, z0, z1))
13
14 /*
15 ** max_f32_m_tied2:
16 ** mov (z[0-9]+)\.d, z0\.d
17 ** movprfx z0, z1
18 ** fmax z0\.s, p0/m, z0\.s, \1\.s
19 ** ret
20 */
21 TEST_UNIFORM_Z (max_f32_m_tied2, svfloat32_t,
22 z0 = svmax_f32_m (p0, z1, z0),
23 z0 = svmax_m (p0, z1, z0))
24
25 /*
26 ** max_f32_m_untied:
27 ** movprfx z0, z1
28 ** fmax z0\.s, p0/m, z0\.s, z2\.s
29 ** ret
30 */
31 TEST_UNIFORM_Z (max_f32_m_untied, svfloat32_t,
32 z0 = svmax_f32_m (p0, z1, z2),
33 z0 = svmax_m (p0, z1, z2))
34
35 /*
36 ** max_s4_f32_m_tied1:
37 ** mov (z[0-9]+\.s), s4
38 ** fmax z0\.s, p0/m, z0\.s, \1
39 ** ret
40 */
41 TEST_UNIFORM_ZD (max_s4_f32_m_tied1, svfloat32_t, float,
42 z0 = svmax_n_f32_m (p0, z0, d4),
43 z0 = svmax_m (p0, z0, d4))
44
45 /*
46 ** max_s4_f32_m_untied:
47 ** mov (z[0-9]+\.s), s4
48 ** movprfx z0, z1
49 ** fmax z0\.s, p0/m, z0\.s, \1
50 ** ret
51 */
52 TEST_UNIFORM_ZD (max_s4_f32_m_untied, svfloat32_t, float,
53 z0 = svmax_n_f32_m (p0, z1, d4),
54 z0 = svmax_m (p0, z1, d4))
55
56 /*
57 ** max_0_f32_m_tied1:
58 ** fmax z0\.s, p0/m, z0\.s, #0\.0
59 ** ret
60 */
61 TEST_UNIFORM_Z (max_0_f32_m_tied1, svfloat32_t,
62 z0 = svmax_n_f32_m (p0, z0, 0),
63 z0 = svmax_m (p0, z0, 0))
64
65 /*
66 ** max_0_f32_m_untied:
67 ** movprfx z0, z1
68 ** fmax z0\.s, p0/m, z0\.s, #0\.0
69 ** ret
70 */
71 TEST_UNIFORM_Z (max_0_f32_m_untied, svfloat32_t,
72 z0 = svmax_n_f32_m (p0, z1, 0),
73 z0 = svmax_m (p0, z1, 0))
74
75 /*
76 ** max_1_f32_m_tied1:
77 ** fmax z0\.s, p0/m, z0\.s, #1\.0
78 ** ret
79 */
80 TEST_UNIFORM_Z (max_1_f32_m_tied1, svfloat32_t,
81 z0 = svmax_n_f32_m (p0, z0, 1),
82 z0 = svmax_m (p0, z0, 1))
83
84 /*
85 ** max_1_f32_m_untied:
86 ** movprfx z0, z1
87 ** fmax z0\.s, p0/m, z0\.s, #1\.0
88 ** ret
89 */
90 TEST_UNIFORM_Z (max_1_f32_m_untied, svfloat32_t,
91 z0 = svmax_n_f32_m (p0, z1, 1),
92 z0 = svmax_m (p0, z1, 1))
93
94 /*
95 ** max_2_f32_m:
96 ** fmov (z[0-9]+\.s), #2\.0(?:e\+0)?
