]>
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 "hash-table.h" |
37 | #include "vec.h" | |
38 | #include "ggc.h" | |
39 | #include "basic-block.h" | |
40 | #include "tree-ssa-alias.h" | |
41 | #include "internal-fn.h" | |
42 | #include "gimple-fold.h" | |
43 | #include "tree-eh.h" | |
44 | #include "gimple-expr.h" | |
45 | #include "is-a.h" | |
0ac198d3 | 46 | #include "gimple.h" |
5be5c238 | 47 | #include "gimple-iterator.h" |
43e9d192 | 48 | |
bc5e395d JG |
49 | #define v8qi_UP V8QImode |
50 | #define v4hi_UP V4HImode | |
51 | #define v2si_UP V2SImode | |
52 | #define v2sf_UP V2SFmode | |
53 | #define v1df_UP V1DFmode | |
54 | #define di_UP DImode | |
55 | #define df_UP DFmode | |
56 | #define v16qi_UP V16QImode | |
57 | #define v8hi_UP V8HImode | |
58 | #define v4si_UP V4SImode | |
59 | #define v4sf_UP V4SFmode | |
60 | #define v2di_UP V2DImode | |
61 | #define v2df_UP V2DFmode | |
62 | #define ti_UP TImode | |
63 | #define ei_UP EImode | |
64 | #define oi_UP OImode | |
65 | #define ci_UP CImode | |
66 | #define xi_UP XImode | |
67 | #define si_UP SImode | |
68 | #define sf_UP SFmode | |
69 | #define hi_UP HImode | |
70 | #define qi_UP QImode | |
43e9d192 IB |
71 | #define UP(X) X##_UP |
72 | ||
b5828b4b JG |
73 | #define SIMD_MAX_BUILTIN_ARGS 5 |
74 | ||
75 | enum aarch64_type_qualifiers | |
43e9d192 | 76 | { |
b5828b4b JG |
77 | /* T foo. */ |
78 | qualifier_none = 0x0, | |
79 | /* unsigned T foo. */ | |
80 | qualifier_unsigned = 0x1, /* 1 << 0 */ | |
81 | /* const T foo. */ | |
82 | qualifier_const = 0x2, /* 1 << 1 */ | |
83 | /* T *foo. */ | |
84 | qualifier_pointer = 0x4, /* 1 << 2 */ | |
b5828b4b JG |
85 | /* Used when expanding arguments if an operand could |
86 | be an immediate. */ | |
87 | qualifier_immediate = 0x8, /* 1 << 3 */ | |
88 | qualifier_maybe_immediate = 0x10, /* 1 << 4 */ | |
89 | /* void foo (...). */ | |
90 | qualifier_void = 0x20, /* 1 << 5 */ | |
91 | /* Some patterns may have internal operands, this qualifier is an | |
92 | instruction to the initialisation code to skip this operand. */ | |
93 | qualifier_internal = 0x40, /* 1 << 6 */ | |
94 | /* Some builtins should use the T_*mode* encoded in a simd_builtin_datum | |
95 | rather than using the type of the operand. */ | |
96 | qualifier_map_mode = 0x80, /* 1 << 7 */ | |
97 | /* qualifier_pointer | qualifier_map_mode */ | |
98 | qualifier_pointer_map_mode = 0x84, | |
e625e715 | 99 | /* qualifier_const | qualifier_pointer | qualifier_map_mode */ |
6db1ec94 JG |
100 | qualifier_const_pointer_map_mode = 0x86, |
101 | /* Polynomial types. */ | |
102 | qualifier_poly = 0x100 | |
b5828b4b | 103 | }; |
43e9d192 IB |
104 | |
105 | typedef struct | |
106 | { | |
107 | const char *name; | |
bc5e395d | 108 | enum machine_mode mode; |
342be7f7 JG |
109 | const enum insn_code code; |
110 | unsigned int fcode; | |
b5828b4b | 111 | enum aarch64_type_qualifiers *qualifiers; |
43e9d192 IB |
112 | } aarch64_simd_builtin_datum; |
113 | ||
096c59be AL |
114 | /* The qualifier_internal allows generation of a unary builtin from |
115 | a pattern with a third pseudo-operand such as a match_scratch. */ | |
b5828b4b JG |
116 | static enum aarch64_type_qualifiers |
117 | aarch64_types_unop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
096c59be | 118 | = { qualifier_none, qualifier_none, qualifier_internal }; |
b5828b4b | 119 | #define TYPES_UNOP (aarch64_types_unop_qualifiers) |
5a7a4e80 TB |
120 | static enum aarch64_type_qualifiers |
121 | aarch64_types_unopu_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
122 | = { qualifier_unsigned, qualifier_unsigned }; | |
123 | #define TYPES_UNOPU (aarch64_types_unopu_qualifiers) | |
b5828b4b | 124 | #define TYPES_CREATE (aarch64_types_unop_qualifiers) |
bcd48995 AV |
125 | #define TYPES_REINTERP_SS (aarch64_types_unop_qualifiers) |
126 | static enum aarch64_type_qualifiers | |
127 | aarch64_types_unop_su_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
128 | = { qualifier_none, qualifier_unsigned }; | |
129 | #define TYPES_REINTERP_SU (aarch64_types_unop_su_qualifiers) | |
130 | static enum aarch64_type_qualifiers | |
131 | aarch64_types_unop_sp_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
132 | = { qualifier_none, qualifier_poly }; | |
133 | #define TYPES_REINTERP_SP (aarch64_types_unop_sp_qualifiers) | |
134 | static enum aarch64_type_qualifiers | |
135 | aarch64_types_unop_us_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
136 | = { qualifier_unsigned, qualifier_none }; | |
137 | #define TYPES_REINTERP_US (aarch64_types_unop_us_qualifiers) | |
138 | static enum aarch64_type_qualifiers | |
139 | aarch64_types_unop_ps_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
140 | = { qualifier_poly, qualifier_none }; | |
141 | #define TYPES_REINTERP_PS (aarch64_types_unop_ps_qualifiers) | |
b5828b4b JG |
142 | static enum aarch64_type_qualifiers |
143 | aarch64_types_binop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
144 | = { qualifier_none, qualifier_none, qualifier_maybe_immediate }; | |
145 | #define TYPES_BINOP (aarch64_types_binop_qualifiers) | |
146 | static enum aarch64_type_qualifiers | |
ae0533da AL |
147 | aarch64_types_binopv_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
148 | = { qualifier_void, qualifier_none, qualifier_none }; | |
149 | #define TYPES_BINOPV (aarch64_types_binopv_qualifiers) | |
150 | static enum aarch64_type_qualifiers | |
5a7a4e80 TB |
151 | aarch64_types_binopu_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
152 | = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned }; | |
153 | #define TYPES_BINOPU (aarch64_types_binopu_qualifiers) | |
7baa225d | 154 | static enum aarch64_type_qualifiers |
de10bcce AL |
155 | aarch64_types_binop_uus_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
156 | = { qualifier_unsigned, qualifier_unsigned, qualifier_none }; | |
157 | #define TYPES_BINOP_UUS (aarch64_types_binop_uus_qualifiers) | |
158 | static enum aarch64_type_qualifiers | |
918621d3 AL |
159 | aarch64_types_binop_ssu_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
160 | = { qualifier_none, qualifier_none, qualifier_unsigned }; | |
161 | #define TYPES_BINOP_SSU (aarch64_types_binop_ssu_qualifiers) | |
162 | static enum aarch64_type_qualifiers | |
7baa225d TB |
163 | aarch64_types_binopp_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
164 | = { qualifier_poly, qualifier_poly, qualifier_poly }; | |
165 | #define TYPES_BINOPP (aarch64_types_binopp_qualifiers) | |
166 | ||
5a7a4e80 | 167 | static enum aarch64_type_qualifiers |
b5828b4b JG |
168 | aarch64_types_ternop_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
169 | = { qualifier_none, qualifier_none, qualifier_none, qualifier_none }; | |
170 | #define TYPES_TERNOP (aarch64_types_ternop_qualifiers) | |
30442682 TB |
171 | static enum aarch64_type_qualifiers |
172 | aarch64_types_ternopu_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
173 | = { qualifier_unsigned, qualifier_unsigned, | |
174 | qualifier_unsigned, qualifier_unsigned }; | |
175 | #define TYPES_TERNOPU (aarch64_types_ternopu_qualifiers) | |
176 | ||
b5828b4b JG |
177 | static enum aarch64_type_qualifiers |
178 | aarch64_types_quadop_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
179 | = { qualifier_none, qualifier_none, qualifier_none, | |
180 | qualifier_none, qualifier_none }; | |
181 | #define TYPES_QUADOP (aarch64_types_quadop_qualifiers) | |
182 | ||
183 | static enum aarch64_type_qualifiers | |
184 | aarch64_types_getlane_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
185 | = { qualifier_none, qualifier_none, qualifier_immediate }; | |
186 | #define TYPES_GETLANE (aarch64_types_getlane_qualifiers) | |
187 | #define TYPES_SHIFTIMM (aarch64_types_getlane_qualifiers) | |
188 | static enum aarch64_type_qualifiers | |
de10bcce AL |
189 | aarch64_types_shift_to_unsigned_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
190 | = { qualifier_unsigned, qualifier_none, qualifier_immediate }; | |
191 | #define TYPES_SHIFTIMM_USS (aarch64_types_shift_to_unsigned_qualifiers) | |
