145
|
1 /* ACLE support for AArch64 SVE (function_base classes)
|
|
2 Copyright (C) 2018-2020 Free Software Foundation, Inc.
|
|
3
|
|
4 This file is part of GCC.
|
|
5
|
|
6 GCC is free software; you can redistribute it and/or modify it
|
|
7 under the terms of the GNU General Public License as published by
|
|
8 the Free Software Foundation; either version 3, or (at your option)
|
|
9 any later version.
|
|
10
|
|
11 GCC is distributed in the hope that it will be useful, but
|
|
12 WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
|
14 General Public License for more details.
|
|
15
|
|
16 You should have received a copy of the GNU General Public License
|
|
17 along with GCC; see the file COPYING3. If not see
|
|
18 <http://www.gnu.org/licenses/>. */
|
|
19
|
|
20 #ifndef GCC_AARCH64_SVE_BUILTINS_FUNCTIONS_H
|
|
21 #define GCC_AARCH64_SVE_BUILTINS_FUNCTIONS_H
|
|
22
|
|
23 namespace aarch64_sve {
|
|
24
|
|
25 /* Wrap T, which is derived from function_base, and indicate that the
|
|
26 function never has side effects. It is only necessary to use this
|
|
27 wrapper on functions that might have floating-point suffixes, since
|
|
28 otherwise we assume by default that the function has no side effects. */
|
|
29 template<typename T>
|
|
30 class quiet : public T
|
|
31 {
|
|
32 public:
|
|
33 CONSTEXPR quiet () : T () {}
|
|
34
|
|
35 /* Unfortunately we can't use parameter packs yet. */
|
|
36 template<typename T1>
|
|
37 CONSTEXPR quiet (const T1 &t1) : T (t1) {}
|
|
38
|
|
39 template<typename T1, typename T2>
|
|
40 CONSTEXPR quiet (const T1 &t1, const T2 &t2) : T (t1, t2) {}
|
|
41
|
|
42 template<typename T1, typename T2, typename T3>
|
|
43 CONSTEXPR quiet (const T1 &t1, const T2 &t2, const T3 &t3)
|
|
44 : T (t1, t2, t3) {}
|
|
45
|
|
46 unsigned int
|
|
47 call_properties (const function_instance &) const OVERRIDE
|
|
48 {
|
|
49 return 0;
|
|
50 }
|
|
51 };
|
|
52
|
|
53 /* A function_base that sometimes or always operates on tuples of
|
|
54 vectors. */
|
|
55 class multi_vector_function : public function_base
|
|
56 {
|
|
57 public:
|
|
58 CONSTEXPR multi_vector_function (unsigned int vectors_per_tuple)
|
|
59 : m_vectors_per_tuple (vectors_per_tuple) {}
|
|
60
|
|
61 unsigned int
|
|
62 vectors_per_tuple () const OVERRIDE
|
|
63 {
|
|
64 return m_vectors_per_tuple;
|
|
65 }
|
|
66
|
|
67 /* The number of vectors in a tuple, or 1 if the function only operates
|
|
68 on single vectors. */
|
|
69 unsigned int m_vectors_per_tuple;
|
|
70 };
|
|
71
|
|
72 /* A function_base that loads or stores contiguous memory elements
|
|
73 without extending or truncating them. */
|
|
74 class full_width_access : public multi_vector_function
|
|
75 {
|
|
76 public:
|
|
77 CONSTEXPR full_width_access (unsigned int vectors_per_tuple = 1)
|
|
78 : multi_vector_function (vectors_per_tuple) {}
|
|
79
|
|
80 tree
|
|
81 memory_scalar_type (const function_instance &fi) const OVERRIDE
|
|
82 {
|
|
83 return fi.scalar_type (0);
|
|
84 }
|
|
85
|
|
86 machine_mode
|
|
87 memory_vector_mode (const function_instance &fi) const OVERRIDE
|
|
88 {
|
|
89 machine_mode mode = fi.vector_mode (0);
|
|
90 if (m_vectors_per_tuple != 1)
|
|
