]>
Commit | Line | Data |
---|---|---|
43e9d192 | 1 | /* Builtins' description for AArch64 SIMD architecture. |
23a5b65a | 2 | Copyright (C) 2011-2014 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, | |
c6a29a09 | 56 | T_V1DF, |
342be7f7 JG |
57 | T_DI, |
58 | T_DF, | |
59 | T_V16QI, | |
60 | T_V8HI, | |
61 | T_V4SI, | |
62 | T_V4SF, | |
63 | T_V2DI, | |
64 | T_V2DF, | |
65 | T_TI, | |
66 | T_EI, | |
67 | T_OI, | |
68 | T_XI, | |
69 | T_SI, | |
0050faf8 | 70 | T_SF, |
342be7f7 JG |
71 | T_HI, |
72 | T_QI, | |
73 | T_MAX | |
43e9d192 IB |
74 | }; |
75 | ||
76 | #define v8qi_UP T_V8QI | |
77 | #define v4hi_UP T_V4HI | |
78 | #define v2si_UP T_V2SI | |
79 | #define v2sf_UP T_V2SF | |
c6a29a09 | 80 | #define v1df_UP T_V1DF |
43e9d192 IB |
81 | #define di_UP T_DI |
82 | #define df_UP T_DF | |
83 | #define v16qi_UP T_V16QI | |
84 | #define v8hi_UP T_V8HI | |
85 | #define v4si_UP T_V4SI | |
86 | #define v4sf_UP T_V4SF | |
87 | #define v2di_UP T_V2DI | |
88 | #define v2df_UP T_V2DF | |
89 | #define ti_UP T_TI | |
90 | #define ei_UP T_EI | |
91 | #define oi_UP T_OI | |
92 | #define xi_UP T_XI | |
93 | #define si_UP T_SI | |
0050faf8 | 94 | #define sf_UP T_SF |
43e9d192 IB |
95 | #define hi_UP T_HI |
96 | #define qi_UP T_QI | |
97 | ||
98 | #define UP(X) X##_UP | |
99 | ||
b5828b4b JG |
100 | #define SIMD_MAX_BUILTIN_ARGS 5 |
101 | ||
102 | enum aarch64_type_qualifiers | |
43e9d192 | 103 | { |
b5828b4b JG |
104 | /* T foo. */ |
105 | qualifier_none = 0x0, | |
106 | /* unsigned T foo. */ | |
107 | qualifier_unsigned = 0x1, /* 1 << 0 */ | |
108 | /* const T foo. */ | |
109 | qualifier_const = 0x2, /* 1 << 1 */ | |
110 | /* T *foo. */ | |
111 | qualifier_pointer = 0x4, /* 1 << 2 */ | |
112 | /* const T *foo. */ | |
113 | qualifier_const_pointer = 0x6, /* qualifier_const | qualifier_pointer */ | |
114 | /* Used when expanding arguments if an operand could | |
115 | be an immediate. */ | |
116 | qualifier_immediate = 0x8, /* 1 << 3 */ | |
117 | qualifier_maybe_immediate = 0x10, /* 1 << 4 */ | |
118 | /* void foo (...). */ | |
119 | qualifier_void = 0x20, /* 1 << 5 */ | |
120 | /* Some patterns may have internal operands, this qualifier is an | |
121 | instruction to the initialisation code to skip this operand. */ | |
122 | qualifier_internal = 0x40, /* 1 << 6 */ | |
123 | /* Some builtins should use the T_*mode* encoded in a simd_builtin_datum | |
124 | rather than using the type of the operand. */ | |
125 | qualifier_map_mode = 0x80, /* 1 << 7 */ | |
126 | /* qualifier_pointer | qualifier_map_mode */ | |
127 | qualifier_pointer_map_mode = 0x84, | |
128 | /* qualifier_const_pointer | qualifier_map_mode */ | |
6db1ec94 JG |
129 | qualifier_const_pointer_map_mode = 0x86, |
130 | /* Polynomial types. */ | |
131 | qualifier_poly = 0x100 | |
b5828b4b | 132 | }; |
43e9d192 IB |
133 | |
134 | typedef struct | |
135 | { | |
136 | const char *name; | |
342be7f7 JG |
137 | enum aarch64_simd_builtin_type_mode mode; |
138 | const enum insn_code code; | |
139 | unsigned int fcode; | |
b5828b4b | 140 | enum aarch64_type_qualifiers *qualifiers; |
43e9d192 IB |
141 | } aarch64_simd_builtin_datum; |
142 | ||
096c59be AL |
143 | /* The qualifier_internal allows generation of a unary builtin from |
144 | a pattern with a third pseudo-operand such as a match_scratch. */ | |
b5828b4b JG |
145 | static enum aarch64_type_qualifiers |
146 | aarch64_types_unop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
096c59be | 147 | = { qualifier_none, qualifier_none, qualifier_internal }; |
b5828b4b | 148 | #define TYPES_UNOP (aarch64_types_unop_qualifiers) |
5a7a4e80 TB |
149 | static enum aarch64_type_qualifiers |
150 | aarch64_types_unopu_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
151 | = { qualifier_unsigned, qualifier_unsigned }; | |
152 | #define TYPES_UNOPU (aarch64_types_unopu_qualifiers) | |
b5828b4b | 153 | #define TYPES_CREATE (aarch64_types_unop_qualifiers) |
bcd48995 AV |
154 | #define TYPES_REINTERP_SS (aarch64_types_unop_qualifiers) |
155 | static enum aarch64_type_qualifiers | |
156 | aarch64_types_unop_su_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
157 | = { qualifier_none, qualifier_unsigned }; | |
158 | #define TYPES_REINTERP_SU (aarch64_types_unop_su_qualifiers) | |
159 | static enum aarch64_type_qualifiers | |
160 | aarch64_types_unop_sp_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
161 | = { qualifier_none, qualifier_poly }; | |
162 | #define TYPES_REINTERP_SP (aarch64_types_unop_sp_qualifiers) | |
163 | static enum aarch64_type_qualifiers | |
164 | aarch64_types_unop_us_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
165 | = { qualifier_unsigned, qualifier_none }; | |
166 | #define TYPES_REINTERP_US (aarch64_types_unop_us_qualifiers) | |
167 | static enum aarch64_type_qualifiers | |
168 | aarch64_types_unop_ps_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
169 | = { qualifier_poly, qualifier_none }; | |
170 | #define TYPES_REINTERP_PS (aarch64_types_unop_ps_qualifiers) | |
b5828b4b JG |
171 | static enum aarch64_type_qualifiers |
172 | aarch64_types_binop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
173 | = { qualifier_none, qualifier_none, qualifier_maybe_immediate }; | |
174 | #define TYPES_BINOP (aarch64_types_binop_qualifiers) | |
175 | static enum aarch64_type_qualifiers | |
ae0533da AL |
176 | aarch64_types_binopv_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
177 | = { qualifier_void, qualifier_none, qualifier_none }; | |
178 | #define TYPES_BINOPV (aarch64_types_binopv_qualifiers) | |
179 | static enum aarch64_type_qualifiers | |
5a7a4e80 TB |
180 | aarch64_types_binopu_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
181 | = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned }; | |
182 | #define TYPES_BINOPU (aarch64_types_binopu_qualifiers) | |
7baa225d | 183 | static enum aarch64_type_qualifiers |
de10bcce AL |
184 | aarch64_types_binop_uus_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
185 | = { qualifier_unsigned, qualifier_unsigned, qualifier_none }; | |
186 | #define TYPES_BINOP_UUS (aarch64_types_binop_uus_qualifiers) | |
187 | static enum aarch64_type_qualifiers | |
918621d3 AL |
188 | aarch64_types_binop_ssu_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
189 | = { qualifier_none, qualifier_none, qualifier_unsigned }; | |
190 | #define TYPES_BINOP_SSU (aarch64_types_binop_ssu_qualifiers) | |
191 | static enum aarch64_type_qualifiers | |
7baa225d TB |
192 | aarch64_types_binopp_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
193 | = { qualifier_poly, qualifier_poly, qualifier_poly }; | |
194 | #define TYPES_BINOPP (aarch64_types_binopp_qualifiers) | |
195 | ||
5a7a4e80 | 196 | static enum aarch64_type_qualifiers |
b5828b4b JG |
197 | aarch64_types_ternop_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
198 | = { qualifier_none, qualifier_none, qualifier_none, qualifier_none }; | |
199 | #define TYPES_TERNOP (aarch64_types_ternop_qualifiers) | |
30442682 TB |
200 | static enum aarch64_type_qualifiers |
201 | aarch64_types_ternopu_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
202 | = { qualifier_unsigned, qualifier_unsigned, | |
203 | qualifier_unsigned, qualifier_unsigned }; | |
204 | #define TYPES_TERNOPU (aarch64_types_ternopu_qualifiers) | |
205 | ||
b5828b4b JG |
206 | static enum aarch64_type_qualifiers |
207 | aarch64_types_quadop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
208 | = { qualifier_none, qualifier_none, qualifier_none, | |
209 | qualifier_none, qualifier_none }; | |
210 | #define TYPES_QUADOP (aarch64_types_quadop_qualifiers) | |
211 | ||
212 | static enum aarch64_type_qualifiers | |
213 | aarch64_types_getlane_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
214 | = { qualifier_none, qualifier_none, qualifier_immediate }; | |
215 | #define TYPES_GETLANE (aarch64_types_getlane_qualifiers) | |
216 | #define TYPES_SHIFTIMM (aarch64_types_getlane_qualifiers) | |
217 | static enum aarch64_type_qualifiers | |
de10bcce AL |
218 | aarch64_types_shift_to_unsigned_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
219 | = { qualifier_unsigned, qualifier_none, qualifier_immediate }; | |
220 | #define TYPES_SHIFTIMM_USS (aarch64_types_shift_to_unsigned_qualifiers) | |
221 | static enum aarch64_type_qualifiers | |
