]>
Commit | Line | Data |
---|---|---|
bdb7bf8a | 1 | /* Machine description for AArch64 architecture. |
23a5b65a | 2 | Copyright (C) 2009-2014 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 "insn-codes.h" | |
26 | #include "rtl.h" | |
27 | #include "insn-attr.h" | |
28 | #include "tree.h" | |
d8a2d370 DN |
29 | #include "stringpool.h" |
30 | #include "stor-layout.h" | |
31 | #include "calls.h" | |
32 | #include "varasm.h" | |
43e9d192 IB |
33 | #include "regs.h" |
34 | #include "df.h" | |
35 | #include "hard-reg-set.h" | |
36 | #include "output.h" | |
37 | #include "expr.h" | |
38 | #include "reload.h" | |
39 | #include "toplev.h" | |
40 | #include "target.h" | |
41 | #include "target-def.h" | |
42 | #include "targhooks.h" | |
43 | #include "ggc.h" | |
44 | #include "function.h" | |
45 | #include "tm_p.h" | |
46 | #include "recog.h" | |
47 | #include "langhooks.h" | |
48 | #include "diagnostic-core.h" | |
2fb9a547 AM |
49 | #include "pointer-set.h" |
50 | #include "hash-table.h" | |
51 | #include "vec.h" | |
52 | #include "basic-block.h" | |
53 | #include "tree-ssa-alias.h" | |
54 | #include "internal-fn.h" | |
55 | #include "gimple-fold.h" | |
56 | #include "tree-eh.h" | |
57 | #include "gimple-expr.h" | |
58 | #include "is-a.h" | |
18f429e2 | 59 | #include "gimple.h" |
45b0be94 | 60 | #include "gimplify.h" |
43e9d192 IB |
61 | #include "optabs.h" |
62 | #include "dwarf2.h" | |
8990e73a TB |
63 | #include "cfgloop.h" |
64 | #include "tree-vectorizer.h" | |
73250c4c | 65 | #include "config/arm/aarch-cost-tables.h" |
43e9d192 | 66 | |
28514dda YZ |
67 | /* Defined for convenience. */ |
68 | #define POINTER_BYTES (POINTER_SIZE / BITS_PER_UNIT) | |
69 | ||
43e9d192 IB |
70 | /* Classifies an address. |
71 | ||
72 | ADDRESS_REG_IMM | |
73 | A simple base register plus immediate offset. | |
74 | ||
75 | ADDRESS_REG_WB | |
76 | A base register indexed by immediate offset with writeback. | |
77 | ||
78 | ADDRESS_REG_REG | |
79 | A base register indexed by (optionally scaled) register. | |
80 | ||
81 | ADDRESS_REG_UXTW | |
82 | A base register indexed by (optionally scaled) zero-extended register. | |
83 | ||
84 | ADDRESS_REG_SXTW | |
85 | A base register indexed by (optionally scaled) sign-extended register. | |
86 | ||
87 | ADDRESS_LO_SUM | |
88 | A LO_SUM rtx with a base register and "LO12" symbol relocation. | |
89 | ||
90 | ADDRESS_SYMBOLIC: | |
91 | A constant symbolic address, in pc-relative literal pool. */ | |
92 | ||
93 | enum aarch64_address_type { | |
94 | ADDRESS_REG_IMM, | |
95 | ADDRESS_REG_WB, | |
96 | ADDRESS_REG_REG, | |
97 | ADDRESS_REG_UXTW, | |
98 | ADDRESS_REG_SXTW, | |
99 | ADDRESS_LO_SUM, | |
100 | ADDRESS_SYMBOLIC | |
101 | }; | |
102 | ||
103 | struct aarch64_address_info { | |
104 | enum aarch64_address_type type; | |
105 | rtx base; | |
106 | rtx offset; | |
107 | int shift; | |
108 | enum aarch64_symbol_type symbol_type; | |
109 | }; | |
110 | ||
48063b9d IB |
111 | struct simd_immediate_info |
112 | { | |
113 | rtx value; | |
114 | int shift; | |
115 | int element_width; | |
48063b9d | 116 | bool mvn; |
e4f0f84d | 117 | bool msl; |
48063b9d IB |
118 | }; |
119 | ||
43e9d192 IB |
120 | /* The current code model. */ |
121 | enum aarch64_code_model aarch64_cmodel; | |
122 | ||
123 | #ifdef HAVE_AS_TLS | |
124 | #undef TARGET_HAVE_TLS | |
125 | #define TARGET_HAVE_TLS 1 | |
126 | #endif | |
127 | ||
38e8f663 | 128 | static bool aarch64_lra_p (void); |
43e9d192 IB |
129 | static bool aarch64_composite_type_p (const_tree, enum machine_mode); |
130 | static bool aarch64_vfp_is_call_or_return_candidate (enum machine_mode, | |
131 | const_tree, | |
132 | enum machine_mode *, int *, | |
133 | bool *); | |
134 | static void aarch64_elf_asm_constructor (rtx, int) ATTRIBUTE_UNUSED; | |
135 | static void aarch64_elf_asm_destructor (rtx, int) ATTRIBUTE_UNUSED; | |
43e9d192 | 136 | static void aarch64_override_options_after_change (void); |
43e9d192 IB |
137 | static bool aarch64_vector_mode_supported_p (enum machine_mode); |
138 | static unsigned bit_count (unsigned HOST_WIDE_INT); | |
139 | static bool aarch64_const_vec_all_same_int_p (rtx, | |
140 | HOST_WIDE_INT, HOST_WIDE_INT); | |
141 | ||
88b08073 JG |
142 | static bool aarch64_vectorize_vec_perm_const_ok (enum machine_mode vmode, |
143 | const unsigned char *sel); | |
144 | ||
43e9d192 | 145 | /* The processor for which instructions should be scheduled. */ |
02fdbd5b | 146 | enum aarch64_processor aarch64_tune = cortexa53; |
43e9d192 IB |
147 | |
148 | /* The current tuning set. */ | |
149 | const struct tune_params *aarch64_tune_params; | |
150 | ||
151 | /* Mask to specify which instructions we are allowed to generate. */ | |
152 | unsigned long aarch64_isa_flags = 0; | |
153 | ||
154 | /* Mask to specify which instruction scheduling options should be used. */ | |
155 | unsigned long aarch64_tune_flags = 0; | |
156 | ||
157 | /* Tuning parameters. */ | |
158 | ||
159 | #if HAVE_DESIGNATED_INITIALIZERS | |
160 | #define NAMED_PARAM(NAME, VAL) .NAME = (VAL) | |
161 | #else | |
162 | #define NAMED_PARAM(NAME, VAL) (VAL) | |
163 | #endif | |
164 | ||
165 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
166 | __extension__ | |
167 | #endif | |
43e9d192 IB |
168 | |
169 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
170 | __extension__ | |
171 | #endif | |
172 | static const struct cpu_addrcost_table generic_addrcost_table = | |
173 | { | |
174 | NAMED_PARAM (pre_modify, 0), | |
175 | NAMED_PARAM (post_modify, 0), | |
176 | NAMED_PARAM (register_offset, 0), | |
177 | NAMED_PARAM (register_extend, 0), | |
178 | NAMED_PARAM (imm_offset, 0) | |
179 | }; | |
180 | ||
181 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
182 | __extension__ | |
183 | #endif | |
184 | static const struct cpu_regmove_cost generic_regmove_cost = | |
185 | { | |
186 | NAMED_PARAM (GP2GP, 1), | |
187 | NAMED_PARAM (GP2FP, 2), | |
188 | NAMED_PARAM (FP2GP, 2), | |
189 | /* We currently do not provide direct support for TFmode Q->Q move. | |
190 | Therefore we need to raise the cost above 2 in order to have | |
191 | reload handle the situation. */ | |
192 | NAMED_PARAM (FP2FP, 4) | |
193 | }; | |
194 | ||
8990e73a TB |
195 | /* Generic costs for vector insn classes. */ |
196 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 | |
197 | __extension__ | |
198 | #endif | |
199 | static const struct cpu_vector_cost generic_vector_cost = | |
200 | { | |
201 | NAMED_PARAM (scalar_stmt_cost, 1), | |
202 | NAMED_PARAM (scalar_load_cost, 1), | |
203 | NAMED_PARAM (scalar_store_cost, 1), | |
204 | NAMED_PARAM (vec_stmt_cost, 1), | |
205 | NAMED_PARAM (vec_to_scalar_cost, 1), | |
206 | NAMED_PARAM (scalar_to_vec_cost, 1), | |
207 | NAMED_PARAM (vec_align_load_cost, 1), | |
208 | NAMED_PARAM (vec_unalign_load_cost, 1), | |
209 | NAMED_PARAM (vec_unalign_store_cost, 1), | |
210 | NAMED_PARAM (vec_store_cost, 1), | |
211 | NAMED_PARAM (cond_taken_branch_cost, 3), | |
212 | NAMED_PARAM (cond_not_taken_branch_cost, 1) | |
213 | }; | |
214 | ||
43e9d192 IB |
215 | #if HAVE_DESIGNATED_INITIALIZERS && GCC_VERSION >= 2007 |
216 | __extension__ | |
217 | #endif | |
218 | static const struct tune_params generic_tunings = | |
219 | { | |
73250c4c | 220 | &generic_extra_costs, |
43e9d192 IB |
221 | &generic_addrcost_table, |
222 | &generic_regmove_cost, | |
8990e73a | 223 | &generic_vector_cost, |
d126a4ae AP |
224 | NAMED_PARAM (memmov_cost, 4), |
225 | NAMED_PARAM (issue_rate, 2) | |
43e9d192 IB |
226 | }; |
227 | ||
984239ad KT |
228 | static const struct tune_params cortexa53_tunings = |
229 | { | |
230 | &cortexa53_extra_costs, | |
231 | &generic_addrcost_table, | |
232 | &generic_regmove_cost, | |
233 | &generic_vector_cost, | |
d126a4ae AP |
234 | NAMED_PARAM (memmov_cost, 4), |
235 | NAMED_PARAM (issue_rate, 2) | |
984239ad KT |
236 | }; |
237 | ||
43e9d192 IB |
238 | /* A processor implementing AArch64. */ |
239 | struct processor | |
240 | { | |
241 | const char *const name; | |
242 | enum aarch64_processor core; | |
243 | const char *arch; | |
244 | const unsigned long flags; | |
245 | const struct tune_params *const tune; | |
246 | }; | |
247 | ||
248 | /* Processor cores implementing AArch64. */ | |
249 | static const struct processor all_cores[] = | |
250 | { | |
192ed1dd | 251 | #define AARCH64_CORE(NAME, X, IDENT, ARCH, FLAGS, COSTS) \ |
43e9d192 IB |
252 | {NAME, IDENT, #ARCH, FLAGS | AARCH64_FL_FOR_ARCH##ARCH, &COSTS##_tunings}, |
253 | #include "aarch64-cores.def" | |
254 | #undef AARCH64_CORE | |
02fdbd5b | 255 | {"generic", cortexa53, "8", AARCH64_FL_FPSIMD | AARCH64_FL_FOR_ARCH8, &generic_tunings}, |
43e9d192 IB |
256 | {NULL, aarch64_none, NULL, 0, NULL} |
257 | }; | |
258 | ||
259 | /* Architectures implementing AArch64. */ | |
260 | static const struct processor all_architectures[] = | |
261 | { | |
262 | #define AARCH64_ARCH(NAME, CORE, ARCH, FLAGS) \ | |
263 | {NAME, CORE, #ARCH, FLAGS, NULL}, | |
264 | #include "aarch64-arches.def" | |
265 | #undef AARCH64_ARCH | |
43e9d192 IB |
266 | {NULL, aarch64_none, NULL, 0, NULL} |
267 | }; | |
268 | ||
269 | /* Target specification. These are populated as commandline arguments | |
270 | are processed, or NULL if not specified. */ | |
271 | static const struct processor *selected_arch; | |
272 | static const struct processor *selected_cpu; | |
273 | static const struct processor *selected_tune; | |
274 | ||
275 | #define AARCH64_CPU_DEFAULT_FLAGS ((selected_cpu) ? selected_cpu->flags : 0) | |
276 | ||
277 | /* An ISA extension in the co-processor and main instruction set space. */ | |
278 | struct aarch64_option_extension | |
279 | { | |
280 | const char *const name; | |
281 | const unsigned long flags_on; | |
282 | const unsigned long flags_off; | |
283 | }; | |
284 | ||
285 | /* ISA extensions in AArch64. */ | |
286 | static const struct aarch64_option_extension all_extensions[] = | |
287 | { | |
288 | #define AARCH64_OPT_EXTENSION(NAME, FLAGS_ON, FLAGS_OFF) \ | |
289 | {NAME, FLAGS_ON, FLAGS_OFF}, | |
290 | #include "aarch64-option-extensions.def" | |
291 | #undef AARCH64_OPT_EXTENSION | |
292 | {NULL, 0, 0} | |
293 | }; | |
294 | ||
295 | /* Used to track the size of an address when generating a pre/post | |
296 | increment address. */ | |
297 | static enum machine_mode aarch64_memory_reference_mode; | |
298 | ||
299 | /* Used to force GTY into this file. */ | |
300 | static GTY(()) int gty_dummy; | |
301 | ||
302 | /* A table of valid AArch64 "bitmask immediate" values for | |
303 | logical instructions. */ | |
304 | ||
305 | #define AARCH64_NUM_BITMASKS 5334 | |
306 | static unsigned HOST_WIDE_INT aarch64_bitmasks[AARCH64_NUM_BITMASKS]; | |
307 | ||
308 | /* Did we set flag_omit_frame_pointer just so | |
309 | aarch64_frame_pointer_required would be called? */ | |
310 | static bool faked_omit_frame_pointer; | |
311 | ||
312 | typedef enum aarch64_cond_code | |
313 | { | |
314 | AARCH64_EQ = 0, AARCH64_NE, AARCH64_CS, AARCH64_CC, AARCH64_MI, AARCH64_PL, | |
315 | AARCH64_VS, AARCH64_VC, AARCH64_HI, AARCH64_LS, AARCH64_GE, AARCH64_LT, | |
316 | AARCH64_GT, AARCH64_LE, AARCH64_AL, AARCH64_NV | |
317 | } | |
318 | aarch64_cc; | |
319 | ||
320 | #define AARCH64_INVERSE_CONDITION_CODE(X) ((aarch64_cc) (((int) X) ^ 1)) | |
321 | ||
322 | /* The condition codes of the processor, and the inverse function. */ | |
323 | static const char * const aarch64_condition_codes[] = | |
324 | { | |
325 | "eq", "ne", "cs", "cc", "mi", "pl", "vs", "vc", | |
326 | "hi", "ls", "ge", "lt", "gt", "le", "al", "nv" | |
327 | }; | |
328 | ||
329 | /* Provide a mapping from gcc register numbers to dwarf register numbers. */ | |
330 | unsigned | |
331 | aarch64_dbx_register_number (unsigned regno) | |
332 | { | |
333 | if (GP_REGNUM_P (regno)) | |
334 | return AARCH64_DWARF_R0 + regno - R0_REGNUM; | |
335 | else if (regno == SP_REGNUM) | |
336 | return AARCH64_DWARF_SP; | |
337 | else if (FP_REGNUM_P (regno)) | |
338 | return AARCH64_DWARF_V0 + regno - V0_REGNUM; | |
339 | ||
340 | /* Return values >= DWARF_FRAME_REGISTERS indicate that there is no | |
341 | equivalent DWARF register. */ | |
342 | return DWARF_FRAME_REGISTERS; | |
343 | } | |
344 | ||
345 | /* Return TRUE if MODE is any of the large INT modes. */ | |
346 | static bool | |
347 | aarch64_vect_struct_mode_p (enum machine_mode mode) | |
348 | { | |
349 | return mode == OImode || mode == CImode || mode == XImode; | |
350 | } | |
351 | ||
352 | /* Return TRUE if MODE is any of the vector modes. */ | |
353 | static bool | |
354 | aarch64_vector_mode_p (enum machine_mode mode) | |
355 | { | |
356 | return aarch64_vector_mode_supported_p (mode) | |
357 | || aarch64_vect_struct_mode_p (mode); | |
358 | } | |
359 | ||
360 | /* Implement target hook TARGET_ARRAY_MODE_SUPPORTED_P. */ | |
361 | static bool | |
362 | aarch64_array_mode_supported_p (enum machine_mode mode, | |
363 | unsigned HOST_WIDE_INT nelems) | |
364 | { | |
365 | if (TARGET_SIMD | |
366 | && AARCH64_VALID_SIMD_QREG_MODE (mode) | |
367 | && (nelems >= 2 && nelems <= 4)) | |
368 | return true; | |
369 | ||
370 | return false; | |
371 | } | |
372 | ||
373 | /* Implement HARD_REGNO_NREGS. */ | |
374 | ||
375 | int | |
376 | aarch64_hard_regno_nregs (unsigned regno, enum machine_mode mode) | |
377 | { | |
378 | switch (aarch64_regno_regclass (regno)) | |
379 | { | |
380 | case FP_REGS: | |
381 | case FP_LO_REGS: | |
382 | return (GET_MODE_SIZE (mode) + UNITS_PER_VREG - 1) / UNITS_PER_VREG; | |
383 | default: | |
384 | return (GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
385 | } | |
386 | gcc_unreachable (); | |
387 | } | |
388 | ||
389 | /* Implement HARD_REGNO_MODE_OK. */ | |
390 | ||
391 | int | |
392 | aarch64_hard_regno_mode_ok (unsigned regno, enum machine_mode mode) | |
393 | { | |
394 | if (GET_MODE_CLASS (mode) == MODE_CC) | |
395 | return regno == CC_REGNUM; | |
396 | ||
9259db42 YZ |
397 | if (regno == SP_REGNUM) |
398 | /* The purpose of comparing with ptr_mode is to support the | |
399 | global register variable associated with the stack pointer | |
400 | register via the syntax of asm ("wsp") in ILP32. */ | |
401 | return mode == Pmode || mode == ptr_mode; | |
402 | ||
403 | if (regno == FRAME_POINTER_REGNUM || regno == ARG_POINTER_REGNUM) | |
43e9d192 IB |
404 | return mode == Pmode; |
405 | ||
406 | if (GP_REGNUM_P (regno) && ! aarch64_vect_struct_mode_p (mode)) | |
407 | return 1; | |
408 | ||
409 | if (FP_REGNUM_P (regno)) | |
410 | { | |
411 | if (aarch64_vect_struct_mode_p (mode)) | |
412 | return | |
413 | (regno + aarch64_hard_regno_nregs (regno, mode) - 1) <= V31_REGNUM; | |
414 | else | |
415 | return 1; | |
416 | } | |
417 | ||
418 | return 0; | |
419 | } | |
420 | ||
421 | /* Return true if calls to DECL should be treated as | |
422 | long-calls (ie called via a register). */ | |
423 | static bool | |
424 | aarch64_decl_is_long_call_p (const_tree decl ATTRIBUTE_UNUSED) | |
425 | { | |
426 | return false; | |
427 | } | |
428 | ||
429 | /* Return true if calls to symbol-ref SYM should be treated as | |
430 | long-calls (ie called via a register). */ | |
431 | bool | |
432 | aarch64_is_long_call_p (rtx sym) | |
433 | { | |
434 | return aarch64_decl_is_long_call_p (SYMBOL_REF_DECL (sym)); | |
435 | } | |
436 | ||
437 | /* Return true if the offsets to a zero/sign-extract operation | |
438 | represent an expression that matches an extend operation. The | |
439 | operands represent the paramters from | |
440 | ||
441 | (extract (mult (reg) (mult_imm)) (extract_imm) (const_int 0)). */ | |
442 | bool | |
443 | aarch64_is_extend_from_extract (enum machine_mode mode, rtx mult_imm, | |
444 | rtx extract_imm) | |
445 | { | |
446 | HOST_WIDE_INT mult_val, extract_val; | |
447 | ||
448 | if (! CONST_INT_P (mult_imm) || ! CONST_INT_P (extract_imm)) | |
449 | return false; | |
450 | ||
451 | mult_val = INTVAL (mult_imm); | |
452 | extract_val = INTVAL (extract_imm); | |
453 | ||
454 | if (extract_val > 8 | |
455 | && extract_val < GET_MODE_BITSIZE (mode) | |
456 | && exact_log2 (extract_val & ~7) > 0 | |
457 | && (extract_val & 7) <= 4 | |
458 | && mult_val == (1 << (extract_val & 7))) | |
459 | return true; | |
460 | ||
461 | return false; | |
462 | } | |
463 | ||
464 | /* Emit an insn that's a simple single-set. Both the operands must be | |
465 | known to be valid. */ | |
466 | inline static rtx | |
467 | emit_set_insn (rtx x, rtx y) | |
468 | { | |
469 | return emit_insn (gen_rtx_SET (VOIDmode, x, y)); | |
470 | } | |
471 | ||
472 | /* X and Y are two things to compare using CODE. Emit the compare insn and | |
473 | return the rtx for register 0 in the proper mode. */ | |
474 | rtx | |
475 | aarch64_gen_compare_reg (RTX_CODE code, rtx x, rtx y) | |
476 | { | |
477 | enum machine_mode mode = SELECT_CC_MODE (code, x, y); | |
478 | rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM); | |
479 | ||
480 | emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y)); | |
481 | return cc_reg; | |
482 | } | |
483 | ||
484 | /* Build the SYMBOL_REF for __tls_get_addr. */ | |
485 | ||
486 | static GTY(()) rtx tls_get_addr_libfunc; | |
487 | ||
488 | rtx | |
489 | aarch64_tls_get_addr (void) | |
490 | { | |
491 | if (!tls_get_addr_libfunc) | |
492 | tls_get_addr_libfunc = init_one_libfunc ("__tls_get_addr"); | |
493 | return tls_get_addr_libfunc; | |
494 | } | |
495 | ||
496 | /* Return the TLS model to use for ADDR. */ | |
497 | ||
498 | static enum tls_model | |
499 | tls_symbolic_operand_type (rtx addr) | |
500 | { | |
501 | enum tls_model tls_kind = TLS_MODEL_NONE; | |
502 | rtx sym, addend; | |
503 | ||
504 | if (GET_CODE (addr) == CONST) | |
505 | { | |
506 | split_const (addr, &sym, &addend); | |
507 | if (GET_CODE (sym) == SYMBOL_REF) | |
508 | tls_kind = SYMBOL_REF_TLS_MODEL (sym); | |
509 | } | |
510 | else if (GET_CODE (addr) == SYMBOL_REF) | |
511 | tls_kind = SYMBOL_REF_TLS_MODEL (addr); | |
512 | ||
513 | return tls_kind; | |
514 | } | |
515 | ||
516 | /* We'll allow lo_sum's in addresses in our legitimate addresses | |
517 | so that combine would take care of combining addresses where | |
518 | necessary, but for generation purposes, we'll generate the address | |
519 | as : | |
520 | RTL Absolute | |
521 | tmp = hi (symbol_ref); adrp x1, foo | |
522 | dest = lo_sum (tmp, symbol_ref); add dest, x1, :lo_12:foo | |
523 | nop | |
524 | ||
525 | PIC TLS | |
526 | adrp x1, :got:foo adrp tmp, :tlsgd:foo | |
527 | ldr x1, [:got_lo12:foo] add dest, tmp, :tlsgd_lo12:foo | |
528 | bl __tls_get_addr | |
529 | nop | |
530 | ||
531 | Load TLS symbol, depending on TLS mechanism and TLS access model. | |
532 | ||
533 | Global Dynamic - Traditional TLS: | |
534 | adrp tmp, :tlsgd:imm | |
535 | add dest, tmp, #:tlsgd_lo12:imm | |
536 | bl __tls_get_addr | |
537 | ||
538 | Global Dynamic - TLS Descriptors: | |
539 | adrp dest, :tlsdesc:imm | |
540 | ldr tmp, [dest, #:tlsdesc_lo12:imm] | |
541 | add dest, dest, #:tlsdesc_lo12:imm | |
542 | blr tmp | |
543 | mrs tp, tpidr_el0 | |
544 | add dest, dest, tp | |
545 | ||
546 | Initial Exec: | |
547 | mrs tp, tpidr_el0 | |
548 | adrp tmp, :gottprel:imm | |
549 | ldr dest, [tmp, #:gottprel_lo12:imm] | |
550 | add dest, dest, tp | |
551 | ||
552 | Local Exec: | |
553 | mrs tp, tpidr_el0 | |
554 | add t0, tp, #:tprel_hi12:imm | |
555 | add t0, #:tprel_lo12_nc:imm | |
556 | */ | |
557 | ||
558 | static void | |
559 | aarch64_load_symref_appropriately (rtx dest, rtx imm, | |
560 | enum aarch64_symbol_type type) | |
561 | { | |
562 | switch (type) | |
563 | { | |
564 | case SYMBOL_SMALL_ABSOLUTE: | |
565 | { | |
28514dda | 566 | /* In ILP32, the mode of dest can be either SImode or DImode. */ |
43e9d192 | 567 | rtx tmp_reg = dest; |
28514dda YZ |
568 | enum machine_mode mode = GET_MODE (dest); |
569 | ||
570 | gcc_assert (mode == Pmode || mode == ptr_mode); | |
571 | ||
43e9d192 | 572 | if (can_create_pseudo_p ()) |
28514dda | 573 | tmp_reg = gen_reg_rtx (mode); |
43e9d192 | 574 | |
28514dda | 575 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); |
43e9d192 IB |
576 | emit_insn (gen_add_losym (dest, tmp_reg, imm)); |
577 | return; | |
578 | } | |
579 | ||
a5350ddc CSS |
580 | case SYMBOL_TINY_ABSOLUTE: |
581 | emit_insn (gen_rtx_SET (Pmode, dest, imm)); | |
582 | return; | |
583 | ||
43e9d192 IB |
584 | case SYMBOL_SMALL_GOT: |
585 | { | |
28514dda YZ |
586 | /* In ILP32, the mode of dest can be either SImode or DImode, |
587 | while the got entry is always of SImode size. The mode of | |
588 | dest depends on how dest is used: if dest is assigned to a | |
589 | pointer (e.g. in the memory), it has SImode; it may have | |
590 | DImode if dest is dereferenced to access the memeory. | |
591 | This is why we have to handle three different ldr_got_small | |
592 | patterns here (two patterns for ILP32). */ | |
43e9d192 | 593 | rtx tmp_reg = dest; |
28514dda YZ |
594 | enum machine_mode mode = GET_MODE (dest); |
595 | ||
43e9d192 | 596 | if (can_create_pseudo_p ()) |
28514dda YZ |
597 | tmp_reg = gen_reg_rtx (mode); |
598 | ||
599 | emit_move_insn (tmp_reg, gen_rtx_HIGH (mode, imm)); | |
600 | if (mode == ptr_mode) | |
601 | { | |
602 | if (mode == DImode) | |
603 | emit_insn (gen_ldr_got_small_di (dest, tmp_reg, imm)); | |
604 | else | |
605 | emit_insn (gen_ldr_got_small_si (dest, tmp_reg, imm)); | |
606 | } | |
607 | else | |
608 | { | |
609 | gcc_assert (mode == Pmode); | |
610 | emit_insn (gen_ldr_got_small_sidi (dest, tmp_reg, imm)); | |
611 | } | |
612 | ||
43e9d192 IB |
613 | return; |
614 | } | |
615 | ||
616 | case SYMBOL_SMALL_TLSGD: | |
617 | { | |
618 | rtx insns; | |
619 | rtx result = gen_rtx_REG (Pmode, R0_REGNUM); | |
620 | ||
621 | start_sequence (); | |
622 | emit_call_insn (gen_tlsgd_small (result, imm)); | |
623 | insns = get_insns (); | |
624 | end_sequence (); | |
625 | ||
626 | RTL_CONST_CALL_P (insns) = 1; | |
627 | emit_libcall_block (insns, dest, result, imm); | |
628 | return; | |
629 | } | |
630 | ||
631 | case SYMBOL_SMALL_TLSDESC: | |
632 | { | |
633 | rtx x0 = gen_rtx_REG (Pmode, R0_REGNUM); | |
634 | rtx tp; | |
635 | ||
636 | emit_insn (gen_tlsdesc_small (imm)); | |
637 | tp = aarch64_load_tp (NULL); | |
638 | emit_insn (gen_rtx_SET (Pmode, dest, gen_rtx_PLUS (Pmode, tp, x0))); | |
639 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
640 | return; | |
641 | } | |
642 | ||
643 | case SYMBOL_SMALL_GOTTPREL: | |
644 | { | |
645 | rtx tmp_reg = gen_reg_rtx (Pmode); | |
646 | rtx tp = aarch64_load_tp (NULL); | |
647 | emit_insn (gen_tlsie_small (tmp_reg, imm)); | |
648 | emit_insn (gen_rtx_SET (Pmode, dest, gen_rtx_PLUS (Pmode, tp, tmp_reg))); | |
649 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
650 | return; | |
651 | } | |
652 | ||
653 | case SYMBOL_SMALL_TPREL: | |
654 | { | |
655 | rtx tp = aarch64_load_tp (NULL); | |
656 | emit_insn (gen_tlsle_small (dest, tp, imm)); | |
657 | set_unique_reg_note (get_last_insn (), REG_EQUIV, imm); | |
658 | return; | |
659 | } | |
660 | ||
87dd8ab0 MS |
661 | case SYMBOL_TINY_GOT: |
662 | emit_insn (gen_ldr_got_tiny (dest, imm)); | |
663 | return; | |
664 | ||
43e9d192 IB |
665 | default: |
666 | gcc_unreachable (); | |
667 | } | |
668 | } | |
669 | ||
670 | /* Emit a move from SRC to DEST. Assume that the move expanders can | |
671 | handle all moves if !can_create_pseudo_p (). The distinction is | |
672 | important because, unlike emit_move_insn, the move expanders know | |
673 | how to force Pmode objects into the constant pool even when the | |
674 | constant pool address is not itself legitimate. */ | |
675 | static rtx | |
676 | aarch64_emit_move (rtx dest, rtx src) | |
677 | { | |
678 | return (can_create_pseudo_p () | |
679 | ? emit_move_insn (dest, src) | |
680 | : emit_move_insn_1 (dest, src)); | |
681 | } | |
682 | ||
030d03b8 RE |
683 | /* Split a 128-bit move operation into two 64-bit move operations, |
684 | taking care to handle partial overlap of register to register | |
685 | copies. Special cases are needed when moving between GP regs and | |
686 | FP regs. SRC can be a register, constant or memory; DST a register | |
687 | or memory. If either operand is memory it must not have any side | |
688 | effects. */ | |
43e9d192 IB |
689 | void |
690 | aarch64_split_128bit_move (rtx dst, rtx src) | |
691 | { | |
030d03b8 RE |
692 | rtx dst_lo, dst_hi; |
693 | rtx src_lo, src_hi; | |
43e9d192 | 694 | |
030d03b8 | 695 | enum machine_mode mode = GET_MODE (dst); |
12dc6974 | 696 | |
030d03b8 RE |
697 | gcc_assert (mode == TImode || mode == TFmode); |
698 | gcc_assert (!(side_effects_p (src) || side_effects_p (dst))); | |
699 | gcc_assert (mode == GET_MODE (src) || GET_MODE (src) == VOIDmode); | |
43e9d192 IB |
700 | |
701 | if (REG_P (dst) && REG_P (src)) | |
702 | { | |
030d03b8 RE |
703 | int src_regno = REGNO (src); |
704 | int dst_regno = REGNO (dst); | |
43e9d192 | 705 | |
030d03b8 | 706 | /* Handle FP <-> GP regs. */ |
43e9d192 IB |
707 | if (FP_REGNUM_P (dst_regno) && GP_REGNUM_P (src_regno)) |
708 | { | |
030d03b8 RE |
709 | src_lo = gen_lowpart (word_mode, src); |
710 | src_hi = gen_highpart (word_mode, src); | |
711 | ||
712 | if (mode == TImode) | |
713 | { | |
714 | emit_insn (gen_aarch64_movtilow_di (dst, src_lo)); | |
715 | emit_insn (gen_aarch64_movtihigh_di (dst, src_hi)); | |
716 | } | |
717 | else | |
718 | { | |
719 | emit_insn (gen_aarch64_movtflow_di (dst, src_lo)); | |
720 | emit_insn (gen_aarch64_movtfhigh_di (dst, src_hi)); | |
721 | } | |
722 | return; | |
43e9d192 IB |
723 | } |
724 | else if (GP_REGNUM_P (dst_regno) && FP_REGNUM_P (src_regno)) | |
725 | { | |
030d03b8 RE |
726 | dst_lo = gen_lowpart (word_mode, dst); |
727 | dst_hi = gen_highpart (word_mode, dst); | |
728 | ||
729 | if (mode == TImode) | |
730 | { | |
731 | emit_insn (gen_aarch64_movdi_tilow (dst_lo, src)); | |
732 | emit_insn (gen_aarch64_movdi_tihigh (dst_hi, src)); | |
733 | } | |
734 | else | |
735 | { | |
736 | emit_insn (gen_aarch64_movdi_tflow (dst_lo, src)); | |
737 | emit_insn (gen_aarch64_movdi_tfhigh (dst_hi, src)); | |
738 | } | |
739 | return; | |
43e9d192 | 740 | } |
43e9d192 IB |
741 | } |
742 | ||
030d03b8 RE |
743 | dst_lo = gen_lowpart (word_mode, dst); |
744 | dst_hi = gen_highpart (word_mode, dst); | |
745 | src_lo = gen_lowpart (word_mode, src); | |
746 | src_hi = gen_highpart_mode (word_mode, mode, src); | |
747 | ||
748 | /* At most one pairing may overlap. */ | |
749 | if (reg_overlap_mentioned_p (dst_lo, src_hi)) | |
750 | { | |
751 | aarch64_emit_move (dst_hi, src_hi); | |
752 | aarch64_emit_move (dst_lo, src_lo); | |
753 | } | |
754 | else | |
755 | { | |
756 | aarch64_emit_move (dst_lo, src_lo); | |
757 | aarch64_emit_move (dst_hi, src_hi); | |
758 | } | |
43e9d192 IB |
759 | } |
760 | ||
761 | bool | |
762 | aarch64_split_128bit_move_p (rtx dst, rtx src) | |
763 | { | |
764 | return (! REG_P (src) | |
765 | || ! (FP_REGNUM_P (REGNO (dst)) && FP_REGNUM_P (REGNO (src)))); | |
766 | } | |
767 | ||
8b033a8a SN |
768 | /* Split a complex SIMD combine. */ |
769 | ||
770 | void | |
771 | aarch64_split_simd_combine (rtx dst, rtx src1, rtx src2) | |
772 | { | |
773 | enum machine_mode src_mode = GET_MODE (src1); | |
774 | enum machine_mode dst_mode = GET_MODE (dst); | |
775 | ||
776 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
777 | ||
778 | if (REG_P (dst) && REG_P (src1) && REG_P (src2)) | |
779 | { | |
780 | rtx (*gen) (rtx, rtx, rtx); | |
781 | ||
782 | switch (src_mode) | |
783 | { | |
784 | case V8QImode: | |
785 | gen = gen_aarch64_simd_combinev8qi; | |
786 | break; | |
787 | case V4HImode: | |
788 | gen = gen_aarch64_simd_combinev4hi; | |
789 | break; | |
790 | case V2SImode: | |
791 | gen = gen_aarch64_simd_combinev2si; | |
792 | break; | |
793 | case V2SFmode: | |
794 | gen = gen_aarch64_simd_combinev2sf; | |
795 | break; | |
796 | case DImode: | |
797 | gen = gen_aarch64_simd_combinedi; | |
798 | break; | |
799 | case DFmode: | |
800 | gen = gen_aarch64_simd_combinedf; | |
801 | break; | |
802 | default: | |
803 | gcc_unreachable (); | |
804 | } | |
805 | ||
806 | emit_insn (gen (dst, src1, src2)); | |
807 | return; | |
808 | } | |
809 | } | |
810 | ||
fd4842cd SN |
811 | /* Split a complex SIMD move. */ |
812 | ||
813 | void | |
814 | aarch64_split_simd_move (rtx dst, rtx src) | |
815 | { | |
816 | enum machine_mode src_mode = GET_MODE (src); | |
817 | enum machine_mode dst_mode = GET_MODE (dst); | |
818 | ||
819 | gcc_assert (VECTOR_MODE_P (dst_mode)); | |
820 | ||
821 | if (REG_P (dst) && REG_P (src)) | |
822 | { | |
c59b7e28 SN |
823 | rtx (*gen) (rtx, rtx); |
824 | ||
fd4842cd SN |
825 | gcc_assert (VECTOR_MODE_P (src_mode)); |
826 | ||
827 | switch (src_mode) | |
828 | { | |
829 | case V16QImode: | |
c59b7e28 | 830 | gen = gen_aarch64_split_simd_movv16qi; |
fd4842cd SN |
831 | break; |
832 | case V8HImode: | |
c59b7e28 | 833 | gen = gen_aarch64_split_simd_movv8hi; |
fd4842cd SN |
834 | break; |
835 | case V4SImode: | |
c59b7e28 | 836 | gen = gen_aarch64_split_simd_movv4si; |
fd4842cd SN |
837 | break; |
838 | case V2DImode: | |
c59b7e28 | 839 | gen = gen_aarch64_split_simd_movv2di; |
fd4842cd SN |
840 | break; |
841 | case V4SFmode: | |
c59b7e28 | 842 | gen = gen_aarch64_split_simd_movv4sf; |
fd4842cd SN |
843 | break; |
844 | case V2DFmode: | |
c59b7e28 | 845 | gen = gen_aarch64_split_simd_movv2df; |
fd4842cd SN |
846 | break; |
847 | default: | |
848 | gcc_unreachable (); | |
849 | } | |
c59b7e28 SN |
850 | |
851 | emit_insn (gen (dst, src)); | |
fd4842cd SN |
852 | return; |
853 | } | |
854 | } | |
855 | ||
43e9d192 | 856 | static rtx |
e18b4a81 | 857 | aarch64_force_temporary (enum machine_mode mode, rtx x, rtx value) |
43e9d192 IB |
858 | { |
859 | if (can_create_pseudo_p ()) | |
e18b4a81 | 860 | return force_reg (mode, value); |
43e9d192 IB |
861 | else |
862 | { | |
863 | x = aarch64_emit_move (x, value); | |
864 | return x; | |
865 | } | |
866 | } | |
867 | ||
868 | ||
869 | static rtx | |
870 | aarch64_add_offset (enum machine_mode mode, rtx temp, rtx reg, HOST_WIDE_INT offset) | |
871 | { | |
9c023bf0 | 872 | if (!aarch64_plus_immediate (GEN_INT (offset), mode)) |
43e9d192 IB |
873 | { |
874 | rtx high; | |
875 | /* Load the full offset into a register. This | |
876 | might be improvable in the future. */ | |
877 | high = GEN_INT (offset); | |
878 | offset = 0; | |
e18b4a81 YZ |
879 | high = aarch64_force_temporary (mode, temp, high); |
880 | reg = aarch64_force_temporary (mode, temp, | |
881 | gen_rtx_PLUS (mode, high, reg)); | |
43e9d192 IB |
882 | } |
883 | return plus_constant (mode, reg, offset); | |
884 | } | |
885 | ||
886 | void | |
887 | aarch64_expand_mov_immediate (rtx dest, rtx imm) | |
888 | { | |
889 | enum machine_mode mode = GET_MODE (dest); | |
890 | unsigned HOST_WIDE_INT mask; | |
891 | int i; | |
892 | bool first; | |
893 | unsigned HOST_WIDE_INT val; | |
894 | bool subtargets; | |
895 | rtx subtarget; | |
896 | int one_match, zero_match; | |
897 | ||
898 | gcc_assert (mode == SImode || mode == DImode); | |
899 | ||
900 | /* Check on what type of symbol it is. */ | |
901 | if (GET_CODE (imm) == SYMBOL_REF | |
902 | || GET_CODE (imm) == LABEL_REF | |
903 | || GET_CODE (imm) == CONST) | |
904 | { | |
905 | rtx mem, base, offset; | |
906 | enum aarch64_symbol_type sty; | |
907 | ||
908 | /* If we have (const (plus symbol offset)), separate out the offset | |
909 | before we start classifying the symbol. */ | |
910 | split_const (imm, &base, &offset); | |
911 | ||
912 | sty = aarch64_classify_symbol (base, SYMBOL_CONTEXT_ADR); | |
913 | switch (sty) | |
914 | { | |
915 | case SYMBOL_FORCE_TO_MEM: | |
916 | if (offset != const0_rtx | |
917 | && targetm.cannot_force_const_mem (mode, imm)) | |
918 | { | |
919 | gcc_assert(can_create_pseudo_p ()); | |
e18b4a81 | 920 | base = aarch64_force_temporary (mode, dest, base); |
43e9d192 IB |
921 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); |
922 | aarch64_emit_move (dest, base); | |
923 | return; | |
924 | } | |
28514dda | 925 | mem = force_const_mem (ptr_mode, imm); |
43e9d192 | 926 | gcc_assert (mem); |
28514dda YZ |
927 | if (mode != ptr_mode) |
928 | mem = gen_rtx_ZERO_EXTEND (mode, mem); | |
43e9d192 IB |
929 | emit_insn (gen_rtx_SET (VOIDmode, dest, mem)); |
930 | return; | |
931 | ||
932 | case SYMBOL_SMALL_TLSGD: | |
933 | case SYMBOL_SMALL_TLSDESC: | |
934 | case SYMBOL_SMALL_GOTTPREL: | |
935 | case SYMBOL_SMALL_GOT: | |
87dd8ab0 | 936 | case SYMBOL_TINY_GOT: |
43e9d192 IB |
937 | if (offset != const0_rtx) |
938 | { | |
939 | gcc_assert(can_create_pseudo_p ()); | |
e18b4a81 | 940 | base = aarch64_force_temporary (mode, dest, base); |
43e9d192 IB |
941 | base = aarch64_add_offset (mode, NULL, base, INTVAL (offset)); |
942 | aarch64_emit_move (dest, base); | |
943 | return; | |
944 | } | |
945 | /* FALLTHRU */ | |
946 | ||
947 | case SYMBOL_SMALL_TPREL: | |
948 | case SYMBOL_SMALL_ABSOLUTE: | |
a5350ddc | 949 | case SYMBOL_TINY_ABSOLUTE: |
43e9d192 IB |
950 | aarch64_load_symref_appropriately (dest, imm, sty); |
951 | return; | |
952 | ||
953 | default: | |
954 | gcc_unreachable (); | |
955 | } | |
956 | } | |
957 | ||
958 | if (CONST_INT_P (imm) && aarch64_move_imm (INTVAL (imm), mode)) | |
959 | { | |
960 | emit_insn (gen_rtx_SET (VOIDmode, dest, imm)); | |
961 | return; | |
962 | } | |
963 | ||
964 | if (!CONST_INT_P (imm)) | |
965 | { | |
966 | if (GET_CODE (imm) == HIGH) | |
967 | emit_insn (gen_rtx_SET (VOIDmode, dest, imm)); | |
968 | else | |
969 | { | |
970 | rtx mem = force_const_mem (mode, imm); | |
971 | gcc_assert (mem); | |
972 | emit_insn (gen_rtx_SET (VOIDmode, dest, mem)); | |
973 | } | |
974 | ||
975 | return; | |
976 | } | |
977 | ||
978 | if (mode == SImode) | |
979 | { | |
980 | /* We know we can't do this in 1 insn, and we must be able to do it | |
981 | in two; so don't mess around looking for sequences that don't buy | |
982 | us anything. */ | |
983 | emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_INT (INTVAL (imm) & 0xffff))); | |
984 | emit_insn (gen_insv_immsi (dest, GEN_INT (16), | |
985 | GEN_INT ((INTVAL (imm) >> 16) & 0xffff))); | |
986 | return; | |
987 | } | |
988 | ||
989 | /* Remaining cases are all for DImode. */ | |
990 | ||
991 | val = INTVAL (imm); | |
992 | subtargets = optimize && can_create_pseudo_p (); | |
993 | ||
994 | one_match = 0; | |
995 | zero_match = 0; | |
996 | mask = 0xffff; | |
997 | ||
998 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
999 | { | |
1000 | if ((val & mask) == 0) | |
1001 | zero_match++; | |
1002 | else if ((val & mask) == mask) | |
1003 | one_match++; | |
1004 | } | |
1005 | ||
1006 | if (one_match == 2) | |
1007 | { | |
1008 | mask = 0xffff; | |
1009 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1010 | { | |
1011 | if ((val & mask) != mask) | |
1012 | { | |
1013 | emit_insn (gen_rtx_SET (VOIDmode, dest, GEN_INT (val | mask))); | |
1014 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1015 | GEN_INT ((val >> i) & 0xffff))); | |
1016 | return; | |
1017 | } | |
1018 | } | |
1019 | gcc_unreachable (); | |
1020 | } | |
1021 | ||
1022 | if (zero_match == 2) | |
1023 | goto simple_sequence; | |
1024 | ||
1025 | mask = 0x0ffff0000UL; | |
1026 | for (i = 16; i < 64; i += 16, mask <<= 16) | |
1027 | { | |
1028 | HOST_WIDE_INT comp = mask & ~(mask - 1); | |
1029 | ||
1030 | if (aarch64_uimm12_shift (val - (val & mask))) | |
1031 | { | |
1032 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1033 | ||
1034 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, GEN_INT (val & mask))); | |
1035 | emit_insn (gen_adddi3 (dest, subtarget, | |
1036 | GEN_INT (val - (val & mask)))); | |
1037 | return; | |
1038 | } | |
1039 | else if (aarch64_uimm12_shift (-(val - ((val + comp) & mask)))) | |
1040 | { | |
1041 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1042 | ||
1043 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1044 | GEN_INT ((val + comp) & mask))); | |
1045 | emit_insn (gen_adddi3 (dest, subtarget, | |
1046 | GEN_INT (val - ((val + comp) & mask)))); | |
1047 | return; | |
1048 | } | |
1049 | else if (aarch64_uimm12_shift (val - ((val - comp) | ~mask))) | |
1050 | { | |
1051 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1052 | ||
1053 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1054 | GEN_INT ((val - comp) | ~mask))); | |
1055 | emit_insn (gen_adddi3 (dest, subtarget, | |
1056 | GEN_INT (val - ((val - comp) | ~mask)))); | |
1057 | return; | |
1058 | } | |
1059 | else if (aarch64_uimm12_shift (-(val - (val | ~mask)))) | |
1060 | { | |
1061 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1062 | ||
1063 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1064 | GEN_INT (val | ~mask))); | |
1065 | emit_insn (gen_adddi3 (dest, subtarget, | |
1066 | GEN_INT (val - (val | ~mask)))); | |
1067 | return; | |
1068 | } | |
1069 | } | |
1070 | ||
1071 | /* See if we can do it by arithmetically combining two | |
1072 | immediates. */ | |
1073 | for (i = 0; i < AARCH64_NUM_BITMASKS; i++) | |
1074 | { | |
1075 | int j; | |
1076 | mask = 0xffff; | |
1077 | ||
1078 | if (aarch64_uimm12_shift (val - aarch64_bitmasks[i]) | |
1079 | || aarch64_uimm12_shift (-val + aarch64_bitmasks[i])) | |
1080 | { | |
1081 | subtarget = subtargets ? gen_reg_rtx (DImode) : dest; | |
1082 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1083 | GEN_INT (aarch64_bitmasks[i]))); | |
1084 | emit_insn (gen_adddi3 (dest, subtarget, | |
1085 | GEN_INT (val - aarch64_bitmasks[i]))); | |
1086 | return; | |
1087 | } | |
1088 | ||
1089 | for (j = 0; j < 64; j += 16, mask <<= 16) | |
1090 | { | |
1091 | if ((aarch64_bitmasks[i] & ~mask) == (val & ~mask)) | |
1092 | { | |
1093 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1094 | GEN_INT (aarch64_bitmasks[i]))); | |
1095 | emit_insn (gen_insv_immdi (dest, GEN_INT (j), | |
1096 | GEN_INT ((val >> j) & 0xffff))); | |
1097 | return; | |
1098 | } | |
1099 | } | |
1100 | } | |
1101 | ||
1102 | /* See if we can do it by logically combining two immediates. */ | |
1103 | for (i = 0; i < AARCH64_NUM_BITMASKS; i++) | |
1104 | { | |
1105 | if ((aarch64_bitmasks[i] & val) == aarch64_bitmasks[i]) | |
1106 | { | |
1107 | int j; | |
1108 | ||
1109 | for (j = i + 1; j < AARCH64_NUM_BITMASKS; j++) | |
1110 | if (val == (aarch64_bitmasks[i] | aarch64_bitmasks[j])) | |
1111 | { | |
1112 | subtarget = subtargets ? gen_reg_rtx (mode) : dest; | |
1113 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1114 | GEN_INT (aarch64_bitmasks[i]))); | |
1115 | emit_insn (gen_iordi3 (dest, subtarget, | |
1116 | GEN_INT (aarch64_bitmasks[j]))); | |
1117 | return; | |
1118 | } | |
1119 | } | |
1120 | else if ((val & aarch64_bitmasks[i]) == val) | |
