]>
Commit | Line | Data |
---|---|---|
1 | /* Subroutines used for code generation of Andes NDS32 cpu for GNU compiler | |
2 | Copyright (C) 2012-2018 Free Software Foundation, Inc. | |
3 | Contributed by Andes Technology Corporation. | |
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 | |
9 | by the Free Software Foundation; either version 3, or (at your | |
10 | option) any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
14 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
15 | 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 | /* ------------------------------------------------------------------------ */ | |
22 | ||
23 | #define IN_TARGET_CODE 1 | |
24 | ||
25 | #include "config.h" | |
26 | #include "system.h" | |
27 | #include "coretypes.h" | |
28 | #include "backend.h" | |
29 | #include "target.h" | |
30 | #include "rtl.h" | |
31 | #include "tree.h" | |
32 | #include "tree-pass.h" | |
33 | #include "stringpool.h" | |
34 | #include "attribs.h" | |
35 | #include "df.h" | |
36 | #include "memmodel.h" | |
37 | #include "tm_p.h" | |
38 | #include "optabs.h" /* For GEN_FCN. */ | |
39 | #include "regs.h" | |
40 | #include "emit-rtl.h" | |
41 | #include "recog.h" | |
42 | #include "diagnostic-core.h" | |
43 | #include "stor-layout.h" | |
44 | #include "varasm.h" | |
45 | #include "calls.h" | |
46 | #include "output.h" | |
47 | #include "explow.h" | |
48 | #include "expr.h" | |
49 | #include "tm-constrs.h" | |
50 | #include "builtins.h" | |
51 | #include "cpplib.h" | |
52 | #include "context.h" | |
53 | ||
54 | /* This file should be included last. */ | |
55 | #include "target-def.h" | |
56 | ||
57 | /* ------------------------------------------------------------------------ */ | |
58 | ||
59 | /* This file is divided into five parts: | |
60 | ||
61 | PART 1: Auxiliary static variable definitions and | |
62 | target hook static variable definitions. | |
63 | ||
64 | PART 2: Auxiliary static function definitions. | |
65 | ||
66 | PART 3: Implement target hook stuff definitions. | |
67 | ||
68 | PART 4: Implemet extern function definitions, | |
69 | the prototype is in nds32-protos.h. | |
70 | ||
71 | PART 5: Initialize target hook structure and definitions. */ | |
72 | ||
73 | /* ------------------------------------------------------------------------ */ | |
74 | ||
75 | /* PART 1: Auxiliary static variable definitions and | |
76 | target hook static variable definitions. */ | |
77 | ||
78 | /* Define intrinsic register names. | |
79 | Please refer to nds32_intrinsic.h file, the index is corresponding to | |
80 | 'enum nds32_intrinsic_registers' data type values. | |
81 | NOTE that the base value starting from 1024. */ | |
82 | static const char * const nds32_intrinsic_register_names[] = | |
83 | { | |
84 | "$CPU_VER", | |
85 | "$ICM_CFG", | |
86 | "$DCM_CFG", | |
87 | "$MMU_CFG", | |
88 | "$MSC_CFG", | |
89 | "$MSC_CFG2", | |
90 | "$CORE_ID", | |
91 | "$FUCOP_EXIST", | |
92 | ||
93 | "$PSW", | |
94 | "$IPSW", | |
95 | "$P_IPSW", | |
96 | "$IVB", | |
97 | "$EVA", | |
98 | "$P_EVA", | |
99 | "$ITYPE", | |
100 | "$P_ITYPE", | |
101 | ||
102 | "$MERR", | |
103 | "$IPC", | |
104 | "$P_IPC", | |
105 | "$OIPC", | |
106 | "$P_P0", | |
107 | "$P_P1", | |
108 | ||
109 | "$INT_MASK", | |
110 | "$INT_MASK2", | |
111 | "$INT_MASK3", | |
112 | "$INT_PEND", | |
113 | "$INT_PEND2", | |
114 | "$INT_PEND3", | |
115 | "$SP_USR", | |
116 | "$SP_PRIV", | |
117 | "$INT_PRI", | |
118 | "$INT_PRI2", | |
119 | "$INT_PRI3", | |
120 | "$INT_PRI4", | |
121 | "$INT_CTRL", | |
122 | "$INT_TRIGGER", | |
123 | "$INT_TRIGGER2", | |
124 | "$INT_GPR_PUSH_DIS", | |
125 | ||
126 | "$MMU_CTL", | |
127 | "$L1_PPTB", | |
128 | "$TLB_VPN", | |
129 | "$TLB_DATA", | |
130 | "$TLB_MISC", | |
131 | "$VLPT_IDX", | |
132 | "$ILMB", | |
133 | "$DLMB", | |
134 | ||
135 | "$CACHE_CTL", | |
136 | "$HSMP_SADDR", | |
137 | "$HSMP_EADDR", | |
138 | "$SDZ_CTL", | |
139 | "$N12MISC_CTL", | |
140 | "$MISC_CTL", | |
141 | "$ECC_MISC", | |
142 | ||
143 | "$BPC0", | |
144 | "$BPC1", | |
145 | "$BPC2", | |
146 | "$BPC3", | |
147 | "$BPC4", | |
148 | "$BPC5", | |
149 | "$BPC6", | |
150 | "$BPC7", | |
151 | ||
152 | "$BPA0", | |
153 | "$BPA1", | |
154 | "$BPA2", | |
155 | "$BPA3", | |
156 | "$BPA4", | |
157 | "$BPA5", | |
158 | "$BPA6", | |
159 | "$BPA7", | |
160 | ||
161 | "$BPAM0", | |
162 | "$BPAM1", | |
163 | "$BPAM2", | |
164 | "$BPAM3", | |
165 | "$BPAM4", | |
166 | "$BPAM5", | |
167 | "$BPAM6", | |
168 | "$BPAM7", | |
169 | ||
170 | "$BPV0", | |
171 | "$BPV1", | |
172 | "$BPV2", | |
173 | "$BPV3", | |
174 | "$BPV4", | |
175 | "$BPV5", | |
176 | "$BPV6", | |
177 | "$BPV7", | |
178 | ||
179 | "$BPCID0", | |
180 | "$BPCID1", | |
181 | "$BPCID2", | |
182 | "$BPCID3", | |
183 | "$BPCID4", | |
184 | "$BPCID5", | |
185 | "$BPCID6", | |
186 | "$BPCID7", | |
187 | ||
188 | "$EDM_CFG", | |
189 | "$EDMSW", | |
190 | "$EDM_CTL", | |
191 | "$EDM_DTR", | |
192 | "$BPMTC", | |
193 | "$DIMBR", | |
194 | ||
195 | "$TECR0", | |
196 | "$TECR1", | |
197 | "$PFMC0", | |
198 | "$PFMC1", | |
199 | "$PFMC2", | |
200 | "$PFM_CTL", | |
201 | "$PFT_CTL", | |
202 | "$HSP_CTL", | |
203 | "$SP_BOUND", | |
204 | "$SP_BOUND_PRIV", | |
205 | "$SP_BASE", | |
206 | "$SP_BASE_PRIV", | |
207 | "$FUCOP_CTL", | |
208 | "$PRUSR_ACC_CTL", | |
209 | ||
210 | "$DMA_CFG", | |
211 | "$DMA_GCSW", | |
212 | "$DMA_CHNSEL", | |
213 | "$DMA_ACT", | |
214 | "$DMA_SETUP", | |
215 | "$DMA_ISADDR", | |
216 | "$DMA_ESADDR", | |
217 | "$DMA_TCNT", | |
218 | "$DMA_STATUS", | |
219 | "$DMA_2DSET", | |
220 | "$DMA_2DSCTL", | |
221 | "$DMA_RCNT", | |
222 | "$DMA_HSTATUS", | |
223 | ||
224 | "$PC", | |
225 | "$SP_USR1", | |
226 | "$SP_USR2", | |
227 | "$SP_USR3", | |
228 | "$SP_PRIV1", | |
229 | "$SP_PRIV2", | |
230 | "$SP_PRIV3", | |
231 | "$BG_REGION", | |
232 | "$SFCR", | |
233 | "$SIGN", | |
234 | "$ISIGN", | |
235 | "$P_ISIGN", | |
236 | "$IFC_LP", | |
237 | "$ITB" | |
238 | }; | |
239 | ||
240 | ||
241 | /* Defining register allocation order for performance. | |
242 | We want to allocate callee-saved registers after others. | |
243 | It may be used by nds32_adjust_reg_alloc_order(). */ | |
244 | static const int nds32_reg_alloc_order_for_speed[] = | |
245 | { | |
246 | 0, 1, 2, 3, 4, 5, 16, 17, | |
247 | 18, 19, 20, 21, 22, 23, 24, 25, | |
248 | 26, 27, 6, 7, 8, 9, 10, 11, | |
249 | 12, 13, 14, 15 | |
250 | }; | |
251 | ||
252 | /* Defining target-specific uses of __attribute__. */ | |
253 | static const struct attribute_spec nds32_attribute_table[] = | |
254 | { | |
255 | /* Syntax: { name, min_len, max_len, decl_required, type_required, | |
256 | function_type_required, affects_type_identity, handler, | |
257 | exclude } */ | |
258 | ||
259 | /* The interrupt vid: [0-63]+ (actual vector number starts from 9 to 72). */ | |
260 | { "interrupt", 1, 64, false, false, false, false, NULL, NULL }, | |
261 | /* The exception vid: [1-8]+ (actual vector number starts from 1 to 8). */ | |
262 | { "exception", 1, 8, false, false, false, false, NULL, NULL }, | |
263 | /* Argument is user's interrupt numbers. The vector number is always 0. */ | |
264 | { "reset", 1, 1, false, false, false, false, NULL, NULL }, | |
265 | ||
266 | /* The attributes describing isr nested type. */ | |
267 | { "nested", 0, 0, false, false, false, false, NULL, NULL }, | |
268 | { "not_nested", 0, 0, false, false, false, false, NULL, NULL }, | |
269 | { "nested_ready", 0, 0, false, false, false, false, NULL, NULL }, | |
270 | ||
271 | /* The attributes describing isr register save scheme. */ | |
272 | { "save_all", 0, 0, false, false, false, false, NULL, NULL }, | |
273 | { "partial_save", 0, 0, false, false, false, false, NULL, NULL }, | |
274 | ||
275 | /* The attributes used by reset attribute. */ | |