252c7556 AV |
222 | aarch64_types_unsigned_shift_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
223 | = { qualifier_unsigned, qualifier_unsigned, qualifier_immediate }; | |
224 | #define TYPES_USHIFTIMM (aarch64_types_unsigned_shift_qualifiers) | |
de10bcce | 225 | |
252c7556 | 226 | static enum aarch64_type_qualifiers |
b5828b4b JG |
227 | aarch64_types_setlane_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
228 | = { qualifier_none, qualifier_none, qualifier_none, qualifier_immediate }; | |
229 | #define TYPES_SETLANE (aarch64_types_setlane_qualifiers) | |
230 | #define TYPES_SHIFTINSERT (aarch64_types_setlane_qualifiers) | |
231 | #define TYPES_SHIFTACC (aarch64_types_setlane_qualifiers) | |
232 | ||
de10bcce AL |
233 | static enum aarch64_type_qualifiers |
234 | aarch64_types_unsigned_shiftacc_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
235 | = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned, | |
236 | qualifier_immediate }; | |
237 | #define TYPES_USHIFTACC (aarch64_types_unsigned_shiftacc_qualifiers) | |
238 | ||
239 | ||
b5828b4b JG |
240 | static enum aarch64_type_qualifiers |
241 | aarch64_types_combine_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
242 | = { qualifier_none, qualifier_none, qualifier_none }; | |
243 | #define TYPES_COMBINE (aarch64_types_combine_qualifiers) | |
244 | ||
245 | static enum aarch64_type_qualifiers | |
246 | aarch64_types_load1_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
247 | = { qualifier_none, qualifier_const_pointer_map_mode }; | |
248 | #define TYPES_LOAD1 (aarch64_types_load1_qualifiers) | |
249 | #define TYPES_LOADSTRUCT (aarch64_types_load1_qualifiers) | |
250 | ||
46e778c4 JG |
251 | static enum aarch64_type_qualifiers |
252 | aarch64_types_bsl_p_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
253 | = { qualifier_poly, qualifier_unsigned, | |
254 | qualifier_poly, qualifier_poly }; | |
255 | #define TYPES_BSL_P (aarch64_types_bsl_p_qualifiers) | |
256 | static enum aarch64_type_qualifiers | |
257 | aarch64_types_bsl_s_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
258 | = { qualifier_none, qualifier_unsigned, | |
259 | qualifier_none, qualifier_none }; | |
260 | #define TYPES_BSL_S (aarch64_types_bsl_s_qualifiers) | |
261 | static enum aarch64_type_qualifiers | |
262 | aarch64_types_bsl_u_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
263 | = { qualifier_unsigned, qualifier_unsigned, | |
264 | qualifier_unsigned, qualifier_unsigned }; | |
265 | #define TYPES_BSL_U (aarch64_types_bsl_u_qualifiers) | |
266 | ||
b5828b4b JG |
267 | /* The first argument (return type) of a store should be void type, |
268 | which we represent with qualifier_void. Their first operand will be | |
269 | a DImode pointer to the location to store to, so we must use | |
270 | qualifier_map_mode | qualifier_pointer to build a pointer to the | |
271 | element type of the vector. */ | |
272 | static enum aarch64_type_qualifiers | |
273 | aarch64_types_store1_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
274 | = { qualifier_void, qualifier_pointer_map_mode, qualifier_none }; | |
275 | #define TYPES_STORE1 (aarch64_types_store1_qualifiers) | |
276 | #define TYPES_STORESTRUCT (aarch64_types_store1_qualifiers) | |
ba081b77 JG |
277 | static enum aarch64_type_qualifiers |
278 | aarch64_types_storestruct_lane_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
279 | = { qualifier_void, qualifier_pointer_map_mode, | |
280 | qualifier_none, qualifier_none }; | |
281 | #define TYPES_STORESTRUCT_LANE (aarch64_types_storestruct_lane_qualifiers) | |
b5828b4b | 282 | |
0ddec79f JG |
283 | #define CF0(N, X) CODE_FOR_aarch64_##N##X |
284 | #define CF1(N, X) CODE_FOR_##N##X##1 | |
285 | #define CF2(N, X) CODE_FOR_##N##X##2 | |
286 | #define CF3(N, X) CODE_FOR_##N##X##3 | |
287 | #define CF4(N, X) CODE_FOR_##N##X##4 | |
288 | #define CF10(N, X) CODE_FOR_##N##X | |
289 | ||
290 | #define VAR1(T, N, MAP, A) \ | |
b5828b4b | 291 | {#N, UP (A), CF##MAP (N, A), 0, TYPES_##T}, |
0ddec79f JG |
292 | #define VAR2(T, N, MAP, A, B) \ |
293 | VAR1 (T, N, MAP, A) \ | |
294 | VAR1 (T, N, MAP, B) | |
295 | #define VAR3(T, N, MAP, A, B, C) \ | |
296 | VAR2 (T, N, MAP, A, B) \ | |
297 | VAR1 (T, N, MAP, C) | |
298 | #define VAR4(T, N, MAP, A, B, C, D) \ | |
299 | VAR3 (T, N, MAP, A, B, C) \ | |
300 | VAR1 (T, N, MAP, D) | |
301 | #define VAR5(T, N, MAP, A, B, C, D, E) \ | |
302 | VAR4 (T, N, MAP, A, B, C, D) \ | |
303 | VAR1 (T, N, MAP, E) | |
304 | #define VAR6(T, N, MAP, A, B, C, D, E, F) \ | |
305 | VAR5 (T, N, MAP, A, B, C, D, E) \ | |
306 | VAR1 (T, N, MAP, F) | |
307 | #define VAR7(T, N, MAP, A, B, C, D, E, F, G) \ | |
308 | VAR6 (T, N, MAP, A, B, C, D, E, F) \ | |
309 | VAR1 (T, N, MAP, G) | |
310 | #define VAR8(T, N, MAP, A, B, C, D, E, F, G, H) \ | |
311 | VAR7 (T, N, MAP, A, B, C, D, E, F, G) \ | |
312 | VAR1 (T, N, MAP, H) | |
313 | #define VAR9(T, N, MAP, A, B, C, D, E, F, G, H, I) \ | |
314 | VAR8 (T, N, MAP, A, B, C, D, E, F, G, H) \ | |
315 | VAR1 (T, N, MAP, I) | |
316 | #define VAR10(T, N, MAP, A, B, C, D, E, F, G, H, I, J) \ | |
317 | VAR9 (T, N, MAP, A, B, C, D, E, F, G, H, I) \ | |
318 | VAR1 (T, N, MAP, J) | |
319 | #define VAR11(T, N, MAP, A, B, C, D, E, F, G, H, I, J, K) \ | |
320 | VAR10 (T, N, MAP, A, B, C, D, E, F, G, H, I, J) \ | |
321 | VAR1 (T, N, MAP, K) | |
322 | #define VAR12(T, N, MAP, A, B, C, D, E, F, G, H, I, J, K, L) \ | |
323 | VAR11 (T, N, MAP, A, B, C, D, E, F, G, H, I, J, K) \ | |
324 | VAR1 (T, N, MAP, L) | |
342be7f7 JG |
325 | |
326 | /* BUILTIN_<ITERATOR> macros should expand to cover the same range of | |
327 | modes as is given for each define_mode_iterator in | |
328 | config/aarch64/iterators.md. */ | |
329 | ||
0ddec79f JG |
330 | #define BUILTIN_DX(T, N, MAP) \ |
331 | VAR2 (T, N, MAP, di, df) | |
332 | #define BUILTIN_GPF(T, N, MAP) \ | |
333 | VAR2 (T, N, MAP, sf, df) | |
334 | #define BUILTIN_SDQ_I(T, N, MAP) \ | |
335 | VAR4 (T, N, MAP, qi, hi, si, di) | |
336 | #define BUILTIN_SD_HSI(T, N, MAP) \ | |
337 | VAR2 (T, N, MAP, hi, si) | |
338 | #define BUILTIN_V2F(T, N, MAP) \ | |
339 | VAR2 (T, N, MAP, v2sf, v2df) | |
340 | #define BUILTIN_VALL(T, N, MAP) \ | |
341 | VAR10 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
342 | v4si, v2di, v2sf, v4sf, v2df) | |
bb60efd9 JG |
343 | #define BUILTIN_VALLDI(T, N, MAP) \ |
344 | VAR11 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
345 | v4si, v2di, v2sf, v4sf, v2df, di) | |
46e778c4 JG |
346 | #define BUILTIN_VALLDIF(T, N, MAP) \ |
347 | VAR12 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
348 | v4si, v2di, v2sf, v4sf, v2df, di, df) | |
0ddec79f JG |
349 | #define BUILTIN_VB(T, N, MAP) \ |
350 | VAR2 (T, N, MAP, v8qi, v16qi) | |
351 | #define BUILTIN_VD(T, N, MAP) \ | |
352 | VAR4 (T, N, MAP, v8qi, v4hi, v2si, v2sf) | |
c6a29a09 AL |
353 | #define BUILTIN_VD1(T, N, MAP) \ |
354 | VAR5 (T, N, MAP, v8qi, v4hi, v2si, v2sf, v1df) | |
0ddec79f JG |
355 | #define BUILTIN_VDC(T, N, MAP) \ |
356 | VAR6 (T, N, MAP, v8qi, v4hi, v2si, v2sf, di, df) | |
357 | #define BUILTIN_VDIC(T, N, MAP) \ | |
358 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
359 | #define BUILTIN_VDN(T, N, MAP) \ | |
360 | VAR3 (T, N, MAP, v4hi, v2si, di) | |
361 | #define BUILTIN_VDQ(T, N, MAP) \ | |
362 | VAR7 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di) | |
363 | #define BUILTIN_VDQF(T, N, MAP) \ | |
364 | VAR3 (T, N, MAP, v2sf, v4sf, v2df) | |
74dc11ed AV |
365 | #define BUILTIN_VDQF_DF(T, N, MAP) \ |
366 | VAR4 (T, N, MAP, v2sf, v4sf, v2df, df) | |
0ddec79f JG |
367 | #define BUILTIN_VDQH(T, N, MAP) \ |
368 | VAR2 (T, N, MAP, v4hi, v8hi) | |
369 | #define BUILTIN_VDQHS(T, N, MAP) \ | |
370 | VAR4 (T, N, MAP, v4hi, v8hi, v2si, v4si) | |
371 | #define BUILTIN_VDQIF(T, N, MAP) \ | |
372 | VAR9 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2sf, v4sf, v2df) | |
373 | #define BUILTIN_VDQM(T, N, MAP) \ | |
374 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
375 | #define BUILTIN_VDQV(T, N, MAP) \ | |
376 | VAR5 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v4si) | |
46e778c4 JG |
377 | #define BUILTIN_VDQQH(T, N, MAP) \ |
378 | VAR4 (T, N, MAP, v8qi, v16qi, v4hi, v8hi) | |
0ddec79f JG |
379 | #define BUILTIN_VDQ_BHSI(T, N, MAP) \ |
380 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