1121 | { | |
1122 | int j; | |
1123 | ||
1124 | for (j = i + 1; j < AARCH64_NUM_BITMASKS; j++) | |
1125 | if (val == (aarch64_bitmasks[j] & aarch64_bitmasks[i])) | |
1126 | { | |
1127 | ||
1128 | subtarget = subtargets ? gen_reg_rtx (mode) : dest; | |
1129 | emit_insn (gen_rtx_SET (VOIDmode, subtarget, | |
1130 | GEN_INT (aarch64_bitmasks[j]))); | |
1131 | emit_insn (gen_anddi3 (dest, subtarget, | |
1132 | GEN_INT (aarch64_bitmasks[i]))); | |
1133 | return; | |
1134 | } | |
1135 | } | |
1136 | } | |
1137 | ||
1138 | simple_sequence: | |
1139 | first = true; | |
1140 | mask = 0xffff; | |
1141 | for (i = 0; i < 64; i += 16, mask <<= 16) | |
1142 | { | |
1143 | if ((val & mask) != 0) | |
1144 | { | |
1145 | if (first) | |
1146 | { | |
1147 | emit_insn (gen_rtx_SET (VOIDmode, dest, | |
1148 | GEN_INT (val & mask))); | |
1149 | first = false; | |
1150 | } | |
1151 | else | |
1152 | emit_insn (gen_insv_immdi (dest, GEN_INT (i), | |
1153 | GEN_INT ((val >> i) & 0xffff))); | |
1154 | } | |
1155 | } | |
1156 | } | |
1157 | ||
1158 | static bool | |
1159 | aarch64_function_ok_for_sibcall (tree decl, tree exp ATTRIBUTE_UNUSED) | |
1160 | { | |
1161 | /* Indirect calls are not currently supported. */ | |
1162 | if (decl == NULL) | |
1163 | return false; | |
1164 | ||
1165 | /* Cannot tail-call to long-calls, since these are outside of the | |
1166 | range of a branch instruction (we could handle this if we added | |
1167 | support for indirect tail-calls. */ | |
1168 | if (aarch64_decl_is_long_call_p (decl)) | |
1169 | return false; | |
1170 | ||
1171 | return true; | |
1172 | } | |
1173 | ||
1174 | /* Implement TARGET_PASS_BY_REFERENCE. */ | |
1175 | ||
1176 | static bool | |
1177 | aarch64_pass_by_reference (cumulative_args_t pcum ATTRIBUTE_UNUSED, | |
1178 | enum machine_mode mode, | |
1179 | const_tree type, | |
1180 | bool named ATTRIBUTE_UNUSED) | |
1181 | { | |
1182 | HOST_WIDE_INT size; | |
1183 | enum machine_mode dummymode; | |
1184 | int nregs; | |
1185 | ||
1186 | /* GET_MODE_SIZE (BLKmode) is useless since it is 0. */ | |
1187 | size = (mode == BLKmode && type) | |
1188 | ? int_size_in_bytes (type) : (int) GET_MODE_SIZE (mode); | |
1189 | ||
1190 | if (type) | |
1191 | { | |
1192 | /* Arrays always passed by reference. */ | |
1193 | if (TREE_CODE (type) == ARRAY_TYPE) | |
1194 | return true; | |
1195 | /* Other aggregates based on their size. */ | |
1196 | if (AGGREGATE_TYPE_P (type)) | |
1197 | size = int_size_in_bytes (type); | |
1198 | } | |
1199 | ||
1200 | /* Variable sized arguments are always returned by reference. */ | |
1201 | if (size < 0) | |
1202 | return true; | |
1203 | ||
1204 | /* Can this be a candidate to be passed in fp/simd register(s)? */ | |
1205 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1206 | &dummymode, &nregs, | |
1207 | NULL)) | |
1208 | return false; | |
1209 | ||
1210 | /* Arguments which are variable sized or larger than 2 registers are | |
1211 | passed by reference unless they are a homogenous floating point | |
1212 | aggregate. */ | |
1213 | return size > 2 * UNITS_PER_WORD; | |
1214 | } | |
1215 | ||
1216 | /* Return TRUE if VALTYPE is padded to its least significant bits. */ | |
1217 | static bool | |
1218 | aarch64_return_in_msb (const_tree valtype) | |
1219 | { | |
1220 | enum machine_mode dummy_mode; | |
1221 | int dummy_int; | |
1222 | ||
1223 | /* Never happens in little-endian mode. */ | |
1224 | if (!BYTES_BIG_ENDIAN) | |
1225 | return false; | |
1226 | ||
1227 | /* Only composite types smaller than or equal to 16 bytes can | |
1228 | be potentially returned in registers. */ | |
1229 | if (!aarch64_composite_type_p (valtype, TYPE_MODE (valtype)) | |
1230 | || int_size_in_bytes (valtype) <= 0 | |
1231 | || int_size_in_bytes (valtype) > 16) | |
1232 | return false; | |
1233 | ||
1234 | /* But not a composite that is an HFA (Homogeneous Floating-point Aggregate) | |
1235 | or an HVA (Homogeneous Short-Vector Aggregate); such a special composite | |
1236 | is always passed/returned in the least significant bits of fp/simd | |
1237 | register(s). */ | |
1238 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (valtype), valtype, | |
1239 | &dummy_mode, &dummy_int, NULL)) | |
1240 | return false; | |
1241 | ||
1242 | return true; | |
1243 | } | |
1244 | ||
1245 | /* Implement TARGET_FUNCTION_VALUE. | |
1246 | Define how to find the value returned by a function. */ | |
1247 | ||
1248 | static rtx | |
1249 | aarch64_function_value (const_tree type, const_tree func, | |
1250 | bool outgoing ATTRIBUTE_UNUSED) | |
1251 | { | |
1252 | enum machine_mode mode; | |
1253 | int unsignedp; | |
1254 | int count; | |
1255 | enum machine_mode ag_mode; | |
1256 | ||
1257 | mode = TYPE_MODE (type); | |
1258 | if (INTEGRAL_TYPE_P (type)) | |
1259 | mode = promote_function_mode (type, mode, &unsignedp, func, 1); | |
1260 | ||
1261 | if (aarch64_return_in_msb (type)) | |
1262 | { | |
1263 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
1264 | ||
1265 | if (size % UNITS_PER_WORD != 0) | |
1266 | { | |
1267 | size += UNITS_PER_WORD - size % UNITS_PER_WORD; | |
1268 | mode = mode_for_size (size * BITS_PER_UNIT, MODE_INT, 0); | |
1269 | } | |
1270 | } | |
1271 | ||
1272 | if (aarch64_vfp_is_call_or_return_candidate (mode, type, | |
1273 | &ag_mode, &count, NULL)) | |
1274 | { | |
1275 | if (!aarch64_composite_type_p (type, mode)) | |
1276 | { | |
1277 | gcc_assert (count == 1 && mode == ag_mode); | |
1278 | return gen_rtx_REG (mode, V0_REGNUM); | |
1279 | } | |
1280 | else | |
1281 | { | |
1282 | int i; | |
1283 | rtx par; | |
1284 | ||
1285 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (count)); | |
1286 | for (i = 0; i < count; i++) | |
1287 | { | |
1288 | rtx tmp = gen_rtx_REG (ag_mode, V0_REGNUM + i); | |
1289 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
1290 | GEN_INT (i * GET_MODE_SIZE (ag_mode))); | |
1291 | XVECEXP (par, 0, i) = tmp; | |
1292 | } | |
1293 | return par; | |
1294 | } | |
1295 | } | |
1296 | else | |
1297 | return gen_rtx_REG (mode, R0_REGNUM); | |
1298 | } | |
1299 | ||
1300 | /* Implements TARGET_FUNCTION_VALUE_REGNO_P. | |
1301 | Return true if REGNO is the number of a hard register in which the values | |
1302 | of called function may come back. */ | |
1303 | ||
1304 | static bool | |
1305 | aarch64_function_value_regno_p (const unsigned int regno) | |
1306 | { | |
1307 | /* Maximum of 16 bytes can be returned in the general registers. Examples | |
1308 | of 16-byte return values are: 128-bit integers and 16-byte small | |
1309 | structures (excluding homogeneous floating-point aggregates). */ | |
1310 | if (regno == R0_REGNUM || regno == R1_REGNUM) | |
1311 | return true; | |
1312 | ||
1313 | /* Up to four fp/simd registers can return a function value, e.g. a | |
1314 | homogeneous floating-point aggregate having four members. */ | |
1315 | if (regno >= V0_REGNUM && regno < V0_REGNUM + HA_MAX_NUM_FLDS) | |
1316 | return !TARGET_GENERAL_REGS_ONLY; | |
1317 | ||
1318 | return false; | |
1319 | } | |
1320 | ||
1321 | /* Implement TARGET_RETURN_IN_MEMORY. | |
1322 | ||
1323 | If the type T of the result of a function is such that | |
1324 | void func (T arg) | |
1325 | would require that arg be passed as a value in a register (or set of | |
1326 | registers) according to the parameter passing rules, then the result | |
1327 | is returned in the same registers as would be used for such an | |
1328 | argument. */ | |
1329 | ||
1330 | static bool | |
1331 | aarch64_return_in_memory (const_tree type, const_tree fndecl ATTRIBUTE_UNUSED) | |
1332 | { | |
1333 | HOST_WIDE_INT size; | |
1334 | enum machine_mode ag_mode; | |
1335 | int count; | |
1336 | ||
1337 | if (!AGGREGATE_TYPE_P (type) | |
1338 | && TREE_CODE (type) != COMPLEX_TYPE | |
1339 | && TREE_CODE (type) != VECTOR_TYPE) | |
1340 | /* Simple scalar types always returned in registers. */ | |
1341 | return false; | |
1342 | ||
1343 | if (aarch64_vfp_is_call_or_return_candidate (TYPE_MODE (type), | |
1344 | type, | |
1345 | &ag_mode, | |
1346 | &count, | |
1347 | NULL)) | |
1348 | return false; | |
1349 | ||
1350 | /* Types larger than 2 registers returned in memory. */ | |
1351 | size = int_size_in_bytes (type); | |
1352 | return (size < 0 || size > 2 * UNITS_PER_WORD); | |
1353 | } | |
1354 | ||
1355 | static bool | |
1356 | aarch64_vfp_is_call_candidate (cumulative_args_t pcum_v, enum machine_mode mode, | |
1357 | const_tree type, int *nregs) | |
1358 | { | |
1359 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1360 | return aarch64_vfp_is_call_or_return_candidate (mode, | |
1361 | type, | |
1362 | &pcum->aapcs_vfp_rmode, | |
1363 | nregs, | |
1364 | NULL); | |
1365 | } | |
1366 | ||
1367 | /* Given MODE and TYPE of a function argument, return the alignment in | |
1368 | bits. The idea is to suppress any stronger alignment requested by | |
1369 | the user and opt for the natural alignment (specified in AAPCS64 \S 4.1). | |
1370 | This is a helper function for local use only. */ | |
1371 | ||
1372 | static unsigned int | |
1373 | aarch64_function_arg_alignment (enum machine_mode mode, const_tree type) | |
1374 | { | |
1375 | unsigned int alignment; | |
1376 | ||
1377 | if (type) | |
1378 | { | |
1379 | if (!integer_zerop (TYPE_SIZE (type))) | |
1380 | { | |
1381 | if (TYPE_MODE (type) == mode) | |
1382 | alignment = TYPE_ALIGN (type); | |
1383 | else | |
1384 | alignment = GET_MODE_ALIGNMENT (mode); | |
1385 | } | |
1386 | else | |
1387 | alignment = 0; | |
1388 | } | |
1389 | else | |
1390 | alignment = GET_MODE_ALIGNMENT (mode); | |
1391 | ||
1392 | return alignment; | |
1393 | } | |
1394 | ||
1395 | /* Layout a function argument according to the AAPCS64 rules. The rule | |
1396 | numbers refer to the rule numbers in the AAPCS64. */ | |
1397 | ||
1398 | static void | |
1399 | aarch64_layout_arg (cumulative_args_t pcum_v, enum machine_mode mode, | |
1400 | const_tree type, | |
1401 | bool named ATTRIBUTE_UNUSED) | |
1402 | { | |
1403 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1404 | int ncrn, nvrn, nregs; | |
1405 | bool allocate_ncrn, allocate_nvrn; | |
1406 | ||
1407 | /* We need to do this once per argument. */ | |
1408 | if (pcum->aapcs_arg_processed) | |
1409 | return; | |
1410 | ||
1411 | pcum->aapcs_arg_processed = true; | |
1412 | ||
1413 | allocate_ncrn = (type) ? !(FLOAT_TYPE_P (type)) : !FLOAT_MODE_P (mode); | |
1414 | allocate_nvrn = aarch64_vfp_is_call_candidate (pcum_v, | |
1415 | mode, | |
1416 | type, | |
1417 | &nregs); | |
1418 | ||
1419 | /* allocate_ncrn may be false-positive, but allocate_nvrn is quite reliable. | |
1420 | The following code thus handles passing by SIMD/FP registers first. */ | |
1421 | ||
1422 | nvrn = pcum->aapcs_nvrn; | |
1423 | ||
1424 | /* C1 - C5 for floating point, homogenous floating point aggregates (HFA) | |
1425 | and homogenous short-vector aggregates (HVA). */ | |
1426 | if (allocate_nvrn) | |
1427 | { | |
1428 | if (nvrn + nregs <= NUM_FP_ARG_REGS) | |
1429 | { | |
1430 | pcum->aapcs_nextnvrn = nvrn + nregs; | |
1431 | if (!aarch64_composite_type_p (type, mode)) | |
1432 | { | |
1433 | gcc_assert (nregs == 1); | |
1434 | pcum->aapcs_reg = gen_rtx_REG (mode, V0_REGNUM + nvrn); | |
1435 | } | |
1436 | else | |
1437 | { | |
1438 | rtx par; | |
1439 | int i; | |
1440 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
1441 | for (i = 0; i < nregs; i++) | |
1442 | { | |
1443 | rtx tmp = gen_rtx_REG (pcum->aapcs_vfp_rmode, | |
1444 | V0_REGNUM + nvrn + i); | |
1445 | tmp = gen_rtx_EXPR_LIST | |
1446 | (VOIDmode, tmp, | |
1447 | GEN_INT (i * GET_MODE_SIZE (pcum->aapcs_vfp_rmode))); | |
1448 | XVECEXP (par, 0, i) = tmp; | |
1449 | } | |
1450 | pcum->aapcs_reg = par; | |
1451 | } | |
1452 | return; | |
1453 | } | |
1454 | else | |
1455 | { | |
1456 | /* C.3 NSRN is set to 8. */ | |
1457 | pcum->aapcs_nextnvrn = NUM_FP_ARG_REGS; | |
1458 | goto on_stack; | |
1459 | } | |
1460 | } | |
1461 | ||
1462 | ncrn = pcum->aapcs_ncrn; | |
1463 | nregs = ((type ? int_size_in_bytes (type) : GET_MODE_SIZE (mode)) | |
1464 | + UNITS_PER_WORD - 1) / UNITS_PER_WORD; | |
1465 | ||
1466 | ||
1467 | /* C6 - C9. though the sign and zero extension semantics are | |
1468 | handled elsewhere. This is the case where the argument fits | |
1469 | entirely general registers. */ | |
1470 | if (allocate_ncrn && (ncrn + nregs <= NUM_ARG_REGS)) | |
1471 | { | |
1472 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
1473 | ||
1474 | gcc_assert (nregs == 0 || nregs == 1 || nregs == 2); | |
1475 | ||
1476 | /* C.8 if the argument has an alignment of 16 then the NGRN is | |
1477 | rounded up to the next even number. */ | |
1478 | if (nregs == 2 && alignment == 16 * BITS_PER_UNIT && ncrn % 2) | |
1479 | { | |
1480 | ++ncrn; | |
1481 | gcc_assert (ncrn + nregs <= NUM_ARG_REGS); | |
1482 | } | |
1483 | /* NREGS can be 0 when e.g. an empty structure is to be passed. | |
1484 | A reg is still generated for it, but the caller should be smart | |
1485 | enough not to use it. */ | |
1486 | if (nregs == 0 || nregs == 1 || GET_MODE_CLASS (mode) == MODE_INT) | |
1487 | { | |
1488 | pcum->aapcs_reg = gen_rtx_REG (mode, R0_REGNUM + ncrn); | |
1489 | } | |
1490 | else | |
1491 | { | |
1492 | rtx par; | |
1493 | int i; | |
1494 | ||
1495 | par = gen_rtx_PARALLEL (mode, rtvec_alloc (nregs)); | |
1496 | for (i = 0; i < nregs; i++) | |
1497 | { | |
1498 | rtx tmp = gen_rtx_REG (word_mode, R0_REGNUM + ncrn + i); | |
1499 | tmp = gen_rtx_EXPR_LIST (VOIDmode, tmp, | |
1500 | GEN_INT (i * UNITS_PER_WORD)); | |
1501 | XVECEXP (par, 0, i) = tmp; | |
1502 | } | |
1503 | pcum->aapcs_reg = par; | |
1504 | } | |
1505 | ||
1506 | pcum->aapcs_nextncrn = ncrn + nregs; | |
1507 | return; | |
1508 | } | |
1509 | ||
1510 | /* C.11 */ | |
1511 | pcum->aapcs_nextncrn = NUM_ARG_REGS; | |
1512 | ||
1513 | /* The argument is passed on stack; record the needed number of words for | |
1514 | this argument (we can re-use NREGS) and align the total size if | |
1515 | necessary. */ | |
1516 | on_stack: | |
1517 | pcum->aapcs_stack_words = nregs; | |
1518 | if (aarch64_function_arg_alignment (mode, type) == 16 * BITS_PER_UNIT) | |
1519 | pcum->aapcs_stack_size = AARCH64_ROUND_UP (pcum->aapcs_stack_size, | |
1520 | 16 / UNITS_PER_WORD) + 1; | |
1521 | return; | |
1522 | } | |
1523 | ||
1524 | /* Implement TARGET_FUNCTION_ARG. */ | |
1525 | ||
1526 | static rtx | |
1527 | aarch64_function_arg (cumulative_args_t pcum_v, enum machine_mode mode, | |
1528 | const_tree type, bool named) | |
1529 | { | |
1530 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1531 | gcc_assert (pcum->pcs_variant == ARM_PCS_AAPCS64); | |
1532 | ||
1533 | if (mode == VOIDmode) | |
1534 | return NULL_RTX; | |
1535 | ||
1536 | aarch64_layout_arg (pcum_v, mode, type, named); | |
1537 | return pcum->aapcs_reg; | |
1538 | } | |
1539 | ||
1540 | void | |
1541 | aarch64_init_cumulative_args (CUMULATIVE_ARGS *pcum, | |
1542 | const_tree fntype ATTRIBUTE_UNUSED, | |
1543 | rtx libname ATTRIBUTE_UNUSED, | |
1544 | const_tree fndecl ATTRIBUTE_UNUSED, | |
1545 | unsigned n_named ATTRIBUTE_UNUSED) | |
1546 | { | |
1547 | pcum->aapcs_ncrn = 0; | |
1548 | pcum->aapcs_nvrn = 0; | |
1549 | pcum->aapcs_nextncrn = 0; | |
1550 | pcum->aapcs_nextnvrn = 0; | |
1551 | pcum->pcs_variant = ARM_PCS_AAPCS64; | |
1552 | pcum->aapcs_reg = NULL_RTX; | |
1553 | pcum->aapcs_arg_processed = false; | |
1554 | pcum->aapcs_stack_words = 0; | |
1555 | pcum->aapcs_stack_size = 0; | |
1556 | ||
1557 | return; | |
1558 | } | |
1559 | ||
1560 | static void | |
1561 | aarch64_function_arg_advance (cumulative_args_t pcum_v, | |
1562 | enum machine_mode mode, | |
1563 | const_tree type, | |
1564 | bool named) | |
1565 | { | |
1566 | CUMULATIVE_ARGS *pcum = get_cumulative_args (pcum_v); | |
1567 | if (pcum->pcs_variant == ARM_PCS_AAPCS64) | |
1568 | { | |
1569 | aarch64_layout_arg (pcum_v, mode, type, named); | |
1570 | gcc_assert ((pcum->aapcs_reg != NULL_RTX) | |
1571 | != (pcum->aapcs_stack_words != 0)); | |
1572 | pcum->aapcs_arg_processed = false; | |
1573 | pcum->aapcs_ncrn = pcum->aapcs_nextncrn; | |
1574 | pcum->aapcs_nvrn = pcum->aapcs_nextnvrn; | |
1575 | pcum->aapcs_stack_size += pcum->aapcs_stack_words; | |
1576 | pcum->aapcs_stack_words = 0; | |
1577 | pcum->aapcs_reg = NULL_RTX; | |
1578 | } | |
1579 | } | |
1580 | ||
1581 | bool | |
1582 | aarch64_function_arg_regno_p (unsigned regno) | |
1583 | { | |
1584 | return ((GP_REGNUM_P (regno) && regno < R0_REGNUM + NUM_ARG_REGS) | |
1585 | || (FP_REGNUM_P (regno) && regno < V0_REGNUM + NUM_FP_ARG_REGS)); | |
1586 | } | |
1587 | ||
1588 | /* Implement FUNCTION_ARG_BOUNDARY. Every parameter gets at least | |
1589 | PARM_BOUNDARY bits of alignment, but will be given anything up | |
1590 | to STACK_BOUNDARY bits if the type requires it. This makes sure | |
1591 | that both before and after the layout of each argument, the Next | |
1592 | Stacked Argument Address (NSAA) will have a minimum alignment of | |
1593 | 8 bytes. */ | |
1594 | ||
1595 | static unsigned int | |
1596 | aarch64_function_arg_boundary (enum machine_mode mode, const_tree type) | |
1597 | { | |
1598 | unsigned int alignment = aarch64_function_arg_alignment (mode, type); | |
1599 | ||
1600 | if (alignment < PARM_BOUNDARY) | |
1601 | alignment = PARM_BOUNDARY; | |
1602 | if (alignment > STACK_BOUNDARY) | |
1603 | alignment = STACK_BOUNDARY; | |
1604 | return alignment; | |
1605 | } | |
1606 | ||
1607 | /* For use by FUNCTION_ARG_PADDING (MODE, TYPE). | |
1608 | ||
1609 | Return true if an argument passed on the stack should be padded upwards, | |
1610 | i.e. if the least-significant byte of the stack slot has useful data. | |
1611 | ||
1612 | Small aggregate types are placed in the lowest memory address. | |
1613 | ||
1614 | The related parameter passing rules are B.4, C.3, C.5 and C.14. */ | |
1615 | ||
1616 | bool | |
1617 | aarch64_pad_arg_upward (enum machine_mode mode, const_tree type) | |
1618 | { | |
1619 | /* On little-endian targets, the least significant byte of every stack | |
1620 | argument is passed at the lowest byte address of the stack slot. */ | |
1621 | if (!BYTES_BIG_ENDIAN) | |
1622 | return true; | |
1623 | ||
00edcfbe | 1624 | /* Otherwise, integral, floating-point and pointer types are padded downward: |
43e9d192 IB |
1625 | the least significant byte of a stack argument is passed at the highest |
1626 | byte address of the stack slot. */ | |
1627 | if (type | |
00edcfbe YZ |
1628 | ? (INTEGRAL_TYPE_P (type) || SCALAR_FLOAT_TYPE_P (type) |
1629 | || POINTER_TYPE_P (type)) | |
43e9d192 IB |
1630 | : (SCALAR_INT_MODE_P (mode) || SCALAR_FLOAT_MODE_P (mode))) |
1631 | return false; | |
1632 | ||
1633 | /* Everything else padded upward, i.e. data in first byte of stack slot. */ | |
1634 | return true; | |
1635 | } | |
1636 | ||
1637 | /* Similarly, for use by BLOCK_REG_PADDING (MODE, TYPE, FIRST). | |
1638 | ||
1639 | It specifies padding for the last (may also be the only) | |
1640 | element of a block move between registers and memory. If | |
1641 | assuming the block is in the memory, padding upward means that | |
1642 | the last element is padded after its highest significant byte, | |
1643 | while in downward padding, the last element is padded at the | |
1644 | its least significant byte side. | |
1645 | ||
1646 | Small aggregates and small complex types are always padded | |
1647 | upwards. | |
1648 | ||
1649 | We don't need to worry about homogeneous floating-point or | |
1650 | short-vector aggregates; their move is not affected by the | |
1651 | padding direction determined here. Regardless of endianness, | |
1652 | each element of such an aggregate is put in the least | |
1653 | significant bits of a fp/simd register. | |
1654 | ||
1655 | Return !BYTES_BIG_ENDIAN if the least significant byte of the | |
1656 | register has useful data, and return the opposite if the most | |
1657 | significant byte does. */ | |
1658 | ||
1659 | bool | |
1660 | aarch64_pad_reg_upward (enum machine_mode mode, const_tree type, | |
1661 | bool first ATTRIBUTE_UNUSED) | |
1662 | { | |
1663 | ||
1664 | /* Small composite types are always padded upward. */ | |
1665 | if (BYTES_BIG_ENDIAN && aarch64_composite_type_p (type, mode)) | |
1666 | { | |
1667 | HOST_WIDE_INT size = (type ? int_size_in_bytes (type) | |
1668 | : GET_MODE_SIZE (mode)); | |
1669 | if (size < 2 * UNITS_PER_WORD) | |
1670 | return true; | |
1671 | } | |
1672 | ||
1673 | /* Otherwise, use the default padding. */ | |
1674 | return !BYTES_BIG_ENDIAN; | |
1675 | } | |
1676 | ||
1677 | static enum machine_mode | |
1678 | aarch64_libgcc_cmp_return_mode (void) | |
1679 | { | |
1680 | return SImode; | |
1681 | } | |
1682 | ||
1683 | static bool | |
1684 | aarch64_frame_pointer_required (void) | |
1685 | { | |
1686 | /* If the function contains dynamic stack allocations, we need to | |
1687 | use the frame pointer to access the static parts of the frame. */ | |
1688 | if (cfun->calls_alloca) | |
1689 | return true; | |
1690 | ||
1691 | /* We may have turned flag_omit_frame_pointer on in order to have this | |
1692 | function called; if we did, we also set the 'faked_omit_frame_pointer' flag | |
1693 | and we'll check it here. | |
1694 | If we really did set flag_omit_frame_pointer normally, then we return false | |
1695 | (no frame pointer required) in all cases. */ | |
1696 | ||
1697 | if (flag_omit_frame_pointer && !faked_omit_frame_pointer) | |
1698 | return false; | |
1699 | else if (flag_omit_leaf_frame_pointer) | |
77436791 | 1700 | return !crtl->is_leaf || df_regs_ever_live_p (LR_REGNUM); |
43e9d192 IB |
1701 | return true; |
1702 | } | |
1703 | ||
1704 | /* Mark the registers that need to be saved by the callee and calculate | |
1705 | the size of the callee-saved registers area and frame record (both FP | |
1706 | and LR may be omitted). */ | |
1707 | static void | |
1708 | aarch64_layout_frame (void) | |
1709 | { | |
1710 | HOST_WIDE_INT offset = 0; | |
1711 | int regno; | |
1712 | ||
1713 | if (reload_completed && cfun->machine->frame.laid_out) | |
1714 | return; | |
1715 | ||
1716 | cfun->machine->frame.fp_lr_offset = 0; | |
1717 | ||
1718 | /* First mark all the registers that really need to be saved... */ | |
1719 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
1720 | cfun->machine->frame.reg_offset[regno] = -1; | |
1721 | ||
1722 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1723 | cfun->machine->frame.reg_offset[regno] = -1; | |
1724 | ||
1725 | /* ... that includes the eh data registers (if needed)... */ | |
1726 | if (crtl->calls_eh_return) | |
1727 | for (regno = 0; EH_RETURN_DATA_REGNO (regno) != INVALID_REGNUM; regno++) | |
1728 | cfun->machine->frame.reg_offset[EH_RETURN_DATA_REGNO (regno)] = 0; | |
1729 | ||
1730 | /* ... and any callee saved register that dataflow says is live. */ | |
1731 | for (regno = R0_REGNUM; regno <= R30_REGNUM; regno++) | |
1732 | if (df_regs_ever_live_p (regno) | |
1733 | && !call_used_regs[regno]) | |
1734 | cfun->machine->frame.reg_offset[regno] = 0; | |
1735 | ||
1736 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1737 | if (df_regs_ever_live_p (regno) | |
1738 | && !call_used_regs[regno]) | |
1739 | cfun->machine->frame.reg_offset[regno] = 0; | |
1740 | ||
1741 | if (frame_pointer_needed) | |
1742 | { | |
1743 | cfun->machine->frame.reg_offset[R30_REGNUM] = 0; | |
1744 | cfun->machine->frame.reg_offset[R29_REGNUM] = 0; | |
1745 | cfun->machine->frame.hardfp_offset = 2 * UNITS_PER_WORD; | |
1746 | } | |
1747 | ||
1748 | /* Now assign stack slots for them. */ | |
1749 | for (regno = R0_REGNUM; regno <= R28_REGNUM; regno++) | |
1750 | if (cfun->machine->frame.reg_offset[regno] != -1) | |
1751 | { | |
1752 | cfun->machine->frame.reg_offset[regno] = offset; | |
1753 | offset += UNITS_PER_WORD; | |
1754 | } | |
1755 | ||
1756 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1757 | if (cfun->machine->frame.reg_offset[regno] != -1) | |
1758 | { | |
1759 | cfun->machine->frame.reg_offset[regno] = offset; | |
1760 | offset += UNITS_PER_WORD; | |
1761 | } | |
1762 | ||
1763 | if (frame_pointer_needed) | |
1764 | { | |
1765 | cfun->machine->frame.reg_offset[R29_REGNUM] = offset; | |
1766 | offset += UNITS_PER_WORD; | |
1767 | cfun->machine->frame.fp_lr_offset = UNITS_PER_WORD; | |
1768 | } | |
1769 | ||
1770 | if (cfun->machine->frame.reg_offset[R30_REGNUM] != -1) | |
1771 | { | |
1772 | cfun->machine->frame.reg_offset[R30_REGNUM] = offset; | |
1773 | offset += UNITS_PER_WORD; | |
1774 | cfun->machine->frame.fp_lr_offset += UNITS_PER_WORD; | |
1775 | } | |
1776 | ||
1777 | cfun->machine->frame.padding0 = | |
1778 | (AARCH64_ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT) - offset); | |
1779 | offset = AARCH64_ROUND_UP (offset, STACK_BOUNDARY / BITS_PER_UNIT); | |
1780 | ||
1781 | cfun->machine->frame.saved_regs_size = offset; | |
1782 | cfun->machine->frame.laid_out = true; | |
1783 | } | |
1784 | ||
1785 | /* Make the last instruction frame-related and note that it performs | |
1786 | the operation described by FRAME_PATTERN. */ | |
1787 | ||
1788 | static void | |
1789 | aarch64_set_frame_expr (rtx frame_pattern) | |
1790 | { | |
1791 | rtx insn; | |
1792 | ||
1793 | insn = get_last_insn (); | |
1794 | RTX_FRAME_RELATED_P (insn) = 1; | |
1795 | RTX_FRAME_RELATED_P (frame_pattern) = 1; | |
1796 | REG_NOTES (insn) = alloc_EXPR_LIST (REG_FRAME_RELATED_EXPR, | |
1797 | frame_pattern, | |
1798 | REG_NOTES (insn)); | |
1799 | } | |
1800 | ||
1801 | static bool | |
1802 | aarch64_register_saved_on_entry (int regno) | |
1803 | { | |
1804 | return cfun->machine->frame.reg_offset[regno] != -1; | |
1805 | } | |
1806 | ||
1807 | ||
1808 | static void | |
1809 | aarch64_save_or_restore_fprs (int start_offset, int increment, | |
1810 | bool restore, rtx base_rtx) | |
1811 | ||
1812 | { | |
1813 | unsigned regno; | |
1814 | unsigned regno2; | |
1815 | rtx insn; | |
e0f396bc MS |
1816 | rtx (*gen_mem_ref)(enum machine_mode, rtx) |
1817 | = (frame_pointer_needed)? gen_frame_mem : gen_rtx_MEM; | |
43e9d192 IB |
1818 | |
1819 | ||
1820 | for (regno = V0_REGNUM; regno <= V31_REGNUM; regno++) | |
1821 | { | |
1822 | if (aarch64_register_saved_on_entry (regno)) | |
1823 | { | |
1824 | rtx mem; | |
1825 | mem = gen_mem_ref (DFmode, | |
1826 | plus_constant (Pmode, | |
1827 | base_rtx, | |
1828 | start_offset)); | |
1829 | ||
1830 | for (regno2 = regno + 1; | |
1831 | regno2 <= V31_REGNUM | |
1832 | && !aarch64_register_saved_on_entry (regno2); | |
1833 | regno2++) | |
1834 | { | |
1835 | /* Empty loop. */ | |
1836 | } | |
1837 | if (regno2 <= V31_REGNUM && | |
1838 | aarch64_register_saved_on_entry (regno2)) | |
1839 | { | |
1840 | rtx mem2; | |
1841 | /* Next highest register to be saved. */ | |
1842 | mem2 = gen_mem_ref (DFmode, | |
1843 | plus_constant | |
1844 | (Pmode, | |
1845 | base_rtx, | |
1846 | start_offset + increment)); | |
1847 | if (restore == false) | |
1848 | { | |
1849 | insn = emit_insn | |
1850 | ( gen_store_pairdf (mem, gen_rtx_REG (DFmode, regno), | |
1851 | mem2, gen_rtx_REG (DFmode, regno2))); | |
1852 | ||
1853 | } | |
1854 | else | |
1855 | { | |
1856 | insn = emit_insn | |
1857 | ( gen_load_pairdf (gen_rtx_REG (DFmode, regno), mem, | |
1858 | gen_rtx_REG (DFmode, regno2), mem2)); | |
1859 | ||
e0f396bc MS |
1860 | add_reg_note (insn, REG_CFA_RESTORE, |
1861 | gen_rtx_REG (DFmode, regno)); | |
1862 | add_reg_note (insn, REG_CFA_RESTORE, | |
1863 | gen_rtx_REG (DFmode, regno2)); | |
43e9d192 IB |
1864 | } |
1865 | ||
1866 | /* The first part of a frame-related parallel insn | |
1867 | is always assumed to be relevant to the frame | |
1868 | calculations; subsequent parts, are only | |
1869 | frame-related if explicitly marked. */ | |
e0f396bc | 1870 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; |
43e9d192 IB |
1871 | regno = regno2; |
1872 | start_offset += increment * 2; | |
1873 | } | |
1874 | else | |
1875 | { | |
1876 | if (restore == false) | |
1877 | insn = emit_move_insn (mem, gen_rtx_REG (DFmode, regno)); | |
1878 | else | |
1879 | { | |
1880 | insn = emit_move_insn (gen_rtx_REG (DFmode, regno), mem); | |
e0f396bc MS |
1881 | add_reg_note (insn, REG_CFA_RESTORE, |
1882 | gen_rtx_REG (DImode, regno)); | |
43e9d192 IB |
1883 | } |
1884 | start_offset += increment; | |
1885 | } | |
1886 | RTX_FRAME_RELATED_P (insn) = 1; | |
1887 | } | |
1888 | } | |
1889 | ||
1890 | } | |
1891 | ||
1892 | ||
1893 | /* offset from the stack pointer of where the saves and | |
1894 | restore's have to happen. */ | |
1895 | static void | |
1896 | aarch64_save_or_restore_callee_save_registers (HOST_WIDE_INT offset, | |
1897 | bool restore) | |
1898 | { | |
1899 | rtx insn; | |
1900 | rtx base_rtx = stack_pointer_rtx; | |
1901 | HOST_WIDE_INT start_offset = offset; | |
1902 | HOST_WIDE_INT increment = UNITS_PER_WORD; | |
1903 | rtx (*gen_mem_ref)(enum machine_mode, rtx) = (frame_pointer_needed)? gen_frame_mem : gen_rtx_MEM; | |
1904 | unsigned limit = (frame_pointer_needed)? R28_REGNUM: R30_REGNUM; | |
1905 | unsigned regno; | |
1906 | unsigned regno2; | |
1907 | ||
1908 | for (regno = R0_REGNUM; regno <= limit; regno++) | |
1909 | { | |
1910 | if (aarch64_register_saved_on_entry (regno)) | |
1911 | { | |
1912 | rtx mem; | |
1913 | mem = gen_mem_ref (Pmode, | |
1914 | plus_constant (Pmode, | |
1915 | base_rtx, | |
1916 | start_offset)); | |
1917 | ||
1918 | for (regno2 = regno + 1; | |
1919 | regno2 <= limit | |
1920 | && !aarch64_register_saved_on_entry (regno2); | |
1921 | regno2++) | |
1922 | { | |
1923 | /* Empty loop. */ | |
1924 | } | |
1925 | if (regno2 <= limit && | |
1926 | aarch64_register_saved_on_entry (regno2)) | |
1927 | { | |
1928 | rtx mem2; | |
1929 | /* Next highest register to be saved. */ | |
1930 | mem2 = gen_mem_ref (Pmode, | |
1931 | plus_constant | |
1932 | (Pmode, | |
1933 | base_rtx, | |
1934 | start_offset + increment)); | |
1935 | if (restore == false) | |
1936 | { | |
1937 | insn = emit_insn | |
1938 | ( gen_store_pairdi (mem, gen_rtx_REG (DImode, regno), | |
1939 | mem2, gen_rtx_REG (DImode, regno2))); | |
1940 | ||
1941 | } | |
1942 | else | |
1943 | { | |
1944 | insn = emit_insn | |
1945 | ( gen_load_pairdi (gen_rtx_REG (DImode, regno), mem, | |
1946 | gen_rtx_REG (DImode, regno2), mem2)); | |
1947 | ||
1948 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, regno)); | |
1949 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, regno2)); | |
1950 | } | |
1951 | ||
1952 | /* The first part of a frame-related parallel insn | |
1953 | is always assumed to be relevant to the frame | |
1954 | calculations; subsequent parts, are only | |
1955 | frame-related if explicitly marked. */ | |
1956 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, | |
1957 | 1)) = 1; | |
1958 | regno = regno2; | |
1959 | start_offset += increment * 2; | |
1960 | } | |
1961 | else | |
1962 | { | |
1963 | if (restore == false) | |
1964 | insn = emit_move_insn (mem, gen_rtx_REG (DImode, regno)); | |
1965 | else | |
1966 | { | |
1967 | insn = emit_move_insn (gen_rtx_REG (DImode, regno), mem); | |
1968 | add_reg_note (insn, REG_CFA_RESTORE, gen_rtx_REG (DImode, regno)); | |
1969 | } | |
1970 | start_offset += increment; | |
1971 | } | |
1972 | RTX_FRAME_RELATED_P (insn) = 1; | |
1973 | } | |
1974 | } | |
1975 | ||
1976 | aarch64_save_or_restore_fprs (start_offset, increment, restore, base_rtx); | |
1977 | ||
1978 | } | |
1979 | ||
1980 | /* AArch64 stack frames generated by this compiler look like: | |
1981 | ||
1982 | +-------------------------------+ | |
1983 | | | | |
1984 | | incoming stack arguments | | |
1985 | | | | |
1986 | +-------------------------------+ <-- arg_pointer_rtx | |
1987 | | | | |
1988 | | callee-allocated save area | | |
1989 | | for register varargs | | |
1990 | | | | |
1991 | +-------------------------------+ | |
1992 | | | | |
1993 | | local variables | | |
1994 | | | | |
1995 | +-------------------------------+ <-- frame_pointer_rtx | |
1996 | | | | |
1997 | | callee-saved registers | | |
1998 | | | | |
1999 | +-------------------------------+ | |
2000 | | LR' | | |
2001 | +-------------------------------+ | |
2002 | | FP' | | |
2003 | P +-------------------------------+ <-- hard_frame_pointer_rtx | |
2004 | | dynamic allocation | | |
2005 | +-------------------------------+ | |
2006 | | | | |
2007 | | outgoing stack arguments | | |
2008 | | | | |
2009 | +-------------------------------+ <-- stack_pointer_rtx | |
2010 | ||
2011 | Dynamic stack allocations such as alloca insert data at point P. | |
2012 | They decrease stack_pointer_rtx but leave frame_pointer_rtx and | |
2013 | hard_frame_pointer_rtx unchanged. */ | |
2014 | ||
2015 | /* Generate the prologue instructions for entry into a function. | |
2016 | Establish the stack frame by decreasing the stack pointer with a | |
2017 | properly calculated size and, if necessary, create a frame record | |
2018 | filled with the values of LR and previous frame pointer. The | |
6991c977 | 2019 | current FP is also set up if it is in use. */ |
43e9d192 IB |
2020 | |
2021 | void | |
2022 | aarch64_expand_prologue (void) | |
2023 | { | |
2024 | /* sub sp, sp, #<frame_size> | |
2025 | stp {fp, lr}, [sp, #<frame_size> - 16] | |
2026 | add fp, sp, #<frame_size> - hardfp_offset | |
2027 | stp {cs_reg}, [fp, #-16] etc. | |
2028 | ||
2029 | sub sp, sp, <final_adjustment_if_any> | |
2030 | */ | |
2031 | HOST_WIDE_INT original_frame_size; /* local variables + vararg save */ | |
2032 | HOST_WIDE_INT frame_size, offset; | |
2033 | HOST_WIDE_INT fp_offset; /* FP offset from SP */ | |
2034 | rtx insn; | |
2035 | ||
2036 | aarch64_layout_frame (); | |
2037 | original_frame_size = get_frame_size () + cfun->machine->saved_varargs_size; | |
2038 | gcc_assert ((!cfun->machine->saved_varargs_size || cfun->stdarg) | |
2039 | && (cfun->stdarg || !cfun->machine->saved_varargs_size)); | |
2040 | frame_size = (original_frame_size + cfun->machine->frame.saved_regs_size | |
2041 | + crtl->outgoing_args_size); | |
2042 | offset = frame_size = AARCH64_ROUND_UP (frame_size, | |
2043 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2044 | ||
2045 | if (flag_stack_usage_info) | |
2046 | current_function_static_stack_size = frame_size; | |
2047 | ||
2048 | fp_offset = (offset | |
2049 | - original_frame_size | |
2050 | - cfun->machine->frame.saved_regs_size); | |
2051 | ||
44c0e7b9 | 2052 | /* Store pairs and load pairs have a range only -512 to 504. */ |
43e9d192 IB |
2053 | if (offset >= 512) |
2054 | { | |
2055 | /* When the frame has a large size, an initial decrease is done on | |
2056 | the stack pointer to jump over the callee-allocated save area for | |
2057 | register varargs, the local variable area and/or the callee-saved | |
2058 | register area. This will allow the pre-index write-back | |
2059 | store pair instructions to be used for setting up the stack frame | |
2060 | efficiently. */ | |
2061 | offset = original_frame_size + cfun->machine->frame.saved_regs_size; | |
2062 | if (offset >= 512) | |
2063 | offset = cfun->machine->frame.saved_regs_size; | |
2064 | ||
2065 | frame_size -= (offset + crtl->outgoing_args_size); | |
2066 | fp_offset = 0; | |
2067 | ||
2068 | if (frame_size >= 0x1000000) | |
2069 | { | |
2070 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2071 | emit_move_insn (op0, GEN_INT (-frame_size)); | |
2072 | emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); | |
2073 | aarch64_set_frame_expr (gen_rtx_SET | |
2074 | (Pmode, stack_pointer_rtx, | |
f6fe771a RL |
2075 | plus_constant (Pmode, |
2076 | stack_pointer_rtx, | |
2077 | -frame_size))); | |
43e9d192 IB |
2078 | } |
2079 | else if (frame_size > 0) | |
2080 | { | |
2081 | if ((frame_size & 0xfff) != frame_size) | |
2082 | { | |
2083 | insn = emit_insn (gen_add2_insn | |
2084 | (stack_pointer_rtx, | |
2085 | GEN_INT (-(frame_size | |
2086 | & ~(HOST_WIDE_INT)0xfff)))); | |
2087 | RTX_FRAME_RELATED_P (insn) = 1; | |
2088 | } | |
2089 | if ((frame_size & 0xfff) != 0) | |
2090 | { | |
2091 | insn = emit_insn (gen_add2_insn | |
2092 | (stack_pointer_rtx, | |
2093 | GEN_INT (-(frame_size | |
2094 | & (HOST_WIDE_INT)0xfff)))); | |
2095 | RTX_FRAME_RELATED_P (insn) = 1; | |
2096 | } | |
2097 | } | |
2098 | } | |
2099 | else | |
2100 | frame_size = -1; | |
2101 | ||
2102 | if (offset > 0) | |
2103 | { | |
2104 | /* Save the frame pointer and lr if the frame pointer is needed | |
2105 | first. Make the frame pointer point to the location of the | |
2106 | old frame pointer on the stack. */ | |
2107 | if (frame_pointer_needed) | |
2108 | { | |
2109 | rtx mem_fp, mem_lr; | |
2110 | ||
2111 | if (fp_offset) | |
2112 | { | |
2113 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2114 | GEN_INT (-offset))); | |
2115 | RTX_FRAME_RELATED_P (insn) = 1; | |
2116 | aarch64_set_frame_expr (gen_rtx_SET | |
2117 | (Pmode, stack_pointer_rtx, | |
2118 | gen_rtx_MINUS (Pmode, | |
2119 | stack_pointer_rtx, | |
2120 | GEN_INT (offset)))); | |
2121 | mem_fp = gen_frame_mem (DImode, | |
2122 | plus_constant (Pmode, | |
2123 | stack_pointer_rtx, | |
2124 | fp_offset)); | |
2125 | mem_lr = gen_frame_mem (DImode, | |
2126 | plus_constant (Pmode, | |
2127 | stack_pointer_rtx, | |
2128 | fp_offset | |
2129 | + UNITS_PER_WORD)); | |
2130 | insn = emit_insn (gen_store_pairdi (mem_fp, | |
2131 | hard_frame_pointer_rtx, | |
2132 | mem_lr, | |
2133 | gen_rtx_REG (DImode, | |
2134 | LR_REGNUM))); | |
2135 | } | |
2136 | else | |
2137 | { | |
2138 | insn = emit_insn (gen_storewb_pairdi_di | |
2139 | (stack_pointer_rtx, stack_pointer_rtx, | |
2140 | hard_frame_pointer_rtx, | |
2141 | gen_rtx_REG (DImode, LR_REGNUM), | |
2142 | GEN_INT (-offset), | |
2143 | GEN_INT (GET_MODE_SIZE (DImode) - offset))); | |
2144 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
2145 | } | |
2146 | ||
2147 | /* The first part of a frame-related parallel insn is always | |
2148 | assumed to be relevant to the frame calculations; | |
2149 | subsequent parts, are only frame-related if explicitly | |
2150 | marked. */ | |
2151 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
2152 | RTX_FRAME_RELATED_P (insn) = 1; | |
2153 | ||
2154 | /* Set up frame pointer to point to the location of the | |
2155 | previous frame pointer on the stack. */ | |
2156 | insn = emit_insn (gen_add3_insn (hard_frame_pointer_rtx, | |
2157 | stack_pointer_rtx, | |
2158 | GEN_INT (fp_offset))); | |
2159 | aarch64_set_frame_expr (gen_rtx_SET | |
2160 | (Pmode, hard_frame_pointer_rtx, | |
f6fe771a RL |
2161 | plus_constant (Pmode, |
2162 | stack_pointer_rtx, | |
2163 | fp_offset))); | |
43e9d192 IB |
2164 | RTX_FRAME_RELATED_P (insn) = 1; |
