]>
Commit | Line | Data |
---|---|---|
43e9d192 | 1 | /* Builtins' description for AArch64 SIMD architecture. |
d1e082c2 | 2 | Copyright (C) 2011-2013 Free Software Foundation, Inc. |
43e9d192 IB |
3 | Contributed by ARM Ltd. |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU General Public License as published by | |
9 | the Free Software Foundation; either version 3, or (at your option) | |
10 | any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but | |
13 | WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | |
15 | General Public License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "tm.h" | |
25 | #include "rtl.h" | |
26 | #include "tree.h" | |
d8a2d370 DN |
27 | #include "stor-layout.h" |
28 | #include "stringpool.h" | |
29 | #include "calls.h" | |
43e9d192 IB |
30 | #include "expr.h" |
31 | #include "tm_p.h" | |
32 | #include "recog.h" | |
33 | #include "langhooks.h" | |
34 | #include "diagnostic-core.h" | |
35 | #include "optabs.h" | |
2fb9a547 AM |
36 | #include "pointer-set.h" |
37 | #include "hash-table.h" | |
38 | #include "vec.h" | |
39 | #include "ggc.h" | |
40 | #include "basic-block.h" | |
41 | #include "tree-ssa-alias.h" | |
42 | #include "internal-fn.h" | |
43 | #include "gimple-fold.h" | |
44 | #include "tree-eh.h" | |
45 | #include "gimple-expr.h" | |
46 | #include "is-a.h" | |
0ac198d3 | 47 | #include "gimple.h" |
5be5c238 | 48 | #include "gimple-iterator.h" |
43e9d192 | 49 | |
342be7f7 | 50 | enum aarch64_simd_builtin_type_mode |
43e9d192 | 51 | { |
342be7f7 JG |
52 | T_V8QI, |
53 | T_V4HI, | |
54 | T_V2SI, | |
55 | T_V2SF, | |
56 | T_DI, | |
57 | T_DF, | |
58 | T_V16QI, | |
59 | T_V8HI, | |
60 | T_V4SI, | |
61 | T_V4SF, | |
62 | T_V2DI, | |
63 | T_V2DF, | |
64 | T_TI, | |
65 | T_EI, | |
66 | T_OI, | |
67 | T_XI, | |
68 | T_SI, | |
0050faf8 | 69 | T_SF, |
342be7f7 JG |
70 | T_HI, |
71 | T_QI, | |
72 | T_MAX | |
43e9d192 IB |
73 | }; |
74 | ||
75 | #define v8qi_UP T_V8QI | |
76 | #define v4hi_UP T_V4HI | |
77 | #define v2si_UP T_V2SI | |
78 | #define v2sf_UP T_V2SF | |
79 | #define di_UP T_DI | |
80 | #define df_UP T_DF | |
81 | #define v16qi_UP T_V16QI | |
82 | #define v8hi_UP T_V8HI | |
83 | #define v4si_UP T_V4SI | |
84 | #define v4sf_UP T_V4SF | |
85 | #define v2di_UP T_V2DI | |
86 | #define v2df_UP T_V2DF | |
87 | #define ti_UP T_TI | |
88 | #define ei_UP T_EI | |
89 | #define oi_UP T_OI | |
90 | #define xi_UP T_XI | |
91 | #define si_UP T_SI | |
0050faf8 | 92 | #define sf_UP T_SF |
43e9d192 IB |
93 | #define hi_UP T_HI |
94 | #define qi_UP T_QI | |
95 | ||
96 | #define UP(X) X##_UP | |
97 | ||
b5828b4b JG |
98 | #define SIMD_MAX_BUILTIN_ARGS 5 |
99 | ||
100 | enum aarch64_type_qualifiers | |
43e9d192 | 101 | { |
b5828b4b JG |
102 | /* T foo. */ |
103 | qualifier_none = 0x0, | |
104 | /* unsigned T foo. */ | |
105 | qualifier_unsigned = 0x1, /* 1 << 0 */ | |
106 | /* const T foo. */ | |
107 | qualifier_const = 0x2, /* 1 << 1 */ | |
108 | /* T *foo. */ | |
109 | qualifier_pointer = 0x4, /* 1 << 2 */ | |
110 | /* const T *foo. */ | |
111 | qualifier_const_pointer = 0x6, /* qualifier_const | qualifier_pointer */ | |
112 | /* Used when expanding arguments if an operand could | |
113 | be an immediate. */ | |
114 | qualifier_immediate = 0x8, /* 1 << 3 */ | |
115 | qualifier_maybe_immediate = 0x10, /* 1 << 4 */ | |
116 | /* void foo (...). */ | |
117 | qualifier_void = 0x20, /* 1 << 5 */ | |
118 | /* Some patterns may have internal operands, this qualifier is an | |
119 | instruction to the initialisation code to skip this operand. */ | |
120 | qualifier_internal = 0x40, /* 1 << 6 */ | |
121 | /* Some builtins should use the T_*mode* encoded in a simd_builtin_datum | |
122 | rather than using the type of the operand. */ | |
123 | qualifier_map_mode = 0x80, /* 1 << 7 */ | |
124 | /* qualifier_pointer | qualifier_map_mode */ | |
125 | qualifier_pointer_map_mode = 0x84, | |
126 | /* qualifier_const_pointer | qualifier_map_mode */ | |
127 | qualifier_const_pointer_map_mode = 0x86 | |
128 | }; | |
43e9d192 IB |
129 | |
130 | typedef struct | |
131 | { | |
132 | const char *name; | |
342be7f7 JG |
133 | enum aarch64_simd_builtin_type_mode mode; |
134 | const enum insn_code code; | |
135 | unsigned int fcode; | |
b5828b4b | 136 | enum aarch64_type_qualifiers *qualifiers; |
43e9d192 IB |
137 | } aarch64_simd_builtin_datum; |
138 | ||
b5828b4b JG |
139 | static enum aarch64_type_qualifiers |
140 | aarch64_types_unop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
141 | = { qualifier_none, qualifier_none }; | |
142 | #define TYPES_UNOP (aarch64_types_unop_qualifiers) | |
143 | #define TYPES_CREATE (aarch64_types_unop_qualifiers) | |
144 | #define TYPES_REINTERP (aarch64_types_unop_qualifiers) | |
145 | static enum aarch64_type_qualifiers | |
146 | aarch64_types_binop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
147 | = { qualifier_none, qualifier_none, qualifier_maybe_immediate }; | |
148 | #define TYPES_BINOP (aarch64_types_binop_qualifiers) | |
149 | static enum aarch64_type_qualifiers | |
150 | aarch64_types_ternop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
151 | = { qualifier_none, qualifier_none, qualifier_none, qualifier_none }; | |
152 | #define TYPES_TERNOP (aarch64_types_ternop_qualifiers) | |
153 | static enum aarch64_type_qualifiers | |
154 | aarch64_types_quadop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
155 | = { qualifier_none, qualifier_none, qualifier_none, | |
156 | qualifier_none, qualifier_none }; | |
157 | #define TYPES_QUADOP (aarch64_types_quadop_qualifiers) | |
158 | ||
159 | static enum aarch64_type_qualifiers | |
160 | aarch64_types_getlane_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