381 | #define BUILTIN_VDQ_I(T, N, MAP) \ | |
382 | VAR7 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di) | |
383 | #define BUILTIN_VDW(T, N, MAP) \ | |
384 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
385 | #define BUILTIN_VD_BHSI(T, N, MAP) \ | |
386 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
387 | #define BUILTIN_VD_HSI(T, N, MAP) \ | |
388 | VAR2 (T, N, MAP, v4hi, v2si) | |
0ddec79f JG |
389 | #define BUILTIN_VQ(T, N, MAP) \ |
390 | VAR6 (T, N, MAP, v16qi, v8hi, v4si, v2di, v4sf, v2df) | |
391 | #define BUILTIN_VQN(T, N, MAP) \ | |
392 | VAR3 (T, N, MAP, v8hi, v4si, v2di) | |
393 | #define BUILTIN_VQW(T, N, MAP) \ | |
394 | VAR3 (T, N, MAP, v16qi, v8hi, v4si) | |
395 | #define BUILTIN_VQ_HSI(T, N, MAP) \ | |
396 | VAR2 (T, N, MAP, v8hi, v4si) | |
397 | #define BUILTIN_VQ_S(T, N, MAP) \ | |
398 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
399 | #define BUILTIN_VSDQ_HSI(T, N, MAP) \ | |
400 | VAR6 (T, N, MAP, v4hi, v8hi, v2si, v4si, hi, si) | |
401 | #define BUILTIN_VSDQ_I(T, N, MAP) \ | |
402 | VAR11 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, qi, hi, si, di) | |
403 | #define BUILTIN_VSDQ_I_BHSI(T, N, MAP) \ | |
404 | VAR10 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, qi, hi, si) | |
405 | #define BUILTIN_VSDQ_I_DI(T, N, MAP) \ | |
406 | VAR8 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, di) | |
407 | #define BUILTIN_VSD_HSI(T, N, MAP) \ | |
408 | VAR4 (T, N, MAP, v4hi, v2si, hi, si) | |
409 | #define BUILTIN_VSQN_HSDI(T, N, MAP) \ | |
410 | VAR6 (T, N, MAP, v8hi, v4si, v2di, hi, si, di) | |
411 | #define BUILTIN_VSTRUCT(T, N, MAP) \ | |
412 | VAR3 (T, N, MAP, oi, ci, xi) | |
43e9d192 IB |
413 | |
414 | static aarch64_simd_builtin_datum aarch64_simd_builtin_data[] = { | |
342be7f7 JG |
415 | #include "aarch64-simd-builtins.def" |
416 | }; | |
417 | ||
5d357f26 KT |
418 | /* There's only 8 CRC32 builtins. Probably not worth their own .def file. */ |
419 | #define AARCH64_CRC32_BUILTINS \ | |
420 | CRC32_BUILTIN (crc32b, QI) \ | |
421 | CRC32_BUILTIN (crc32h, HI) \ | |
422 | CRC32_BUILTIN (crc32w, SI) \ | |
423 | CRC32_BUILTIN (crc32x, DI) \ | |
424 | CRC32_BUILTIN (crc32cb, QI) \ | |
425 | CRC32_BUILTIN (crc32ch, HI) \ | |
426 | CRC32_BUILTIN (crc32cw, SI) \ | |
427 | CRC32_BUILTIN (crc32cx, DI) | |
428 | ||
429 | typedef struct | |
430 | { | |
431 | const char *name; | |
432 | enum machine_mode mode; | |
433 | const enum insn_code icode; | |
434 | unsigned int fcode; | |
435 | } aarch64_crc_builtin_datum; | |
436 | ||
437 | #define CRC32_BUILTIN(N, M) \ | |
438 | AARCH64_BUILTIN_##N, | |
439 | ||
342be7f7 | 440 | #undef VAR1 |
0ddec79f | 441 | #define VAR1(T, N, MAP, A) \ |
e993fea1 | 442 | AARCH64_SIMD_BUILTIN_##T##_##N##A, |
342be7f7 JG |
443 | |
444 | enum aarch64_builtins | |
445 | { | |
446 | AARCH64_BUILTIN_MIN, | |
aa87aced KV |
447 | |
448 | AARCH64_BUILTIN_GET_FPCR, | |
449 | AARCH64_BUILTIN_SET_FPCR, | |
450 | AARCH64_BUILTIN_GET_FPSR, | |
451 | AARCH64_BUILTIN_SET_FPSR, | |
452 | ||
342be7f7 JG |
453 | AARCH64_SIMD_BUILTIN_BASE, |
454 | #include "aarch64-simd-builtins.def" | |
455 | AARCH64_SIMD_BUILTIN_MAX = AARCH64_SIMD_BUILTIN_BASE | |
456 | + ARRAY_SIZE (aarch64_simd_builtin_data), | |
5d357f26 KT |
457 | AARCH64_CRC32_BUILTIN_BASE, |
458 | AARCH64_CRC32_BUILTINS | |
459 | AARCH64_CRC32_BUILTIN_MAX, | |
342be7f7 | 460 | AARCH64_BUILTIN_MAX |
43e9d192 IB |
461 | }; |
462 | ||
5d357f26 KT |
463 | #undef CRC32_BUILTIN |
464 | #define CRC32_BUILTIN(N, M) \ | |
465 | {"__builtin_aarch64_"#N, M##mode, CODE_FOR_aarch64_##N, AARCH64_BUILTIN_##N}, | |
466 | ||
467 | static aarch64_crc_builtin_datum aarch64_crc_builtin_data[] = { | |
468 | AARCH64_CRC32_BUILTINS | |
469 | }; | |
470 | ||
471 | #undef CRC32_BUILTIN | |
472 | ||
119103ca JG |
473 | static GTY(()) tree aarch64_builtin_decls[AARCH64_BUILTIN_MAX]; |
474 | ||
43e9d192 IB |
475 | #define NUM_DREG_TYPES 6 |
476 | #define NUM_QREG_TYPES 6 | |
477 | ||
b5828b4b JG |
478 | /* Return a tree for a signed or unsigned argument of either |
479 | the mode specified by MODE, or the inner mode of MODE. */ | |
480 | tree | |
6db1ec94 JG |
481 | aarch64_build_scalar_type (enum machine_mode mode, |
482 | bool unsigned_p, | |
483 | bool poly_p) | |
b5828b4b JG |
484 | { |
485 | #undef INT_TYPES | |
486 | #define INT_TYPES \ | |
487 | AARCH64_TYPE_BUILDER (QI) \ | |
488 | AARCH64_TYPE_BUILDER (HI) \ | |
489 | AARCH64_TYPE_BUILDER (SI) \ | |
490 | AARCH64_TYPE_BUILDER (DI) \ | |
491 | AARCH64_TYPE_BUILDER (EI) \ | |
492 | AARCH64_TYPE_BUILDER (OI) \ | |
493 | AARCH64_TYPE_BUILDER (CI) \ | |
494 | AARCH64_TYPE_BUILDER (XI) \ | |
495 | AARCH64_TYPE_BUILDER (TI) \ | |
496 | ||
497 | /* Statically declare all the possible types we might need. */ | |
498 | #undef AARCH64_TYPE_BUILDER | |
499 | #define AARCH64_TYPE_BUILDER(X) \ | |
6db1ec94 | 500 | static tree X##_aarch64_type_node_p = NULL; \ |
b5828b4b JG |
501 | static tree X##_aarch64_type_node_s = NULL; \ |
502 | static tree X##_aarch64_type_node_u = NULL; | |
503 | ||
504 | INT_TYPES | |
505 | ||
506 | static tree float_aarch64_type_node = NULL; | |
507 | static tree double_aarch64_type_node = NULL; | |
508 | ||
509 | gcc_assert (!VECTOR_MODE_P (mode)); | |
510 | ||
511 | /* If we've already initialised this type, don't initialise it again, | |
512 | otherwise ask for a new type of the correct size. */ | |
513 | #undef AARCH64_TYPE_BUILDER | |
514 | #define AARCH64_TYPE_BUILDER(X) \ | |
515 | case X##mode: \ | |
516 | if (unsigned_p) \ | |
517 | return (X##_aarch64_type_node_u \ | |
518 | ? X##_aarch64_type_node_u \ | |
519 | : X##_aarch64_type_node_u \ | |
520 | = make_unsigned_type (GET_MODE_PRECISION (mode))); \ | |
6db1ec94 JG |
521 | else if (poly_p) \ |
522 | return (X##_aarch64_type_node_p \ | |
523 | ? X##_aarch64_type_node_p \ | |
524 | : X##_aarch64_type_node_p \ | |
525 | = make_unsigned_type (GET_MODE_PRECISION (mode))); \ | |
b5828b4b JG |
526 | else \ |
527 | return (X##_aarch64_type_node_s \ | |
528 | ? X##_aarch64_type_node_s \ | |
529 | : X##_aarch64_type_node_s \ | |
530 | = make_signed_type (GET_MODE_PRECISION (mode))); \ | |
531 | break; | |
532 | ||
533 | switch (mode) | |
534 | { | |
535 | INT_TYPES | |
536 | case SFmode: | |
537 | if (!float_aarch64_type_node) | |
538 | { | |
539 | float_aarch64_type_node = make_node (REAL_TYPE); | |
540 | TYPE_PRECISION (float_aarch64_type_node) = FLOAT_TYPE_SIZE; | |
541 | layout_type (float_aarch64_type_node); | |
542 | } | |
543 | return float_aarch64_type_node; | |
544 | break; | |
545 | case DFmode: | |
546 | if (!double_aarch64_type_node) | |
547 | { | |
548 | double_aarch64_type_node = make_node (REAL_TYPE); | |
549 | TYPE_PRECISION (double_aarch64_type_node) = DOUBLE_TYPE_SIZE; | |
550 | layout_type (double_aarch64_type_node); | |
551 | } | |
552 | return double_aarch64_type_node; | |
553 | break; | |
554 | default: | |
555 | gcc_unreachable (); | |
556 | } | |
557 | } | |
558 | ||
559 | tree | |
6db1ec94 JG |
560 | aarch64_build_vector_type (enum machine_mode mode, |
561 | bool unsigned_p, | |
562 | bool poly_p) | |
b5828b4b JG |
563 | { |
564 | tree eltype; | |
565 | ||
566 | #define VECTOR_TYPES \ | |
567 | AARCH64_TYPE_BUILDER (V16QI) \ | |
568 | AARCH64_TYPE_BUILDER (V8HI) \ | |
569 | AARCH64_TYPE_BUILDER (V4SI) \ | |
570 | AARCH64_TYPE_BUILDER (V2DI) \ | |
571 | AARCH64_TYPE_BUILDER (V8QI) \ | |
572 | AARCH64_TYPE_BUILDER (V4HI) \ | |
573 | AARCH64_TYPE_BUILDER (V2SI) \ | |
574 | \ | |
575 | AARCH64_TYPE_BUILDER (V4SF) \ | |
576 | AARCH64_TYPE_BUILDER (V2DF) \ | |
577 | AARCH64_TYPE_BUILDER (V2SF) \ | |
578 | /* Declare our "cache" of values. */ | |
579 | #undef AARCH64_TYPE_BUILDER | |
580 | #define AARCH64_TYPE_BUILDER(X) \ | |
581 | static tree X##_aarch64_type_node_s = NULL; \ | |
6db1ec94 JG |
582 | static tree X##_aarch64_type_node_u = NULL; \ |
583 | static tree X##_aarch64_type_node_p = NULL; | |
b5828b4b JG |
584 | |
585 | VECTOR_TYPES | |
586 | ||
587 | gcc_assert (VECTOR_MODE_P (mode)); | |
588 | ||
589 | #undef AARCH64_TYPE_BUILDER | |
590 | #define AARCH64_TYPE_BUILDER(X) \ | |
591 | case X##mode: \ | |
592 | if (unsigned_p) \ | |
593 | return X##_aarch64_type_node_u \ | |
594 | ? X##_aarch64_type_node_u \ | |
595 | : X##_aarch64_type_node_u \ | |