276 | { "nmi", 1, 1, false, false, false, false, NULL, NULL }, | |
277 | { "warm", 1, 1, false, false, false, false, NULL, NULL }, | |
278 | ||
279 | /* The attribute telling no prologue/epilogue. */ | |
280 | { "naked", 0, 0, false, false, false, false, NULL, NULL }, | |
281 | ||
282 | /* The last attribute spec is set to be NULL. */ | |
283 | { NULL, 0, 0, false, false, false, false, NULL, NULL } | |
284 | }; | |
285 | ||
286 | ||
287 | /* ------------------------------------------------------------------------ */ | |
288 | ||
289 | /* PART 2: Auxiliary static function definitions. */ | |
290 | ||
291 | /* Function to save and restore machine-specific function data. */ | |
292 | static struct machine_function * | |
293 | nds32_init_machine_status (void) | |
294 | { | |
295 | struct machine_function *machine; | |
296 | machine = ggc_cleared_alloc<machine_function> (); | |
297 | ||
298 | /* Initially assume this function needs prologue/epilogue. */ | |
299 | machine->naked_p = 0; | |
300 | ||
301 | /* Initially assume this function does NOT use fp_as_gp optimization. */ | |
302 | machine->fp_as_gp_p = 0; | |
303 | ||
304 | return machine; | |
305 | } | |
306 | ||
307 | /* Function to compute stack frame size and | |
308 | store into cfun->machine structure. */ | |
309 | static void | |
310 | nds32_compute_stack_frame (void) | |
311 | { | |
312 | int r; | |
313 | int block_size; | |
314 | bool v3pushpop_p; | |
315 | ||
316 | /* Because nds32_compute_stack_frame() will be called from different place, | |
317 | everytime we enter this function, we have to assume this function | |
318 | needs prologue/epilogue. */ | |
319 | cfun->machine->naked_p = 0; | |
320 | ||
321 | /* Get variadic arguments size to prepare pretend arguments and | |
322 | we will push them into stack at prologue by ourself. */ | |
323 | cfun->machine->va_args_size = crtl->args.pretend_args_size; | |
324 | if (cfun->machine->va_args_size != 0) | |
325 | { | |
326 | cfun->machine->va_args_first_regno | |
327 | = NDS32_GPR_ARG_FIRST_REGNUM | |
328 | + NDS32_MAX_GPR_REGS_FOR_ARGS | |
329 | - (crtl->args.pretend_args_size / UNITS_PER_WORD); | |
330 | cfun->machine->va_args_last_regno | |
331 | = NDS32_GPR_ARG_FIRST_REGNUM + NDS32_MAX_GPR_REGS_FOR_ARGS - 1; | |
332 | } | |
333 | else | |
334 | { | |
335 | cfun->machine->va_args_first_regno = SP_REGNUM; | |
336 | cfun->machine->va_args_last_regno = SP_REGNUM; | |
337 | } | |
338 | ||
339 | /* Important: We need to make sure that varargs area is 8-byte alignment. */ | |
340 | block_size = cfun->machine->va_args_size; | |
341 | if (!NDS32_DOUBLE_WORD_ALIGN_P (block_size)) | |
342 | { | |
343 | cfun->machine->va_args_area_padding_bytes | |
344 | = NDS32_ROUND_UP_DOUBLE_WORD (block_size) - block_size; | |
345 | } | |
346 | ||
347 | /* Get local variables, incoming variables, and temporary variables size. | |
348 | Note that we need to make sure it is 8-byte alignment because | |
349 | there may be no padding bytes if we are using LRA. */ | |
350 | cfun->machine->local_size = NDS32_ROUND_UP_DOUBLE_WORD (get_frame_size ()); | |
351 | ||
352 | /* Get outgoing arguments size. */ | |
353 | cfun->machine->out_args_size = crtl->outgoing_args_size; | |
354 | ||
355 | /* If $fp value is required to be saved on stack, it needs 4 bytes space. | |
356 | Check whether $fp is ever live. */ | |
357 | cfun->machine->fp_size = (df_regs_ever_live_p (FP_REGNUM)) ? 4 : 0; | |
358 | ||
359 | /* If $gp value is required to be saved on stack, it needs 4 bytes space. | |
360 | Check whether we are using PIC code genration. */ | |
361 | cfun->machine->gp_size = (flag_pic) ? 4 : 0; | |
362 | ||
363 | /* If $lp value is required to be saved on stack, it needs 4 bytes space. | |
364 | Check whether $lp is ever live. */ | |
365 | cfun->machine->lp_size = (df_regs_ever_live_p (LP_REGNUM)) ? 4 : 0; | |
366 | ||
367 | /* Initially there is no padding bytes. */ | |
368 | cfun->machine->callee_saved_area_gpr_padding_bytes = 0; | |
369 | ||
370 | /* Calculate the bytes of saving callee-saved registers on stack. */ | |
371 | cfun->machine->callee_saved_gpr_regs_size = 0; | |
372 | cfun->machine->callee_saved_first_gpr_regno = SP_REGNUM; | |
373 | cfun->machine->callee_saved_last_gpr_regno = SP_REGNUM; | |
374 | cfun->machine->callee_saved_fpr_regs_size = 0; | |
375 | cfun->machine->callee_saved_first_fpr_regno = SP_REGNUM; | |
376 | cfun->machine->callee_saved_last_fpr_regno = SP_REGNUM; | |
377 | ||
378 | /* Currently, there is no need to check $r28~$r31 | |
379 | because we will save them in another way. */ | |
380 | for (r = 0; r < 28; r++) | |
381 | { | |
382 | if (NDS32_REQUIRED_CALLEE_SAVED_P (r)) | |
383 | { | |
384 | /* Mark the first required callee-saved register | |
385 | (only need to set it once). | |
386 | If first regno == SP_REGNUM, we can tell that | |
387 | it is the first time to be here. */ | |
388 | if (cfun->machine->callee_saved_first_gpr_regno == SP_REGNUM) | |
389 | cfun->machine->callee_saved_first_gpr_regno = r; | |
390 | /* Mark the last required callee-saved register. */ | |
391 | cfun->machine->callee_saved_last_gpr_regno = r; | |
392 | } | |
393 | } | |
394 | ||
395 | /* Recording fpu callee-saved register. */ | |
396 | if (TARGET_HARD_FLOAT) | |
397 | { | |
398 | for (r = NDS32_FIRST_FPR_REGNUM; r < NDS32_LAST_FPR_REGNUM; r++) | |
399 | { | |
400 | if (NDS32_REQUIRED_CALLEE_SAVED_P (r)) | |
401 | { | |
402 | /* Mark the first required callee-saved register. */ | |
403 | if (cfun->machine->callee_saved_first_fpr_regno == SP_REGNUM) | |
404 | { | |
405 | /* Make first callee-saved number is even, | |
406 | bacause we use doubleword access, and this way | |
407 | promise 8-byte alignemt. */ | |
408 | if (!NDS32_FPR_REGNO_OK_FOR_DOUBLE (r)) | |
409 | cfun->machine->callee_saved_first_fpr_regno = r - 1; | |
410 | else | |
411 | cfun->machine->callee_saved_first_fpr_regno = r; | |
412 | } | |
413 | cfun->machine->callee_saved_last_fpr_regno = r; | |
414 | } | |
415 | } | |
416 | ||
417 | /* Make last callee-saved register number is odd, | |
418 | we hope callee-saved register is even. */ | |
419 | int last_fpr = cfun->machine->callee_saved_last_fpr_regno; | |
420 | if (NDS32_FPR_REGNO_OK_FOR_DOUBLE (last_fpr)) | |
421 | cfun->machine->callee_saved_last_fpr_regno++; | |
422 | } | |
423 | ||
424 | /* Check if this function can omit prologue/epilogue code fragment. | |
425 | If there is 'naked' attribute in this function, | |
426 | we can set 'naked_p' flag to indicate that | |
427 | we do not have to generate prologue/epilogue. | |
428 | Or, if all the following conditions succeed, | |
429 | we can set this function 'naked_p' as well: | |
430 | condition 1: first_regno == last_regno == SP_REGNUM, | |
431 | which means we do not have to save | |
432 | any callee-saved registers. | |
433 | condition 2: Both $lp and $fp are NOT live in this function, | |
434 | which means we do not need to save them and there | |
435 | is no outgoing size. | |
436 | condition 3: There is no local_size, which means | |
437 | we do not need to adjust $sp. */ | |
438 | if (lookup_attribute ("naked", DECL_ATTRIBUTES (current_function_decl)) | |
439 | || (cfun->machine->callee_saved_first_gpr_regno == SP_REGNUM | |
440 | && cfun->machine->callee_saved_last_gpr_regno == SP_REGNUM | |
441 | && cfun->machine->callee_saved_first_fpr_regno == SP_REGNUM | |
442 | && cfun->machine->callee_saved_last_fpr_regno == SP_REGNUM | |
443 | && !df_regs_ever_live_p (FP_REGNUM) | |
444 | && !df_regs_ever_live_p (LP_REGNUM) | |
445 | && cfun->machine->local_size == 0)) | |
446 | { | |
447 | /* Set this function 'naked_p' and other functions can check this flag. | |