91 mode = targetm.array_mode (mode, m_vectors_per_tuple).require ();
|
|
92 return mode;
|
|
93 }
|
|
94 };
|
|
95
|
|
96 /* A function_base that loads elements from memory and extends them
|
|
97 to a wider element. The memory element type is a fixed part of
|
|
98 the function base name. */
|
|
99 class extending_load : public function_base
|
|
100 {
|
|
101 public:
|
|
102 CONSTEXPR extending_load (type_suffix_index memory_type)
|
|
103 : m_memory_type (memory_type) {}
|
|
104
|
|
105 unsigned int
|
|
106 call_properties (const function_instance &) const OVERRIDE
|
|
107 {
|
|
108 return CP_READ_MEMORY;
|
|
109 }
|
|
110
|
|
111 tree
|
|
112 memory_scalar_type (const function_instance &) const OVERRIDE
|
|
113 {
|
|
114 return scalar_types[type_suffixes[m_memory_type].vector_type];
|
|
115 }
|
|
116
|
|
117 machine_mode
|
|
118 memory_vector_mode (const function_instance &fi) const OVERRIDE
|
|
119 {
|
|
120 machine_mode mem_mode = type_suffixes[m_memory_type].vector_mode;
|
|
121 machine_mode reg_mode = fi.vector_mode (0);
|
|
122 return aarch64_sve_data_mode (GET_MODE_INNER (mem_mode),
|
|
123 GET_MODE_NUNITS (reg_mode)).require ();
|
|
124 }
|
|
125
|
|
126 /* Return the rtx code associated with the kind of extension that
|
|
127 the load performs. */
|
|
128 rtx_code
|
|
129 extend_rtx_code () const
|
|
130 {
|
|
131 return (type_suffixes[m_memory_type].unsigned_p
|
|
132 ? ZERO_EXTEND : SIGN_EXTEND);
|
|
133 }
|
|
134
|
|
135 /* The type of the memory elements. This is part of the function base
|
|
136 name rather than a true type suffix. */
|
|
137 type_suffix_index m_memory_type;
|
|
138 };
|
|
139
|
|
140 /* A function_base that truncates vector elements and stores them to memory.
|
|
141 The memory element width is a fixed part of the function base name. */
|
|
142 class truncating_store : public function_base
|
|
143 {
|
|
144 public:
|
|
145 CONSTEXPR truncating_store (scalar_int_mode to_mode) : m_to_mode (to_mode) {}
|
|
146
|
|
147 unsigned int
|
|
148 call_properties (const function_instance &) const OVERRIDE
|
|
149 {
|
|
150 return CP_WRITE_MEMORY;
|
|
151 }
|
|
152
|
|
153 tree
|
|
154 memory_scalar_type (const function_instance &fi) const OVERRIDE
|
|
155 {
|
|
156 /* In truncating stores, the signedness of the memory element is defined
|
|
157 to be the same as the signedness of the vector element. The signedness
|
|
158 doesn't make any difference to the behavior of the function. */
|
|
159 type_class_index tclass = fi.type_suffix (0).tclass;
|
|
160 unsigned int element_bits = GET_MODE_BITSIZE (m_to_mode);
|
|
161 type_suffix_index suffix = find_type_suffix (tclass, element_bits);
|
|
162 return scalar_types[type_suffixes[suffix].vector_type];
|
|
163 }
|
|
164
|
|
165 machine_mode
|
|
166 memory_vector_mode (const function_instance &fi) const OVERRIDE
|
|
167 {
|
|
168 poly_uint64 nunits = GET_MODE_NUNITS (fi.vector_mode (0));
|
|
169 return aarch64_sve_data_mode (m_to_mode, nunits).require ();
|
|
170 }
|
|
171
|
|
172 /* The mode of a single memory element. */
|
|
173 scalar_int_mode m_to_mode;
|
|
174 };
|
|
175
|
|
176 /* An incomplete function_base for functions that have an associated rtx code.