596 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
597 | (GET_MODE_INNER (mode), \ | |
6db1ec94 JG |
598 | unsigned_p, poly_p), mode); \ |
599 | else if (poly_p) \ | |
600 | return X##_aarch64_type_node_p \ | |
601 | ? X##_aarch64_type_node_p \ | |
602 | : X##_aarch64_type_node_p \ | |
603 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
604 | (GET_MODE_INNER (mode), \ | |
605 | unsigned_p, poly_p), mode); \ | |
b5828b4b JG |
606 | else \ |
607 | return X##_aarch64_type_node_s \ | |
608 | ? X##_aarch64_type_node_s \ | |
609 | : X##_aarch64_type_node_s \ | |
610 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
611 | (GET_MODE_INNER (mode), \ | |
6db1ec94 | 612 | unsigned_p, poly_p), mode); \ |
b5828b4b JG |
613 | break; |
614 | ||
615 | switch (mode) | |
616 | { | |
617 | default: | |
6db1ec94 JG |
618 | eltype = aarch64_build_scalar_type (GET_MODE_INNER (mode), |
619 | unsigned_p, poly_p); | |
b5828b4b JG |
620 | return build_vector_type_for_mode (eltype, mode); |
621 | break; | |
622 | VECTOR_TYPES | |
623 | } | |
624 | } | |
625 | ||
626 | tree | |
6db1ec94 | 627 | aarch64_build_type (enum machine_mode mode, bool unsigned_p, bool poly_p) |
b5828b4b JG |
628 | { |
629 | if (VECTOR_MODE_P (mode)) | |
6db1ec94 | 630 | return aarch64_build_vector_type (mode, unsigned_p, poly_p); |
b5828b4b | 631 | else |
6db1ec94 JG |
632 | return aarch64_build_scalar_type (mode, unsigned_p, poly_p); |
633 | } | |
634 | ||
635 | tree | |
636 | aarch64_build_signed_type (enum machine_mode mode) | |
637 | { | |
638 | return aarch64_build_type (mode, false, false); | |
639 | } | |
640 | ||
641 | tree | |
642 | aarch64_build_unsigned_type (enum machine_mode mode) | |
643 | { | |
644 | return aarch64_build_type (mode, true, false); | |
645 | } | |
646 | ||
647 | tree | |
648 | aarch64_build_poly_type (enum machine_mode mode) | |
649 | { | |
650 | return aarch64_build_type (mode, false, true); | |
b5828b4b JG |
651 | } |
652 | ||
af55e82d | 653 | static void |
342be7f7 | 654 | aarch64_init_simd_builtins (void) |
43e9d192 | 655 | { |
342be7f7 | 656 | unsigned int i, fcode = AARCH64_SIMD_BUILTIN_BASE + 1; |
43e9d192 | 657 | |
6db1ec94 JG |
658 | /* Signed scalar type nodes. */ |
659 | tree aarch64_simd_intQI_type_node = aarch64_build_signed_type (QImode); | |
660 | tree aarch64_simd_intHI_type_node = aarch64_build_signed_type (HImode); | |
661 | tree aarch64_simd_intSI_type_node = aarch64_build_signed_type (SImode); | |
662 | tree aarch64_simd_intDI_type_node = aarch64_build_signed_type (DImode); | |
663 | tree aarch64_simd_intTI_type_node = aarch64_build_signed_type (TImode); | |
664 | tree aarch64_simd_intEI_type_node = aarch64_build_signed_type (EImode); | |
665 | tree aarch64_simd_intOI_type_node = aarch64_build_signed_type (OImode); | |
666 | tree aarch64_simd_intCI_type_node = aarch64_build_signed_type (CImode); | |
667 | tree aarch64_simd_intXI_type_node = aarch64_build_signed_type (XImode); | |
668 | ||
669 | /* Unsigned scalar type nodes. */ | |
670 | tree aarch64_simd_intUQI_type_node = aarch64_build_unsigned_type (QImode); | |
671 | tree aarch64_simd_intUHI_type_node = aarch64_build_unsigned_type (HImode); | |
672 | tree aarch64_simd_intUSI_type_node = aarch64_build_unsigned_type (SImode); | |
673 | tree aarch64_simd_intUDI_type_node = aarch64_build_unsigned_type (DImode); | |
674 | ||
675 | /* Poly scalar type nodes. */ | |
676 | tree aarch64_simd_polyQI_type_node = aarch64_build_poly_type (QImode); | |
677 | tree aarch64_simd_polyHI_type_node = aarch64_build_poly_type (HImode); | |
7baa225d TB |
678 | tree aarch64_simd_polyDI_type_node = aarch64_build_poly_type (DImode); |
679 | tree aarch64_simd_polyTI_type_node = aarch64_build_poly_type (TImode); | |
b5828b4b JG |
680 | |
681 | /* Float type nodes. */ | |
6db1ec94 JG |
682 | tree aarch64_simd_float_type_node = aarch64_build_signed_type (SFmode); |
683 | tree aarch64_simd_double_type_node = aarch64_build_signed_type (DFmode); | |
43e9d192 IB |
684 | |
685 | /* Define typedefs which exactly correspond to the modes we are basing vector | |
686 | types on. If you change these names you'll need to change | |
687 | the table used by aarch64_mangle_type too. */ | |
688 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intQI_type_node, | |
689 | "__builtin_aarch64_simd_qi"); | |
690 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intHI_type_node, | |
691 | "__builtin_aarch64_simd_hi"); | |
692 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intSI_type_node, | |
693 | "__builtin_aarch64_simd_si"); | |
694 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_float_type_node, | |
695 | "__builtin_aarch64_simd_sf"); | |
696 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intDI_type_node, | |
697 | "__builtin_aarch64_simd_di"); | |
698 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_double_type_node, | |
699 | "__builtin_aarch64_simd_df"); | |
700 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyQI_type_node, | |
701 | "__builtin_aarch64_simd_poly8"); | |
702 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyHI_type_node, | |
703 | "__builtin_aarch64_simd_poly16"); | |
7baa225d TB |
704 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyDI_type_node, |
705 | "__builtin_aarch64_simd_poly64"); | |
706 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyTI_type_node, | |
707 | "__builtin_aarch64_simd_poly128"); | |
b5828b4b | 708 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intTI_type_node, |
43e9d192 | 709 | "__builtin_aarch64_simd_ti"); |
b5828b4b | 710 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intEI_type_node, |
43e9d192 | 711 | "__builtin_aarch64_simd_ei"); |
b5828b4b | 712 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intOI_type_node, |
43e9d192 | 713 | "__builtin_aarch64_simd_oi"); |
b5828b4b | 714 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intCI_type_node, |
43e9d192 | 715 | "__builtin_aarch64_simd_ci"); |
b5828b4b | 716 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intXI_type_node, |
43e9d192 IB |
717 | "__builtin_aarch64_simd_xi"); |
718 | ||
b5828b4b JG |
719 | /* Unsigned integer types for various mode sizes. */ |
720 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUQI_type_node, | |
721 | "__builtin_aarch64_simd_uqi"); | |
722 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUHI_type_node, | |
723 | "__builtin_aarch64_simd_uhi"); | |
724 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUSI_type_node, | |
725 | "__builtin_aarch64_simd_usi"); | |
726 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUDI_type_node, | |
727 | "__builtin_aarch64_simd_udi"); | |
43e9d192 | 728 | |
342be7f7 | 729 | for (i = 0; i < ARRAY_SIZE (aarch64_simd_builtin_data); i++, fcode++) |
43e9d192 | 730 | { |
b5828b4b JG |
731 | bool print_type_signature_p = false; |
732 | char type_signature[SIMD_MAX_BUILTIN_ARGS] = { 0 }; | |
43e9d192 | 733 | aarch64_simd_builtin_datum *d = &aarch64_simd_builtin_data[i]; |
342be7f7 | 734 | const char *const modenames[] = |
b5828b4b | 735 | { |
c6a29a09 | 736 | "v8qi", "v4hi", "v2si", "v2sf", "v1df", "di", "df", |
b5828b4b JG |
737 | "v16qi", "v8hi", "v4si", "v4sf", "v2di", "v2df", |
738 | "ti", "ei", "oi", "xi", "si", "sf", "hi", "qi" | |
739 | }; | |
740 | const enum machine_mode modes[] = | |
741 | { | |
c6a29a09 | 742 | V8QImode, V4HImode, V2SImode, V2SFmode, V1DFmode, DImode, DFmode, |
b5828b4b JG |
743 | V16QImode, V8HImode, V4SImode, V4SFmode, V2DImode, |
744 | V2DFmode, TImode, EImode, OImode, XImode, SImode, | |
745 | SFmode, HImode, QImode | |
746 | }; | |
342be7f7 JG |
747 | char namebuf[60]; |
748 | tree ftype = NULL; | |
119103ca | 749 | tree fndecl = NULL; |
342be7f7 JG |
750 | |
751 | gcc_assert (ARRAY_SIZE (modenames) == T_MAX); | |
43e9d192 | 752 | |
342be7f7 | 753 | d->fcode = fcode; |
43e9d192 | 754 | |
b5828b4b JG |
755 | /* We must track two variables here. op_num is |
756 | the operand number as in the RTL pattern. This is | |
757 | required to access the mode (e.g. V4SF mode) of the | |
758 | argument, from which the base type can be derived. | |
759 | arg_num is an index in to the qualifiers data, which | |
760 | gives qualifiers to the type (e.g. const unsigned). | |
761 | The reason these two variables may differ by one is the | |
762 | void return type. While all return types take the 0th entry | |