192 | static enum aarch64_type_qualifiers | |
252c7556 AV |
193 | aarch64_types_unsigned_shift_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
194 | = { qualifier_unsigned, qualifier_unsigned, qualifier_immediate }; | |
195 | #define TYPES_USHIFTIMM (aarch64_types_unsigned_shift_qualifiers) | |
de10bcce | 196 | |
252c7556 | 197 | static enum aarch64_type_qualifiers |
b5828b4b JG |
198 | aarch64_types_setlane_qualifiers[SIMD_MAX_BUILTIN_ARGS] |
199 | = { qualifier_none, qualifier_none, qualifier_none, qualifier_immediate }; | |
200 | #define TYPES_SETLANE (aarch64_types_setlane_qualifiers) | |
201 | #define TYPES_SHIFTINSERT (aarch64_types_setlane_qualifiers) | |
202 | #define TYPES_SHIFTACC (aarch64_types_setlane_qualifiers) | |
203 | ||
de10bcce AL |
204 | static enum aarch64_type_qualifiers |
205 | aarch64_types_unsigned_shiftacc_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
206 | = { qualifier_unsigned, qualifier_unsigned, qualifier_unsigned, | |
207 | qualifier_immediate }; | |
208 | #define TYPES_USHIFTACC (aarch64_types_unsigned_shiftacc_qualifiers) | |
209 | ||
210 | ||
b5828b4b JG |
211 | static enum aarch64_type_qualifiers |
212 | aarch64_types_combine_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
213 | = { qualifier_none, qualifier_none, qualifier_none }; | |
214 | #define TYPES_COMBINE (aarch64_types_combine_qualifiers) | |
215 | ||
216 | static enum aarch64_type_qualifiers | |
217 | aarch64_types_load1_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
218 | = { qualifier_none, qualifier_const_pointer_map_mode }; | |
219 | #define TYPES_LOAD1 (aarch64_types_load1_qualifiers) | |
220 | #define TYPES_LOADSTRUCT (aarch64_types_load1_qualifiers) | |
221 | ||
46e778c4 JG |
222 | static enum aarch64_type_qualifiers |
223 | aarch64_types_bsl_p_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
224 | = { qualifier_poly, qualifier_unsigned, | |
225 | qualifier_poly, qualifier_poly }; | |
226 | #define TYPES_BSL_P (aarch64_types_bsl_p_qualifiers) | |
227 | static enum aarch64_type_qualifiers | |
228 | aarch64_types_bsl_s_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
229 | = { qualifier_none, qualifier_unsigned, | |
230 | qualifier_none, qualifier_none }; | |
231 | #define TYPES_BSL_S (aarch64_types_bsl_s_qualifiers) | |
232 | static enum aarch64_type_qualifiers | |
233 | aarch64_types_bsl_u_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
234 | = { qualifier_unsigned, qualifier_unsigned, | |
235 | qualifier_unsigned, qualifier_unsigned }; | |
236 | #define TYPES_BSL_U (aarch64_types_bsl_u_qualifiers) | |
237 | ||
b5828b4b JG |
238 | /* The first argument (return type) of a store should be void type, |
239 | which we represent with qualifier_void. Their first operand will be | |
240 | a DImode pointer to the location to store to, so we must use | |
241 | qualifier_map_mode | qualifier_pointer to build a pointer to the | |
242 | element type of the vector. */ | |
243 | static enum aarch64_type_qualifiers | |
244 | aarch64_types_store1_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
245 | = { qualifier_void, qualifier_pointer_map_mode, qualifier_none }; | |
246 | #define TYPES_STORE1 (aarch64_types_store1_qualifiers) | |
247 | #define TYPES_STORESTRUCT (aarch64_types_store1_qualifiers) | |
ba081b77 JG |
248 | static enum aarch64_type_qualifiers |
249 | aarch64_types_storestruct_lane_qualifiers[SIMD_MAX_BUILTIN_ARGS] | |
250 | = { qualifier_void, qualifier_pointer_map_mode, | |
251 | qualifier_none, qualifier_none }; | |
252 | #define TYPES_STORESTRUCT_LANE (aarch64_types_storestruct_lane_qualifiers) | |
b5828b4b | 253 | |
0ddec79f JG |
254 | #define CF0(N, X) CODE_FOR_aarch64_##N##X |
255 | #define CF1(N, X) CODE_FOR_##N##X##1 | |
256 | #define CF2(N, X) CODE_FOR_##N##X##2 | |
257 | #define CF3(N, X) CODE_FOR_##N##X##3 | |
258 | #define CF4(N, X) CODE_FOR_##N##X##4 | |
259 | #define CF10(N, X) CODE_FOR_##N##X | |
260 | ||
261 | #define VAR1(T, N, MAP, A) \ | |
bc5e395d | 262 | {#N #A, UP (A), CF##MAP (N, A), 0, TYPES_##T}, |
0ddec79f JG |
263 | #define VAR2(T, N, MAP, A, B) \ |
264 | VAR1 (T, N, MAP, A) \ | |
265 | VAR1 (T, N, MAP, B) | |
266 | #define VAR3(T, N, MAP, A, B, C) \ | |
267 | VAR2 (T, N, MAP, A, B) \ | |
268 | VAR1 (T, N, MAP, C) | |
269 | #define VAR4(T, N, MAP, A, B, C, D) \ | |
270 | VAR3 (T, N, MAP, A, B, C) \ | |
271 | VAR1 (T, N, MAP, D) | |
272 | #define VAR5(T, N, MAP, A, B, C, D, E) \ | |
273 | VAR4 (T, N, MAP, A, B, C, D) \ | |
274 | VAR1 (T, N, MAP, E) | |
275 | #define VAR6(T, N, MAP, A, B, C, D, E, F) \ | |
276 | VAR5 (T, N, MAP, A, B, C, D, E) \ | |
277 | VAR1 (T, N, MAP, F) | |
278 | #define VAR7(T, N, MAP, A, B, C, D, E, F, G) \ | |
279 | VAR6 (T, N, MAP, A, B, C, D, E, F) \ | |
280 | VAR1 (T, N, MAP, G) | |
281 | #define VAR8(T, N, MAP, A, B, C, D, E, F, G, H) \ | |
282 | VAR7 (T, N, MAP, A, B, C, D, E, F, G) \ | |
283 | VAR1 (T, N, MAP, H) | |
284 | #define VAR9(T, N, MAP, A, B, C, D, E, F, G, H, I) \ | |
285 | VAR8 (T, N, MAP, A, B, C, D, E, F, G, H) \ | |
286 | VAR1 (T, N, MAP, I) | |
287 | #define VAR10(T, N, MAP, A, B, C, D, E, F, G, H, I, J) \ | |
288 | VAR9 (T, N, MAP, A, B, C, D, E, F, G, H, I) \ | |
289 | VAR1 (T, N, MAP, J) | |
290 | #define VAR11(T, N, MAP, A, B, C, D, E, F, G, H, I, J, K) \ | |
291 | VAR10 (T, N, MAP, A, B, C, D, E, F, G, H, I, J) \ | |
292 | VAR1 (T, N, MAP, K) | |
293 | #define VAR12(T, N, MAP, A, B, C, D, E, F, G, H, I, J, K, L) \ | |
294 | VAR11 (T, N, MAP, A, B, C, D, E, F, G, H, I, J, K) \ | |
295 | VAR1 (T, N, MAP, L) | |
342be7f7 JG |
296 | |
297 | /* BUILTIN_<ITERATOR> macros should expand to cover the same range of | |
298 | modes as is given for each define_mode_iterator in | |
299 | config/aarch64/iterators.md. */ | |
300 | ||
0ddec79f JG |
301 | #define BUILTIN_DX(T, N, MAP) \ |
302 | VAR2 (T, N, MAP, di, df) | |
303 | #define BUILTIN_GPF(T, N, MAP) \ | |
304 | VAR2 (T, N, MAP, sf, df) | |
305 | #define BUILTIN_SDQ_I(T, N, MAP) \ | |
306 | VAR4 (T, N, MAP, qi, hi, si, di) | |
307 | #define BUILTIN_SD_HSI(T, N, MAP) \ | |
308 | VAR2 (T, N, MAP, hi, si) | |
309 | #define BUILTIN_V2F(T, N, MAP) \ | |
310 | VAR2 (T, N, MAP, v2sf, v2df) | |
311 | #define BUILTIN_VALL(T, N, MAP) \ | |
312 | VAR10 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
313 | v4si, v2di, v2sf, v4sf, v2df) | |
bb60efd9 JG |
314 | #define BUILTIN_VALLDI(T, N, MAP) \ |
315 | VAR11 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
316 | v4si, v2di, v2sf, v4sf, v2df, di) | |
46e778c4 JG |
317 | #define BUILTIN_VALLDIF(T, N, MAP) \ |
318 | VAR12 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, \ | |
319 | v4si, v2di, v2sf, v4sf, v2df, di, df) | |
0ddec79f JG |
320 | #define BUILTIN_VB(T, N, MAP) \ |
321 | VAR2 (T, N, MAP, v8qi, v16qi) | |
322 | #define BUILTIN_VD(T, N, MAP) \ | |
323 | VAR4 (T, N, MAP, v8qi, v4hi, v2si, v2sf) | |
c6a29a09 AL |
324 | #define BUILTIN_VD1(T, N, MAP) \ |
325 | VAR5 (T, N, MAP, v8qi, v4hi, v2si, v2sf, v1df) | |
0ddec79f JG |
326 | #define BUILTIN_VDC(T, N, MAP) \ |
327 | VAR6 (T, N, MAP, v8qi, v4hi, v2si, v2sf, di, df) | |
328 | #define BUILTIN_VDIC(T, N, MAP) \ | |
329 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
330 | #define BUILTIN_VDN(T, N, MAP) \ | |
331 | VAR3 (T, N, MAP, v4hi, v2si, di) | |
332 | #define BUILTIN_VDQ(T, N, MAP) \ | |
333 | VAR7 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di) | |
334 | #define BUILTIN_VDQF(T, N, MAP) \ | |
335 | VAR3 (T, N, MAP, v2sf, v4sf, v2df) | |
74dc11ed AV |
336 | #define BUILTIN_VDQF_DF(T, N, MAP) \ |
337 | VAR4 (T, N, MAP, v2sf, v4sf, v2df, df) | |
0ddec79f JG |
338 | #define BUILTIN_VDQH(T, N, MAP) \ |
339 | VAR2 (T, N, MAP, v4hi, v8hi) | |
340 | #define BUILTIN_VDQHS(T, N, MAP) \ | |
341 | VAR4 (T, N, MAP, v4hi, v8hi, v2si, v4si) | |
342 | #define BUILTIN_VDQIF(T, N, MAP) \ | |
343 | VAR9 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2sf, v4sf, v2df) | |
344 | #define BUILTIN_VDQM(T, N, MAP) \ | |
345 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
346 | #define BUILTIN_VDQV(T, N, MAP) \ | |
347 | VAR5 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v4si) | |