161 | = { qualifier_none, qualifier_none, qualifier_immediate }; | |
162 | #define TYPES_GETLANE (aarch64_types_getlane_qualifiers) | |
163 | #define TYPES_SHIFTIMM (aarch64_types_getlane_qualifiers) | |
164 | static enum aarch64_type_qualifiers | |
165 | aarch64_types_setlane_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
166 | = { qualifier_none, qualifier_none, qualifier_none, qualifier_immediate }; | |
167 | #define TYPES_SETLANE (aarch64_types_setlane_qualifiers) | |
168 | #define TYPES_SHIFTINSERT (aarch64_types_setlane_qualifiers) | |
169 | #define TYPES_SHIFTACC (aarch64_types_setlane_qualifiers) | |
170 | ||
171 | static enum aarch64_type_qualifiers | |
172 | aarch64_types_combine_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
173 | = { qualifier_none, qualifier_none, qualifier_none }; | |
174 | #define TYPES_COMBINE (aarch64_types_combine_qualifiers) | |
175 | ||
176 | static enum aarch64_type_qualifiers | |
177 | aarch64_types_load1_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
178 | = { qualifier_none, qualifier_const_pointer_map_mode }; | |
179 | #define TYPES_LOAD1 (aarch64_types_load1_qualifiers) | |
180 | #define TYPES_LOADSTRUCT (aarch64_types_load1_qualifiers) | |
181 | ||
182 | /* The first argument (return type) of a store should be void type, | |
183 | which we represent with qualifier_void. Their first operand will be | |
184 | a DImode pointer to the location to store to, so we must use | |
185 | qualifier_map_mode | qualifier_pointer to build a pointer to the | |
186 | element type of the vector. */ | |
187 | static enum aarch64_type_qualifiers | |
188 | aarch64_types_store1_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
189 | = { qualifier_void, qualifier_pointer_map_mode, qualifier_none }; | |
190 | #define TYPES_STORE1 (aarch64_types_store1_qualifiers) | |
191 | #define TYPES_STORESTRUCT (aarch64_types_store1_qualifiers) | |
192 | ||
0ddec79f JG |
193 | #define CF0(N, X) CODE_FOR_aarch64_##N##X |
194 | #define CF1(N, X) CODE_FOR_##N##X##1 | |
195 | #define CF2(N, X) CODE_FOR_##N##X##2 | |
196 | #define CF3(N, X) CODE_FOR_##N##X##3 | |
197 | #define CF4(N, X) CODE_FOR_##N##X##4 | |
198 | #define CF10(N, X) CODE_FOR_##N##X | |
199 | ||
200 | #define VAR1(T, N, MAP, A) \ | |
b5828b4b | 201 | {#N, UP (A), CF##MAP (N, A), 0, TYPES_##T}, |
0ddec79f JG |
202 | #define VAR2(T, N, MAP, A, B) \ |
203 | VAR1 (T, N, MAP, A) \ | |
204 | VAR1 (T, N, MAP, B) | |
205 | #define VAR3(T, N, MAP, A, B, C) \ | |
206 | VAR2 (T, N, MAP, A, B) \ | |
207 | VAR1 (T, N, MAP, C) | |
208 | #define VAR4(T, N, MAP, A, B, C, D) \ | |
209 | VAR3 (T, N, MAP, A, B, C) \ | |
210 | VAR1 (T, N, MAP, D) | |
211 | #define VAR5(T, N, MAP, A, B, C, D, E) \ | |
212 | VAR4 (T, N, MAP, A, B, C, D) \ | |
213 | VAR1 (T, N, MAP, E) | |
214 | #define VAR6(T, N, MAP, A, B, C, D, E, F) \ | |
215 | VAR5 (T, N, MAP, A, B, C, D, E) \ | |
216 | VAR1 (T, N, MAP, F) | |
217 | #define VAR7(T, N, MAP, A, B, C, D, E, F, G) \ | |
218 | VAR6 (T, N, MAP, A, B, C, D, E, F) \ | |
219 | VAR1 (T, N, MAP, G) | |
220 | #define VAR8(T, N, MAP, A, B, C, D, E, F, G, H) \ | |
221 | VAR7 (T, N, MAP, A, B, C, D, E, F, G) \ | |
222 | VAR1 (T, N, MAP, H) | |
223 | #define VAR9(T, N, MAP, A, B, C, D, E, F, G, H, I) \ | |
224 | VAR8 (T, N, MAP, A, B, C, D, E, F, G, H) \ | |
225 | VAR1 (T, N, MAP, I) | |
226 | #define VAR10(T, N, MAP, A, B, C, D, E, F, G, H, I, J) \ | |
227 | VAR9 (T, N, MAP, A, B, C, D, E, F, G, H, I) \ | |
228 | VAR1 (T, N, MAP, J) | |
229 | #define VAR11(T, N, MAP, A, B, C, D, E, F, G, H, I, J, K) \ | |
230 | VAR10 (T, N, MAP, A, B, C, D, E, F, G, H, I, J) \ | |
231 | VAR1 (T, N, MAP, K) | |
232 | #define VAR12(T, N, MAP, A, B, C, D, E, F, G, H, I, J, K, L) \ | |
233 | VAR11 (T, N, MAP, A, B, C, D, E, F, G, H, I, J, K) \ | |
234 | VAR1 (T, N, MAP, L) | |
342be7f7 JG |
235 | |
236 | /* BUILTIN_<ITERATOR> macros should expand to cover the same range of | |
237 | modes as is given for each define_mode_iterator in | |
238 | config/aarch64/iterators.md. */ | |
239 | ||
0ddec79f JG |
240 | #define BUILTIN_DX(T, N, MAP) \ |
241 | VAR2 (T, N, MAP, di, df) | |
242 | #define BUILTIN_GPF(T, N, MAP) \ | |
243 | VAR2 (T, N, MAP, sf, df) | |
244 | #define BUILTIN_SDQ_I(T, N, MAP) \ | |
245 | VAR4 (T, N, MAP, qi, hi, si, di) | |
246 | #define BUILTIN_SD_HSI(T, N, MAP) \ | |
247 | VAR2 (T, N, MAP, hi, si) | |
248 | #define BUILTIN_V2F(T, N, MAP) \ | |
249 | VAR2 (T, N, MAP, v2sf, v2df) | |
250 | #define BUILTIN_VALL(T, N, MAP) \ | |
251 | VAR10 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
252 | v4si, v2di, v2sf, v4sf, v2df) | |
bb60efd9 JG |
253 | #define BUILTIN_VALLDI(T, N, MAP) \ |
254 | VAR11 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
255 | v4si, v2di, v2sf, v4sf, v2df, di) | |
0ddec79f JG |
256 | #define BUILTIN_VB(T, N, MAP) \ |
257 | VAR2 (T, N, MAP, v8qi, v16qi) | |
258 | #define BUILTIN_VD(T, N, MAP) \ | |
259 | VAR4 (T, N, MAP, v8qi, v4hi, v2si, v2sf) | |
260 | #define BUILTIN_VDC(T, N, MAP) \ | |
261 | VAR6 (T, N, MAP, v8qi, v4hi, v2si, v2sf, di, df) | |
262 | #define BUILTIN_VDIC(T, N, MAP) \ | |
263 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
264 | #define BUILTIN_VDN(T, N, MAP) \ | |
265 | VAR3 (T, N, MAP, v4hi, v2si, di) | |
266 | #define BUILTIN_VDQ(T, N, MAP) \ | |
267 | VAR7 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di) | |
268 | #define BUILTIN_VDQF(T, N, MAP) \ | |
269 | VAR3 (T, N, MAP, v2sf, v4sf, v2df) | |
270 | #define BUILTIN_VDQH(T, N, MAP) \ | |
271 | VAR2 (T, N, MAP, v4hi, v8hi) | |
272 | #define BUILTIN_VDQHS(T, N, MAP) \ | |
273 | VAR4 (T, N, MAP, v4hi, v8hi, v2si, v4si) | |