763 | in the qualifiers array, there is no operand for them in the | |
764 | RTL pattern. */ | |
765 | int op_num = insn_data[d->code].n_operands - 1; | |
766 | int arg_num = d->qualifiers[0] & qualifier_void | |
767 | ? op_num + 1 | |
768 | : op_num; | |
769 | tree return_type = void_type_node, args = void_list_node; | |
770 | tree eltype; | |
771 | ||
772 | /* Build a function type directly from the insn_data for this | |
773 | builtin. The build_function_type () function takes care of | |
774 | removing duplicates for us. */ | |
775 | for (; op_num >= 0; arg_num--, op_num--) | |
43e9d192 | 776 | { |
b5828b4b JG |
777 | enum machine_mode op_mode = insn_data[d->code].operand[op_num].mode; |
778 | enum aarch64_type_qualifiers qualifiers = d->qualifiers[arg_num]; | |
43e9d192 | 779 | |
b5828b4b JG |
780 | if (qualifiers & qualifier_unsigned) |
781 | { | |
782 | type_signature[arg_num] = 'u'; | |
783 | print_type_signature_p = true; | |
784 | } | |
6db1ec94 JG |
785 | else if (qualifiers & qualifier_poly) |
786 | { | |
787 | type_signature[arg_num] = 'p'; | |
788 | print_type_signature_p = true; | |
789 | } | |
b5828b4b JG |
790 | else |
791 | type_signature[arg_num] = 's'; | |
792 | ||
793 | /* Skip an internal operand for vget_{low, high}. */ | |
794 | if (qualifiers & qualifier_internal) | |
795 | continue; | |
796 | ||
797 | /* Some builtins have different user-facing types | |
798 | for certain arguments, encoded in d->mode. */ | |
799 | if (qualifiers & qualifier_map_mode) | |
800 | op_mode = modes[d->mode]; | |
801 | ||
802 | /* For pointers, we want a pointer to the basic type | |
803 | of the vector. */ | |
804 | if (qualifiers & qualifier_pointer && VECTOR_MODE_P (op_mode)) | |
805 | op_mode = GET_MODE_INNER (op_mode); | |
806 | ||
807 | eltype = aarch64_build_type (op_mode, | |
6db1ec94 JG |
808 | qualifiers & qualifier_unsigned, |
809 | qualifiers & qualifier_poly); | |
b5828b4b JG |
810 | |
811 | /* Add qualifiers. */ | |
812 | if (qualifiers & qualifier_const) | |
813 | eltype = build_qualified_type (eltype, TYPE_QUAL_CONST); | |
814 | ||
815 | if (qualifiers & qualifier_pointer) | |
816 | eltype = build_pointer_type (eltype); | |
817 | ||
818 | /* If we have reached arg_num == 0, we are at a non-void | |
819 | return type. Otherwise, we are still processing | |
820 | arguments. */ | |
821 | if (arg_num == 0) | |
822 | return_type = eltype; | |
823 | else | |
824 | args = tree_cons (NULL_TREE, eltype, args); | |
825 | } | |
342be7f7 | 826 | |
b5828b4b | 827 | ftype = build_function_type (return_type, args); |
43e9d192 | 828 | |
342be7f7 | 829 | gcc_assert (ftype != NULL); |
43e9d192 | 830 | |
b5828b4b JG |
831 | if (print_type_signature_p) |
832 | snprintf (namebuf, sizeof (namebuf), "__builtin_aarch64_%s%s_%s", | |
833 | d->name, modenames[d->mode], type_signature); | |
834 | else | |
835 | snprintf (namebuf, sizeof (namebuf), "__builtin_aarch64_%s%s", | |
836 | d->name, modenames[d->mode]); | |
43e9d192 | 837 | |
119103ca JG |
838 | fndecl = add_builtin_function (namebuf, ftype, fcode, BUILT_IN_MD, |
839 | NULL, NULL_TREE); | |
840 | aarch64_builtin_decls[fcode] = fndecl; | |
43e9d192 IB |
841 | } |
842 | } | |
843 | ||
5d357f26 KT |
844 | static void |
845 | aarch64_init_crc32_builtins () | |
846 | { | |
847 | tree usi_type = aarch64_build_unsigned_type (SImode); | |
848 | unsigned int i = 0; | |
849 | ||
850 | for (i = 0; i < ARRAY_SIZE (aarch64_crc_builtin_data); ++i) | |
851 | { | |
852 | aarch64_crc_builtin_datum* d = &aarch64_crc_builtin_data[i]; | |
853 | tree argtype = aarch64_build_unsigned_type (d->mode); | |
854 | tree ftype = build_function_type_list (usi_type, usi_type, argtype, NULL_TREE); | |
855 | tree fndecl = add_builtin_function (d->name, ftype, d->fcode, | |
856 | BUILT_IN_MD, NULL, NULL_TREE); | |
857 | ||
858 | aarch64_builtin_decls[d->fcode] = fndecl; | |
859 | } | |
860 | } | |
861 | ||
342be7f7 JG |
862 | void |
863 | aarch64_init_builtins (void) | |
43e9d192 | 864 | { |
aa87aced KV |
865 | tree ftype_set_fpr |
866 | = build_function_type_list (void_type_node, unsigned_type_node, NULL); | |
867 | tree ftype_get_fpr | |
868 | = build_function_type_list (unsigned_type_node, NULL); | |
869 | ||
870 | aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPCR] | |
871 | = add_builtin_function ("__builtin_aarch64_get_fpcr", ftype_get_fpr, | |
872 | AARCH64_BUILTIN_GET_FPCR, BUILT_IN_MD, NULL, NULL_TREE); | |
873 | aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPCR] | |
874 | = add_builtin_function ("__builtin_aarch64_set_fpcr", ftype_set_fpr, | |
875 | AARCH64_BUILTIN_SET_FPCR, BUILT_IN_MD, NULL, NULL_TREE); | |
876 | aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPSR] | |
877 | = add_builtin_function ("__builtin_aarch64_get_fpsr", ftype_get_fpr, | |
878 | AARCH64_BUILTIN_GET_FPSR, BUILT_IN_MD, NULL, NULL_TREE); | |
879 | aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPSR] | |
880 | = add_builtin_function ("__builtin_aarch64_set_fpsr", ftype_set_fpr, | |
881 | AARCH64_BUILTIN_SET_FPSR, BUILT_IN_MD, NULL, NULL_TREE); | |
882 | ||
342be7f7 JG |
883 | if (TARGET_SIMD) |
884 | aarch64_init_simd_builtins (); | |
5d357f26 KT |
885 | if (TARGET_CRC32) |
886 | aarch64_init_crc32_builtins (); | |
43e9d192 IB |
887 | } |
888 | ||
119103ca JG |
889 | tree |
890 | aarch64_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
891 | { | |
892 | if (code >= AARCH64_BUILTIN_MAX) | |
893 | return error_mark_node; | |
894 | ||
895 | return aarch64_builtin_decls[code]; | |
896 | } | |
897 | ||
43e9d192 IB |
898 | typedef enum |
899 | { | |
900 | SIMD_ARG_COPY_TO_REG, | |
901 | SIMD_ARG_CONSTANT, | |
902 | SIMD_ARG_STOP | |
903 | } builtin_simd_arg; | |
904 | ||
43e9d192 IB |
905 | static rtx |
906 | aarch64_simd_expand_args (rtx target, int icode, int have_retval, | |
907 | tree exp, ...) | |
908 | { | |
909 | va_list ap; | |
910 | rtx pat; | |
911 | tree arg[SIMD_MAX_BUILTIN_ARGS]; | |
912 | rtx op[SIMD_MAX_BUILTIN_ARGS]; | |
913 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
914 | enum machine_mode mode[SIMD_MAX_BUILTIN_ARGS]; | |
915 | int argc = 0; | |
916 | ||
917 | if (have_retval | |
918 | && (!target | |
919 | || GET_MODE (target) != tmode | |
920 | || !(*insn_data[icode].operand[0].predicate) (target, tmode))) | |
921 | target = gen_reg_rtx (tmode); | |
922 | ||
923 | va_start (ap, exp); | |
924 | ||
925 | for (;;) | |
926 | { | |
927 | builtin_simd_arg thisarg = (builtin_simd_arg) va_arg (ap, int); | |
928 | ||
929 | if (thisarg == SIMD_ARG_STOP) | |
930 | break; | |
931 | else | |
932 | { | |
933 | arg[argc] = CALL_EXPR_ARG (exp, argc); | |
934 | op[argc] = expand_normal (arg[argc]); | |
935 | mode[argc] = insn_data[icode].operand[argc + have_retval].mode; | |
936 | ||
937 | switch (thisarg) | |
938 | { | |
939 | case SIMD_ARG_COPY_TO_REG: | |
2888c331 YZ |
940 | if (POINTER_TYPE_P (TREE_TYPE (arg[argc]))) |
941 | op[argc] = convert_memory_address (Pmode, op[argc]); | |
43e9d192 IB |
942 | /*gcc_assert (GET_MODE (op[argc]) == mode[argc]); */ |
943 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
944 | (op[argc], mode[argc])) | |
945 | op[argc] = copy_to_mode_reg (mode[argc], op[argc]); | |
946 | break; | |
947 | ||
948 | case SIMD_ARG_CONSTANT: | |
949 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
950 | (op[argc], mode[argc])) | |
951 | error_at (EXPR_LOCATION (exp), "incompatible type for argument %d, " | |
952 | "expected %<const int%>", argc + 1); | |
953 | break; | |
954 | ||
955 | case SIMD_ARG_STOP: | |
956 | gcc_unreachable (); | |
957 | } | |
958 | ||
959 | argc++; | |
960 | } | |
961 | } | |
962 | ||
963 | va_end (ap); | |
964 | ||
965 | if (have_retval) | |
966 | switch (argc) | |
967 | { | |
968 | case 1: | |
969 | pat = GEN_FCN (icode) (target, op[0]); | |
970 | break; | |
971 | ||
972 | case 2: | |
973 | pat = GEN_FCN (icode) (target, op[0], op[1]); | |
974 | break; | |
975 | ||
976 | case 3: | |
977 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2]); | |
978 | break; | |
979 | ||
980 | case 4: | |
981 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3]); | |
982 | break; | |
983 | ||
984 | case 5: | |
985 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3], op[4]); | |
986 | break; | |
987 | ||
988 | default: | |