448 | Note that in nds32 port, the 'naked_p = 1' JUST means there is no | |
449 | callee-saved, local size, and outgoing size. | |
450 | The varargs space and ret instruction may still present in | |
451 | the prologue/epilogue expanding. */ | |
452 | cfun->machine->naked_p = 1; | |
453 | ||
454 | /* No need to save $fp, $gp, and $lp. | |
455 | We should set these value to be zero | |
456 | so that nds32_initial_elimination_offset() can work properly. */ | |
457 | cfun->machine->fp_size = 0; | |
458 | cfun->machine->gp_size = 0; | |
459 | cfun->machine->lp_size = 0; | |
460 | ||
461 | /* If stack usage computation is required, | |
462 | we need to provide the static stack size. */ | |
463 | if (flag_stack_usage_info) | |
464 | current_function_static_stack_size = 0; | |
465 | ||
466 | /* No need to do following adjustment, return immediately. */ | |
467 | return; | |
468 | } | |
469 | ||
470 | v3pushpop_p = NDS32_V3PUSH_AVAILABLE_P; | |
471 | ||
472 | /* Adjustment for v3push instructions: | |
473 | If we are using v3push (push25/pop25) instructions, | |
474 | we need to make sure Rb is $r6 and Re is | |
475 | located on $r6, $r8, $r10, or $r14. | |
476 | Some results above will be discarded and recomputed. | |
477 | Note that it is only available under V3/V3M ISA and we | |
478 | DO NOT setup following stuff for isr or variadic function. */ | |
479 | if (v3pushpop_p) | |
480 | { | |
481 | /* Recompute: | |
482 | cfun->machine->fp_size | |
483 | cfun->machine->gp_size | |
484 | cfun->machine->lp_size | |
485 | cfun->machine->callee_saved_first_gpr_regno | |
486 | cfun->machine->callee_saved_last_gpr_regno */ | |
487 | ||
488 | /* For v3push instructions, $fp, $gp, and $lp are always saved. */ | |
489 | cfun->machine->fp_size = 4; | |
490 | cfun->machine->gp_size = 4; | |
491 | cfun->machine->lp_size = 4; | |
492 | ||
493 | /* Remember to set Rb = $r6. */ | |
494 | cfun->machine->callee_saved_first_gpr_regno = 6; | |
495 | ||
496 | if (cfun->machine->callee_saved_last_gpr_regno <= 6) | |
497 | { | |
498 | /* Re = $r6 */ | |
499 | cfun->machine->callee_saved_last_gpr_regno = 6; | |
500 | } | |
501 | else if (cfun->machine->callee_saved_last_gpr_regno <= 8) | |
502 | { | |
503 | /* Re = $r8 */ | |
504 | cfun->machine->callee_saved_last_gpr_regno = 8; | |
505 | } | |
506 | else if (cfun->machine->callee_saved_last_gpr_regno <= 10) | |
507 | { | |
508 | /* Re = $r10 */ | |
509 | cfun->machine->callee_saved_last_gpr_regno = 10; | |
510 | } | |
511 | else if (cfun->machine->callee_saved_last_gpr_regno <= 14) | |
512 | { | |
513 | /* Re = $r14 */ | |
514 | cfun->machine->callee_saved_last_gpr_regno = 14; | |
515 | } | |
516 | else if (cfun->machine->callee_saved_last_gpr_regno == SP_REGNUM) | |
517 | { | |
518 | /* If last_regno is SP_REGNUM, which means | |
519 | it is never changed, so set it to Re = $r6. */ | |
520 | cfun->machine->callee_saved_last_gpr_regno = 6; | |
521 | } | |
522 | else | |
523 | { | |
524 | /* The program flow should not go here. */ | |
525 | gcc_unreachable (); | |
526 | } | |
527 | } | |
528 | ||
529 | int sp_adjust = cfun->machine->local_size | |
530 | + cfun->machine->out_args_size | |
531 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
532 | + cfun->machine->callee_saved_fpr_regs_size; | |
533 | ||
534 | if (!v3pushpop_p | |
535 | && sp_adjust == 0 | |
536 | && !frame_pointer_needed) | |
537 | { | |
538 | block_size = cfun->machine->fp_size | |
539 | + cfun->machine->gp_size | |
540 | + cfun->machine->lp_size | |
541 | + (4 * (cfun->machine->callee_saved_last_gpr_regno | |
542 | - cfun->machine->callee_saved_first_gpr_regno | |
543 | + 1)); | |
544 | ||
545 | if (!NDS32_DOUBLE_WORD_ALIGN_P (block_size)) | |
546 | { | |
547 | /* $r14 is last callee save register. */ | |
548 | if (cfun->machine->callee_saved_last_gpr_regno | |
549 | < NDS32_LAST_CALLEE_SAVE_GPR_REGNUM) | |
550 | { | |
551 | cfun->machine->callee_saved_last_gpr_regno++; | |
552 | } | |
553 | else if (cfun->machine->callee_saved_first_gpr_regno == SP_REGNUM) | |
554 | { | |
555 | cfun->machine->callee_saved_first_gpr_regno | |
556 | = NDS32_FIRST_CALLEE_SAVE_GPR_REGNUM; | |
557 | cfun->machine->callee_saved_last_gpr_regno | |
558 | = NDS32_FIRST_CALLEE_SAVE_GPR_REGNUM; | |
559 | } | |
560 | } | |
561 | } | |
562 | ||
563 | /* We have correctly set callee_saved_first_gpr_regno | |
564 | and callee_saved_last_gpr_regno. | |
565 | Initially, the callee_saved_gpr_regs_size is supposed to be 0. | |
566 | As long as callee_saved_last_gpr_regno is not SP_REGNUM, | |
567 | we can update callee_saved_gpr_regs_size with new size. */ | |
568 | if (cfun->machine->callee_saved_last_gpr_regno != SP_REGNUM) | |
569 | { | |
570 | /* Compute pushed size of callee-saved registers. */ | |
571 | cfun->machine->callee_saved_gpr_regs_size | |
572 | = 4 * (cfun->machine->callee_saved_last_gpr_regno | |
573 | - cfun->machine->callee_saved_first_gpr_regno | |
574 | + 1); | |
575 | } | |
576 | ||
577 | if (TARGET_HARD_FLOAT) | |
578 | { | |
579 | /* Compute size of callee svaed floating-point registers. */ | |
580 | if (cfun->machine->callee_saved_last_fpr_regno != SP_REGNUM) | |
581 | { | |
582 | cfun->machine->callee_saved_fpr_regs_size | |
583 | = 4 * (cfun->machine->callee_saved_last_fpr_regno | |
584 | - cfun->machine->callee_saved_first_fpr_regno | |
585 | + 1); | |
586 | } | |
587 | } | |
588 | ||
589 | /* Important: We need to make sure that | |
590 | (fp_size + gp_size + lp_size + callee_saved_gpr_regs_size) | |
591 | is 8-byte alignment. | |
592 | If it is not, calculate the padding bytes. */ | |
593 | block_size = cfun->machine->fp_size | |
594 | + cfun->machine->gp_size | |
595 | + cfun->machine->lp_size | |
596 | + cfun->machine->callee_saved_gpr_regs_size; | |
597 | if (!NDS32_DOUBLE_WORD_ALIGN_P (block_size)) | |
598 | { | |
599 | cfun->machine->callee_saved_area_gpr_padding_bytes | |
600 | = NDS32_ROUND_UP_DOUBLE_WORD (block_size) - block_size; | |
601 | } | |
602 | ||
603 | /* If stack usage computation is required, | |
604 | we need to provide the static stack size. */ | |
605 | if (flag_stack_usage_info) | |
606 | { | |
607 | current_function_static_stack_size | |
608 | = NDS32_ROUND_UP_DOUBLE_WORD (block_size) | |
609 | + cfun->machine->local_size | |
610 | + cfun->machine->out_args_size; | |
611 | } | |
612 | } | |
613 | ||
614 | /* Function to create a parallel rtx pattern | |
615 | which presents stack push multiple behavior. | |
616 | The overall concept are: | |
617 | "push registers to memory", | |
618 | "adjust stack pointer". */ | |
619 | static void | |
620 | nds32_emit_stack_push_multiple (unsigned Rb, unsigned Re, | |
621 | bool save_fp_p, bool save_gp_p, bool save_lp_p, | |
622 | bool vaarg_p) | |
623 | { | |
624 | unsigned regno; | |
625 | int extra_count; | |
626 | int num_use_regs; | |
627 | int par_index; | |
628 | int offset; | |
629 | ||
630 | rtx reg; | |
631 | rtx mem; | |
632 | rtx push_rtx; | |
633 | rtx adjust_sp_rtx; | |
634 | rtx parallel_insn; | |
635 | rtx dwarf; | |
636 | ||
637 | /* We need to provide a customized rtx which contains | |
638 | necessary information for data analysis, | |
639 | so we create a parallel rtx like this: | |
640 | (parallel [(set (mem (plus (reg:SI SP_REGNUM) (const_int -32))) | |
641 | (reg:SI Rb)) | |
642 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -28))) | |
643 | (reg:SI Rb+1)) | |
644 | ... | |
645 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -16))) | |
646 | (reg:SI Re)) | |
647 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -12))) | |
648 | (reg:SI FP_REGNUM)) | |
649 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -8))) | |
650 | (reg:SI GP_REGNUM)) | |
651 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -4))) | |