274 | #define BUILTIN_VDQIF(T, N, MAP) \ | |
275 | VAR9 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2sf, v4sf, v2df) | |
276 | #define BUILTIN_VDQM(T, N, MAP) \ | |
277 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
278 | #define BUILTIN_VDQV(T, N, MAP) \ | |
279 | VAR5 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v4si) | |
280 | #define BUILTIN_VDQ_BHSI(T, N, MAP) \ | |
281 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
282 | #define BUILTIN_VDQ_I(T, N, MAP) \ | |
283 | VAR7 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di) | |
284 | #define BUILTIN_VDW(T, N, MAP) \ | |
285 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
286 | #define BUILTIN_VD_BHSI(T, N, MAP) \ | |
287 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
288 | #define BUILTIN_VD_HSI(T, N, MAP) \ | |
289 | VAR2 (T, N, MAP, v4hi, v2si) | |
290 | #define BUILTIN_VD_RE(T, N, MAP) \ | |
291 | VAR6 (T, N, MAP, v8qi, v4hi, v2si, v2sf, di, df) | |
292 | #define BUILTIN_VQ(T, N, MAP) \ | |
293 | VAR6 (T, N, MAP, v16qi, v8hi, v4si, v2di, v4sf, v2df) | |
294 | #define BUILTIN_VQN(T, N, MAP) \ | |
295 | VAR3 (T, N, MAP, v8hi, v4si, v2di) | |
296 | #define BUILTIN_VQW(T, N, MAP) \ | |
297 | VAR3 (T, N, MAP, v16qi, v8hi, v4si) | |
298 | #define BUILTIN_VQ_HSI(T, N, MAP) \ | |
299 | VAR2 (T, N, MAP, v8hi, v4si) | |
300 | #define BUILTIN_VQ_S(T, N, MAP) \ | |
301 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
302 | #define BUILTIN_VSDQ_HSI(T, N, MAP) \ | |
303 | VAR6 (T, N, MAP, v4hi, v8hi, v2si, v4si, hi, si) | |
304 | #define BUILTIN_VSDQ_I(T, N, MAP) \ | |
305 | VAR11 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, qi, hi, si, di) | |
306 | #define BUILTIN_VSDQ_I_BHSI(T, N, MAP) \ | |
307 | VAR10 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, qi, hi, si) | |
308 | #define BUILTIN_VSDQ_I_DI(T, N, MAP) \ | |
309 | VAR8 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, di) | |
310 | #define BUILTIN_VSD_HSI(T, N, MAP) \ | |
311 | VAR4 (T, N, MAP, v4hi, v2si, hi, si) | |
312 | #define BUILTIN_VSQN_HSDI(T, N, MAP) \ | |
313 | VAR6 (T, N, MAP, v8hi, v4si, v2di, hi, si, di) | |
314 | #define BUILTIN_VSTRUCT(T, N, MAP) \ | |
315 | VAR3 (T, N, MAP, oi, ci, xi) | |
43e9d192 IB |
316 | |
317 | static aarch64_simd_builtin_datum aarch64_simd_builtin_data[] = { | |
342be7f7 JG |
318 | #include "aarch64-simd-builtins.def" |
319 | }; | |
320 | ||
321 | #undef VAR1 | |
0ddec79f | 322 | #define VAR1(T, N, MAP, A) \ |
342be7f7 JG |
323 | AARCH64_SIMD_BUILTIN_##N##A, |
324 | ||
325 | enum aarch64_builtins | |
326 | { | |
327 | AARCH64_BUILTIN_MIN, | |
328 | AARCH64_SIMD_BUILTIN_BASE, | |
329 | #include "aarch64-simd-builtins.def" | |
330 | AARCH64_SIMD_BUILTIN_MAX = AARCH64_SIMD_BUILTIN_BASE | |
331 | + ARRAY_SIZE (aarch64_simd_builtin_data), | |
332 | AARCH64_BUILTIN_MAX | |
43e9d192 IB |
333 | }; |
334 | ||
119103ca JG |
335 | static GTY(()) tree aarch64_builtin_decls[AARCH64_BUILTIN_MAX]; |
336 | ||
43e9d192 IB |
337 | #define NUM_DREG_TYPES 6 |
338 | #define NUM_QREG_TYPES 6 | |
339 | ||
b5828b4b JG |
340 | /* Return a tree for a signed or unsigned argument of either |
341 | the mode specified by MODE, or the inner mode of MODE. */ | |
342 | tree | |
343 | aarch64_build_scalar_type (enum machine_mode mode, bool unsigned_p) | |
344 | { | |
345 | #undef INT_TYPES | |
346 | #define INT_TYPES \ | |
347 | AARCH64_TYPE_BUILDER (QI) \ | |
348 | AARCH64_TYPE_BUILDER (HI) \ | |
349 | AARCH64_TYPE_BUILDER (SI) \ | |
350 | AARCH64_TYPE_BUILDER (DI) \ | |
351 | AARCH64_TYPE_BUILDER (EI) \ | |
352 | AARCH64_TYPE_BUILDER (OI) \ | |
353 | AARCH64_TYPE_BUILDER (CI) \ | |
354 | AARCH64_TYPE_BUILDER (XI) \ | |
355 | AARCH64_TYPE_BUILDER (TI) \ | |
356 | ||
357 | /* Statically declare all the possible types we might need. */ | |
358 | #undef AARCH64_TYPE_BUILDER | |
359 | #define AARCH64_TYPE_BUILDER(X) \ | |
360 | static tree X##_aarch64_type_node_s = NULL; \ | |
361 | static tree X##_aarch64_type_node_u = NULL; | |
362 | ||
363 | INT_TYPES | |
364 | ||
365 | static tree float_aarch64_type_node = NULL; | |
366 | static tree double_aarch64_type_node = NULL; | |
367 | ||
368 | gcc_assert (!VECTOR_MODE_P (mode)); | |
369 | ||
370 | /* If we've already initialised this type, don't initialise it again, | |
371 | otherwise ask for a new type of the correct size. */ | |
372 | #undef AARCH64_TYPE_BUILDER | |
373 | #define AARCH64_TYPE_BUILDER(X) \ | |
374 | case X##mode: \ | |
375 | if (unsigned_p) \ | |
376 | return (X##_aarch64_type_node_u \ | |
377 | ? X##_aarch64_type_node_u \ | |
378 | : X##_aarch64_type_node_u \ | |
379 | = make_unsigned_type (GET_MODE_PRECISION (mode))); \ | |
380 | else \ | |
381 | return (X##_aarch64_type_node_s \ | |
382 | ? X##_aarch64_type_node_s \ | |
383 | : X##_aarch64_type_node_s \ | |
384 | = make_signed_type (GET_MODE_PRECISION (mode))); \ | |
385 | break; | |
386 | ||
387 | switch (mode) | |
388 | { | |
389 | INT_TYPES | |
390 | case SFmode: | |
391 | if (!float_aarch64_type_node) | |
392 | { | |
393 | float_aarch64_type_node = make_node (REAL_TYPE); | |
394 | TYPE_PRECISION (float_aarch64_type_node) = FLOAT_TYPE_SIZE; | |
395 | layout_type (float_aarch64_type_node); | |
396 | } | |
397 | return float_aarch64_type_node; | |
398 | break; | |
399 | case DFmode: | |
400 | if (!double_aarch64_type_node) | |
401 | { | |
402 | double_aarch64_type_node = make_node (REAL_TYPE); | |
403 | TYPE_PRECISION (double_aarch64_type_node) = DOUBLE_TYPE_SIZE; | |
404 | layout_type (double_aarch64_type_node); | |
405 | } | |
406 | return double_aarch64_type_node; | |
407 | break; | |
408 | default: | |
409 | gcc_unreachable (); | |