|
|
177 It simply records information about the mapping for derived classes
|
|
178 to use. */
|
|
179 class rtx_code_function_base : public function_base
|
|
180 {
|
|
181 public:
|
|
182 CONSTEXPR rtx_code_function_base (rtx_code code_for_sint,
|
|
183 rtx_code code_for_uint,
|
|
184 int unspec_for_fp = -1)
|
|
185 : m_code_for_sint (code_for_sint), m_code_for_uint (code_for_uint),
|
|
186 m_unspec_for_fp (unspec_for_fp) {}
|
|
187
|
|
188 /* The rtx code to use for signed and unsigned integers respectively.
|
|
189 Can be UNKNOWN for functions that don't have integer forms. */
|
|
190 rtx_code m_code_for_sint;
|
|
191 rtx_code m_code_for_uint;
|
|
192
|
|
193 /* The UNSPEC_COND_* to use for floating-point operations. Can be -1
|
|
194 for functions that only operate on integers. */
|
|
195 int m_unspec_for_fp;
|
|
196 };
|
|
197
|
|
198 /* A function_base for functions that have an associated rtx code.
|
|
199 It supports all forms of predication except PRED_implicit. */
|
|
200 class rtx_code_function : public rtx_code_function_base
|
|
201 {
|
|
202 public:
|
|
203 CONSTEXPR rtx_code_function (rtx_code code_for_sint, rtx_code code_for_uint,
|
|
204 int unspec_for_fp = -1)
|
|
205 : rtx_code_function_base (code_for_sint, code_for_uint, unspec_for_fp) {}
|
|
206
|
|
207 rtx
|
|
208 expand (function_expander &e) const OVERRIDE
|
|
209 {
|
|
210 return e.map_to_rtx_codes (m_code_for_sint, m_code_for_uint,
|
|
211 m_unspec_for_fp);
|
|
212 }
|
|
213 };
|
|
214
|
|
215 /* Like rtx_code_function, but for functions that take what is normally
|
|
216 the final argument first. One use of this class is to handle binary
|
|
217 reversed operations; another is to handle MLA-style operations that
|
|
218 are normally expressed in GCC as MAD-style operations. */
|
|
219 class rtx_code_function_rotated : public rtx_code_function_base
|
|
220 {
|
|
221 public:
|
|
222 CONSTEXPR rtx_code_function_rotated (rtx_code code_for_sint,
|
|
223 rtx_code code_for_uint,
|
|
224 int unspec_for_fp = -1)
|
|
225 : rtx_code_function_base (code_for_sint, code_for_uint, unspec_for_fp) {}
|
|
226
|
|
227 rtx
|
|
228 expand (function_expander &e) const OVERRIDE
|
|
229 {
|
|
230 /* Rotate the inputs into their normal order, but continue to make _m
|
|
231 functions merge with what was originally the first vector argument. */
|
|
232 unsigned int nargs = e.args.length ();
|
|
233 e.rotate_inputs_left (e.pred != PRED_none ? 1 : 0, nargs);
|
|
234 return e.map_to_rtx_codes (m_code_for_sint, m_code_for_uint,
|
|
235 m_unspec_for_fp, nargs - 1);
|
|
236 }
|
|
237 };
|
|
238
|
|
239 /* An incomplete function_base for functions that have an associated
|
|
240 unspec code, with separate codes for signed integers, unsigned
|
|
241 integers and floating-point values. The class simply records
|
|
242 information about the mapping for derived classes to use. */
|
|
243 class unspec_based_function_base : public function_base
|
|
244 {
|
|
245 public:
|
|
246 CONSTEXPR unspec_based_function_base (int unspec_for_sint,
|
|
247 int unspec_for_uint,
|
|
248 int unspec_for_fp)
|
|
249 : m_unspec_for_sint (unspec_for_sint),
|
|
250 m_unspec_for_uint (unspec_for_uint),
|
|
251 m_unspec_for_fp (unspec_for_fp)
|
|
252 {}
|
|
253
|
|
254 /* Return the unspec code to use for INSTANCE, based on type suffix 0. */
|
|
255 int
|
|
256 unspec_for (const function_instance &instance) const
|
|
257 {
|
|
258 return (!instance.type_suffix (0).integer_p ? m_unspec_for_fp
|
|
259 : instance.type_suffix (0).unsigned_p ? m_unspec_for_uint
|
|
260 : m_unspec_for_sint);
|
|
261 }
|
|
262
|
|
263 /* The unspec code associated with signed-integer, unsigned-integer
|
|
264 and floating-point operations respectively. */
|
|
265 int m_unspec_for_sint;
|
|
266 int m_unspec_for_uint;
|
|
267 int m_unspec_for_fp;
|
|
268 };
|
|
269
|
|
270 /* A function_base for functions that have an associated unspec code.