97 ** fmax z0\.s, p0/m, z0\.s, \1
98 ** ret
99 */
100 TEST_UNIFORM_Z (max_2_f32_m, svfloat32_t,
101 z0 = svmax_n_f32_m (p0, z0, 2),
102 z0 = svmax_m (p0, z0, 2))
103
104 /*
105 ** max_f32_z_tied1:
106 ** movprfx z0\.s, p0/z, z0\.s
107 ** fmax z0\.s, p0/m, z0\.s, z1\.s
108 ** ret
109 */
110 TEST_UNIFORM_Z (max_f32_z_tied1, svfloat32_t,
111 z0 = svmax_f32_z (p0, z0, z1),
112 z0 = svmax_z (p0, z0, z1))
113
114 /*
115 ** max_f32_z_tied2:
116 ** movprfx z0\.s, p0/z, z0\.s
117 ** fmax z0\.s, p0/m, z0\.s, z1\.s
118 ** ret
119 */
120 TEST_UNIFORM_Z (max_f32_z_tied2, svfloat32_t,
121 z0 = svmax_f32_z (p0, z1, z0),
122 z0 = svmax_z (p0, z1, z0))
123
124 /*
125 ** max_f32_z_untied:
126 ** (
127 ** movprfx z0\.s, p0/z, z1\.s
128 ** fmax z0\.s, p0/m, z0\.s, z2\.s
129 ** |
130 ** movprfx z0\.s, p0/z, z2\.s
131 ** fmax z0\.s, p0/m, z0\.s, z1\.s
132 ** )
133 ** ret
134 */
135 TEST_UNIFORM_Z (max_f32_z_untied, svfloat32_t,
136 z0 = svmax_f32_z (p0, z1, z2),
137 z0 = svmax_z (p0, z1, z2))
138
139 /*
140 ** max_s4_f32_z_tied1:
141 ** mov (z[0-9]+\.s), s4
142 ** movprfx z0\.s, p0/z, z0\.s
143 ** fmax z0\.s, p0/m, z0\.s, \1
144 ** ret
145 */
146 TEST_UNIFORM_ZD (max_s4_f32_z_tied1, svfloat32_t, float,
147 z0 = svmax_n_f32_z (p0, z0, d4),
148 z0 = svmax_z (p0, z0, d4))
149
150 /*
151 ** max_s4_f32_z_untied:
152 ** mov (z[0-9]+\.s), s4
153 ** (
154 ** movprfx z0\.s, p0/z, z1\.s
155 ** fmax z0\.s, p0/m, z0\.s, \1
156 ** |
157 ** movprfx z0\.s, p0/z, \1
158 ** fmax z0\.s, p0/m, z0\.s, z1\.s
159 ** )
160 ** ret
161 */
162 TEST_UNIFORM_ZD (max_s4_f32_z_untied, svfloat32_t, float,
163 z0 = svmax_n_f32_z (p0, z1, d4),
164 z0 = svmax_z (p0, z1, d4))
165
166 /*
167 ** max_0_f32_z_tied1:
168 ** movprfx z0\.s, p0/z, z0\.s
169 ** fmax z0\.s, p0/m, z0\.s, #0\.0
170 ** ret
171 */
172 TEST_UNIFORM_Z (max_0_f32_z_tied1, svfloat32_t,
173 z0 = svmax_n_f32_z (p0, z0, 0),
174 z0 = svmax_z (p0, z0, 0))
175
176 /*
177 ** max_0_f32_z_untied:
178 ** movprfx z0\.s, p0/z, z1\.s
179 ** fmax z0\.s, p0/m, z0\.s, #0\.0
180 ** ret
181 */
182 TEST_UNIFORM_Z (max_0_f32_z_untied, svfloat32_t,
183 z0 = svmax_n_f32_z (p0, z1, 0),
184 z0 = svmax_z (p0, z1, 0))
185
186 /*
187 ** max_1_f32_z_tied1:
188 ** movprfx z0\.s, p0/z, z0\.s
189 ** fmax z0\.s, p0/m, z0\.s, #1\.0
190 ** ret
191 */
192 TEST_UNIFORM_Z (max_1_f32_z_tied1, svfloat32_t,
193 z0 = svmax_n_f32_z (p0, z0, 1),
194 z0 = svmax_z (p0, z0, 1))
195
196 /*
197 ** max_1_f32_z_untied:
198 ** movprfx z0\.s, p0/z, z1\.s
199 ** fmax z0\.s, p0/m, z0\.s, #1\.0
200 ** ret
201 */
202 TEST_UNIFORM_Z (max_1_f32_z_untied, svfloat32_t,
203 z0 = svmax_n_f32_z (p0, z1, 1),
204 z0 = svmax_z (p0, z1, 1))
205
206 /*
207 ** max_2_f32_z:
208 ** fmov (z[0-9]+\.s), #2\.0(?:e\+0)?