46e778c4 JG |
348 | #define BUILTIN_VDQQH(T, N, MAP) \ |
349 | VAR4 (T, N, MAP, v8qi, v16qi, v4hi, v8hi) | |
0ddec79f JG |
350 | #define BUILTIN_VDQ_BHSI(T, N, MAP) \ |
351 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
352 | #define BUILTIN_VDQ_I(T, N, MAP) \ | |
353 | VAR7 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di) | |
354 | #define BUILTIN_VDW(T, N, MAP) \ | |
355 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
356 | #define BUILTIN_VD_BHSI(T, N, MAP) \ | |
357 | VAR3 (T, N, MAP, v8qi, v4hi, v2si) | |
358 | #define BUILTIN_VD_HSI(T, N, MAP) \ | |
359 | VAR2 (T, N, MAP, v4hi, v2si) | |
0ddec79f JG |
360 | #define BUILTIN_VQ(T, N, MAP) \ |
361 | VAR6 (T, N, MAP, v16qi, v8hi, v4si, v2di, v4sf, v2df) | |
362 | #define BUILTIN_VQN(T, N, MAP) \ | |
363 | VAR3 (T, N, MAP, v8hi, v4si, v2di) | |
364 | #define BUILTIN_VQW(T, N, MAP) \ | |
365 | VAR3 (T, N, MAP, v16qi, v8hi, v4si) | |
366 | #define BUILTIN_VQ_HSI(T, N, MAP) \ | |
367 | VAR2 (T, N, MAP, v8hi, v4si) | |
368 | #define BUILTIN_VQ_S(T, N, MAP) \ | |
369 | VAR6 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si) | |
370 | #define BUILTIN_VSDQ_HSI(T, N, MAP) \ | |
371 | VAR6 (T, N, MAP, v4hi, v8hi, v2si, v4si, hi, si) | |
372 | #define BUILTIN_VSDQ_I(T, N, MAP) \ | |
373 | VAR11 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, qi, hi, si, di) | |
374 | #define BUILTIN_VSDQ_I_BHSI(T, N, MAP) \ | |
375 | VAR10 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, qi, hi, si) | |
376 | #define BUILTIN_VSDQ_I_DI(T, N, MAP) \ | |
377 | VAR8 (T, N, MAP, v8qi, v16qi, v4hi, v8hi, v2si, v4si, v2di, di) | |
378 | #define BUILTIN_VSD_HSI(T, N, MAP) \ | |
379 | VAR4 (T, N, MAP, v4hi, v2si, hi, si) | |
380 | #define BUILTIN_VSQN_HSDI(T, N, MAP) \ | |
381 | VAR6 (T, N, MAP, v8hi, v4si, v2di, hi, si, di) | |
382 | #define BUILTIN_VSTRUCT(T, N, MAP) \ | |
383 | VAR3 (T, N, MAP, oi, ci, xi) | |
43e9d192 IB |
384 | |
385 | static aarch64_simd_builtin_datum aarch64_simd_builtin_data[] = { | |
342be7f7 JG |
386 | #include "aarch64-simd-builtins.def" |
387 | }; | |
388 | ||
5d357f26 KT |
389 | /* There's only 8 CRC32 builtins. Probably not worth their own .def file. */ |
390 | #define AARCH64_CRC32_BUILTINS \ | |
391 | CRC32_BUILTIN (crc32b, QI) \ | |
392 | CRC32_BUILTIN (crc32h, HI) \ | |
393 | CRC32_BUILTIN (crc32w, SI) \ | |
394 | CRC32_BUILTIN (crc32x, DI) \ | |
395 | CRC32_BUILTIN (crc32cb, QI) \ | |
396 | CRC32_BUILTIN (crc32ch, HI) \ | |
397 | CRC32_BUILTIN (crc32cw, SI) \ | |
398 | CRC32_BUILTIN (crc32cx, DI) | |
399 | ||
400 | typedef struct | |
401 | { | |
402 | const char *name; | |
403 | enum machine_mode mode; | |
404 | const enum insn_code icode; | |
405 | unsigned int fcode; | |
406 | } aarch64_crc_builtin_datum; | |
407 | ||
408 | #define CRC32_BUILTIN(N, M) \ | |
409 | AARCH64_BUILTIN_##N, | |
410 | ||
342be7f7 | 411 | #undef VAR1 |
0ddec79f | 412 | #define VAR1(T, N, MAP, A) \ |
e993fea1 | 413 | AARCH64_SIMD_BUILTIN_##T##_##N##A, |
342be7f7 JG |
414 | |
415 | enum aarch64_builtins | |
416 | { | |
417 | AARCH64_BUILTIN_MIN, | |
aa87aced KV |
418 | |
419 | AARCH64_BUILTIN_GET_FPCR, | |
420 | AARCH64_BUILTIN_SET_FPCR, | |
421 | AARCH64_BUILTIN_GET_FPSR, | |
422 | AARCH64_BUILTIN_SET_FPSR, | |
423 | ||
342be7f7 JG |
424 | AARCH64_SIMD_BUILTIN_BASE, |
425 | #include "aarch64-simd-builtins.def" | |
426 | AARCH64_SIMD_BUILTIN_MAX = AARCH64_SIMD_BUILTIN_BASE | |
427 | + ARRAY_SIZE (aarch64_simd_builtin_data), | |
5d357f26 KT |
428 | AARCH64_CRC32_BUILTIN_BASE, |
429 | AARCH64_CRC32_BUILTINS | |
430 | AARCH64_CRC32_BUILTIN_MAX, | |
342be7f7 | 431 | AARCH64_BUILTIN_MAX |
43e9d192 IB |
432 | }; |
433 | ||
5d357f26 KT |
434 | #undef CRC32_BUILTIN |
435 | #define CRC32_BUILTIN(N, M) \ | |
436 | {"__builtin_aarch64_"#N, M##mode, CODE_FOR_aarch64_##N, AARCH64_BUILTIN_##N}, | |
437 | ||
438 | static aarch64_crc_builtin_datum aarch64_crc_builtin_data[] = { | |
439 | AARCH64_CRC32_BUILTINS | |
440 | }; | |
441 | ||
442 | #undef CRC32_BUILTIN | |
443 | ||
119103ca JG |
444 | static GTY(()) tree aarch64_builtin_decls[AARCH64_BUILTIN_MAX]; |
445 | ||
43e9d192 IB |
446 | #define NUM_DREG_TYPES 6 |
447 | #define NUM_QREG_TYPES 6 | |
448 | ||
b5828b4b JG |
449 | /* Return a tree for a signed or unsigned argument of either |
450 | the mode specified by MODE, or the inner mode of MODE. */ | |
451 | tree | |
6db1ec94 JG |
452 | aarch64_build_scalar_type (enum machine_mode mode, |
453 | bool unsigned_p, | |
454 | bool poly_p) | |
b5828b4b JG |
455 | { |
456 | #undef INT_TYPES | |
457 | #define INT_TYPES \ | |
458 | AARCH64_TYPE_BUILDER (QI) \ | |
459 | AARCH64_TYPE_BUILDER (HI) \ | |
460 | AARCH64_TYPE_BUILDER (SI) \ | |
461 | AARCH64_TYPE_BUILDER (DI) \ | |
462 | AARCH64_TYPE_BUILDER (EI) \ | |
463 | AARCH64_TYPE_BUILDER (OI) \ | |
464 | AARCH64_TYPE_BUILDER (CI) \ | |
465 | AARCH64_TYPE_BUILDER (XI) \ | |
466 | AARCH64_TYPE_BUILDER (TI) \ | |
467 | ||
468 | /* Statically declare all the possible types we might need. */ | |
469 | #undef AARCH64_TYPE_BUILDER | |
470 | #define AARCH64_TYPE_BUILDER(X) \ | |
6db1ec94 | 471 | static tree X##_aarch64_type_node_p = NULL; \ |
b5828b4b JG |
472 | static tree X##_aarch64_type_node_s = NULL; \ |
473 | static tree X##_aarch64_type_node_u = NULL; | |
474 | ||
475 | INT_TYPES | |
476 | ||
477 | static tree float_aarch64_type_node = NULL; | |
478 | static tree double_aarch64_type_node = NULL; | |
479 | ||
480 | gcc_assert (!VECTOR_MODE_P (mode)); | |
481 | ||
482 | /* If we've already initialised this type, don't initialise it again, | |
483 | otherwise ask for a new type of the correct size. */ | |
484 | #undef AARCH64_TYPE_BUILDER | |
485 | #define AARCH64_TYPE_BUILDER(X) \ | |
486 | case X##mode: \ | |
487 | if (unsigned_p) \ | |
488 | return (X##_aarch64_type_node_u \ | |
489 | ? X##_aarch64_type_node_u \ | |
490 | : X##_aarch64_type_node_u \ | |
491 | = make_unsigned_type (GET_MODE_PRECISION (mode))); \ | |
6db1ec94 JG |
492 | else if (poly_p) \ |
493 | return (X##_aarch64_type_node_p \ | |
494 | ? X##_aarch64_type_node_p \ | |
495 | : X##_aarch64_type_node_p \ | |
496 | = make_unsigned_type (GET_MODE_PRECISION (mode))); \ | |
b5828b4b JG |
497 | else \ |
498 | return (X##_aarch64_type_node_s \ | |
499 | ? X##_aarch64_type_node_s \ | |
500 | : X##_aarch64_type_node_s \ | |
501 | = make_signed_type (GET_MODE_PRECISION (mode))); \ | |
502 | break; | |
503 | ||
504 | switch (mode) | |
505 | { | |
506 | INT_TYPES | |
507 | case SFmode: | |
508 | if (!float_aarch64_type_node) | |
509 | { | |
510 | float_aarch64_type_node = make_node (REAL_TYPE); | |
511 | TYPE_PRECISION (float_aarch64_type_node) = FLOAT_TYPE_SIZE; | |
512 | layout_type (float_aarch64_type_node); | |
513 | } | |
514 | return float_aarch64_type_node; | |
515 | break; | |
516 | case DFmode: | |
517 | if (!double_aarch64_type_node) | |
518 | { | |
519 | double_aarch64_type_node = make_node (REAL_TYPE); | |
520 | TYPE_PRECISION (double_aarch64_type_node) = DOUBLE_TYPE_SIZE; | |
521 | layout_type (double_aarch64_type_node); | |
522 | } | |
523 | return double_aarch64_type_node; | |
524 | break; | |
525 | default: | |
526 | gcc_unreachable (); | |
527 | } | |
528 | } | |
529 | ||
530 | tree | |
6db1ec94 JG |
531 | aarch64_build_vector_type (enum machine_mode mode, |
532 | bool unsigned_p, | |
533 | bool poly_p) | |
b5828b4b JG |
534 | { |
535 | tree eltype; | |
536 | ||
537 | #define VECTOR_TYPES \ | |
538 | AARCH64_TYPE_BUILDER (V16QI) \ | |
539 | AARCH64_TYPE_BUILDER (V8HI) \ | |
540 | AARCH64_TYPE_BUILDER (V4SI) \ | |
541 | AARCH64_TYPE_BUILDER (V2DI) \ | |
542 | AARCH64_TYPE_BUILDER (V8QI) \ | |
543 | AARCH64_TYPE_BUILDER (V4HI) \ | |
544 | AARCH64_TYPE_BUILDER (V2SI) \ | |
545 | \ | |
546 | AARCH64_TYPE_BUILDER (V4SF) \ | |
547 | AARCH64_TYPE_BUILDER (V2DF) \ | |
548 | AARCH64_TYPE_BUILDER (V2SF) \ | |
549 | /* Declare our "cache" of values. */ | |
550 | #undef AARCH64_TYPE_BUILDER | |
551 | #define AARCH64_TYPE_BUILDER(X) \ | |
552 | static tree X##_aarch64_type_node_s = NULL; \ | |
6db1ec94 JG |
553 | static tree X##_aarch64_type_node_u = NULL; \ |