2165 | insn = emit_insn (gen_stack_tie (stack_pointer_rtx, | |
2166 | hard_frame_pointer_rtx)); | |
2167 | } | |
2168 | else | |
2169 | { | |
2170 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2171 | GEN_INT (-offset))); | |
2172 | RTX_FRAME_RELATED_P (insn) = 1; | |
2173 | } | |
2174 | ||
2175 | aarch64_save_or_restore_callee_save_registers | |
2176 | (fp_offset + cfun->machine->frame.hardfp_offset, 0); | |
2177 | } | |
2178 | ||
2179 | /* when offset >= 512, | |
2180 | sub sp, sp, #<outgoing_args_size> */ | |
2181 | if (frame_size > -1) | |
2182 | { | |
2183 | if (crtl->outgoing_args_size > 0) | |
2184 | { | |
2185 | insn = emit_insn (gen_add2_insn | |
2186 | (stack_pointer_rtx, | |
2187 | GEN_INT (- crtl->outgoing_args_size))); | |
2188 | RTX_FRAME_RELATED_P (insn) = 1; | |
2189 | } | |
2190 | } | |
2191 | } | |
2192 | ||
2193 | /* Generate the epilogue instructions for returning from a function. */ | |
2194 | void | |
2195 | aarch64_expand_epilogue (bool for_sibcall) | |
2196 | { | |
2197 | HOST_WIDE_INT original_frame_size, frame_size, offset; | |
2198 | HOST_WIDE_INT fp_offset; | |
2199 | rtx insn; | |
44c0e7b9 | 2200 | rtx cfa_reg; |
43e9d192 IB |
2201 | |
2202 | aarch64_layout_frame (); | |
2203 | original_frame_size = get_frame_size () + cfun->machine->saved_varargs_size; | |
2204 | frame_size = (original_frame_size + cfun->machine->frame.saved_regs_size | |
2205 | + crtl->outgoing_args_size); | |
2206 | offset = frame_size = AARCH64_ROUND_UP (frame_size, | |
2207 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2208 | ||
2209 | fp_offset = (offset | |
2210 | - original_frame_size | |
2211 | - cfun->machine->frame.saved_regs_size); | |
2212 | ||
44c0e7b9 YZ |
2213 | cfa_reg = frame_pointer_needed ? hard_frame_pointer_rtx : stack_pointer_rtx; |
2214 | ||
2215 | /* Store pairs and load pairs have a range only -512 to 504. */ | |
43e9d192 IB |
2216 | if (offset >= 512) |
2217 | { | |
2218 | offset = original_frame_size + cfun->machine->frame.saved_regs_size; | |
2219 | if (offset >= 512) | |
2220 | offset = cfun->machine->frame.saved_regs_size; | |
2221 | ||
2222 | frame_size -= (offset + crtl->outgoing_args_size); | |
2223 | fp_offset = 0; | |
2224 | if (!frame_pointer_needed && crtl->outgoing_args_size > 0) | |
2225 | { | |
2226 | insn = emit_insn (gen_add2_insn | |
2227 | (stack_pointer_rtx, | |
2228 | GEN_INT (crtl->outgoing_args_size))); | |
2229 | RTX_FRAME_RELATED_P (insn) = 1; | |
2230 | } | |
2231 | } | |
2232 | else | |
2233 | frame_size = -1; | |
2234 | ||
2235 | /* If there were outgoing arguments or we've done dynamic stack | |
2236 | allocation, then restore the stack pointer from the frame | |
2237 | pointer. This is at most one insn and more efficient than using | |
2238 | GCC's internal mechanism. */ | |
2239 | if (frame_pointer_needed | |
2240 | && (crtl->outgoing_args_size || cfun->calls_alloca)) | |
2241 | { | |
2242 | insn = emit_insn (gen_add3_insn (stack_pointer_rtx, | |
2243 | hard_frame_pointer_rtx, | |
2244 | GEN_INT (- fp_offset))); | |
2245 | RTX_FRAME_RELATED_P (insn) = 1; | |
44c0e7b9 YZ |
2246 | /* As SP is set to (FP - fp_offset), according to the rules in |
2247 | dwarf2cfi.c:dwarf2out_frame_debug_expr, CFA should be calculated | |
2248 | from the value of SP from now on. */ | |
2249 | cfa_reg = stack_pointer_rtx; | |
43e9d192 IB |
2250 | } |
2251 | ||
2252 | aarch64_save_or_restore_callee_save_registers | |
2253 | (fp_offset + cfun->machine->frame.hardfp_offset, 1); | |
2254 | ||
2255 | /* Restore the frame pointer and lr if the frame pointer is needed. */ | |
2256 | if (offset > 0) | |
2257 | { | |
2258 | if (frame_pointer_needed) | |
2259 | { | |
2260 | rtx mem_fp, mem_lr; | |
2261 | ||
2262 | if (fp_offset) | |
2263 | { | |
2264 | mem_fp = gen_frame_mem (DImode, | |
2265 | plus_constant (Pmode, | |
2266 | stack_pointer_rtx, | |
2267 | fp_offset)); | |
2268 | mem_lr = gen_frame_mem (DImode, | |
2269 | plus_constant (Pmode, | |
2270 | stack_pointer_rtx, | |
2271 | fp_offset | |
2272 | + UNITS_PER_WORD)); | |
2273 | insn = emit_insn (gen_load_pairdi (hard_frame_pointer_rtx, | |
2274 | mem_fp, | |
2275 | gen_rtx_REG (DImode, | |
2276 | LR_REGNUM), | |
2277 | mem_lr)); | |
2278 | } | |
2279 | else | |
2280 | { | |
2281 | insn = emit_insn (gen_loadwb_pairdi_di | |
2282 | (stack_pointer_rtx, | |
2283 | stack_pointer_rtx, | |
2284 | hard_frame_pointer_rtx, | |
2285 | gen_rtx_REG (DImode, LR_REGNUM), | |
2286 | GEN_INT (offset), | |
2287 | GEN_INT (GET_MODE_SIZE (DImode) + offset))); | |
2288 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 2)) = 1; | |
44c0e7b9 YZ |
2289 | add_reg_note (insn, REG_CFA_ADJUST_CFA, |
2290 | (gen_rtx_SET (Pmode, stack_pointer_rtx, | |
dc2d3c67 YZ |
2291 | plus_constant (Pmode, cfa_reg, |
2292 | offset)))); | |
43e9d192 IB |
2293 | } |
2294 | ||
2295 | /* The first part of a frame-related parallel insn | |
2296 | is always assumed to be relevant to the frame | |
2297 | calculations; subsequent parts, are only | |
2298 | frame-related if explicitly marked. */ | |
2299 | RTX_FRAME_RELATED_P (XVECEXP (PATTERN (insn), 0, 1)) = 1; | |
2300 | RTX_FRAME_RELATED_P (insn) = 1; | |
2301 | add_reg_note (insn, REG_CFA_RESTORE, hard_frame_pointer_rtx); | |
2302 | add_reg_note (insn, REG_CFA_RESTORE, | |
2303 | gen_rtx_REG (DImode, LR_REGNUM)); | |
2304 | ||
2305 | if (fp_offset) | |
2306 | { | |
2307 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2308 | GEN_INT (offset))); | |
2309 | RTX_FRAME_RELATED_P (insn) = 1; | |
2310 | } | |
2311 | } | |
43e9d192 IB |
2312 | else |
2313 | { | |
2314 | insn = emit_insn (gen_add2_insn (stack_pointer_rtx, | |
2315 | GEN_INT (offset))); | |
2316 | RTX_FRAME_RELATED_P (insn) = 1; | |
2317 | } | |
2318 | } | |
2319 | ||
2320 | /* Stack adjustment for exception handler. */ | |
2321 | if (crtl->calls_eh_return) | |
2322 | { | |
2323 | /* We need to unwind the stack by the offset computed by | |
2324 | EH_RETURN_STACKADJ_RTX. However, at this point the CFA is | |
2325 | based on SP. Ideally we would update the SP and define the | |
2326 | CFA along the lines of: | |
2327 | ||
2328 | SP = SP + EH_RETURN_STACKADJ_RTX | |
2329 | (regnote CFA = SP - EH_RETURN_STACKADJ_RTX) | |
2330 | ||
2331 | However the dwarf emitter only understands a constant | |
2332 | register offset. | |
2333 | ||
631b20a7 | 2334 | The solution chosen here is to use the otherwise unused IP0 |
43e9d192 IB |
2335 | as a temporary register to hold the current SP value. The |
2336 | CFA is described using IP0 then SP is modified. */ | |
2337 | ||
2338 | rtx ip0 = gen_rtx_REG (DImode, IP0_REGNUM); | |
2339 | ||
2340 | insn = emit_move_insn (ip0, stack_pointer_rtx); | |
2341 | add_reg_note (insn, REG_CFA_DEF_CFA, ip0); | |
2342 | RTX_FRAME_RELATED_P (insn) = 1; | |
2343 | ||
2344 | emit_insn (gen_add2_insn (stack_pointer_rtx, EH_RETURN_STACKADJ_RTX)); | |
2345 | ||
2346 | /* Ensure the assignment to IP0 does not get optimized away. */ | |
2347 | emit_use (ip0); | |
2348 | } | |
2349 | ||
2350 | if (frame_size > -1) | |
2351 | { | |
2352 | if (frame_size >= 0x1000000) | |
2353 | { | |
2354 | rtx op0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2355 | emit_move_insn (op0, GEN_INT (frame_size)); | |
2356 | emit_insn (gen_add2_insn (stack_pointer_rtx, op0)); | |
2357 | aarch64_set_frame_expr (gen_rtx_SET | |
2358 | (Pmode, stack_pointer_rtx, | |
f6fe771a RL |
2359 | plus_constant (Pmode, |
2360 | stack_pointer_rtx, | |
2361 | frame_size))); | |
43e9d192 IB |
2362 | } |
2363 | else if (frame_size > 0) | |
2364 | { | |
2365 | if ((frame_size & 0xfff) != 0) | |
2366 | { | |
2367 | insn = emit_insn (gen_add2_insn | |
2368 | (stack_pointer_rtx, | |
2369 | GEN_INT ((frame_size | |
2370 | & (HOST_WIDE_INT) 0xfff)))); | |
2371 | RTX_FRAME_RELATED_P (insn) = 1; | |
2372 | } | |
2373 | if ((frame_size & 0xfff) != frame_size) | |
2374 | { | |
2375 | insn = emit_insn (gen_add2_insn | |
2376 | (stack_pointer_rtx, | |
2377 | GEN_INT ((frame_size | |
2378 | & ~ (HOST_WIDE_INT) 0xfff)))); | |
2379 | RTX_FRAME_RELATED_P (insn) = 1; | |
2380 | } | |
2381 | } | |
2382 | ||
f6fe771a RL |
2383 | aarch64_set_frame_expr (gen_rtx_SET (Pmode, stack_pointer_rtx, |
2384 | plus_constant (Pmode, | |
2385 | stack_pointer_rtx, | |
2386 | offset))); | |
43e9d192 IB |
2387 | } |
2388 | ||
2389 | emit_use (gen_rtx_REG (DImode, LR_REGNUM)); | |
2390 | if (!for_sibcall) | |
2391 | emit_jump_insn (ret_rtx); | |
2392 | } | |
2393 | ||
2394 | /* Return the place to copy the exception unwinding return address to. | |
2395 | This will probably be a stack slot, but could (in theory be the | |
2396 | return register). */ | |
2397 | rtx | |
2398 | aarch64_final_eh_return_addr (void) | |
2399 | { | |
2400 | HOST_WIDE_INT original_frame_size, frame_size, offset, fp_offset; | |
2401 | aarch64_layout_frame (); | |
2402 | original_frame_size = get_frame_size () + cfun->machine->saved_varargs_size; | |
2403 | frame_size = (original_frame_size + cfun->machine->frame.saved_regs_size | |
2404 | + crtl->outgoing_args_size); | |
2405 | offset = frame_size = AARCH64_ROUND_UP (frame_size, | |
2406 | STACK_BOUNDARY / BITS_PER_UNIT); | |
2407 | fp_offset = offset | |
2408 | - original_frame_size | |
2409 | - cfun->machine->frame.saved_regs_size; | |
2410 | ||
2411 | if (cfun->machine->frame.reg_offset[LR_REGNUM] < 0) | |
2412 | return gen_rtx_REG (DImode, LR_REGNUM); | |
2413 | ||
2414 | /* DSE and CSELIB do not detect an alias between sp+k1 and fp+k2. This can | |
2415 | result in a store to save LR introduced by builtin_eh_return () being | |
2416 | incorrectly deleted because the alias is not detected. | |
2417 | So in the calculation of the address to copy the exception unwinding | |
2418 | return address to, we note 2 cases. | |
2419 | If FP is needed and the fp_offset is 0, it means that SP = FP and hence | |
2420 | we return a SP-relative location since all the addresses are SP-relative | |
2421 | in this case. This prevents the store from being optimized away. | |
2422 | If the fp_offset is not 0, then the addresses will be FP-relative and | |
2423 | therefore we return a FP-relative location. */ | |
2424 | ||
2425 | if (frame_pointer_needed) | |
2426 | { | |
2427 | if (fp_offset) | |
2428 | return gen_frame_mem (DImode, | |
2429 | plus_constant (Pmode, hard_frame_pointer_rtx, UNITS_PER_WORD)); | |
2430 | else | |
2431 | return gen_frame_mem (DImode, | |
2432 | plus_constant (Pmode, stack_pointer_rtx, UNITS_PER_WORD)); | |
2433 | } | |
2434 | ||
2435 | /* If FP is not needed, we calculate the location of LR, which would be | |
2436 | at the top of the saved registers block. */ | |
2437 | ||
2438 | return gen_frame_mem (DImode, | |
2439 | plus_constant (Pmode, | |
2440 | stack_pointer_rtx, | |
2441 | fp_offset | |
2442 | + cfun->machine->frame.saved_regs_size | |
2443 | - 2 * UNITS_PER_WORD)); | |
2444 | } | |
2445 | ||
2446 | /* Output code to build up a constant in a register. */ | |
2447 | static void | |
d9600ae5 | 2448 | aarch64_build_constant (int regnum, HOST_WIDE_INT val) |
43e9d192 IB |
2449 | { |
2450 | if (aarch64_bitmask_imm (val, DImode)) | |
d9600ae5 | 2451 | emit_move_insn (gen_rtx_REG (Pmode, regnum), GEN_INT (val)); |
43e9d192 IB |
2452 | else |
2453 | { | |
2454 | int i; | |
2455 | int ncount = 0; | |
2456 | int zcount = 0; | |
2457 | HOST_WIDE_INT valp = val >> 16; | |
2458 | HOST_WIDE_INT valm; | |
2459 | HOST_WIDE_INT tval; | |
2460 | ||
2461 | for (i = 16; i < 64; i += 16) | |
2462 | { | |
2463 | valm = (valp & 0xffff); | |
2464 | ||
2465 | if (valm != 0) | |
2466 | ++ zcount; | |
2467 | ||
2468 | if (valm != 0xffff) | |
2469 | ++ ncount; | |
2470 | ||
2471 | valp >>= 16; | |
2472 | } | |
2473 | ||
2474 | /* zcount contains the number of additional MOVK instructions | |
2475 | required if the constant is built up with an initial MOVZ instruction, | |
2476 | while ncount is the number of MOVK instructions required if starting | |
2477 | with a MOVN instruction. Choose the sequence that yields the fewest | |
2478 | number of instructions, preferring MOVZ instructions when they are both | |
2479 | the same. */ | |
2480 | if (ncount < zcount) | |
2481 | { | |
d9600ae5 | 2482 | emit_move_insn (gen_rtx_REG (Pmode, regnum), |
d103f29b | 2483 | GEN_INT (val | ~(HOST_WIDE_INT) 0xffff)); |
43e9d192 IB |
2484 | tval = 0xffff; |
2485 | } | |
2486 | else | |
2487 | { | |
d9600ae5 SN |
2488 | emit_move_insn (gen_rtx_REG (Pmode, regnum), |
2489 | GEN_INT (val & 0xffff)); | |
43e9d192 IB |
2490 | tval = 0; |
2491 | } | |
2492 | ||
2493 | val >>= 16; | |
2494 | ||
2495 | for (i = 16; i < 64; i += 16) | |
2496 | { | |
2497 | if ((val & 0xffff) != tval) | |
d9600ae5 SN |
2498 | emit_insn (gen_insv_immdi (gen_rtx_REG (Pmode, regnum), |
2499 | GEN_INT (i), GEN_INT (val & 0xffff))); | |
43e9d192 IB |
2500 | val >>= 16; |
2501 | } | |
2502 | } | |
2503 | } | |
2504 | ||
2505 | static void | |
d9600ae5 | 2506 | aarch64_add_constant (int regnum, int scratchreg, HOST_WIDE_INT delta) |
43e9d192 IB |
2507 | { |
2508 | HOST_WIDE_INT mdelta = delta; | |
d9600ae5 SN |
2509 | rtx this_rtx = gen_rtx_REG (Pmode, regnum); |
2510 | rtx scratch_rtx = gen_rtx_REG (Pmode, scratchreg); | |
43e9d192 IB |
2511 | |
2512 | if (mdelta < 0) | |
2513 | mdelta = -mdelta; | |
2514 | ||
2515 | if (mdelta >= 4096 * 4096) | |
2516 | { | |
d9600ae5 SN |
2517 | aarch64_build_constant (scratchreg, delta); |
2518 | emit_insn (gen_add3_insn (this_rtx, this_rtx, scratch_rtx)); | |
43e9d192 IB |
2519 | } |
2520 | else if (mdelta > 0) | |
2521 | { | |
43e9d192 | 2522 | if (mdelta >= 4096) |
d9600ae5 SN |
2523 | { |
2524 | emit_insn (gen_rtx_SET (Pmode, scratch_rtx, GEN_INT (mdelta / 4096))); | |
2525 | rtx shift = gen_rtx_ASHIFT (Pmode, scratch_rtx, GEN_INT (12)); | |
2526 | if (delta < 0) | |
2527 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2528 | gen_rtx_MINUS (Pmode, this_rtx, shift))); | |
2529 | else | |
2530 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2531 | gen_rtx_PLUS (Pmode, this_rtx, shift))); | |
2532 | } | |
43e9d192 | 2533 | if (mdelta % 4096 != 0) |
d9600ae5 SN |
2534 | { |
2535 | scratch_rtx = GEN_INT ((delta < 0 ? -1 : 1) * (mdelta % 4096)); | |
2536 | emit_insn (gen_rtx_SET (Pmode, this_rtx, | |
2537 | gen_rtx_PLUS (Pmode, this_rtx, scratch_rtx))); | |
2538 | } | |
43e9d192 IB |
2539 | } |
2540 | } | |
2541 | ||
2542 | /* Output code to add DELTA to the first argument, and then jump | |
2543 | to FUNCTION. Used for C++ multiple inheritance. */ | |
2544 | static void | |
2545 | aarch64_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
2546 | HOST_WIDE_INT delta, | |
2547 | HOST_WIDE_INT vcall_offset, | |
2548 | tree function) | |
2549 | { | |
2550 | /* The this pointer is always in x0. Note that this differs from | |
2551 | Arm where the this pointer maybe bumped to r1 if r0 is required | |
2552 | to return a pointer to an aggregate. On AArch64 a result value | |
2553 | pointer will be in x8. */ | |
2554 | int this_regno = R0_REGNUM; | |
75f1d6fc | 2555 | rtx this_rtx, temp0, temp1, addr, insn, funexp; |
43e9d192 | 2556 | |
75f1d6fc SN |
2557 | reload_completed = 1; |
2558 | emit_note (NOTE_INSN_PROLOGUE_END); | |
43e9d192 IB |
2559 | |
2560 | if (vcall_offset == 0) | |
d9600ae5 | 2561 | aarch64_add_constant (this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
2562 | else |
2563 | { | |
28514dda | 2564 | gcc_assert ((vcall_offset & (POINTER_BYTES - 1)) == 0); |
43e9d192 | 2565 | |
75f1d6fc SN |
2566 | this_rtx = gen_rtx_REG (Pmode, this_regno); |
2567 | temp0 = gen_rtx_REG (Pmode, IP0_REGNUM); | |
2568 | temp1 = gen_rtx_REG (Pmode, IP1_REGNUM); | |
43e9d192 | 2569 | |
75f1d6fc SN |
2570 | addr = this_rtx; |
2571 | if (delta != 0) | |
2572 | { | |
2573 | if (delta >= -256 && delta < 256) | |
2574 | addr = gen_rtx_PRE_MODIFY (Pmode, this_rtx, | |
2575 | plus_constant (Pmode, this_rtx, delta)); | |
2576 | else | |
d9600ae5 | 2577 | aarch64_add_constant (this_regno, IP1_REGNUM, delta); |
43e9d192 IB |
2578 | } |
2579 | ||
28514dda YZ |
2580 | if (Pmode == ptr_mode) |
2581 | aarch64_emit_move (temp0, gen_rtx_MEM (ptr_mode, addr)); | |
2582 | else | |
2583 | aarch64_emit_move (temp0, | |
2584 | gen_rtx_ZERO_EXTEND (Pmode, | |
2585 | gen_rtx_MEM (ptr_mode, addr))); | |
75f1d6fc | 2586 | |
28514dda | 2587 | if (vcall_offset >= -256 && vcall_offset < 4096 * POINTER_BYTES) |
75f1d6fc | 2588 | addr = plus_constant (Pmode, temp0, vcall_offset); |
43e9d192 IB |
2589 | else |
2590 | { | |
d9600ae5 | 2591 | aarch64_build_constant (IP1_REGNUM, vcall_offset); |
75f1d6fc | 2592 | addr = gen_rtx_PLUS (Pmode, temp0, temp1); |
43e9d192 IB |
2593 | } |
2594 | ||
28514dda YZ |
2595 | if (Pmode == ptr_mode) |
2596 | aarch64_emit_move (temp1, gen_rtx_MEM (ptr_mode,addr)); | |
2597 | else | |
2598 | aarch64_emit_move (temp1, | |
2599 | gen_rtx_SIGN_EXTEND (Pmode, | |
2600 | gen_rtx_MEM (ptr_mode, addr))); | |
2601 | ||
75f1d6fc | 2602 | emit_insn (gen_add2_insn (this_rtx, temp1)); |
43e9d192 IB |
2603 | } |
2604 | ||
75f1d6fc SN |
2605 | /* Generate a tail call to the target function. */ |
2606 | if (!TREE_USED (function)) | |
2607 | { | |
2608 | assemble_external (function); | |
2609 | TREE_USED (function) = 1; | |
2610 | } | |
2611 | funexp = XEXP (DECL_RTL (function), 0); | |
2612 | funexp = gen_rtx_MEM (FUNCTION_MODE, funexp); | |
2613 | insn = emit_call_insn (gen_sibcall (funexp, const0_rtx, NULL_RTX)); | |
2614 | SIBLING_CALL_P (insn) = 1; | |
2615 | ||
2616 | insn = get_insns (); | |
2617 | shorten_branches (insn); | |
2618 | final_start_function (insn, file, 1); | |
2619 | final (insn, file, 1); | |
43e9d192 | 2620 | final_end_function (); |
75f1d6fc SN |
2621 | |
2622 | /* Stop pretending to be a post-reload pass. */ | |
2623 | reload_completed = 0; | |
43e9d192 IB |
2624 | } |
2625 | ||
43e9d192 IB |
2626 | static int |
2627 | aarch64_tls_operand_p_1 (rtx *x, void *data ATTRIBUTE_UNUSED) | |
2628 | { | |
2629 | if (GET_CODE (*x) == SYMBOL_REF) | |
2630 | return SYMBOL_REF_TLS_MODEL (*x) != 0; | |
2631 | ||
2632 | /* Don't recurse into UNSPEC_TLS looking for TLS symbols; these are | |
2633 | TLS offsets, not real symbol references. */ | |
2634 | if (GET_CODE (*x) == UNSPEC | |
2635 | && XINT (*x, 1) == UNSPEC_TLS) | |
2636 | return -1; | |
2637 | ||
2638 | return 0; | |
2639 | } | |
2640 | ||
2641 | static bool | |
2642 | aarch64_tls_referenced_p (rtx x) | |
2643 | { | |
2644 | if (!TARGET_HAVE_TLS) | |
2645 | return false; | |
2646 | ||
2647 | return for_each_rtx (&x, aarch64_tls_operand_p_1, NULL); | |
2648 | } | |
2649 | ||
2650 | ||
2651 | static int | |
2652 | aarch64_bitmasks_cmp (const void *i1, const void *i2) | |
2653 | { | |
2654 | const unsigned HOST_WIDE_INT *imm1 = (const unsigned HOST_WIDE_INT *) i1; | |
2655 | const unsigned HOST_WIDE_INT *imm2 = (const unsigned HOST_WIDE_INT *) i2; | |
2656 | ||
2657 | if (*imm1 < *imm2) | |
2658 | return -1; | |
2659 | if (*imm1 > *imm2) | |
2660 | return +1; | |
2661 | return 0; | |
2662 | } | |
2663 | ||
2664 | ||
2665 | static void | |
2666 | aarch64_build_bitmask_table (void) | |
2667 | { | |
2668 | unsigned HOST_WIDE_INT mask, imm; | |
2669 | unsigned int log_e, e, s, r; | |
2670 | unsigned int nimms = 0; | |
2671 | ||
2672 | for (log_e = 1; log_e <= 6; log_e++) | |
2673 | { | |
2674 | e = 1 << log_e; | |
2675 | if (e == 64) | |
2676 | mask = ~(HOST_WIDE_INT) 0; | |
2677 | else | |
2678 | mask = ((HOST_WIDE_INT) 1 << e) - 1; | |
2679 | for (s = 1; s < e; s++) | |
2680 | { | |
2681 | for (r = 0; r < e; r++) | |
2682 | { | |
2683 | /* set s consecutive bits to 1 (s < 64) */ | |
2684 | imm = ((unsigned HOST_WIDE_INT)1 << s) - 1; | |
2685 | /* rotate right by r */ | |
2686 | if (r != 0) | |
2687 | imm = ((imm >> r) | (imm << (e - r))) & mask; | |
2688 | /* replicate the constant depending on SIMD size */ | |
2689 | switch (log_e) { | |
2690 | case 1: imm |= (imm << 2); | |
2691 | case 2: imm |= (imm << 4); | |
2692 | case 3: imm |= (imm << 8); | |
2693 | case 4: imm |= (imm << 16); | |
2694 | case 5: imm |= (imm << 32); | |
2695 | case 6: | |
2696 | break; | |
2697 | default: | |
2698 | gcc_unreachable (); | |
2699 | } | |
2700 | gcc_assert (nimms < AARCH64_NUM_BITMASKS); | |
2701 | aarch64_bitmasks[nimms++] = imm; | |
2702 | } | |
2703 | } | |
2704 | } | |
2705 | ||
2706 | gcc_assert (nimms == AARCH64_NUM_BITMASKS); | |
2707 | qsort (aarch64_bitmasks, nimms, sizeof (aarch64_bitmasks[0]), | |
2708 | aarch64_bitmasks_cmp); | |
2709 | } | |
2710 | ||
2711 | ||
2712 | /* Return true if val can be encoded as a 12-bit unsigned immediate with | |
2713 | a left shift of 0 or 12 bits. */ | |
2714 | bool | |
2715 | aarch64_uimm12_shift (HOST_WIDE_INT val) | |
2716 | { | |
2717 | return ((val & (((HOST_WIDE_INT) 0xfff) << 0)) == val | |
2718 | || (val & (((HOST_WIDE_INT) 0xfff) << 12)) == val | |
2719 | ); | |
2720 | } | |
2721 | ||
2722 | ||
2723 | /* Return true if val is an immediate that can be loaded into a | |
2724 | register by a MOVZ instruction. */ | |
2725 | static bool | |
2726 | aarch64_movw_imm (HOST_WIDE_INT val, enum machine_mode mode) | |
2727 | { | |
2728 | if (GET_MODE_SIZE (mode) > 4) | |
2729 | { | |
2730 | if ((val & (((HOST_WIDE_INT) 0xffff) << 32)) == val | |
2731 | || (val & (((HOST_WIDE_INT) 0xffff) << 48)) == val) | |
2732 | return 1; | |
2733 | } | |
2734 | else | |
2735 | { | |
2736 | /* Ignore sign extension. */ | |
2737 | val &= (HOST_WIDE_INT) 0xffffffff; | |
2738 | } | |
2739 | return ((val & (((HOST_WIDE_INT) 0xffff) << 0)) == val | |
2740 | || (val & (((HOST_WIDE_INT) 0xffff) << 16)) == val); | |
2741 | } | |
2742 | ||
2743 | ||
2744 | /* Return true if val is a valid bitmask immediate. */ | |
2745 | bool | |
2746 | aarch64_bitmask_imm (HOST_WIDE_INT val, enum machine_mode mode) | |
2747 | { | |
2748 | if (GET_MODE_SIZE (mode) < 8) | |
2749 | { | |
2750 | /* Replicate bit pattern. */ | |
2751 | val &= (HOST_WIDE_INT) 0xffffffff; | |
2752 | val |= val << 32; | |
2753 | } | |
2754 | return bsearch (&val, aarch64_bitmasks, AARCH64_NUM_BITMASKS, | |
2755 | sizeof (aarch64_bitmasks[0]), aarch64_bitmasks_cmp) != NULL; | |
2756 | } | |
2757 | ||
2758 | ||
2759 | /* Return true if val is an immediate that can be loaded into a | |
2760 | register in a single instruction. */ | |
2761 | bool | |
2762 | aarch64_move_imm (HOST_WIDE_INT val, enum machine_mode mode) | |
2763 | { | |
2764 | if (aarch64_movw_imm (val, mode) || aarch64_movw_imm (~val, mode)) | |
2765 | return 1; | |
2766 | return aarch64_bitmask_imm (val, mode); | |
2767 | } | |
2768 | ||
2769 | static bool | |
2770 | aarch64_cannot_force_const_mem (enum machine_mode mode ATTRIBUTE_UNUSED, rtx x) | |
2771 | { | |
2772 | rtx base, offset; | |
7eda14e1 | 2773 | |
43e9d192 IB |
2774 | if (GET_CODE (x) == HIGH) |
2775 | return true; | |
2776 | ||
2777 | split_const (x, &base, &offset); | |
2778 | if (GET_CODE (base) == SYMBOL_REF || GET_CODE (base) == LABEL_REF) | |
28514dda YZ |
2779 | { |
2780 | if (aarch64_classify_symbol (base, SYMBOL_CONTEXT_ADR) | |
2781 | != SYMBOL_FORCE_TO_MEM) | |
2782 | return true; | |
2783 | else | |
2784 | /* Avoid generating a 64-bit relocation in ILP32; leave | |
2785 | to aarch64_expand_mov_immediate to handle it properly. */ | |
2786 | return mode != ptr_mode; | |
2787 | } | |
43e9d192 IB |
2788 | |
2789 | return aarch64_tls_referenced_p (x); | |
2790 | } | |
2791 | ||
2792 | /* Return true if register REGNO is a valid index register. | |
2793 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2794 | ||
2795 | bool | |
2796 | aarch64_regno_ok_for_index_p (int regno, bool strict_p) | |
2797 | { | |
2798 | if (!HARD_REGISTER_NUM_P (regno)) | |
2799 | { | |
2800 | if (!strict_p) | |
2801 | return true; | |
2802 | ||
2803 | if (!reg_renumber) | |
2804 | return false; | |
2805 | ||
2806 | regno = reg_renumber[regno]; | |
2807 | } | |
2808 | return GP_REGNUM_P (regno); | |
2809 | } | |
2810 | ||
2811 | /* Return true if register REGNO is a valid base register for mode MODE. | |
2812 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2813 | ||
2814 | bool | |
2815 | aarch64_regno_ok_for_base_p (int regno, bool strict_p) | |
2816 | { | |
2817 | if (!HARD_REGISTER_NUM_P (regno)) | |
2818 | { | |
2819 | if (!strict_p) | |
2820 | return true; | |
2821 | ||
2822 | if (!reg_renumber) | |
2823 | return false; | |
2824 | ||
2825 | regno = reg_renumber[regno]; | |
2826 | } | |
2827 | ||
2828 | /* The fake registers will be eliminated to either the stack or | |
2829 | hard frame pointer, both of which are usually valid base registers. | |
2830 | Reload deals with the cases where the eliminated form isn't valid. */ | |
2831 | return (GP_REGNUM_P (regno) | |
2832 | || regno == SP_REGNUM | |
2833 | || regno == FRAME_POINTER_REGNUM | |
2834 | || regno == ARG_POINTER_REGNUM); | |
2835 | } | |
2836 | ||
2837 | /* Return true if X is a valid base register for mode MODE. | |
2838 | STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2839 | ||
2840 | static bool | |
2841 | aarch64_base_register_rtx_p (rtx x, bool strict_p) | |
2842 | { | |
2843 | if (!strict_p && GET_CODE (x) == SUBREG) | |
2844 | x = SUBREG_REG (x); | |
2845 | ||
2846 | return (REG_P (x) && aarch64_regno_ok_for_base_p (REGNO (x), strict_p)); | |
2847 | } | |
2848 | ||
2849 | /* Return true if address offset is a valid index. If it is, fill in INFO | |
2850 | appropriately. STRICT_P is true if REG_OK_STRICT is in effect. */ | |
2851 | ||
2852 | static bool | |
2853 | aarch64_classify_index (struct aarch64_address_info *info, rtx x, | |
2854 | enum machine_mode mode, bool strict_p) | |
2855 | { | |
2856 | enum aarch64_address_type type; | |
2857 | rtx index; | |
2858 | int shift; | |
2859 | ||
2860 | /* (reg:P) */ | |
2861 | if ((REG_P (x) || GET_CODE (x) == SUBREG) | |
2862 | && GET_MODE (x) == Pmode) | |
2863 | { | |
2864 | type = ADDRESS_REG_REG; | |
2865 | index = x; | |
2866 | shift = 0; | |
2867 | } | |
2868 | /* (sign_extend:DI (reg:SI)) */ | |
2869 | else if ((GET_CODE (x) == SIGN_EXTEND | |
2870 | || GET_CODE (x) == ZERO_EXTEND) | |
2871 | && GET_MODE (x) == DImode | |
2872 | && GET_MODE (XEXP (x, 0)) == SImode) | |
2873 | { | |
2874 | type = (GET_CODE (x) == SIGN_EXTEND) | |
2875 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
2876 | index = XEXP (x, 0); | |
2877 | shift = 0; | |
2878 | } | |
2879 | /* (mult:DI (sign_extend:DI (reg:SI)) (const_int scale)) */ | |
2880 | else if (GET_CODE (x) == MULT | |
2881 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
2882 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
2883 | && GET_MODE (XEXP (x, 0)) == DImode | |
2884 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
2885 | && CONST_INT_P (XEXP (x, 1))) | |
2886 | { | |
2887 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
2888 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
2889 | index = XEXP (XEXP (x, 0), 0); | |
2890 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
2891 | } | |
2892 | /* (ashift:DI (sign_extend:DI (reg:SI)) (const_int shift)) */ | |
2893 | else if (GET_CODE (x) == ASHIFT | |
2894 | && (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND | |
2895 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND) | |
2896 | && GET_MODE (XEXP (x, 0)) == DImode | |
2897 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == SImode | |
2898 | && CONST_INT_P (XEXP (x, 1))) | |
2899 | { | |
2900 | type = (GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
2901 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
2902 | index = XEXP (XEXP (x, 0), 0); | |
2903 | shift = INTVAL (XEXP (x, 1)); | |
2904 | } | |
2905 | /* (sign_extract:DI (mult:DI (reg:DI) (const_int scale)) 32+shift 0) */ | |
2906 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
2907 | || GET_CODE (x) == ZERO_EXTRACT) | |
2908 | && GET_MODE (x) == DImode | |
2909 | && GET_CODE (XEXP (x, 0)) == MULT | |
2910 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
2911 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
2912 | { | |
2913 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
2914 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
2915 | index = XEXP (XEXP (x, 0), 0); | |
2916 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
2917 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
2918 | || INTVAL (XEXP (x, 2)) != 0) | |
2919 | shift = -1; | |
2920 | } | |
2921 | /* (and:DI (mult:DI (reg:DI) (const_int scale)) | |
2922 | (const_int 0xffffffff<<shift)) */ | |
2923 | else if (GET_CODE (x) == AND | |
2924 | && GET_MODE (x) == DImode | |
2925 | && GET_CODE (XEXP (x, 0)) == MULT | |
2926 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
2927 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
2928 | && CONST_INT_P (XEXP (x, 1))) | |
2929 | { | |
2930 | type = ADDRESS_REG_UXTW; | |
2931 | index = XEXP (XEXP (x, 0), 0); | |
2932 | shift = exact_log2 (INTVAL (XEXP (XEXP (x, 0), 1))); | |
2933 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
2934 | shift = -1; | |
2935 | } | |
2936 | /* (sign_extract:DI (ashift:DI (reg:DI) (const_int shift)) 32+shift 0) */ | |
2937 | else if ((GET_CODE (x) == SIGN_EXTRACT | |
2938 | || GET_CODE (x) == ZERO_EXTRACT) | |
2939 | && GET_MODE (x) == DImode | |
2940 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
2941 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
2942 | && CONST_INT_P (XEXP (XEXP (x, 0), 1))) | |
2943 | { | |
2944 | type = (GET_CODE (x) == SIGN_EXTRACT) | |
2945 | ? ADDRESS_REG_SXTW : ADDRESS_REG_UXTW; | |
2946 | index = XEXP (XEXP (x, 0), 0); | |
2947 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
2948 | if (INTVAL (XEXP (x, 1)) != 32 + shift | |
2949 | || INTVAL (XEXP (x, 2)) != 0) | |
2950 | shift = -1; | |
2951 | } | |
2952 | /* (and:DI (ashift:DI (reg:DI) (const_int shift)) | |
2953 | (const_int 0xffffffff<<shift)) */ | |
2954 | else if (GET_CODE (x) == AND | |
2955 | && GET_MODE (x) == DImode | |
2956 | && GET_CODE (XEXP (x, 0)) == ASHIFT | |
2957 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == DImode | |
2958 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
2959 | && CONST_INT_P (XEXP (x, 1))) | |
2960 | { | |
2961 | type = ADDRESS_REG_UXTW; | |
2962 | index = XEXP (XEXP (x, 0), 0); | |
2963 | shift = INTVAL (XEXP (XEXP (x, 0), 1)); | |
2964 | if (INTVAL (XEXP (x, 1)) != (HOST_WIDE_INT)0xffffffff << shift) | |
2965 | shift = -1; | |
2966 | } | |
2967 | /* (mult:P (reg:P) (const_int scale)) */ | |
2968 | else if (GET_CODE (x) == MULT | |
2969 | && GET_MODE (x) == Pmode | |
2970 | && GET_MODE (XEXP (x, 0)) == Pmode | |
2971 | && CONST_INT_P (XEXP (x, 1))) | |
2972 | { | |
2973 | type = ADDRESS_REG_REG; | |
2974 | index = XEXP (x, 0); | |
2975 | shift = exact_log2 (INTVAL (XEXP (x, 1))); | |
2976 | } | |
2977 | /* (ashift:P (reg:P) (const_int shift)) */ | |
2978 | else if (GET_CODE (x) == ASHIFT | |
2979 | && GET_MODE (x) == Pmode | |
2980 | && GET_MODE (XEXP (x, 0)) == Pmode | |
2981 | && CONST_INT_P (XEXP (x, 1))) | |
2982 | { | |
2983 | type = ADDRESS_REG_REG; | |
2984 | index = XEXP (x, 0); | |
2985 | shift = INTVAL (XEXP (x, 1)); | |
2986 | } | |
2987 | else | |
2988 | return false; | |
2989 | ||
2990 | if (GET_CODE (index) == SUBREG) | |
2991 | index = SUBREG_REG (index); | |
2992 | ||
2993 | if ((shift == 0 || | |
2994 | (shift > 0 && shift <= 3 | |
2995 | && (1 << shift) == GET_MODE_SIZE (mode))) | |
2996 | && REG_P (index) | |
2997 | && aarch64_regno_ok_for_index_p (REGNO (index), strict_p)) | |
2998 | { | |
2999 | info->type = type; | |
3000 | info->offset = index; | |
3001 | info->shift = shift; | |
3002 | return true; | |
3003 | } | |
3004 | ||
3005 | return false; | |
3006 | } | |
3007 | ||
3008 | static inline bool | |
3009 | offset_7bit_signed_scaled_p (enum machine_mode mode, HOST_WIDE_INT offset) | |
3010 | { | |
3011 | return (offset >= -64 * GET_MODE_SIZE (mode) | |
3012 | && offset < 64 * GET_MODE_SIZE (mode) | |
3013 | && offset % GET_MODE_SIZE (mode) == 0); | |
3014 | } | |
3015 | ||
3016 | static inline bool | |
3017 | offset_9bit_signed_unscaled_p (enum machine_mode mode ATTRIBUTE_UNUSED, | |
3018 | HOST_WIDE_INT offset) | |
3019 | { | |
3020 | return offset >= -256 && offset < 256; | |
3021 | } | |
3022 | ||
3023 | static inline bool | |
3024 | offset_12bit_unsigned_scaled_p (enum machine_mode mode, HOST_WIDE_INT offset) | |
3025 | { | |
3026 | return (offset >= 0 | |
3027 | && offset < 4096 * GET_MODE_SIZE (mode) | |
3028 | && offset % GET_MODE_SIZE (mode) == 0); | |
3029 | } | |
3030 | ||
3031 | /* Return true if X is a valid address for machine mode MODE. If it is, | |
3032 | fill in INFO appropriately. STRICT_P is true if REG_OK_STRICT is in | |
3033 | effect. OUTER_CODE is PARALLEL for a load/store pair. */ | |
3034 | ||
3035 | static bool | |
3036 | aarch64_classify_address (struct aarch64_address_info *info, | |
3037 | rtx x, enum machine_mode mode, | |
3038 | RTX_CODE outer_code, bool strict_p) | |
3039 | { | |
3040 | enum rtx_code code = GET_CODE (x); | |
3041 | rtx op0, op1; | |
3042 | bool allow_reg_index_p = | |
3043 | outer_code != PARALLEL && GET_MODE_SIZE(mode) != 16; | |
3044 | ||
3045 | /* Don't support anything other than POST_INC or REG addressing for | |
3046 | AdvSIMD. */ | |
3047 | if (aarch64_vector_mode_p (mode) | |
3048 | && (code != POST_INC && code != REG)) | |
3049 | return false; | |
3050 | ||
3051 | switch (code) | |
3052 | { | |
3053 | case REG: | |
3054 | case SUBREG: | |
3055 | info->type = ADDRESS_REG_IMM; | |
3056 | info->base = x; | |
3057 | info->offset = const0_rtx; | |
3058 | return aarch64_base_register_rtx_p (x, strict_p); | |
3059 | ||
3060 | case PLUS: | |
3061 | op0 = XEXP (x, 0); | |
3062 | op1 = XEXP (x, 1); | |
3063 | if (GET_MODE_SIZE (mode) != 0 | |
3064 | && CONST_INT_P (op1) | |
3065 | && aarch64_base_register_rtx_p (op0, strict_p)) | |
3066 | { | |
3067 | HOST_WIDE_INT offset = INTVAL (op1); | |
3068 | ||
3069 | info->type = ADDRESS_REG_IMM; | |
3070 | info->base = op0; | |
3071 | info->offset = op1; | |
3072 | ||
3073 | /* TImode and TFmode values are allowed in both pairs of X | |
3074 | registers and individual Q registers. The available | |
3075 | address modes are: | |
3076 | X,X: 7-bit signed scaled offset | |
3077 | Q: 9-bit signed offset | |
3078 | We conservatively require an offset representable in either mode. | |
3079 | */ | |
3080 | if (mode == TImode || mode == TFmode) | |
3081 | return (offset_7bit_signed_scaled_p (mode, offset) | |
3082 | && offset_9bit_signed_unscaled_p (mode, offset)); | |
3083 | ||
3084 | if (outer_code == PARALLEL) | |
3085 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
3086 | && offset_7bit_signed_scaled_p (mode, offset)); | |
3087 | else | |
3088 | return (offset_9bit_signed_unscaled_p (mode, offset) | |
3089 | || offset_12bit_unsigned_scaled_p (mode, offset)); | |
3090 | } | |
3091 | ||
3092 | if (allow_reg_index_p) | |
3093 | { | |
3094 | /* Look for base + (scaled/extended) index register. */ | |
3095 | if (aarch64_base_register_rtx_p (op0, strict_p) | |
3096 | && aarch64_classify_index (info, op1, mode, strict_p)) | |
3097 | { | |
3098 | info->base = op0; | |
3099 | return true; | |
3100 | } | |
3101 | if (aarch64_base_register_rtx_p (op1, strict_p) | |
3102 | && aarch64_classify_index (info, op0, mode, strict_p)) | |
3103 | { | |
3104 | info->base = op1; | |
3105 | return true; | |
3106 | } | |
3107 | } | |
3108 | ||
3109 | return false; | |
3110 | ||
3111 | case POST_INC: | |
3112 | case POST_DEC: | |
3113 | case PRE_INC: | |
3114 | case PRE_DEC: | |
3115 | info->type = ADDRESS_REG_WB; | |
3116 | info->base = XEXP (x, 0); | |
3117 | info->offset = NULL_RTX; | |
3118 | return aarch64_base_register_rtx_p (info->base, strict_p); | |
3119 | ||
3120 | case POST_MODIFY: | |
3121 | case PRE_MODIFY: | |
3122 | info->type = ADDRESS_REG_WB; | |
3123 | info->base = XEXP (x, 0); | |
3124 | if (GET_CODE (XEXP (x, 1)) == PLUS | |
3125 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) | |
3126 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), info->base) | |
3127 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
3128 | { | |
3129 | HOST_WIDE_INT offset; | |
3130 | info->offset = XEXP (XEXP (x, 1), 1); | |
3131 | offset = INTVAL (info->offset); | |
3132 | ||
3133 | /* TImode and TFmode values are allowed in both pairs of X | |
3134 | registers and individual Q registers. The available | |
3135 | address modes are: | |
3136 | X,X: 7-bit signed scaled offset | |
3137 | Q: 9-bit signed offset | |
3138 | We conservatively require an offset representable in either mode. | |
3139 | */ | |
3140 | if (mode == TImode || mode == TFmode) | |
3141 | return (offset_7bit_signed_scaled_p (mode, offset) | |
3142 | && offset_9bit_signed_unscaled_p (mode, offset)); | |
3143 | ||
3144 | if (outer_code == PARALLEL) | |
3145 | return ((GET_MODE_SIZE (mode) == 4 || GET_MODE_SIZE (mode) == 8) | |
3146 | && offset_7bit_signed_scaled_p (mode, offset)); | |
3147 | else | |
3148 | return offset_9bit_signed_unscaled_p (mode, offset); | |
3149 | } | |
3150 | return false; | |
3151 | ||
3152 | case CONST: | |
3153 | case SYMBOL_REF: | |
3154 | case LABEL_REF: | |
79517551 SN |
3155 | /* load literal: pc-relative constant pool entry. Only supported |
3156 | for SI mode or larger. */ | |
43e9d192 | 3157 | info->type = ADDRESS_SYMBOLIC; |
79517551 | 3158 | if (outer_code != PARALLEL && GET_MODE_SIZE (mode) >= 4) |
43e9d192 IB |
3159 | { |
3160 | rtx sym, addend; | |
3161 | ||
3162 | split_const (x, &sym, &addend); | |
3163 | return (GET_CODE (sym) == LABEL_REF | |
3164 | || (GET_CODE (sym) == SYMBOL_REF | |
3165 | && CONSTANT_POOL_ADDRESS_P (sym))); | |
3166 | } | |
3167 | return false; | |
3168 | ||
3169 | case LO_SUM: | |
3170 | info->type = ADDRESS_LO_SUM; | |
3171 | info->base = XEXP (x, 0); | |
3172 | info->offset = XEXP (x, 1); | |
3173 | if (allow_reg_index_p | |
3174 | && aarch64_base_register_rtx_p (info->base, strict_p)) | |
3175 | { | |
3176 | rtx sym, offs; | |
3177 | split_const (info->offset, &sym, &offs); | |
3178 | if (GET_CODE (sym) == SYMBOL_REF | |
3179 | && (aarch64_classify_symbol (sym, SYMBOL_CONTEXT_MEM) | |
3180 | == SYMBOL_SMALL_ABSOLUTE)) | |
3181 | { | |
3182 | /* The symbol and offset must be aligned to the access size. */ | |
3183 | unsigned int align; | |
3184 | unsigned int ref_size; | |
3185 | ||
3186 | if (CONSTANT_POOL_ADDRESS_P (sym)) | |
3187 | align = GET_MODE_ALIGNMENT (get_pool_mode (sym)); | |
3188 | else if (TREE_CONSTANT_POOL_ADDRESS_P (sym)) | |
3189 | { | |
3190 | tree exp = SYMBOL_REF_DECL (sym); | |
3191 | align = TYPE_ALIGN (TREE_TYPE (exp)); | |
3192 | align = CONSTANT_ALIGNMENT (exp, align); | |
3193 | } | |
3194 | else if (SYMBOL_REF_DECL (sym)) | |