209 ** movprfx z0\.s, p0/z, z0\.s
210 ** fmax z0\.s, p0/m, z0\.s, \1
211 ** ret
212 */
213 TEST_UNIFORM_Z (max_2_f32_z, svfloat32_t,
214 z0 = svmax_n_f32_z (p0, z0, 2),
215 z0 = svmax_z (p0, z0, 2))
216
217 /*
218 ** max_f32_x_tied1:
219 ** fmax z0\.s, p0/m, z0\.s, z1\.s
220 ** ret
221 */
222 TEST_UNIFORM_Z (max_f32_x_tied1, svfloat32_t,
223 z0 = svmax_f32_x (p0, z0, z1),
224 z0 = svmax_x (p0, z0, z1))
225
226 /*
227 ** max_f32_x_tied2:
228 ** fmax z0\.s, p0/m, z0\.s, z1\.s
229 ** ret
230 */
231 TEST_UNIFORM_Z (max_f32_x_tied2, svfloat32_t,
232 z0 = svmax_f32_x (p0, z1, z0),
233 z0 = svmax_x (p0, z1, z0))
234
235 /*
236 ** max_f32_x_untied:
237 ** (
238 ** movprfx z0, z1
239 ** fmax z0\.s, p0/m, z0\.s, z2\.s
240 ** |
241 ** movprfx z0, z2
242 ** fmax z0\.s, p0/m, z0\.s, z1\.s
243 ** )
244 ** ret
245 */
246 TEST_UNIFORM_Z (max_f32_x_untied, svfloat32_t,
247 z0 = svmax_f32_x (p0, z1, z2),
248 z0 = svmax_x (p0, z1, z2))
249
250 /*
251 ** max_s4_f32_x_tied1:
252 ** mov (z[0-9]+\.s), s4
253 ** fmax z0\.s, p0/m, z0\.s, \1
254 ** ret
255 */
256 TEST_UNIFORM_ZD (max_s4_f32_x_tied1, svfloat32_t, float,
257 z0 = svmax_n_f32_x (p0, z0, d4),
258 z0 = svmax_x (p0, z0, d4))
259
260 /*
261 ** max_s4_f32_x_untied:
262 ** mov z0\.s, s4
263 ** fmax z0\.s, p0/m, z0\.s, z1\.s
264 ** ret
265 */
266 TEST_UNIFORM_ZD (max_s4_f32_x_untied, svfloat32_t, float,
267 z0 = svmax_n_f32_x (p0, z1, d4),
268 z0 = svmax_x (p0, z1, d4))
269
270 /*
271 ** max_0_f32_x_tied1:
272 ** fmax z0\.s, p0/m, z0\.s, #0\.0
273 ** ret
274 */
275 TEST_UNIFORM_Z (max_0_f32_x_tied1, svfloat32_t,
276 z0 = svmax_n_f32_x (p0, z0, 0),
277 z0 = svmax_x (p0, z0, 0))
278
279 /*
280 ** max_0_f32_x_untied:
281 ** movprfx z0, z1
282 ** fmax z0\.s, p0/m, z0\.s, #0\.0
283 ** ret
284 */
285 TEST_UNIFORM_Z (max_0_f32_x_untied, svfloat32_t,
286 z0 = svmax_n_f32_x (p0, z1, 0),
287 z0 = svmax_x (p0, z1, 0))
288
289 /*
290 ** max_1_f32_x_tied1:
291 ** fmax z0\.s, p0/m, z0\.s, #1\.0
292 ** ret
293 */
294 TEST_UNIFORM_Z (max_1_f32_x_tied1, svfloat32_t,
295 z0 = svmax_n_f32_x (p0, z0, 1),
296 z0 = svmax_x (p0, z0, 1))
297
298 /*
299 ** max_1_f32_x_untied:
300 ** movprfx z0, z1
301 ** fmax z0\.s, p0/m, z0\.s, #1\.0
302 ** ret
303 */
304 TEST_UNIFORM_Z (max_1_f32_x_untied, svfloat32_t,
305 z0 = svmax_n_f32_x (p0, z1, 1),
306 z0 = svmax_x (p0, z1, 1))
307
308 /*
309 ** max_2_f32_x_tied1:
310 ** fmov (z[0-9]+\.s), #2\.0(?:e\+0)?
311 ** fmax z0\.s, p0/m, z0\.s, \1
312 ** ret
313 */
314 TEST_UNIFORM_Z (max_2_f32_x_tied1, svfloat32_t,
315 z0 = svmax_n_f32_x (p0, z0, 2),
316 z0 = svmax_x (p0, z0, 2))
317
318 /*
319 ** max_2_f32_x_untied:
320 ** fmov z0\.s, #2\.0(?:e\+0)?