989 | gcc_unreachable (); | |
990 | } | |
991 | else | |
992 | switch (argc) | |
993 | { | |
994 | case 1: | |
995 | pat = GEN_FCN (icode) (op[0]); | |
996 | break; | |
997 | ||
998 | case 2: | |
999 | pat = GEN_FCN (icode) (op[0], op[1]); | |
1000 | break; | |
1001 | ||
1002 | case 3: | |
1003 | pat = GEN_FCN (icode) (op[0], op[1], op[2]); | |
1004 | break; | |
1005 | ||
1006 | case 4: | |
1007 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]); | |
1008 | break; | |
1009 | ||
1010 | case 5: | |
1011 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3], op[4]); | |
1012 | break; | |
1013 | ||
1014 | default: | |
1015 | gcc_unreachable (); | |
1016 | } | |
1017 | ||
1018 | if (!pat) | |
1019 | return 0; | |
1020 | ||
1021 | emit_insn (pat); | |
1022 | ||
1023 | return target; | |
1024 | } | |
1025 | ||
1026 | /* Expand an AArch64 AdvSIMD builtin(intrinsic). */ | |
1027 | rtx | |
1028 | aarch64_simd_expand_builtin (int fcode, tree exp, rtx target) | |
1029 | { | |
342be7f7 JG |
1030 | aarch64_simd_builtin_datum *d = |
1031 | &aarch64_simd_builtin_data[fcode - (AARCH64_SIMD_BUILTIN_BASE + 1)]; | |
342be7f7 | 1032 | enum insn_code icode = d->code; |
b5828b4b JG |
1033 | builtin_simd_arg args[SIMD_MAX_BUILTIN_ARGS]; |
1034 | int num_args = insn_data[d->code].n_operands; | |
1035 | int is_void = 0; | |
1036 | int k; | |
43e9d192 | 1037 | |
b5828b4b | 1038 | is_void = !!(d->qualifiers[0] & qualifier_void); |
43e9d192 | 1039 | |
b5828b4b JG |
1040 | num_args += is_void; |
1041 | ||
1042 | for (k = 1; k < num_args; k++) | |
1043 | { | |
1044 | /* We have four arrays of data, each indexed in a different fashion. | |
1045 | qualifiers - element 0 always describes the function return type. | |
1046 | operands - element 0 is either the operand for return value (if | |
1047 | the function has a non-void return type) or the operand for the | |
1048 | first argument. | |
1049 | expr_args - element 0 always holds the first argument. | |
1050 | args - element 0 is always used for the return type. */ | |
1051 | int qualifiers_k = k; | |
1052 | int operands_k = k - is_void; | |
1053 | int expr_args_k = k - 1; | |
1054 | ||
1055 | if (d->qualifiers[qualifiers_k] & qualifier_immediate) | |
1056 | args[k] = SIMD_ARG_CONSTANT; | |
1057 | else if (d->qualifiers[qualifiers_k] & qualifier_maybe_immediate) | |
1058 | { | |
1059 | rtx arg | |
1060 | = expand_normal (CALL_EXPR_ARG (exp, | |
1061 | (expr_args_k))); | |
1062 | /* Handle constants only if the predicate allows it. */ | |
1063 | bool op_const_int_p = | |
1064 | (CONST_INT_P (arg) | |
1065 | && (*insn_data[icode].operand[operands_k].predicate) | |
1066 | (arg, insn_data[icode].operand[operands_k].mode)); | |
1067 | args[k] = op_const_int_p ? SIMD_ARG_CONSTANT : SIMD_ARG_COPY_TO_REG; | |
1068 | } | |
1069 | else | |
1070 | args[k] = SIMD_ARG_COPY_TO_REG; | |
43e9d192 | 1071 | |
43e9d192 | 1072 | } |
b5828b4b JG |
1073 | args[k] = SIMD_ARG_STOP; |
1074 | ||
1075 | /* The interface to aarch64_simd_expand_args expects a 0 if | |
1076 | the function is void, and a 1 if it is not. */ | |
1077 | return aarch64_simd_expand_args | |
1078 | (target, icode, !is_void, exp, | |
1079 | args[1], | |
1080 | args[2], | |
1081 | args[3], | |
1082 | args[4], | |
1083 | SIMD_ARG_STOP); | |
43e9d192 | 1084 | } |
342be7f7 | 1085 | |
5d357f26 KT |
1086 | rtx |
1087 | aarch64_crc32_expand_builtin (int fcode, tree exp, rtx target) | |
1088 | { | |
1089 | rtx pat; | |
1090 | aarch64_crc_builtin_datum *d | |
1091 | = &aarch64_crc_builtin_data[fcode - (AARCH64_CRC32_BUILTIN_BASE + 1)]; | |
1092 | enum insn_code icode = d->icode; | |
1093 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
1094 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
1095 | rtx op0 = expand_normal (arg0); | |
1096 | rtx op1 = expand_normal (arg1); | |
1097 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
1098 | enum machine_mode mode0 = insn_data[icode].operand[1].mode; | |
1099 | enum machine_mode mode1 = insn_data[icode].operand[2].mode; | |
1100 | ||
1101 | if (! target | |
1102 | || GET_MODE (target) != tmode | |
1103 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
1104 | target = gen_reg_rtx (tmode); | |
1105 | ||
1106 | gcc_assert ((GET_MODE (op0) == mode0 || GET_MODE (op0) == VOIDmode) | |
1107 | && (GET_MODE (op1) == mode1 || GET_MODE (op1) == VOIDmode)); | |
1108 | ||
1109 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
1110 | op0 = copy_to_mode_reg (mode0, op0); | |
1111 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
1112 | op1 = copy_to_mode_reg (mode1, op1); | |
1113 | ||
1114 | pat = GEN_FCN (icode) (target, op0, op1); | |
1115 | if (! pat) | |
1116 | return 0; | |
1117 | emit_insn (pat); | |
1118 | return target; | |
1119 | } | |
1120 | ||
342be7f7 JG |
1121 | /* Expand an expression EXP that calls a built-in function, |
1122 | with result going to TARGET if that's convenient. */ | |
1123 | rtx | |
1124 | aarch64_expand_builtin (tree exp, | |
1125 | rtx target, | |
1126 | rtx subtarget ATTRIBUTE_UNUSED, | |
1127 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
1128 | int ignore ATTRIBUTE_UNUSED) | |
1129 | { | |
1130 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
1131 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
aa87aced KV |
1132 | int icode; |
1133 | rtx pat, op0; | |
1134 | tree arg0; | |
1135 | ||
1136 | switch (fcode) | |
1137 | { | |
1138 | case AARCH64_BUILTIN_GET_FPCR: | |
1139 | case AARCH64_BUILTIN_SET_FPCR: | |
1140 | case AARCH64_BUILTIN_GET_FPSR: | |
1141 | case AARCH64_BUILTIN_SET_FPSR: | |
1142 | if ((fcode == AARCH64_BUILTIN_GET_FPCR) | |
1143 | || (fcode == AARCH64_BUILTIN_GET_FPSR)) | |
1144 | { | |
1145 | icode = (fcode == AARCH64_BUILTIN_GET_FPSR) ? | |
1146 | CODE_FOR_get_fpsr : CODE_FOR_get_fpcr; | |
1147 | target = gen_reg_rtx (SImode); | |
1148 | pat = GEN_FCN (icode) (target); | |
1149 | } | |
1150 | else | |
1151 | { | |
1152 | target = NULL_RTX; | |
1153 | icode = (fcode == AARCH64_BUILTIN_SET_FPSR) ? | |
1154 | CODE_FOR_set_fpsr : CODE_FOR_set_fpcr; | |
1155 | arg0 = CALL_EXPR_ARG (exp, 0); | |
1156 | op0 = expand_normal (arg0); | |
1157 | pat = GEN_FCN (icode) (op0); | |
1158 | } | |
1159 | emit_insn (pat); | |
1160 | return target; | |
1161 | } | |
342be7f7 | 1162 | |
5d357f26 | 1163 | if (fcode >= AARCH64_SIMD_BUILTIN_BASE && fcode <= AARCH64_SIMD_BUILTIN_MAX) |
342be7f7 | 1164 | return aarch64_simd_expand_builtin (fcode, exp, target); |
5d357f26 KT |
1165 | else if (fcode >= AARCH64_CRC32_BUILTIN_BASE && fcode <= AARCH64_CRC32_BUILTIN_MAX) |
1166 | return aarch64_crc32_expand_builtin (fcode, exp, target); | |
342be7f7 JG |
1167 | |
1168 | return NULL_RTX; | |
1169 | } | |
42fc9a7f JG |
1170 | |
1171 | tree | |
1172 | aarch64_builtin_vectorized_function (tree fndecl, tree type_out, tree type_in) | |
1173 | { | |
1174 | enum machine_mode in_mode, out_mode; | |
1175 | int in_n, out_n; | |
1176 | ||
1177 | if (TREE_CODE (type_out) != VECTOR_TYPE | |
1178 | || TREE_CODE (type_in) != VECTOR_TYPE) | |
1179 | return NULL_TREE; | |
1180 | ||
1181 | out_mode = TYPE_MODE (TREE_TYPE (type_out)); | |
1182 | out_n = TYPE_VECTOR_SUBPARTS (type_out); | |
1183 | in_mode = TYPE_MODE (TREE_TYPE (type_in)); | |
1184 | in_n = TYPE_VECTOR_SUBPARTS (type_in); | |
1185 | ||
1186 | #undef AARCH64_CHECK_BUILTIN_MODE | |
1187 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) 1 | |
1188 | #define AARCH64_FIND_FRINT_VARIANT(N) \ | |
1189 | (AARCH64_CHECK_BUILTIN_MODE (2, D) \ | |
e993fea1 | 1190 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v2df] \ |
42fc9a7f | 1191 | : (AARCH64_CHECK_BUILTIN_MODE (4, S) \ |
e993fea1 | 1192 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v4sf] \ |
42fc9a7f | 1193 | : (AARCH64_CHECK_BUILTIN_MODE (2, S) \ |
e993fea1 | 1194 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v2sf] \ |
42fc9a7f JG |
1195 | : NULL_TREE))) |
1196 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
1197 | { | |
1198 | enum built_in_function fn = DECL_FUNCTION_CODE (fndecl); | |
1199 | switch (fn) | |
1200 | { | |
1201 | #undef AARCH64_CHECK_BUILTIN_MODE | |
1202 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ | |
1203 | (out_mode == N##Fmode && out_n == C \ | |
1204 | && in_mode == N##Fmode && in_n == C) | |
1205 | case BUILT_IN_FLOOR: | |
1206 | case BUILT_IN_FLOORF: | |