3195 | align = DECL_ALIGN (SYMBOL_REF_DECL (sym)); | |
3196 | else | |
3197 | align = BITS_PER_UNIT; | |
3198 | ||
3199 | ref_size = GET_MODE_SIZE (mode); | |
3200 | if (ref_size == 0) | |
3201 | ref_size = GET_MODE_SIZE (DImode); | |
3202 | ||
3203 | return ((INTVAL (offs) & (ref_size - 1)) == 0 | |
3204 | && ((align / BITS_PER_UNIT) & (ref_size - 1)) == 0); | |
3205 | } | |
3206 | } | |
3207 | return false; | |
3208 | ||
3209 | default: | |
3210 | return false; | |
3211 | } | |
3212 | } | |
3213 | ||
3214 | bool | |
3215 | aarch64_symbolic_address_p (rtx x) | |
3216 | { | |
3217 | rtx offset; | |
3218 | ||
3219 | split_const (x, &x, &offset); | |
3220 | return GET_CODE (x) == SYMBOL_REF || GET_CODE (x) == LABEL_REF; | |
3221 | } | |
3222 | ||
3223 | /* Classify the base of symbolic expression X, given that X appears in | |
3224 | context CONTEXT. */ | |
da4f13a4 MS |
3225 | |
3226 | enum aarch64_symbol_type | |
3227 | aarch64_classify_symbolic_expression (rtx x, | |
3228 | enum aarch64_symbol_context context) | |
43e9d192 IB |
3229 | { |
3230 | rtx offset; | |
da4f13a4 | 3231 | |
43e9d192 IB |
3232 | split_const (x, &x, &offset); |
3233 | return aarch64_classify_symbol (x, context); | |
3234 | } | |
3235 | ||
3236 | ||
3237 | /* Return TRUE if X is a legitimate address for accessing memory in | |
3238 | mode MODE. */ | |
3239 | static bool | |
3240 | aarch64_legitimate_address_hook_p (enum machine_mode mode, rtx x, bool strict_p) | |
3241 | { | |
3242 | struct aarch64_address_info addr; | |
3243 | ||
3244 | return aarch64_classify_address (&addr, x, mode, MEM, strict_p); | |
3245 | } | |
3246 | ||
3247 | /* Return TRUE if X is a legitimate address for accessing memory in | |
3248 | mode MODE. OUTER_CODE will be PARALLEL if this is a load/store | |
3249 | pair operation. */ | |
3250 | bool | |
3251 | aarch64_legitimate_address_p (enum machine_mode mode, rtx x, | |
3252 | RTX_CODE outer_code, bool strict_p) | |
3253 | { | |
3254 | struct aarch64_address_info addr; | |
3255 | ||
3256 | return aarch64_classify_address (&addr, x, mode, outer_code, strict_p); | |
3257 | } | |
3258 | ||
3259 | /* Return TRUE if rtx X is immediate constant 0.0 */ | |
3260 | bool | |
3520f7cc | 3261 | aarch64_float_const_zero_rtx_p (rtx x) |
43e9d192 IB |
3262 | { |
3263 | REAL_VALUE_TYPE r; | |
3264 | ||
3265 | if (GET_MODE (x) == VOIDmode) | |
3266 | return false; | |
3267 | ||
3268 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3269 | if (REAL_VALUE_MINUS_ZERO (r)) | |
3270 | return !HONOR_SIGNED_ZEROS (GET_MODE (x)); | |
3271 | return REAL_VALUES_EQUAL (r, dconst0); | |
3272 | } | |
3273 | ||
70f09188 AP |
3274 | /* Return the fixed registers used for condition codes. */ |
3275 | ||
3276 | static bool | |
3277 | aarch64_fixed_condition_code_regs (unsigned int *p1, unsigned int *p2) | |
3278 | { | |
3279 | *p1 = CC_REGNUM; | |
3280 | *p2 = INVALID_REGNUM; | |
3281 | return true; | |
3282 | } | |
3283 | ||
43e9d192 IB |
3284 | enum machine_mode |
3285 | aarch64_select_cc_mode (RTX_CODE code, rtx x, rtx y) | |
3286 | { | |
3287 | /* All floating point compares return CCFP if it is an equality | |
3288 | comparison, and CCFPE otherwise. */ | |
3289 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_FLOAT) | |
3290 | { | |
3291 | switch (code) | |
3292 | { | |
3293 | case EQ: | |
3294 | case NE: | |
3295 | case UNORDERED: | |
3296 | case ORDERED: | |
3297 | case UNLT: | |
3298 | case UNLE: | |
3299 | case UNGT: | |
3300 | case UNGE: | |
3301 | case UNEQ: | |
3302 | case LTGT: | |
3303 | return CCFPmode; | |
3304 | ||
3305 | case LT: | |
3306 | case LE: | |
3307 | case GT: | |
3308 | case GE: | |
3309 | return CCFPEmode; | |
3310 | ||
3311 | default: | |
3312 | gcc_unreachable (); | |
3313 | } | |
3314 | } | |
3315 | ||
3316 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3317 | && y == const0_rtx | |
3318 | && (code == EQ || code == NE || code == LT || code == GE) | |
b056c910 N |
3319 | && (GET_CODE (x) == PLUS || GET_CODE (x) == MINUS || GET_CODE (x) == AND |
3320 | || GET_CODE (x) == NEG)) | |
43e9d192 IB |
3321 | return CC_NZmode; |
3322 | ||
1c992d1e | 3323 | /* A compare with a shifted operand. Because of canonicalization, |
43e9d192 IB |
3324 | the comparison will have to be swapped when we emit the assembly |
3325 | code. */ | |
3326 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3327 | && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG) | |
3328 | && (GET_CODE (x) == ASHIFT || GET_CODE (x) == ASHIFTRT | |
3329 | || GET_CODE (x) == LSHIFTRT | |
1c992d1e | 3330 | || GET_CODE (x) == ZERO_EXTEND || GET_CODE (x) == SIGN_EXTEND)) |
43e9d192 IB |
3331 | return CC_SWPmode; |
3332 | ||
1c992d1e RE |
3333 | /* Similarly for a negated operand, but we can only do this for |
3334 | equalities. */ | |
3335 | if ((GET_MODE (x) == SImode || GET_MODE (x) == DImode) | |
3336 | && (GET_CODE (y) == REG || GET_CODE (y) == SUBREG) | |
3337 | && (code == EQ || code == NE) | |
3338 | && GET_CODE (x) == NEG) | |
3339 | return CC_Zmode; | |
3340 | ||
43e9d192 IB |
3341 | /* A compare of a mode narrower than SI mode against zero can be done |
3342 | by extending the value in the comparison. */ | |
3343 | if ((GET_MODE (x) == QImode || GET_MODE (x) == HImode) | |
3344 | && y == const0_rtx) | |
3345 | /* Only use sign-extension if we really need it. */ | |
3346 | return ((code == GT || code == GE || code == LE || code == LT) | |
3347 | ? CC_SESWPmode : CC_ZESWPmode); | |
3348 | ||
3349 | /* For everything else, return CCmode. */ | |
3350 | return CCmode; | |
3351 | } | |
3352 | ||
3353 | static unsigned | |
3354 | aarch64_get_condition_code (rtx x) | |
3355 | { | |
3356 | enum machine_mode mode = GET_MODE (XEXP (x, 0)); | |
3357 | enum rtx_code comp_code = GET_CODE (x); | |
3358 | ||
3359 | if (GET_MODE_CLASS (mode) != MODE_CC) | |
3360 | mode = SELECT_CC_MODE (comp_code, XEXP (x, 0), XEXP (x, 1)); | |
3361 | ||
3362 | switch (mode) | |
3363 | { | |
3364 | case CCFPmode: | |
3365 | case CCFPEmode: | |
3366 | switch (comp_code) | |
3367 | { | |
3368 | case GE: return AARCH64_GE; | |
3369 | case GT: return AARCH64_GT; | |
3370 | case LE: return AARCH64_LS; | |
3371 | case LT: return AARCH64_MI; | |
3372 | case NE: return AARCH64_NE; | |
3373 | case EQ: return AARCH64_EQ; | |
3374 | case ORDERED: return AARCH64_VC; | |
3375 | case UNORDERED: return AARCH64_VS; | |
3376 | case UNLT: return AARCH64_LT; | |
3377 | case UNLE: return AARCH64_LE; | |
3378 | case UNGT: return AARCH64_HI; | |
3379 | case UNGE: return AARCH64_PL; | |
3380 | default: gcc_unreachable (); | |
3381 | } | |
3382 | break; | |
3383 | ||
3384 | case CCmode: | |
3385 | switch (comp_code) | |
3386 | { | |
3387 | case NE: return AARCH64_NE; | |
3388 | case EQ: return AARCH64_EQ; | |
3389 | case GE: return AARCH64_GE; | |
3390 | case GT: return AARCH64_GT; | |
3391 | case LE: return AARCH64_LE; | |
3392 | case LT: return AARCH64_LT; | |
3393 | case GEU: return AARCH64_CS; | |
3394 | case GTU: return AARCH64_HI; | |
3395 | case LEU: return AARCH64_LS; | |
3396 | case LTU: return AARCH64_CC; | |
3397 | default: gcc_unreachable (); | |
3398 | } | |
3399 | break; | |
3400 | ||
3401 | case CC_SWPmode: | |
3402 | case CC_ZESWPmode: | |
3403 | case CC_SESWPmode: | |
3404 | switch (comp_code) | |
3405 | { | |
3406 | case NE: return AARCH64_NE; | |
3407 | case EQ: return AARCH64_EQ; | |
3408 | case GE: return AARCH64_LE; | |
3409 | case GT: return AARCH64_LT; | |
3410 | case LE: return AARCH64_GE; | |
3411 | case LT: return AARCH64_GT; | |
3412 | case GEU: return AARCH64_LS; | |
3413 | case GTU: return AARCH64_CC; | |
3414 | case LEU: return AARCH64_CS; | |
3415 | case LTU: return AARCH64_HI; | |
3416 | default: gcc_unreachable (); | |
3417 | } | |
3418 | break; | |
3419 | ||
3420 | case CC_NZmode: | |
3421 | switch (comp_code) | |
3422 | { | |
3423 | case NE: return AARCH64_NE; | |
3424 | case EQ: return AARCH64_EQ; | |
3425 | case GE: return AARCH64_PL; | |
3426 | case LT: return AARCH64_MI; | |
3427 | default: gcc_unreachable (); | |
3428 | } | |
3429 | break; | |
3430 | ||
1c992d1e RE |
3431 | case CC_Zmode: |
3432 | switch (comp_code) | |
3433 | { | |
3434 | case NE: return AARCH64_NE; | |
3435 | case EQ: return AARCH64_EQ; | |
3436 | default: gcc_unreachable (); | |
3437 | } | |
3438 | break; | |
3439 | ||
43e9d192 IB |
3440 | default: |
3441 | gcc_unreachable (); | |
3442 | break; | |
3443 | } | |
3444 | } | |
3445 | ||
3446 | static unsigned | |
3447 | bit_count (unsigned HOST_WIDE_INT value) | |
3448 | { | |
3449 | unsigned count = 0; | |
3450 | ||
3451 | while (value) | |
3452 | { | |
3453 | count++; | |
3454 | value &= value - 1; | |
3455 | } | |
3456 | ||
3457 | return count; | |
3458 | } | |
3459 | ||
3460 | void | |
3461 | aarch64_print_operand (FILE *f, rtx x, char code) | |
3462 | { | |
3463 | switch (code) | |
3464 | { | |
f541a481 KT |
3465 | /* An integer or symbol address without a preceding # sign. */ |
3466 | case 'c': | |
3467 | switch (GET_CODE (x)) | |
3468 | { | |
3469 | case CONST_INT: | |
3470 | fprintf (f, HOST_WIDE_INT_PRINT_DEC, INTVAL (x)); | |
3471 | break; | |
3472 | ||
3473 | case SYMBOL_REF: | |
3474 | output_addr_const (f, x); | |
3475 | break; | |
3476 | ||
3477 | case CONST: | |
3478 | if (GET_CODE (XEXP (x, 0)) == PLUS | |
3479 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF) | |
3480 | { | |
3481 | output_addr_const (f, x); | |
3482 | break; | |
3483 | } | |
3484 | /* Fall through. */ | |
3485 | ||
3486 | default: | |
3487 | output_operand_lossage ("Unsupported operand for code '%c'", code); | |
3488 | } | |
3489 | break; | |
3490 | ||
43e9d192 IB |
3491 | case 'e': |
3492 | /* Print the sign/zero-extend size as a character 8->b, 16->h, 32->w. */ | |
3493 | { | |
3494 | int n; | |
3495 | ||
3496 | if (GET_CODE (x) != CONST_INT | |
3497 | || (n = exact_log2 (INTVAL (x) & ~7)) <= 0) | |
3498 | { | |
3499 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3500 | return; | |
3501 | } | |
3502 | ||
3503 | switch (n) | |
3504 | { | |
3505 | case 3: | |
3506 | fputc ('b', f); | |
3507 | break; | |
3508 | case 4: | |
3509 | fputc ('h', f); | |
3510 | break; | |
3511 | case 5: | |
3512 | fputc ('w', f); | |
3513 | break; | |
3514 | default: | |
3515 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3516 | return; | |
3517 | } | |
3518 | } | |
3519 | break; | |
3520 | ||
3521 | case 'p': | |
3522 | { | |
3523 | int n; | |
3524 | ||
3525 | /* Print N such that 2^N == X. */ | |
3526 | if (GET_CODE (x) != CONST_INT || (n = exact_log2 (INTVAL (x))) < 0) | |
3527 | { | |
3528 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3529 | return; | |
3530 | } | |
3531 | ||
3532 | asm_fprintf (f, "%d", n); | |
3533 | } | |
3534 | break; | |
3535 | ||
3536 | case 'P': | |
3537 | /* Print the number of non-zero bits in X (a const_int). */ | |
3538 | if (GET_CODE (x) != CONST_INT) | |
3539 | { | |
3540 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3541 | return; | |
3542 | } | |
3543 | ||
3544 | asm_fprintf (f, "%u", bit_count (INTVAL (x))); | |
3545 | break; | |
3546 | ||
3547 | case 'H': | |
3548 | /* Print the higher numbered register of a pair (TImode) of regs. */ | |
3549 | if (GET_CODE (x) != REG || !GP_REGNUM_P (REGNO (x) + 1)) | |
3550 | { | |
3551 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3552 | return; | |
3553 | } | |
3554 | ||
01a3a324 | 3555 | asm_fprintf (f, "%s", reg_names [REGNO (x) + 1]); |
43e9d192 IB |
3556 | break; |
3557 | ||
43e9d192 IB |
3558 | case 'm': |
3559 | /* Print a condition (eq, ne, etc). */ | |
3560 | ||
3561 | /* CONST_TRUE_RTX means always -- that's the default. */ | |
3562 | if (x == const_true_rtx) | |
3563 | return; | |
3564 | ||
3565 | if (!COMPARISON_P (x)) | |
3566 | { | |
3567 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3568 | return; | |
3569 | } | |
3570 | ||
3571 | fputs (aarch64_condition_codes[aarch64_get_condition_code (x)], f); | |
3572 | break; | |
3573 | ||
3574 | case 'M': | |
3575 | /* Print the inverse of a condition (eq <-> ne, etc). */ | |
3576 | ||
3577 | /* CONST_TRUE_RTX means never -- that's the default. */ | |
3578 | if (x == const_true_rtx) | |
3579 | { | |
3580 | fputs ("nv", f); | |
3581 | return; | |
3582 | } | |
3583 | ||
3584 | if (!COMPARISON_P (x)) | |
3585 | { | |
3586 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3587 | return; | |
3588 | } | |
3589 | ||
3590 | fputs (aarch64_condition_codes[AARCH64_INVERSE_CONDITION_CODE | |
3591 | (aarch64_get_condition_code (x))], f); | |
3592 | break; | |
3593 | ||
3594 | case 'b': | |
3595 | case 'h': | |
3596 | case 's': | |
3597 | case 'd': | |
3598 | case 'q': | |
3599 | /* Print a scalar FP/SIMD register name. */ | |
3600 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
3601 | { | |
3602 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
3603 | return; | |
3604 | } | |
50ce6f88 | 3605 | asm_fprintf (f, "%c%d", code, REGNO (x) - V0_REGNUM); |
43e9d192 IB |
3606 | break; |
3607 | ||
3608 | case 'S': | |
3609 | case 'T': | |
3610 | case 'U': | |
3611 | case 'V': | |
3612 | /* Print the first FP/SIMD register name in a list. */ | |
3613 | if (!REG_P (x) || !FP_REGNUM_P (REGNO (x))) | |
3614 | { | |
3615 | output_operand_lossage ("incompatible floating point / vector register operand for '%%%c'", code); | |
3616 | return; | |
3617 | } | |
50ce6f88 | 3618 | asm_fprintf (f, "v%d", REGNO (x) - V0_REGNUM + (code - 'S')); |
43e9d192 IB |
3619 | break; |
3620 | ||
a05c0ddf | 3621 | case 'X': |
50d38551 | 3622 | /* Print bottom 16 bits of integer constant in hex. */ |
a05c0ddf IB |
3623 | if (GET_CODE (x) != CONST_INT) |
3624 | { | |
3625 | output_operand_lossage ("invalid operand for '%%%c'", code); | |
3626 | return; | |
3627 | } | |
50d38551 | 3628 | asm_fprintf (f, "0x%wx", UINTVAL (x) & 0xffff); |
a05c0ddf IB |
3629 | break; |
3630 | ||
43e9d192 IB |
3631 | case 'w': |
3632 | case 'x': | |
3633 | /* Print a general register name or the zero register (32-bit or | |
3634 | 64-bit). */ | |
3520f7cc JG |
3635 | if (x == const0_rtx |
3636 | || (CONST_DOUBLE_P (x) && aarch64_float_const_zero_rtx_p (x))) | |
43e9d192 | 3637 | { |
50ce6f88 | 3638 | asm_fprintf (f, "%czr", code); |
43e9d192 IB |
3639 | break; |
3640 | } | |
3641 | ||
3642 | if (REG_P (x) && GP_REGNUM_P (REGNO (x))) | |
3643 | { | |
50ce6f88 | 3644 | asm_fprintf (f, "%c%d", code, REGNO (x) - R0_REGNUM); |
43e9d192 IB |
3645 | break; |
3646 | } | |
3647 | ||
3648 | if (REG_P (x) && REGNO (x) == SP_REGNUM) | |
3649 | { | |
50ce6f88 | 3650 | asm_fprintf (f, "%ssp", code == 'w' ? "w" : ""); |
43e9d192 IB |
3651 | break; |
3652 | } | |
3653 | ||
3654 | /* Fall through */ | |
3655 | ||
3656 | case 0: | |
3657 | /* Print a normal operand, if it's a general register, then we | |
3658 | assume DImode. */ | |
3659 | if (x == NULL) | |
3660 | { | |
3661 | output_operand_lossage ("missing operand"); | |
3662 | return; | |
3663 | } | |
3664 | ||
3665 | switch (GET_CODE (x)) | |
3666 | { | |
3667 | case REG: | |
01a3a324 | 3668 | asm_fprintf (f, "%s", reg_names [REGNO (x)]); |
43e9d192 IB |
3669 | break; |
3670 | ||
3671 | case MEM: | |
3672 | aarch64_memory_reference_mode = GET_MODE (x); | |
3673 | output_address (XEXP (x, 0)); | |
3674 | break; | |
3675 | ||
3676 | case LABEL_REF: | |
3677 | case SYMBOL_REF: | |
3678 | output_addr_const (asm_out_file, x); | |
3679 | break; | |
3680 | ||
3681 | case CONST_INT: | |
3682 | asm_fprintf (f, "%wd", INTVAL (x)); | |
3683 | break; | |
3684 | ||
3685 | case CONST_VECTOR: | |
3520f7cc JG |
3686 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_VECTOR_INT) |
3687 | { | |
3688 | gcc_assert (aarch64_const_vec_all_same_int_p (x, | |
3689 | HOST_WIDE_INT_MIN, | |
3690 | HOST_WIDE_INT_MAX)); | |
3691 | asm_fprintf (f, "%wd", INTVAL (CONST_VECTOR_ELT (x, 0))); | |
3692 | } | |
3693 | else if (aarch64_simd_imm_zero_p (x, GET_MODE (x))) | |
3694 | { | |
3695 | fputc ('0', f); | |
3696 | } | |
3697 | else | |
3698 | gcc_unreachable (); | |
43e9d192 IB |
3699 | break; |
3700 | ||
3520f7cc JG |
3701 | case CONST_DOUBLE: |
3702 | /* CONST_DOUBLE can represent a double-width integer. | |
3703 | In this case, the mode of x is VOIDmode. */ | |
3704 | if (GET_MODE (x) == VOIDmode) | |
3705 | ; /* Do Nothing. */ | |
3706 | else if (aarch64_float_const_zero_rtx_p (x)) | |
3707 | { | |
3708 | fputc ('0', f); | |
3709 | break; | |
3710 | } | |
3711 | else if (aarch64_float_const_representable_p (x)) | |
3712 | { | |
3713 | #define buf_size 20 | |
3714 | char float_buf[buf_size] = {'\0'}; | |
3715 | REAL_VALUE_TYPE r; | |
3716 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
3717 | real_to_decimal_for_mode (float_buf, &r, | |
3718 | buf_size, buf_size, | |
3719 | 1, GET_MODE (x)); | |
3720 | asm_fprintf (asm_out_file, "%s", float_buf); | |
3721 | break; | |
3722 | #undef buf_size | |
3723 | } | |
3724 | output_operand_lossage ("invalid constant"); | |
3725 | return; | |
43e9d192 IB |
3726 | default: |
3727 | output_operand_lossage ("invalid operand"); | |
3728 | return; | |
3729 | } | |
3730 | break; | |
3731 | ||
3732 | case 'A': | |
3733 | if (GET_CODE (x) == HIGH) | |
3734 | x = XEXP (x, 0); | |
3735 | ||
3736 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
3737 | { | |
3738 | case SYMBOL_SMALL_GOT: | |
3739 | asm_fprintf (asm_out_file, ":got:"); | |
3740 | break; | |
3741 | ||
3742 | case SYMBOL_SMALL_TLSGD: | |
3743 | asm_fprintf (asm_out_file, ":tlsgd:"); | |
3744 | break; | |
3745 | ||
3746 | case SYMBOL_SMALL_TLSDESC: | |
3747 | asm_fprintf (asm_out_file, ":tlsdesc:"); | |
3748 | break; | |
3749 | ||
3750 | case SYMBOL_SMALL_GOTTPREL: | |
3751 | asm_fprintf (asm_out_file, ":gottprel:"); | |
3752 | break; | |
3753 | ||
3754 | case SYMBOL_SMALL_TPREL: | |
3755 | asm_fprintf (asm_out_file, ":tprel:"); | |
3756 | break; | |
3757 | ||
87dd8ab0 MS |
3758 | case SYMBOL_TINY_GOT: |
3759 | gcc_unreachable (); | |
3760 | break; | |
3761 | ||
43e9d192 IB |
3762 | default: |
3763 | break; | |
3764 | } | |
3765 | output_addr_const (asm_out_file, x); | |
3766 | break; | |
3767 | ||
3768 | case 'L': | |
3769 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
3770 | { | |
3771 | case SYMBOL_SMALL_GOT: | |
3772 | asm_fprintf (asm_out_file, ":lo12:"); | |
3773 | break; | |
3774 | ||
3775 | case SYMBOL_SMALL_TLSGD: | |
3776 | asm_fprintf (asm_out_file, ":tlsgd_lo12:"); | |
3777 | break; | |
3778 | ||
3779 | case SYMBOL_SMALL_TLSDESC: | |
3780 | asm_fprintf (asm_out_file, ":tlsdesc_lo12:"); | |
3781 | break; | |
3782 | ||
3783 | case SYMBOL_SMALL_GOTTPREL: | |
3784 | asm_fprintf (asm_out_file, ":gottprel_lo12:"); | |
3785 | break; | |
3786 | ||
3787 | case SYMBOL_SMALL_TPREL: | |
3788 | asm_fprintf (asm_out_file, ":tprel_lo12_nc:"); | |
3789 | break; | |
3790 | ||
87dd8ab0 MS |
3791 | case SYMBOL_TINY_GOT: |
3792 | asm_fprintf (asm_out_file, ":got:"); | |
3793 | break; | |
3794 | ||
43e9d192 IB |
3795 | default: |
3796 | break; | |
3797 | } | |
3798 | output_addr_const (asm_out_file, x); | |
3799 | break; | |
3800 | ||
3801 | case 'G': | |
3802 | ||
3803 | switch (aarch64_classify_symbolic_expression (x, SYMBOL_CONTEXT_ADR)) | |
3804 | { | |
3805 | case SYMBOL_SMALL_TPREL: | |
3806 | asm_fprintf (asm_out_file, ":tprel_hi12:"); | |
3807 | break; | |
3808 | default: | |
3809 | break; | |
3810 | } | |
3811 | output_addr_const (asm_out_file, x); | |
3812 | break; | |
3813 | ||
3814 | default: | |
3815 | output_operand_lossage ("invalid operand prefix '%%%c'", code); | |
3816 | return; | |
3817 | } | |
3818 | } | |
3819 | ||
3820 | void | |
3821 | aarch64_print_operand_address (FILE *f, rtx x) | |
3822 | { | |
3823 | struct aarch64_address_info addr; | |
3824 | ||
3825 | if (aarch64_classify_address (&addr, x, aarch64_memory_reference_mode, | |
3826 | MEM, true)) | |
3827 | switch (addr.type) | |
3828 | { | |
3829 | case ADDRESS_REG_IMM: | |
3830 | if (addr.offset == const0_rtx) | |
01a3a324 | 3831 | asm_fprintf (f, "[%s]", reg_names [REGNO (addr.base)]); |
43e9d192 | 3832 | else |
01a3a324 | 3833 | asm_fprintf (f, "[%s,%wd]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3834 | INTVAL (addr.offset)); |
3835 | return; | |
3836 | ||
3837 | case ADDRESS_REG_REG: | |
3838 | if (addr.shift == 0) | |
01a3a324 N |
3839 | asm_fprintf (f, "[%s,%s]", reg_names [REGNO (addr.base)], |
3840 | reg_names [REGNO (addr.offset)]); | |
43e9d192 | 3841 | else |
01a3a324 N |
3842 | asm_fprintf (f, "[%s,%s,lsl %u]", reg_names [REGNO (addr.base)], |
3843 | reg_names [REGNO (addr.offset)], addr.shift); | |
43e9d192 IB |
3844 | return; |
3845 | ||
3846 | case ADDRESS_REG_UXTW: | |
3847 | if (addr.shift == 0) | |
01a3a324 | 3848 | asm_fprintf (f, "[%s,w%d,uxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3849 | REGNO (addr.offset) - R0_REGNUM); |
3850 | else | |
01a3a324 | 3851 | asm_fprintf (f, "[%s,w%d,uxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3852 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
3853 | return; | |
3854 | ||
3855 | case ADDRESS_REG_SXTW: | |
3856 | if (addr.shift == 0) | |
01a3a324 | 3857 | asm_fprintf (f, "[%s,w%d,sxtw]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3858 | REGNO (addr.offset) - R0_REGNUM); |
3859 | else | |
01a3a324 | 3860 | asm_fprintf (f, "[%s,w%d,sxtw %u]", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3861 | REGNO (addr.offset) - R0_REGNUM, addr.shift); |
3862 | return; | |
3863 | ||
3864 | case ADDRESS_REG_WB: | |
3865 | switch (GET_CODE (x)) | |
3866 | { | |
3867 | case PRE_INC: | |
01a3a324 | 3868 | asm_fprintf (f, "[%s,%d]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3869 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
3870 | return; | |
3871 | case POST_INC: | |
01a3a324 | 3872 | asm_fprintf (f, "[%s],%d", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3873 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
3874 | return; | |
3875 | case PRE_DEC: | |
01a3a324 | 3876 | asm_fprintf (f, "[%s,-%d]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3877 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
3878 | return; | |
3879 | case POST_DEC: | |
01a3a324 | 3880 | asm_fprintf (f, "[%s],-%d", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3881 | GET_MODE_SIZE (aarch64_memory_reference_mode)); |
3882 | return; | |
3883 | case PRE_MODIFY: | |
01a3a324 | 3884 | asm_fprintf (f, "[%s,%wd]!", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3885 | INTVAL (addr.offset)); |
3886 | return; | |
3887 | case POST_MODIFY: | |
01a3a324 | 3888 | asm_fprintf (f, "[%s],%wd", reg_names [REGNO (addr.base)], |
43e9d192 IB |
3889 | INTVAL (addr.offset)); |
3890 | return; | |
3891 | default: | |
3892 | break; | |
3893 | } | |
3894 | break; | |
3895 | ||
3896 | case ADDRESS_LO_SUM: | |
01a3a324 | 3897 | asm_fprintf (f, "[%s,#:lo12:", reg_names [REGNO (addr.base)]); |
43e9d192 IB |
3898 | output_addr_const (f, addr.offset); |
3899 | asm_fprintf (f, "]"); | |
3900 | return; | |
3901 | ||
3902 | case ADDRESS_SYMBOLIC: | |
3903 | break; | |
3904 | } | |
3905 | ||
3906 | output_addr_const (f, x); | |
3907 | } | |
3908 | ||
43e9d192 IB |
3909 | bool |
3910 | aarch64_label_mentioned_p (rtx x) | |
3911 | { | |
3912 | const char *fmt; | |
3913 | int i; | |
3914 | ||
3915 | if (GET_CODE (x) == LABEL_REF) | |
3916 | return true; | |
3917 | ||
3918 | /* UNSPEC_TLS entries for a symbol include a LABEL_REF for the | |
3919 | referencing instruction, but they are constant offsets, not | |
3920 | symbols. */ | |
3921 | if (GET_CODE (x) == UNSPEC && XINT (x, 1) == UNSPEC_TLS) | |
3922 | return false; | |
3923 | ||
3924 | fmt = GET_RTX_FORMAT (GET_CODE (x)); | |
3925 | for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--) | |
3926 | { | |
3927 | if (fmt[i] == 'E') | |
3928 | { | |
3929 | int j; | |
3930 | ||
3931 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3932 | if (aarch64_label_mentioned_p (XVECEXP (x, i, j))) | |
3933 | return 1; | |
3934 | } | |
3935 | else if (fmt[i] == 'e' && aarch64_label_mentioned_p (XEXP (x, i))) | |
3936 | return 1; | |
3937 | } | |
3938 | ||
3939 | return 0; | |
3940 | } | |
3941 | ||
3942 | /* Implement REGNO_REG_CLASS. */ | |
3943 | ||
3944 | enum reg_class | |
3945 | aarch64_regno_regclass (unsigned regno) | |
3946 | { | |
3947 | if (GP_REGNUM_P (regno)) | |
3948 | return CORE_REGS; | |
3949 | ||
3950 | if (regno == SP_REGNUM) | |
3951 | return STACK_REG; | |
3952 | ||
3953 | if (regno == FRAME_POINTER_REGNUM | |
3954 | || regno == ARG_POINTER_REGNUM) | |
f24bb080 | 3955 | return POINTER_REGS; |
43e9d192 IB |
3956 | |
3957 | if (FP_REGNUM_P (regno)) | |
3958 | return FP_LO_REGNUM_P (regno) ? FP_LO_REGS : FP_REGS; | |
3959 | ||
3960 | return NO_REGS; | |
3961 | } | |
3962 | ||
3963 | /* Try a machine-dependent way of reloading an illegitimate address | |
3964 | operand. If we find one, push the reload and return the new rtx. */ | |
3965 | ||
3966 | rtx | |
3967 | aarch64_legitimize_reload_address (rtx *x_p, | |
3968 | enum machine_mode mode, | |
3969 | int opnum, int type, | |
3970 | int ind_levels ATTRIBUTE_UNUSED) | |
3971 | { | |
3972 | rtx x = *x_p; | |
3973 | ||
3974 | /* Do not allow mem (plus (reg, const)) if vector mode. */ | |
3975 | if (aarch64_vector_mode_p (mode) | |
3976 | && GET_CODE (x) == PLUS | |
3977 | && REG_P (XEXP (x, 0)) | |
3978 | && CONST_INT_P (XEXP (x, 1))) | |
3979 | { | |
3980 | rtx orig_rtx = x; | |
3981 | x = copy_rtx (x); | |
3982 | push_reload (orig_rtx, NULL_RTX, x_p, NULL, | |
3983 | BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0, | |
3984 | opnum, (enum reload_type) type); | |
3985 | return x; | |
3986 | } | |
3987 | ||
3988 | /* We must recognize output that we have already generated ourselves. */ | |
3989 | if (GET_CODE (x) == PLUS | |
3990 | && GET_CODE (XEXP (x, 0)) == PLUS | |
3991 | && REG_P (XEXP (XEXP (x, 0), 0)) | |
3992 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) | |
3993 | && CONST_INT_P (XEXP (x, 1))) | |
3994 | { | |
3995 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
3996 | BASE_REG_CLASS, GET_MODE (x), VOIDmode, 0, 0, | |
3997 | opnum, (enum reload_type) type); | |
3998 | return x; | |
3999 | } | |
4000 | ||
4001 | /* We wish to handle large displacements off a base register by splitting | |
4002 | the addend across an add and the mem insn. This can cut the number of | |
4003 | extra insns needed from 3 to 1. It is only useful for load/store of a | |
4004 | single register with 12 bit offset field. */ | |
4005 | if (GET_CODE (x) == PLUS | |
4006 | && REG_P (XEXP (x, 0)) | |
4007 | && CONST_INT_P (XEXP (x, 1)) | |
4008 | && HARD_REGISTER_P (XEXP (x, 0)) | |
4009 | && mode != TImode | |
4010 | && mode != TFmode | |
4011 | && aarch64_regno_ok_for_base_p (REGNO (XEXP (x, 0)), true)) | |
4012 | { | |
4013 | HOST_WIDE_INT val = INTVAL (XEXP (x, 1)); | |
4014 | HOST_WIDE_INT low = val & 0xfff; | |
4015 | HOST_WIDE_INT high = val - low; | |
4016 | HOST_WIDE_INT offs; | |
4017 | rtx cst; | |
28514dda YZ |
4018 | enum machine_mode xmode = GET_MODE (x); |
4019 | ||
4020 | /* In ILP32, xmode can be either DImode or SImode. */ | |
4021 | gcc_assert (xmode == DImode || xmode == SImode); | |
43e9d192 IB |
4022 | |
4023 | /* Reload non-zero BLKmode offsets. This is because we cannot ascertain | |
4024 | BLKmode alignment. */ | |
4025 | if (GET_MODE_SIZE (mode) == 0) | |
4026 | return NULL_RTX; | |
4027 | ||
4028 | offs = low % GET_MODE_SIZE (mode); | |
4029 | ||
4030 | /* Align misaligned offset by adjusting high part to compensate. */ | |
4031 | if (offs != 0) | |
4032 | { | |
4033 | if (aarch64_uimm12_shift (high + offs)) | |
4034 | { | |
4035 | /* Align down. */ | |
4036 | low = low - offs; | |
4037 | high = high + offs; | |
4038 | } | |
4039 | else | |
4040 | { | |
4041 | /* Align up. */ | |
4042 | offs = GET_MODE_SIZE (mode) - offs; | |
4043 | low = low + offs; | |
4044 | high = high + (low & 0x1000) - offs; | |
4045 | low &= 0xfff; | |
4046 | } | |
4047 | } | |
4048 | ||
4049 | /* Check for overflow. */ | |
4050 | if (high + low != val) | |
4051 | return NULL_RTX; | |
4052 | ||
4053 | cst = GEN_INT (high); | |
4054 | if (!aarch64_uimm12_shift (high)) | |
28514dda | 4055 | cst = force_const_mem (xmode, cst); |
43e9d192 IB |
4056 | |
4057 | /* Reload high part into base reg, leaving the low part | |
acf67b8e KT |
4058 | in the mem instruction. |
4059 | Note that replacing this gen_rtx_PLUS with plus_constant is | |
4060 | wrong in this case because we rely on the | |
4061 | (plus (plus reg c1) c2) structure being preserved so that | |
4062 | XEXP (*p, 0) in push_reload below uses the correct term. */ | |
4d0e5b54 JG |
4063 | x = gen_rtx_PLUS (xmode, |
4064 | gen_rtx_PLUS (xmode, XEXP (x, 0), cst), | |
4065 | GEN_INT (low)); | |
43e9d192 IB |
4066 | |
4067 | push_reload (XEXP (x, 0), NULL_RTX, &XEXP (x, 0), NULL, | |
28514dda | 4068 | BASE_REG_CLASS, xmode, VOIDmode, 0, 0, |
43e9d192 IB |
4069 | opnum, (enum reload_type) type); |
4070 | return x; | |
4071 | } | |
4072 | ||
4073 | return NULL_RTX; | |
4074 | } | |
4075 | ||
4076 | ||
4077 | static reg_class_t | |
4078 | aarch64_secondary_reload (bool in_p ATTRIBUTE_UNUSED, rtx x, | |
4079 | reg_class_t rclass, | |
4080 | enum machine_mode mode, | |
4081 | secondary_reload_info *sri) | |
4082 | { | |
43e9d192 IB |
4083 | /* Without the TARGET_SIMD instructions we cannot move a Q register |
4084 | to a Q register directly. We need a scratch. */ | |
4085 | if (REG_P (x) && (mode == TFmode || mode == TImode) && mode == GET_MODE (x) | |
4086 | && FP_REGNUM_P (REGNO (x)) && !TARGET_SIMD | |
4087 | && reg_class_subset_p (rclass, FP_REGS)) | |
4088 | { | |
4089 | if (mode == TFmode) | |
4090 | sri->icode = CODE_FOR_aarch64_reload_movtf; | |
4091 | else if (mode == TImode) | |
4092 | sri->icode = CODE_FOR_aarch64_reload_movti; | |
4093 | return NO_REGS; | |
4094 | } | |
4095 | ||
4096 | /* A TFmode or TImode memory access should be handled via an FP_REGS | |
4097 | because AArch64 has richer addressing modes for LDR/STR instructions | |
4098 | than LDP/STP instructions. */ | |
4099 | if (!TARGET_GENERAL_REGS_ONLY && rclass == CORE_REGS | |
4100 | && GET_MODE_SIZE (mode) == 16 && MEM_P (x)) | |
4101 | return FP_REGS; | |
4102 | ||
4103 | if (rclass == FP_REGS && (mode == TImode || mode == TFmode) && CONSTANT_P(x)) | |
4104 | return CORE_REGS; | |
4105 | ||
4106 | return NO_REGS; | |
4107 | } | |
4108 | ||
4109 | static bool | |
4110 | aarch64_can_eliminate (const int from, const int to) | |
4111 | { | |
4112 | /* If we need a frame pointer, we must eliminate FRAME_POINTER_REGNUM into | |
4113 | HARD_FRAME_POINTER_REGNUM and not into STACK_POINTER_REGNUM. */ | |
4114 | ||
4115 | if (frame_pointer_needed) | |
4116 | { | |
4117 | if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
4118 | return true; | |
4119 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
4120 | return false; | |
4121 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM | |
4122 | && !cfun->calls_alloca) | |
4123 | return true; | |
4124 | if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM) | |
4125 | return true; | |
4126 | return false; | |
4127 | } | |
4128 | else | |
4129 | { | |
777e6976 IB |
4130 | /* If we decided that we didn't need a leaf frame pointer but then used |
4131 | LR in the function, then we'll want a frame pointer after all, so | |
4132 | prevent this elimination to ensure a frame pointer is used. | |
4133 | ||
4134 | NOTE: the original value of flag_omit_frame_pointer gets trashed | |
4135 | IFF flag_omit_leaf_frame_pointer is true, so we check the value | |
4136 | of faked_omit_frame_pointer here (which is true when we always | |
4137 | wish to keep non-leaf frame pointers but only wish to keep leaf frame | |
4138 | pointers when LR is clobbered). */ | |
77436791 | 4139 | if (to == STACK_POINTER_REGNUM |
777e6976 IB |
4140 | && df_regs_ever_live_p (LR_REGNUM) |
4141 | && faked_omit_frame_pointer) | |
43e9d192 IB |
4142 | return false; |
4143 | } | |
777e6976 | 4144 | |
43e9d192 IB |
4145 | return true; |
4146 | } | |
4147 | ||
4148 | HOST_WIDE_INT | |
4149 | aarch64_initial_elimination_offset (unsigned from, unsigned to) | |
4150 | { | |
4151 | HOST_WIDE_INT frame_size; | |
4152 | HOST_WIDE_INT offset; | |
4153 | ||
4154 | aarch64_layout_frame (); | |
4155 | frame_size = (get_frame_size () + cfun->machine->frame.saved_regs_size | |
4156 | + crtl->outgoing_args_size | |
4157 | + cfun->machine->saved_varargs_size); | |
4158 | ||
4159 | frame_size = AARCH64_ROUND_UP (frame_size, STACK_BOUNDARY / BITS_PER_UNIT); | |
4160 | offset = frame_size; | |
4161 | ||
4162 | if (to == HARD_FRAME_POINTER_REGNUM) | |
4163 | { | |
4164 | if (from == ARG_POINTER_REGNUM) | |
4165 | return offset - crtl->outgoing_args_size; | |
4166 | ||
4167 | if (from == FRAME_POINTER_REGNUM) | |
6991c977 | 4168 | return cfun->machine->frame.saved_regs_size + get_frame_size (); |
43e9d192 IB |
4169 | } |
4170 | ||
4171 | if (to == STACK_POINTER_REGNUM) | |
4172 | { | |
4173 | if (from == FRAME_POINTER_REGNUM) | |
4174 | { | |
4175 | HOST_WIDE_INT elim = crtl->outgoing_args_size | |
4176 | + cfun->machine->frame.saved_regs_size | |
6991c977 | 4177 | + get_frame_size () |
43e9d192 IB |
4178 | - cfun->machine->frame.fp_lr_offset; |
4179 | elim = AARCH64_ROUND_UP (elim, STACK_BOUNDARY / BITS_PER_UNIT); | |
4180 | return elim; | |
4181 | } | |
4182 | } | |
4183 | ||
4184 | return offset; | |
4185 | } | |
4186 | ||
4187 | ||
4188 | /* Implement RETURN_ADDR_RTX. We do not support moving back to a | |
4189 | previous frame. */ | |
4190 | ||
4191 | rtx | |
4192 | aarch64_return_addr (int count, rtx frame ATTRIBUTE_UNUSED) | |
4193 | { | |
4194 | if (count != 0) | |
4195 | return const0_rtx; | |
4196 | return get_hard_reg_initial_val (Pmode, LR_REGNUM); | |
4197 | } | |
4198 | ||
4199 | ||
4200 | static void | |
4201 | aarch64_asm_trampoline_template (FILE *f) | |
4202 | { | |
28514dda YZ |
4203 | if (TARGET_ILP32) |
4204 | { | |
4205 | asm_fprintf (f, "\tldr\tw%d, .+16\n", IP1_REGNUM - R0_REGNUM); | |
4206 | asm_fprintf (f, "\tldr\tw%d, .+16\n", STATIC_CHAIN_REGNUM - R0_REGNUM); | |
4207 | } | |
4208 | else | |
4209 | { | |
4210 | asm_fprintf (f, "\tldr\t%s, .+16\n", reg_names [IP1_REGNUM]); | |
4211 | asm_fprintf (f, "\tldr\t%s, .+20\n", reg_names [STATIC_CHAIN_REGNUM]); | |
4212 | } | |
01a3a324 | 4213 | asm_fprintf (f, "\tbr\t%s\n", reg_names [IP1_REGNUM]); |
43e9d192 | 4214 | assemble_aligned_integer (4, const0_rtx); |
28514dda YZ |
4215 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); |
4216 | assemble_aligned_integer (POINTER_BYTES, const0_rtx); | |
43e9d192 IB |
4217 | } |
4218 | ||
4219 | static void | |
4220 | aarch64_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
4221 | { | |
4222 | rtx fnaddr, mem, a_tramp; | |
28514dda | 4223 | const int tramp_code_sz = 16; |
43e9d192 IB |
4224 | |
4225 | /* Don't need to copy the trailing D-words, we fill those in below. */ | |
4226 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
28514dda YZ |
4227 | GEN_INT (tramp_code_sz), BLOCK_OP_NORMAL); |
4228 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz); | |
43e9d192 | 4229 | fnaddr = XEXP (DECL_RTL (fndecl), 0); |
28514dda YZ |
4230 | if (GET_MODE (fnaddr) != ptr_mode) |
4231 | fnaddr = convert_memory_address (ptr_mode, fnaddr); | |
43e9d192 IB |
4232 | emit_move_insn (mem, fnaddr); |
4233 | ||
28514dda | 4234 | mem = adjust_address (m_tramp, ptr_mode, tramp_code_sz + POINTER_BYTES); |
43e9d192 IB |
4235 | emit_move_insn (mem, chain_value); |
4236 | ||
4237 | /* XXX We should really define a "clear_cache" pattern and use | |
4238 | gen_clear_cache(). */ | |
4239 | a_tramp = XEXP (m_tramp, 0); | |
4240 | emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "__clear_cache"), | |
28514dda YZ |
4241 | LCT_NORMAL, VOIDmode, 2, a_tramp, ptr_mode, |
4242 | plus_constant (ptr_mode, a_tramp, TRAMPOLINE_SIZE), | |
4243 | ptr_mode); | |
43e9d192 IB |
4244 | } |
4245 | ||
4246 | static unsigned char | |
4247 | aarch64_class_max_nregs (reg_class_t regclass, enum machine_mode mode) | |
4248 | { | |
4249 | switch (regclass) | |
4250 | { | |
4251 | case CORE_REGS: | |
4252 | case POINTER_REGS: | |
4253 | case GENERAL_REGS: | |
4254 | case ALL_REGS: | |
4255 | case FP_REGS: | |
4256 | case FP_LO_REGS: | |
4257 | return | |
4258 | aarch64_vector_mode_p (mode) ? (GET_MODE_SIZE (mode) + 15) / 16 : | |
4259 | (GET_MODE_SIZE (mode) + 7) / 8; | |
4260 | case STACK_REG: | |
4261 | return 1; | |
4262 | ||
4263 | case NO_REGS: | |
4264 | return 0; | |
4265 | ||
4266 | default: | |
4267 | break; | |
4268 | } | |
4269 | gcc_unreachable (); | |
4270 | } | |
4271 | ||
4272 | static reg_class_t | |
78d8b9f0 | 4273 | aarch64_preferred_reload_class (rtx x, reg_class_t regclass) |
43e9d192 | 4274 | { |
51bb310d | 4275 | if (regclass == POINTER_REGS) |
78d8b9f0 IB |
4276 | return GENERAL_REGS; |
4277 | ||
51bb310d MS |
4278 | if (regclass == STACK_REG) |
4279 | { | |
4280 | if (REG_P(x) | |
4281 | && reg_class_subset_p (REGNO_REG_CLASS (REGNO (x)), POINTER_REGS)) | |
4282 | return regclass; | |
4283 | ||
4284 | return NO_REGS; | |
4285 | } | |
4286 | ||
78d8b9f0 IB |
4287 | /* If it's an integer immediate that MOVI can't handle, then |
4288 | FP_REGS is not an option, so we return NO_REGS instead. */ | |
4289 | if (CONST_INT_P (x) && reg_class_subset_p (regclass, FP_REGS) | |
4290 | && !aarch64_simd_imm_scalar_p (x, GET_MODE (x))) | |
4291 | return NO_REGS; | |
4292 | ||
27bd251b IB |
4293 | /* Register eliminiation can result in a request for |
4294 | SP+constant->FP_REGS. We cannot support such operations which | |
4295 | use SP as source and an FP_REG as destination, so reject out | |
4296 | right now. */ | |
4297 | if (! reg_class_subset_p (regclass, GENERAL_REGS) && GET_CODE (x) == PLUS) | |
4298 | { | |
4299 | rtx lhs = XEXP (x, 0); | |
4300 | ||
4301 | /* Look through a possible SUBREG introduced by ILP32. */ | |
4302 | if (GET_CODE (lhs) == SUBREG) | |
4303 | lhs = SUBREG_REG (lhs); | |
4304 | ||
4305 | gcc_assert (REG_P (lhs)); | |
4306 | gcc_assert (reg_class_subset_p (REGNO_REG_CLASS (REGNO (lhs)), | |
4307 | POINTER_REGS)); | |
4308 | return NO_REGS; | |
4309 | } | |
4310 | ||
78d8b9f0 | 4311 | return regclass; |
43e9d192 IB |
4312 | } |
4313 | ||
4314 | void | |
4315 | aarch64_asm_output_labelref (FILE* f, const char *name) | |
4316 | { | |
4317 | asm_fprintf (f, "%U%s", name); | |
4318 | } | |
4319 | ||
4320 | static void | |
4321 | aarch64_elf_asm_constructor (rtx symbol, int priority) | |
4322 | { | |
4323 | if (priority == DEFAULT_INIT_PRIORITY) | |
4324 | default_ctor_section_asm_out_constructor (symbol, priority); | |
4325 | else | |
4326 | { | |