410 | } | |
411 | } | |
412 | ||
413 | tree | |
414 | aarch64_build_vector_type (enum machine_mode mode, bool unsigned_p) | |
415 | { | |
416 | tree eltype; | |
417 | ||
418 | #define VECTOR_TYPES \ | |
419 | AARCH64_TYPE_BUILDER (V16QI) \ | |
420 | AARCH64_TYPE_BUILDER (V8HI) \ | |
421 | AARCH64_TYPE_BUILDER (V4SI) \ | |
422 | AARCH64_TYPE_BUILDER (V2DI) \ | |
423 | AARCH64_TYPE_BUILDER (V8QI) \ | |
424 | AARCH64_TYPE_BUILDER (V4HI) \ | |
425 | AARCH64_TYPE_BUILDER (V2SI) \ | |
426 | \ | |
427 | AARCH64_TYPE_BUILDER (V4SF) \ | |
428 | AARCH64_TYPE_BUILDER (V2DF) \ | |
429 | AARCH64_TYPE_BUILDER (V2SF) \ | |
430 | /* Declare our "cache" of values. */ | |
431 | #undef AARCH64_TYPE_BUILDER | |
432 | #define AARCH64_TYPE_BUILDER(X) \ | |
433 | static tree X##_aarch64_type_node_s = NULL; \ | |
434 | static tree X##_aarch64_type_node_u = NULL; | |
435 | ||
436 | VECTOR_TYPES | |
437 | ||
438 | gcc_assert (VECTOR_MODE_P (mode)); | |
439 | ||
440 | #undef AARCH64_TYPE_BUILDER | |
441 | #define AARCH64_TYPE_BUILDER(X) \ | |
442 | case X##mode: \ | |
443 | if (unsigned_p) \ | |
444 | return X##_aarch64_type_node_u \ | |
445 | ? X##_aarch64_type_node_u \ | |
446 | : X##_aarch64_type_node_u \ | |
447 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
448 | (GET_MODE_INNER (mode), \ | |
449 | unsigned_p), mode); \ | |
450 | else \ | |
451 | return X##_aarch64_type_node_s \ | |
452 | ? X##_aarch64_type_node_s \ | |
453 | : X##_aarch64_type_node_s \ | |
454 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
455 | (GET_MODE_INNER (mode), \ | |
456 | unsigned_p), mode); \ | |
457 | break; | |
458 | ||
459 | switch (mode) | |
460 | { | |
461 | default: | |
462 | eltype = aarch64_build_scalar_type (GET_MODE_INNER (mode), unsigned_p); | |
463 | return build_vector_type_for_mode (eltype, mode); | |
464 | break; | |
465 | VECTOR_TYPES | |
466 | } | |
467 | } | |
468 | ||
469 | tree | |
470 | aarch64_build_type (enum machine_mode mode, bool unsigned_p) | |
471 | { | |
472 | if (VECTOR_MODE_P (mode)) | |
473 | return aarch64_build_vector_type (mode, unsigned_p); | |
474 | else | |
475 | return aarch64_build_scalar_type (mode, unsigned_p); | |
476 | } | |
477 | ||
af55e82d | 478 | static void |
342be7f7 | 479 | aarch64_init_simd_builtins (void) |
43e9d192 | 480 | { |
342be7f7 | 481 | unsigned int i, fcode = AARCH64_SIMD_BUILTIN_BASE + 1; |
43e9d192 | 482 | |
b5828b4b JG |
483 | /* In order that 'poly' types mangle correctly they must not share |
484 | a base tree with the other scalar types, thus we must generate them | |
485 | as a special case. */ | |
486 | tree aarch64_simd_polyQI_type_node = | |
43e9d192 | 487 | make_signed_type (GET_MODE_PRECISION (QImode)); |
b5828b4b | 488 | tree aarch64_simd_polyHI_type_node = |
43e9d192 | 489 | make_signed_type (GET_MODE_PRECISION (HImode)); |
b5828b4b JG |
490 | |
491 | /* Scalar type nodes. */ | |
492 | tree aarch64_simd_intQI_type_node = aarch64_build_type (QImode, false); | |
493 | tree aarch64_simd_intHI_type_node = aarch64_build_type (HImode, false); | |
494 | tree aarch64_simd_intSI_type_node = aarch64_build_type (SImode, false); | |
495 | tree aarch64_simd_intDI_type_node = aarch64_build_type (DImode, false); | |
496 | tree aarch64_simd_intTI_type_node = aarch64_build_type (TImode, false); | |
497 | tree aarch64_simd_intEI_type_node = aarch64_build_type (EImode, false); | |
498 | tree aarch64_simd_intOI_type_node = aarch64_build_type (OImode, false); | |
499 | tree aarch64_simd_intCI_type_node = aarch64_build_type (CImode, false); | |
500 | tree aarch64_simd_intXI_type_node = aarch64_build_type (XImode, false); | |
501 | tree aarch64_simd_intUQI_type_node = aarch64_build_type (QImode, true); | |
502 | tree aarch64_simd_intUHI_type_node = aarch64_build_type (HImode, true); | |
503 | tree aarch64_simd_intUSI_type_node = aarch64_build_type (SImode, true); | |
504 | tree aarch64_simd_intUDI_type_node = aarch64_build_type (DImode, true); | |
505 | ||
506 | /* Float type nodes. */ | |
507 | tree aarch64_simd_float_type_node = aarch64_build_type (SFmode, false); | |
508 | tree aarch64_simd_double_type_node = aarch64_build_type (DFmode, false); | |
43e9d192 IB |
509 | |
510 | /* Define typedefs which exactly correspond to the modes we are basing vector | |
511 | types on. If you change these names you'll need to change | |
512 | the table used by aarch64_mangle_type too. */ | |
513 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intQI_type_node, | |
514 | "__builtin_aarch64_simd_qi"); | |
515 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intHI_type_node, | |
516 | "__builtin_aarch64_simd_hi"); | |
517 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intSI_type_node, | |
518 | "__builtin_aarch64_simd_si"); | |
519 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_float_type_node, | |
520 | "__builtin_aarch64_simd_sf"); | |
521 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intDI_type_node, | |
522 | "__builtin_aarch64_simd_di"); | |
523 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_double_type_node, | |
524 | "__builtin_aarch64_simd_df"); | |
525 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyQI_type_node, | |
526 | "__builtin_aarch64_simd_poly8"); | |
527 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyHI_type_node, | |
528 | "__builtin_aarch64_simd_poly16"); | |
b5828b4b | 529 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intTI_type_node, |
43e9d192 | 530 | "__builtin_aarch64_simd_ti"); |
b5828b4b | 531 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intEI_type_node, |
43e9d192 | 532 | "__builtin_aarch64_simd_ei"); |