554 | static tree X##_aarch64_type_node_p = NULL; | |
b5828b4b JG |
555 | |
556 | VECTOR_TYPES | |
557 | ||
558 | gcc_assert (VECTOR_MODE_P (mode)); | |
559 | ||
560 | #undef AARCH64_TYPE_BUILDER | |
561 | #define AARCH64_TYPE_BUILDER(X) \ | |
562 | case X##mode: \ | |
563 | if (unsigned_p) \ | |
564 | return X##_aarch64_type_node_u \ | |
565 | ? X##_aarch64_type_node_u \ | |
566 | : X##_aarch64_type_node_u \ | |
567 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
568 | (GET_MODE_INNER (mode), \ | |
6db1ec94 JG |
569 | unsigned_p, poly_p), mode); \ |
570 | else if (poly_p) \ | |
571 | return X##_aarch64_type_node_p \ | |
572 | ? X##_aarch64_type_node_p \ | |
573 | : X##_aarch64_type_node_p \ | |
574 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
575 | (GET_MODE_INNER (mode), \ | |
576 | unsigned_p, poly_p), mode); \ | |
b5828b4b JG |
577 | else \ |
578 | return X##_aarch64_type_node_s \ | |
579 | ? X##_aarch64_type_node_s \ | |
580 | : X##_aarch64_type_node_s \ | |
581 | = build_vector_type_for_mode (aarch64_build_scalar_type \ | |
582 | (GET_MODE_INNER (mode), \ | |
6db1ec94 | 583 | unsigned_p, poly_p), mode); \ |
b5828b4b JG |
584 | break; |
585 | ||
586 | switch (mode) | |
587 | { | |
588 | default: | |
6db1ec94 JG |
589 | eltype = aarch64_build_scalar_type (GET_MODE_INNER (mode), |
590 | unsigned_p, poly_p); | |
b5828b4b JG |
591 | return build_vector_type_for_mode (eltype, mode); |
592 | break; | |
593 | VECTOR_TYPES | |
594 | } | |
595 | } | |
596 | ||
597 | tree | |
6db1ec94 | 598 | aarch64_build_type (enum machine_mode mode, bool unsigned_p, bool poly_p) |
b5828b4b JG |
599 | { |
600 | if (VECTOR_MODE_P (mode)) | |
6db1ec94 | 601 | return aarch64_build_vector_type (mode, unsigned_p, poly_p); |
b5828b4b | 602 | else |
6db1ec94 JG |
603 | return aarch64_build_scalar_type (mode, unsigned_p, poly_p); |
604 | } | |
605 | ||
606 | tree | |
607 | aarch64_build_signed_type (enum machine_mode mode) | |
608 | { | |
609 | return aarch64_build_type (mode, false, false); | |
610 | } | |
611 | ||
612 | tree | |
613 | aarch64_build_unsigned_type (enum machine_mode mode) | |
614 | { | |
615 | return aarch64_build_type (mode, true, false); | |
616 | } | |
617 | ||
618 | tree | |
619 | aarch64_build_poly_type (enum machine_mode mode) | |
620 | { | |
621 | return aarch64_build_type (mode, false, true); | |
b5828b4b JG |
622 | } |
623 | ||
af55e82d | 624 | static void |
342be7f7 | 625 | aarch64_init_simd_builtins (void) |
43e9d192 | 626 | { |
342be7f7 | 627 | unsigned int i, fcode = AARCH64_SIMD_BUILTIN_BASE + 1; |
43e9d192 | 628 | |
6db1ec94 JG |
629 | /* Signed scalar type nodes. */ |
630 | tree aarch64_simd_intQI_type_node = aarch64_build_signed_type (QImode); | |
631 | tree aarch64_simd_intHI_type_node = aarch64_build_signed_type (HImode); | |
632 | tree aarch64_simd_intSI_type_node = aarch64_build_signed_type (SImode); | |
633 | tree aarch64_simd_intDI_type_node = aarch64_build_signed_type (DImode); | |
634 | tree aarch64_simd_intTI_type_node = aarch64_build_signed_type (TImode); | |
635 | tree aarch64_simd_intEI_type_node = aarch64_build_signed_type (EImode); | |
636 | tree aarch64_simd_intOI_type_node = aarch64_build_signed_type (OImode); | |
637 | tree aarch64_simd_intCI_type_node = aarch64_build_signed_type (CImode); | |
638 | tree aarch64_simd_intXI_type_node = aarch64_build_signed_type (XImode); | |
639 | ||
640 | /* Unsigned scalar type nodes. */ | |
641 | tree aarch64_simd_intUQI_type_node = aarch64_build_unsigned_type (QImode); | |
642 | tree aarch64_simd_intUHI_type_node = aarch64_build_unsigned_type (HImode); | |
643 | tree aarch64_simd_intUSI_type_node = aarch64_build_unsigned_type (SImode); | |
644 | tree aarch64_simd_intUDI_type_node = aarch64_build_unsigned_type (DImode); | |
645 | ||
646 | /* Poly scalar type nodes. */ | |
647 | tree aarch64_simd_polyQI_type_node = aarch64_build_poly_type (QImode); | |
648 | tree aarch64_simd_polyHI_type_node = aarch64_build_poly_type (HImode); | |
7baa225d TB |
649 | tree aarch64_simd_polyDI_type_node = aarch64_build_poly_type (DImode); |
650 | tree aarch64_simd_polyTI_type_node = aarch64_build_poly_type (TImode); | |
b5828b4b JG |
651 | |
652 | /* Float type nodes. */ | |
6db1ec94 JG |
653 | tree aarch64_simd_float_type_node = aarch64_build_signed_type (SFmode); |
654 | tree aarch64_simd_double_type_node = aarch64_build_signed_type (DFmode); | |
43e9d192 IB |
655 | |
656 | /* Define typedefs which exactly correspond to the modes we are basing vector | |
657 | types on. If you change these names you'll need to change | |
658 | the table used by aarch64_mangle_type too. */ | |
659 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intQI_type_node, | |
660 | "__builtin_aarch64_simd_qi"); | |
661 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intHI_type_node, | |
662 | "__builtin_aarch64_simd_hi"); | |
663 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intSI_type_node, | |
664 | "__builtin_aarch64_simd_si"); | |
665 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_float_type_node, | |
666 | "__builtin_aarch64_simd_sf"); | |
667 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intDI_type_node, | |
668 | "__builtin_aarch64_simd_di"); | |
669 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_double_type_node, | |
670 | "__builtin_aarch64_simd_df"); | |
671 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyQI_type_node, | |
672 | "__builtin_aarch64_simd_poly8"); | |
673 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyHI_type_node, | |
674 | "__builtin_aarch64_simd_poly16"); | |
7baa225d TB |
675 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyDI_type_node, |
676 | "__builtin_aarch64_simd_poly64"); | |
677 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_polyTI_type_node, | |
678 | "__builtin_aarch64_simd_poly128"); | |
b5828b4b | 679 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intTI_type_node, |
43e9d192 | 680 | "__builtin_aarch64_simd_ti"); |
b5828b4b | 681 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intEI_type_node, |
43e9d192 | 682 | "__builtin_aarch64_simd_ei"); |
b5828b4b | 683 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intOI_type_node, |
43e9d192 | 684 | "__builtin_aarch64_simd_oi"); |
b5828b4b | 685 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intCI_type_node, |
43e9d192 | 686 | "__builtin_aarch64_simd_ci"); |
b5828b4b | 687 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intXI_type_node, |
43e9d192 IB |
688 | "__builtin_aarch64_simd_xi"); |
689 | ||
b5828b4b JG |
690 | /* Unsigned integer types for various mode sizes. */ |
691 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUQI_type_node, | |
692 | "__builtin_aarch64_simd_uqi"); | |
693 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUHI_type_node, | |
694 | "__builtin_aarch64_simd_uhi"); | |
695 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUSI_type_node, | |
696 | "__builtin_aarch64_simd_usi"); | |
697 | (*lang_hooks.types.register_builtin_type) (aarch64_simd_intUDI_type_node, | |
698 | "__builtin_aarch64_simd_udi"); | |
43e9d192 | 699 | |
342be7f7 | 700 | for (i = 0; i < ARRAY_SIZE (aarch64_simd_builtin_data); i++, fcode++) |
43e9d192 | 701 | { |
b5828b4b JG |
702 | bool print_type_signature_p = false; |
703 | char type_signature[SIMD_MAX_BUILTIN_ARGS] = { 0 }; | |
43e9d192 | 704 | aarch64_simd_builtin_datum *d = &aarch64_simd_builtin_data[i]; |
342be7f7 JG |
705 | char namebuf[60]; |
706 | tree ftype = NULL; | |
119103ca | 707 | tree fndecl = NULL; |
342be7f7 | 708 | |
342be7f7 | 709 | d->fcode = fcode; |
43e9d192 | 710 | |
b5828b4b JG |
711 | /* We must track two variables here. op_num is |
712 | the operand number as in the RTL pattern. This is | |
713 | required to access the mode (e.g. V4SF mode) of the | |
714 | argument, from which the base type can be derived. | |
715 | arg_num is an index in to the qualifiers data, which | |
716 | gives qualifiers to the type (e.g. const unsigned). | |
717 | The reason these two variables may differ by one is the | |
718 | void return type. While all return types take the 0th entry | |