4327 | section *s; | |
4328 | char buf[18]; | |
4329 | snprintf (buf, sizeof (buf), ".init_array.%.5u", priority); | |
4330 | s = get_section (buf, SECTION_WRITE, NULL); | |
4331 | switch_to_section (s); | |
4332 | assemble_align (POINTER_SIZE); | |
28514dda | 4333 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
4334 | } |
4335 | } | |
4336 | ||
4337 | static void | |
4338 | aarch64_elf_asm_destructor (rtx symbol, int priority) | |
4339 | { | |
4340 | if (priority == DEFAULT_INIT_PRIORITY) | |
4341 | default_dtor_section_asm_out_destructor (symbol, priority); | |
4342 | else | |
4343 | { | |
4344 | section *s; | |
4345 | char buf[18]; | |
4346 | snprintf (buf, sizeof (buf), ".fini_array.%.5u", priority); | |
4347 | s = get_section (buf, SECTION_WRITE, NULL); | |
4348 | switch_to_section (s); | |
4349 | assemble_align (POINTER_SIZE); | |
28514dda | 4350 | assemble_aligned_integer (POINTER_BYTES, symbol); |
43e9d192 IB |
4351 | } |
4352 | } | |
4353 | ||
4354 | const char* | |
4355 | aarch64_output_casesi (rtx *operands) | |
4356 | { | |
4357 | char buf[100]; | |
4358 | char label[100]; | |
592a16fc | 4359 | rtx diff_vec = PATTERN (NEXT_INSN (operands[2])); |
43e9d192 IB |
4360 | int index; |
4361 | static const char *const patterns[4][2] = | |
4362 | { | |
4363 | { | |
4364 | "ldrb\t%w3, [%0,%w1,uxtw]", | |
4365 | "add\t%3, %4, %w3, sxtb #2" | |
4366 | }, | |
4367 | { | |
4368 | "ldrh\t%w3, [%0,%w1,uxtw #1]", | |
4369 | "add\t%3, %4, %w3, sxth #2" | |
4370 | }, | |
4371 | { | |
4372 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
4373 | "add\t%3, %4, %w3, sxtw #2" | |
4374 | }, | |
4375 | /* We assume that DImode is only generated when not optimizing and | |
4376 | that we don't really need 64-bit address offsets. That would | |
4377 | imply an object file with 8GB of code in a single function! */ | |
4378 | { | |
4379 | "ldr\t%w3, [%0,%w1,uxtw #2]", | |
4380 | "add\t%3, %4, %w3, sxtw #2" | |
4381 | } | |
4382 | }; | |
4383 | ||
4384 | gcc_assert (GET_CODE (diff_vec) == ADDR_DIFF_VEC); | |
4385 | ||
4386 | index = exact_log2 (GET_MODE_SIZE (GET_MODE (diff_vec))); | |
4387 | ||
4388 | gcc_assert (index >= 0 && index <= 3); | |
4389 | ||
4390 | /* Need to implement table size reduction, by chaning the code below. */ | |
4391 | output_asm_insn (patterns[index][0], operands); | |
4392 | ASM_GENERATE_INTERNAL_LABEL (label, "Lrtx", CODE_LABEL_NUMBER (operands[2])); | |
4393 | snprintf (buf, sizeof (buf), | |
4394 | "adr\t%%4, %s", targetm.strip_name_encoding (label)); | |
4395 | output_asm_insn (buf, operands); | |
4396 | output_asm_insn (patterns[index][1], operands); | |
4397 | output_asm_insn ("br\t%3", operands); | |
4398 | assemble_label (asm_out_file, label); | |
4399 | return ""; | |
4400 | } | |
4401 | ||
4402 | ||
4403 | /* Return size in bits of an arithmetic operand which is shifted/scaled and | |
4404 | masked such that it is suitable for a UXTB, UXTH, or UXTW extend | |
4405 | operator. */ | |
4406 | ||
4407 | int | |
4408 | aarch64_uxt_size (int shift, HOST_WIDE_INT mask) | |
4409 | { | |
4410 | if (shift >= 0 && shift <= 3) | |
4411 | { | |
4412 | int size; | |
4413 | for (size = 8; size <= 32; size *= 2) | |
4414 | { | |
4415 | HOST_WIDE_INT bits = ((HOST_WIDE_INT)1U << size) - 1; | |
4416 | if (mask == bits << shift) | |
4417 | return size; | |
4418 | } | |
4419 | } | |
4420 | return 0; | |
4421 | } | |
4422 | ||
4423 | static bool | |
4424 | aarch64_use_blocks_for_constant_p (enum machine_mode mode ATTRIBUTE_UNUSED, | |
4425 | const_rtx x ATTRIBUTE_UNUSED) | |
4426 | { | |
4427 | /* We can't use blocks for constants when we're using a per-function | |
4428 | constant pool. */ | |
4429 | return false; | |
4430 | } | |
4431 | ||
4432 | static section * | |
4433 | aarch64_select_rtx_section (enum machine_mode mode ATTRIBUTE_UNUSED, | |
4434 | rtx x ATTRIBUTE_UNUSED, | |
4435 | unsigned HOST_WIDE_INT align ATTRIBUTE_UNUSED) | |
4436 | { | |
4437 | /* Force all constant pool entries into the current function section. */ | |
4438 | return function_section (current_function_decl); | |
4439 | } | |
4440 | ||
4441 | ||
4442 | /* Costs. */ | |
4443 | ||
4444 | /* Helper function for rtx cost calculation. Strip a shift expression | |
4445 | from X. Returns the inner operand if successful, or the original | |
4446 | expression on failure. */ | |
4447 | static rtx | |
4448 | aarch64_strip_shift (rtx x) | |
4449 | { | |
4450 | rtx op = x; | |
4451 | ||
4452 | if ((GET_CODE (op) == ASHIFT | |
4453 | || GET_CODE (op) == ASHIFTRT | |
4454 | || GET_CODE (op) == LSHIFTRT) | |
4455 | && CONST_INT_P (XEXP (op, 1))) | |
4456 | return XEXP (op, 0); | |
4457 | ||
4458 | if (GET_CODE (op) == MULT | |
4459 | && CONST_INT_P (XEXP (op, 1)) | |
4460 | && ((unsigned) exact_log2 (INTVAL (XEXP (op, 1)))) < 64) | |
4461 | return XEXP (op, 0); | |
4462 | ||
4463 | return x; | |
4464 | } | |
4465 | ||
4466 | /* Helper function for rtx cost calculation. Strip a shift or extend | |
4467 | expression from X. Returns the inner operand if successful, or the | |
4468 | original expression on failure. We deal with a number of possible | |
4469 | canonicalization variations here. */ | |
4470 | static rtx | |
4471 | aarch64_strip_shift_or_extend (rtx x) | |
4472 | { | |
4473 | rtx op = x; | |
4474 | ||
4475 | /* Zero and sign extraction of a widened value. */ | |
4476 | if ((GET_CODE (op) == ZERO_EXTRACT || GET_CODE (op) == SIGN_EXTRACT) | |
4477 | && XEXP (op, 2) == const0_rtx | |
4478 | && aarch64_is_extend_from_extract (GET_MODE (op), XEXP (XEXP (op, 0), 1), | |
4479 | XEXP (op, 1))) | |
4480 | return XEXP (XEXP (op, 0), 0); | |
4481 | ||
4482 | /* It can also be represented (for zero-extend) as an AND with an | |
4483 | immediate. */ | |
4484 | if (GET_CODE (op) == AND | |
4485 | && GET_CODE (XEXP (op, 0)) == MULT | |
4486 | && CONST_INT_P (XEXP (XEXP (op, 0), 1)) | |
4487 | && CONST_INT_P (XEXP (op, 1)) | |
4488 | && aarch64_uxt_size (exact_log2 (INTVAL (XEXP (XEXP (op, 0), 1))), | |
4489 | INTVAL (XEXP (op, 1))) != 0) | |
4490 | return XEXP (XEXP (op, 0), 0); | |
4491 | ||
4492 | /* Now handle extended register, as this may also have an optional | |
4493 | left shift by 1..4. */ | |
4494 | if (GET_CODE (op) == ASHIFT | |
4495 | && CONST_INT_P (XEXP (op, 1)) | |
4496 | && ((unsigned HOST_WIDE_INT) INTVAL (XEXP (op, 1))) <= 4) | |
4497 | op = XEXP (op, 0); | |
4498 | ||
4499 | if (GET_CODE (op) == ZERO_EXTEND | |
4500 | || GET_CODE (op) == SIGN_EXTEND) | |
4501 | op = XEXP (op, 0); | |
4502 | ||
4503 | if (op != x) | |
4504 | return op; | |
4505 | ||
4506 | return aarch64_strip_shift (x); | |
4507 | } | |
4508 | ||
4509 | /* Calculate the cost of calculating X, storing it in *COST. Result | |
4510 | is true if the total cost of the operation has now been calculated. */ | |
4511 | static bool | |
4512 | aarch64_rtx_costs (rtx x, int code, int outer ATTRIBUTE_UNUSED, | |
4513 | int param ATTRIBUTE_UNUSED, int *cost, bool speed) | |
4514 | { | |
4515 | rtx op0, op1; | |
73250c4c | 4516 | const struct cpu_cost_table *extra_cost |
43e9d192 IB |
4517 | = aarch64_tune_params->insn_extra_cost; |
4518 | ||
4519 | switch (code) | |
4520 | { | |
4521 | case SET: | |
4522 | op0 = SET_DEST (x); | |
4523 | op1 = SET_SRC (x); | |
4524 | ||
4525 | switch (GET_CODE (op0)) | |
4526 | { | |
4527 | case MEM: | |
4528 | if (speed) | |
73250c4c | 4529 | *cost += extra_cost->ldst.store; |
43e9d192 IB |
4530 | |
4531 | if (op1 != const0_rtx) | |
4532 | *cost += rtx_cost (op1, SET, 1, speed); | |
4533 | return true; | |
4534 | ||
4535 | case SUBREG: | |
4536 | if (! REG_P (SUBREG_REG (op0))) | |
4537 | *cost += rtx_cost (SUBREG_REG (op0), SET, 0, speed); | |
4538 | /* Fall through. */ | |
4539 | case REG: | |
4540 | /* Cost is just the cost of the RHS of the set. */ | |
4541 | *cost += rtx_cost (op1, SET, 1, true); | |
4542 | return true; | |
4543 | ||
4544 | case ZERO_EXTRACT: /* Bit-field insertion. */ | |
4545 | case SIGN_EXTRACT: | |
4546 | /* Strip any redundant widening of the RHS to meet the width of | |
4547 | the target. */ | |
4548 | if (GET_CODE (op1) == SUBREG) | |
4549 | op1 = SUBREG_REG (op1); | |
4550 | if ((GET_CODE (op1) == ZERO_EXTEND | |
4551 | || GET_CODE (op1) == SIGN_EXTEND) | |
4552 | && GET_CODE (XEXP (op0, 1)) == CONST_INT | |
4553 | && (GET_MODE_BITSIZE (GET_MODE (XEXP (op1, 0))) | |
4554 | >= INTVAL (XEXP (op0, 1)))) | |
4555 | op1 = XEXP (op1, 0); | |
4556 | *cost += rtx_cost (op1, SET, 1, speed); | |
4557 | return true; | |
4558 | ||
4559 | default: | |
4560 | break; | |
4561 | } | |
4562 | return false; | |
4563 | ||
4564 | case MEM: | |
4565 | if (speed) | |
73250c4c | 4566 | *cost += extra_cost->ldst.load; |
43e9d192 IB |
4567 | |
4568 | return true; | |
4569 | ||
4570 | case NEG: | |
4571 | op0 = CONST0_RTX (GET_MODE (x)); | |
4572 | op1 = XEXP (x, 0); | |
4573 | goto cost_minus; | |
4574 | ||
4575 | case COMPARE: | |
4576 | op0 = XEXP (x, 0); | |
4577 | op1 = XEXP (x, 1); | |
4578 | ||
4579 | if (op1 == const0_rtx | |
4580 | && GET_CODE (op0) == AND) | |
4581 | { | |
4582 | x = op0; | |
4583 | goto cost_logic; | |
4584 | } | |
4585 | ||
4586 | /* Comparisons can work if the order is swapped. | |
4587 | Canonicalization puts the more complex operation first, but | |
4588 | we want it in op1. */ | |
4589 | if (! (REG_P (op0) | |
4590 | || (GET_CODE (op0) == SUBREG && REG_P (SUBREG_REG (op0))))) | |
4591 | { | |
4592 | op0 = XEXP (x, 1); | |
4593 | op1 = XEXP (x, 0); | |
4594 | } | |
4595 | goto cost_minus; | |
4596 | ||
4597 | case MINUS: | |
4598 | op0 = XEXP (x, 0); | |
4599 | op1 = XEXP (x, 1); | |
4600 | ||
4601 | cost_minus: | |
4602 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT | |
4603 | || (GET_MODE_CLASS (GET_MODE (x)) == MODE_CC | |
4604 | && GET_MODE_CLASS (GET_MODE (op0)) == MODE_INT)) | |
4605 | { | |
4606 | if (op0 != const0_rtx) | |
4607 | *cost += rtx_cost (op0, MINUS, 0, speed); | |
4608 | ||
4609 | if (CONST_INT_P (op1)) | |
4610 | { | |
4611 | if (!aarch64_uimm12_shift (INTVAL (op1))) | |
4612 | *cost += rtx_cost (op1, MINUS, 1, speed); | |
4613 | } | |
4614 | else | |
4615 | { | |
4616 | op1 = aarch64_strip_shift_or_extend (op1); | |
4617 | *cost += rtx_cost (op1, MINUS, 1, speed); | |
4618 | } | |
4619 | return true; | |
4620 | } | |
4621 | ||
4622 | return false; | |
4623 | ||
4624 | case PLUS: | |
4625 | op0 = XEXP (x, 0); | |
4626 | op1 = XEXP (x, 1); | |
4627 | ||
4628 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
4629 | { | |
4630 | if (CONST_INT_P (op1) && aarch64_uimm12_shift (INTVAL (op1))) | |
4631 | { | |
4632 | *cost += rtx_cost (op0, PLUS, 0, speed); | |
4633 | } | |
4634 | else | |
4635 | { | |
4636 | rtx new_op0 = aarch64_strip_shift_or_extend (op0); | |
4637 | ||
4638 | if (new_op0 == op0 | |
4639 | && GET_CODE (op0) == MULT) | |
4640 | { | |
4641 | if ((GET_CODE (XEXP (op0, 0)) == ZERO_EXTEND | |
4642 | && GET_CODE (XEXP (op0, 1)) == ZERO_EXTEND) | |
4643 | || (GET_CODE (XEXP (op0, 0)) == SIGN_EXTEND | |
4644 | && GET_CODE (XEXP (op0, 1)) == SIGN_EXTEND)) | |
4645 | { | |
4646 | *cost += (rtx_cost (XEXP (XEXP (op0, 0), 0), MULT, 0, | |
4647 | speed) | |
4648 | + rtx_cost (XEXP (XEXP (op0, 1), 0), MULT, 1, | |
4649 | speed) | |
4650 | + rtx_cost (op1, PLUS, 1, speed)); | |
4651 | if (speed) | |
73250c4c KT |
4652 | *cost += |
4653 | extra_cost->mult[GET_MODE (x) == DImode].extend_add; | |
43e9d192 IB |
4654 | return true; |
4655 | } | |
328402a9 | 4656 | |
43e9d192 IB |
4657 | *cost += (rtx_cost (XEXP (op0, 0), MULT, 0, speed) |
4658 | + rtx_cost (XEXP (op0, 1), MULT, 1, speed) | |
4659 | + rtx_cost (op1, PLUS, 1, speed)); | |
4660 | ||
4661 | if (speed) | |
73250c4c | 4662 | *cost += extra_cost->mult[GET_MODE (x) == DImode].add; |
328402a9 RE |
4663 | |
4664 | return true; | |
43e9d192 IB |
4665 | } |
4666 | ||
4667 | *cost += (rtx_cost (new_op0, PLUS, 0, speed) | |
4668 | + rtx_cost (op1, PLUS, 1, speed)); | |
4669 | } | |
4670 | return true; | |
4671 | } | |
4672 | ||
4673 | return false; | |
4674 | ||
4675 | case IOR: | |
4676 | case XOR: | |
4677 | case AND: | |
4678 | cost_logic: | |
4679 | op0 = XEXP (x, 0); | |
4680 | op1 = XEXP (x, 1); | |
4681 | ||
4682 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
4683 | { | |
4684 | if (CONST_INT_P (op1) | |
4685 | && aarch64_bitmask_imm (INTVAL (op1), GET_MODE (x))) | |
4686 | { | |
4687 | *cost += rtx_cost (op0, AND, 0, speed); | |
4688 | } | |
4689 | else | |
4690 | { | |
4691 | if (GET_CODE (op0) == NOT) | |
4692 | op0 = XEXP (op0, 0); | |
4693 | op0 = aarch64_strip_shift (op0); | |
4694 | *cost += (rtx_cost (op0, AND, 0, speed) | |
4695 | + rtx_cost (op1, AND, 1, speed)); | |
4696 | } | |
4697 | return true; | |
4698 | } | |
4699 | return false; | |
4700 | ||
4701 | case ZERO_EXTEND: | |
4702 | if ((GET_MODE (x) == DImode | |
4703 | && GET_MODE (XEXP (x, 0)) == SImode) | |
4704 | || GET_CODE (XEXP (x, 0)) == MEM) | |
4705 | { | |
4706 | *cost += rtx_cost (XEXP (x, 0), ZERO_EXTEND, 0, speed); | |
4707 | return true; | |
4708 | } | |
4709 | return false; | |
4710 | ||
4711 | case SIGN_EXTEND: | |
4712 | if (GET_CODE (XEXP (x, 0)) == MEM) | |
4713 | { | |
4714 | *cost += rtx_cost (XEXP (x, 0), SIGN_EXTEND, 0, speed); | |
4715 | return true; | |
4716 | } | |
4717 | return false; | |
4718 | ||
4719 | case ROTATE: | |
4720 | if (!CONST_INT_P (XEXP (x, 1))) | |
4721 | *cost += COSTS_N_INSNS (2); | |
4722 | /* Fall through. */ | |
4723 | case ROTATERT: | |
4724 | case LSHIFTRT: | |
4725 | case ASHIFT: | |
4726 | case ASHIFTRT: | |
4727 | ||
4728 | /* Shifting by a register often takes an extra cycle. */ | |
4729 | if (speed && !CONST_INT_P (XEXP (x, 1))) | |
73250c4c | 4730 | *cost += extra_cost->alu.arith_shift_reg; |
43e9d192 IB |
4731 | |
4732 | *cost += rtx_cost (XEXP (x, 0), ASHIFT, 0, speed); | |
4733 | return true; | |
4734 | ||
4735 | case HIGH: | |
4736 | if (!CONSTANT_P (XEXP (x, 0))) | |
4737 | *cost += rtx_cost (XEXP (x, 0), HIGH, 0, speed); | |
4738 | return true; | |
4739 | ||
4740 | case LO_SUM: | |
4741 | if (!CONSTANT_P (XEXP (x, 1))) | |
4742 | *cost += rtx_cost (XEXP (x, 1), LO_SUM, 1, speed); | |
4743 | *cost += rtx_cost (XEXP (x, 0), LO_SUM, 0, speed); | |
4744 | return true; | |
4745 | ||
4746 | case ZERO_EXTRACT: | |
4747 | case SIGN_EXTRACT: | |
4748 | *cost += rtx_cost (XEXP (x, 0), ZERO_EXTRACT, 0, speed); | |
4749 | return true; | |
4750 | ||
4751 | case MULT: | |
4752 | op0 = XEXP (x, 0); | |
4753 | op1 = XEXP (x, 1); | |
4754 | ||
4755 | *cost = COSTS_N_INSNS (1); | |
4756 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
4757 | { | |
4758 | if (CONST_INT_P (op1) | |
4759 | && exact_log2 (INTVAL (op1)) > 0) | |
4760 | { | |
4761 | *cost += rtx_cost (op0, ASHIFT, 0, speed); | |
4762 | return true; | |
4763 | } | |
4764 | ||
4765 | if ((GET_CODE (op0) == ZERO_EXTEND | |
4766 | && GET_CODE (op1) == ZERO_EXTEND) | |
4767 | || (GET_CODE (op0) == SIGN_EXTEND | |
4768 | && GET_CODE (op1) == SIGN_EXTEND)) | |
4769 | { | |
4770 | *cost += (rtx_cost (XEXP (op0, 0), MULT, 0, speed) | |
4771 | + rtx_cost (XEXP (op1, 0), MULT, 1, speed)); | |
4772 | if (speed) | |
73250c4c | 4773 | *cost += extra_cost->mult[GET_MODE (x) == DImode].extend; |
43e9d192 IB |
4774 | return true; |
4775 | } | |
4776 | ||
4777 | if (speed) | |
73250c4c | 4778 | *cost += extra_cost->mult[GET_MODE (x) == DImode].simple; |
43e9d192 IB |
4779 | } |
4780 | else if (speed) | |
4781 | { | |
4782 | if (GET_MODE (x) == DFmode) | |
73250c4c | 4783 | *cost += extra_cost->fp[1].mult; |
43e9d192 | 4784 | else if (GET_MODE (x) == SFmode) |
73250c4c | 4785 | *cost += extra_cost->fp[0].mult; |
43e9d192 IB |
4786 | } |
4787 | ||
4788 | return false; /* All arguments need to be in registers. */ | |
4789 | ||
4790 | case MOD: | |
4791 | case UMOD: | |
4792 | *cost = COSTS_N_INSNS (2); | |
4793 | if (speed) | |
4794 | { | |
4795 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
73250c4c KT |
4796 | *cost += (extra_cost->mult[GET_MODE (x) == DImode].add |
4797 | + extra_cost->mult[GET_MODE (x) == DImode].idiv); | |
43e9d192 | 4798 | else if (GET_MODE (x) == DFmode) |
73250c4c KT |
4799 | *cost += (extra_cost->fp[1].mult |
4800 | + extra_cost->fp[1].div); | |
43e9d192 | 4801 | else if (GET_MODE (x) == SFmode) |
73250c4c KT |
4802 | *cost += (extra_cost->fp[0].mult |
4803 | + extra_cost->fp[0].div); | |
43e9d192 IB |
4804 | } |
4805 | return false; /* All arguments need to be in registers. */ | |
4806 | ||
4807 | case DIV: | |
4808 | case UDIV: | |
4809 | *cost = COSTS_N_INSNS (1); | |
4810 | if (speed) | |
4811 | { | |
4812 | if (GET_MODE_CLASS (GET_MODE (x)) == MODE_INT) | |
73250c4c | 4813 | *cost += extra_cost->mult[GET_MODE (x) == DImode].idiv; |
43e9d192 | 4814 | else if (GET_MODE (x) == DFmode) |
73250c4c | 4815 | *cost += extra_cost->fp[1].div; |
43e9d192 | 4816 | else if (GET_MODE (x) == SFmode) |
73250c4c | 4817 | *cost += extra_cost->fp[0].div; |
43e9d192 IB |
4818 | } |
4819 | return false; /* All arguments need to be in registers. */ | |
4820 | ||
4821 | default: | |
4822 | break; | |
4823 | } | |
4824 | return false; | |
4825 | } | |
4826 | ||
4827 | static int | |
4828 | aarch64_address_cost (rtx x ATTRIBUTE_UNUSED, | |
4829 | enum machine_mode mode ATTRIBUTE_UNUSED, | |
4830 | addr_space_t as ATTRIBUTE_UNUSED, bool speed ATTRIBUTE_UNUSED) | |
4831 | { | |
4832 | enum rtx_code c = GET_CODE (x); | |
4833 | const struct cpu_addrcost_table *addr_cost = aarch64_tune_params->addr_cost; | |
4834 | ||
4835 | if (c == PRE_INC || c == PRE_DEC || c == PRE_MODIFY) | |
4836 | return addr_cost->pre_modify; | |
4837 | ||
4838 | if (c == POST_INC || c == POST_DEC || c == POST_MODIFY) | |
4839 | return addr_cost->post_modify; | |
4840 | ||
4841 | if (c == PLUS) | |
4842 | { | |
4843 | if (GET_CODE (XEXP (x, 1)) == CONST_INT) | |
4844 | return addr_cost->imm_offset; | |
4845 | else if (GET_CODE (XEXP (x, 0)) == MULT | |
4846 | || GET_CODE (XEXP (x, 0)) == ZERO_EXTEND | |
4847 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) | |
4848 | return addr_cost->register_extend; | |
4849 | ||
4850 | return addr_cost->register_offset; | |
4851 | } | |
4852 | else if (c == MEM || c == LABEL_REF || c == SYMBOL_REF) | |
4853 | return addr_cost->imm_offset; | |
4854 | ||
4855 | return 0; | |
4856 | } | |
4857 | ||
4858 | static int | |
4859 | aarch64_register_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED, | |
4860 | reg_class_t from, reg_class_t to) | |
4861 | { | |
4862 | const struct cpu_regmove_cost *regmove_cost | |
4863 | = aarch64_tune_params->regmove_cost; | |
4864 | ||
6ee70f81 AP |
4865 | /* Moving between GPR and stack cost is the same as GP2GP. */ |
4866 | if ((from == GENERAL_REGS && to == STACK_REG) | |
4867 | || (to == GENERAL_REGS && from == STACK_REG)) | |
4868 | return regmove_cost->GP2GP; | |
4869 | ||
4870 | /* To/From the stack register, we move via the gprs. */ | |
4871 | if (to == STACK_REG || from == STACK_REG) | |
4872 | return aarch64_register_move_cost (mode, from, GENERAL_REGS) | |
4873 | + aarch64_register_move_cost (mode, GENERAL_REGS, to); | |
4874 | ||
43e9d192 IB |
4875 | if (from == GENERAL_REGS && to == GENERAL_REGS) |
4876 | return regmove_cost->GP2GP; | |
4877 | else if (from == GENERAL_REGS) | |
4878 | return regmove_cost->GP2FP; | |
4879 | else if (to == GENERAL_REGS) | |
4880 | return regmove_cost->FP2GP; | |
4881 | ||
4882 | /* When AdvSIMD instructions are disabled it is not possible to move | |
4883 | a 128-bit value directly between Q registers. This is handled in | |
4884 | secondary reload. A general register is used as a scratch to move | |
4885 | the upper DI value and the lower DI value is moved directly, | |
4886 | hence the cost is the sum of three moves. */ | |
4887 | ||
4888 | if (! TARGET_SIMD && GET_MODE_SIZE (from) == 128 && GET_MODE_SIZE (to) == 128) | |
4889 | return regmove_cost->GP2FP + regmove_cost->FP2GP + regmove_cost->FP2FP; | |
4890 | ||
4891 | return regmove_cost->FP2FP; | |
4892 | } | |
4893 | ||
4894 | static int | |
4895 | aarch64_memory_move_cost (enum machine_mode mode ATTRIBUTE_UNUSED, | |
4896 | reg_class_t rclass ATTRIBUTE_UNUSED, | |
4897 | bool in ATTRIBUTE_UNUSED) | |
4898 | { | |
4899 | return aarch64_tune_params->memmov_cost; | |
4900 | } | |
4901 | ||
d126a4ae AP |
4902 | /* Return the number of instructions that can be issued per cycle. */ |
4903 | static int | |
4904 | aarch64_sched_issue_rate (void) | |
4905 | { | |
4906 | return aarch64_tune_params->issue_rate; | |
4907 | } | |
4908 | ||
8990e73a TB |
4909 | /* Vectorizer cost model target hooks. */ |
4910 | ||
4911 | /* Implement targetm.vectorize.builtin_vectorization_cost. */ | |
4912 | static int | |
4913 | aarch64_builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, | |
4914 | tree vectype, | |
4915 | int misalign ATTRIBUTE_UNUSED) | |
4916 | { | |
4917 | unsigned elements; | |
4918 | ||
4919 | switch (type_of_cost) | |
4920 | { | |
4921 | case scalar_stmt: | |
4922 | return aarch64_tune_params->vec_costs->scalar_stmt_cost; | |
4923 | ||
4924 | case scalar_load: | |
4925 | return aarch64_tune_params->vec_costs->scalar_load_cost; | |
4926 | ||
4927 | case scalar_store: | |
4928 | return aarch64_tune_params->vec_costs->scalar_store_cost; | |
4929 | ||
4930 | case vector_stmt: | |
4931 | return aarch64_tune_params->vec_costs->vec_stmt_cost; | |
4932 | ||
4933 | case vector_load: | |
4934 | return aarch64_tune_params->vec_costs->vec_align_load_cost; | |
4935 | ||
4936 | case vector_store: | |
4937 | return aarch64_tune_params->vec_costs->vec_store_cost; | |
4938 | ||
4939 | case vec_to_scalar: | |
4940 | return aarch64_tune_params->vec_costs->vec_to_scalar_cost; | |
4941 | ||
4942 | case scalar_to_vec: | |
4943 | return aarch64_tune_params->vec_costs->scalar_to_vec_cost; | |
4944 | ||
4945 | case unaligned_load: | |
4946 | return aarch64_tune_params->vec_costs->vec_unalign_load_cost; | |
4947 | ||
4948 | case unaligned_store: | |
4949 | return aarch64_tune_params->vec_costs->vec_unalign_store_cost; | |
4950 | ||
4951 | case cond_branch_taken: | |
4952 | return aarch64_tune_params->vec_costs->cond_taken_branch_cost; | |
4953 | ||
4954 | case cond_branch_not_taken: | |
4955 | return aarch64_tune_params->vec_costs->cond_not_taken_branch_cost; | |
4956 | ||
4957 | case vec_perm: | |
4958 | case vec_promote_demote: | |
4959 | return aarch64_tune_params->vec_costs->vec_stmt_cost; | |
4960 | ||
4961 | case vec_construct: | |
4962 | elements = TYPE_VECTOR_SUBPARTS (vectype); | |
4963 | return elements / 2 + 1; | |
4964 | ||
4965 | default: | |
4966 | gcc_unreachable (); | |
4967 | } | |
4968 | } | |
4969 | ||
4970 | /* Implement targetm.vectorize.add_stmt_cost. */ | |
4971 | static unsigned | |
4972 | aarch64_add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, | |
4973 | struct _stmt_vec_info *stmt_info, int misalign, | |
4974 | enum vect_cost_model_location where) | |
4975 | { | |
4976 | unsigned *cost = (unsigned *) data; | |
4977 | unsigned retval = 0; | |
4978 | ||
4979 | if (flag_vect_cost_model) | |
4980 | { | |
4981 | tree vectype = stmt_info ? stmt_vectype (stmt_info) : NULL_TREE; | |
4982 | int stmt_cost = | |
4983 | aarch64_builtin_vectorization_cost (kind, vectype, misalign); | |
4984 | ||
4985 | /* Statements in an inner loop relative to the loop being | |
4986 | vectorized are weighted more heavily. The value here is | |
4987 | a function (linear for now) of the loop nest level. */ | |
4988 | if (where == vect_body && stmt_info && stmt_in_inner_loop_p (stmt_info)) | |
4989 | { | |
4990 | loop_vec_info loop_info = STMT_VINFO_LOOP_VINFO (stmt_info); | |
4991 | struct loop *loop = LOOP_VINFO_LOOP (loop_info); | |
4992 | unsigned nest_level = loop_depth (loop); | |
4993 | ||
4994 | count *= nest_level; | |
4995 | } | |
4996 | ||
4997 | retval = (unsigned) (count * stmt_cost); | |
4998 | cost[where] += retval; | |
4999 | } | |
5000 | ||
5001 | return retval; | |
5002 | } | |
5003 | ||
43e9d192 IB |
5004 | static void initialize_aarch64_code_model (void); |
5005 | ||
5006 | /* Parse the architecture extension string. */ | |
5007 | ||
5008 | static void | |
5009 | aarch64_parse_extension (char *str) | |
5010 | { | |
5011 | /* The extension string is parsed left to right. */ | |
5012 | const struct aarch64_option_extension *opt = NULL; | |
5013 | ||
5014 | /* Flag to say whether we are adding or removing an extension. */ | |
5015 | int adding_ext = -1; | |
5016 | ||
5017 | while (str != NULL && *str != 0) | |
5018 | { | |
5019 | char *ext; | |
5020 | size_t len; | |
5021 | ||
5022 | str++; | |
5023 | ext = strchr (str, '+'); | |
5024 | ||
5025 | if (ext != NULL) | |
5026 | len = ext - str; | |
5027 | else | |
5028 | len = strlen (str); | |
5029 | ||
5030 | if (len >= 2 && strncmp (str, "no", 2) == 0) | |
5031 | { | |
5032 | adding_ext = 0; | |
5033 | len -= 2; | |
5034 | str += 2; | |
5035 | } | |
5036 | else if (len > 0) | |
5037 | adding_ext = 1; | |
5038 | ||
5039 | if (len == 0) | |
5040 | { | |
5041 | error ("missing feature modifier after %qs", "+no"); | |
5042 | return; | |
5043 | } | |
5044 | ||
5045 | /* Scan over the extensions table trying to find an exact match. */ | |
5046 | for (opt = all_extensions; opt->name != NULL; opt++) | |
5047 | { | |
5048 | if (strlen (opt->name) == len && strncmp (opt->name, str, len) == 0) | |
5049 | { | |
5050 | /* Add or remove the extension. */ | |
5051 | if (adding_ext) | |
5052 | aarch64_isa_flags |= opt->flags_on; | |
5053 | else | |
5054 | aarch64_isa_flags &= ~(opt->flags_off); | |
5055 | break; | |
5056 | } | |
5057 | } | |
5058 | ||
5059 | if (opt->name == NULL) | |
5060 | { | |
5061 | /* Extension not found in list. */ | |
5062 | error ("unknown feature modifier %qs", str); | |
5063 | return; | |
5064 | } | |
5065 | ||
5066 | str = ext; | |
5067 | }; | |
5068 | ||
5069 | return; | |
5070 | } | |
5071 | ||
5072 | /* Parse the ARCH string. */ | |
5073 | ||
5074 | static void | |
5075 | aarch64_parse_arch (void) | |
5076 | { | |
5077 | char *ext; | |
5078 | const struct processor *arch; | |
5079 | char *str = (char *) alloca (strlen (aarch64_arch_string) + 1); | |
5080 | size_t len; | |
5081 | ||
5082 | strcpy (str, aarch64_arch_string); | |
5083 | ||
5084 | ext = strchr (str, '+'); | |
5085 | ||
5086 | if (ext != NULL) | |
5087 | len = ext - str; | |
5088 | else | |
5089 | len = strlen (str); | |
5090 | ||
5091 | if (len == 0) | |
5092 | { | |
5093 | error ("missing arch name in -march=%qs", str); | |
5094 | return; | |
5095 | } | |
5096 | ||
5097 | /* Loop through the list of supported ARCHs to find a match. */ | |
5098 | for (arch = all_architectures; arch->name != NULL; arch++) | |
5099 | { | |
5100 | if (strlen (arch->name) == len && strncmp (arch->name, str, len) == 0) | |
5101 | { | |
5102 | selected_arch = arch; | |
5103 | aarch64_isa_flags = selected_arch->flags; | |
ffee7aa9 JG |
5104 | |
5105 | if (!selected_cpu) | |
5106 | selected_cpu = &all_cores[selected_arch->core]; | |
43e9d192 IB |
5107 | |
5108 | if (ext != NULL) | |
5109 | { | |
5110 | /* ARCH string contains at least one extension. */ | |
5111 | aarch64_parse_extension (ext); | |
5112 | } | |
5113 | ||
ffee7aa9 JG |
5114 | if (strcmp (selected_arch->arch, selected_cpu->arch)) |
5115 | { | |
5116 | warning (0, "switch -mcpu=%s conflicts with -march=%s switch", | |
5117 | selected_cpu->name, selected_arch->name); | |
5118 | } | |
5119 | ||
43e9d192 IB |
5120 | return; |
5121 | } | |
5122 | } | |
5123 | ||
5124 | /* ARCH name not found in list. */ | |
5125 | error ("unknown value %qs for -march", str); | |
5126 | return; | |
5127 | } | |
5128 | ||
5129 | /* Parse the CPU string. */ | |
5130 | ||
5131 | static void | |
5132 | aarch64_parse_cpu (void) | |
5133 | { | |
5134 | char *ext; | |
5135 | const struct processor *cpu; | |
5136 | char *str = (char *) alloca (strlen (aarch64_cpu_string) + 1); | |
5137 | size_t len; | |
5138 | ||
5139 | strcpy (str, aarch64_cpu_string); | |
5140 | ||
5141 | ext = strchr (str, '+'); | |
5142 | ||
5143 | if (ext != NULL) | |
5144 | len = ext - str; | |
5145 | else | |
5146 | len = strlen (str); | |
5147 | ||
5148 | if (len == 0) | |
5149 | { | |
5150 | error ("missing cpu name in -mcpu=%qs", str); | |
5151 | return; | |
5152 | } | |
5153 | ||
5154 | /* Loop through the list of supported CPUs to find a match. */ | |
5155 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
5156 | { | |
5157 | if (strlen (cpu->name) == len && strncmp (cpu->name, str, len) == 0) | |
5158 | { | |
5159 | selected_cpu = cpu; | |
192ed1dd | 5160 | selected_tune = cpu; |
43e9d192 IB |
5161 | aarch64_isa_flags = selected_cpu->flags; |
5162 | ||
5163 | if (ext != NULL) | |
5164 | { | |
5165 | /* CPU string contains at least one extension. */ | |
5166 | aarch64_parse_extension (ext); | |
5167 | } | |
5168 | ||
5169 | return; | |
5170 | } | |
5171 | } | |
5172 | ||
5173 | /* CPU name not found in list. */ | |
5174 | error ("unknown value %qs for -mcpu", str); | |
5175 | return; | |
5176 | } | |
5177 | ||
5178 | /* Parse the TUNE string. */ | |
5179 | ||
5180 | static void | |
5181 | aarch64_parse_tune (void) | |
5182 | { | |
5183 | const struct processor *cpu; | |
5184 | char *str = (char *) alloca (strlen (aarch64_tune_string) + 1); | |
5185 | strcpy (str, aarch64_tune_string); | |
5186 | ||
5187 | /* Loop through the list of supported CPUs to find a match. */ | |
5188 | for (cpu = all_cores; cpu->name != NULL; cpu++) | |
5189 | { | |
5190 | if (strcmp (cpu->name, str) == 0) | |
5191 | { | |
5192 | selected_tune = cpu; | |
5193 | return; | |
5194 | } | |
5195 | } | |
5196 | ||
5197 | /* CPU name not found in list. */ | |
5198 | error ("unknown value %qs for -mtune", str); | |
5199 | return; | |
5200 | } | |
5201 | ||
5202 | ||
5203 | /* Implement TARGET_OPTION_OVERRIDE. */ | |
5204 | ||
5205 | static void | |
5206 | aarch64_override_options (void) | |
5207 | { | |
ffee7aa9 JG |
5208 | /* -mcpu=CPU is shorthand for -march=ARCH_FOR_CPU, -mtune=CPU. |
5209 | If either of -march or -mtune is given, they override their | |
5210 | respective component of -mcpu. | |
43e9d192 | 5211 | |
ffee7aa9 JG |
5212 | So, first parse AARCH64_CPU_STRING, then the others, be careful |
5213 | with -march as, if -mcpu is not present on the command line, march | |
5214 | must set a sensible default CPU. */ | |
5215 | if (aarch64_cpu_string) | |
43e9d192 | 5216 | { |
ffee7aa9 | 5217 | aarch64_parse_cpu (); |
43e9d192 IB |
5218 | } |
5219 | ||
ffee7aa9 | 5220 | if (aarch64_arch_string) |
43e9d192 | 5221 | { |
ffee7aa9 | 5222 | aarch64_parse_arch (); |
43e9d192 IB |
5223 | } |
5224 | ||
5225 | if (aarch64_tune_string) | |
5226 | { | |
5227 | aarch64_parse_tune (); | |
5228 | } | |
5229 | ||
63892fa2 KV |
5230 | #ifndef HAVE_AS_MABI_OPTION |
5231 | /* The compiler may have been configured with 2.23.* binutils, which does | |
5232 | not have support for ILP32. */ | |
5233 | if (TARGET_ILP32) | |
5234 | error ("Assembler does not support -mabi=ilp32"); | |
5235 | #endif | |
5236 | ||
43e9d192 IB |
5237 | initialize_aarch64_code_model (); |
5238 | ||
5239 | aarch64_build_bitmask_table (); | |
5240 | ||
5241 | /* This target defaults to strict volatile bitfields. */ | |
5242 | if (flag_strict_volatile_bitfields < 0 && abi_version_at_least (2)) | |
5243 | flag_strict_volatile_bitfields = 1; | |
5244 | ||
5245 | /* If the user did not specify a processor, choose the default | |
5246 | one for them. This will be the CPU set during configuration using | |
02fdbd5b | 5247 | --with-cpu, otherwise it is "coretex-a53". */ |
43e9d192 IB |
5248 | if (!selected_cpu) |
5249 | { | |
5250 | selected_cpu = &all_cores[TARGET_CPU_DEFAULT & 0x3f]; | |
5251 | aarch64_isa_flags = TARGET_CPU_DEFAULT >> 6; | |
5252 | } | |
5253 | ||
5254 | gcc_assert (selected_cpu); | |
5255 | ||
5256 | /* The selected cpu may be an architecture, so lookup tuning by core ID. */ | |
5257 | if (!selected_tune) | |
5258 | selected_tune = &all_cores[selected_cpu->core]; | |
5259 | ||
5260 | aarch64_tune_flags = selected_tune->flags; | |
5261 | aarch64_tune = selected_tune->core; | |
5262 | aarch64_tune_params = selected_tune->tune; | |
5263 | ||
5264 | aarch64_override_options_after_change (); | |
5265 | } | |
5266 | ||
5267 | /* Implement targetm.override_options_after_change. */ | |
5268 | ||
5269 | static void | |
5270 | aarch64_override_options_after_change (void) | |
5271 | { | |
5272 | faked_omit_frame_pointer = false; | |
5273 | ||
5274 | /* To omit leaf frame pointers, we need to turn flag_omit_frame_pointer on so | |
5275 | that aarch64_frame_pointer_required will be called. We need to remember | |
5276 | whether flag_omit_frame_pointer was turned on normally or just faked. */ | |
5277 | ||
5278 | if (flag_omit_leaf_frame_pointer && !flag_omit_frame_pointer) | |
5279 | { | |
5280 | flag_omit_frame_pointer = true; | |
5281 | faked_omit_frame_pointer = true; | |
5282 | } | |
5283 | } | |
5284 | ||
5285 | static struct machine_function * | |
5286 | aarch64_init_machine_status (void) | |
5287 | { | |
5288 | struct machine_function *machine; | |
5289 | machine = ggc_alloc_cleared_machine_function (); | |
5290 | return machine; | |
5291 | } | |
5292 | ||
5293 | void | |
5294 | aarch64_init_expanders (void) | |
5295 | { | |
5296 | init_machine_status = aarch64_init_machine_status; | |
5297 | } | |
5298 | ||
5299 | /* A checking mechanism for the implementation of the various code models. */ | |
5300 | static void | |
5301 | initialize_aarch64_code_model (void) | |
5302 | { | |
5303 | if (flag_pic) | |
5304 | { | |
5305 | switch (aarch64_cmodel_var) | |
5306 | { | |
5307 | case AARCH64_CMODEL_TINY: | |
5308 | aarch64_cmodel = AARCH64_CMODEL_TINY_PIC; | |
5309 | break; | |
5310 | case AARCH64_CMODEL_SMALL: | |
5311 | aarch64_cmodel = AARCH64_CMODEL_SMALL_PIC; | |
5312 | break; | |
5313 | case AARCH64_CMODEL_LARGE: | |
5314 | sorry ("code model %qs with -f%s", "large", | |
5315 | flag_pic > 1 ? "PIC" : "pic"); | |
5316 | default: | |
5317 | gcc_unreachable (); | |
5318 | } | |
5319 | } | |
5320 | else | |
5321 | aarch64_cmodel = aarch64_cmodel_var; | |
5322 | } | |
5323 | ||
5324 | /* Return true if SYMBOL_REF X binds locally. */ | |
5325 | ||
5326 | static bool | |
5327 | aarch64_symbol_binds_local_p (const_rtx x) | |
5328 | { | |
5329 | return (SYMBOL_REF_DECL (x) | |
5330 | ? targetm.binds_local_p (SYMBOL_REF_DECL (x)) | |
5331 | : SYMBOL_REF_LOCAL_P (x)); | |
5332 | } | |
5333 | ||
5334 | /* Return true if SYMBOL_REF X is thread local */ | |
5335 | static bool | |
5336 | aarch64_tls_symbol_p (rtx x) | |
5337 | { | |
5338 | if (! TARGET_HAVE_TLS) | |
5339 | return false; | |
5340 | ||
5341 | if (GET_CODE (x) != SYMBOL_REF) | |
5342 | return false; | |
5343 | ||
5344 | return SYMBOL_REF_TLS_MODEL (x) != 0; | |
5345 | } | |
5346 | ||
5347 | /* Classify a TLS symbol into one of the TLS kinds. */ | |
5348 | enum aarch64_symbol_type | |
5349 | aarch64_classify_tls_symbol (rtx x) | |
5350 | { | |
5351 | enum tls_model tls_kind = tls_symbolic_operand_type (x); | |
5352 | ||
5353 | switch (tls_kind) | |
5354 | { | |
5355 | case TLS_MODEL_GLOBAL_DYNAMIC: | |
5356 | case TLS_MODEL_LOCAL_DYNAMIC: | |
5357 | return TARGET_TLS_DESC ? SYMBOL_SMALL_TLSDESC : SYMBOL_SMALL_TLSGD; | |
5358 | ||
5359 | case TLS_MODEL_INITIAL_EXEC: | |
5360 | return SYMBOL_SMALL_GOTTPREL; | |
5361 | ||
5362 | case TLS_MODEL_LOCAL_EXEC: | |
5363 | return SYMBOL_SMALL_TPREL; | |
5364 | ||
5365 | case TLS_MODEL_EMULATED: | |
5366 | case TLS_MODEL_NONE: | |
5367 | return SYMBOL_FORCE_TO_MEM; | |
5368 | ||
5369 | default: | |
5370 | gcc_unreachable (); | |
5371 | } | |
5372 | } | |
5373 | ||
5374 | /* Return the method that should be used to access SYMBOL_REF or | |
5375 | LABEL_REF X in context CONTEXT. */ | |
17f4d4bf | 5376 | |
43e9d192 IB |
5377 | enum aarch64_symbol_type |
5378 | aarch64_classify_symbol (rtx x, | |
5379 | enum aarch64_symbol_context context ATTRIBUTE_UNUSED) | |
5380 | { | |
5381 | if (GET_CODE (x) == LABEL_REF) | |
5382 | { | |
5383 | switch (aarch64_cmodel) | |
5384 | { | |
5385 | case AARCH64_CMODEL_LARGE: | |
5386 | return SYMBOL_FORCE_TO_MEM; | |
5387 | ||
5388 | case AARCH64_CMODEL_TINY_PIC: | |
5389 | case AARCH64_CMODEL_TINY: | |
a5350ddc CSS |
5390 | return SYMBOL_TINY_ABSOLUTE; |
5391 | ||
43e9d192 IB |
5392 | case AARCH64_CMODEL_SMALL_PIC: |
5393 | case AARCH64_CMODEL_SMALL: | |
5394 | return SYMBOL_SMALL_ABSOLUTE; | |
5395 | ||
5396 | default: | |
5397 | gcc_unreachable (); | |
5398 | } | |
5399 | } | |
5400 | ||
17f4d4bf | 5401 | if (GET_CODE (x) == SYMBOL_REF) |
43e9d192 | 5402 | { |
17f4d4bf CSS |
5403 | if (aarch64_cmodel == AARCH64_CMODEL_LARGE |
5404 | || CONSTANT_POOL_ADDRESS_P (x)) | |
43e9d192 IB |
5405 | return SYMBOL_FORCE_TO_MEM; |
5406 | ||
5407 | if (aarch64_tls_symbol_p (x)) | |
5408 | return aarch64_classify_tls_symbol (x); | |
5409 | ||
17f4d4bf CSS |
5410 | switch (aarch64_cmodel) |
5411 | { | |
5412 | case AARCH64_CMODEL_TINY: | |
a5350ddc CSS |
5413 | if (SYMBOL_REF_WEAK (x)) |
5414 | return SYMBOL_FORCE_TO_MEM; | |
5415 | return SYMBOL_TINY_ABSOLUTE; | |
5416 | ||
17f4d4bf CSS |
5417 | case AARCH64_CMODEL_SMALL: |
5418 | if (SYMBOL_REF_WEAK (x)) | |
5419 | return SYMBOL_FORCE_TO_MEM; | |
5420 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 5421 | |
17f4d4bf | 5422 | case AARCH64_CMODEL_TINY_PIC: |
38e6c9a6 | 5423 | if (!aarch64_symbol_binds_local_p (x)) |
87dd8ab0 | 5424 | return SYMBOL_TINY_GOT; |
38e6c9a6 MS |
5425 | return SYMBOL_TINY_ABSOLUTE; |
5426 | ||
17f4d4bf CSS |
5427 | case AARCH64_CMODEL_SMALL_PIC: |
5428 | if (!aarch64_symbol_binds_local_p (x)) | |
5429 | return SYMBOL_SMALL_GOT; | |
5430 | return SYMBOL_SMALL_ABSOLUTE; | |
43e9d192 | 5431 | |
17f4d4bf CSS |
5432 | default: |
5433 | gcc_unreachable (); | |
5434 | } | |
43e9d192 | 5435 | } |
17f4d4bf | 5436 | |
43e9d192 IB |
5437 | /* By default push everything into the constant pool. */ |
5438 | return SYMBOL_FORCE_TO_MEM; | |
5439 | } | |
5440 | ||
43e9d192 IB |
5441 | bool |
5442 | aarch64_constant_address_p (rtx x) | |
5443 | { | |
5444 | return (CONSTANT_P (x) && memory_address_p (DImode, x)); | |
5445 | } | |
5446 | ||
5447 | bool | |
5448 | aarch64_legitimate_pic_operand_p (rtx x) | |
5449 | { | |
5450 | if (GET_CODE (x) == SYMBOL_REF | |
5451 | || (GET_CODE (x) == CONST | |
5452 | && GET_CODE (XEXP (x, 0)) == PLUS | |
5453 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == SYMBOL_REF)) | |
5454 | return false; | |
5455 | ||
5456 | return true; | |
5457 | } | |
5458 | ||
3520f7cc JG |
5459 | /* Return true if X holds either a quarter-precision or |
5460 | floating-point +0.0 constant. */ | |
5461 | static bool | |