321 ** fmax z0\.s, p0/m, z0\.s, z1\.s
322 ** ret
323 */
324 TEST_UNIFORM_Z (max_2_f32_x_untied, svfloat32_t,
325 z0 = svmax_n_f32_x (p0, z1, 2),
326 z0 = svmax_x (p0, z1, 2))
327
328 /*
329 ** ptrue_max_f32_x_tied1:
330 ** ...
331 ** ptrue p[0-9]+\.b[^\n]*
332 ** ...
333 ** ret
334 */
335 TEST_UNIFORM_Z (ptrue_max_f32_x_tied1, svfloat32_t,
336 z0 = svmax_f32_x (svptrue_b32 (), z0, z1),
337 z0 = svmax_x (svptrue_b32 (), z0, z1))
338
339 /*
340 ** ptrue_max_f32_x_tied2:
341 ** ...
342 ** ptrue p[0-9]+\.b[^\n]*
343 ** ...
344 ** ret
345 */
346 TEST_UNIFORM_Z (ptrue_max_f32_x_tied2, svfloat32_t,
347 z0 = svmax_f32_x (svptrue_b32 (), z1, z0),
348 z0 = svmax_x (svptrue_b32 (), z1, z0))
349
350 /*
351 ** ptrue_max_f32_x_untied:
352 ** ...
353 ** ptrue p[0-9]+\.b[^\n]*
354 ** ...
355 ** ret
356 */
357 TEST_UNIFORM_Z (ptrue_max_f32_x_untied, svfloat32_t,
358 z0 = svmax_f32_x (svptrue_b32 (), z1, z2),
359 z0 = svmax_x (svptrue_b32 (), z1, z2))
360
361 /*
362 ** ptrue_max_0_f32_x_tied1:
363 ** ...
364 ** ptrue p[0-9]+\.b[^\n]*
365 ** ...
366 ** ret
367 */
368 TEST_UNIFORM_Z (ptrue_max_0_f32_x_tied1, svfloat32_t,
369 z0 = svmax_n_f32_x (svptrue_b32 (), z0, 0),
370 z0 = svmax_x (svptrue_b32 (), z0, 0))
371
372 /*
373 ** ptrue_max_0_f32_x_untied:
374 ** ...
375 ** ptrue p[0-9]+\.b[^\n]*
376 ** ...
377 ** ret
378 */
379 TEST_UNIFORM_Z (ptrue_max_0_f32_x_untied, svfloat32_t,
380 z0 = svmax_n_f32_x (svptrue_b32 (), z1, 0),
381 z0 = svmax_x (svptrue_b32 (), z1, 0))
382
383 /*
384 ** ptrue_max_1_f32_x_tied1:
385 ** ...
386 ** ptrue p[0-9]+\.b[^\n]*
387 ** ...
388 ** ret
389 */
390 TEST_UNIFORM_Z (ptrue_max_1_f32_x_tied1, svfloat32_t,
391 z0 = svmax_n_f32_x (svptrue_b32 (), z0, 1),
392 z0 = svmax_x (svptrue_b32 (), z0, 1))
393
394 /*
395 ** ptrue_max_1_f32_x_untied:
396 ** ...
397 ** ptrue p[0-9]+\.b[^\n]*
398 ** ...
399 ** ret
400 */
401 TEST_UNIFORM_Z (ptrue_max_1_f32_x_untied, svfloat32_t,
402 z0 = svmax_n_f32_x (svptrue_b32 (), z1, 1),
403 z0 = svmax_x (svptrue_b32 (), z1, 1))
404
405 /*
406 ** ptrue_max_2_f32_x_tied1:
407 ** ...
408 ** ptrue p[0-9]+\.b[^\n]*
409 ** ...
410 ** ret
411 */
412 TEST_UNIFORM_Z (ptrue_max_2_f32_x_tied1, svfloat32_t,
413 z0 = svmax_n_f32_x (svptrue_b32 (), z0, 2),
414 z0 = svmax_x (svptrue_b32 (), z0, 2))
415
416 /*
417 ** ptrue_max_2_f32_x_untied:
418 ** ...
419 ** ptrue p[0-9]+\.b[^\n]*
420 ** ...
421 ** ret
422 */
423 TEST_UNIFORM_Z (ptrue_max_2_f32_x_untied, svfloat32_t,
424 z0 = svmax_n_f32_x (svptrue_b32 (), z1, 2),
425 z0 = svmax_x (svptrue_b32 (), z1, 2))