719 | in the qualifiers array, there is no operand for them in the | |
720 | RTL pattern. */ | |
721 | int op_num = insn_data[d->code].n_operands - 1; | |
722 | int arg_num = d->qualifiers[0] & qualifier_void | |
723 | ? op_num + 1 | |
724 | : op_num; | |
725 | tree return_type = void_type_node, args = void_list_node; | |
726 | tree eltype; | |
727 | ||
728 | /* Build a function type directly from the insn_data for this | |
729 | builtin. The build_function_type () function takes care of | |
730 | removing duplicates for us. */ | |
731 | for (; op_num >= 0; arg_num--, op_num--) | |
43e9d192 | 732 | { |
b5828b4b JG |
733 | enum machine_mode op_mode = insn_data[d->code].operand[op_num].mode; |
734 | enum aarch64_type_qualifiers qualifiers = d->qualifiers[arg_num]; | |
43e9d192 | 735 | |
b5828b4b JG |
736 | if (qualifiers & qualifier_unsigned) |
737 | { | |
738 | type_signature[arg_num] = 'u'; | |
739 | print_type_signature_p = true; | |
740 | } | |
6db1ec94 JG |
741 | else if (qualifiers & qualifier_poly) |
742 | { | |
743 | type_signature[arg_num] = 'p'; | |
744 | print_type_signature_p = true; | |
745 | } | |
b5828b4b JG |
746 | else |
747 | type_signature[arg_num] = 's'; | |
748 | ||
749 | /* Skip an internal operand for vget_{low, high}. */ | |
750 | if (qualifiers & qualifier_internal) | |
751 | continue; | |
752 | ||
753 | /* Some builtins have different user-facing types | |
754 | for certain arguments, encoded in d->mode. */ | |
755 | if (qualifiers & qualifier_map_mode) | |
bc5e395d | 756 | op_mode = d->mode; |
b5828b4b JG |
757 | |
758 | /* For pointers, we want a pointer to the basic type | |
759 | of the vector. */ | |
760 | if (qualifiers & qualifier_pointer && VECTOR_MODE_P (op_mode)) | |
761 | op_mode = GET_MODE_INNER (op_mode); | |
762 | ||
763 | eltype = aarch64_build_type (op_mode, | |
6db1ec94 JG |
764 | qualifiers & qualifier_unsigned, |
765 | qualifiers & qualifier_poly); | |
b5828b4b JG |
766 | |
767 | /* Add qualifiers. */ | |
768 | if (qualifiers & qualifier_const) | |
769 | eltype = build_qualified_type (eltype, TYPE_QUAL_CONST); | |
770 | ||
771 | if (qualifiers & qualifier_pointer) | |
772 | eltype = build_pointer_type (eltype); | |
773 | ||
774 | /* If we have reached arg_num == 0, we are at a non-void | |
775 | return type. Otherwise, we are still processing | |
776 | arguments. */ | |
777 | if (arg_num == 0) | |
778 | return_type = eltype; | |
779 | else | |
780 | args = tree_cons (NULL_TREE, eltype, args); | |
781 | } | |
342be7f7 | 782 | |
b5828b4b | 783 | ftype = build_function_type (return_type, args); |
43e9d192 | 784 | |
342be7f7 | 785 | gcc_assert (ftype != NULL); |
43e9d192 | 786 | |
b5828b4b | 787 | if (print_type_signature_p) |
bc5e395d JG |
788 | snprintf (namebuf, sizeof (namebuf), "__builtin_aarch64_%s_%s", |
789 | d->name, type_signature); | |
b5828b4b | 790 | else |
bc5e395d JG |
791 | snprintf (namebuf, sizeof (namebuf), "__builtin_aarch64_%s", |
792 | d->name); | |
43e9d192 | 793 | |
119103ca JG |
794 | fndecl = add_builtin_function (namebuf, ftype, fcode, BUILT_IN_MD, |
795 | NULL, NULL_TREE); | |
796 | aarch64_builtin_decls[fcode] = fndecl; | |
43e9d192 IB |
797 | } |
798 | } | |
799 | ||
5d357f26 KT |
800 | static void |
801 | aarch64_init_crc32_builtins () | |
802 | { | |
803 | tree usi_type = aarch64_build_unsigned_type (SImode); | |
804 | unsigned int i = 0; | |
805 | ||
806 | for (i = 0; i < ARRAY_SIZE (aarch64_crc_builtin_data); ++i) | |
807 | { | |
808 | aarch64_crc_builtin_datum* d = &aarch64_crc_builtin_data[i]; | |
809 | tree argtype = aarch64_build_unsigned_type (d->mode); | |
810 | tree ftype = build_function_type_list (usi_type, usi_type, argtype, NULL_TREE); | |
811 | tree fndecl = add_builtin_function (d->name, ftype, d->fcode, | |
812 | BUILT_IN_MD, NULL, NULL_TREE); | |
813 | ||
814 | aarch64_builtin_decls[d->fcode] = fndecl; | |
815 | } | |
816 | } | |
817 | ||
342be7f7 JG |
818 | void |
819 | aarch64_init_builtins (void) | |
43e9d192 | 820 | { |
aa87aced KV |
821 | tree ftype_set_fpr |
822 | = build_function_type_list (void_type_node, unsigned_type_node, NULL); | |
823 | tree ftype_get_fpr | |
824 | = build_function_type_list (unsigned_type_node, NULL); | |
825 | ||
826 | aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPCR] | |
827 | = add_builtin_function ("__builtin_aarch64_get_fpcr", ftype_get_fpr, | |
828 | AARCH64_BUILTIN_GET_FPCR, BUILT_IN_MD, NULL, NULL_TREE); | |
829 | aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPCR] | |
830 | = add_builtin_function ("__builtin_aarch64_set_fpcr", ftype_set_fpr, | |
831 | AARCH64_BUILTIN_SET_FPCR, BUILT_IN_MD, NULL, NULL_TREE); | |
832 | aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPSR] | |
833 | = add_builtin_function ("__builtin_aarch64_get_fpsr", ftype_get_fpr, | |
834 | AARCH64_BUILTIN_GET_FPSR, BUILT_IN_MD, NULL, NULL_TREE); | |
835 | aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPSR] | |
836 | = add_builtin_function ("__builtin_aarch64_set_fpsr", ftype_set_fpr, | |
837 | AARCH64_BUILTIN_SET_FPSR, BUILT_IN_MD, NULL, NULL_TREE); | |
838 | ||
342be7f7 JG |
839 | if (TARGET_SIMD) |
840 | aarch64_init_simd_builtins (); | |
5d357f26 KT |
841 | if (TARGET_CRC32) |
842 | aarch64_init_crc32_builtins (); | |
43e9d192 IB |
843 | } |
844 | ||
119103ca JG |
845 | tree |
846 | aarch64_builtin_decl (unsigned code, bool initialize_p ATTRIBUTE_UNUSED) | |
847 | { | |
848 | if (code >= AARCH64_BUILTIN_MAX) | |
849 | return error_mark_node; | |
850 | ||
851 | return aarch64_builtin_decls[code]; | |
852 | } | |
853 | ||
43e9d192 IB |
854 | typedef enum |
855 | { | |
856 | SIMD_ARG_COPY_TO_REG, | |
857 | SIMD_ARG_CONSTANT, | |
858 | SIMD_ARG_STOP | |
859 | } builtin_simd_arg; | |
860 | ||
43e9d192 IB |
861 | static rtx |
862 | aarch64_simd_expand_args (rtx target, int icode, int have_retval, | |
8d3d350a | 863 | tree exp, builtin_simd_arg *args) |
43e9d192 | 864 | { |
43e9d192 IB |
865 | rtx pat; |
866 | tree arg[SIMD_MAX_BUILTIN_ARGS]; | |
867 | rtx op[SIMD_MAX_BUILTIN_ARGS]; | |
868 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
869 | enum machine_mode mode[SIMD_MAX_BUILTIN_ARGS]; | |
870 | int argc = 0; | |
871 | ||
872 | if (have_retval | |
873 | && (!target | |
874 | || GET_MODE (target) != tmode | |
875 | || !(*insn_data[icode].operand[0].predicate) (target, tmode))) | |
876 | target = gen_reg_rtx (tmode); | |
877 | ||
43e9d192 IB |
878 | for (;;) |
879 | { | |
8d3d350a | 880 | builtin_simd_arg thisarg = args[argc]; |
43e9d192 IB |
881 | |
882 | if (thisarg == SIMD_ARG_STOP) | |
883 | break; | |
884 | else | |
885 | { | |
886 | arg[argc] = CALL_EXPR_ARG (exp, argc); | |
887 | op[argc] = expand_normal (arg[argc]); | |
888 | mode[argc] = insn_data[icode].operand[argc + have_retval].mode; | |
889 | ||
890 | switch (thisarg) | |
891 | { | |
892 | case SIMD_ARG_COPY_TO_REG: | |
2888c331 YZ |
893 | if (POINTER_TYPE_P (TREE_TYPE (arg[argc]))) |
894 | op[argc] = convert_memory_address (Pmode, op[argc]); | |
43e9d192 IB |
895 | /*gcc_assert (GET_MODE (op[argc]) == mode[argc]); */ |
896 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
897 | (op[argc], mode[argc])) | |
898 | op[argc] = copy_to_mode_reg (mode[argc], op[argc]); | |
899 | break; | |
900 | ||
901 | case SIMD_ARG_CONSTANT: | |
902 | if (!(*insn_data[icode].operand[argc + have_retval].predicate) | |
903 | (op[argc], mode[argc])) | |
904 | error_at (EXPR_LOCATION (exp), "incompatible type for argument %d, " | |
905 | "expected %<const int%>", argc + 1); | |
906 | break; | |
907 | ||
908 | case SIMD_ARG_STOP: | |
909 | gcc_unreachable (); | |
910 | } | |
911 | ||
912 | argc++; | |
913 | } | |
914 | } | |
915 | ||
43e9d192 IB |
916 | if (have_retval) |
917 | switch (argc) | |
918 | { | |
919 | case 1: | |
920 | pat = GEN_FCN (icode) (target, op[0]); | |
921 | break; | |
922 | ||
923 | case 2: | |
924 | pat = GEN_FCN (icode) (target, op[0], op[1]); | |
925 | break; | |
926 | ||
927 | case 3: | |
928 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2]); | |
929 | break; | |
930 | ||
931 | case 4: | |
932 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3]); | |