b5828b4b | 533 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intOI_type_node, |
43e9d192 | 534 | "__builtin_aarch64_simd_oi"); |
b5828b4b | 535 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intCI_type_node, |
43e9d192 | 536 | "__builtin_aarch64_simd_ci"); |
b5828b4b | 537 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intXI_type_node, |
43e9d192 IB |
538 | "__builtin_aarch64_simd_xi"); |
539 | ||
b5828b4b JG |
540 | /* Unsigned integer types for various mode sizes. */ |
541 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUQI_type_node, | |
542 | "__builtin_aarch64_simd_uqi"); | |
543 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUHI_type_node, | |
544 | "__builtin_aarch64_simd_uhi"); | |
545 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUSI_type_node, | |
546 | "__builtin_aarch64_simd_usi"); | |
547 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUDI_type_node, | |
548 | "__builtin_aarch64_simd_udi"); | |
43e9d192 | 549 | |
342be7f7 | 550 | for (i = 0; i < ARRAY_SIZE (aarch64_simd_builtin_data); i++, fcode++) |
43e9d192 | 551 | { |
b5828b4b JG |
552 | bool print_type_signature_p = false; |
553 | char type_signature[SIMD_MAX_BUILTIN_ARGS] = { 0 }; | |
43e9d192 | 554 | aarch64_simd_builtin_datum *d = &aarch64_simd_builtin_data[i]; |
342be7f7 | 555 | const char *const modenames[] = |
b5828b4b JG |
556 | { |
557 | "v8qi", "v4hi", "v2si", "v2sf", "di", "df", | |
558 | "v16qi", "v8hi", "v4si", "v4sf", "v2di", "v2df", | |
559 | "ti", "ei", "oi", "xi", "si", "sf", "hi", "qi" | |
560 | }; | |
561 | const enum machine_mode modes[] = | |
562 | { | |
563 | V8QImode, V4HImode, V2SImode, V2SFmode, DImode, DFmode, | |
564 | V16QImode, V8HImode, V4SImode, V4SFmode, V2DImode, | |
565 | V2DFmode, TImode, EImode, OImode, XImode, SImode, | |
566 | SFmode, HImode, QImode | |
567 | }; | |
342be7f7 JG |
568 | char namebuf[60]; |
569 | tree ftype = NULL; | |
119103ca | 570 | tree fndecl = NULL; |
342be7f7 JG |
571 | |
572 | gcc_assert (ARRAY_SIZE (modenames) == T_MAX); | |
43e9d192 | 573 | |
342be7f7 | 574 | d->fcode = fcode; |
43e9d192 | 575 | |
b5828b4b JG |
576 | /* We must track two variables here. op_num is |
577 | the operand number as in the RTL pattern. This is | |
578 | required to access the mode (e.g. V4SF mode) of the | |
579 | argument, from which the base type can be derived. | |
580 | arg_num is an index in to the qualifiers data, which | |
581 | gives qualifiers to the type (e.g. const unsigned). | |
582 | The reason these two variables may differ by one is the | |
583 | void return type. While all return types take the 0th entry | |
584 | in the qualifiers array, there is no operand for them in the | |
585 | RTL pattern. */ | |
586 | int op_num = insn_data[d->code].n_operands - 1; | |
587 | int arg_num = d->qualifiers[0] & qualifier_void | |
588 | ? op_num + 1 | |
589 | : op_num; | |
590 | tree return_type = void_type_node, args = void_list_node; | |
591 | tree eltype; | |
592 | ||
593 | /* Build a function type directly from the insn_data for this | |
594 | builtin. The build_function_type () function takes care of | |
595 | removing duplicates for us. */ | |
596 | for (; op_num >= 0; arg_num--, op_num--) | |
43e9d192 | 597 | { |
b5828b4b JG |
598 | enum machine_mode op_mode = insn_data[d->code].operand[op_num].mode; |
599 | enum aarch64_type_qualifiers qualifiers = d->qualifiers[arg_num]; | |
43e9d192 | 600 | |
b5828b4b JG |
601 | if (qualifiers & qualifier_unsigned) |
602 | { | |
603 | type_signature[arg_num] = 'u'; | |
604 | print_type_signature_p = true; | |
605 | } | |
606 | else | |
607 | type_signature[arg_num] = 's'; | |
608 | ||
609 | /* Skip an internal operand for vget_{low, high}. */ | |
610 | if (qualifiers & qualifier_internal) | |
611 | continue; | |
612 | ||
613 | /* Some builtins have different user-facing types | |
614 | for certain arguments, encoded in d->mode. */ | |
615 | if (qualifiers & qualifier_map_mode) | |
616 | op_mode = modes[d->mode]; | |
617 | ||
618 | /* For pointers, we want a pointer to the basic type | |
619 | of the vector. */ | |
620 | if (qualifiers & qualifier_pointer && VECTOR_MODE_P (op_mode)) | |
621 | op_mode = GET_MODE_INNER (op_mode); | |
622 | ||
623 | eltype = aarch64_build_type (op_mode, | |
624 | qualifiers & qualifier_unsigned); | |
625 | ||
626 | /* Add qualifiers. */ | |
627 | if (qualifiers & qualifier_const) | |
628 | eltype = build_qualified_type (eltype, TYPE_QUAL_CONST); | |
629 | ||
630 | if (qualifiers & qualifier_pointer) | |
631 | eltype = build_pointer_type (eltype); | |
632 | ||
633 | /* If we have reached arg_num == 0, we are at a non-void | |
634 | return type. Otherwise, we are still processing | |
635 | arguments. */ | |
636 | if (arg_num == 0) | |
637 | return_type = eltype; | |
638 | else | |
639 | args = tree_cons (NULL_TREE, eltype, args); | |
640 | } | |
342be7f7 | 641 | |
b5828b4b | 642 | ftype = build_function_type (return_type, args); |
43e9d192 | 643 | |
342be7f7 | 644 | gcc_assert (ftype != NULL); |
43e9d192 | 645 | |
b5828b4b JG |
646 | if (print_type_signature_p) |
647 | snprintf (namebuf, sizeof (namebuf), "__builtin_aarch64_%s%s_%s", | |
648 | d->name, modenames[d->mode], type_signature); | |
649 | else | |
650 | snprintf (namebuf, sizeof (namebuf), "__builtin_aarch64_%s%s", | |
651 | d->name, modenames[d->mode]); | |
43e9d192 | 652 | |
119103ca JG |
653 | fndecl = add_builtin_function (namebuf, ftype, fcode, BUILT_IN_MD, |
654 | NULL, NULL_TREE); | |
655 | aarch64_builtin_decls[fcode] = fndecl; | |
43e9d192 IB |
656 | } |
657 | } | |
658 | ||
342be7f7 JG |
659 | void |
660 | aarch64_init_builtins (void) | |
43e9d192 | 661 | { |
342be7f7 JG |
662 | if (TARGET_SIMD) |