652 | (reg:SI LP_REGNUM)) | |
653 | (set (reg:SI SP_REGNUM) | |
654 | (plus (reg:SI SP_REGNUM) (const_int -32)))]) */ | |
655 | ||
656 | /* Calculate the number of registers that will be pushed. */ | |
657 | extra_count = 0; | |
658 | if (save_fp_p) | |
659 | extra_count++; | |
660 | if (save_gp_p) | |
661 | extra_count++; | |
662 | if (save_lp_p) | |
663 | extra_count++; | |
664 | /* Note that Rb and Re may be SP_REGNUM. DO NOT count it in. */ | |
665 | if (Rb == SP_REGNUM && Re == SP_REGNUM) | |
666 | num_use_regs = extra_count; | |
667 | else | |
668 | num_use_regs = Re - Rb + 1 + extra_count; | |
669 | ||
670 | /* In addition to used registers, | |
671 | we need one more space for (set sp sp-x) rtx. */ | |
672 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
673 | rtvec_alloc (num_use_regs + 1)); | |
674 | par_index = 0; | |
675 | ||
676 | /* Initialize offset and start to create push behavior. */ | |
677 | offset = -(num_use_regs * 4); | |
678 | ||
679 | /* Create (set mem regX) from Rb, Rb+1 up to Re. */ | |
680 | for (regno = Rb; regno <= Re; regno++) | |
681 | { | |
682 | /* Rb and Re may be SP_REGNUM. | |
683 | We need to break this loop immediately. */ | |
684 | if (regno == SP_REGNUM) | |
685 | break; | |
686 | ||
687 | reg = gen_rtx_REG (SImode, regno); | |
688 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
689 | stack_pointer_rtx, | |
690 | offset)); | |
691 | push_rtx = gen_rtx_SET (mem, reg); | |
692 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
693 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
694 | offset = offset + 4; | |
695 | par_index++; | |
696 | } | |
697 | ||
698 | /* Create (set mem fp), (set mem gp), and (set mem lp) if necessary. */ | |
699 | if (save_fp_p) | |
700 | { | |
701 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
702 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
703 | stack_pointer_rtx, | |
704 | offset)); | |
705 | push_rtx = gen_rtx_SET (mem, reg); | |
706 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
707 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
708 | offset = offset + 4; | |
709 | par_index++; | |
710 | } | |
711 | if (save_gp_p) | |
712 | { | |
713 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
714 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
715 | stack_pointer_rtx, | |
716 | offset)); | |
717 | push_rtx = gen_rtx_SET (mem, reg); | |
718 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
719 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
720 | offset = offset + 4; | |
721 | par_index++; | |
722 | } | |
723 | if (save_lp_p) | |
724 | { | |
725 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
726 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
727 | stack_pointer_rtx, | |
728 | offset)); | |
729 | push_rtx = gen_rtx_SET (mem, reg); | |
730 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
731 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
732 | offset = offset + 4; | |
733 | par_index++; | |
734 | } | |
735 | ||
736 | /* Create (set sp sp-x). */ | |
737 | ||
738 | /* We need to re-calculate the offset value again for adjustment. */ | |
739 | offset = -(num_use_regs * 4); | |
740 | adjust_sp_rtx | |
741 | = gen_rtx_SET (stack_pointer_rtx, | |
742 | plus_constant (Pmode, stack_pointer_rtx, offset)); | |
743 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
744 | RTX_FRAME_RELATED_P (adjust_sp_rtx) = 1; | |
745 | ||
746 | parallel_insn = emit_insn (parallel_insn); | |
747 | ||
748 | /* The insn rtx 'parallel_insn' will change frame layout. | |
749 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
750 | generate CFI (Call Frame Information) stuff. */ | |
751 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
752 | ||
753 | /* Don't use GCC's logic for CFI info if we are generate a push for VAARG | |
754 | since we will not restore those register at epilogue. */ | |
755 | if (vaarg_p) | |
756 | { | |
757 | dwarf = alloc_reg_note (REG_CFA_ADJUST_CFA, | |
758 | copy_rtx (adjust_sp_rtx), NULL_RTX); | |
759 | REG_NOTES (parallel_insn) = dwarf; | |
760 | } | |
761 | } | |
762 | ||
763 | /* Function to create a parallel rtx pattern | |
764 | which presents stack pop multiple behavior. | |
765 | The overall concept are: | |
766 | "pop registers from memory", | |
767 | "adjust stack pointer". */ | |
768 | static void | |
769 | nds32_emit_stack_pop_multiple (unsigned Rb, unsigned Re, | |
770 | bool save_fp_p, bool save_gp_p, bool save_lp_p) | |
771 | { | |
772 | unsigned regno; | |
773 | int extra_count; | |
774 | int num_use_regs; | |
775 | int par_index; | |
776 | int offset; | |
777 | ||
778 | rtx reg; | |
779 | rtx mem; | |
780 | rtx pop_rtx; | |
781 | rtx adjust_sp_rtx; | |
782 | rtx parallel_insn; | |
783 | rtx dwarf = NULL_RTX; | |
784 | ||
785 | /* We need to provide a customized rtx which contains | |
786 | necessary information for data analysis, | |
787 | so we create a parallel rtx like this: | |
788 | (parallel [(set (reg:SI Rb) | |
789 | (mem (reg:SI SP_REGNUM))) | |
790 | (set (reg:SI Rb+1) | |
791 | (mem (plus (reg:SI SP_REGNUM) (const_int 4)))) | |
792 | ... | |
793 | (set (reg:SI Re) | |
794 | (mem (plus (reg:SI SP_REGNUM) (const_int 16)))) | |
795 | (set (reg:SI FP_REGNUM) | |
796 | (mem (plus (reg:SI SP_REGNUM) (const_int 20)))) | |
797 | (set (reg:SI GP_REGNUM) | |
798 | (mem (plus (reg:SI SP_REGNUM) (const_int 24)))) | |
799 | (set (reg:SI LP_REGNUM) | |
800 | (mem (plus (reg:SI SP_REGNUM) (const_int 28)))) | |
801 | (set (reg:SI SP_REGNUM) | |
802 | (plus (reg:SI SP_REGNUM) (const_int 32)))]) */ | |
803 | ||
804 | /* Calculate the number of registers that will be poped. */ | |
805 | extra_count = 0; | |
806 | if (save_fp_p) | |
807 | extra_count++; | |
808 | if (save_gp_p) | |
809 | extra_count++; | |
810 | if (save_lp_p) | |
811 | extra_count++; | |
812 | /* Note that Rb and Re may be SP_REGNUM. DO NOT count it in. */ | |
813 | if (Rb == SP_REGNUM && Re == SP_REGNUM) | |
814 | num_use_regs = extra_count; | |
815 | else | |
816 | num_use_regs = Re - Rb + 1 + extra_count; | |
817 | ||
818 | /* In addition to used registers, | |
819 | we need one more space for (set sp sp+x) rtx. */ | |
820 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
821 | rtvec_alloc (num_use_regs + 1)); | |
822 | par_index = 0; | |
823 | ||
824 | /* Initialize offset and start to create pop behavior. */ | |
825 | offset = 0; | |
826 | ||
827 | /* Create (set regX mem) from Rb, Rb+1 up to Re. */ | |
828 | for (regno = Rb; regno <= Re; regno++) | |
829 | { | |
830 | /* Rb and Re may be SP_REGNUM. | |
831 | We need to break this loop immediately. */ | |
832 | if (regno == SP_REGNUM) | |
833 | break; | |
834 | ||
835 | reg = gen_rtx_REG (SImode, regno); | |
836 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
837 | stack_pointer_rtx, | |
838 | offset)); | |
839 | pop_rtx = gen_rtx_SET (reg, mem); | |
840 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
841 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
842 | offset = offset + 4; | |
843 | par_index++; | |
844 | ||
845 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
846 | } | |
847 | ||
848 | /* Create (set fp mem), (set gp mem), and (set lp mem) if necessary. */ | |
849 | if (save_fp_p) | |
850 | { | |
851 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
852 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
853 | stack_pointer_rtx, | |
854 | offset)); | |
855 | pop_rtx = gen_rtx_SET (reg, mem); | |
856 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
857 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
858 | offset = offset + 4; | |
859 | par_index++; | |
860 | ||
861 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