933 | break; | |
934 | ||
935 | case 5: | |
936 | pat = GEN_FCN (icode) (target, op[0], op[1], op[2], op[3], op[4]); | |
937 | break; | |
938 | ||
939 | default: | |
940 | gcc_unreachable (); | |
941 | } | |
942 | else | |
943 | switch (argc) | |
944 | { | |
945 | case 1: | |
946 | pat = GEN_FCN (icode) (op[0]); | |
947 | break; | |
948 | ||
949 | case 2: | |
950 | pat = GEN_FCN (icode) (op[0], op[1]); | |
951 | break; | |
952 | ||
953 | case 3: | |
954 | pat = GEN_FCN (icode) (op[0], op[1], op[2]); | |
955 | break; | |
956 | ||
957 | case 4: | |
958 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3]); | |
959 | break; | |
960 | ||
961 | case 5: | |
962 | pat = GEN_FCN (icode) (op[0], op[1], op[2], op[3], op[4]); | |
963 | break; | |
964 | ||
965 | default: | |
966 | gcc_unreachable (); | |
967 | } | |
968 | ||
969 | if (!pat) | |
970 | return 0; | |
971 | ||
972 | emit_insn (pat); | |
973 | ||
974 | return target; | |
975 | } | |
976 | ||
977 | /* Expand an AArch64 AdvSIMD builtin(intrinsic). */ | |
978 | rtx | |
979 | aarch64_simd_expand_builtin (int fcode, tree exp, rtx target) | |
980 | { | |
342be7f7 JG |
981 | aarch64_simd_builtin_datum *d = |
982 | &aarch64_simd_builtin_data[fcode - (AARCH64_SIMD_BUILTIN_BASE + 1)]; | |
342be7f7 | 983 | enum insn_code icode = d->code; |
b5828b4b JG |
984 | builtin_simd_arg args[SIMD_MAX_BUILTIN_ARGS]; |
985 | int num_args = insn_data[d->code].n_operands; | |
986 | int is_void = 0; | |
987 | int k; | |
43e9d192 | 988 | |
b5828b4b | 989 | is_void = !!(d->qualifiers[0] & qualifier_void); |
43e9d192 | 990 | |
b5828b4b JG |
991 | num_args += is_void; |
992 | ||
993 | for (k = 1; k < num_args; k++) | |
994 | { | |
995 | /* We have four arrays of data, each indexed in a different fashion. | |
996 | qualifiers - element 0 always describes the function return type. | |
997 | operands - element 0 is either the operand for return value (if | |
998 | the function has a non-void return type) or the operand for the | |
999 | first argument. | |
1000 | expr_args - element 0 always holds the first argument. | |
1001 | args - element 0 is always used for the return type. */ | |
1002 | int qualifiers_k = k; | |
1003 | int operands_k = k - is_void; | |
1004 | int expr_args_k = k - 1; | |
1005 | ||
1006 | if (d->qualifiers[qualifiers_k] & qualifier_immediate) | |
1007 | args[k] = SIMD_ARG_CONSTANT; | |
1008 | else if (d->qualifiers[qualifiers_k] & qualifier_maybe_immediate) | |
1009 | { | |
1010 | rtx arg | |
1011 | = expand_normal (CALL_EXPR_ARG (exp, | |
1012 | (expr_args_k))); | |
1013 | /* Handle constants only if the predicate allows it. */ | |
1014 | bool op_const_int_p = | |
1015 | (CONST_INT_P (arg) | |
1016 | && (*insn_data[icode].operand[operands_k].predicate) | |
1017 | (arg, insn_data[icode].operand[operands_k].mode)); | |
1018 | args[k] = op_const_int_p ? SIMD_ARG_CONSTANT : SIMD_ARG_COPY_TO_REG; | |
1019 | } | |
1020 | else | |
1021 | args[k] = SIMD_ARG_COPY_TO_REG; | |
43e9d192 | 1022 | |
43e9d192 | 1023 | } |
b5828b4b JG |
1024 | args[k] = SIMD_ARG_STOP; |
1025 | ||
1026 | /* The interface to aarch64_simd_expand_args expects a 0 if | |
1027 | the function is void, and a 1 if it is not. */ | |
1028 | return aarch64_simd_expand_args | |
8d3d350a | 1029 | (target, icode, !is_void, exp, &args[1]); |
43e9d192 | 1030 | } |
342be7f7 | 1031 | |
5d357f26 KT |
1032 | rtx |
1033 | aarch64_crc32_expand_builtin (int fcode, tree exp, rtx target) | |
1034 | { | |
1035 | rtx pat; | |
1036 | aarch64_crc_builtin_datum *d | |
1037 | = &aarch64_crc_builtin_data[fcode - (AARCH64_CRC32_BUILTIN_BASE + 1)]; | |
1038 | enum insn_code icode = d->icode; | |
1039 | tree arg0 = CALL_EXPR_ARG (exp, 0); | |
1040 | tree arg1 = CALL_EXPR_ARG (exp, 1); | |
1041 | rtx op0 = expand_normal (arg0); | |
1042 | rtx op1 = expand_normal (arg1); | |
1043 | enum machine_mode tmode = insn_data[icode].operand[0].mode; | |
1044 | enum machine_mode mode0 = insn_data[icode].operand[1].mode; | |
1045 | enum machine_mode mode1 = insn_data[icode].operand[2].mode; | |
1046 | ||
1047 | if (! target | |
1048 | || GET_MODE (target) != tmode | |
1049 | || ! (*insn_data[icode].operand[0].predicate) (target, tmode)) | |
1050 | target = gen_reg_rtx (tmode); | |
1051 | ||
1052 | gcc_assert ((GET_MODE (op0) == mode0 || GET_MODE (op0) == VOIDmode) | |
1053 | && (GET_MODE (op1) == mode1 || GET_MODE (op1) == VOIDmode)); | |
1054 | ||
1055 | if (! (*insn_data[icode].operand[1].predicate) (op0, mode0)) | |
1056 | op0 = copy_to_mode_reg (mode0, op0); | |
1057 | if (! (*insn_data[icode].operand[2].predicate) (op1, mode1)) | |
1058 | op1 = copy_to_mode_reg (mode1, op1); | |
1059 | ||
1060 | pat = GEN_FCN (icode) (target, op0, op1); | |
1061 | if (! pat) | |
1062 | return 0; | |
1063 | emit_insn (pat); | |
1064 | return target; | |
1065 | } | |
1066 | ||
342be7f7 JG |
1067 | /* Expand an expression EXP that calls a built-in function, |
1068 | with result going to TARGET if that's convenient. */ | |
1069 | rtx | |
1070 | aarch64_expand_builtin (tree exp, | |
1071 | rtx target, | |
1072 | rtx subtarget ATTRIBUTE_UNUSED, | |
1073 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
1074 | int ignore ATTRIBUTE_UNUSED) | |
1075 | { | |
1076 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
1077 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
aa87aced KV |
1078 | int icode; |
1079 | rtx pat, op0; | |
1080 | tree arg0; | |
1081 | ||
1082 | switch (fcode) | |
1083 | { | |
1084 | case AARCH64_BUILTIN_GET_FPCR: | |
1085 | case AARCH64_BUILTIN_SET_FPCR: | |
1086 | case AARCH64_BUILTIN_GET_FPSR: | |
1087 | case AARCH64_BUILTIN_SET_FPSR: | |
1088 | if ((fcode == AARCH64_BUILTIN_GET_FPCR) | |
1089 | || (fcode == AARCH64_BUILTIN_GET_FPSR)) | |
1090 | { | |
1091 | icode = (fcode == AARCH64_BUILTIN_GET_FPSR) ? | |
1092 | CODE_FOR_get_fpsr : CODE_FOR_get_fpcr; | |
1093 | target = gen_reg_rtx (SImode); | |
1094 | pat = GEN_FCN (icode) (target); | |
1095 | } | |
1096 | else | |
1097 | { | |
1098 | target = NULL_RTX; | |
1099 | icode = (fcode == AARCH64_BUILTIN_SET_FPSR) ? | |
1100 | CODE_FOR_set_fpsr : CODE_FOR_set_fpcr; | |
1101 | arg0 = CALL_EXPR_ARG (exp, 0); | |
1102 | op0 = expand_normal (arg0); | |
1103 | pat = GEN_FCN (icode) (op0); | |
1104 | } | |
1105 | emit_insn (pat); | |
1106 | return target; | |
1107 | } | |
342be7f7 | 1108 | |
5d357f26 | 1109 | if (fcode >= AARCH64_SIMD_BUILTIN_BASE && fcode <= AARCH64_SIMD_BUILTIN_MAX) |
342be7f7 | 1110 | return aarch64_simd_expand_builtin (fcode, exp, target); |
5d357f26 KT |
1111 | else if (fcode >= AARCH64_CRC32_BUILTIN_BASE && fcode <= AARCH64_CRC32_BUILTIN_MAX) |
1112 | return aarch64_crc32_expand_builtin (fcode, exp, target); | |
342be7f7 JG |
1113 | |
1114 | return NULL_RTX; | |
1115 | } | |
42fc9a7f JG |
1116 | |
1117 | tree | |
1118 | aarch64_builtin_vectorized_function (tree fndecl, tree type_out, tree type_in) | |
1119 | { | |
1120 | enum machine_mode in_mode, out_mode; | |
1121 | int in_n, out_n; | |
1122 | ||
1123 | if (TREE_CODE (type_out) != VECTOR_TYPE | |
1124 | || TREE_CODE (type_in) != VECTOR_TYPE) | |
1125 | return NULL_TREE; | |
1126 | ||
1127 | out_mode = TYPE_MODE (TREE_TYPE (type_out)); | |
1128 | out_n = TYPE_VECTOR_SUBPARTS (type_out); | |
1129 | in_mode = TYPE_MODE (TREE_TYPE (type_in)); | |
1130 | in_n = TYPE_VECTOR_SUBPARTS (type_in); | |
1131 | ||
1132 | #undef AARCH64_CHECK_BUILTIN_MODE | |
1133 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) 1 | |
1134 | #define AARCH64_FIND_FRINT_VARIANT(N) \ | |
1135 | (AARCH64_CHECK_BUILTIN_MODE (2, D) \ | |
e993fea1 | 1136 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v2df] \ |
42fc9a7f | 1137 | : (AARCH64_CHECK_BUILTIN_MODE (4, S) \ |
e993fea1 | 1138 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v4sf] \ |
42fc9a7f | 1139 | : (AARCH64_CHECK_BUILTIN_MODE (2, S) \ |
e993fea1 | 1140 | ? aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_##N##v2sf] \ |
42fc9a7f JG |
1141 | : NULL_TREE))) |
1142 | if (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL) | |
1143 | { | |
1144 | enum built_in_function fn = DECL_FUNCTION_CODE (fndecl); | |