663 | aarch64_init_simd_builtins (); | |
43e9d192 IB |
664 | } |
665 | ||
119103ca JG |
666 | tree |
667 | aarch64_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
668 | { | |
669 | if (code >= AARCH64_BUILTIN_MAX) | |
670 | return error_mark_node; | |
671 | ||
672 | return aarch64_builtin_decls[code]; | |
673 | } | |
674 | ||
43e9d192 IB |
675 | typedef enum |
676 | { | |
677 | SIMD_ARG_COPY_TO_REG, | |
678 | SIMD_ARG_CONSTANT, | |
679 | SIMD_ARG_STOP | |
680 | } builtin_simd_arg; | |
681 | ||
43e9d192 IB |
682 | static rtx |
683 | aarch64_simd_expand_args (rtx target, int icode, int have_retval, | |
684 | tree exp, ...) | |
685 | { | |
686 | va_list ap; | |
687 | rtx pat; | |
688 | tree arg[SIMD_MAX_BUILTIN_ARGS]; | |
689 | rtx op[SIMD_MAX_BUILTIN_ARGS]; | |
690 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
691 | enum machine_mode mode[SIMD_MAX_BUILTIN_ARGS]; | |
692 | int argc = 0; | |
693 | ||
694 | if (have_retval | |
695 | && (!target | |
696 | || GET_MODE (target) != tmode | |
697 | || !(*insn_data[icode].operand[0].predicate) (target, tmode))) | |
698 | target = gen_reg_rtx (tmode); | |
699 | ||
700 | va_start (ap, exp); | |
701 | ||
702 | for (;;) | |
703 | { | |
704 | builtin_simd_arg thisarg = (builtin_simd_arg) va_arg (ap, int); | |
705 | ||
706 | if (thisarg == SIMD_ARG_STOP) | |
707 | break; | |
708 | else | |
709 | { | |
710 | arg[argc] = CALL_EXPR_ARG (exp, argc); | |
711 | op[argc] = expand_normal (arg[argc]); | |
712 | mode[argc] = insn_data[icode].operand[argc + have_retval].mode; | |
713 | ||
714 | switch (thisarg) | |
715 | { | |
716 | case SIMD_ARG_COPY_TO_REG: | |
2888c331 YZ |
717 | if (POINTER_TYPE_P (TREE_TYPE (arg[argc]))) |
718 | op[argc] = convert_memory_address (Pmode, op[argc]); | |
43e9d192 IB |
719 | /*gcc_assert (GET_MODE (op[argc]) == mode[argc]); */ |
720 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
721 | (op[argc], mode[argc])) | |
722 | op[argc] = copy_to_mode_reg (mode[argc], op[argc]); | |
723 | break; | |
724 | ||
725 | case SIMD_ARG_CONSTANT: | |
726 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
727 | (op[argc], mode[argc])) | |
728 | error_at (EXPR_LOCATION (exp), "incompatible type for argument %d, " | |
729 | "expected %<const int%>", argc + 1); | |
730 | break; | |
731 | ||
732 | case SIMD_ARG_STOP: | |
733 | gcc_unreachable (); | |
734 | } | |
735 | ||
736 | argc++; | |
737 | } | |
738 | } | |
739 | ||
740 | va_end (ap); | |
741 | ||
742 | if (have_retval) | |
743 | switch (argc) | |
744 | { | |
745 | case 1: | |
746 | pat = GEN_FCN (icode) (target, op[0]); | |
747 | break; | |
748 | ||
749 | case 2: | |
750 | pat = GEN_FCN (icode) (target, op[0], op[1]); | |
751 | break; | |
752 | ||
753 | case 3: | |
754 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2]); | |
755 | break; | |
756 | ||
757 | case 4: | |
758 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3]); | |
759 | break; | |
760 | ||
761 | case 5: | |
762 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3], op[4]); | |
763 | break; | |
764 | ||
765 | default: | |
766 | gcc_unreachable (); | |
767 | } | |
768 | else | |
769 | switch (argc) | |
770 | { | |
771 | case 1: | |
772 | pat = GEN_FCN (icode) (op[0]); | |
773 | break; | |
774 | ||
775 | case 2: | |
776 | pat = GEN_FCN (icode) (op[0], op[1]); | |
777 | break; | |
778 | ||
779 | case 3: | |
780 | pat = GEN_FCN (icode) (op[0], op[1], op[2]); | |
781 | break; | |
782 | ||
783 | case 4: | |
784 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]); | |
785 | break; | |
786 | ||
787 | case 5: | |
788 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3], op[4]); | |
789 | break; | |
790 | ||
791 | default: | |
792 | gcc_unreachable (); | |
793 | } | |
794 | ||
795 | if (!pat) | |
796 | return 0; | |
797 | ||
798 | emit_insn (pat); | |
799 | ||
800 | return target; | |
801 | } | |
802 | ||
803 | /* Expand an AArch64 AdvSIMD builtin(intrinsic). */ | |
804 | rtx | |
805 | aarch64_simd_expand_builtin (int fcode, tree exp, rtx target) | |
806 | { | |
342be7f7 JG |
807 | aarch64_simd_builtin_datum *d = |
808 | &aarch64_simd_builtin_data[fcode - (AARCH64_SIMD_BUILTIN_BASE + 1)]; | |
342be7f7 | 809 | enum insn_code icode = d->code; |
b5828b4b JG |
810 | builtin_simd_arg args[SIMD_MAX_BUILTIN_ARGS]; |
811 | int num_args = insn_data[d->code].n_operands; | |
812 | int is_void = 0; | |
813 | int k; | |
43e9d192 | 814 | |
b5828b4b | 815 | is_void = !!(d->qualifiers[0] & qualifier_void); |
43e9d192 | 816 | |
b5828b4b JG |
817 | num_args += is_void; |
818 | ||
819 | for (k = 1; k < num_args; k++) | |
820 | { | |
821 | /* We have four arrays of data, each indexed in a different fashion. | |
822 | qualifiers - element 0 always describes the function return type. | |
823 | operands - element 0 is either the operand for return value (if | |
824 | the function has a non-void return type) or the operand for the | |
825 | first argument. | |
826 | expr_args - element 0 always holds the first argument. | |
827 | args - element 0 is always used for the return type. */ | |
828 | int qualifiers_k = k; | |
829 | int operands_k = k - is_void; | |
830 | int expr_args_k = k - 1; | |
831 | ||
832 | if (d->qualifiers[qualifiers_k] & qualifier_immediate) | |
833 | args[k] = SIMD_ARG_CONSTANT; | |
834 | else if (d->qualifiers[qualifiers_k] & qualifier_maybe_immediate) | |
835 | { | |
836 | rtx arg | |
837 | = expand_normal (CALL_EXPR_ARG (exp, | |
838 | (expr_args_k))); | |
839 | /* Handle constants only if the predicate allows it. */ | |
840 | bool op_const_int_p = | |
841 | (CONST_INT_P (arg) | |
842 | && (*insn_data[icode].operand[operands_k].predicate) | |