862 | } | |
863 | if (save_gp_p) | |
864 | { | |
865 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
866 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
867 | stack_pointer_rtx, | |
868 | offset)); | |
869 | pop_rtx = gen_rtx_SET (reg, mem); | |
870 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
871 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
872 | offset = offset + 4; | |
873 | par_index++; | |
874 | ||
875 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
876 | } | |
877 | if (save_lp_p) | |
878 | { | |
879 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
880 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
881 | stack_pointer_rtx, | |
882 | offset)); | |
883 | pop_rtx = gen_rtx_SET (reg, mem); | |
884 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
885 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
886 | offset = offset + 4; | |
887 | par_index++; | |
888 | ||
889 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
890 | } | |
891 | ||
892 | /* Create (set sp sp+x). */ | |
893 | ||
894 | /* The offset value is already in place. No need to re-calculate it. */ | |
895 | adjust_sp_rtx | |
896 | = gen_rtx_SET (stack_pointer_rtx, | |
897 | plus_constant (Pmode, stack_pointer_rtx, offset)); | |
898 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
899 | ||
900 | /* Tell gcc we adjust SP in this insn. */ | |
901 | dwarf = alloc_reg_note (REG_CFA_ADJUST_CFA, copy_rtx (adjust_sp_rtx), dwarf); | |
902 | ||
903 | parallel_insn = emit_insn (parallel_insn); | |
904 | ||
905 | /* The insn rtx 'parallel_insn' will change frame layout. | |
906 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
907 | generate CFI (Call Frame Information) stuff. */ | |
908 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
909 | ||
910 | /* Add CFI info by manual. */ | |
911 | REG_NOTES (parallel_insn) = dwarf; | |
912 | } | |
913 | ||
914 | /* Function to create a parallel rtx pattern | |
915 | which presents stack v3push behavior. | |
916 | The overall concept are: | |
917 | "push registers to memory", | |
918 | "adjust stack pointer". */ | |
919 | static void | |
920 | nds32_emit_stack_v3push (unsigned Rb, | |
921 | unsigned Re, | |
922 | unsigned imm8u) | |
923 | { | |
924 | unsigned regno; | |
925 | int num_use_regs; | |
926 | int par_index; | |
927 | int offset; | |
928 | ||
929 | rtx reg; | |
930 | rtx mem; | |
931 | rtx push_rtx; | |
932 | rtx adjust_sp_rtx; | |
933 | rtx parallel_insn; | |
934 | ||
935 | /* We need to provide a customized rtx which contains | |
936 | necessary information for data analysis, | |
937 | so we create a parallel rtx like this: | |
938 | (parallel [(set (mem (plus (reg:SI SP_REGNUM) (const_int -32))) | |
939 | (reg:SI Rb)) | |
940 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -28))) | |
941 | (reg:SI Rb+1)) | |
942 | ... | |
943 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -16))) | |
944 | (reg:SI Re)) | |
945 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -12))) | |
946 | (reg:SI FP_REGNUM)) | |
947 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -8))) | |
948 | (reg:SI GP_REGNUM)) | |
949 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -4))) | |
950 | (reg:SI LP_REGNUM)) | |
951 | (set (reg:SI SP_REGNUM) | |
952 | (plus (reg:SI SP_REGNUM) (const_int -32-imm8u)))]) */ | |
953 | ||
954 | /* Calculate the number of registers that will be pushed. | |
955 | Since $fp, $gp, and $lp is always pushed with v3push instruction, | |
956 | we need to count these three registers. | |
957 | Under v3push, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
958 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
959 | num_use_regs = Re - Rb + 1 + 3; | |
960 | ||
961 | /* In addition to used registers, | |
962 | we need one more space for (set sp sp-x-imm8u) rtx. */ | |
963 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
964 | rtvec_alloc (num_use_regs + 1)); | |
965 | par_index = 0; | |
966 | ||
967 | /* Initialize offset and start to create push behavior. */ | |
968 | offset = -(num_use_regs * 4); | |
969 | ||
970 | /* Create (set mem regX) from Rb, Rb+1 up to Re. | |
971 | Under v3push, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
972 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
973 | for (regno = Rb; regno <= Re; regno++) | |
974 | { | |
975 | reg = gen_rtx_REG (SImode, regno); | |
976 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
977 | stack_pointer_rtx, | |
978 | offset)); | |
979 | push_rtx = gen_rtx_SET (mem, reg); | |
980 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
981 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
982 | offset = offset + 4; | |
983 | par_index++; | |
984 | } | |
985 | ||
986 | /* Create (set mem fp). */ | |
987 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
988 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
989 | stack_pointer_rtx, | |
990 | offset)); | |
991 | push_rtx = gen_rtx_SET (mem, reg); | |
992 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
993 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
994 | offset = offset + 4; | |
995 | par_index++; | |
996 | /* Create (set mem gp). */ | |
997 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
998 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
999 | stack_pointer_rtx, | |
1000 | offset)); | |
1001 | push_rtx = gen_rtx_SET (mem, reg); | |
1002 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
1003 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
1004 | offset = offset + 4; | |
1005 | par_index++; | |
1006 | /* Create (set mem lp). */ | |
1007 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
1008 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
1009 | stack_pointer_rtx, | |
1010 | offset)); | |
1011 | push_rtx = gen_rtx_SET (mem, reg); | |
1012 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
1013 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
1014 | offset = offset + 4; | |
1015 | par_index++; | |
1016 | ||
1017 | /* Create (set sp sp-x-imm8u). */ | |
1018 | ||
1019 | /* We need to re-calculate the offset value again for adjustment. */ | |
1020 | offset = -(num_use_regs * 4); | |
1021 | adjust_sp_rtx | |
1022 | = gen_rtx_SET (stack_pointer_rtx, | |
1023 | plus_constant (Pmode, | |
1024 | stack_pointer_rtx, | |
1025 | offset - imm8u)); | |
1026 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
1027 | RTX_FRAME_RELATED_P (adjust_sp_rtx) = 1; | |
1028 | ||
1029 | parallel_insn = emit_insn (parallel_insn); | |
1030 | ||
1031 | /* The insn rtx 'parallel_insn' will change frame layout. | |
1032 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
1033 | generate CFI (Call Frame Information) stuff. */ | |
1034 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
1035 | } | |
1036 | ||
1037 | /* Function to create a parallel rtx pattern | |
1038 | which presents stack v3pop behavior. | |
1039 | The overall concept are: | |
1040 | "pop registers from memory", | |
1041 | "adjust stack pointer". */ | |
1042 | static void | |
1043 | nds32_emit_stack_v3pop (unsigned Rb, | |
1044 | unsigned Re, | |
1045 | unsigned imm8u) | |
1046 | { | |
1047 | unsigned regno; | |
1048 | int num_use_regs; | |
1049 | int par_index; | |
1050 | int offset; | |
1051 | ||
1052 | rtx reg; | |
1053 | rtx mem; | |
1054 | rtx pop_rtx; | |
1055 | rtx adjust_sp_rtx; | |
1056 | rtx parallel_insn; | |
1057 | rtx dwarf = NULL_RTX; | |
1058 | ||
1059 | /* We need to provide a customized rtx which contains | |
1060 | necessary information for data analysis, | |
1061 | so we create a parallel rtx like this: | |
1062 | (parallel [(set (reg:SI Rb) | |
1063 | (mem (reg:SI SP_REGNUM))) | |