|
|
271 It supports all forms of predication except PRED_implicit. */
|
|
272 class unspec_based_function : public unspec_based_function_base
|
|
273 {
|
|
274 public:
|
|
275 CONSTEXPR unspec_based_function (int unspec_for_sint, int unspec_for_uint,
|
|
276 int unspec_for_fp)
|
|
277 : unspec_based_function_base (unspec_for_sint, unspec_for_uint,
|
|
278 unspec_for_fp)
|
|
279 {}
|
|
280
|
|
281 rtx
|
|
282 expand (function_expander &e) const OVERRIDE
|
|
283 {
|
|
284 return e.map_to_unspecs (m_unspec_for_sint, m_unspec_for_uint,
|
|
285 m_unspec_for_fp);
|
|
286 }
|
|
287 };
|
|
288
|
|
289 /* Like unspec_based_function, but for functions that take what is normally
|
|
290 the final argument first. One use of this class is to handle binary
|
|
291 reversed operations; another is to handle MLA-style operations that
|
|
292 are normally expressed in GCC as MAD-style operations. */
|
|
293 class unspec_based_function_rotated : public unspec_based_function_base
|
|
294 {
|
|
295 public:
|
|
296 CONSTEXPR unspec_based_function_rotated (int unspec_for_sint,
|
|
297 int unspec_for_uint,
|
|
298 int unspec_for_fp)
|
|
299 : unspec_based_function_base (unspec_for_sint, unspec_for_uint,
|
|
300 unspec_for_fp)
|
|
301 {}
|
|
302
|
|
303 rtx
|
|
304 expand (function_expander &e) const OVERRIDE
|
|
305 {
|
|
306 /* Rotate the inputs into their normal order, but continue to make _m
|
|
307 functions merge with what was originally the first vector argument. */
|
|
308 unsigned int nargs = e.args.length ();
|
|
309 e.rotate_inputs_left (e.pred != PRED_none ? 1 : 0, nargs);
|
|
310 return e.map_to_unspecs (m_unspec_for_sint, m_unspec_for_uint,
|
|
311 m_unspec_for_fp, nargs - 1);
|
|
312 }
|
|
313 };
|
|
314
|
|
315 /* Like unspec_based_function, but map the function directly to
|
|
316 CODE (UNSPEC, M) instead of using the generic predication-based
|
|
317 expansion. where M is the vector mode associated with type suffix 0.