5462 | aarch64_valid_floating_const (enum machine_mode mode, rtx x) | |
5463 | { | |
5464 | if (!CONST_DOUBLE_P (x)) | |
5465 | return false; | |
5466 | ||
5467 | /* TODO: We could handle moving 0.0 to a TFmode register, | |
5468 | but first we would like to refactor the movtf_aarch64 | |
5469 | to be more amicable to split moves properly and | |
5470 | correctly gate on TARGET_SIMD. For now - reject all | |
5471 | constants which are not to SFmode or DFmode registers. */ | |
5472 | if (!(mode == SFmode || mode == DFmode)) | |
5473 | return false; | |
5474 | ||
5475 | if (aarch64_float_const_zero_rtx_p (x)) | |
5476 | return true; | |
5477 | return aarch64_float_const_representable_p (x); | |
5478 | } | |
5479 | ||
43e9d192 IB |
5480 | static bool |
5481 | aarch64_legitimate_constant_p (enum machine_mode mode, rtx x) | |
5482 | { | |
5483 | /* Do not allow vector struct mode constants. We could support | |
5484 | 0 and -1 easily, but they need support in aarch64-simd.md. */ | |
5485 | if (TARGET_SIMD && aarch64_vect_struct_mode_p (mode)) | |
5486 | return false; | |
5487 | ||
5488 | /* This could probably go away because | |
5489 | we now decompose CONST_INTs according to expand_mov_immediate. */ | |
5490 | if ((GET_CODE (x) == CONST_VECTOR | |
48063b9d | 5491 | && aarch64_simd_valid_immediate (x, mode, false, NULL)) |
3520f7cc JG |
5492 | || CONST_INT_P (x) || aarch64_valid_floating_const (mode, x)) |
5493 | return !targetm.cannot_force_const_mem (mode, x); | |
43e9d192 IB |
5494 | |
5495 | if (GET_CODE (x) == HIGH | |
5496 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
5497 | return true; | |
5498 | ||
5499 | return aarch64_constant_address_p (x); | |
5500 | } | |
5501 | ||
a5bc806c | 5502 | rtx |
43e9d192 IB |
5503 | aarch64_load_tp (rtx target) |
5504 | { | |
5505 | if (!target | |
5506 | || GET_MODE (target) != Pmode | |
5507 | || !register_operand (target, Pmode)) | |
5508 | target = gen_reg_rtx (Pmode); | |
5509 | ||
5510 | /* Can return in any reg. */ | |
5511 | emit_insn (gen_aarch64_load_tp_hard (target)); | |
5512 | return target; | |
5513 | } | |
5514 | ||
43e9d192 IB |
5515 | /* On AAPCS systems, this is the "struct __va_list". */ |
5516 | static GTY(()) tree va_list_type; | |
5517 | ||
5518 | /* Implement TARGET_BUILD_BUILTIN_VA_LIST. | |
5519 | Return the type to use as __builtin_va_list. | |
5520 | ||
5521 | AAPCS64 \S 7.1.4 requires that va_list be a typedef for a type defined as: | |
5522 | ||
5523 | struct __va_list | |
5524 | { | |
5525 | void *__stack; | |
5526 | void *__gr_top; | |
5527 | void *__vr_top; | |
5528 | int __gr_offs; | |
5529 | int __vr_offs; | |
5530 | }; */ | |
5531 | ||
5532 | static tree | |
5533 | aarch64_build_builtin_va_list (void) | |
5534 | { | |
5535 | tree va_list_name; | |
5536 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
5537 | ||
5538 | /* Create the type. */ | |
5539 | va_list_type = lang_hooks.types.make_type (RECORD_TYPE); | |
5540 | /* Give it the required name. */ | |
5541 | va_list_name = build_decl (BUILTINS_LOCATION, | |
5542 | TYPE_DECL, | |
5543 | get_identifier ("__va_list"), | |
5544 | va_list_type); | |
5545 | DECL_ARTIFICIAL (va_list_name) = 1; | |
5546 | TYPE_NAME (va_list_type) = va_list_name; | |
665c56c6 | 5547 | TYPE_STUB_DECL (va_list_type) = va_list_name; |
43e9d192 IB |
5548 | |
5549 | /* Create the fields. */ | |
5550 | f_stack = build_decl (BUILTINS_LOCATION, | |
5551 | FIELD_DECL, get_identifier ("__stack"), | |
5552 | ptr_type_node); | |
5553 | f_grtop = build_decl (BUILTINS_LOCATION, | |
5554 | FIELD_DECL, get_identifier ("__gr_top"), | |
5555 | ptr_type_node); | |
5556 | f_vrtop = build_decl (BUILTINS_LOCATION, | |
5557 | FIELD_DECL, get_identifier ("__vr_top"), | |
5558 | ptr_type_node); | |
5559 | f_groff = build_decl (BUILTINS_LOCATION, | |
5560 | FIELD_DECL, get_identifier ("__gr_offs"), | |
5561 | integer_type_node); | |
5562 | f_vroff = build_decl (BUILTINS_LOCATION, | |
5563 | FIELD_DECL, get_identifier ("__vr_offs"), | |
5564 | integer_type_node); | |
5565 | ||
5566 | DECL_ARTIFICIAL (f_stack) = 1; | |
5567 | DECL_ARTIFICIAL (f_grtop) = 1; | |
5568 | DECL_ARTIFICIAL (f_vrtop) = 1; | |
5569 | DECL_ARTIFICIAL (f_groff) = 1; | |
5570 | DECL_ARTIFICIAL (f_vroff) = 1; | |
5571 | ||
5572 | DECL_FIELD_CONTEXT (f_stack) = va_list_type; | |
5573 | DECL_FIELD_CONTEXT (f_grtop) = va_list_type; | |
5574 | DECL_FIELD_CONTEXT (f_vrtop) = va_list_type; | |
5575 | DECL_FIELD_CONTEXT (f_groff) = va_list_type; | |
5576 | DECL_FIELD_CONTEXT (f_vroff) = va_list_type; | |
5577 | ||
5578 | TYPE_FIELDS (va_list_type) = f_stack; | |
5579 | DECL_CHAIN (f_stack) = f_grtop; | |
5580 | DECL_CHAIN (f_grtop) = f_vrtop; | |
5581 | DECL_CHAIN (f_vrtop) = f_groff; | |
5582 | DECL_CHAIN (f_groff) = f_vroff; | |
5583 | ||
5584 | /* Compute its layout. */ | |
5585 | layout_type (va_list_type); | |
5586 | ||
5587 | return va_list_type; | |
5588 | } | |
5589 | ||
5590 | /* Implement TARGET_EXPAND_BUILTIN_VA_START. */ | |
5591 | static void | |
5592 | aarch64_expand_builtin_va_start (tree valist, rtx nextarg ATTRIBUTE_UNUSED) | |
5593 | { | |
5594 | const CUMULATIVE_ARGS *cum; | |
5595 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
5596 | tree stack, grtop, vrtop, groff, vroff; | |
5597 | tree t; | |
5598 | int gr_save_area_size; | |
5599 | int vr_save_area_size; | |
5600 | int vr_offset; | |
5601 | ||
5602 | cum = &crtl->args.info; | |
5603 | gr_save_area_size | |
5604 | = (NUM_ARG_REGS - cum->aapcs_ncrn) * UNITS_PER_WORD; | |
5605 | vr_save_area_size | |
5606 | = (NUM_FP_ARG_REGS - cum->aapcs_nvrn) * UNITS_PER_VREG; | |
5607 | ||
5608 | if (TARGET_GENERAL_REGS_ONLY) | |
5609 | { | |
5610 | if (cum->aapcs_nvrn > 0) | |
5611 | sorry ("%qs and floating point or vector arguments", | |
5612 | "-mgeneral-regs-only"); | |
5613 | vr_save_area_size = 0; | |
5614 | } | |
5615 | ||
5616 | f_stack = TYPE_FIELDS (va_list_type_node); | |
5617 | f_grtop = DECL_CHAIN (f_stack); | |
5618 | f_vrtop = DECL_CHAIN (f_grtop); | |
5619 | f_groff = DECL_CHAIN (f_vrtop); | |
5620 | f_vroff = DECL_CHAIN (f_groff); | |
5621 | ||
5622 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), valist, f_stack, | |
5623 | NULL_TREE); | |
5624 | grtop = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), valist, f_grtop, | |
5625 | NULL_TREE); | |
5626 | vrtop = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), valist, f_vrtop, | |
5627 | NULL_TREE); | |
5628 | groff = build3 (COMPONENT_REF, TREE_TYPE (f_groff), valist, f_groff, | |
5629 | NULL_TREE); | |
5630 | vroff = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), valist, f_vroff, | |
5631 | NULL_TREE); | |
5632 | ||
5633 | /* Emit code to initialize STACK, which points to the next varargs stack | |
5634 | argument. CUM->AAPCS_STACK_SIZE gives the number of stack words used | |
5635 | by named arguments. STACK is 8-byte aligned. */ | |
5636 | t = make_tree (TREE_TYPE (stack), virtual_incoming_args_rtx); | |
5637 | if (cum->aapcs_stack_size > 0) | |
5638 | t = fold_build_pointer_plus_hwi (t, cum->aapcs_stack_size * UNITS_PER_WORD); | |
5639 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), stack, t); | |
5640 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
5641 | ||
5642 | /* Emit code to initialize GRTOP, the top of the GR save area. | |
5643 | virtual_incoming_args_rtx should have been 16 byte aligned. */ | |
5644 | t = make_tree (TREE_TYPE (grtop), virtual_incoming_args_rtx); | |
5645 | t = build2 (MODIFY_EXPR, TREE_TYPE (grtop), grtop, t); | |
5646 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
5647 | ||
5648 | /* Emit code to initialize VRTOP, the top of the VR save area. | |
5649 | This address is gr_save_area_bytes below GRTOP, rounded | |
5650 | down to the next 16-byte boundary. */ | |
5651 | t = make_tree (TREE_TYPE (vrtop), virtual_incoming_args_rtx); | |
5652 | vr_offset = AARCH64_ROUND_UP (gr_save_area_size, | |
5653 | STACK_BOUNDARY / BITS_PER_UNIT); | |
5654 | ||
5655 | if (vr_offset) | |
5656 | t = fold_build_pointer_plus_hwi (t, -vr_offset); | |
5657 | t = build2 (MODIFY_EXPR, TREE_TYPE (vrtop), vrtop, t); | |
5658 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
5659 | ||
5660 | /* Emit code to initialize GROFF, the offset from GRTOP of the | |
5661 | next GPR argument. */ | |
5662 | t = build2 (MODIFY_EXPR, TREE_TYPE (groff), groff, | |
5663 | build_int_cst (TREE_TYPE (groff), -gr_save_area_size)); | |
5664 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
5665 | ||
5666 | /* Likewise emit code to initialize VROFF, the offset from FTOP | |
5667 | of the next VR argument. */ | |
5668 | t = build2 (MODIFY_EXPR, TREE_TYPE (vroff), vroff, | |
5669 | build_int_cst (TREE_TYPE (vroff), -vr_save_area_size)); | |
5670 | expand_expr (t, const0_rtx, VOIDmode, EXPAND_NORMAL); | |
5671 | } | |
5672 | ||
5673 | /* Implement TARGET_GIMPLIFY_VA_ARG_EXPR. */ | |
5674 | ||
5675 | static tree | |
5676 | aarch64_gimplify_va_arg_expr (tree valist, tree type, gimple_seq *pre_p, | |
5677 | gimple_seq *post_p ATTRIBUTE_UNUSED) | |
5678 | { | |
5679 | tree addr; | |
5680 | bool indirect_p; | |
5681 | bool is_ha; /* is HFA or HVA. */ | |
5682 | bool dw_align; /* double-word align. */ | |
5683 | enum machine_mode ag_mode = VOIDmode; | |
5684 | int nregs; | |
5685 | enum machine_mode mode; | |
5686 | ||
5687 | tree f_stack, f_grtop, f_vrtop, f_groff, f_vroff; | |
5688 | tree stack, f_top, f_off, off, arg, roundup, on_stack; | |
5689 | HOST_WIDE_INT size, rsize, adjust, align; | |
5690 | tree t, u, cond1, cond2; | |
5691 | ||
5692 | indirect_p = pass_by_reference (NULL, TYPE_MODE (type), type, false); | |
5693 | if (indirect_p) | |
5694 | type = build_pointer_type (type); | |
5695 | ||
5696 | mode = TYPE_MODE (type); | |
5697 | ||
5698 | f_stack = TYPE_FIELDS (va_list_type_node); | |
5699 | f_grtop = DECL_CHAIN (f_stack); | |
5700 | f_vrtop = DECL_CHAIN (f_grtop); | |
5701 | f_groff = DECL_CHAIN (f_vrtop); | |
5702 | f_vroff = DECL_CHAIN (f_groff); | |
5703 | ||
5704 | stack = build3 (COMPONENT_REF, TREE_TYPE (f_stack), unshare_expr (valist), | |
5705 | f_stack, NULL_TREE); | |
5706 | size = int_size_in_bytes (type); | |
5707 | align = aarch64_function_arg_alignment (mode, type) / BITS_PER_UNIT; | |
5708 | ||
5709 | dw_align = false; | |
5710 | adjust = 0; | |
5711 | if (aarch64_vfp_is_call_or_return_candidate (mode, | |
5712 | type, | |
5713 | &ag_mode, | |
5714 | &nregs, | |
5715 | &is_ha)) | |
5716 | { | |
5717 | /* TYPE passed in fp/simd registers. */ | |
5718 | if (TARGET_GENERAL_REGS_ONLY) | |
5719 | sorry ("%qs and floating point or vector arguments", | |
5720 | "-mgeneral-regs-only"); | |
5721 | ||
5722 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_vrtop), | |
5723 | unshare_expr (valist), f_vrtop, NULL_TREE); | |
5724 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_vroff), | |
5725 | unshare_expr (valist), f_vroff, NULL_TREE); | |
5726 | ||
5727 | rsize = nregs * UNITS_PER_VREG; | |
5728 | ||
5729 | if (is_ha) | |
5730 | { | |
5731 | if (BYTES_BIG_ENDIAN && GET_MODE_SIZE (ag_mode) < UNITS_PER_VREG) | |
5732 | adjust = UNITS_PER_VREG - GET_MODE_SIZE (ag_mode); | |
5733 | } | |
5734 | else if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
5735 | && size < UNITS_PER_VREG) | |
5736 | { | |
5737 | adjust = UNITS_PER_VREG - size; | |
5738 | } | |
5739 | } | |
5740 | else | |
5741 | { | |
5742 | /* TYPE passed in general registers. */ | |
5743 | f_top = build3 (COMPONENT_REF, TREE_TYPE (f_grtop), | |
5744 | unshare_expr (valist), f_grtop, NULL_TREE); | |
5745 | f_off = build3 (COMPONENT_REF, TREE_TYPE (f_groff), | |
5746 | unshare_expr (valist), f_groff, NULL_TREE); | |
5747 | rsize = (size + UNITS_PER_WORD - 1) & -UNITS_PER_WORD; | |
5748 | nregs = rsize / UNITS_PER_WORD; | |
5749 | ||
5750 | if (align > 8) | |
5751 | dw_align = true; | |
5752 | ||
5753 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
5754 | && size < UNITS_PER_WORD) | |
5755 | { | |
5756 | adjust = UNITS_PER_WORD - size; | |
5757 | } | |
5758 | } | |
5759 | ||
5760 | /* Get a local temporary for the field value. */ | |
5761 | off = get_initialized_tmp_var (f_off, pre_p, NULL); | |
5762 | ||
5763 | /* Emit code to branch if off >= 0. */ | |
5764 | t = build2 (GE_EXPR, boolean_type_node, off, | |
5765 | build_int_cst (TREE_TYPE (off), 0)); | |
5766 | cond1 = build3 (COND_EXPR, ptr_type_node, t, NULL_TREE, NULL_TREE); | |
5767 | ||
5768 | if (dw_align) | |
5769 | { | |
5770 | /* Emit: offs = (offs + 15) & -16. */ | |
5771 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
5772 | build_int_cst (TREE_TYPE (off), 15)); | |
5773 | t = build2 (BIT_AND_EXPR, TREE_TYPE (off), t, | |
5774 | build_int_cst (TREE_TYPE (off), -16)); | |
5775 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (off), off, t); | |
5776 | } | |
5777 | else | |
5778 | roundup = NULL; | |
5779 | ||
5780 | /* Update ap.__[g|v]r_offs */ | |
5781 | t = build2 (PLUS_EXPR, TREE_TYPE (off), off, | |
5782 | build_int_cst (TREE_TYPE (off), rsize)); | |
5783 | t = build2 (MODIFY_EXPR, TREE_TYPE (f_off), unshare_expr (f_off), t); | |
5784 | ||
5785 | /* String up. */ | |
5786 | if (roundup) | |
5787 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
5788 | ||
5789 | /* [cond2] if (ap.__[g|v]r_offs > 0) */ | |
5790 | u = build2 (GT_EXPR, boolean_type_node, unshare_expr (f_off), | |
5791 | build_int_cst (TREE_TYPE (f_off), 0)); | |
5792 | cond2 = build3 (COND_EXPR, ptr_type_node, u, NULL_TREE, NULL_TREE); | |
5793 | ||
5794 | /* String up: make sure the assignment happens before the use. */ | |
5795 | t = build2 (COMPOUND_EXPR, TREE_TYPE (cond2), t, cond2); | |
5796 | COND_EXPR_ELSE (cond1) = t; | |
5797 | ||
5798 | /* Prepare the trees handling the argument that is passed on the stack; | |
5799 | the top level node will store in ON_STACK. */ | |
5800 | arg = get_initialized_tmp_var (stack, pre_p, NULL); | |
5801 | if (align > 8) | |
5802 | { | |
5803 | /* if (alignof(type) > 8) (arg = arg + 15) & -16; */ | |
5804 | t = fold_convert (intDI_type_node, arg); | |
5805 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
5806 | build_int_cst (TREE_TYPE (t), 15)); | |
5807 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
5808 | build_int_cst (TREE_TYPE (t), -16)); | |
5809 | t = fold_convert (TREE_TYPE (arg), t); | |
5810 | roundup = build2 (MODIFY_EXPR, TREE_TYPE (arg), arg, t); | |
5811 | } | |
5812 | else | |
5813 | roundup = NULL; | |
5814 | /* Advance ap.__stack */ | |
5815 | t = fold_convert (intDI_type_node, arg); | |
5816 | t = build2 (PLUS_EXPR, TREE_TYPE (t), t, | |
5817 | build_int_cst (TREE_TYPE (t), size + 7)); | |
5818 | t = build2 (BIT_AND_EXPR, TREE_TYPE (t), t, | |
5819 | build_int_cst (TREE_TYPE (t), -8)); | |
5820 | t = fold_convert (TREE_TYPE (arg), t); | |
5821 | t = build2 (MODIFY_EXPR, TREE_TYPE (stack), unshare_expr (stack), t); | |
5822 | /* String up roundup and advance. */ | |
5823 | if (roundup) | |
5824 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), roundup, t); | |
5825 | /* String up with arg */ | |
5826 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), t, arg); | |
5827 | /* Big-endianness related address adjustment. */ | |
5828 | if (BLOCK_REG_PADDING (mode, type, 1) == downward | |
5829 | && size < UNITS_PER_WORD) | |
5830 | { | |
5831 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (arg), arg, | |
5832 | size_int (UNITS_PER_WORD - size)); | |
5833 | on_stack = build2 (COMPOUND_EXPR, TREE_TYPE (arg), on_stack, t); | |
5834 | } | |
5835 | ||
5836 | COND_EXPR_THEN (cond1) = unshare_expr (on_stack); | |
5837 | COND_EXPR_THEN (cond2) = unshare_expr (on_stack); | |
5838 | ||
5839 | /* Adjustment to OFFSET in the case of BIG_ENDIAN. */ | |
5840 | t = off; | |
5841 | if (adjust) | |
5842 | t = build2 (PREINCREMENT_EXPR, TREE_TYPE (off), off, | |
5843 | build_int_cst (TREE_TYPE (off), adjust)); | |
5844 | ||
5845 | t = fold_convert (sizetype, t); | |
5846 | t = build2 (POINTER_PLUS_EXPR, TREE_TYPE (f_top), f_top, t); | |
5847 | ||
5848 | if (is_ha) | |
5849 | { | |
5850 | /* type ha; // treat as "struct {ftype field[n];}" | |
5851 | ... [computing offs] | |
5852 | for (i = 0; i <nregs; ++i, offs += 16) | |
5853 | ha.field[i] = *((ftype *)(ap.__vr_top + offs)); | |
5854 | return ha; */ | |
5855 | int i; | |
5856 | tree tmp_ha, field_t, field_ptr_t; | |
5857 | ||
5858 | /* Declare a local variable. */ | |
5859 | tmp_ha = create_tmp_var_raw (type, "ha"); | |
5860 | gimple_add_tmp_var (tmp_ha); | |
5861 | ||
5862 | /* Establish the base type. */ | |
5863 | switch (ag_mode) | |
5864 | { | |
5865 | case SFmode: | |
5866 | field_t = float_type_node; | |
5867 | field_ptr_t = float_ptr_type_node; | |
5868 | break; | |
5869 | case DFmode: | |
5870 | field_t = double_type_node; | |
5871 | field_ptr_t = double_ptr_type_node; | |
5872 | break; | |
5873 | case TFmode: | |
5874 | field_t = long_double_type_node; | |
5875 | field_ptr_t = long_double_ptr_type_node; | |
5876 | break; | |
5877 | /* The half precision and quad precision are not fully supported yet. Enable | |
5878 | the following code after the support is complete. Need to find the correct | |
5879 | type node for __fp16 *. */ | |
5880 | #if 0 | |
5881 | case HFmode: | |
5882 | field_t = float_type_node; | |
5883 | field_ptr_t = float_ptr_type_node; | |
5884 | break; | |
5885 | #endif | |
5886 | case V2SImode: | |
5887 | case V4SImode: | |
5888 | { | |
5889 | tree innertype = make_signed_type (GET_MODE_PRECISION (SImode)); | |
5890 | field_t = build_vector_type_for_mode (innertype, ag_mode); | |
5891 | field_ptr_t = build_pointer_type (field_t); | |
5892 | } | |
5893 | break; | |
5894 | default: | |
5895 | gcc_assert (0); | |
5896 | } | |
5897 | ||
5898 | /* *(field_ptr_t)&ha = *((field_ptr_t)vr_saved_area */ | |
5899 | tmp_ha = build1 (ADDR_EXPR, field_ptr_t, tmp_ha); | |
5900 | addr = t; | |
5901 | t = fold_convert (field_ptr_t, addr); | |
5902 | t = build2 (MODIFY_EXPR, field_t, | |
5903 | build1 (INDIRECT_REF, field_t, tmp_ha), | |
5904 | build1 (INDIRECT_REF, field_t, t)); | |
5905 | ||
5906 | /* ha.field[i] = *((field_ptr_t)vr_saved_area + i) */ | |
5907 | for (i = 1; i < nregs; ++i) | |
5908 | { | |
5909 | addr = fold_build_pointer_plus_hwi (addr, UNITS_PER_VREG); | |
5910 | u = fold_convert (field_ptr_t, addr); | |
5911 | u = build2 (MODIFY_EXPR, field_t, | |
5912 | build2 (MEM_REF, field_t, tmp_ha, | |
5913 | build_int_cst (field_ptr_t, | |
5914 | (i * | |
5915 | int_size_in_bytes (field_t)))), | |
5916 | build1 (INDIRECT_REF, field_t, u)); | |
5917 | t = build2 (COMPOUND_EXPR, TREE_TYPE (t), t, u); | |
5918 | } | |
5919 | ||
5920 | u = fold_convert (TREE_TYPE (f_top), tmp_ha); | |
5921 | t = build2 (COMPOUND_EXPR, TREE_TYPE (f_top), t, u); | |
5922 | } | |
5923 | ||
5924 | COND_EXPR_ELSE (cond2) = t; | |
5925 | addr = fold_convert (build_pointer_type (type), cond1); | |
5926 | addr = build_va_arg_indirect_ref (addr); | |
5927 | ||
5928 | if (indirect_p) | |
5929 | addr = build_va_arg_indirect_ref (addr); | |
5930 | ||
5931 | return addr; | |
5932 | } | |
5933 | ||
5934 | /* Implement TARGET_SETUP_INCOMING_VARARGS. */ | |
5935 | ||
5936 | static void | |
5937 | aarch64_setup_incoming_varargs (cumulative_args_t cum_v, enum machine_mode mode, | |
5938 | tree type, int *pretend_size ATTRIBUTE_UNUSED, | |
5939 | int no_rtl) | |
5940 | { | |
5941 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
5942 | CUMULATIVE_ARGS local_cum; | |
5943 | int gr_saved, vr_saved; | |
5944 | ||
5945 | /* The caller has advanced CUM up to, but not beyond, the last named | |
5946 | argument. Advance a local copy of CUM past the last "real" named | |
5947 | argument, to find out how many registers are left over. */ | |
5948 | local_cum = *cum; | |
5949 | aarch64_function_arg_advance (pack_cumulative_args(&local_cum), mode, type, true); | |
5950 | ||
5951 | /* Found out how many registers we need to save. */ | |
5952 | gr_saved = NUM_ARG_REGS - local_cum.aapcs_ncrn; | |
5953 | vr_saved = NUM_FP_ARG_REGS - local_cum.aapcs_nvrn; | |
5954 | ||
5955 | if (TARGET_GENERAL_REGS_ONLY) | |
5956 | { | |
5957 | if (local_cum.aapcs_nvrn > 0) | |
5958 | sorry ("%qs and floating point or vector arguments", | |
5959 | "-mgeneral-regs-only"); | |
5960 | vr_saved = 0; | |
5961 | } | |
5962 | ||
5963 | if (!no_rtl) | |
5964 | { | |
5965 | if (gr_saved > 0) | |
5966 | { | |
5967 | rtx ptr, mem; | |
5968 | ||
5969 | /* virtual_incoming_args_rtx should have been 16-byte aligned. */ | |
5970 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, | |
5971 | - gr_saved * UNITS_PER_WORD); | |
5972 | mem = gen_frame_mem (BLKmode, ptr); | |
5973 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
5974 | ||
5975 | move_block_from_reg (local_cum.aapcs_ncrn + R0_REGNUM, | |
5976 | mem, gr_saved); | |
5977 | } | |
5978 | if (vr_saved > 0) | |
5979 | { | |
5980 | /* We can't use move_block_from_reg, because it will use | |
5981 | the wrong mode, storing D regs only. */ | |
5982 | enum machine_mode mode = TImode; | |
5983 | int off, i; | |
5984 | ||
5985 | /* Set OFF to the offset from virtual_incoming_args_rtx of | |
5986 | the first vector register. The VR save area lies below | |
5987 | the GR one, and is aligned to 16 bytes. */ | |
5988 | off = -AARCH64_ROUND_UP (gr_saved * UNITS_PER_WORD, | |
5989 | STACK_BOUNDARY / BITS_PER_UNIT); | |
5990 | off -= vr_saved * UNITS_PER_VREG; | |
5991 | ||
5992 | for (i = local_cum.aapcs_nvrn; i < NUM_FP_ARG_REGS; ++i) | |
5993 | { | |
5994 | rtx ptr, mem; | |
5995 | ||
5996 | ptr = plus_constant (Pmode, virtual_incoming_args_rtx, off); | |
5997 | mem = gen_frame_mem (mode, ptr); | |
5998 | set_mem_alias_set (mem, get_varargs_alias_set ()); | |
5999 | aarch64_emit_move (mem, gen_rtx_REG (mode, V0_REGNUM + i)); | |
6000 | off += UNITS_PER_VREG; | |
6001 | } | |
6002 | } | |
6003 | } | |
6004 | ||
6005 | /* We don't save the size into *PRETEND_SIZE because we want to avoid | |
6006 | any complication of having crtl->args.pretend_args_size changed. */ | |
6007 | cfun->machine->saved_varargs_size | |
6008 | = (AARCH64_ROUND_UP (gr_saved * UNITS_PER_WORD, | |
6009 | STACK_BOUNDARY / BITS_PER_UNIT) | |
6010 | + vr_saved * UNITS_PER_VREG); | |
6011 | } | |
6012 | ||
6013 | static void | |
6014 | aarch64_conditional_register_usage (void) | |
6015 | { | |
6016 | int i; | |
6017 | if (!TARGET_FLOAT) | |
6018 | { | |
6019 | for (i = V0_REGNUM; i <= V31_REGNUM; i++) | |
6020 | { | |
6021 | fixed_regs[i] = 1; | |
6022 | call_used_regs[i] = 1; | |
6023 | } | |
6024 | } | |
6025 | } | |
6026 | ||
6027 | /* Walk down the type tree of TYPE counting consecutive base elements. | |
6028 | If *MODEP is VOIDmode, then set it to the first valid floating point | |
6029 | type. If a non-floating point type is found, or if a floating point | |
6030 | type that doesn't match a non-VOIDmode *MODEP is found, then return -1, | |
6031 | otherwise return the count in the sub-tree. */ | |
6032 | static int | |
6033 | aapcs_vfp_sub_candidate (const_tree type, enum machine_mode *modep) | |
6034 | { | |
6035 | enum machine_mode mode; | |
6036 | HOST_WIDE_INT size; | |
6037 | ||
6038 | switch (TREE_CODE (type)) | |
6039 | { | |
6040 | case REAL_TYPE: | |
6041 | mode = TYPE_MODE (type); | |
6042 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
6043 | return -1; | |
6044 | ||
6045 | if (*modep == VOIDmode) | |
6046 | *modep = mode; | |
6047 | ||
6048 | if (*modep == mode) | |
6049 | return 1; | |
6050 | ||
6051 | break; | |
6052 | ||
6053 | case COMPLEX_TYPE: | |
6054 | mode = TYPE_MODE (TREE_TYPE (type)); | |
6055 | if (mode != DFmode && mode != SFmode && mode != TFmode) | |
6056 | return -1; | |
6057 | ||
6058 | if (*modep == VOIDmode) | |
6059 | *modep = mode; | |
6060 | ||
6061 | if (*modep == mode) | |
6062 | return 2; | |
6063 | ||
6064 | break; | |
6065 | ||
6066 | case VECTOR_TYPE: | |
6067 | /* Use V2SImode and V4SImode as representatives of all 64-bit | |
6068 | and 128-bit vector types. */ | |
6069 | size = int_size_in_bytes (type); | |
6070 | switch (size) | |
6071 | { | |
6072 | case 8: | |
6073 | mode = V2SImode; | |
6074 | break; | |
6075 | case 16: | |
6076 | mode = V4SImode; | |
6077 | break; | |
6078 | default: | |
6079 | return -1; | |
6080 | } | |
6081 | ||
6082 | if (*modep == VOIDmode) | |
6083 | *modep = mode; | |
6084 | ||
6085 | /* Vector modes are considered to be opaque: two vectors are | |
6086 | equivalent for the purposes of being homogeneous aggregates | |
6087 | if they are the same size. */ | |
6088 | if (*modep == mode) | |
6089 | return 1; | |
6090 | ||
6091 | break; | |
6092 | ||
6093 | case ARRAY_TYPE: | |
6094 | { | |
6095 | int count; | |
6096 | tree index = TYPE_DOMAIN (type); | |
6097 | ||
6098 | /* Can't handle incomplete types. */ | |
6099 | if (!COMPLETE_TYPE_P (type)) | |
6100 | return -1; | |
6101 | ||
6102 | count = aapcs_vfp_sub_candidate (TREE_TYPE (type), modep); | |
6103 | if (count == -1 | |
6104 | || !index | |
6105 | || !TYPE_MAX_VALUE (index) | |
cc269bb6 | 6106 | || !tree_fits_uhwi_p (TYPE_MAX_VALUE (index)) |
43e9d192 | 6107 | || !TYPE_MIN_VALUE (index) |
cc269bb6 | 6108 | || !tree_fits_uhwi_p (TYPE_MIN_VALUE (index)) |
43e9d192 IB |
6109 | || count < 0) |
6110 | return -1; | |
6111 | ||
ae7e9ddd RS |
6112 | count *= (1 + tree_to_uhwi (TYPE_MAX_VALUE (index)) |
6113 | - tree_to_uhwi (TYPE_MIN_VALUE (index))); | |
43e9d192 IB |
6114 | |
6115 | /* There must be no padding. */ | |
cc269bb6 | 6116 | if (!tree_fits_uhwi_p (TYPE_SIZE (type)) |
7d362f6c | 6117 | || ((HOST_WIDE_INT) tree_to_uhwi (TYPE_SIZE (type)) |
43e9d192 IB |
6118 | != count * GET_MODE_BITSIZE (*modep))) |
6119 | return -1; | |
6120 | ||
6121 | return count; | |
6122 | } | |
6123 | ||
6124 | case RECORD_TYPE: | |
6125 | { | |
6126 | int count = 0; | |
6127 | int sub_count; | |
6128 | tree field; | |
6129 | ||
6130 | /* Can't handle incomplete types. */ | |
6131 | if (!COMPLETE_TYPE_P (type)) | |
6132 | return -1; | |
6133 | ||
6134 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
6135 | { | |
6136 | if (TREE_CODE (field) != FIELD_DECL) | |
6137 | continue; | |
6138 | ||
6139 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
6140 | if (sub_count < 0) | |
6141 | return -1; | |
6142 | count += sub_count; | |
6143 | } | |
6144 | ||
6145 | /* There must be no padding. */ | |
cc269bb6 | 6146 | if (!tree_fits_uhwi_p (TYPE_SIZE (type)) |
7d362f6c | 6147 | || ((HOST_WIDE_INT) tree_to_uhwi (TYPE_SIZE (type)) |
43e9d192 IB |
6148 | != count * GET_MODE_BITSIZE (*modep))) |
6149 | return -1; | |
6150 | ||
6151 | return count; | |
6152 | } | |
6153 | ||
6154 | case UNION_TYPE: | |
6155 | case QUAL_UNION_TYPE: | |
6156 | { | |
6157 | /* These aren't very interesting except in a degenerate case. */ | |
6158 | int count = 0; | |
6159 | int sub_count; | |
6160 | tree field; | |
6161 | ||
6162 | /* Can't handle incomplete types. */ | |
6163 | if (!COMPLETE_TYPE_P (type)) | |
6164 | return -1; | |
6165 | ||
6166 | for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field)) | |
6167 | { | |
6168 | if (TREE_CODE (field) != FIELD_DECL) | |
6169 | continue; | |
6170 | ||
6171 | sub_count = aapcs_vfp_sub_candidate (TREE_TYPE (field), modep); | |
6172 | if (sub_count < 0) | |
6173 | return -1; | |
6174 | count = count > sub_count ? count : sub_count; | |
6175 | } | |
6176 | ||
6177 | /* There must be no padding. */ | |
cc269bb6 | 6178 | if (!tree_fits_uhwi_p (TYPE_SIZE (type)) |
7d362f6c | 6179 | || ((HOST_WIDE_INT) tree_to_uhwi (TYPE_SIZE (type)) |
43e9d192 IB |
6180 | != count * GET_MODE_BITSIZE (*modep))) |
6181 | return -1; | |
6182 | ||
6183 | return count; | |
6184 | } | |
6185 | ||
6186 | default: | |
6187 | break; | |
6188 | } | |
6189 | ||
6190 | return -1; | |
6191 | } | |
6192 | ||
38e8f663 YR |
6193 | /* Return true if we use LRA instead of reload pass. */ |
6194 | static bool | |
6195 | aarch64_lra_p (void) | |
6196 | { | |
6197 | return aarch64_lra_flag; | |
6198 | } | |
6199 | ||
43e9d192 IB |
6200 | /* Return TRUE if the type, as described by TYPE and MODE, is a composite |
6201 | type as described in AAPCS64 \S 4.3. This includes aggregate, union and | |
6202 | array types. The C99 floating-point complex types are also considered | |
6203 | as composite types, according to AAPCS64 \S 7.1.1. The complex integer | |
6204 | types, which are GCC extensions and out of the scope of AAPCS64, are | |
6205 | treated as composite types here as well. | |
6206 | ||
6207 | Note that MODE itself is not sufficient in determining whether a type | |
6208 | is such a composite type or not. This is because | |
6209 | stor-layout.c:compute_record_mode may have already changed the MODE | |
6210 | (BLKmode) of a RECORD_TYPE TYPE to some other mode. For example, a | |
6211 | structure with only one field may have its MODE set to the mode of the | |
6212 | field. Also an integer mode whose size matches the size of the | |
6213 | RECORD_TYPE type may be used to substitute the original mode | |
6214 | (i.e. BLKmode) in certain circumstances. In other words, MODE cannot be | |
6215 | solely relied on. */ | |
6216 | ||
6217 | static bool | |
6218 | aarch64_composite_type_p (const_tree type, | |
6219 | enum machine_mode mode) | |
6220 | { | |
6221 | if (type && (AGGREGATE_TYPE_P (type) || TREE_CODE (type) == COMPLEX_TYPE)) | |
6222 | return true; | |
6223 | ||
6224 | if (mode == BLKmode | |
6225 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT | |
6226 | || GET_MODE_CLASS (mode) == MODE_COMPLEX_INT) | |
6227 | return true; | |
6228 | ||
6229 | return false; | |
6230 | } | |
6231 | ||
6232 | /* Return TRUE if the type, as described by TYPE and MODE, is a short vector | |
6233 | type as described in AAPCS64 \S 4.1.2. | |
6234 | ||
6235 | See the comment above aarch64_composite_type_p for the notes on MODE. */ | |
6236 | ||
6237 | static bool | |
6238 | aarch64_short_vector_p (const_tree type, | |
6239 | enum machine_mode mode) | |
6240 | { | |
6241 | HOST_WIDE_INT size = -1; | |
6242 | ||
6243 | if (type && TREE_CODE (type) == VECTOR_TYPE) | |
6244 | size = int_size_in_bytes (type); | |
6245 | else if (!aarch64_composite_type_p (type, mode) | |
6246 | && (GET_MODE_CLASS (mode) == MODE_VECTOR_INT | |
6247 | || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)) | |
6248 | size = GET_MODE_SIZE (mode); | |
6249 | ||
6250 | return (size == 8 || size == 16) ? true : false; | |
6251 | } | |
6252 | ||
6253 | /* Return TRUE if an argument, whose type is described by TYPE and MODE, | |
6254 | shall be passed or returned in simd/fp register(s) (providing these | |
6255 | parameter passing registers are available). | |
6256 | ||
6257 | Upon successful return, *COUNT returns the number of needed registers, | |
6258 | *BASE_MODE returns the mode of the individual register and when IS_HAF | |
6259 | is not NULL, *IS_HA indicates whether or not the argument is a homogeneous | |
6260 | floating-point aggregate or a homogeneous short-vector aggregate. */ | |
6261 | ||
6262 | static bool | |
6263 | aarch64_vfp_is_call_or_return_candidate (enum machine_mode mode, | |
6264 | const_tree type, | |
6265 | enum machine_mode *base_mode, | |
6266 | int *count, | |
6267 | bool *is_ha) | |
6268 | { | |
6269 | enum machine_mode new_mode = VOIDmode; | |
6270 | bool composite_p = aarch64_composite_type_p (type, mode); | |
6271 | ||
6272 | if (is_ha != NULL) *is_ha = false; | |
6273 | ||
6274 | if ((!composite_p && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
6275 | || aarch64_short_vector_p (type, mode)) | |
6276 | { | |
6277 | *count = 1; | |
6278 | new_mode = mode; | |
6279 | } | |
6280 | else if (GET_MODE_CLASS (mode) == MODE_COMPLEX_FLOAT) | |
6281 | { | |
6282 | if (is_ha != NULL) *is_ha = true; | |
6283 | *count = 2; | |
6284 | new_mode = GET_MODE_INNER (mode); | |
6285 | } | |
6286 | else if (type && composite_p) | |
6287 | { | |
6288 | int ag_count = aapcs_vfp_sub_candidate (type, &new_mode); | |
6289 | ||
6290 | if (ag_count > 0 && ag_count <= HA_MAX_NUM_FLDS) | |
6291 | { | |
6292 | if (is_ha != NULL) *is_ha = true; | |
6293 | *count = ag_count; | |
6294 | } | |
6295 | else | |
6296 | return false; | |
6297 | } | |
6298 | else | |
6299 | return false; | |
6300 | ||
6301 | *base_mode = new_mode; | |
6302 | return true; | |
6303 | } | |
6304 | ||
6305 | /* Implement TARGET_STRUCT_VALUE_RTX. */ | |
6306 | ||
6307 | static rtx | |
6308 | aarch64_struct_value_rtx (tree fndecl ATTRIBUTE_UNUSED, | |
6309 | int incoming ATTRIBUTE_UNUSED) | |
6310 | { | |
6311 | return gen_rtx_REG (Pmode, AARCH64_STRUCT_VALUE_REGNUM); | |
6312 | } | |
6313 | ||
6314 | /* Implements target hook vector_mode_supported_p. */ | |
6315 | static bool | |
6316 | aarch64_vector_mode_supported_p (enum machine_mode mode) | |
6317 | { | |
6318 | if (TARGET_SIMD | |
6319 | && (mode == V4SImode || mode == V8HImode | |
6320 | || mode == V16QImode || mode == V2DImode | |
6321 | || mode == V2SImode || mode == V4HImode | |
6322 | || mode == V8QImode || mode == V2SFmode | |
6323 | || mode == V4SFmode || mode == V2DFmode)) | |
6324 | return true; | |
6325 | ||
6326 | return false; | |
6327 | } | |
6328 | ||
b7342d25 IB |
6329 | /* Return appropriate SIMD container |
6330 | for MODE within a vector of WIDTH bits. */ | |
43e9d192 | 6331 | static enum machine_mode |
b7342d25 | 6332 | aarch64_simd_container_mode (enum machine_mode mode, unsigned width) |
43e9d192 | 6333 | { |
b7342d25 | 6334 | gcc_assert (width == 64 || width == 128); |
43e9d192 | 6335 | if (TARGET_SIMD) |
b7342d25 IB |
6336 | { |
6337 | if (width == 128) | |
6338 | switch (mode) | |
6339 | { | |
6340 | case DFmode: | |
6341 | return V2DFmode; | |
6342 | case SFmode: | |
6343 | return V4SFmode; | |
6344 | case SImode: | |
6345 | return V4SImode; | |
6346 | case HImode: | |
6347 | return V8HImode; | |
6348 | case QImode: | |
6349 | return V16QImode; | |
6350 | case DImode: | |
6351 | return V2DImode; | |
6352 | default: | |
6353 | break; | |
6354 | } | |
6355 | else | |
6356 | switch (mode) | |
6357 | { | |
6358 | case SFmode: | |
6359 | return V2SFmode; | |
6360 | case SImode: | |
6361 | return V2SImode; | |
6362 | case HImode: | |
6363 | return V4HImode; | |
6364 | case QImode: | |
6365 | return V8QImode; | |
6366 | default: | |
6367 | break; | |
6368 | } | |
6369 | } | |
43e9d192 IB |
6370 | return word_mode; |
6371 | } | |
6372 | ||
b7342d25 IB |
6373 | /* Return 128-bit container as the preferred SIMD mode for MODE. */ |
6374 | static enum machine_mode | |
6375 | aarch64_preferred_simd_mode (enum machine_mode mode) | |
6376 | { | |
6377 | return aarch64_simd_container_mode (mode, 128); | |
6378 | } | |
6379 | ||
3b357264 JG |
6380 | /* Return the bitmask of possible vector sizes for the vectorizer |
6381 | to iterate over. */ | |
6382 | static unsigned int | |
6383 | aarch64_autovectorize_vector_sizes (void) | |
6384 | { | |
6385 | return (16 | 8); | |
6386 | } | |
6387 | ||
c6fc9e43 YZ |
6388 | /* A table to help perform AArch64-specific name mangling for AdvSIMD |
6389 | vector types in order to conform to the AAPCS64 (see "Procedure | |
6390 | Call Standard for the ARM 64-bit Architecture", Appendix A). To | |
6391 | qualify for emission with the mangled names defined in that document, | |
6392 | a vector type must not only be of the correct mode but also be | |
6393 | composed of AdvSIMD vector element types (e.g. | |
6394 | _builtin_aarch64_simd_qi); these types are registered by | |
6395 | aarch64_init_simd_builtins (). In other words, vector types defined | |
6396 | in other ways e.g. via vector_size attribute will get default | |
6397 | mangled names. */ | |
6398 | typedef struct | |
6399 | { | |
6400 | enum machine_mode mode; | |
6401 | const char *element_type_name; | |
6402 | const char *mangled_name; | |
6403 | } aarch64_simd_mangle_map_entry; | |
6404 | ||
6405 | static aarch64_simd_mangle_map_entry aarch64_simd_mangle_map[] = { | |
6406 | /* 64-bit containerized types. */ | |
6407 | { V8QImode, "__builtin_aarch64_simd_qi", "10__Int8x8_t" }, | |
6408 | { V8QImode, "__builtin_aarch64_simd_uqi", "11__Uint8x8_t" }, | |
6409 | { V4HImode, "__builtin_aarch64_simd_hi", "11__Int16x4_t" }, | |
6410 | { V4HImode, "__builtin_aarch64_simd_uhi", "12__Uint16x4_t" }, | |
6411 | { V2SImode, "__builtin_aarch64_simd_si", "11__Int32x2_t" }, | |
6412 | { V2SImode, "__builtin_aarch64_simd_usi", "12__Uint32x2_t" }, | |
6413 | { V2SFmode, "__builtin_aarch64_simd_sf", "13__Float32x2_t" }, | |
6414 | { V8QImode, "__builtin_aarch64_simd_poly8", "11__Poly8x8_t" }, | |
6415 | { V4HImode, "__builtin_aarch64_simd_poly16", "12__Poly16x4_t" }, | |
6416 | /* 128-bit containerized types. */ | |
6417 | { V16QImode, "__builtin_aarch64_simd_qi", "11__Int8x16_t" }, | |
6418 | { V16QImode, "__builtin_aarch64_simd_uqi", "12__Uint8x16_t" }, | |
6419 | { V8HImode, "__builtin_aarch64_simd_hi", "11__Int16x8_t" }, | |
6420 | { V8HImode, "__builtin_aarch64_simd_uhi", "12__Uint16x8_t" }, | |
6421 | { V4SImode, "__builtin_aarch64_simd_si", "11__Int32x4_t" }, | |
6422 | { V4SImode, "__builtin_aarch64_simd_usi", "12__Uint32x4_t" }, | |
6423 | { V2DImode, "__builtin_aarch64_simd_di", "11__Int64x2_t" }, | |
6424 | { V2DImode, "__builtin_aarch64_simd_udi", "12__Uint64x2_t" }, | |
6425 | { V4SFmode, "__builtin_aarch64_simd_sf", "13__Float32x4_t" }, | |
6426 | { V2DFmode, "__builtin_aarch64_simd_df", "13__Float64x2_t" }, | |
6427 | { V16QImode, "__builtin_aarch64_simd_poly8", "12__Poly8x16_t" }, | |
6428 | { V8HImode, "__builtin_aarch64_simd_poly16", "12__Poly16x8_t" }, | |
7baa225d | 6429 | { V2DImode, "__builtin_aarch64_simd_poly64", "12__Poly64x2_t" }, |
c6fc9e43 YZ |
6430 | { VOIDmode, NULL, NULL } |
6431 | }; | |
6432 | ||
ac2b960f YZ |
6433 | /* Implement TARGET_MANGLE_TYPE. */ |
6434 | ||
6f549691 | 6435 | static const char * |
ac2b960f YZ |
6436 | aarch64_mangle_type (const_tree type) |
6437 | { | |
6438 | /* The AArch64 ABI documents say that "__va_list" has to be | |
6439 | managled as if it is in the "std" namespace. */ | |
6440 | if (lang_hooks.types_compatible_p (CONST_CAST_TREE (type), va_list_type)) | |
6441 | return "St9__va_list"; | |
6442 | ||
c6fc9e43 YZ |
6443 | /* Check the mode of the vector type, and the name of the vector |
6444 | element type, against the table. */ | |
6445 | if (TREE_CODE (type) == VECTOR_TYPE) | |
6446 | { | |
6447 | aarch64_simd_mangle_map_entry *pos = aarch64_simd_mangle_map; | |
6448 | ||
6449 | while (pos->mode != VOIDmode) | |
6450 | { | |
6451 | tree elt_type = TREE_TYPE (type); | |
6452 | ||
6453 | if (pos->mode == TYPE_MODE (type) | |
6454 | && TREE_CODE (TYPE_NAME (elt_type)) == TYPE_DECL | |
6455 | && !strcmp (IDENTIFIER_POINTER (DECL_NAME (TYPE_NAME (elt_type))), | |