843 | (arg, insn_data[icode].operand[operands_k].mode)); | |
844 | args[k] = op_const_int_p ? SIMD_ARG_CONSTANT : SIMD_ARG_COPY_TO_REG; | |
845 | } | |
846 | else | |
847 | args[k] = SIMD_ARG_COPY_TO_REG; | |
43e9d192 | 848 | |
43e9d192 | 849 | } |
b5828b4b JG |
850 | args[k] = SIMD_ARG_STOP; |
851 | ||
852 | /* The interface to aarch64_simd_expand_args expects a 0 if | |
853 | the function is void, and a 1 if it is not. */ | |
854 | return aarch64_simd_expand_args | |
855 | (target, icode, !is_void, exp, | |
856 | args[1], | |
857 | args[2], | |
858 | args[3], | |
859 | args[4], | |
860 | SIMD_ARG_STOP); | |
43e9d192 | 861 | } |
342be7f7 JG |
862 | |
863 | /* Expand an expression EXP that calls a built-in function, | |
864 | with result going to TARGET if that's convenient. */ | |
865 | rtx | |
866 | aarch64_expand_builtin (tree exp, | |
867 | rtx target, | |
868 | rtx subtarget ATTRIBUTE_UNUSED, | |
869 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
870 | int ignore ATTRIBUTE_UNUSED) | |
871 | { | |
872 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
873 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
874 | ||
875 | if (fcode >= AARCH64_SIMD_BUILTIN_BASE) | |
876 | return aarch64_simd_expand_builtin (fcode, exp, target); | |
877 | ||
878 | return NULL_RTX; | |
879 | } | |
42fc9a7f JG |
880 | |
881 | tree | |
882 | aarch64_builtin_vectorized_function (tree fndecl, tree type_out, tree type_in) | |
883 | { | |
884 | enum machine_mode in_mode, out_mode; | |
885 | int in_n, out_n; | |
886 | ||
887 | if (TREE_CODE (type_out) != VECTOR_TYPE | |
888 | || TREE_CODE (type_in) != VECTOR_TYPE) | |
889 | return NULL_TREE; | |
890 | ||
891 | out_mode = TYPE_MODE (TREE_TYPE (type_out)); | |
892 | out_n = TYPE_VECTOR_SUBPARTS (type_out); | |
893 | in_mode = TYPE_MODE (TREE_TYPE (type_in)); | |
894 | in_n = TYPE_VECTOR_SUBPARTS (type_in); | |
895 | ||
896 | #undef AARCH64_CHECK_BUILTIN_MODE | |
897 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) 1 | |
898 | #define AARCH64_FIND_FRINT_VARIANT(N) \ | |
899 | (AARCH64_CHECK_BUILTIN_MODE (2, D) \ | |
900 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_##N##v2df] \ | |
901 | : (AARCH64_CHECK_BUILTIN_MODE (4, S) \ | |
902 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_##N##v4sf] \ | |
903 | : (AARCH64_CHECK_BUILTIN_MODE (2, S) \ | |
904 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_##N##v2sf] \ | |
905 | : NULL_TREE))) | |
906 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
907 | { | |
908 | enum built_in_function fn = DECL_FUNCTION_CODE (fndecl); | |
909 | switch (fn) | |
910 | { | |
911 | #undef AARCH64_CHECK_BUILTIN_MODE | |
912 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ | |
913 | (out_mode == N##Fmode && out_n == C \ | |
914 | && in_mode == N##Fmode && in_n == C) | |
915 | case BUILT_IN_FLOOR: | |
916 | case BUILT_IN_FLOORF: | |
0659ce6f | 917 | return AARCH64_FIND_FRINT_VARIANT (floor); |
42fc9a7f JG |
918 | case BUILT_IN_CEIL: |
919 | case BUILT_IN_CEILF: | |
0659ce6f | 920 | return AARCH64_FIND_FRINT_VARIANT (ceil); |
42fc9a7f JG |
921 | case BUILT_IN_TRUNC: |
922 | case BUILT_IN_TRUNCF: | |
0659ce6f | 923 | return AARCH64_FIND_FRINT_VARIANT (btrunc); |
42fc9a7f JG |
924 | case BUILT_IN_ROUND: |
925 | case BUILT_IN_ROUNDF: | |
0659ce6f | 926 | return AARCH64_FIND_FRINT_VARIANT (round); |
42fc9a7f JG |
927 | case BUILT_IN_NEARBYINT: |
928 | case BUILT_IN_NEARBYINTF: | |
0659ce6f | 929 | return AARCH64_FIND_FRINT_VARIANT (nearbyint); |
4dcd1054 JG |
930 | case BUILT_IN_SQRT: |
931 | case BUILT_IN_SQRTF: | |
932 | return AARCH64_FIND_FRINT_VARIANT (sqrt); | |
42fc9a7f | 933 | #undef AARCH64_CHECK_BUILTIN_MODE |
b5574232 VP |
934 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ |
935 | (out_mode == SImode && out_n == C \ | |
936 | && in_mode == N##Imode && in_n == C) | |
937 | case BUILT_IN_CLZ: | |
938 | { | |
939 | if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
940 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_clzv4si]; | |
941 | return NULL_TREE; | |
942 | } | |
943 | #undef AARCH64_CHECK_BUILTIN_MODE | |
42fc9a7f JG |
944 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ |
945 | (out_mode == N##Imode && out_n == C \ | |
946 | && in_mode == N##Fmode && in_n == C) | |
947 | case BUILT_IN_LFLOOR: | |
0386b123 | 948 | case BUILT_IN_IFLOORF: |
ce966824 JG |
949 | { |
950 | tree new_tree = NULL_TREE; | |
951 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) | |
952 | new_tree = | |
953 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lfloorv2dfv2di]; | |
954 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
955 | new_tree = | |
956 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lfloorv4sfv4si]; | |
957 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) | |
958 | new_tree = | |
959 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lfloorv2sfv2si]; | |
960 | return new_tree; | |
961 | } | |
42fc9a7f | 962 | case BUILT_IN_LCEIL: |
0386b123 | 963 | case BUILT_IN_ICEILF: |
ce966824 JG |
964 | { |
965 | tree new_tree = NULL_TREE; | |
966 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) | |
967 | new_tree = | |
968 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lceilv2dfv2di]; | |
969 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
970 | new_tree = | |
971 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lceilv4sfv4si]; | |
972 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) | |
973 | new_tree = | |
974 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lceilv2sfv2si]; | |
975 | return new_tree; | |
976 | } | |
0386b123 JG |
977 | case BUILT_IN_LROUND: |
978 | case BUILT_IN_IROUNDF: | |