1064 | (set (reg:SI Rb+1) | |
1065 | (mem (plus (reg:SI SP_REGNUM) (const_int 4)))) | |
1066 | ... | |
1067 | (set (reg:SI Re) | |
1068 | (mem (plus (reg:SI SP_REGNUM) (const_int 16)))) | |
1069 | (set (reg:SI FP_REGNUM) | |
1070 | (mem (plus (reg:SI SP_REGNUM) (const_int 20)))) | |
1071 | (set (reg:SI GP_REGNUM) | |
1072 | (mem (plus (reg:SI SP_REGNUM) (const_int 24)))) | |
1073 | (set (reg:SI LP_REGNUM) | |
1074 | (mem (plus (reg:SI SP_REGNUM) (const_int 28)))) | |
1075 | (set (reg:SI SP_REGNUM) | |
1076 | (plus (reg:SI SP_REGNUM) (const_int 32+imm8u)))]) */ | |
1077 | ||
1078 | /* Calculate the number of registers that will be poped. | |
1079 | Since $fp, $gp, and $lp is always poped with v3pop instruction, | |
1080 | we need to count these three registers. | |
1081 | Under v3push, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
1082 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
1083 | num_use_regs = Re - Rb + 1 + 3; | |
1084 | ||
1085 | /* In addition to used registers, | |
1086 | we need one more space for (set sp sp+x+imm8u) rtx. */ | |
1087 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
1088 | rtvec_alloc (num_use_regs + 1)); | |
1089 | par_index = 0; | |
1090 | ||
1091 | /* Initialize offset and start to create pop behavior. */ | |
1092 | offset = 0; | |
1093 | ||
1094 | /* Create (set regX mem) from Rb, Rb+1 up to Re. | |
1095 | Under v3pop, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
1096 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
1097 | for (regno = Rb; regno <= Re; regno++) | |
1098 | { | |
1099 | reg = gen_rtx_REG (SImode, regno); | |
1100 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
1101 | stack_pointer_rtx, | |
1102 | offset)); | |
1103 | pop_rtx = gen_rtx_SET (reg, mem); | |
1104 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
1105 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
1106 | offset = offset + 4; | |
1107 | par_index++; | |
1108 | ||
1109 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
1110 | } | |
1111 | ||
1112 | /* Create (set fp mem). */ | |
1113 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
1114 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
1115 | stack_pointer_rtx, | |
1116 | offset)); | |
1117 | pop_rtx = gen_rtx_SET (reg, mem); | |
1118 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
1119 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
1120 | offset = offset + 4; | |
1121 | par_index++; | |
1122 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
1123 | ||
1124 | /* Create (set gp mem). */ | |
1125 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
1126 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
1127 | stack_pointer_rtx, | |
1128 | offset)); | |
1129 | pop_rtx = gen_rtx_SET (reg, mem); | |
1130 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
1131 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
1132 | offset = offset + 4; | |
1133 | par_index++; | |
1134 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
1135 | ||
1136 | /* Create (set lp mem ). */ | |
1137 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
1138 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
1139 | stack_pointer_rtx, | |
1140 | offset)); | |
1141 | pop_rtx = gen_rtx_SET (reg, mem); | |
1142 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
1143 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
1144 | offset = offset + 4; | |
1145 | par_index++; | |
1146 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
1147 | ||
1148 | /* Create (set sp sp+x+imm8u). */ | |
1149 | ||
1150 | /* The offset value is already in place. No need to re-calculate it. */ | |
1151 | adjust_sp_rtx | |
1152 | = gen_rtx_SET (stack_pointer_rtx, | |
1153 | plus_constant (Pmode, | |
1154 | stack_pointer_rtx, | |
1155 | offset + imm8u)); | |
1156 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
1157 | ||
1158 | if (frame_pointer_needed) | |
1159 | { | |
1160 | /* (expr_list:REG_CFA_DEF_CFA (plus:SI (reg/f:SI $sp) | |
1161 | (const_int 0)) | |
1162 | mean reset frame pointer to $sp and reset to offset 0. */ | |
1163 | rtx cfa_adjust_rtx = gen_rtx_PLUS (Pmode, stack_pointer_rtx, | |
1164 | const0_rtx); | |
1165 | dwarf = alloc_reg_note (REG_CFA_DEF_CFA, cfa_adjust_rtx, dwarf); | |
1166 | } | |
1167 | else | |
1168 | { | |
1169 | /* Tell gcc we adjust SP in this insn. */ | |
1170 | dwarf = alloc_reg_note (REG_CFA_ADJUST_CFA, | |
1171 | copy_rtx (adjust_sp_rtx), dwarf); | |
1172 | } | |
1173 | ||
1174 | parallel_insn = emit_insn (parallel_insn); | |
1175 | ||
1176 | /* The insn rtx 'parallel_insn' will change frame layout. | |
1177 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
1178 | generate CFI (Call Frame Information) stuff. */ | |
1179 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
1180 | ||
1181 | /* Add CFI info by manual. */ | |
1182 | REG_NOTES (parallel_insn) = dwarf; | |
1183 | } | |
1184 | ||
1185 | /* Function that may creates more instructions | |
1186 | for large value on adjusting stack pointer. | |
1187 | ||
1188 | In nds32 target, 'addi' can be used for stack pointer | |
1189 | adjustment in prologue/epilogue stage. | |
1190 | However, sometimes there are too many local variables so that | |
1191 | the adjustment value is not able to be fit in the 'addi' instruction. | |
1192 | One solution is to move value into a register | |
1193 | and then use 'add' instruction. | |
1194 | In practice, we use TA_REGNUM ($r15) to accomplish this purpose. */ | |
1195 | static void | |
1196 | nds32_emit_adjust_frame (rtx to_reg, rtx from_reg, int adjust_value) | |
1197 | { | |
1198 | rtx tmp_reg; | |
1199 | rtx frame_adjust_insn; | |
1200 | rtx adjust_value_rtx = GEN_INT (adjust_value); | |
1201 | ||
1202 | if (adjust_value == 0) | |
1203 | return; | |
1204 | ||
1205 | if (!satisfies_constraint_Is15 (adjust_value_rtx)) | |
1206 | { | |
1207 | /* The value is not able to fit in single addi instruction. | |
1208 | Create more instructions of moving value into a register | |
1209 | and then add stack pointer with it. */ | |
1210 | ||
1211 | /* $r15 is going to be temporary register to hold the value. */ | |
1212 | tmp_reg = gen_rtx_REG (SImode, TA_REGNUM); | |
1213 | ||
1214 | /* Create one more instruction to move value | |
1215 | into the temporary register. */ | |
1216 | emit_move_insn (tmp_reg, adjust_value_rtx); | |
1217 | ||
1218 | /* Create new 'add' rtx. */ | |
1219 | frame_adjust_insn = gen_addsi3 (to_reg, | |
1220 | from_reg, | |
1221 | tmp_reg); | |
1222 | /* Emit rtx into insn list and receive its transformed insn rtx. */ | |
1223 | frame_adjust_insn = emit_insn (frame_adjust_insn); | |
1224 | ||
1225 | /* Because (tmp_reg <- full_value) may be split into two | |
1226 | rtl patterns, we can not set its RTX_FRAME_RELATED_P. | |
1227 | We need to construct another (sp <- sp + full_value) | |
1228 | and then insert it into sp_adjust_insn's reg note to | |
1229 | represent a frame related expression. | |
1230 | GCC knows how to refer it and output debug information. */ | |
1231 | ||
1232 | rtx plus_rtx; | |
1233 | rtx set_rtx; | |
1234 | ||
1235 | plus_rtx = plus_constant (Pmode, from_reg, adjust_value); | |
1236 | set_rtx = gen_rtx_SET (to_reg, plus_rtx); | |
1237 | add_reg_note (frame_adjust_insn, REG_FRAME_RELATED_EXPR, set_rtx); | |
1238 | } | |
1239 | else | |
1240 | { | |
1241 | /* Generate sp adjustment instruction if and only if sp_adjust != 0. */ | |
1242 | frame_adjust_insn = gen_addsi3 (to_reg, | |
1243 | from_reg, | |
1244 | adjust_value_rtx); | |
1245 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
1246 | frame_adjust_insn = emit_insn (frame_adjust_insn); | |
1247 | } | |
1248 | ||