6456 | pos->element_type_name)) | |
6457 | return pos->mangled_name; | |
6458 | ||
6459 | pos++; | |
6460 | } | |
6461 | } | |
6462 | ||
ac2b960f YZ |
6463 | /* Use the default mangling. */ |
6464 | return NULL; | |
6465 | } | |
6466 | ||
43e9d192 | 6467 | /* Return the equivalent letter for size. */ |
81c2dfb9 | 6468 | static char |
43e9d192 IB |
6469 | sizetochar (int size) |
6470 | { | |
6471 | switch (size) | |
6472 | { | |
6473 | case 64: return 'd'; | |
6474 | case 32: return 's'; | |
6475 | case 16: return 'h'; | |
6476 | case 8 : return 'b'; | |
6477 | default: gcc_unreachable (); | |
6478 | } | |
6479 | } | |
6480 | ||
3520f7cc JG |
6481 | /* Return true iff x is a uniform vector of floating-point |
6482 | constants, and the constant can be represented in | |
6483 | quarter-precision form. Note, as aarch64_float_const_representable | |
6484 | rejects both +0.0 and -0.0, we will also reject +0.0 and -0.0. */ | |
6485 | static bool | |
6486 | aarch64_vect_float_const_representable_p (rtx x) | |
6487 | { | |
6488 | int i = 0; | |
6489 | REAL_VALUE_TYPE r0, ri; | |
6490 | rtx x0, xi; | |
6491 | ||
6492 | if (GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_FLOAT) | |
6493 | return false; | |
6494 | ||
6495 | x0 = CONST_VECTOR_ELT (x, 0); | |
6496 | if (!CONST_DOUBLE_P (x0)) | |
6497 | return false; | |
6498 | ||
6499 | REAL_VALUE_FROM_CONST_DOUBLE (r0, x0); | |
6500 | ||
6501 | for (i = 1; i < CONST_VECTOR_NUNITS (x); i++) | |
6502 | { | |
6503 | xi = CONST_VECTOR_ELT (x, i); | |
6504 | if (!CONST_DOUBLE_P (xi)) | |
6505 | return false; | |
6506 | ||
6507 | REAL_VALUE_FROM_CONST_DOUBLE (ri, xi); | |
6508 | if (!REAL_VALUES_EQUAL (r0, ri)) | |
6509 | return false; | |
6510 | } | |
6511 | ||
6512 | return aarch64_float_const_representable_p (x0); | |
6513 | } | |
6514 | ||
d8edd899 | 6515 | /* Return true for valid and false for invalid. */ |
3ea63f60 | 6516 | bool |
48063b9d IB |
6517 | aarch64_simd_valid_immediate (rtx op, enum machine_mode mode, bool inverse, |
6518 | struct simd_immediate_info *info) | |
43e9d192 IB |
6519 | { |
6520 | #define CHECK(STRIDE, ELSIZE, CLASS, TEST, SHIFT, NEG) \ | |
6521 | matches = 1; \ | |
6522 | for (i = 0; i < idx; i += (STRIDE)) \ | |
6523 | if (!(TEST)) \ | |
6524 | matches = 0; \ | |
6525 | if (matches) \ | |
6526 | { \ | |
6527 | immtype = (CLASS); \ | |
6528 | elsize = (ELSIZE); \ | |
43e9d192 IB |
6529 | eshift = (SHIFT); \ |
6530 | emvn = (NEG); \ | |
6531 | break; \ | |
6532 | } | |
6533 | ||
6534 | unsigned int i, elsize = 0, idx = 0, n_elts = CONST_VECTOR_NUNITS (op); | |
6535 | unsigned int innersize = GET_MODE_SIZE (GET_MODE_INNER (mode)); | |
6536 | unsigned char bytes[16]; | |
43e9d192 IB |
6537 | int immtype = -1, matches; |
6538 | unsigned int invmask = inverse ? 0xff : 0; | |
6539 | int eshift, emvn; | |
6540 | ||
43e9d192 | 6541 | if (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT) |
3520f7cc | 6542 | { |
81c2dfb9 IB |
6543 | if (! (aarch64_simd_imm_zero_p (op, mode) |
6544 | || aarch64_vect_float_const_representable_p (op))) | |
d8edd899 | 6545 | return false; |
3520f7cc | 6546 | |
48063b9d IB |
6547 | if (info) |
6548 | { | |
6549 | info->value = CONST_VECTOR_ELT (op, 0); | |
81c2dfb9 | 6550 | info->element_width = GET_MODE_BITSIZE (GET_MODE (info->value)); |
48063b9d IB |
6551 | info->mvn = false; |
6552 | info->shift = 0; | |
6553 | } | |
3520f7cc | 6554 | |
d8edd899 | 6555 | return true; |
3520f7cc | 6556 | } |
43e9d192 IB |
6557 | |
6558 | /* Splat vector constant out into a byte vector. */ | |
6559 | for (i = 0; i < n_elts; i++) | |
6560 | { | |
6561 | rtx el = CONST_VECTOR_ELT (op, i); | |
6562 | unsigned HOST_WIDE_INT elpart; | |
6563 | unsigned int part, parts; | |
6564 | ||
6565 | if (GET_CODE (el) == CONST_INT) | |
6566 | { | |
6567 | elpart = INTVAL (el); | |
6568 | parts = 1; | |
6569 | } | |
6570 | else if (GET_CODE (el) == CONST_DOUBLE) | |
6571 | { | |
6572 | elpart = CONST_DOUBLE_LOW (el); | |
6573 | parts = 2; | |
6574 | } | |
6575 | else | |
6576 | gcc_unreachable (); | |
6577 | ||
6578 | for (part = 0; part < parts; part++) | |
6579 | { | |
6580 | unsigned int byte; | |
6581 | for (byte = 0; byte < innersize; byte++) | |
6582 | { | |
6583 | bytes[idx++] = (elpart & 0xff) ^ invmask; | |
6584 | elpart >>= BITS_PER_UNIT; | |
6585 | } | |
6586 | if (GET_CODE (el) == CONST_DOUBLE) | |
6587 | elpart = CONST_DOUBLE_HIGH (el); | |
6588 | } | |
6589 | } | |
6590 | ||
6591 | /* Sanity check. */ | |
6592 | gcc_assert (idx == GET_MODE_SIZE (mode)); | |
6593 | ||
6594 | do | |
6595 | { | |
6596 | CHECK (4, 32, 0, bytes[i] == bytes[0] && bytes[i + 1] == 0 | |
6597 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 0, 0); | |
6598 | ||
6599 | CHECK (4, 32, 1, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
6600 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); | |
6601 | ||
6602 | CHECK (4, 32, 2, bytes[i] == 0 && bytes[i + 1] == 0 | |
6603 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); | |
6604 | ||
6605 | CHECK (4, 32, 3, bytes[i] == 0 && bytes[i + 1] == 0 | |
6606 | && bytes[i + 2] == 0 && bytes[i + 3] == bytes[3], 24, 0); | |
6607 | ||
6608 | CHECK (2, 16, 4, bytes[i] == bytes[0] && bytes[i + 1] == 0, 0, 0); | |
6609 | ||
6610 | CHECK (2, 16, 5, bytes[i] == 0 && bytes[i + 1] == bytes[1], 8, 0); | |
6611 | ||
6612 | CHECK (4, 32, 6, bytes[i] == bytes[0] && bytes[i + 1] == 0xff | |
6613 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 0, 1); | |
6614 | ||
6615 | CHECK (4, 32, 7, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
6616 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); | |
6617 | ||
6618 | CHECK (4, 32, 8, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6619 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); | |
6620 | ||
6621 | CHECK (4, 32, 9, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
6622 | && bytes[i + 2] == 0xff && bytes[i + 3] == bytes[3], 24, 1); | |
6623 | ||
6624 | CHECK (2, 16, 10, bytes[i] == bytes[0] && bytes[i + 1] == 0xff, 0, 1); | |
6625 | ||
6626 | CHECK (2, 16, 11, bytes[i] == 0xff && bytes[i + 1] == bytes[1], 8, 1); | |
6627 | ||
6628 | CHECK (4, 32, 12, bytes[i] == 0xff && bytes[i + 1] == bytes[1] | |
e4f0f84d | 6629 | && bytes[i + 2] == 0 && bytes[i + 3] == 0, 8, 0); |
43e9d192 IB |
6630 | |
6631 | CHECK (4, 32, 13, bytes[i] == 0 && bytes[i + 1] == bytes[1] | |
e4f0f84d | 6632 | && bytes[i + 2] == 0xff && bytes[i + 3] == 0xff, 8, 1); |
43e9d192 IB |
6633 | |
6634 | CHECK (4, 32, 14, bytes[i] == 0xff && bytes[i + 1] == 0xff | |
e4f0f84d | 6635 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0, 16, 0); |
43e9d192 IB |
6636 | |
6637 | CHECK (4, 32, 15, bytes[i] == 0 && bytes[i + 1] == 0 | |
e4f0f84d | 6638 | && bytes[i + 2] == bytes[2] && bytes[i + 3] == 0xff, 16, 1); |
43e9d192 IB |
6639 | |
6640 | CHECK (1, 8, 16, bytes[i] == bytes[0], 0, 0); | |
6641 | ||
6642 | CHECK (1, 64, 17, (bytes[i] == 0 || bytes[i] == 0xff) | |
6643 | && bytes[i] == bytes[(i + 8) % idx], 0, 0); | |
6644 | } | |
6645 | while (0); | |
6646 | ||
e4f0f84d | 6647 | if (immtype == -1) |
d8edd899 | 6648 | return false; |
43e9d192 | 6649 | |
48063b9d | 6650 | if (info) |
43e9d192 | 6651 | { |
48063b9d | 6652 | info->element_width = elsize; |
48063b9d IB |
6653 | info->mvn = emvn != 0; |
6654 | info->shift = eshift; | |
6655 | ||
43e9d192 IB |
6656 | unsigned HOST_WIDE_INT imm = 0; |
6657 | ||
e4f0f84d TB |
6658 | if (immtype >= 12 && immtype <= 15) |
6659 | info->msl = true; | |
6660 | ||
43e9d192 IB |
6661 | /* Un-invert bytes of recognized vector, if necessary. */ |
6662 | if (invmask != 0) | |
6663 | for (i = 0; i < idx; i++) | |
6664 | bytes[i] ^= invmask; | |
6665 | ||
6666 | if (immtype == 17) | |
6667 | { | |
6668 | /* FIXME: Broken on 32-bit H_W_I hosts. */ | |
6669 | gcc_assert (sizeof (HOST_WIDE_INT) == 8); | |
6670 | ||
6671 | for (i = 0; i < 8; i++) | |
6672 | imm |= (unsigned HOST_WIDE_INT) (bytes[i] ? 0xff : 0) | |
6673 | << (i * BITS_PER_UNIT); | |
6674 | ||
43e9d192 | 6675 | |
48063b9d IB |
6676 | info->value = GEN_INT (imm); |
6677 | } | |
6678 | else | |
6679 | { | |
6680 | for (i = 0; i < elsize / BITS_PER_UNIT; i++) | |
6681 | imm |= (unsigned HOST_WIDE_INT) bytes[i] << (i * BITS_PER_UNIT); | |
43e9d192 IB |
6682 | |
6683 | /* Construct 'abcdefgh' because the assembler cannot handle | |
48063b9d IB |
6684 | generic constants. */ |
6685 | if (info->mvn) | |
43e9d192 | 6686 | imm = ~imm; |
48063b9d IB |
6687 | imm = (imm >> info->shift) & 0xff; |
6688 | info->value = GEN_INT (imm); | |
6689 | } | |
43e9d192 IB |
6690 | } |
6691 | ||
48063b9d | 6692 | return true; |
43e9d192 IB |
6693 | #undef CHECK |
6694 | } | |
6695 | ||
43e9d192 IB |
6696 | static bool |
6697 | aarch64_const_vec_all_same_int_p (rtx x, | |
6698 | HOST_WIDE_INT minval, | |
6699 | HOST_WIDE_INT maxval) | |
6700 | { | |
6701 | HOST_WIDE_INT firstval; | |
6702 | int count, i; | |
6703 | ||
6704 | if (GET_CODE (x) != CONST_VECTOR | |
6705 | || GET_MODE_CLASS (GET_MODE (x)) != MODE_VECTOR_INT) | |
6706 | return false; | |
6707 | ||
6708 | firstval = INTVAL (CONST_VECTOR_ELT (x, 0)); | |
6709 | if (firstval < minval || firstval > maxval) | |
6710 | return false; | |
6711 | ||
6712 | count = CONST_VECTOR_NUNITS (x); | |
6713 | for (i = 1; i < count; i++) | |
6714 | if (INTVAL (CONST_VECTOR_ELT (x, i)) != firstval) | |
6715 | return false; | |
6716 | ||
6717 | return true; | |
6718 | } | |
6719 | ||
6720 | /* Check of immediate shift constants are within range. */ | |
6721 | bool | |
6722 | aarch64_simd_shift_imm_p (rtx x, enum machine_mode mode, bool left) | |
6723 | { | |
6724 | int bit_width = GET_MODE_UNIT_SIZE (mode) * BITS_PER_UNIT; | |
6725 | if (left) | |
6726 | return aarch64_const_vec_all_same_int_p (x, 0, bit_width - 1); | |
6727 | else | |
6728 | return aarch64_const_vec_all_same_int_p (x, 1, bit_width); | |
6729 | } | |
6730 | ||
3520f7cc JG |
6731 | /* Return true if X is a uniform vector where all elements |
6732 | are either the floating-point constant 0.0 or the | |
6733 | integer constant 0. */ | |
43e9d192 IB |
6734 | bool |
6735 | aarch64_simd_imm_zero_p (rtx x, enum machine_mode mode) | |
6736 | { | |
3520f7cc | 6737 | return x == CONST0_RTX (mode); |
43e9d192 IB |
6738 | } |
6739 | ||
6740 | bool | |
6741 | aarch64_simd_imm_scalar_p (rtx x, enum machine_mode mode ATTRIBUTE_UNUSED) | |
6742 | { | |
6743 | HOST_WIDE_INT imm = INTVAL (x); | |
6744 | int i; | |
6745 | ||
6746 | for (i = 0; i < 8; i++) | |
6747 | { | |
6748 | unsigned int byte = imm & 0xff; | |
6749 | if (byte != 0xff && byte != 0) | |
6750 | return false; | |
6751 | imm >>= 8; | |
6752 | } | |
6753 | ||
6754 | return true; | |
6755 | } | |
6756 | ||
83f8c414 CSS |
6757 | bool |
6758 | aarch64_mov_operand_p (rtx x, | |
a5350ddc | 6759 | enum aarch64_symbol_context context, |
83f8c414 CSS |
6760 | enum machine_mode mode) |
6761 | { | |
83f8c414 CSS |
6762 | if (GET_CODE (x) == HIGH |
6763 | && aarch64_valid_symref (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) | |
6764 | return true; | |
6765 | ||
6766 | if (CONST_INT_P (x) && aarch64_move_imm (INTVAL (x), mode)) | |
6767 | return true; | |
6768 | ||
6769 | if (GET_CODE (x) == SYMBOL_REF && mode == DImode && CONSTANT_ADDRESS_P (x)) | |
6770 | return true; | |
6771 | ||
a5350ddc CSS |
6772 | return aarch64_classify_symbolic_expression (x, context) |
6773 | == SYMBOL_TINY_ABSOLUTE; | |
83f8c414 CSS |
6774 | } |
6775 | ||
43e9d192 IB |
6776 | /* Return a const_int vector of VAL. */ |
6777 | rtx | |
6778 | aarch64_simd_gen_const_vector_dup (enum machine_mode mode, int val) | |
6779 | { | |
6780 | int nunits = GET_MODE_NUNITS (mode); | |
6781 | rtvec v = rtvec_alloc (nunits); | |
6782 | int i; | |
6783 | ||
6784 | for (i=0; i < nunits; i++) | |
6785 | RTVEC_ELT (v, i) = GEN_INT (val); | |
6786 | ||
6787 | return gen_rtx_CONST_VECTOR (mode, v); | |
6788 | } | |
6789 | ||
051d0e2f SN |
6790 | /* Check OP is a legal scalar immediate for the MOVI instruction. */ |
6791 | ||
6792 | bool | |
6793 | aarch64_simd_scalar_immediate_valid_for_move (rtx op, enum machine_mode mode) | |
6794 | { | |
6795 | enum machine_mode vmode; | |
6796 | ||
6797 | gcc_assert (!VECTOR_MODE_P (mode)); | |
6798 | vmode = aarch64_preferred_simd_mode (mode); | |
6799 | rtx op_v = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (op)); | |
48063b9d | 6800 | return aarch64_simd_valid_immediate (op_v, vmode, false, NULL); |
051d0e2f SN |
6801 | } |
6802 | ||
43e9d192 IB |
6803 | /* Construct and return a PARALLEL RTX vector. */ |
6804 | rtx | |
6805 | aarch64_simd_vect_par_cnst_half (enum machine_mode mode, bool high) | |
6806 | { | |
6807 | int nunits = GET_MODE_NUNITS (mode); | |
6808 | rtvec v = rtvec_alloc (nunits / 2); | |
6809 | int base = high ? nunits / 2 : 0; | |
6810 | rtx t1; | |
6811 | int i; | |
6812 | ||
6813 | for (i=0; i < nunits / 2; i++) | |
6814 | RTVEC_ELT (v, i) = GEN_INT (base + i); | |
6815 | ||
6816 | t1 = gen_rtx_PARALLEL (mode, v); | |
6817 | return t1; | |
6818 | } | |
6819 | ||
6820 | /* Bounds-check lanes. Ensure OPERAND lies between LOW (inclusive) and | |
6821 | HIGH (exclusive). */ | |
6822 | void | |
6823 | aarch64_simd_lane_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high) | |
6824 | { | |
6825 | HOST_WIDE_INT lane; | |
6826 | gcc_assert (GET_CODE (operand) == CONST_INT); | |
6827 | lane = INTVAL (operand); | |
6828 | ||
6829 | if (lane < low || lane >= high) | |
6830 | error ("lane out of range"); | |
6831 | } | |
6832 | ||
6833 | void | |
6834 | aarch64_simd_const_bounds (rtx operand, HOST_WIDE_INT low, HOST_WIDE_INT high) | |
6835 | { | |
6836 | gcc_assert (GET_CODE (operand) == CONST_INT); | |
6837 | HOST_WIDE_INT lane = INTVAL (operand); | |
6838 | ||
6839 | if (lane < low || lane >= high) | |
6840 | error ("constant out of range"); | |
6841 | } | |
6842 | ||
6843 | /* Emit code to reinterpret one AdvSIMD type as another, | |
6844 | without altering bits. */ | |
6845 | void | |
6846 | aarch64_simd_reinterpret (rtx dest, rtx src) | |
6847 | { | |
6848 | emit_move_insn (dest, gen_lowpart (GET_MODE (dest), src)); | |
6849 | } | |
6850 | ||
6851 | /* Emit code to place a AdvSIMD pair result in memory locations (with equal | |
6852 | registers). */ | |
6853 | void | |
6854 | aarch64_simd_emit_pair_result_insn (enum machine_mode mode, | |
6855 | rtx (*intfn) (rtx, rtx, rtx), rtx destaddr, | |
6856 | rtx op1) | |
6857 | { | |
6858 | rtx mem = gen_rtx_MEM (mode, destaddr); | |
6859 | rtx tmp1 = gen_reg_rtx (mode); | |
6860 | rtx tmp2 = gen_reg_rtx (mode); | |
6861 | ||
6862 | emit_insn (intfn (tmp1, op1, tmp2)); | |
6863 | ||
6864 | emit_move_insn (mem, tmp1); | |
6865 | mem = adjust_address (mem, mode, GET_MODE_SIZE (mode)); | |
6866 | emit_move_insn (mem, tmp2); | |
6867 | } | |
6868 | ||
6869 | /* Return TRUE if OP is a valid vector addressing mode. */ | |
6870 | bool | |
6871 | aarch64_simd_mem_operand_p (rtx op) | |
6872 | { | |
6873 | return MEM_P (op) && (GET_CODE (XEXP (op, 0)) == POST_INC | |
6874 | || GET_CODE (XEXP (op, 0)) == REG); | |
6875 | } | |
6876 | ||
6877 | /* Set up OPERANDS for a register copy from SRC to DEST, taking care | |
6878 | not to early-clobber SRC registers in the process. | |
6879 | ||
6880 | We assume that the operands described by SRC and DEST represent a | |
6881 | decomposed copy of OPERANDS[1] into OPERANDS[0]. COUNT is the | |
6882 | number of components into which the copy has been decomposed. */ | |
6883 | void | |
6884 | aarch64_simd_disambiguate_copy (rtx *operands, rtx *dest, | |
6885 | rtx *src, unsigned int count) | |
6886 | { | |
6887 | unsigned int i; | |
6888 | ||
6889 | if (!reg_overlap_mentioned_p (operands[0], operands[1]) | |
6890 | || REGNO (operands[0]) < REGNO (operands[1])) | |
6891 | { | |
6892 | for (i = 0; i < count; i++) | |
6893 | { | |
6894 | operands[2 * i] = dest[i]; | |
6895 | operands[2 * i + 1] = src[i]; | |
6896 | } | |
6897 | } | |
6898 | else | |
6899 | { | |
6900 | for (i = 0; i < count; i++) | |
6901 | { | |
6902 | operands[2 * i] = dest[count - i - 1]; | |
6903 | operands[2 * i + 1] = src[count - i - 1]; | |
6904 | } | |
6905 | } | |
6906 | } | |
6907 | ||
6908 | /* Compute and return the length of aarch64_simd_mov<mode>, where <mode> is | |
6909 | one of VSTRUCT modes: OI, CI or XI. */ | |
6910 | int | |
6911 | aarch64_simd_attr_length_move (rtx insn) | |
6912 | { | |
43e9d192 IB |
6913 | enum machine_mode mode; |
6914 | ||
6915 | extract_insn_cached (insn); | |
6916 | ||
6917 | if (REG_P (recog_data.operand[0]) && REG_P (recog_data.operand[1])) | |
6918 | { | |
6919 | mode = GET_MODE (recog_data.operand[0]); | |
6920 | switch (mode) | |
6921 | { | |
6922 | case OImode: | |
6923 | return 8; | |
6924 | case CImode: | |
6925 | return 12; | |
6926 | case XImode: | |
6927 | return 16; | |
6928 | default: | |
6929 | gcc_unreachable (); | |
6930 | } | |
6931 | } | |
6932 | return 4; | |
6933 | } | |
6934 | ||
db0253a4 TB |
6935 | /* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum |
6936 | alignment of a vector to 128 bits. */ | |
6937 | static HOST_WIDE_INT | |
6938 | aarch64_simd_vector_alignment (const_tree type) | |
6939 | { | |
9439e9a1 | 6940 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); |
db0253a4 TB |
6941 | return MIN (align, 128); |
6942 | } | |
6943 | ||
6944 | /* Implement target hook TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE. */ | |
6945 | static bool | |
6946 | aarch64_simd_vector_alignment_reachable (const_tree type, bool is_packed) | |
6947 | { | |
6948 | if (is_packed) | |
6949 | return false; | |
6950 | ||
6951 | /* We guarantee alignment for vectors up to 128-bits. */ | |
6952 | if (tree_int_cst_compare (TYPE_SIZE (type), | |
6953 | bitsize_int (BIGGEST_ALIGNMENT)) > 0) | |
6954 | return false; | |
6955 | ||
6956 | /* Vectors whose size is <= BIGGEST_ALIGNMENT are naturally aligned. */ | |
6957 | return true; | |
6958 | } | |
6959 | ||
4369c11e TB |
6960 | /* If VALS is a vector constant that can be loaded into a register |
6961 | using DUP, generate instructions to do so and return an RTX to | |
6962 | assign to the register. Otherwise return NULL_RTX. */ | |
6963 | static rtx | |
6964 | aarch64_simd_dup_constant (rtx vals) | |
6965 | { | |
6966 | enum machine_mode mode = GET_MODE (vals); | |
6967 | enum machine_mode inner_mode = GET_MODE_INNER (mode); | |
6968 | int n_elts = GET_MODE_NUNITS (mode); | |
6969 | bool all_same = true; | |
6970 | rtx x; | |
6971 | int i; | |
6972 | ||
6973 | if (GET_CODE (vals) != CONST_VECTOR) | |
6974 | return NULL_RTX; | |
6975 | ||
6976 | for (i = 1; i < n_elts; ++i) | |
6977 | { | |
6978 | x = CONST_VECTOR_ELT (vals, i); | |
6979 | if (!rtx_equal_p (x, CONST_VECTOR_ELT (vals, 0))) | |
6980 | all_same = false; | |
6981 | } | |
6982 | ||
6983 | if (!all_same) | |
6984 | return NULL_RTX; | |
6985 | ||
6986 | /* We can load this constant by using DUP and a constant in a | |
6987 | single ARM register. This will be cheaper than a vector | |
6988 | load. */ | |
6989 | x = copy_to_mode_reg (inner_mode, CONST_VECTOR_ELT (vals, 0)); | |
6990 | return gen_rtx_VEC_DUPLICATE (mode, x); | |
6991 | } | |
6992 | ||
6993 | ||
6994 | /* Generate code to load VALS, which is a PARALLEL containing only | |
6995 | constants (for vec_init) or CONST_VECTOR, efficiently into a | |
6996 | register. Returns an RTX to copy into the register, or NULL_RTX | |
6997 | for a PARALLEL that can not be converted into a CONST_VECTOR. */ | |
1df3f464 | 6998 | static rtx |
4369c11e TB |
6999 | aarch64_simd_make_constant (rtx vals) |
7000 | { | |
7001 | enum machine_mode mode = GET_MODE (vals); | |
7002 | rtx const_dup; | |
7003 | rtx const_vec = NULL_RTX; | |
7004 | int n_elts = GET_MODE_NUNITS (mode); | |
7005 | int n_const = 0; | |
7006 | int i; | |
7007 | ||
7008 | if (GET_CODE (vals) == CONST_VECTOR) | |
7009 | const_vec = vals; | |
7010 | else if (GET_CODE (vals) == PARALLEL) | |
7011 | { | |
7012 | /* A CONST_VECTOR must contain only CONST_INTs and | |
7013 | CONST_DOUBLEs, but CONSTANT_P allows more (e.g. SYMBOL_REF). | |
7014 | Only store valid constants in a CONST_VECTOR. */ | |
7015 | for (i = 0; i < n_elts; ++i) | |
7016 | { | |
7017 | rtx x = XVECEXP (vals, 0, i); | |
7018 | if (CONST_INT_P (x) || CONST_DOUBLE_P (x)) | |
7019 | n_const++; | |
7020 | } | |
7021 | if (n_const == n_elts) | |
7022 | const_vec = gen_rtx_CONST_VECTOR (mode, XVEC (vals, 0)); | |
7023 | } | |
7024 | else | |
7025 | gcc_unreachable (); | |
7026 | ||
7027 | if (const_vec != NULL_RTX | |
48063b9d | 7028 | && aarch64_simd_valid_immediate (const_vec, mode, false, NULL)) |
4369c11e TB |
7029 | /* Load using MOVI/MVNI. */ |
7030 | return const_vec; | |
7031 | else if ((const_dup = aarch64_simd_dup_constant (vals)) != NULL_RTX) | |
7032 | /* Loaded using DUP. */ | |
7033 | return const_dup; | |
7034 | else if (const_vec != NULL_RTX) | |
7035 | /* Load from constant pool. We can not take advantage of single-cycle | |
7036 | LD1 because we need a PC-relative addressing mode. */ | |
7037 | return const_vec; | |
7038 | else | |
7039 | /* A PARALLEL containing something not valid inside CONST_VECTOR. | |
7040 | We can not construct an initializer. */ | |
7041 | return NULL_RTX; | |
7042 | } | |
7043 | ||
7044 | void | |
7045 | aarch64_expand_vector_init (rtx target, rtx vals) | |
7046 | { | |
7047 | enum machine_mode mode = GET_MODE (target); | |
7048 | enum machine_mode inner_mode = GET_MODE_INNER (mode); | |
7049 | int n_elts = GET_MODE_NUNITS (mode); | |
7050 | int n_var = 0, one_var = -1; | |
7051 | bool all_same = true; | |
7052 | rtx x, mem; | |
7053 | int i; | |
7054 | ||
7055 | x = XVECEXP (vals, 0, 0); | |
7056 | if (!CONST_INT_P (x) && !CONST_DOUBLE_P (x)) | |
7057 | n_var = 1, one_var = 0; | |
7058 | ||
7059 | for (i = 1; i < n_elts; ++i) | |
7060 | { | |
7061 | x = XVECEXP (vals, 0, i); | |
7062 | if (!CONST_INT_P (x) && !CONST_DOUBLE_P (x)) | |
7063 | ++n_var, one_var = i; | |
7064 | ||
7065 | if (!rtx_equal_p (x, XVECEXP (vals, 0, 0))) | |
7066 | all_same = false; | |
7067 | } | |
7068 | ||
7069 | if (n_var == 0) | |
7070 | { | |
7071 | rtx constant = aarch64_simd_make_constant (vals); | |
7072 | if (constant != NULL_RTX) | |
7073 | { | |
7074 | emit_move_insn (target, constant); | |
7075 | return; | |
7076 | } | |
7077 | } | |
7078 | ||
7079 | /* Splat a single non-constant element if we can. */ | |
7080 | if (all_same) | |
7081 | { | |
7082 | x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, 0)); | |
7083 | aarch64_emit_move (target, gen_rtx_VEC_DUPLICATE (mode, x)); | |
7084 | return; | |
7085 | } | |
7086 | ||
7087 | /* One field is non-constant. Load constant then overwrite varying | |
7088 | field. This is more efficient than using the stack. */ | |
7089 | if (n_var == 1) | |
7090 | { | |
7091 | rtx copy = copy_rtx (vals); | |
7092 | rtx index = GEN_INT (one_var); | |
7093 | enum insn_code icode; | |
7094 | ||
7095 | /* Load constant part of vector, substitute neighboring value for | |
7096 | varying element. */ | |
7097 | XVECEXP (copy, 0, one_var) = XVECEXP (vals, 0, one_var ^ 1); | |
7098 | aarch64_expand_vector_init (target, copy); | |
7099 | ||
7100 | /* Insert variable. */ | |
7101 | x = copy_to_mode_reg (inner_mode, XVECEXP (vals, 0, one_var)); | |
7102 | icode = optab_handler (vec_set_optab, mode); | |
7103 | gcc_assert (icode != CODE_FOR_nothing); | |
7104 | emit_insn (GEN_FCN (icode) (target, x, index)); | |
7105 | return; | |
7106 | } | |
7107 | ||
7108 | /* Construct the vector in memory one field at a time | |
7109 | and load the whole vector. */ | |
7110 | mem = assign_stack_temp (mode, GET_MODE_SIZE (mode)); | |
7111 | for (i = 0; i < n_elts; i++) | |
7112 | emit_move_insn (adjust_address_nv (mem, inner_mode, | |
7113 | i * GET_MODE_SIZE (inner_mode)), | |
7114 | XVECEXP (vals, 0, i)); | |
7115 | emit_move_insn (target, mem); | |
7116 | ||
7117 | } | |
7118 | ||
43e9d192 IB |
7119 | static unsigned HOST_WIDE_INT |
7120 | aarch64_shift_truncation_mask (enum machine_mode mode) | |
7121 | { | |
7122 | return | |
7123 | (aarch64_vector_mode_supported_p (mode) | |
7124 | || aarch64_vect_struct_mode_p (mode)) ? 0 : (GET_MODE_BITSIZE (mode) - 1); | |
7125 | } | |
7126 | ||
7127 | #ifndef TLS_SECTION_ASM_FLAG | |
7128 | #define TLS_SECTION_ASM_FLAG 'T' | |
7129 | #endif | |
7130 | ||
7131 | void | |
7132 | aarch64_elf_asm_named_section (const char *name, unsigned int flags, | |
7133 | tree decl ATTRIBUTE_UNUSED) | |
7134 | { | |
7135 | char flagchars[10], *f = flagchars; | |
7136 | ||
7137 | /* If we have already declared this section, we can use an | |
7138 | abbreviated form to switch back to it -- unless this section is | |
7139 | part of a COMDAT groups, in which case GAS requires the full | |
7140 | declaration every time. */ | |
7141 | if (!(HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
7142 | && (flags & SECTION_DECLARED)) | |
7143 | { | |
7144 | fprintf (asm_out_file, "\t.section\t%s\n", name); | |
7145 | return; | |
7146 | } | |
7147 | ||
7148 | if (!(flags & SECTION_DEBUG)) | |
7149 | *f++ = 'a'; | |
7150 | if (flags & SECTION_WRITE) | |
7151 | *f++ = 'w'; | |
7152 | if (flags & SECTION_CODE) | |
7153 | *f++ = 'x'; | |
7154 | if (flags & SECTION_SMALL) | |
7155 | *f++ = 's'; | |
7156 | if (flags & SECTION_MERGE) | |
7157 | *f++ = 'M'; | |
7158 | if (flags & SECTION_STRINGS) | |
7159 | *f++ = 'S'; | |
7160 | if (flags & SECTION_TLS) | |
7161 | *f++ = TLS_SECTION_ASM_FLAG; | |
7162 | if (HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
7163 | *f++ = 'G'; | |
7164 | *f = '\0'; | |
7165 | ||
7166 | fprintf (asm_out_file, "\t.section\t%s,\"%s\"", name, flagchars); | |
7167 | ||
7168 | if (!(flags & SECTION_NOTYPE)) | |
7169 | { | |
7170 | const char *type; | |
7171 | const char *format; | |
7172 | ||
7173 | if (flags & SECTION_BSS) | |
7174 | type = "nobits"; | |
7175 | else | |
7176 | type = "progbits"; | |
7177 | ||
7178 | #ifdef TYPE_OPERAND_FMT | |
7179 | format = "," TYPE_OPERAND_FMT; | |
7180 | #else | |
7181 | format = ",@%s"; | |
7182 | #endif | |
7183 | ||
7184 | fprintf (asm_out_file, format, type); | |
7185 | ||
7186 | if (flags & SECTION_ENTSIZE) | |
7187 | fprintf (asm_out_file, ",%d", flags & SECTION_ENTSIZE); | |
7188 | if (HAVE_COMDAT_GROUP && (flags & SECTION_LINKONCE)) | |
7189 | { | |
7190 | if (TREE_CODE (decl) == IDENTIFIER_NODE) | |
7191 | fprintf (asm_out_file, ",%s,comdat", IDENTIFIER_POINTER (decl)); | |
7192 | else | |
7193 | fprintf (asm_out_file, ",%s,comdat", | |
7194 | IDENTIFIER_POINTER (DECL_COMDAT_GROUP (decl))); | |
7195 | } | |
7196 | } | |
7197 | ||
7198 | putc ('\n', asm_out_file); | |
7199 | } | |
7200 | ||
7201 | /* Select a format to encode pointers in exception handling data. */ | |
7202 | int | |
7203 | aarch64_asm_preferred_eh_data_format (int code ATTRIBUTE_UNUSED, int global) | |
7204 | { | |
7205 | int type; | |
7206 | switch (aarch64_cmodel) | |
7207 | { | |
7208 | case AARCH64_CMODEL_TINY: | |
7209 | case AARCH64_CMODEL_TINY_PIC: | |
7210 | case AARCH64_CMODEL_SMALL: | |
7211 | case AARCH64_CMODEL_SMALL_PIC: | |
7212 | /* text+got+data < 4Gb. 4-byte signed relocs are sufficient | |
7213 | for everything. */ | |
7214 | type = DW_EH_PE_sdata4; | |
7215 | break; | |
7216 | default: | |
7217 | /* No assumptions here. 8-byte relocs required. */ | |
7218 | type = DW_EH_PE_sdata8; | |
7219 | break; | |
7220 | } | |
7221 | return (global ? DW_EH_PE_indirect : 0) | DW_EH_PE_pcrel | type; | |
7222 | } | |
7223 | ||
0462169c SN |
7224 | /* Emit load exclusive. */ |
7225 | ||
7226 | static void | |
7227 | aarch64_emit_load_exclusive (enum machine_mode mode, rtx rval, | |
7228 | rtx mem, rtx model_rtx) | |
7229 | { | |
7230 | rtx (*gen) (rtx, rtx, rtx); | |
7231 | ||
7232 | switch (mode) | |
7233 | { | |
7234 | case QImode: gen = gen_aarch64_load_exclusiveqi; break; | |
7235 | case HImode: gen = gen_aarch64_load_exclusivehi; break; | |
7236 | case SImode: gen = gen_aarch64_load_exclusivesi; break; | |
7237 | case DImode: gen = gen_aarch64_load_exclusivedi; break; | |
7238 | default: | |
7239 | gcc_unreachable (); | |
7240 | } | |
7241 | ||
7242 | emit_insn (gen (rval, mem, model_rtx)); | |
7243 | } | |
7244 | ||
7245 | /* Emit store exclusive. */ | |
7246 | ||
7247 | static void | |
7248 | aarch64_emit_store_exclusive (enum machine_mode mode, rtx bval, | |
7249 | rtx rval, rtx mem, rtx model_rtx) | |
7250 | { | |
7251 | rtx (*gen) (rtx, rtx, rtx, rtx); | |
7252 | ||
7253 | switch (mode) | |
7254 | { | |
7255 | case QImode: gen = gen_aarch64_store_exclusiveqi; break; | |
7256 | case HImode: gen = gen_aarch64_store_exclusivehi; break; | |
7257 | case SImode: gen = gen_aarch64_store_exclusivesi; break; | |
7258 | case DImode: gen = gen_aarch64_store_exclusivedi; break; | |
7259 | default: | |
7260 | gcc_unreachable (); | |
7261 | } | |
7262 | ||
7263 | emit_insn (gen (bval, rval, mem, model_rtx)); | |
7264 | } | |
7265 | ||
7266 | /* Mark the previous jump instruction as unlikely. */ | |
7267 | ||
7268 | static void | |
7269 | aarch64_emit_unlikely_jump (rtx insn) | |
7270 | { | |
e5af9ddd | 7271 | int very_unlikely = REG_BR_PROB_BASE / 100 - 1; |
0462169c SN |
7272 | |
7273 | insn = emit_jump_insn (insn); | |
e5af9ddd | 7274 | add_int_reg_note (insn, REG_BR_PROB, very_unlikely); |
0462169c SN |
7275 | } |
7276 | ||
7277 | /* Expand a compare and swap pattern. */ | |
7278 | ||
7279 | void | |
7280 | aarch64_expand_compare_and_swap (rtx operands[]) | |
7281 | { | |
7282 | rtx bval, rval, mem, oldval, newval, is_weak, mod_s, mod_f, x; | |
7283 | enum machine_mode mode, cmp_mode; | |
7284 | rtx (*gen) (rtx, rtx, rtx, rtx, rtx, rtx, rtx); | |
7285 | ||
7286 | bval = operands[0]; | |
7287 | rval = operands[1]; | |
7288 | mem = operands[2]; | |
7289 | oldval = operands[3]; | |
7290 | newval = operands[4]; | |
7291 | is_weak = operands[5]; | |
7292 | mod_s = operands[6]; | |
7293 | mod_f = operands[7]; | |
7294 | mode = GET_MODE (mem); | |
7295 | cmp_mode = mode; | |
7296 | ||
7297 | /* Normally the succ memory model must be stronger than fail, but in the | |
7298 | unlikely event of fail being ACQUIRE and succ being RELEASE we need to | |
7299 | promote succ to ACQ_REL so that we don't lose the acquire semantics. */ | |
7300 | ||
7301 | if (INTVAL (mod_f) == MEMMODEL_ACQUIRE | |
7302 | && INTVAL (mod_s) == MEMMODEL_RELEASE) | |
7303 | mod_s = GEN_INT (MEMMODEL_ACQ_REL); | |
7304 | ||
7305 | switch (mode) | |
7306 | { | |
7307 | case QImode: | |
7308 | case HImode: | |
7309 | /* For short modes, we're going to perform the comparison in SImode, | |
7310 | so do the zero-extension now. */ | |
7311 | cmp_mode = SImode; | |
7312 | rval = gen_reg_rtx (SImode); | |
7313 | oldval = convert_modes (SImode, mode, oldval, true); | |
7314 | /* Fall through. */ | |
7315 | ||
7316 | case SImode: | |
7317 | case DImode: | |
7318 | /* Force the value into a register if needed. */ | |
7319 | if (!aarch64_plus_operand (oldval, mode)) | |
7320 | oldval = force_reg (cmp_mode, oldval); | |
7321 | break; | |
7322 | ||
7323 | default: | |
7324 | gcc_unreachable (); | |
7325 | } | |
7326 | ||
7327 | switch (mode) | |
7328 | { | |
7329 | case QImode: gen = gen_atomic_compare_and_swapqi_1; break; | |
7330 | case HImode: gen = gen_atomic_compare_and_swaphi_1; break; | |
7331 | case SImode: gen = gen_atomic_compare_and_swapsi_1; break; | |
7332 | case DImode: gen = gen_atomic_compare_and_swapdi_1; break; | |
7333 | default: | |
7334 | gcc_unreachable (); | |
7335 | } | |
7336 | ||
7337 | emit_insn (gen (rval, mem, oldval, newval, is_weak, mod_s, mod_f)); | |
7338 | ||
7339 | if (mode == QImode || mode == HImode) | |
7340 | emit_move_insn (operands[1], gen_lowpart (mode, rval)); | |
7341 | ||
7342 | x = gen_rtx_REG (CCmode, CC_REGNUM); | |
7343 | x = gen_rtx_EQ (SImode, x, const0_rtx); | |
7344 | emit_insn (gen_rtx_SET (VOIDmode, bval, x)); | |
7345 | } | |
7346 | ||
7347 | /* Split a compare and swap pattern. */ | |
7348 | ||
7349 | void | |
7350 | aarch64_split_compare_and_swap (rtx operands[]) | |
7351 | { | |
7352 | rtx rval, mem, oldval, newval, scratch; | |
7353 | enum machine_mode mode; | |
0462169c SN |
7354 | bool is_weak; |
7355 | rtx label1, label2, x, cond; | |
7356 | ||
7357 | rval = operands[0]; | |
7358 | mem = operands[1]; | |
7359 | oldval = operands[2]; | |
7360 | newval = operands[3]; | |
7361 | is_weak = (operands[4] != const0_rtx); | |
0462169c SN |
7362 | scratch = operands[7]; |
7363 | mode = GET_MODE (mem); | |
7364 | ||
7365 | label1 = NULL_RTX; | |
7366 | if (!is_weak) | |
7367 | { | |
7368 | label1 = gen_label_rtx (); | |
7369 | emit_label (label1); | |
7370 | } | |
7371 | label2 = gen_label_rtx (); | |
7372 | ||
7373 | aarch64_emit_load_exclusive (mode, rval, mem, operands[5]); | |
7374 | ||
7375 | cond = aarch64_gen_compare_reg (NE, rval, oldval); | |
7376 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
7377 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
7378 | gen_rtx_LABEL_REF (Pmode, label2), pc_rtx); | |
7379 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
7380 | ||
7381 | aarch64_emit_store_exclusive (mode, scratch, mem, newval, operands[5]); | |
7382 | ||
7383 | if (!is_weak) | |
7384 | { | |
7385 | x = gen_rtx_NE (VOIDmode, scratch, const0_rtx); | |
7386 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
7387 | gen_rtx_LABEL_REF (Pmode, label1), pc_rtx); | |
7388 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
7389 | } | |
7390 | else | |
7391 | { | |
7392 | cond = gen_rtx_REG (CCmode, CC_REGNUM); | |
7393 | x = gen_rtx_COMPARE (CCmode, scratch, const0_rtx); | |
7394 | emit_insn (gen_rtx_SET (VOIDmode, cond, x)); | |
7395 | } | |
7396 | ||
7397 | emit_label (label2); | |
7398 | } | |
7399 | ||
7400 | /* Split an atomic operation. */ | |
7401 | ||
7402 | void | |
7403 | aarch64_split_atomic_op (enum rtx_code code, rtx old_out, rtx new_out, rtx mem, | |
7404 | rtx value, rtx model_rtx, rtx cond) | |
7405 | { | |
7406 | enum machine_mode mode = GET_MODE (mem); | |
7407 | enum machine_mode wmode = (mode == DImode ? DImode : SImode); | |
7408 | rtx label, x; | |
7409 | ||
7410 | label = gen_label_rtx (); | |
7411 | emit_label (label); | |
7412 | ||
7413 | if (new_out) | |
7414 | new_out = gen_lowpart (wmode, new_out); | |
7415 | if (old_out) | |
7416 | old_out = gen_lowpart (wmode, old_out); | |
7417 | else | |
7418 | old_out = new_out; | |
7419 | value = simplify_gen_subreg (wmode, value, mode, 0); | |
7420 | ||
7421 | aarch64_emit_load_exclusive (mode, old_out, mem, model_rtx); | |
7422 | ||
7423 | switch (code) | |
7424 | { | |
7425 | case SET: | |
7426 | new_out = value; | |
7427 | break; | |
7428 | ||
7429 | case NOT: | |
7430 | x = gen_rtx_AND (wmode, old_out, value); | |
7431 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
7432 | x = gen_rtx_NOT (wmode, new_out); | |
7433 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
7434 | break; | |
7435 | ||
7436 | case MINUS: | |
7437 | if (CONST_INT_P (value)) | |
7438 | { | |
7439 | value = GEN_INT (-INTVAL (value)); | |
7440 | code = PLUS; | |
7441 | } | |
7442 | /* Fall through. */ | |
7443 | ||
7444 | default: | |
7445 | x = gen_rtx_fmt_ee (code, wmode, old_out, value); | |
7446 | emit_insn (gen_rtx_SET (VOIDmode, new_out, x)); | |
7447 | break; | |
7448 | } | |
7449 | ||
7450 | aarch64_emit_store_exclusive (mode, cond, mem, | |
7451 | gen_lowpart (mode, new_out), model_rtx); | |
7452 | ||
7453 | x = gen_rtx_NE (VOIDmode, cond, const0_rtx); | |
7454 | x = gen_rtx_IF_THEN_ELSE (VOIDmode, x, | |
7455 | gen_rtx_LABEL_REF (Pmode, label), pc_rtx); | |
7456 | aarch64_emit_unlikely_jump (gen_rtx_SET (VOIDmode, pc_rtx, x)); | |
7457 | } | |
7458 | ||
95ca411e YZ |
7459 | static void |
7460 | aarch64_print_extension (void) | |
7461 | { | |
7462 | const struct aarch64_option_extension *opt = NULL; | |
7463 | ||
7464 | for (opt = all_extensions; opt->name != NULL; opt++) | |
7465 | if ((aarch64_isa_flags & opt->flags_on) == opt->flags_on) | |
7466 | asm_fprintf (asm_out_file, "+%s", opt->name); | |
7467 | ||
7468 | asm_fprintf (asm_out_file, "\n"); | |
7469 | } | |
7470 | ||
43e9d192 IB |
7471 | static void |
7472 | aarch64_start_file (void) | |
7473 | { | |
7474 | if (selected_arch) | |
95ca411e YZ |
7475 | { |
7476 | asm_fprintf (asm_out_file, "\t.arch %s", selected_arch->name); | |
7477 | aarch64_print_extension (); | |
7478 | } | |
43e9d192 | 7479 | else if (selected_cpu) |
95ca411e | 7480 | { |
682287fb JG |
7481 | const char *truncated_name |
7482 | = aarch64_rewrite_selected_cpu (selected_cpu->name); | |
7483 | asm_fprintf (asm_out_file, "\t.cpu %s", truncated_name); | |
95ca411e YZ |
7484 | aarch64_print_extension (); |
7485 | } | |
43e9d192 IB |
7486 | default_file_start(); |
7487 | } | |
7488 | ||
7489 | /* Target hook for c_mode_for_suffix. */ | |
7490 | static enum machine_mode | |
7491 | aarch64_c_mode_for_suffix (char suffix) | |
7492 | { | |
7493 | if (suffix == 'q') | |
7494 | return TFmode; | |
7495 | ||
7496 | return VOIDmode; | |
7497 | } | |
7498 | ||
3520f7cc JG |
7499 | /* We can only represent floating point constants which will fit in |
7500 | "quarter-precision" values. These values are characterised by | |
7501 | a sign bit, a 4-bit mantissa and a 3-bit exponent. And are given | |
7502 | by: | |
7503 | ||
7504 | (-1)^s * (n/16) * 2^r | |
7505 | ||
7506 | Where: | |
7507 | 's' is the sign bit. | |
7508 | 'n' is an integer in the range 16 <= n <= 31. | |
7509 | 'r' is an integer in the range -3 <= r <= 4. */ | |
7510 | ||
7511 | /* Return true iff X can be represented by a quarter-precision | |
7512 | floating point immediate operand X. Note, we cannot represent 0.0. */ | |
7513 | bool | |
7514 | aarch64_float_const_representable_p (rtx x) | |
7515 | { | |
7516 | /* This represents our current view of how many bits | |
7517 | make up the mantissa. */ | |
7518 | int point_pos = 2 * HOST_BITS_PER_WIDE_INT - 1; | |
ba96cdfb | 7519 | int exponent; |
3520f7cc JG |
7520 | unsigned HOST_WIDE_INT mantissa, mask; |
7521 | HOST_WIDE_INT m1, m2; | |
7522 | REAL_VALUE_TYPE r, m; | |
7523 | ||
7524 | if (!CONST_DOUBLE_P (x)) | |
7525 | return false; | |
7526 | ||
7527 | REAL_VALUE_FROM_CONST_DOUBLE (r, x); | |
7528 | ||
7529 | /* We cannot represent infinities, NaNs or +/-zero. We won't | |
7530 | know if we have +zero until we analyse the mantissa, but we | |
7531 | can reject the other invalid values. */ | |
7532 | if (REAL_VALUE_ISINF (r) || REAL_VALUE_ISNAN (r) | |
7533 | || REAL_VALUE_MINUS_ZERO (r)) | |
7534 | return false; | |
7535 | ||
ba96cdfb | 7536 | /* Extract exponent. */ |
3520f7cc JG |
7537 | r = real_value_abs (&r); |
7538 | exponent = REAL_EXP (&r); | |
7539 | ||