|
|
318 This is useful if the unspec doesn't describe the full operation or
|
|
319 if the usual predication rules don't apply for some reason. */
|
|
320 template<insn_code (*CODE) (int, machine_mode)>
|
|
321 class unspec_based_function_exact_insn : public unspec_based_function_base
|
|
322 {
|
|
323 public:
|
|
324 CONSTEXPR unspec_based_function_exact_insn (int unspec_for_sint,
|
|
325 int unspec_for_uint,
|
|
326 int unspec_for_fp)
|
|
327 : unspec_based_function_base (unspec_for_sint, unspec_for_uint,
|
|
328 unspec_for_fp)
|
|
329 {}
|
|
330
|
|
331 rtx
|
|
332 expand (function_expander &e) const OVERRIDE
|
|
333 {
|
|
334 return e.use_exact_insn (CODE (unspec_for (e), e.vector_mode (0)));
|
|
335 }
|
|
336 };
|
|
337
|
|
338 /* A function that performs an unspec and then adds it to another value. */
|
|
339 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_add>
|
|
340 unspec_based_add_function;
|
|
341 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_add_lane>
|
|
342 unspec_based_add_lane_function;
|
|
343
|
|
344 /* Generic unspec-based _lane function. */
|
|
345 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_lane>
|
|
346 unspec_based_lane_function;
|
|
347
|
|
348 /* A functon that uses aarch64_pred* patterns regardless of the
|
|
349 predication type. */
|
|
350 typedef unspec_based_function_exact_insn<code_for_aarch64_pred>
|
|
351 unspec_based_pred_function;
|
|
352
|
|
353 /* Like unspec_based_add_function and unspec_based_add_lane_function,
|
|
354 but using saturating addition. */
|
|
355 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_qadd>
|
|
356 unspec_based_qadd_function;
|
|
357 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_qadd_lane>
|
|
358 unspec_based_qadd_lane_function;
|
|
359
|
|
360 /* Like unspec_based_sub_function and unspec_based_sub_lane_function,
|
|
361 but using saturating subtraction. */
|
|
362 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_qsub>
|
|
363 unspec_based_qsub_function;
|
|
364 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_qsub_lane>
|
|
365 unspec_based_qsub_lane_function;
|
|
366
|
|
367 /* A function that performs an unspec and then subtracts it from
|
|
368 another value. */
|
|
369 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_sub>
|
|
370 unspec_based_sub_function;
|
|
371 typedef unspec_based_function_exact_insn<code_for_aarch64_sve_sub_lane>
|
|
372 unspec_based_sub_lane_function;
|
|
373
|
|
374 /* A function that acts like unspec_based_function_exact_insn<INT_CODE>
|
|
375 when operating on integers, but that expands to an (fma ...)-style
|
|
376 aarch64_sve* operation when applied to floats. */
|
|
377 template<insn_code (*INT_CODE) (int, machine_mode)>
|
|
378 class unspec_based_fused_function : public unspec_based_function_base
|
|
379 {
|
|
380 public:
|
|
381 CONSTEXPR unspec_based_fused_function (int unspec_for_sint,
|
|
382 int unspec_for_uint,
|
|
383 int unspec_for_fp)
|
|
384 : unspec_based_function_base (unspec_for_sint, unspec_for_uint,
|
|
385 unspec_for_fp)
|
|
386 {}
|
|
387
|
|
388 rtx
|
|
389 expand (function_expander &e) const OVERRIDE
|
|
390 {
|
|
391 int unspec = unspec_for (e);
|
|
392 insn_code icode;
|
|
393 if (e.type_suffix (0).float_p)
|
|
394 {
|
|
395 /* Put the operands in the normal (fma ...) order, with the accumulator
|
|