1145 | switch (fn) | |
1146 | { | |
1147 | #undef AARCH64_CHECK_BUILTIN_MODE | |
1148 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ | |
1149 | (out_mode == N##Fmode && out_n == C \ | |
1150 | && in_mode == N##Fmode && in_n == C) | |
1151 | case BUILT_IN_FLOOR: | |
1152 | case BUILT_IN_FLOORF: | |
0659ce6f | 1153 | return AARCH64_FIND_FRINT_VARIANT (floor); |
42fc9a7f JG |
1154 | case BUILT_IN_CEIL: |
1155 | case BUILT_IN_CEILF: | |
0659ce6f | 1156 | return AARCH64_FIND_FRINT_VARIANT (ceil); |
42fc9a7f JG |
1157 | case BUILT_IN_TRUNC: |
1158 | case BUILT_IN_TRUNCF: | |
0659ce6f | 1159 | return AARCH64_FIND_FRINT_VARIANT (btrunc); |
42fc9a7f JG |
1160 | case BUILT_IN_ROUND: |
1161 | case BUILT_IN_ROUNDF: | |
0659ce6f | 1162 | return AARCH64_FIND_FRINT_VARIANT (round); |
42fc9a7f JG |
1163 | case BUILT_IN_NEARBYINT: |
1164 | case BUILT_IN_NEARBYINTF: | |
0659ce6f | 1165 | return AARCH64_FIND_FRINT_VARIANT (nearbyint); |
4dcd1054 JG |
1166 | case BUILT_IN_SQRT: |
1167 | case BUILT_IN_SQRTF: | |
1168 | return AARCH64_FIND_FRINT_VARIANT (sqrt); | |
42fc9a7f | 1169 | #undef AARCH64_CHECK_BUILTIN_MODE |
b5574232 VP |
1170 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ |
1171 | (out_mode == SImode && out_n == C \ | |
1172 | && in_mode == N##Imode && in_n == C) | |
1173 | case BUILT_IN_CLZ: | |
1174 | { | |
1175 | if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
e993fea1 | 1176 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOP_clzv4si]; |
b5574232 VP |
1177 | return NULL_TREE; |
1178 | } | |
1179 | #undef AARCH64_CHECK_BUILTIN_MODE | |
42fc9a7f JG |
1180 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ |
1181 | (out_mode == N##Imode && out_n == C \ | |
1182 | && in_mode == N##Fmode && in_n == C) | |
1183 | case BUILT_IN_LFLOOR: | |
bf0f324e YZ |
1184 | case BUILT_IN_LFLOORF: |
1185 | case BUILT_IN_LLFLOOR: | |
0386b123 | 1186 | case BUILT_IN_IFLOORF: |
ce966824 | 1187 | { |
e993fea1 | 1188 | enum aarch64_builtins builtin; |
ce966824 | 1189 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) |
e993fea1 | 1190 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv2dfv2di; |
ce966824 | 1191 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) |
e993fea1 | 1192 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv4sfv4si; |
ce966824 | 1193 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) |
e993fea1 JG |
1194 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lfloorv2sfv2si; |
1195 | else | |
1196 | return NULL_TREE; | |
1197 | ||
1198 | return aarch64_builtin_decls[builtin]; | |
ce966824 | 1199 | } |
42fc9a7f | 1200 | case BUILT_IN_LCEIL: |
bf0f324e YZ |
1201 | case BUILT_IN_LCEILF: |
1202 | case BUILT_IN_LLCEIL: | |
0386b123 | 1203 | case BUILT_IN_ICEILF: |
ce966824 | 1204 | { |
e993fea1 | 1205 | enum aarch64_builtins builtin; |
ce966824 | 1206 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) |
e993fea1 | 1207 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv2dfv2di; |
ce966824 | 1208 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) |
e993fea1 | 1209 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv4sfv4si; |
ce966824 | 1210 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) |
e993fea1 JG |
1211 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lceilv2sfv2si; |
1212 | else | |
1213 | return NULL_TREE; | |
1214 | ||
1215 | return aarch64_builtin_decls[builtin]; | |
ce966824 | 1216 | } |
0386b123 JG |
1217 | case BUILT_IN_LROUND: |
1218 | case BUILT_IN_IROUNDF: | |
1219 | { | |
e993fea1 | 1220 | enum aarch64_builtins builtin; |
0386b123 | 1221 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) |
e993fea1 | 1222 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv2dfv2di; |
0386b123 | 1223 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) |
e993fea1 | 1224 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv4sfv4si; |
0386b123 | 1225 | else if (AARCH64_CHECK_BUILTIN_MODE (2, S)) |
e993fea1 JG |
1226 | builtin = AARCH64_SIMD_BUILTIN_UNOP_lroundv2sfv2si; |
1227 | else | |
1228 | return NULL_TREE; | |
1229 | ||
1230 | return aarch64_builtin_decls[builtin]; | |
0386b123 | 1231 | } |
c7f28cd5 KT |
1232 | case BUILT_IN_BSWAP16: |
1233 | #undef AARCH64_CHECK_BUILTIN_MODE | |
1234 | #define AARCH64_CHECK_BUILTIN_MODE(C, N) \ | |
1235 | (out_mode == N##Imode && out_n == C \ | |
1236 | && in_mode == N##Imode && in_n == C) | |
1237 | if (AARCH64_CHECK_BUILTIN_MODE (4, H)) | |
1238 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv4hi]; | |
1239 | else if (AARCH64_CHECK_BUILTIN_MODE (8, H)) | |
1240 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv8hi]; | |
1241 | else | |
1242 | return NULL_TREE; | |
1243 | case BUILT_IN_BSWAP32: | |
1244 | if (AARCH64_CHECK_BUILTIN_MODE (2, S)) | |
1245 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv2si]; | |
1246 | else if (AARCH64_CHECK_BUILTIN_MODE (4, S)) | |
1247 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv4si]; | |
1248 | else | |
1249 | return NULL_TREE; | |
1250 | case BUILT_IN_BSWAP64: | |
1251 | if (AARCH64_CHECK_BUILTIN_MODE (2, D)) | |
1252 | return aarch64_builtin_decls[AARCH64_SIMD_BUILTIN_UNOPU_bswapv2di]; | |
1253 | else | |
1254 | return NULL_TREE; | |
42fc9a7f JG |
1255 | default: |
1256 | return NULL_TREE; | |
1257 | } | |
1258 | } | |
1259 | ||
1260 | return NULL_TREE; | |
1261 | } | |
0ac198d3 JG |
1262 | |
1263 | #undef VAR1 | |
1264 | #define VAR1(T, N, MAP, A) \ | |
e993fea1 | 1265 | case AARCH64_SIMD_BUILTIN_##T##_##N##A: |
0ac198d3 | 1266 | |
9697e620 JG |
1267 | tree |
1268 | aarch64_fold_builtin (tree fndecl, int n_args ATTRIBUTE_UNUSED, tree *args, | |
1269 | bool ignore ATTRIBUTE_UNUSED) | |
1270 | { | |
1271 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
1272 | tree type = TREE_TYPE (TREE_TYPE (fndecl)); | |
1273 | ||
1274 | switch (fcode) | |
1275 | { | |
d05d0709 | 1276 | BUILTIN_VALLDI (UNOP, abs, 2) |
9697e620 JG |
1277 | return fold_build1 (ABS_EXPR, type, args[0]); |
1278 | break; | |
c6a29a09 AL |
1279 | VAR1 (REINTERP_SS, reinterpretdi, 0, v1df) |
1280 | VAR1 (REINTERP_SS, reinterpretv8qi, 0, v1df) | |
1281 | VAR1 (REINTERP_SS, reinterpretv4hi, 0, v1df) | |
1282 | VAR1 (REINTERP_SS, reinterpretv2si, 0, v1df) | |
1283 | VAR1 (REINTERP_SS, reinterpretv2sf, 0, v1df) | |
1284 | BUILTIN_VD (REINTERP_SS, reinterpretv1df, 0) | |
1285 | BUILTIN_VD (REINTERP_SU, reinterpretv1df, 0) | |
1286 | VAR1 (REINTERP_US, reinterpretdi, 0, v1df) | |
1287 | VAR1 (REINTERP_US, reinterpretv8qi, 0, v1df) | |
1288 | VAR1 (REINTERP_US, reinterpretv4hi, 0, v1df) | |
1289 | VAR1 (REINTERP_US, reinterpretv2si, 0, v1df) | |
1290 | VAR1 (REINTERP_US, reinterpretv2sf, 0, v1df) | |
1291 | BUILTIN_VD (REINTERP_SP, reinterpretv1df, 0) | |
1292 | VAR1 (REINTERP_PS, reinterpretdi, 0, v1df) | |
1293 | VAR1 (REINTERP_PS, reinterpretv8qi, 0, v1df) | |
1294 | VAR1 (REINTERP_PS, reinterpretv4hi, 0, v1df) | |
1295 | VAR1 (REINTERP_PS, reinterpretv2sf, 0, v1df) | |
bcd48995 | 1296 | return fold_build1 (VIEW_CONVERT_EXPR, type, args[0]); |
1709ff9b JG |
1297 | VAR1 (UNOP, floatv2si, 2, v2sf) |
1298 | VAR1 (UNOP, floatv4si, 2, v4sf) | |
1299 | VAR1 (UNOP, floatv2di, 2, v2df) | |
1300 | return fold_build1 (FLOAT_EXPR, type, args[0]); | |
9697e620 JG |
1301 | default: |
1302 | break; | |
1303 | } | |
1304 | ||
1305 | return NULL_TREE; | |
1306 | } | |
1307 | ||
0ac198d3 JG |
1308 | bool |
1309 | aarch64_gimple_fold_builtin (gimple_stmt_iterator *gsi) | |
1310 | { | |
1311 | bool changed = false; | |
1312 | gimple stmt = gsi_stmt (*gsi); | |
1313 | tree call = gimple_call_fn (stmt); | |
1314 | tree fndecl; | |
1315 | gimple new_stmt = NULL; | |
22756ccf JG |
1316 | |
1317 | /* The operations folded below are reduction operations. These are | |
1318 | defined to leave their result in the 0'th element (from the perspective | |
1319 | of GCC). The architectural instruction we are folding will leave the | |
1320 | result in the 0'th element (from the perspective of the architecture). | |
1321 | For big-endian systems, these perspectives are not aligned. | |
1322 | ||