7540 | /* For the mantissa, we expand into two HOST_WIDE_INTS, apart from the | |
7541 | highest (sign) bit, with a fixed binary point at bit point_pos. | |
7542 | m1 holds the low part of the mantissa, m2 the high part. | |
7543 | WARNING: If we ever have a representation using more than 2 * H_W_I - 1 | |
7544 | bits for the mantissa, this can fail (low bits will be lost). */ | |
7545 | real_ldexp (&m, &r, point_pos - exponent); | |
7546 | REAL_VALUE_TO_INT (&m1, &m2, m); | |
7547 | ||
7548 | /* If the low part of the mantissa has bits set we cannot represent | |
7549 | the value. */ | |
7550 | if (m1 != 0) | |
7551 | return false; | |
7552 | /* We have rejected the lower HOST_WIDE_INT, so update our | |
7553 | understanding of how many bits lie in the mantissa and | |
7554 | look only at the high HOST_WIDE_INT. */ | |
7555 | mantissa = m2; | |
7556 | point_pos -= HOST_BITS_PER_WIDE_INT; | |
7557 | ||
7558 | /* We can only represent values with a mantissa of the form 1.xxxx. */ | |
7559 | mask = ((unsigned HOST_WIDE_INT)1 << (point_pos - 5)) - 1; | |
7560 | if ((mantissa & mask) != 0) | |
7561 | return false; | |
7562 | ||
7563 | /* Having filtered unrepresentable values, we may now remove all | |
7564 | but the highest 5 bits. */ | |
7565 | mantissa >>= point_pos - 5; | |
7566 | ||
7567 | /* We cannot represent the value 0.0, so reject it. This is handled | |
7568 | elsewhere. */ | |
7569 | if (mantissa == 0) | |
7570 | return false; | |
7571 | ||
7572 | /* Then, as bit 4 is always set, we can mask it off, leaving | |
7573 | the mantissa in the range [0, 15]. */ | |
7574 | mantissa &= ~(1 << 4); | |
7575 | gcc_assert (mantissa <= 15); | |
7576 | ||
7577 | /* GCC internally does not use IEEE754-like encoding (where normalized | |
7578 | significands are in the range [1, 2). GCC uses [0.5, 1) (see real.c). | |
7579 | Our mantissa values are shifted 4 places to the left relative to | |
7580 | normalized IEEE754 so we must modify the exponent returned by REAL_EXP | |
7581 | by 5 places to correct for GCC's representation. */ | |
7582 | exponent = 5 - exponent; | |
7583 | ||
7584 | return (exponent >= 0 && exponent <= 7); | |
7585 | } | |
7586 | ||
7587 | char* | |
81c2dfb9 | 7588 | aarch64_output_simd_mov_immediate (rtx const_vector, |
3520f7cc JG |
7589 | enum machine_mode mode, |
7590 | unsigned width) | |
7591 | { | |
3ea63f60 | 7592 | bool is_valid; |
3520f7cc | 7593 | static char templ[40]; |
3520f7cc | 7594 | const char *mnemonic; |
e4f0f84d | 7595 | const char *shift_op; |
3520f7cc | 7596 | unsigned int lane_count = 0; |
81c2dfb9 | 7597 | char element_char; |
3520f7cc | 7598 | |
e4f0f84d | 7599 | struct simd_immediate_info info = { NULL_RTX, 0, 0, false, false }; |
48063b9d IB |
7600 | |
7601 | /* This will return true to show const_vector is legal for use as either | |
7602 | a AdvSIMD MOVI instruction (or, implicitly, MVNI) immediate. It will | |
7603 | also update INFO to show how the immediate should be generated. */ | |
81c2dfb9 | 7604 | is_valid = aarch64_simd_valid_immediate (const_vector, mode, false, &info); |
3520f7cc JG |
7605 | gcc_assert (is_valid); |
7606 | ||
81c2dfb9 | 7607 | element_char = sizetochar (info.element_width); |
48063b9d IB |
7608 | lane_count = width / info.element_width; |
7609 | ||
3520f7cc JG |
7610 | mode = GET_MODE_INNER (mode); |
7611 | if (mode == SFmode || mode == DFmode) | |
7612 | { | |
48063b9d IB |
7613 | gcc_assert (info.shift == 0 && ! info.mvn); |
7614 | if (aarch64_float_const_zero_rtx_p (info.value)) | |
7615 | info.value = GEN_INT (0); | |
7616 | else | |
7617 | { | |
7618 | #define buf_size 20 | |
7619 | REAL_VALUE_TYPE r; | |
7620 | REAL_VALUE_FROM_CONST_DOUBLE (r, info.value); | |
7621 | char float_buf[buf_size] = {'\0'}; | |
7622 | real_to_decimal_for_mode (float_buf, &r, buf_size, buf_size, 1, mode); | |
7623 | #undef buf_size | |
7624 | ||
7625 | if (lane_count == 1) | |
7626 | snprintf (templ, sizeof (templ), "fmov\t%%d0, %s", float_buf); | |
7627 | else | |
7628 | snprintf (templ, sizeof (templ), "fmov\t%%0.%d%c, %s", | |
81c2dfb9 | 7629 | lane_count, element_char, float_buf); |
48063b9d IB |
7630 | return templ; |
7631 | } | |
3520f7cc | 7632 | } |
3520f7cc | 7633 | |
48063b9d | 7634 | mnemonic = info.mvn ? "mvni" : "movi"; |
e4f0f84d | 7635 | shift_op = info.msl ? "msl" : "lsl"; |
3520f7cc JG |
7636 | |
7637 | if (lane_count == 1) | |
48063b9d IB |
7638 | snprintf (templ, sizeof (templ), "%s\t%%d0, " HOST_WIDE_INT_PRINT_HEX, |
7639 | mnemonic, UINTVAL (info.value)); | |
7640 | else if (info.shift) | |
7641 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX | |
e4f0f84d TB |
7642 | ", %s %d", mnemonic, lane_count, element_char, |
7643 | UINTVAL (info.value), shift_op, info.shift); | |
3520f7cc | 7644 | else |
48063b9d | 7645 | snprintf (templ, sizeof (templ), "%s\t%%0.%d%c, " HOST_WIDE_INT_PRINT_HEX, |
81c2dfb9 | 7646 | mnemonic, lane_count, element_char, UINTVAL (info.value)); |
3520f7cc JG |
7647 | return templ; |
7648 | } | |
7649 | ||
b7342d25 IB |
7650 | char* |
7651 | aarch64_output_scalar_simd_mov_immediate (rtx immediate, | |
7652 | enum machine_mode mode) | |
7653 | { | |
7654 | enum machine_mode vmode; | |
7655 | ||
7656 | gcc_assert (!VECTOR_MODE_P (mode)); | |
7657 | vmode = aarch64_simd_container_mode (mode, 64); | |
7658 | rtx v_op = aarch64_simd_gen_const_vector_dup (vmode, INTVAL (immediate)); | |
7659 | return aarch64_output_simd_mov_immediate (v_op, vmode, 64); | |
7660 | } | |
7661 | ||
88b08073 JG |
7662 | /* Split operands into moves from op[1] + op[2] into op[0]. */ |
7663 | ||
7664 | void | |
7665 | aarch64_split_combinev16qi (rtx operands[3]) | |
7666 | { | |
7667 | unsigned int dest = REGNO (operands[0]); | |
7668 | unsigned int src1 = REGNO (operands[1]); | |
7669 | unsigned int src2 = REGNO (operands[2]); | |
7670 | enum machine_mode halfmode = GET_MODE (operands[1]); | |
7671 | unsigned int halfregs = HARD_REGNO_NREGS (src1, halfmode); | |
7672 | rtx destlo, desthi; | |
7673 | ||
7674 | gcc_assert (halfmode == V16QImode); | |
7675 | ||
7676 | if (src1 == dest && src2 == dest + halfregs) | |
7677 | { | |
7678 | /* No-op move. Can't split to nothing; emit something. */ | |
7679 | emit_note (NOTE_INSN_DELETED); | |
7680 | return; | |
7681 | } | |
7682 | ||
7683 | /* Preserve register attributes for variable tracking. */ | |
7684 | destlo = gen_rtx_REG_offset (operands[0], halfmode, dest, 0); | |
7685 | desthi = gen_rtx_REG_offset (operands[0], halfmode, dest + halfregs, | |
7686 | GET_MODE_SIZE (halfmode)); | |
7687 | ||
7688 | /* Special case of reversed high/low parts. */ | |
7689 | if (reg_overlap_mentioned_p (operands[2], destlo) | |
7690 | && reg_overlap_mentioned_p (operands[1], desthi)) | |
7691 | { | |
7692 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
7693 | emit_insn (gen_xorv16qi3 (operands[2], operands[1], operands[2])); | |
7694 | emit_insn (gen_xorv16qi3 (operands[1], operands[1], operands[2])); | |
7695 | } | |
7696 | else if (!reg_overlap_mentioned_p (operands[2], destlo)) | |
7697 | { | |
7698 | /* Try to avoid unnecessary moves if part of the result | |
7699 | is in the right place already. */ | |
7700 | if (src1 != dest) | |
7701 | emit_move_insn (destlo, operands[1]); | |
7702 | if (src2 != dest + halfregs) | |
7703 | emit_move_insn (desthi, operands[2]); | |
7704 | } | |
7705 | else | |
7706 | { | |
7707 | if (src2 != dest + halfregs) | |
7708 | emit_move_insn (desthi, operands[2]); | |
7709 | if (src1 != dest) | |
7710 | emit_move_insn (destlo, operands[1]); | |
7711 | } | |
7712 | } | |
7713 | ||
7714 | /* vec_perm support. */ | |
7715 | ||
7716 | #define MAX_VECT_LEN 16 | |
7717 | ||
7718 | struct expand_vec_perm_d | |
7719 | { | |
7720 | rtx target, op0, op1; | |
7721 | unsigned char perm[MAX_VECT_LEN]; | |
7722 | enum machine_mode vmode; | |
7723 | unsigned char nelt; | |
7724 | bool one_vector_p; | |
7725 | bool testing_p; | |
7726 | }; | |
7727 | ||
7728 | /* Generate a variable permutation. */ | |
7729 | ||
7730 | static void | |
7731 | aarch64_expand_vec_perm_1 (rtx target, rtx op0, rtx op1, rtx sel) | |
7732 | { | |
7733 | enum machine_mode vmode = GET_MODE (target); | |
7734 | bool one_vector_p = rtx_equal_p (op0, op1); | |
7735 | ||
7736 | gcc_checking_assert (vmode == V8QImode || vmode == V16QImode); | |
7737 | gcc_checking_assert (GET_MODE (op0) == vmode); | |
7738 | gcc_checking_assert (GET_MODE (op1) == vmode); | |
7739 | gcc_checking_assert (GET_MODE (sel) == vmode); | |
7740 | gcc_checking_assert (TARGET_SIMD); | |
7741 | ||
7742 | if (one_vector_p) | |
7743 | { | |
7744 | if (vmode == V8QImode) | |
7745 | { | |
7746 | /* Expand the argument to a V16QI mode by duplicating it. */ | |
7747 | rtx pair = gen_reg_rtx (V16QImode); | |
7748 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op0)); | |
7749 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
7750 | } | |
7751 | else | |
7752 | { | |
7753 | emit_insn (gen_aarch64_tbl1v16qi (target, op0, sel)); | |
7754 | } | |
7755 | } | |
7756 | else | |
7757 | { | |
7758 | rtx pair; | |
7759 | ||
7760 | if (vmode == V8QImode) | |
7761 | { | |
7762 | pair = gen_reg_rtx (V16QImode); | |
7763 | emit_insn (gen_aarch64_combinev8qi (pair, op0, op1)); | |
7764 | emit_insn (gen_aarch64_tbl1v8qi (target, pair, sel)); | |
7765 | } | |
7766 | else | |
7767 | { | |
7768 | pair = gen_reg_rtx (OImode); | |
7769 | emit_insn (gen_aarch64_combinev16qi (pair, op0, op1)); | |
7770 | emit_insn (gen_aarch64_tbl2v16qi (target, pair, sel)); | |
7771 | } | |
7772 | } | |
7773 | } | |
7774 | ||
7775 | void | |
7776 | aarch64_expand_vec_perm (rtx target, rtx op0, rtx op1, rtx sel) | |
7777 | { | |
7778 | enum machine_mode vmode = GET_MODE (target); | |
7779 | unsigned int i, nelt = GET_MODE_NUNITS (vmode); | |
7780 | bool one_vector_p = rtx_equal_p (op0, op1); | |
7781 | rtx rmask[MAX_VECT_LEN], mask; | |
7782 | ||
7783 | gcc_checking_assert (!BYTES_BIG_ENDIAN); | |
7784 | ||
7785 | /* The TBL instruction does not use a modulo index, so we must take care | |
7786 | of that ourselves. */ | |
7787 | mask = GEN_INT (one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
7788 | for (i = 0; i < nelt; ++i) | |
7789 | rmask[i] = mask; | |
7790 | mask = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rmask)); | |
7791 | sel = expand_simple_binop (vmode, AND, sel, mask, NULL, 0, OPTAB_LIB_WIDEN); | |
7792 | ||
7793 | aarch64_expand_vec_perm_1 (target, op0, op1, sel); | |
7794 | } | |
7795 | ||
cc4d934f JG |
7796 | /* Recognize patterns suitable for the TRN instructions. */ |
7797 | static bool | |
7798 | aarch64_evpc_trn (struct expand_vec_perm_d *d) | |
7799 | { | |
7800 | unsigned int i, odd, mask, nelt = d->nelt; | |
7801 | rtx out, in0, in1, x; | |
7802 | rtx (*gen) (rtx, rtx, rtx); | |
7803 | enum machine_mode vmode = d->vmode; | |
7804 | ||
7805 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
7806 | return false; | |
7807 | ||
7808 | /* Note that these are little-endian tests. | |
7809 | We correct for big-endian later. */ | |
7810 | if (d->perm[0] == 0) | |
7811 | odd = 0; | |
7812 | else if (d->perm[0] == 1) | |
7813 | odd = 1; | |
7814 | else | |
7815 | return false; | |
7816 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
7817 | ||
7818 | for (i = 0; i < nelt; i += 2) | |
7819 | { | |
7820 | if (d->perm[i] != i + odd) | |
7821 | return false; | |
7822 | if (d->perm[i + 1] != ((i + nelt + odd) & mask)) | |
7823 | return false; | |
7824 | } | |
7825 | ||
7826 | /* Success! */ | |
7827 | if (d->testing_p) | |
7828 | return true; | |
7829 | ||
7830 | in0 = d->op0; | |
7831 | in1 = d->op1; | |
7832 | if (BYTES_BIG_ENDIAN) | |
7833 | { | |
7834 | x = in0, in0 = in1, in1 = x; | |
7835 | odd = !odd; | |
7836 | } | |
7837 | out = d->target; | |
7838 | ||
7839 | if (odd) | |
7840 | { | |
7841 | switch (vmode) | |
7842 | { | |
7843 | case V16QImode: gen = gen_aarch64_trn2v16qi; break; | |
7844 | case V8QImode: gen = gen_aarch64_trn2v8qi; break; | |
7845 | case V8HImode: gen = gen_aarch64_trn2v8hi; break; | |
7846 | case V4HImode: gen = gen_aarch64_trn2v4hi; break; | |
7847 | case V4SImode: gen = gen_aarch64_trn2v4si; break; | |
7848 | case V2SImode: gen = gen_aarch64_trn2v2si; break; | |
7849 | case V2DImode: gen = gen_aarch64_trn2v2di; break; | |
7850 | case V4SFmode: gen = gen_aarch64_trn2v4sf; break; | |
7851 | case V2SFmode: gen = gen_aarch64_trn2v2sf; break; | |
7852 | case V2DFmode: gen = gen_aarch64_trn2v2df; break; | |
7853 | default: | |
7854 | return false; | |
7855 | } | |
7856 | } | |
7857 | else | |
7858 | { | |
7859 | switch (vmode) | |
7860 | { | |
7861 | case V16QImode: gen = gen_aarch64_trn1v16qi; break; | |
7862 | case V8QImode: gen = gen_aarch64_trn1v8qi; break; | |
7863 | case V8HImode: gen = gen_aarch64_trn1v8hi; break; | |
7864 | case V4HImode: gen = gen_aarch64_trn1v4hi; break; | |
7865 | case V4SImode: gen = gen_aarch64_trn1v4si; break; | |
7866 | case V2SImode: gen = gen_aarch64_trn1v2si; break; | |
7867 | case V2DImode: gen = gen_aarch64_trn1v2di; break; | |
7868 | case V4SFmode: gen = gen_aarch64_trn1v4sf; break; | |
7869 | case V2SFmode: gen = gen_aarch64_trn1v2sf; break; | |
7870 | case V2DFmode: gen = gen_aarch64_trn1v2df; break; | |
7871 | default: | |
7872 | return false; | |
7873 | } | |
7874 | } | |
7875 | ||
7876 | emit_insn (gen (out, in0, in1)); | |
7877 | return true; | |
7878 | } | |
7879 | ||
7880 | /* Recognize patterns suitable for the UZP instructions. */ | |
7881 | static bool | |
7882 | aarch64_evpc_uzp (struct expand_vec_perm_d *d) | |
7883 | { | |
7884 | unsigned int i, odd, mask, nelt = d->nelt; | |
7885 | rtx out, in0, in1, x; | |
7886 | rtx (*gen) (rtx, rtx, rtx); | |
7887 | enum machine_mode vmode = d->vmode; | |
7888 | ||
7889 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
7890 | return false; | |
7891 | ||
7892 | /* Note that these are little-endian tests. | |
7893 | We correct for big-endian later. */ | |
7894 | if (d->perm[0] == 0) | |
7895 | odd = 0; | |
7896 | else if (d->perm[0] == 1) | |
7897 | odd = 1; | |
7898 | else | |
7899 | return false; | |
7900 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
7901 | ||
7902 | for (i = 0; i < nelt; i++) | |
7903 | { | |
7904 | unsigned elt = (i * 2 + odd) & mask; | |
7905 | if (d->perm[i] != elt) | |
7906 | return false; | |
7907 | } | |
7908 | ||
7909 | /* Success! */ | |
7910 | if (d->testing_p) | |
7911 | return true; | |
7912 | ||
7913 | in0 = d->op0; | |
7914 | in1 = d->op1; | |
7915 | if (BYTES_BIG_ENDIAN) | |
7916 | { | |
7917 | x = in0, in0 = in1, in1 = x; | |
7918 | odd = !odd; | |
7919 | } | |
7920 | out = d->target; | |
7921 | ||
7922 | if (odd) | |
7923 | { | |
7924 | switch (vmode) | |
7925 | { | |
7926 | case V16QImode: gen = gen_aarch64_uzp2v16qi; break; | |
7927 | case V8QImode: gen = gen_aarch64_uzp2v8qi; break; | |
7928 | case V8HImode: gen = gen_aarch64_uzp2v8hi; break; | |
7929 | case V4HImode: gen = gen_aarch64_uzp2v4hi; break; | |
7930 | case V4SImode: gen = gen_aarch64_uzp2v4si; break; | |
7931 | case V2SImode: gen = gen_aarch64_uzp2v2si; break; | |
7932 | case V2DImode: gen = gen_aarch64_uzp2v2di; break; | |
7933 | case V4SFmode: gen = gen_aarch64_uzp2v4sf; break; | |
7934 | case V2SFmode: gen = gen_aarch64_uzp2v2sf; break; | |
7935 | case V2DFmode: gen = gen_aarch64_uzp2v2df; break; | |
7936 | default: | |
7937 | return false; | |
7938 | } | |
7939 | } | |
7940 | else | |
7941 | { | |
7942 | switch (vmode) | |
7943 | { | |
7944 | case V16QImode: gen = gen_aarch64_uzp1v16qi; break; | |
7945 | case V8QImode: gen = gen_aarch64_uzp1v8qi; break; | |
7946 | case V8HImode: gen = gen_aarch64_uzp1v8hi; break; | |
7947 | case V4HImode: gen = gen_aarch64_uzp1v4hi; break; | |
7948 | case V4SImode: gen = gen_aarch64_uzp1v4si; break; | |
7949 | case V2SImode: gen = gen_aarch64_uzp1v2si; break; | |
7950 | case V2DImode: gen = gen_aarch64_uzp1v2di; break; | |
7951 | case V4SFmode: gen = gen_aarch64_uzp1v4sf; break; | |
7952 | case V2SFmode: gen = gen_aarch64_uzp1v2sf; break; | |
7953 | case V2DFmode: gen = gen_aarch64_uzp1v2df; break; | |
7954 | default: | |
7955 | return false; | |
7956 | } | |
7957 | } | |
7958 | ||
7959 | emit_insn (gen (out, in0, in1)); | |
7960 | return true; | |
7961 | } | |
7962 | ||
7963 | /* Recognize patterns suitable for the ZIP instructions. */ | |
7964 | static bool | |
7965 | aarch64_evpc_zip (struct expand_vec_perm_d *d) | |
7966 | { | |
7967 | unsigned int i, high, mask, nelt = d->nelt; | |
7968 | rtx out, in0, in1, x; | |
7969 | rtx (*gen) (rtx, rtx, rtx); | |
7970 | enum machine_mode vmode = d->vmode; | |
7971 | ||
7972 | if (GET_MODE_UNIT_SIZE (vmode) > 8) | |
7973 | return false; | |
7974 | ||
7975 | /* Note that these are little-endian tests. | |
7976 | We correct for big-endian later. */ | |
7977 | high = nelt / 2; | |
7978 | if (d->perm[0] == high) | |
7979 | /* Do Nothing. */ | |
7980 | ; | |
7981 | else if (d->perm[0] == 0) | |
7982 | high = 0; | |
7983 | else | |
7984 | return false; | |
7985 | mask = (d->one_vector_p ? nelt - 1 : 2 * nelt - 1); | |
7986 | ||
7987 | for (i = 0; i < nelt / 2; i++) | |
7988 | { | |
7989 | unsigned elt = (i + high) & mask; | |
7990 | if (d->perm[i * 2] != elt) | |
7991 | return false; | |
7992 | elt = (elt + nelt) & mask; | |
7993 | if (d->perm[i * 2 + 1] != elt) | |
7994 | return false; | |
7995 | } | |
7996 | ||
7997 | /* Success! */ | |
7998 | if (d->testing_p) | |
7999 | return true; | |
8000 | ||
8001 | in0 = d->op0; | |
8002 | in1 = d->op1; | |
8003 | if (BYTES_BIG_ENDIAN) | |
8004 | { | |
8005 | x = in0, in0 = in1, in1 = x; | |
8006 | high = !high; | |
8007 | } | |
8008 | out = d->target; | |
8009 | ||
8010 | if (high) | |
8011 | { | |
8012 | switch (vmode) | |
8013 | { | |
8014 | case V16QImode: gen = gen_aarch64_zip2v16qi; break; | |
8015 | case V8QImode: gen = gen_aarch64_zip2v8qi; break; | |
8016 | case V8HImode: gen = gen_aarch64_zip2v8hi; break; | |
8017 | case V4HImode: gen = gen_aarch64_zip2v4hi; break; | |
8018 | case V4SImode: gen = gen_aarch64_zip2v4si; break; | |
8019 | case V2SImode: gen = gen_aarch64_zip2v2si; break; | |
8020 | case V2DImode: gen = gen_aarch64_zip2v2di; break; | |
8021 | case V4SFmode: gen = gen_aarch64_zip2v4sf; break; | |
8022 | case V2SFmode: gen = gen_aarch64_zip2v2sf; break; | |
8023 | case V2DFmode: gen = gen_aarch64_zip2v2df; break; | |
8024 | default: | |
8025 | return false; | |
8026 | } | |
8027 | } | |
8028 | else | |
8029 | { | |
8030 | switch (vmode) | |
8031 | { | |
8032 | case V16QImode: gen = gen_aarch64_zip1v16qi; break; | |
8033 | case V8QImode: gen = gen_aarch64_zip1v8qi; break; | |
8034 | case V8HImode: gen = gen_aarch64_zip1v8hi; break; | |
8035 | case V4HImode: gen = gen_aarch64_zip1v4hi; break; | |
8036 | case V4SImode: gen = gen_aarch64_zip1v4si; break; | |
8037 | case V2SImode: gen = gen_aarch64_zip1v2si; break; | |
8038 | case V2DImode: gen = gen_aarch64_zip1v2di; break; | |
8039 | case V4SFmode: gen = gen_aarch64_zip1v4sf; break; | |
8040 | case V2SFmode: gen = gen_aarch64_zip1v2sf; break; | |
8041 | case V2DFmode: gen = gen_aarch64_zip1v2df; break; | |
8042 | default: | |
8043 | return false; | |
8044 | } | |
8045 | } | |
8046 | ||
8047 | emit_insn (gen (out, in0, in1)); | |
8048 | return true; | |
8049 | } | |
8050 | ||
91bd4114 JG |
8051 | static bool |
8052 | aarch64_evpc_dup (struct expand_vec_perm_d *d) | |
8053 | { | |
8054 | rtx (*gen) (rtx, rtx, rtx); | |
8055 | rtx out = d->target; | |
8056 | rtx in0; | |
8057 | enum machine_mode vmode = d->vmode; | |
8058 | unsigned int i, elt, nelt = d->nelt; | |
8059 | rtx lane; | |
8060 | ||
8061 | /* TODO: This may not be big-endian safe. */ | |
8062 | if (BYTES_BIG_ENDIAN) | |
8063 | return false; | |
8064 | ||
8065 | elt = d->perm[0]; | |
8066 | for (i = 1; i < nelt; i++) | |
8067 | { | |
8068 | if (elt != d->perm[i]) | |
8069 | return false; | |
8070 | } | |
8071 | ||
8072 | /* The generic preparation in aarch64_expand_vec_perm_const_1 | |
8073 | swaps the operand order and the permute indices if it finds | |
8074 | d->perm[0] to be in the second operand. Thus, we can always | |
8075 | use d->op0 and need not do any extra arithmetic to get the | |
8076 | correct lane number. */ | |
8077 | in0 = d->op0; | |
8078 | lane = GEN_INT (elt); | |
8079 | ||
8080 | switch (vmode) | |
8081 | { | |
8082 | case V16QImode: gen = gen_aarch64_dup_lanev16qi; break; | |
8083 | case V8QImode: gen = gen_aarch64_dup_lanev8qi; break; | |
8084 | case V8HImode: gen = gen_aarch64_dup_lanev8hi; break; | |
8085 | case V4HImode: gen = gen_aarch64_dup_lanev4hi; break; | |
8086 | case V4SImode: gen = gen_aarch64_dup_lanev4si; break; | |
8087 | case V2SImode: gen = gen_aarch64_dup_lanev2si; break; | |
8088 | case V2DImode: gen = gen_aarch64_dup_lanev2di; break; | |
8089 | case V4SFmode: gen = gen_aarch64_dup_lanev4sf; break; | |
8090 | case V2SFmode: gen = gen_aarch64_dup_lanev2sf; break; | |
8091 | case V2DFmode: gen = gen_aarch64_dup_lanev2df; break; | |
8092 | default: | |
8093 | return false; | |
8094 | } | |
8095 | ||
8096 | emit_insn (gen (out, in0, lane)); | |
8097 | return true; | |
8098 | } | |
8099 | ||
88b08073 JG |
8100 | static bool |
8101 | aarch64_evpc_tbl (struct expand_vec_perm_d *d) | |
8102 | { | |
8103 | rtx rperm[MAX_VECT_LEN], sel; | |
8104 | enum machine_mode vmode = d->vmode; | |
8105 | unsigned int i, nelt = d->nelt; | |
8106 | ||
8107 | /* TODO: ARM's TBL indexing is little-endian. In order to handle GCC's | |
8108 | numbering of elements for big-endian, we must reverse the order. */ | |
8109 | if (BYTES_BIG_ENDIAN) | |
8110 | return false; | |
8111 | ||
8112 | if (d->testing_p) | |
8113 | return true; | |
8114 | ||
8115 | /* Generic code will try constant permutation twice. Once with the | |
8116 | original mode and again with the elements lowered to QImode. | |
8117 | So wait and don't do the selector expansion ourselves. */ | |
8118 | if (vmode != V8QImode && vmode != V16QImode) | |
8119 | return false; | |
8120 | ||
8121 | for (i = 0; i < nelt; ++i) | |
8122 | rperm[i] = GEN_INT (d->perm[i]); | |
8123 | sel = gen_rtx_CONST_VECTOR (vmode, gen_rtvec_v (nelt, rperm)); | |
8124 | sel = force_reg (vmode, sel); | |
8125 | ||
8126 | aarch64_expand_vec_perm_1 (d->target, d->op0, d->op1, sel); | |
8127 | return true; | |
8128 | } | |
8129 | ||
8130 | static bool | |
8131 | aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d) | |
8132 | { | |
8133 | /* The pattern matching functions above are written to look for a small | |
8134 | number to begin the sequence (0, 1, N/2). If we begin with an index | |
8135 | from the second operand, we can swap the operands. */ | |
8136 | if (d->perm[0] >= d->nelt) | |
8137 | { | |
8138 | unsigned i, nelt = d->nelt; | |
8139 | rtx x; | |
8140 | ||
8141 | for (i = 0; i < nelt; ++i) | |
8142 | d->perm[i] = (d->perm[i] + nelt) & (2 * nelt - 1); | |
8143 | ||
8144 | x = d->op0; | |
8145 | d->op0 = d->op1; | |
8146 | d->op1 = x; | |
8147 | } | |
8148 | ||
8149 | if (TARGET_SIMD) | |
cc4d934f JG |
8150 | { |
8151 | if (aarch64_evpc_zip (d)) | |
8152 | return true; | |
8153 | else if (aarch64_evpc_uzp (d)) | |
8154 | return true; | |
8155 | else if (aarch64_evpc_trn (d)) | |
8156 | return true; | |
91bd4114 JG |
8157 | else if (aarch64_evpc_dup (d)) |
8158 | return true; | |
cc4d934f JG |
8159 | return aarch64_evpc_tbl (d); |
8160 | } | |
88b08073 JG |
8161 | return false; |
8162 | } | |
8163 | ||
8164 | /* Expand a vec_perm_const pattern. */ | |
8165 | ||
8166 | bool | |
8167 | aarch64_expand_vec_perm_const (rtx target, rtx op0, rtx op1, rtx sel) | |
8168 | { | |
8169 | struct expand_vec_perm_d d; | |
8170 | int i, nelt, which; | |
8171 | ||
8172 | d.target = target; | |
8173 | d.op0 = op0; | |
8174 | d.op1 = op1; | |
8175 | ||
8176 | d.vmode = GET_MODE (target); | |
8177 | gcc_assert (VECTOR_MODE_P (d.vmode)); | |
8178 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
8179 | d.testing_p = false; | |
8180 | ||
8181 | for (i = which = 0; i < nelt; ++i) | |
8182 | { | |
8183 | rtx e = XVECEXP (sel, 0, i); | |
8184 | int ei = INTVAL (e) & (2 * nelt - 1); | |
8185 | which |= (ei < nelt ? 1 : 2); | |
8186 | d.perm[i] = ei; | |
8187 | } | |
8188 | ||
8189 | switch (which) | |
8190 | { | |
8191 | default: | |
8192 | gcc_unreachable (); | |
8193 | ||
8194 | case 3: | |
8195 | d.one_vector_p = false; | |
8196 | if (!rtx_equal_p (op0, op1)) | |
8197 | break; | |
8198 | ||
8199 | /* The elements of PERM do not suggest that only the first operand | |
8200 | is used, but both operands are identical. Allow easier matching | |
8201 | of the permutation by folding the permutation into the single | |
8202 | input vector. */ | |
8203 | /* Fall Through. */ | |
8204 | case 2: | |
8205 | for (i = 0; i < nelt; ++i) | |
8206 | d.perm[i] &= nelt - 1; | |
8207 | d.op0 = op1; | |
8208 | d.one_vector_p = true; | |
8209 | break; | |
8210 | ||
8211 | case 1: | |
8212 | d.op1 = op0; | |
8213 | d.one_vector_p = true; | |
8214 | break; | |
8215 | } | |
8216 | ||
8217 | return aarch64_expand_vec_perm_const_1 (&d); | |
8218 | } | |
8219 | ||
8220 | static bool | |
8221 | aarch64_vectorize_vec_perm_const_ok (enum machine_mode vmode, | |
8222 | const unsigned char *sel) | |
8223 | { | |
8224 | struct expand_vec_perm_d d; | |
8225 | unsigned int i, nelt, which; | |
8226 | bool ret; | |
8227 | ||
8228 | d.vmode = vmode; | |
8229 | d.nelt = nelt = GET_MODE_NUNITS (d.vmode); | |
8230 | d.testing_p = true; | |
8231 | memcpy (d.perm, sel, nelt); | |
8232 | ||
8233 | /* Calculate whether all elements are in one vector. */ | |
8234 | for (i = which = 0; i < nelt; ++i) | |
8235 | { | |
8236 | unsigned char e = d.perm[i]; | |
8237 | gcc_assert (e < 2 * nelt); | |
8238 | which |= (e < nelt ? 1 : 2); | |
8239 | } | |
8240 | ||
8241 | /* If all elements are from the second vector, reindex as if from the | |
8242 | first vector. */ | |
8243 | if (which == 2) | |
8244 | for (i = 0; i < nelt; ++i) | |
8245 | d.perm[i] -= nelt; | |
8246 | ||
8247 | /* Check whether the mask can be applied to a single vector. */ | |
8248 | d.one_vector_p = (which != 3); | |
8249 | ||
8250 | d.target = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 1); | |
8251 | d.op1 = d.op0 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 2); | |
8252 | if (!d.one_vector_p) | |
8253 | d.op1 = gen_raw_REG (d.vmode, LAST_VIRTUAL_REGISTER + 3); | |
8254 | ||
8255 | start_sequence (); | |
8256 | ret = aarch64_expand_vec_perm_const_1 (&d); | |
8257 | end_sequence (); | |
8258 | ||
8259 | return ret; | |
8260 | } | |
8261 | ||
69675d50 TB |
8262 | /* Implement target hook CANNOT_CHANGE_MODE_CLASS. */ |
8263 | bool | |
8264 | aarch64_cannot_change_mode_class (enum machine_mode from, | |
8265 | enum machine_mode to, | |
8266 | enum reg_class rclass) | |
8267 | { | |
8268 | /* Full-reg subregs are allowed on general regs or any class if they are | |
8269 | the same size. */ | |
8270 | if (GET_MODE_SIZE (from) == GET_MODE_SIZE (to) | |
8271 | || !reg_classes_intersect_p (FP_REGS, rclass)) | |
8272 | return false; | |
8273 | ||
8274 | /* Limited combinations of subregs are safe on FPREGs. Particularly, | |
8275 | 1. Vector Mode to Scalar mode where 1 unit of the vector is accessed. | |
8276 | 2. Scalar to Scalar for integer modes or same size float modes. | |
8277 | 3. Vector to Vector modes. */ | |
8278 | if (GET_MODE_SIZE (from) > GET_MODE_SIZE (to)) | |
8279 | { | |
8280 | if (aarch64_vector_mode_supported_p (from) | |
8281 | && GET_MODE_SIZE (GET_MODE_INNER (from)) == GET_MODE_SIZE (to)) | |
8282 | return false; | |
8283 | ||
8284 | if (GET_MODE_NUNITS (from) == 1 | |
8285 | && GET_MODE_NUNITS (to) == 1 | |
8286 | && (GET_MODE_CLASS (from) == MODE_INT | |
8287 | || from == to)) | |
8288 | return false; | |
8289 | ||
8290 | if (aarch64_vector_mode_supported_p (from) | |
8291 | && aarch64_vector_mode_supported_p (to)) | |
8292 | return false; | |
8293 | } | |
8294 | ||
8295 | return true; | |
8296 | } | |
8297 | ||
43e9d192 IB |
8298 | #undef TARGET_ADDRESS_COST |
8299 | #define TARGET_ADDRESS_COST aarch64_address_cost | |
8300 | ||
8301 | /* This hook will determines whether unnamed bitfields affect the alignment | |
8302 | of the containing structure. The hook returns true if the structure | |
8303 | should inherit the alignment requirements of an unnamed bitfield's | |
8304 | type. */ | |
8305 | #undef TARGET_ALIGN_ANON_BITFIELD | |
8306 | #define TARGET_ALIGN_ANON_BITFIELD hook_bool_void_true | |
8307 | ||
8308 | #undef TARGET_ASM_ALIGNED_DI_OP | |
8309 | #define TARGET_ASM_ALIGNED_DI_OP "\t.xword\t" | |
8310 | ||
8311 | #undef TARGET_ASM_ALIGNED_HI_OP | |
8312 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
8313 | ||
8314 | #undef TARGET_ASM_ALIGNED_SI_OP | |
8315 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
8316 | ||
8317 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
8318 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK \ | |
8319 | hook_bool_const_tree_hwi_hwi_const_tree_true | |
8320 | ||
8321 | #undef TARGET_ASM_FILE_START | |
8322 | #define TARGET_ASM_FILE_START aarch64_start_file | |
8323 | ||
8324 | #undef TARGET_ASM_OUTPUT_MI_THUNK | |
8325 | #define TARGET_ASM_OUTPUT_MI_THUNK aarch64_output_mi_thunk | |
8326 | ||
8327 | #undef TARGET_ASM_SELECT_RTX_SECTION | |
8328 | #define TARGET_ASM_SELECT_RTX_SECTION aarch64_select_rtx_section | |
8329 | ||
8330 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
8331 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE aarch64_asm_trampoline_template | |
8332 | ||
8333 | #undef TARGET_BUILD_BUILTIN_VA_LIST | |
8334 | #define TARGET_BUILD_BUILTIN_VA_LIST aarch64_build_builtin_va_list | |
8335 | ||
8336 | #undef TARGET_CALLEE_COPIES | |
8337 | #define TARGET_CALLEE_COPIES hook_bool_CUMULATIVE_ARGS_mode_tree_bool_false | |
8338 | ||
8339 | #undef TARGET_CAN_ELIMINATE | |
8340 | #define TARGET_CAN_ELIMINATE aarch64_can_eliminate | |
8341 | ||
8342 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
8343 | #define TARGET_CANNOT_FORCE_CONST_MEM aarch64_cannot_force_const_mem | |
8344 | ||
8345 | #undef TARGET_CONDITIONAL_REGISTER_USAGE | |
8346 | #define TARGET_CONDITIONAL_REGISTER_USAGE aarch64_conditional_register_usage | |
8347 | ||
8348 | /* Only the least significant bit is used for initialization guard | |
8349 | variables. */ | |
8350 | #undef TARGET_CXX_GUARD_MASK_BIT | |
8351 | #define TARGET_CXX_GUARD_MASK_BIT hook_bool_void_true | |
8352 | ||
8353 | #undef TARGET_C_MODE_FOR_SUFFIX | |
8354 | #define TARGET_C_MODE_FOR_SUFFIX aarch64_c_mode_for_suffix | |
8355 | ||
8356 | #ifdef TARGET_BIG_ENDIAN_DEFAULT | |
8357 | #undef TARGET_DEFAULT_TARGET_FLAGS | |
8358 | #define TARGET_DEFAULT_TARGET_FLAGS (MASK_BIG_END) | |
8359 | #endif | |
8360 | ||
8361 | #undef TARGET_CLASS_MAX_NREGS | |
8362 | #define TARGET_CLASS_MAX_NREGS aarch64_class_max_nregs | |
8363 | ||
119103ca JG |
8364 | #undef TARGET_BUILTIN_DECL |
8365 | #define TARGET_BUILTIN_DECL aarch64_builtin_decl | |
8366 | ||
43e9d192 IB |
8367 | #undef TARGET_EXPAND_BUILTIN |
8368 | #define TARGET_EXPAND_BUILTIN aarch64_expand_builtin | |
8369 | ||
8370 | #undef TARGET_EXPAND_BUILTIN_VA_START | |
8371 | #define TARGET_EXPAND_BUILTIN_VA_START aarch64_expand_builtin_va_start | |
8372 | ||
9697e620 JG |
8373 | #undef TARGET_FOLD_BUILTIN |
8374 | #define TARGET_FOLD_BUILTIN aarch64_fold_builtin | |
8375 | ||
43e9d192 IB |
8376 | #undef TARGET_FUNCTION_ARG |
8377 | #define TARGET_FUNCTION_ARG aarch64_function_arg | |
8378 | ||
8379 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
8380 | #define TARGET_FUNCTION_ARG_ADVANCE aarch64_function_arg_advance | |
8381 | ||
8382 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
8383 | #define TARGET_FUNCTION_ARG_BOUNDARY aarch64_function_arg_boundary | |
8384 | ||
8385 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
8386 | #define TARGET_FUNCTION_OK_FOR_SIBCALL aarch64_function_ok_for_sibcall | |
8387 | ||
8388 | #undef TARGET_FUNCTION_VALUE | |
8389 | #define TARGET_FUNCTION_VALUE aarch64_function_value | |
8390 | ||
8391 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
8392 | #define TARGET_FUNCTION_VALUE_REGNO_P aarch64_function_value_regno_p | |
8393 | ||
8394 | #undef TARGET_FRAME_POINTER_REQUIRED | |
8395 | #define TARGET_FRAME_POINTER_REQUIRED aarch64_frame_pointer_required | |
8396 | ||
0ac198d3 JG |
8397 | #undef TARGET_GIMPLE_FOLD_BUILTIN |
8398 | #define TARGET_GIMPLE_FOLD_BUILTIN aarch64_gimple_fold_builtin | |
8399 | ||
43e9d192 IB |
8400 | #undef TARGET_GIMPLIFY_VA_ARG_EXPR |
8401 | #define TARGET_GIMPLIFY_VA_ARG_EXPR aarch64_gimplify_va_arg_expr | |
8402 | ||
8403 | #undef TARGET_INIT_BUILTINS | |
8404 | #define TARGET_INIT_BUILTINS aarch64_init_builtins | |
8405 | ||
8406 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
8407 | #define TARGET_LEGITIMATE_ADDRESS_P aarch64_legitimate_address_hook_p | |
8408 | ||
8409 | #undef TARGET_LEGITIMATE_CONSTANT_P | |
8410 | #define TARGET_LEGITIMATE_CONSTANT_P aarch64_legitimate_constant_p | |
8411 | ||
8412 | #undef TARGET_LIBGCC_CMP_RETURN_MODE | |
8413 | #define TARGET_LIBGCC_CMP_RETURN_MODE aarch64_libgcc_cmp_return_mode | |
8414 | ||
38e8f663 YR |
8415 | #undef TARGET_LRA_P |
8416 | #define TARGET_LRA_P aarch64_lra_p | |
8417 | ||
ac2b960f YZ |
8418 | #undef TARGET_MANGLE_TYPE |
8419 | #define TARGET_MANGLE_TYPE aarch64_mangle_type | |
8420 | ||
43e9d192 IB |
8421 | #undef TARGET_MEMORY_MOVE_COST |
8422 | #define TARGET_MEMORY_MOVE_COST aarch64_memory_move_cost | |
8423 | ||
8424 | #undef TARGET_MUST_PASS_IN_STACK | |
8425 | #define TARGET_MUST_PASS_IN_STACK must_pass_in_stack_var_size | |
8426 | ||
8427 | /* This target hook should return true if accesses to volatile bitfields | |
8428 | should use the narrowest mode possible. It should return false if these | |
8429 | accesses should use the bitfield container type. */ | |
8430 | #undef TARGET_NARROW_VOLATILE_BITFIELD | |
8431 | #define TARGET_NARROW_VOLATILE_BITFIELD hook_bool_void_false | |
8432 | ||
8433 | #undef TARGET_OPTION_OVERRIDE | |
8434 | #define TARGET_OPTION_OVERRIDE aarch64_override_options | |
8435 | ||
8436 | #undef TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE | |
8437 | #define TARGET_OVERRIDE_OPTIONS_AFTER_CHANGE \ | |
8438 | aarch64_override_options_after_change | |
8439 | ||
8440 | #undef TARGET_PASS_BY_REFERENCE | |
8441 | #define TARGET_PASS_BY_REFERENCE aarch64_pass_by_reference | |
8442 | ||
8443 | #undef TARGET_PREFERRED_RELOAD_CLASS | |
8444 | #define TARGET_PREFERRED_RELOAD_CLASS aarch64_preferred_reload_class | |
8445 | ||
8446 | #undef TARGET_SECONDARY_RELOAD | |
8447 | #define TARGET_SECONDARY_RELOAD aarch64_secondary_reload | |
8448 | ||
8449 | #undef TARGET_SHIFT_TRUNCATION_MASK | |
8450 | #define TARGET_SHIFT_TRUNCATION_MASK aarch64_shift_truncation_mask | |
8451 | ||
8452 | #undef TARGET_SETUP_INCOMING_VARARGS | |
8453 | #define TARGET_SETUP_INCOMING_VARARGS aarch64_setup_incoming_varargs | |
8454 | ||
8455 | #undef TARGET_STRUCT_VALUE_RTX | |
8456 | #define TARGET_STRUCT_VALUE_RTX aarch64_struct_value_rtx | |
8457 | ||
8458 | #undef TARGET_REGISTER_MOVE_COST | |
8459 | #define TARGET_REGISTER_MOVE_COST aarch64_register_move_cost | |
8460 | ||
8461 | #undef TARGET_RETURN_IN_MEMORY | |
8462 | #define TARGET_RETURN_IN_MEMORY aarch64_return_in_memory | |
8463 | ||
8464 | #undef TARGET_RETURN_IN_MSB | |
8465 | #define TARGET_RETURN_IN_MSB aarch64_return_in_msb | |
8466 | ||
8467 | #undef TARGET_RTX_COSTS | |
8468 | #define TARGET_RTX_COSTS aarch64_rtx_costs | |
8469 | ||
d126a4ae AP |
8470 | #undef TARGET_SCHED_ISSUE_RATE |
8471 | #define TARGET_SCHED_ISSUE_RATE aarch64_sched_issue_rate | |
8472 | ||
43e9d192 IB |
8473 | #undef TARGET_TRAMPOLINE_INIT |
8474 | #define TARGET_TRAMPOLINE_INIT aarch64_trampoline_init | |
8475 | ||
8476 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
8477 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P aarch64_use_blocks_for_constant_p | |
8478 | ||
8479 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
8480 | #define TARGET_VECTOR_MODE_SUPPORTED_P aarch64_vector_mode_supported_p | |
8481 | ||
8482 | #undef TARGET_ARRAY_MODE_SUPPORTED_P | |
8483 | #define TARGET_ARRAY_MODE_SUPPORTED_P aarch64_array_mode_supported_p | |
8484 | ||
8990e73a TB |
8485 | #undef TARGET_VECTORIZE_ADD_STMT_COST |
8486 | #define TARGET_VECTORIZE_ADD_STMT_COST aarch64_add_stmt_cost | |
8487 | ||
8488 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST | |
8489 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZATION_COST \ | |
8490 | aarch64_builtin_vectorization_cost | |
8491 | ||
43e9d192 IB |
8492 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE |
8493 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE aarch64_preferred_simd_mode | |
8494 | ||
42fc9a7f JG |
8495 | #undef TARGET_VECTORIZE_BUILTINS |
8496 | #define TARGET_VECTORIZE_BUILTINS | |
8497 | ||
8498 | #undef TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION | |
8499 | #define TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION \ | |
8500 | aarch64_builtin_vectorized_function | |
8501 | ||
3b357264 JG |
8502 | #undef TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES |
8503 | #define TARGET_VECTORIZE_AUTOVECTORIZE_VECTOR_SIZES \ | |
8504 | aarch64_autovectorize_vector_sizes | |
8505 | ||
43e9d192 IB |
8506 | /* Section anchor support. */ |
8507 | ||
8508 | #undef TARGET_MIN_ANCHOR_OFFSET | |
8509 | #define TARGET_MIN_ANCHOR_OFFSET -256 | |
8510 | ||
8511 | /* Limit the maximum anchor offset to 4k-1, since that's the limit for a | |
8512 | byte offset; we can do much more for larger data types, but have no way | |
8513 | to determine the size of the access. We assume accesses are aligned. */ | |
8514 | #undef TARGET_MAX_ANCHOR_OFFSET | |
8515 | #define TARGET_MAX_ANCHOR_OFFSET 4095 | |
8516 | ||
db0253a4 TB |
8517 | #undef TARGET_VECTOR_ALIGNMENT |
8518 | #define TARGET_VECTOR_ALIGNMENT aarch64_simd_vector_alignment | |
8519 | ||
8520 | #undef TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE | |
8521 | #define TARGET_VECTORIZE_VECTOR_ALIGNMENT_REACHABLE \ | |
8522 | aarch64_simd_vector_alignment_reachable | |
8523 | ||
88b08073 JG |
8524 | /* vec_perm support. */ |
8525 | ||
8526 | #undef TARGET_VECTORIZE_VEC_PERM_CONST_OK | |
8527 | #define TARGET_VECTORIZE_VEC_PERM_CONST_OK \ | |
8528 | aarch64_vectorize_vec_perm_const_ok | |
8529 | ||
70f09188 | 8530 | |
706b2314 | 8531 | #undef TARGET_FIXED_CONDITION_CODE_REGS |
70f09188 AP |
8532 | #define TARGET_FIXED_CONDITION_CODE_REGS aarch64_fixed_condition_code_regs |
8533 | ||
43e9d192 IB |
8534 | struct gcc_target targetm = TARGET_INITIALIZER; |
8535 | ||
8536 | #include "gt-aarch64.h" |