396 last. This fits naturally since that's also the unprinted operand
|
|
397 in the asm output. */
|
|
398 e.rotate_inputs_left (0, e.pred != PRED_none ? 4 : 3);
|
|
399 icode = code_for_aarch64_sve (unspec, e.vector_mode (0));
|
|
400 }
|
|
401 else
|
|
402 icode = INT_CODE (unspec, e.vector_mode (0));
|
|
403 return e.use_exact_insn (icode);
|
|
404 }
|
|
405 };
|
|
406 typedef unspec_based_fused_function<code_for_aarch64_sve_add>
|
|
407 unspec_based_mla_function;
|
|
408 typedef unspec_based_fused_function<code_for_aarch64_sve_sub>
|
|
409 unspec_based_mls_function;
|
|
410
|
|
411 /* Like unspec_based_fused_function, but for _lane functions. */
|
|
412 template<insn_code (*INT_CODE) (int, machine_mode)>
|
|
413 class unspec_based_fused_lane_function : public unspec_based_function_base
|
|
414 {
|
|
415 public:
|
|
416 CONSTEXPR unspec_based_fused_lane_function (int unspec_for_sint,
|
|
417 int unspec_for_uint,
|
|
418 int unspec_for_fp)
|
|
419 : unspec_based_function_base (unspec_for_sint, unspec_for_uint,
|
|
420 unspec_for_fp)
|
|
421 {}
|
|
422
|
|
423 rtx
|
|
424 expand (function_expander &e) const OVERRIDE
|
|
425 {
|
|
426 int unspec = unspec_for (e);
|
|
427 insn_code icode;
|
|
428 if (e.type_suffix (0).float_p)
|
|
429 {
|
|
430 /* Put the operands in the normal (fma ...) order, with the accumulator
|
|
431 last. This fits naturally since that's also the unprinted operand
|
|
432 in the asm output. */
|
|
433 e.rotate_inputs_left (0, e.pred != PRED_none ? 5 : 4);
|
|
434 icode = code_for_aarch64_lane (unspec, e.vector_mode (0));
|
|
435 }
|
|
436 else
|
|
437 icode = INT_CODE (unspec, e.vector_mode (0));
|
|
438 return e.use_exact_insn (icode);
|
|
439 }
|
|
440 };
|
|
441 typedef unspec_based_fused_lane_function<code_for_aarch64_sve_add_lane>
|
|
442 unspec_based_mla_lane_function;
|
|
443 typedef unspec_based_fused_lane_function<code_for_aarch64_sve_sub_lane>
|
|
444 unspec_based_mls_lane_function;
|
|
445
|
|
446 /* A function_base that uses CODE_FOR_MODE (M) to get the associated
|
|
447 instruction code, where M is the vector mode associated with type
|
|
448 suffix N. */
|
|
449 template<insn_code (*CODE_FOR_MODE) (machine_mode), unsigned int N>
|
|
450 class code_for_mode_function : public function_base
|
|
451 {
|
|
452 public:
|
|
453 rtx
|
|
454 expand (function_expander &e) const OVERRIDE
|
|
455 {
|
|
456 return e.use_exact_insn (CODE_FOR_MODE (e.vector_mode (N)));
|
|
457 }
|
|
458 };
|
|
459
|
|
460 /* A function that uses code_for_<PATTERN> (M), where M is the vector
|
|
461 mode associated with the first type suffix. */
|
|
462 #define CODE_FOR_MODE0(PATTERN) code_for_mode_function<code_for_##PATTERN, 0>
|
|
463
|
|
464 /* Likewise for the second type suffix. */
|
|
465 #define CODE_FOR_MODE1(PATTERN) code_for_mode_function<code_for_##PATTERN, 1>
|
|
466
|
|
467 /* Like CODE_FOR_MODE0, but the function doesn't raise exceptions when
|
|
468 operating on floating-point data. */
|
|
469 #define QUIET_CODE_FOR_MODE0(PATTERN) \
|
|
470 quiet< code_for_mode_function<code_for_##PATTERN, 0> >
|
|
471
|
|
472 /* A function_base for functions that always expand to a fixed insn pattern,
|
|
473 regardless of what the suffixes are. */
|
|
474 class fixed_insn_function : public function_base
|
|
475 {
|
|
476 public:
|
|
477 CONSTEXPR fixed_insn_function (insn_code code) : m_code (code) {}
|
|
478
|
|
479 rtx
|