1323 | It is therefore wrong to perform this fold on big-endian. There | |
1324 | are some tricks we could play with shuffling, but the mid-end is | |
1325 | inconsistent in the way it treats reduction operations, so we will | |
1326 | end up in difficulty. Until we fix the ambiguity - just bail out. */ | |
1327 | if (BYTES_BIG_ENDIAN) | |
1328 | return false; | |
1329 | ||
0ac198d3 JG |
1330 | if (call) |
1331 | { | |
1332 | fndecl = gimple_call_fndecl (stmt); | |
1333 | if (fndecl) | |
1334 | { | |
1335 | int fcode = DECL_FUNCTION_CODE (fndecl); | |
1336 | int nargs = gimple_call_num_args (stmt); | |
1337 | tree *args = (nargs > 0 | |
1338 | ? gimple_call_arg_ptr (stmt, 0) | |
1339 | : &error_mark_node); | |
1340 | ||
1341 | switch (fcode) | |
1342 | { | |
36054fab | 1343 | BUILTIN_VALL (UNOP, reduc_splus_, 10) |
0ac198d3 JG |
1344 | new_stmt = gimple_build_assign_with_ops ( |
1345 | REDUC_PLUS_EXPR, | |
1346 | gimple_call_lhs (stmt), | |
1347 | args[0], | |
1348 | NULL_TREE); | |
1349 | break; | |
1598945b JG |
1350 | BUILTIN_VDQIF (UNOP, reduc_smax_, 10) |
1351 | new_stmt = gimple_build_assign_with_ops ( | |
1352 | REDUC_MAX_EXPR, | |
1353 | gimple_call_lhs (stmt), | |
1354 | args[0], | |
1355 | NULL_TREE); | |
1356 | break; | |
1357 | BUILTIN_VDQIF (UNOP, reduc_smin_, 10) | |
1358 | new_stmt = gimple_build_assign_with_ops ( | |
1359 | REDUC_MIN_EXPR, | |
1360 | gimple_call_lhs (stmt), | |
1361 | args[0], | |
1362 | NULL_TREE); | |
1363 | break; | |
1364 | ||
0ac198d3 JG |
1365 | default: |
1366 | break; | |
1367 | } | |
1368 | } | |
1369 | } | |
1370 | ||
1371 | if (new_stmt) | |
1372 | { | |
1373 | gsi_replace (gsi, new_stmt, true); | |
1374 | changed = true; | |
1375 | } | |
1376 | ||
1377 | return changed; | |
1378 | } | |
1379 | ||
aa87aced KV |
1380 | void |
1381 | aarch64_atomic_assign_expand_fenv (tree *hold, tree *clear, tree *update) | |
1382 | { | |
1383 | const unsigned AARCH64_FE_INVALID = 1; | |
1384 | const unsigned AARCH64_FE_DIVBYZERO = 2; | |
1385 | const unsigned AARCH64_FE_OVERFLOW = 4; | |
1386 | const unsigned AARCH64_FE_UNDERFLOW = 8; | |
1387 | const unsigned AARCH64_FE_INEXACT = 16; | |
1388 | const unsigned HOST_WIDE_INT AARCH64_FE_ALL_EXCEPT = (AARCH64_FE_INVALID | |
1389 | | AARCH64_FE_DIVBYZERO | |
1390 | | AARCH64_FE_OVERFLOW | |
1391 | | AARCH64_FE_UNDERFLOW | |
1392 | | AARCH64_FE_INEXACT); | |
1393 | const unsigned HOST_WIDE_INT AARCH64_FE_EXCEPT_SHIFT = 8; | |
1394 | tree fenv_cr, fenv_sr, get_fpcr, set_fpcr, mask_cr, mask_sr; | |
1395 | tree ld_fenv_cr, ld_fenv_sr, masked_fenv_cr, masked_fenv_sr, hold_fnclex_cr; | |
1396 | tree hold_fnclex_sr, new_fenv_var, reload_fenv, restore_fnenv, get_fpsr, set_fpsr; | |
1397 | tree update_call, atomic_feraiseexcept, hold_fnclex, masked_fenv, ld_fenv; | |
1398 | ||
1399 | /* Generate the equivalence of : | |
1400 | unsigned int fenv_cr; | |
1401 | fenv_cr = __builtin_aarch64_get_fpcr (); | |
1402 | ||
1403 | unsigned int fenv_sr; | |
1404 | fenv_sr = __builtin_aarch64_get_fpsr (); | |
1405 | ||
1406 | Now set all exceptions to non-stop | |
1407 | unsigned int mask_cr | |
1408 | = ~(AARCH64_FE_ALL_EXCEPT << AARCH64_FE_EXCEPT_SHIFT); | |
1409 | unsigned int masked_cr; | |
1410 | masked_cr = fenv_cr & mask_cr; | |
1411 | ||
1412 | And clear all exception flags | |
1413 | unsigned int maske_sr = ~AARCH64_FE_ALL_EXCEPT; | |
1414 | unsigned int masked_cr; | |
1415 | masked_sr = fenv_sr & mask_sr; | |
1416 | ||
1417 | __builtin_aarch64_set_cr (masked_cr); | |
1418 | __builtin_aarch64_set_sr (masked_sr); */ | |
1419 | ||
1420 | fenv_cr = create_tmp_var (unsigned_type_node, NULL); | |
1421 | fenv_sr = create_tmp_var (unsigned_type_node, NULL); | |
1422 | ||
1423 | get_fpcr = aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPCR]; | |
1424 | set_fpcr = aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPCR]; | |
1425 | get_fpsr = aarch64_builtin_decls[AARCH64_BUILTIN_GET_FPSR]; | |
1426 | set_fpsr = aarch64_builtin_decls[AARCH64_BUILTIN_SET_FPSR]; | |
1427 | ||
1428 | mask_cr = build_int_cst (unsigned_type_node, | |
1429 | ~(AARCH64_FE_ALL_EXCEPT << AARCH64_FE_EXCEPT_SHIFT)); | |
1430 | mask_sr = build_int_cst (unsigned_type_node, | |
1431 | ~(AARCH64_FE_ALL_EXCEPT)); | |
1432 | ||
1433 | ld_fenv_cr = build2 (MODIFY_EXPR, unsigned_type_node, | |
1434 | fenv_cr, build_call_expr (get_fpcr, 0)); | |
1435 | ld_fenv_sr = build2 (MODIFY_EXPR, unsigned_type_node, | |
1436 | fenv_sr, build_call_expr (get_fpsr, 0)); | |
1437 | ||
1438 | masked_fenv_cr = build2 (BIT_AND_EXPR, unsigned_type_node, fenv_cr, mask_cr); | |
1439 | masked_fenv_sr = build2 (BIT_AND_EXPR, unsigned_type_node, fenv_sr, mask_sr); | |
1440 | ||
1441 | hold_fnclex_cr = build_call_expr (set_fpcr, 1, masked_fenv_cr); | |
1442 | hold_fnclex_sr = build_call_expr (set_fpsr, 1, masked_fenv_sr); | |
1443 | ||
1444 | hold_fnclex = build2 (COMPOUND_EXPR, void_type_node, hold_fnclex_cr, | |
1445 | hold_fnclex_sr); | |
1446 | masked_fenv = build2 (COMPOUND_EXPR, void_type_node, masked_fenv_cr, | |
1447 | masked_fenv_sr); | |
1448 | ld_fenv = build2 (COMPOUND_EXPR, void_type_node, ld_fenv_cr, ld_fenv_sr); | |
1449 | ||
1450 | *hold = build2 (COMPOUND_EXPR, void_type_node, | |
1451 | build2 (COMPOUND_EXPR, void_type_node, masked_fenv, ld_fenv), | |
1452 | hold_fnclex); | |
1453 | ||
1454 | /* Store the value of masked_fenv to clear the exceptions: | |
1455 | __builtin_aarch64_set_fpsr (masked_fenv_sr); */ | |
1456 | ||
1457 | *clear = build_call_expr (set_fpsr, 1, masked_fenv_sr); | |
1458 | ||
1459 | /* Generate the equivalent of : | |
1460 | unsigned int new_fenv_var; | |
1461 | new_fenv_var = __builtin_aarch64_get_fpsr (); | |
1462 | ||
1463 | __builtin_aarch64_set_fpsr (fenv_sr); | |
1464 | ||
1465 | __atomic_feraiseexcept (new_fenv_var); */ | |
1466 | ||
1467 | new_fenv_var = create_tmp_var (unsigned_type_node, NULL); | |
1468 | reload_fenv = build2 (MODIFY_EXPR, unsigned_type_node, | |
1469 | new_fenv_var, build_call_expr (get_fpsr, 0)); | |
1470 | restore_fnenv = build_call_expr (set_fpsr, 1, fenv_sr); | |
1471 | atomic_feraiseexcept = builtin_decl_implicit (BUILT_IN_ATOMIC_FERAISEEXCEPT); | |
1472 | update_call = build_call_expr (atomic_feraiseexcept, 1, | |
1473 | fold_convert (integer_type_node, new_fenv_var)); | |
1474 | *update = build2 (COMPOUND_EXPR, void_type_node, | |
1475 | build2 (COMPOUND_EXPR, void_type_node, | |
1476 | reload_fenv, restore_fnenv), update_call); | |
1477 | } | |
1478 | ||
1479 | ||
42fc9a7f JG |
1480 | #undef AARCH64_CHECK_BUILTIN_MODE |
1481 | #undef AARCH64_FIND_FRINT_VARIANT | |
0ddec79f JG |
1482 | #undef BUILTIN_DX |
1483 | #undef BUILTIN_SDQ_I | |
1484 | #undef BUILTIN_SD_HSI | |
1485 | #undef BUILTIN_V2F | |
1486 | #undef BUILTIN_VALL | |
1487 | #undef BUILTIN_VB | |
1488 | #undef BUILTIN_VD | |
c6a29a09 | 1489 | #undef BUILTIN_VD1 |
0ddec79f JG |
1490 | #undef BUILTIN_VDC |
1491 | #undef BUILTIN_VDIC | |
1492 | #undef BUILTIN_VDN | |
1493 | #undef BUILTIN_VDQ | |
1494 | #undef BUILTIN_VDQF | |
1495 | #undef BUILTIN_VDQH | |
1496 | #undef BUILTIN_VDQHS | |
1497 | #undef BUILTIN_VDQIF | |
1498 | #undef BUILTIN_VDQM | |
1499 | #undef BUILTIN_VDQV | |
1500 | #undef BUILTIN_VDQ_BHSI | |
1501 | #undef BUILTIN_VDQ_I | |
1502 | #undef BUILTIN_VDW | |
1503 | #undef BUILTIN_VD_BHSI | |
1504 | #undef BUILTIN_VD_HSI | |
0ddec79f JG |
1505 | #undef BUILTIN_VQ |
1506 | #undef BUILTIN_VQN | |
1507 | #undef BUILTIN_VQW | |
1508 | #undef BUILTIN_VQ_HSI | |
1509 | #undef BUILTIN_VQ_S | |
1510 | #undef BUILTIN_VSDQ_HSI | |
1511 | #undef BUILTIN_VSDQ_I | |
1512 | #undef BUILTIN_VSDQ_I_BHSI | |
1513 | #undef BUILTIN_VSDQ_I_DI | |
1514 | #undef BUILTIN_VSD_HSI | |
1515 | #undef BUILTIN_VSQN_HSDI | |
1516 | #undef BUILTIN_VSTRUCT | |
1517 | #undef CF0 | |
1518 | #undef CF1 | |
1519 | #undef CF2 | |
1520 | #undef CF3 | |
1521 | #undef CF4 | |
1522 | #undef CF10 | |
1523 | #undef VAR1 | |
1524 | #undef VAR2 | |
1525 | #undef VAR3 | |
1526 | #undef VAR4 | |
1527 | #undef VAR5 | |
1528 | #undef VAR6 | |
1529 | #undef VAR7 | |
1530 | #undef VAR8 | |
1531 | #undef VAR9 | |
1532 | #undef VAR10 | |
1533 | #undef VAR11 | |
1534 |