1249 | /* The insn rtx 'sp_adjust_insn' will change frame layout. | |
1250 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
1251 | generate CFI (Call Frame Information) stuff. */ | |
1252 | RTX_FRAME_RELATED_P (frame_adjust_insn) = 1; | |
1253 | } | |
1254 | ||
1255 | /* Return true if MODE/TYPE need double word alignment. */ | |
1256 | static bool | |
1257 | nds32_needs_double_word_align (machine_mode mode, const_tree type) | |
1258 | { | |
1259 | unsigned int align; | |
1260 | ||
1261 | /* Pick up the alignment according to the mode or type. */ | |
1262 | align = NDS32_MODE_TYPE_ALIGN (mode, type); | |
1263 | ||
1264 | return (align > PARM_BOUNDARY); | |
1265 | } | |
1266 | ||
1267 | /* Return true if FUNC is a naked function. */ | |
1268 | static bool | |
1269 | nds32_naked_function_p (tree func) | |
1270 | { | |
1271 | tree t; | |
1272 | ||
1273 | if (TREE_CODE (func) != FUNCTION_DECL) | |
1274 | abort (); | |
1275 | ||
1276 | t = lookup_attribute ("naked", DECL_ATTRIBUTES (func)); | |
1277 | ||
1278 | return (t != NULL_TREE); | |
1279 | } | |
1280 | ||
1281 | /* Function that check if 'X' is a valid address register. | |
1282 | The variable 'STRICT' is very important to | |
1283 | make decision for register number. | |
1284 | ||
1285 | STRICT : true | |
1286 | => We are in reload pass or after reload pass. | |
1287 | The register number should be strictly limited in general registers. | |
1288 | ||
1289 | STRICT : false | |
1290 | => Before reload pass, we are free to use any register number. */ | |
1291 | static bool | |
1292 | nds32_address_register_rtx_p (rtx x, bool strict) | |
1293 | { | |
1294 | int regno; | |
1295 | ||
1296 | if (GET_CODE (x) != REG) | |
1297 | return false; | |
1298 | ||
1299 | regno = REGNO (x); | |
1300 | ||
1301 | if (strict) | |
1302 | return REGNO_OK_FOR_BASE_P (regno); | |
1303 | else | |
1304 | return true; | |
1305 | } | |
1306 | ||
1307 | /* Function that check if 'INDEX' is valid to be a index rtx for address. | |
1308 | ||
1309 | OUTER_MODE : Machine mode of outer address rtx. | |
1310 | INDEX : Check if this rtx is valid to be a index for address. | |
1311 | STRICT : If it is true, we are in reload pass or after reload pass. */ | |
1312 | static bool | |
1313 | nds32_legitimate_index_p (machine_mode outer_mode, | |
1314 | rtx index, | |
1315 | bool strict) | |
1316 | { | |
1317 | int regno; | |
1318 | rtx op0; | |
1319 | rtx op1; | |
1320 | ||
1321 | switch (GET_CODE (index)) | |
1322 | { | |
1323 | case REG: | |
1324 | regno = REGNO (index); | |
1325 | /* If we are in reload pass or after reload pass, | |
1326 | we need to limit it to general register. */ | |
1327 | if (strict) | |
1328 | return REGNO_OK_FOR_INDEX_P (regno); | |
1329 | else | |
1330 | return true; | |
1331 | ||
1332 | case CONST_INT: | |
1333 | /* The alignment of the integer value is determined by 'outer_mode'. */ | |
1334 | switch (GET_MODE_SIZE (outer_mode)) | |
1335 | { | |
1336 | case 1: | |
1337 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1338 | if (satisfies_constraint_Is15 (index)) | |
1339 | return true; | |
1340 | break; | |
1341 | ||
1342 | case 2: | |
1343 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1344 | if (satisfies_constraint_Is16 (index)) | |
1345 | { | |
1346 | /* Make sure address is half word alignment. */ | |
1347 | if (NDS32_HALF_WORD_ALIGN_P (INTVAL (index))) | |
1348 | return true; | |
1349 | } | |
1350 | break; | |
1351 | ||
1352 | case 4: | |
1353 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1354 | if (satisfies_constraint_Is17 (index)) | |
1355 | { | |
1356 | if ((TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE)) | |
1357 | { | |
1358 | if (!satisfies_constraint_Is14 (index)) | |
1359 | return false; | |
1360 | } | |
1361 | ||
1362 | /* Make sure address is word alignment. */ | |
1363 | if (NDS32_SINGLE_WORD_ALIGN_P (INTVAL (index))) | |
1364 | return true; | |
1365 | } | |
1366 | break; | |
1367 | ||
1368 | case 8: | |
1369 | if (satisfies_constraint_Is17 (gen_int_mode (INTVAL (index) + 4, | |
1370 | SImode))) | |
1371 | { | |
1372 | if ((TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE)) | |
1373 | { | |
1374 | if (!satisfies_constraint_Is14 (index)) | |
1375 | return false; | |
1376 | } | |
1377 | ||
1378 | /* Make sure address is word alignment. | |
1379 | Currently we do not have 64-bit load/store yet, | |
1380 | so we will use two 32-bit load/store instructions to do | |
1381 | memory access and they are single word alignment. */ | |
1382 | if (NDS32_SINGLE_WORD_ALIGN_P (INTVAL (index))) | |
1383 | return true; | |
1384 | } | |
1385 | break; | |
1386 | ||
1387 | default: | |
1388 | return false; | |
1389 | } | |
1390 | ||
1391 | return false; | |
1392 | ||
1393 | case MULT: | |
1394 | op0 = XEXP (index, 0); | |
1395 | op1 = XEXP (index, 1); | |
1396 | ||
1397 | if (REG_P (op0) && CONST_INT_P (op1)) | |
1398 | { | |
1399 | int multiplier; | |
1400 | multiplier = INTVAL (op1); | |
1401 | ||
1402 | /* We only allow (mult reg const_int_1), (mult reg const_int_2), | |
1403 | (mult reg const_int_4) or (mult reg const_int_8). */ | |
1404 | if (multiplier != 1 && multiplier != 2 | |
1405 | && multiplier != 4 && multiplier != 8) | |
1406 | return false; | |
1407 | ||
1408 | regno = REGNO (op0); | |
1409 | /* Limit it in general registers if we are | |
1410 | in reload pass or after reload pass. */ | |
1411 | if(strict) | |
1412 | return REGNO_OK_FOR_INDEX_P (regno); | |
1413 | else | |
1414 | return true; | |
1415 | } | |
1416 | ||
1417 | return false; | |
1418 | ||
1419 | case ASHIFT: | |
1420 | op0 = XEXP (index, 0); | |
1421 | op1 = XEXP (index, 1); | |
1422 | ||
1423 | if (REG_P (op0) && CONST_INT_P (op1)) | |
1424 | { | |
1425 | int sv; | |
1426 | /* op1 is already the sv value for use to do left shift. */ | |
1427 | sv = INTVAL (op1); | |
1428 | ||
1429 | /* We only allow (ashift reg const_int_0) | |
1430 | or (ashift reg const_int_1) or (ashift reg const_int_2) or | |
1431 | (ashift reg const_int_3). */ | |
1432 | if (sv != 0 && sv != 1 && sv !=2 && sv != 3) | |
1433 | return false; | |
1434 | ||
1435 | regno = REGNO (op0); | |
1436 | /* Limit it in general registers if we are | |
1437 | in reload pass or after reload pass. */ | |
1438 | if(strict) | |
1439 | return REGNO_OK_FOR_INDEX_P (regno); | |
1440 | else | |
1441 | return true; | |
1442 | } | |
1443 | ||
1444 | return false; | |
1445 | ||
1446 | default: | |
1447 | return false; | |
1448 | } | |
1449 | } | |
1450 | ||
1451 | static void | |
1452 | nds32_register_pass ( | |
1453 | rtl_opt_pass *(*make_pass_func) (gcc::context *), | |
1454 | enum pass_positioning_ops pass_pos, | |
1455 | const char *ref_pass_name) | |
1456 | { | |
1457 | opt_pass *new_opt_pass = make_pass_func (g); | |
1458 | ||
1459 | struct register_pass_info insert_pass = | |
1460 | { | |
1461 | new_opt_pass, /* pass */ | |
1462 | ref_pass_name, /* reference_pass_name */ | |
1463 | 1, /* ref_pass_instance_number */ | |
1464 | pass_pos /* po_op */ | |
1465 | }; | |
1466 | ||
1467 | register_pass (&insert_pass); | |
1468 | } | |
1469 | ||
1470 | /* This function is called from nds32_option_override (). | |
1471 | All new passes should be registered here. */ | |
1472 | static void | |
1473 | nds32_register_passes (void) | |
1474 | { | |
1475 | nds32_register_pass ( | |
1476 | make_pass_nds32_relax_opt, | |
1477 | PASS_POS_INSERT_AFTER, | |
1478 | "mach"); | |
1479 | } | |
1480 | ||
1481 | /* ------------------------------------------------------------------------ */ | |
1482 | ||
1483 | /* PART 3: Implement target hook stuff definitions. */ | |
1484 | \f | |
1485 | ||
1486 | /* Computing the Length of an Insn. | |
1487 | Modifies the length assigned to instruction INSN. | |