|
480 expand (function_expander &e) const OVERRIDE
|
|
481 {
|
|
482 return e.use_exact_insn (m_code);
|
|
483 }
|
|
484
|
|
485 /* The instruction to use. */
|
|
486 insn_code m_code;
|
|
487 };
|
|
488
|
|
489 /* A function_base for functions that permute their arguments. */
|
|
490 class permute : public quiet<function_base>
|
|
491 {
|
|
492 public:
|
|
493 /* Fold a unary or binary permute with the permute vector given by
|
|
494 BUILDER. */
|
|
495 gimple *
|
|
496 fold_permute (const gimple_folder &f, const vec_perm_builder &builder) const
|
|
497 {
|
|
498 /* Punt for now on _b16 and wider; we'd need more complex evpc logic
|
|
499 to rerecognize the result. */
|
|
500 if (f.type_suffix (0).bool_p && f.type_suffix (0).element_bits > 8)
|
|
501 return NULL;
|
|
502
|
|
503 unsigned int nargs = gimple_call_num_args (f.call);
|
|
504 poly_uint64 nelts = TYPE_VECTOR_SUBPARTS (TREE_TYPE (f.lhs));
|
|
505 vec_perm_indices indices (builder, nargs, nelts);
|
|
506 tree perm_type = build_vector_type (ssizetype, nelts);
|
|
507 return gimple_build_assign (f.lhs, VEC_PERM_EXPR,
|
|
508 gimple_call_arg (f.call, 0),
|
|
509 gimple_call_arg (f.call, nargs - 1),
|
|
510 vec_perm_indices_to_tree (perm_type, indices));
|
|
511 }
|
|
512 };
|
|
513
|
|
514 /* A function_base for functions that permute two vectors using a fixed
|
|
515 choice of indices. */
|
|
516 class binary_permute : public permute
|
|
517 {
|
|
518 public:
|
|
519 CONSTEXPR binary_permute (int unspec) : m_unspec (unspec) {}
|
|
520
|
|
521 rtx
|
|
522 expand (function_expander &e) const OVERRIDE
|
|
523 {
|
|
524 insn_code icode = code_for_aarch64_sve (m_unspec, e.vector_mode (0));
|
|
525 return e.use_exact_insn (icode);
|
|
526 }
|
|
527
|
|
528 /* The unspec code associated with the operation. */
|
|
529 int m_unspec;
|
|
530 };
|
|
531
|
|
532 /* A function_base for functions that reduce a vector to a scalar. */
|
|
533 class reduction : public function_base
|
|
534 {
|
|
535 public:
|
|
536 CONSTEXPR reduction (int unspec)
|
|
537 : m_unspec_for_sint (unspec),
|
|
538 m_unspec_for_uint (unspec),
|
|
539 m_unspec_for_fp (unspec)
|
|
540 {}
|
|
541
|
|
542 CONSTEXPR reduction (int unspec_for_sint, int unspec_for_uint,
|
|
543 int unspec_for_fp)
|
|
544 : m_unspec_for_sint (unspec_for_sint),
|
|
545 m_unspec_for_uint (unspec_for_uint),
|
|
546 m_unspec_for_fp (unspec_for_fp)
|
|
547 {}
|
|
548
|
|
549 rtx
|
|
550 expand (function_expander &e) const OVERRIDE
|
|
551 {
|
|
552 machine_mode mode = e.vector_mode (0);
|
|
553 int unspec = (!e.type_suffix (0).integer_p ? m_unspec_for_fp
|
|
554 : e.type_suffix (0).unsigned_p ? m_unspec_for_uint
|
|
555 : m_unspec_for_sint);
|
|
556 /* There's no distinction between SADDV and UADDV for 64-bit elements;
|
|
557 the signed versions only exist for narrower elements. */
|
|
558 if (GET_MODE_UNIT_BITSIZE (mode) == 64 && unspec == UNSPEC_SADDV)
|
|
559 unspec = UNSPEC_UADDV;
|
|
560 return e.use_exact_insn (code_for_aarch64_pred_reduc (unspec, mode));
|
|
561 }
|
|
562
|
|
563 /* The unspec code associated with signed-integer, unsigned-integer
|
|
564 and floating-point operations respectively. */
|
|
565 int m_unspec_for_sint;
|
|
566 int m_unspec_for_uint;
|
|
567 int m_unspec_for_fp;
|
|
568 };
|
|
569
|
|
570 /* A function_base for functions that shift narrower-than-64-bit values
|
|
571 by 64-bit amounts. */