0659ce6f | 1207 | return AARCH64_FIND_FRINT_VARIANT (floor); |
42fc9a7f JG |
1208 | case BUILT_IN_CEIL: |
1209 | case BUILT_IN_CEILF: | |
0659ce6f | 1210 | return AARCH64_FIND_FRINT_VARIANT (ceil); |
42fc9a7f JG |
1211 | case BUILT_IN_TRUNC: |
1212 | case BUILT_IN_TRUNCF: | |
0659ce6f | 1213 | return AARCH64_FIND_FRINT_VARIANT (btrunc); |
42fc9a7f JG |
1214 | case BUILT_IN_ROUND: |
1215 | case BUILT_IN_ROUNDF: | |
0659ce6f | 1216 | return AARCH64_FIND_FRINT_VARIANT (round); |
42fc9a7f JG |
1217 | case BUILT_IN_NEARBYINT: |
1218 | case BUILT_IN_NEARBYINTF: | |
0659ce6f | 1219 | return AARCH64_FIND_FRINT_VARIANT (nearbyint); |
4dcd1054 JG |
1220 | case BUILT_IN_SQRT: |
1221 | case BUILT_IN_SQRTF: | |
1222 | return AARCH64_FIND_FRINT_VARIANT (sqrt); | |
42fc9a7f | 1223 | #undef AARCH64_CHECK_BUILTIN_MODE |
b5574232 VP |
1224 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ |
1225 | (out_mode == SImode && out_n == C \ | |
1226 | && in_mode == N##Imode && in_n == C) | |
1227 | case BUILT_IN_CLZ: | |
1228 | { | |
1229 | if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
e993fea1 | 1230 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_clzv4si]; |
b5574232 VP |
1231 | return NULL_TREE; |
1232 | } | |
1233 | #undef AARCH64_CHECK_BUILTIN_MODE | |
42fc9a7f JG |
1234 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ |
1235 | (out_mode == N##Imode && out_n == C \ | |
1236 | && in_mode == N##Fmode && in_n == C) | |
1237 | case BUILT_IN_LFLOOR: | |
bf0f324e YZ |
1238 | case BUILT_IN_LFLOORF: |
1239 | case BUILT_IN_LLFLOOR: | |
0386b123 | 1240 | case BUILT_IN_IFLOORF: |
ce966824 | 1241 | { |
e993fea1 | 1242 | enum aarch64_builtins builtin; |
ce966824 | 1243 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) |
e993fea1 | 1244 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv2dfv2di; |
ce966824 | 1245 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) |
e993fea1 | 1246 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv4sfv4si; |
ce966824 | 1247 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) |
e993fea1 JG |
1248 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv2sfv2si; |
1249 | else | |
1250 | return NULL_TREE; | |
1251 | ||
1252 | return aarch64_builtin_decls[builtin]; | |
ce966824 | 1253 | } |
42fc9a7f | 1254 | case BUILT_IN_LCEIL: |
bf0f324e YZ |
1255 | case BUILT_IN_LCEILF: |
1256 | case BUILT_IN_LLCEIL: | |
0386b123 | 1257 | case BUILT_IN_ICEILF: |
ce966824 | 1258 | { |
e993fea1 | 1259 | enum aarch64_builtins builtin; |
ce966824 | 1260 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) |
e993fea1 | 1261 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv2dfv2di; |
ce966824 | 1262 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) |
e993fea1 | 1263 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv4sfv4si; |
ce966824 | 1264 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) |
e993fea1 JG |
1265 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv2sfv2si; |
1266 | else | |
1267 | return NULL_TREE; | |
1268 | ||
1269 | return aarch64_builtin_decls[builtin]; | |
ce966824 | 1270 | } |
0386b123 JG |
1271 | case BUILT_IN_LROUND: |
1272 | case BUILT_IN_IROUNDF: | |
1273 | { | |
e993fea1 | 1274 | enum aarch64_builtins builtin; |
0386b123 | 1275 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) |
e993fea1 | 1276 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv2dfv2di; |
0386b123 | 1277 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) |
e993fea1 | 1278 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv4sfv4si; |
0386b123 | 1279 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) |
e993fea1 JG |
1280 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv2sfv2si; |
1281 | else | |
1282 | return NULL_TREE; | |
1283 | ||
1284 | return aarch64_builtin_decls[builtin]; | |
0386b123 | 1285 | } |
c7f28cd5 KT |
1286 | case BUILT_IN_BSWAP16: |
1287 | #undef AARCH64_CHECK_BUILTIN_MODE | |
1288 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ | |
1289 | (out_mode == N##Imode && out_n == C \ | |
1290 | && in_mode == N##Imode && in_n == C) | |
1291 | if (AARCH64_CHECK_BUILTIN_MODE (4, H)) | |
1292 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv4hi]; | |
1293 | else if (AARCH64_CHECK_BUILTIN_MODE (8, H)) | |
1294 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv8hi]; | |
1295 | else | |
1296 | return NULL_TREE; | |
1297 | case BUILT_IN_BSWAP32: | |
1298 | if (AARCH64_CHECK_BUILTIN_MODE (2, S)) | |
1299 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv2si]; | |
1300 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
1301 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv4si]; | |
1302 | else | |
1303 | return NULL_TREE; | |
1304 | case BUILT_IN_BSWAP64: | |
1305 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) | |
1306 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv2di]; | |
1307 | else | |
1308 | return NULL_TREE; | |
42fc9a7f JG |
1309 | default: |
1310 | return NULL_TREE; | |
1311 | } | |
1312 | } | |
1313 | ||
1314 | return NULL_TREE; | |
1315 | } | |
0ac198d3 JG |
1316 | |
1317 | #undef VAR1 | |
1318 | #define VAR1(T, N, MAP, A) \ | |
e993fea1 | 1319 | case AARCH64_SIMD_BUILTIN_##T##_##N##A: |
0ac198d3 | 1320 | |
9697e620 JG |
1321 | tree |
1322 | aarch64_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED, tree *args, | |
1323 | bool ignore ATTRIBUTE_UNUSED) | |
1324 | { | |
1325 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
1326 | tree type = TREE_TYPE (TREE_TYPE (fndecl)); | |
1327 | ||
1328 | switch (fcode) | |
1329 | { | |
d05d0709 | 1330 | BUILTIN_VALLDI (UNOP, abs, 2) |
9697e620 JG |
1331 | return fold_build1 (ABS_EXPR, type, args[0]); |
1332 | break; | |
bb60efd9 JG |
1333 | BUILTIN_VALLDI (BINOP, cmge, 0) |
1334 | return fold_build2 (GE_EXPR, type, args[0], args[1]); | |
1335 | break; | |
1336 | BUILTIN_VALLDI (BINOP, cmgt, 0) | |
1337 | return fold_build2 (GT_EXPR, type, args[0], args[1]); | |
1338 | break; | |
1339 | BUILTIN_VALLDI (BINOP, cmeq, 0) | |
1340 | return fold_build2 (EQ_EXPR, type, args[0], args[1]); | |
1341 | break; | |
1342 | BUILTIN_VSDQ_I_DI (BINOP, cmtst, 0) | |
1343 | { | |
1344 | tree and_node = fold_build2 (BIT_AND_EXPR, type, args[0], args[1]); | |
1345 | tree vec_zero_node = build_zero_cst (type); | |
1346 | return fold_build2 (NE_EXPR, type, and_node, vec_zero_node); | |
1347 | break; | |
1348 | } | |
c6a29a09 AL |
1349 | VAR1 (REINTERP_SS, reinterpretdi, 0, v1df) |
1350 | VAR1 (REINTERP_SS, reinterpretv8qi, 0, v1df) | |
1351 | VAR1 (REINTERP_SS, reinterpretv4hi, 0, v1df) | |
1352 | VAR1 (REINTERP_SS, reinterpretv2si, 0, v1df) | |
1353 | VAR1 (REINTERP_SS, reinterpretv2sf, 0, v1df) | |
1354 | BUILTIN_VD (REINTERP_SS, reinterpretv1df, 0) | |
1355 | BUILTIN_VD (REINTERP_SU, reinterpretv1df, 0) | |
1356 | VAR1 (REINTERP_US, reinterpretdi, 0, v1df) | |
1357 | VAR1 (REINTERP_US, reinterpretv8qi, 0, v1df) | |
1358 | VAR1 (REINTERP_US, reinterpretv4hi, 0, v1df) | |
1359 | VAR1 (REINTERP_US, reinterpretv2si, 0, v1df) | |
1360 | VAR1 (REINTERP_US, reinterpretv2sf, 0, v1df) | |
1361 | BUILTIN_VD (REINTERP_SP, reinterpretv1df, 0) | |
1362 | VAR1 (REINTERP_PS, reinterpretdi, 0, v1df) | |
1363 | VAR1 (REINTERP_PS, reinterpretv8qi, 0, v1df) | |
1364 | VAR1 (REINTERP_PS, reinterpretv4hi, 0, v1df) | |
1365 | VAR1 (REINTERP_PS, reinterpretv2sf, 0, v1df) | |
bcd48995 | 1366 | return fold_build1 (VIEW_CONVERT_EXPR, type, args[0]); |
1709ff9b JG |
1367 | VAR1 (UNOP, floatv2si, 2, v2sf) |
1368 | VAR1 (UNOP, floatv4si, 2, v4sf) | |
1369 | VAR1 (UNOP, floatv2di, 2, v2df) | |
1370 | return fold_build1 (FLOAT_EXPR, type, args[0]); | |
9697e620 JG |
1371 | default: |
1372 | break; | |
1373 | } | |
1374 | ||
1375 | return NULL_TREE; | |
1376 | } | |
1377 | ||
0ac198d3 JG |
1378 | bool |
1379 | aarch64_gimple_fold_builtin (gimple_stmt_iterator *gsi) | |
1380 | { | |
1381 | bool changed = false; | |
1382 | gimple stmt = gsi_stmt (*gsi); | |
1383 | tree call = gimple_call_fn (stmt); | |
1384 | tree fndecl; | |
1385 | gimple new_stmt = NULL; | |
22756ccf JG |
1386 | |
1387 | /* The operations folded below are reduction operations. These are | |
1388 | defined to leave their result in the 0'th element (from the perspective | |
1389 | of GCC). The architectural instruction we are folding will leave the | |