979 | { | |
980 | tree new_tree = NULL_TREE; | |
981 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) | |
982 | new_tree = | |
983 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lroundv2dfv2di]; | |
984 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
985 | new_tree = | |
986 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lroundv4sfv4si]; | |
987 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) | |
988 | new_tree = | |
989 | aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_lroundv2sfv2si]; | |
990 | return new_tree; | |
991 | } | |
992 | ||
42fc9a7f JG |
993 | default: |
994 | return NULL_TREE; | |
995 | } | |
996 | } | |
997 | ||
998 | return NULL_TREE; | |
999 | } | |
0ac198d3 JG |
1000 | |
1001 | #undef VAR1 | |
1002 | #define VAR1(T, N, MAP, A) \ | |
1003 | case AARCH64_SIMD_BUILTIN_##N##A: | |
1004 | ||
9697e620 JG |
1005 | tree |
1006 | aarch64_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED, tree *args, | |
1007 | bool ignore ATTRIBUTE_UNUSED) | |
1008 | { | |
1009 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
1010 | tree type = TREE_TYPE (TREE_TYPE (fndecl)); | |
1011 | ||
1012 | switch (fcode) | |
1013 | { | |
d05d0709 | 1014 | BUILTIN_VALLDI (UNOP, abs, 2) |
9697e620 JG |
1015 | return fold_build1 (ABS_EXPR, type, args[0]); |
1016 | break; | |
bb60efd9 JG |
1017 | BUILTIN_VALLDI (BINOP, cmge, 0) |
1018 | return fold_build2 (GE_EXPR, type, args[0], args[1]); | |
1019 | break; | |
1020 | BUILTIN_VALLDI (BINOP, cmgt, 0) | |
1021 | return fold_build2 (GT_EXPR, type, args[0], args[1]); | |
1022 | break; | |
1023 | BUILTIN_VALLDI (BINOP, cmeq, 0) | |
1024 | return fold_build2 (EQ_EXPR, type, args[0], args[1]); | |
1025 | break; | |
1026 | BUILTIN_VSDQ_I_DI (BINOP, cmtst, 0) | |
1027 | { | |
1028 | tree and_node = fold_build2 (BIT_AND_EXPR, type, args[0], args[1]); | |
1029 | tree vec_zero_node = build_zero_cst (type); | |
1030 | return fold_build2 (NE_EXPR, type, and_node, vec_zero_node); | |
1031 | break; | |
1032 | } | |
1709ff9b JG |
1033 | VAR1 (UNOP, floatv2si, 2, v2sf) |
1034 | VAR1 (UNOP, floatv4si, 2, v4sf) | |
1035 | VAR1 (UNOP, floatv2di, 2, v2df) | |
1036 | return fold_build1 (FLOAT_EXPR, type, args[0]); | |
9697e620 JG |
1037 | default: |
1038 | break; | |
1039 | } | |
1040 | ||
1041 | return NULL_TREE; | |
1042 | } | |
1043 | ||
0ac198d3 JG |
1044 | bool |
1045 | aarch64_gimple_fold_builtin (gimple_stmt_iterator *gsi) | |
1046 | { | |
1047 | bool changed = false; | |
1048 | gimple stmt = gsi_stmt (*gsi); | |
1049 | tree call = gimple_call_fn (stmt); | |
1050 | tree fndecl; | |
1051 | gimple new_stmt = NULL; | |
1052 | if (call) | |
1053 | { | |
1054 | fndecl = gimple_call_fndecl (stmt); | |
1055 | if (fndecl) | |
1056 | { | |
1057 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
1058 | int nargs = gimple_call_num_args (stmt); | |
1059 | tree *args = (nargs > 0 | |
1060 | ? gimple_call_arg_ptr (stmt, 0) | |
1061 | : &error_mark_node); | |
1062 | ||
1063 | switch (fcode) | |
1064 | { | |
36054fab | 1065 | BUILTIN_VALL (UNOP, reduc_splus_, 10) |
0ac198d3 JG |
1066 | new_stmt = gimple_build_assign_with_ops ( |
1067 | REDUC_PLUS_EXPR, | |
1068 | gimple_call_lhs (stmt), | |
1069 | args[0], | |
1070 | NULL_TREE); | |
1071 | break; | |
1598945b JG |
1072 | BUILTIN_VDQIF (UNOP, reduc_smax_, 10) |
1073 | new_stmt = gimple_build_assign_with_ops ( | |
1074 | REDUC_MAX_EXPR, | |
1075 | gimple_call_lhs (stmt), | |
1076 | args[0], | |
1077 | NULL_TREE); | |
1078 | break; | |
1079 | BUILTIN_VDQIF (UNOP, reduc_smin_, 10) | |
1080 | new_stmt = gimple_build_assign_with_ops ( | |
1081 | REDUC_MIN_EXPR, | |
1082 | gimple_call_lhs (stmt), | |
1083 | args[0], | |
1084 | NULL_TREE); | |
1085 | break; | |
1086 | ||
0ac198d3 JG |
1087 | default: |
1088 | break; | |
1089 | } | |
1090 | } | |
1091 | } | |
1092 | ||
1093 | if (new_stmt) | |
1094 | { | |
1095 | gsi_replace (gsi, new_stmt, true); | |
1096 | changed = true; | |
1097 | } | |
1098 | ||
1099 | return changed; | |
1100 | } | |
1101 | ||
42fc9a7f JG |
1102 | #undef AARCH64_CHECK_BUILTIN_MODE |
1103 | #undef AARCH64_FIND_FRINT_VARIANT | |
0ddec79f JG |
1104 | #undef BUILTIN_DX |
1105 | #undef BUILTIN_SDQ_I | |
1106 | #undef BUILTIN_SD_HSI | |
1107 | #undef BUILTIN_V2F | |
1108 | #undef BUILTIN_VALL | |
1109 | #undef BUILTIN_VB | |
1110 | #undef BUILTIN_VD | |
1111 | #undef BUILTIN_VDC | |
1112 | #undef BUILTIN_VDIC | |
1113 | #undef BUILTIN_VDN | |
1114 | #undef BUILTIN_VDQ | |
1115 | #undef BUILTIN_VDQF | |
1116 | #undef BUILTIN_VDQH | |
1117 | #undef BUILTIN_VDQHS | |
1118 | #undef BUILTIN_VDQIF | |
1119 | #undef BUILTIN_VDQM | |
1120 | #undef BUILTIN_VDQV | |
1121 | #undef BUILTIN_VDQ_BHSI | |
1122 | #undef BUILTIN_VDQ_I | |
1123 | #undef BUILTIN_VDW | |
1124 | #undef BUILTIN_VD_BHSI | |
1125 | #undef BUILTIN_VD_HSI | |
1126 | #undef BUILTIN_VD_RE | |
1127 | #undef BUILTIN_VQ | |
1128 | #undef BUILTIN_VQN | |
1129 | #undef BUILTIN_VQW | |
1130 | #undef BUILTIN_VQ_HSI | |
1131 | #undef BUILTIN_VQ_S | |
1132 | #undef BUILTIN_VSDQ_HSI | |
1133 | #undef BUILTIN_VSDQ_I | |
1134 | #undef BUILTIN_VSDQ_I_BHSI | |
1135 | #undef BUILTIN_VSDQ_I_DI | |
1136 | #undef BUILTIN_VSD_HSI | |
1137 | #undef BUILTIN_VSQN_HSDI | |
1138 | #undef BUILTIN_VSTRUCT | |
1139 | #undef CF0 | |
1140 | #undef CF1 | |
1141 | #undef CF2 | |
1142 | #undef CF3 | |
1143 | #undef CF4 | |
1144 | #undef CF10 | |
1145 | #undef VAR1 | |
1146 | #undef VAR2 | |
1147 | #undef VAR3 | |
1148 | #undef VAR4 | |
1149 | #undef VAR5 | |
1150 | #undef VAR6 | |
1151 | #undef VAR7 | |
1152 | #undef VAR8 | |
1153 | #undef VAR9 | |
1154 | #undef VAR10 | |
1155 | #undef VAR11 | |
1156 |