|
|
572 class shift_wide : public function_base
|
|
573 {
|
|
574 public:
|
|
575 CONSTEXPR shift_wide (rtx_code code, int wide_unspec)
|
|
576 : m_code (code), m_wide_unspec (wide_unspec) {}
|
|
577
|
|
578 rtx
|
|
579 expand (function_expander &e) const OVERRIDE
|
|
580 {
|
|
581 machine_mode mode = e.vector_mode (0);
|
|
582 machine_mode elem_mode = GET_MODE_INNER (mode);
|
|
583
|
|
584 /* If the argument is a constant that the normal shifts can handle
|
|
585 directly, use them instead. */
|
|
586 rtx shift = unwrap_const_vec_duplicate (e.args.last ());
|
|
587 if (aarch64_simd_shift_imm_p (shift, elem_mode, m_code == ASHIFT))
|
|
588 {
|
|
589 e.args.last () = shift;
|
|
590 return e.map_to_rtx_codes (m_code, m_code, -1);
|
|
591 }
|
|
592
|
|
593 if (e.pred == PRED_x)
|
|
594 return e.use_unpred_insn (code_for_aarch64_sve (m_wide_unspec, mode));
|
|
595
|
|
596 return e.use_cond_insn (code_for_cond (m_wide_unspec, mode));
|
|
597 }
|
|
598
|
|
599 /* The rtx code associated with a "normal" shift. */
|
|
600 rtx_code m_code;
|
|
601
|
|
602 /* The unspec code associated with the wide shift. */
|
|
603 int m_wide_unspec;
|
|
604 };
|
|
605
|
|
606 /* A function_base for unary functions that count bits. */
|
|
607 class unary_count : public quiet<function_base>
|
|
608 {
|
|
609 public:
|
|
610 CONSTEXPR unary_count (rtx_code code) : m_code (code) {}
|
|
611
|
|
612 rtx
|
|
613 expand (function_expander &e) const OVERRIDE
|
|
614 {
|
|
615 /* The md patterns treat the operand as an integer. */
|
|
616 machine_mode mode = aarch64_sve_int_mode (e.vector_mode (0));
|
|
617 e.args.last () = gen_lowpart (mode, e.args.last ());
|
|
618
|
|
619 if (e.pred == PRED_x)
|
|
620 return e.use_pred_x_insn (code_for_aarch64_pred (m_code, mode));
|
|
621
|
|
622 return e.use_cond_insn (code_for_cond (m_code, mode));
|
|
623 }
|
|
624
|
|
625 /* The rtx code associated with the operation. */
|
|
626 rtx_code m_code;
|
|
627 };
|
|
628
|
|
629 /* A function_base for svwhile* functions. */
|
|
630 class while_comparison : public function_base
|
|
631 {
|
|
632 public:
|
|
633 CONSTEXPR while_comparison (int unspec_for_sint, int unspec_for_uint)
|
|
634 : m_unspec_for_sint (unspec_for_sint),
|
|
635 m_unspec_for_uint (unspec_for_uint)
|
|
636 {}
|
|
637
|
|
638 rtx
|
|
639 expand (function_expander &e) const OVERRIDE
|
|
640 {
|
|
641 /* Suffix 0 determines the predicate mode, suffix 1 determines the
|
|
642 scalar mode and signedness. */
|
|
643 int unspec = (e.type_suffix (1).unsigned_p
|
|
644 ? m_unspec_for_uint
|
|
645 : m_unspec_for_sint);
|
|
646 machine_mode pred_mode = e.vector_mode (0);
|
|
647 scalar_mode reg_mode = GET_MODE_INNER (e.vector_mode (1));
|
|
648 return e.use_exact_insn (code_for_while (unspec, reg_mode, pred_mode));
|
|
649 }
|
|
650
|
|
651 /* The unspec codes associated with signed and unsigned operations
|
|
652 respectively. */
|
|
653 int m_unspec_for_sint;
|
|
654 int m_unspec_for_uint;
|
|
655 };
|
|
656
|
|
657 }
|
|
658
|
|
659 /* Declare the global function base NAME, creating it from an instance
|
|
660 of class CLASS with constructor arguments ARGS. */
|
|
661 #define FUNCTION(NAME, CLASS, ARGS) \
|
|
662 namespace { static CONSTEXPR const CLASS NAME##_obj ARGS; } \
|
|
663 namespace functions { const function_base *const NAME = &NAME##_obj; }
|
|
664
|
|
665 #endif
|