1390 | result in the 0'th element (from the perspective of the architecture). | |
1391 | For big-endian systems, these perspectives are not aligned. | |
1392 | ||
1393 | It is therefore wrong to perform this fold on big-endian. There | |
1394 | are some tricks we could play with shuffling, but the mid-end is | |
1395 | inconsistent in the way it treats reduction operations, so we will | |
1396 | end up in difficulty. Until we fix the ambiguity - just bail out. */ | |
1397 | if (BYTES_BIG_ENDIAN) | |
1398 | return false; | |
1399 | ||
0ac198d3 JG |
1400 | if (call) |
1401 | { | |
1402 | fndecl = gimple_call_fndecl (stmt); | |
1403 | if (fndecl) | |
1404 | { | |
1405 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
1406 | int nargs = gimple_call_num_args (stmt); | |
1407 | tree *args = (nargs > 0 | |
1408 | ? gimple_call_arg_ptr (stmt, 0) | |
1409 | : &error_mark_node); | |
1410 | ||
1411 | switch (fcode) | |
1412 | { | |
36054fab | 1413 | BUILTIN_VALL (UNOP, reduc_splus_, 10) |
0ac198d3 JG |
1414 | new_stmt = gimple_build_assign_with_ops ( |
1415 | REDUC_PLUS_EXPR, | |
1416 | gimple_call_lhs (stmt), | |
1417 | args[0], | |
1418 | NULL_TREE); | |
1419 | break; | |
1598945b JG |
1420 | BUILTIN_VDQIF (UNOP, reduc_smax_, 10) |
1421 | new_stmt = gimple_build_assign_with_ops ( | |
1422 | REDUC_MAX_EXPR, | |
1423 | gimple_call_lhs (stmt), | |
1424 | args[0], | |
1425 | NULL_TREE); | |
1426 | break; | |
1427 | BUILTIN_VDQIF (UNOP, reduc_smin_, 10) | |
1428 | new_stmt = gimple_build_assign_with_ops ( | |
1429 | REDUC_MIN_EXPR, | |
1430 | gimple_call_lhs (stmt), | |
1431 | args[0], | |
1432 | NULL_TREE); | |
1433 | break; | |
1434 | ||
0ac198d3 JG |
1435 | default: |
1436 | break; | |
1437 | } | |
1438 | } | |
1439 | } | |
1440 | ||
1441 | if (new_stmt) | |
1442 | { | |
1443 | gsi_replace (gsi, new_stmt, true); | |
1444 | changed = true; | |
1445 | } | |
1446 | ||
1447 | return changed; | |
1448 | } | |
1449 | ||
aa87aced KV |
1450 | void |
1451 | aarch64_atomic_assign_expand_fenv (tree *hold, tree *clear, tree *update) | |
1452 | { | |
1453 | const unsigned AARCH64_FE_INVALID = 1; | |
1454 | const unsigned AARCH64_FE_DIVBYZERO = 2; | |
1455 | const unsigned AARCH64_FE_OVERFLOW = 4; | |
1456 | const unsigned AARCH64_FE_UNDERFLOW = 8; | |
1457 | const unsigned AARCH64_FE_INEXACT = 16; | |
1458 | const unsigned HOST_WIDE_INT AARCH64_FE_ALL_EXCEPT = (AARCH64_FE_INVALID | |
1459 | | AARCH64_FE_DIVBYZERO | |
1460 | | AARCH64_FE_OVERFLOW | |
1461 | | AARCH64_FE_UNDERFLOW | |
1462 | | AARCH64_FE_INEXACT); | |
1463 | const unsigned HOST_WIDE_INT AARCH64_FE_EXCEPT_SHIFT = 8; | |
1464 | tree fenv_cr, fenv_sr, get_fpcr, set_fpcr, mask_cr, mask_sr; | |
1465 | tree ld_fenv_cr, ld_fenv_sr, masked_fenv_cr, masked_fenv_sr, hold_fnclex_cr; | |
1466 | tree hold_fnclex_sr, new_fenv_var, reload_fenv, restore_fnenv, get_fpsr, set_fpsr; | |
1467 | tree update_call, atomic_feraiseexcept, hold_fnclex, masked_fenv, ld_fenv; | |
1468 | ||
1469 | /* Generate the equivalence of : | |
1470 | unsigned int fenv_cr; | |
1471 | fenv_cr = __builtin_aarch64_get_fpcr (); | |
1472 | ||
1473 | unsigned int fenv_sr; | |
1474 | fenv_sr = __builtin_aarch64_get_fpsr (); | |
1475 | ||
1476 | Now set all exceptions to non-stop | |
1477 | unsigned int mask_cr | |
1478 | = ~(AARCH64_FE_ALL_EXCEPT << AARCH64_FE_EXCEPT_SHIFT); | |
1479 | unsigned int masked_cr; | |
1480 | masked_cr = fenv_cr & mask_cr; | |
1481 | ||
1482 | And clear all exception flags | |
1483 | unsigned int maske_sr = ~AARCH64_FE_ALL_EXCEPT; | |
1484 | unsigned int masked_cr; | |
1485 | masked_sr = fenv_sr & mask_sr; | |
1486 | ||
1487 | __builtin_aarch64_set_cr (masked_cr); | |
1488 | __builtin_aarch64_set_sr (masked_sr); */ | |
1489 | ||
1490 | fenv_cr = create_tmp_var (unsigned_type_node, NULL); | |
1491 | fenv_sr = create_tmp_var (unsigned_type_node, NULL); | |
1492 | ||
1493 | get_fpcr = aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPCR]; | |
1494 | set_fpcr = aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPCR]; | |
1495 | get_fpsr = aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPSR]; | |
1496 | set_fpsr = aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPSR]; | |
1497 | ||
1498 | mask_cr = build_int_cst (unsigned_type_node, | |
1499 | ~(AARCH64_FE_ALL_EXCEPT << AARCH64_FE_EXCEPT_SHIFT)); | |
1500 | mask_sr = build_int_cst (unsigned_type_node, | |
1501 | ~(AARCH64_FE_ALL_EXCEPT)); | |
1502 | ||
1503 | ld_fenv_cr = build2 (MODIFY_EXPR, unsigned_type_node, | |
1504 | fenv_cr, build_call_expr (get_fpcr, 0)); | |
1505 | ld_fenv_sr = build2 (MODIFY_EXPR, unsigned_type_node, | |
1506 | fenv_sr, build_call_expr (get_fpsr, 0)); | |
1507 | ||
1508 | masked_fenv_cr = build2 (BIT_AND_EXPR, unsigned_type_node, fenv_cr, mask_cr); | |
1509 | masked_fenv_sr = build2 (BIT_AND_EXPR, unsigned_type_node, fenv_sr, mask_sr); | |
1510 | ||
1511 | hold_fnclex_cr = build_call_expr (set_fpcr, 1, masked_fenv_cr); | |
1512 | hold_fnclex_sr = build_call_expr (set_fpsr, 1, masked_fenv_sr); | |
1513 | ||
1514 | hold_fnclex = build2 (COMPOUND_EXPR, void_type_node, hold_fnclex_cr, | |
1515 | hold_fnclex_sr); | |
1516 | masked_fenv = build2 (COMPOUND_EXPR, void_type_node, masked_fenv_cr, | |
1517 | masked_fenv_sr); | |
1518 | ld_fenv = build2 (COMPOUND_EXPR, void_type_node, ld_fenv_cr, ld_fenv_sr); | |
1519 | ||
1520 | *hold = build2 (COMPOUND_EXPR, void_type_node, | |
1521 | build2 (COMPOUND_EXPR, void_type_node, masked_fenv, ld_fenv), | |
1522 | hold_fnclex); | |
1523 | ||
1524 | /* Store the value of masked_fenv to clear the exceptions: | |
1525 | __builtin_aarch64_set_fpsr (masked_fenv_sr); */ | |
1526 | ||
1527 | *clear = build_call_expr (set_fpsr, 1, masked_fenv_sr); | |
1528 | ||
1529 | /* Generate the equivalent of : | |
1530 | unsigned int new_fenv_var; | |
1531 | new_fenv_var = __builtin_aarch64_get_fpsr (); | |
1532 | ||
1533 | __builtin_aarch64_set_fpsr (fenv_sr); | |
1534 | ||
1535 | __atomic_feraiseexcept (new_fenv_var); */ | |
1536 | ||
1537 | new_fenv_var = create_tmp_var (unsigned_type_node, NULL); | |
1538 | reload_fenv = build2 (MODIFY_EXPR, unsigned_type_node, | |
1539 | new_fenv_var, build_call_expr (get_fpsr, 0)); | |
1540 | restore_fnenv = build_call_expr (set_fpsr, 1, fenv_sr); | |
1541 | atomic_feraiseexcept = builtin_decl_implicit (BUILT_IN_ATOMIC_FERAISEEXCEPT); | |
1542 | update_call = build_call_expr (atomic_feraiseexcept, 1, | |
1543 | fold_convert (integer_type_node, new_fenv_var)); | |
1544 | *update = build2 (COMPOUND_EXPR, void_type_node, | |
1545 | build2 (COMPOUND_EXPR, void_type_node, | |
1546 | reload_fenv, restore_fnenv), update_call); | |
1547 | } | |
1548 | ||
1549 | ||
42fc9a7f JG |
1550 | #undef AARCH64_CHECK_BUILTIN_MODE |
1551 | #undef AARCH64_FIND_FRINT_VARIANT | |
0ddec79f JG |
1552 | #undef BUILTIN_DX |
1553 | #undef BUILTIN_SDQ_I | |
1554 | #undef BUILTIN_SD_HSI | |
1555 | #undef BUILTIN_V2F | |
1556 | #undef BUILTIN_VALL | |
1557 | #undef BUILTIN_VB | |
1558 | #undef BUILTIN_VD | |
c6a29a09 | 1559 | #undef BUILTIN_VD1 |
0ddec79f JG |
1560 | #undef BUILTIN_VDC |
1561 | #undef BUILTIN_VDIC | |
1562 | #undef BUILTIN_VDN | |
1563 | #undef BUILTIN_VDQ | |
1564 | #undef BUILTIN_VDQF | |
1565 | #undef BUILTIN_VDQH | |
1566 | #undef BUILTIN_VDQHS | |
1567 | #undef BUILTIN_VDQIF | |
1568 | #undef BUILTIN_VDQM | |
1569 | #undef BUILTIN_VDQV | |
1570 | #undef BUILTIN_VDQ_BHSI | |
1571 | #undef BUILTIN_VDQ_I | |
1572 | #undef BUILTIN_VDW | |
1573 | #undef BUILTIN_VD_BHSI | |
1574 | #undef BUILTIN_VD_HSI | |
0ddec79f JG |
1575 | #undef BUILTIN_VQ |
1576 | #undef BUILTIN_VQN | |
1577 | #undef BUILTIN_VQW | |
1578 | #undef BUILTIN_VQ_HSI | |
1579 | #undef BUILTIN_VQ_S | |
1580 | #undef BUILTIN_VSDQ_HSI | |
1581 | #undef BUILTIN_VSDQ_I | |
1582 | #undef BUILTIN_VSDQ_I_BHSI | |
1583 | #undef BUILTIN_VSDQ_I_DI | |
1584 | #undef BUILTIN_VSD_HSI | |
1585 | #undef BUILTIN_VSQN_HSDI | |
1586 | #undef BUILTIN_VSTRUCT | |
1587 | #undef CF0 | |
1588 | #undef CF1 | |
1589 | #undef CF2 | |
1590 | #undef CF3 | |
1591 | #undef CF4 | |
1592 | #undef CF10 | |
1593 | #undef VAR1 | |
1594 | #undef VAR2 | |
1595 | #undef VAR3 | |
1596 | #undef VAR4 | |
1597 | #undef VAR5 | |
1598 | #undef VAR6 | |
1599 | #undef VAR7 | |
1600 | #undef VAR8 | |
1601 | #undef VAR9 | |
1602 | #undef VAR10 | |
1603 | #undef VAR11 | |
1604 |