1488 | LEN is the initially computed length of the insn. */ | |
1489 | int | |
1490 | nds32_adjust_insn_length (rtx_insn *insn, int length) | |
1491 | { | |
1492 | int adjust_value = 0; | |
1493 | switch (recog_memoized (insn)) | |
1494 | { | |
1495 | case CODE_FOR_call_internal: | |
1496 | case CODE_FOR_call_value_internal: | |
1497 | { | |
1498 | /* We need insert a nop after a noretun function call | |
1499 | to prevent software breakpoint corrupt the next function. */ | |
1500 | if (find_reg_note (insn, REG_NORETURN, NULL_RTX)) | |
1501 | { | |
1502 | if (TARGET_16_BIT) | |
1503 | adjust_value += 2; | |
1504 | else | |
1505 | adjust_value += 4; | |
1506 | } | |
1507 | } | |
1508 | return length + adjust_value; | |
1509 | ||
1510 | default: | |
1511 | return length; | |
1512 | } | |
1513 | } | |
1514 | ||
1515 | /* Register Usage. */ | |
1516 | ||
1517 | static void | |
1518 | nds32_conditional_register_usage (void) | |
1519 | { | |
1520 | int regno; | |
1521 | ||
1522 | if (TARGET_HARD_FLOAT) | |
1523 | { | |
1524 | for (regno = NDS32_FIRST_FPR_REGNUM; | |
1525 | regno <= NDS32_LAST_FPR_REGNUM; regno++) | |
1526 | { | |
1527 | fixed_regs[regno] = 0; | |
1528 | if (regno < NDS32_FIRST_FPR_REGNUM + NDS32_MAX_FPR_REGS_FOR_ARGS) | |
1529 | call_used_regs[regno] = 1; | |
1530 | else if (regno >= NDS32_FIRST_FPR_REGNUM + 22 | |
1531 | && regno < NDS32_FIRST_FPR_REGNUM + 48) | |
1532 | call_used_regs[regno] = 1; | |
1533 | else | |
1534 | call_used_regs[regno] = 0; | |
1535 | } | |
1536 | } | |
1537 | else if (TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE) | |
1538 | { | |
1539 | for (regno = NDS32_FIRST_FPR_REGNUM; | |
1540 | regno <= NDS32_LAST_FPR_REGNUM; | |
1541 | regno++) | |
1542 | fixed_regs[regno] = 0; | |
1543 | } | |
1544 | } | |
1545 | ||
1546 | \f | |
1547 | /* Register Classes. */ | |
1548 | ||
1549 | static unsigned char | |
1550 | nds32_class_max_nregs (reg_class_t rclass ATTRIBUTE_UNUSED, | |
1551 | machine_mode mode) | |
1552 | { | |
1553 | /* Return the maximum number of consecutive registers | |
1554 | needed to represent "mode" in a register of "rclass". */ | |
1555 | return ((GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD); | |
1556 | } | |
1557 | ||
1558 | static int | |
1559 | nds32_register_priority (int hard_regno) | |
1560 | { | |
1561 | /* Encourage to use r0-r7 for LRA when optimize for size. */ | |
1562 | if (optimize_size) | |
1563 | { | |
1564 | if (hard_regno < 8) | |
1565 | return 4; | |
1566 | else if (hard_regno < 16) | |
1567 | return 3; | |
1568 | else if (hard_regno < 28) | |
1569 | return 2; | |
1570 | else | |
1571 | return 1; | |
1572 | } | |
1573 | else | |
1574 | { | |
1575 | if (hard_regno > 27) | |
1576 | return 1; | |
1577 | else | |
1578 | return 4; | |
1579 | } | |
1580 | } | |
1581 | ||
1582 | static bool | |
1583 | nds32_can_change_mode_class (machine_mode from, | |
1584 | machine_mode to, | |
1585 | reg_class_t rclass) | |
1586 | { | |
1587 | /* Don't spill double-precision register to two singal-precision | |
1588 | registers */ | |
1589 | if ((TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE) | |
1590 | && GET_MODE_SIZE (from) != GET_MODE_SIZE (to)) | |
1591 | { | |
1592 | return !reg_classes_intersect_p (rclass, FP_REGS); | |
1593 | } | |
1594 | ||
1595 | return true; | |
1596 | } | |
1597 | ||
1598 | \f | |
1599 | /* Stack Layout and Calling Conventions. */ | |
1600 | ||
1601 | /* There are three kinds of pointer concepts using in GCC compiler: | |
1602 | ||
1603 | frame pointer: A pointer to the first location of local variables. | |
1604 | stack pointer: A pointer to the top of a stack frame. | |
1605 | argument pointer: A pointer to the incoming arguments. | |
1606 | ||
1607 | In nds32 target calling convention, we are using 8-byte alignment. | |
1608 | Besides, we would like to have each stack frame of a function includes: | |
1609 | ||
1610 | [Block A] | |
1611 | 1. previous hard frame pointer | |
1612 | 2. return address | |
1613 | 3. callee-saved registers | |
1614 | 4. <padding bytes> (we will calculte in nds32_compute_stack_frame() | |
1615 | and save it at | |
1616 | cfun->machine->callee_saved_area_padding_bytes) | |
1617 | ||
1618 | [Block B] | |
1619 | 1. local variables | |
1620 | 2. spilling location | |
1621 | 3. <padding bytes> (it will be calculated by GCC itself) | |
1622 | 4. incoming arguments | |
1623 | 5. <padding bytes> (it will be calculated by GCC itself) | |
1624 | ||
1625 | [Block C] | |
1626 | 1. <padding bytes> (it will be calculated by GCC itself) | |
1627 | 2. outgoing arguments | |
1628 | ||
1629 | We 'wrap' these blocks together with | |
1630 | hard frame pointer ($r28) and stack pointer ($r31). | |
1631 | By applying the basic frame/stack/argument pointers concept, | |
1632 | the layout of a stack frame shoule be like this: | |
1633 | ||
1634 | | | | |
1635 | old stack pointer -> ---- | |
1636 | | | \ | |
1637 | | | saved arguments for | |
1638 | | | vararg functions | |
1639 | | | / | |
1640 | hard frame pointer -> -- | |
1641 | & argument pointer | | \ | |
1642 | | | previous hardware frame pointer | |
1643 | | | return address | |
1644 | | | callee-saved registers | |
1645 | | | / | |
1646 | frame pointer -> -- | |
1647 | | | \ | |
1648 | | | local variables | |
1649 | | | and incoming arguments | |
1650 | | | / | |
1651 | -- | |
1652 | | | \ | |
1653 | | | outgoing | |
1654 | | | arguments | |
1655 | | | / | |
1656 | stack pointer -> ---- | |
1657 | ||
1658 | $SFP and $AP are used to represent frame pointer and arguments pointer, | |
1659 | which will be both eliminated as hard frame pointer. */ | |
1660 | ||
1661 | /* -- Eliminating Frame Pointer and Arg Pointer. */ | |
1662 | ||
1663 | static bool | |
1664 | nds32_can_eliminate (const int from_reg, const int to_reg) | |
1665 | { | |
1666 | if (from_reg == ARG_POINTER_REGNUM && to_reg == STACK_POINTER_REGNUM) | |
1667 | return true; | |
1668 | ||
1669 | if (from_reg == ARG_POINTER_REGNUM && to_reg == HARD_FRAME_POINTER_REGNUM) | |
1670 | return true; | |
1671 | ||
1672 | if (from_reg == FRAME_POINTER_REGNUM && to_reg == STACK_POINTER_REGNUM) | |
1673 | return true; | |
1674 | ||
1675 | if (from_reg == FRAME_POINTER_REGNUM && to_reg == HARD_FRAME_POINTER_REGNUM) | |
1676 | return true; | |
1677 | ||
1678 | return false; | |
1679 | } | |
1680 | ||
1681 | /* -- Passing Arguments in Registers. */ | |
1682 | ||
1683 | static rtx | |
1684 | nds32_function_arg (cumulative_args_t ca, machine_mode mode, | |
1685 | const_tree type, bool named) | |
1686 | { | |
1687 | unsigned int regno; | |
1688 | CUMULATIVE_ARGS *cum = get_cumulative_args (ca); | |
1689 | ||
1690 | /* The last time this hook is called, | |
1691 | it is called with MODE == VOIDmode. */ | |
1692 | if (mode == VOIDmode) | |
1693 | return NULL_RTX; | |
1694 | ||
1695 | /* For nameless arguments, we need to take care it individually. */ | |
1696 | if (!named) | |
1697 | { | |
1698 | /* If we are under hard float abi, we have arguments passed on the | |
1699 | stack and all situation can be handled by GCC itself. */ | |
1700 | if (TARGET_HARD_FLOAT) | |
1701 | return NULL_RTX; | |
1702 | ||
1703 | if (NDS32_ARG_PARTIAL_IN_GPR_REG_P (cum->gpr_offset, mode, type)) | |
1704 | { | |
1705 | /* If we still have enough registers to pass argument, pick up | |
1706 | next available register number. */ | |
1707 | regno | |
1708 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type); | |
1709 | return gen_rtx_REG (mode, regno); | |
1710 | } | |
1711 | ||
1712 | /* No register available, return NULL_RTX. | |
1713 | The compiler will use stack to pass argument instead. */ | |
1714 | return NULL_RTX; | |
1715 | } | |
1716 | ||
1717 | /* The following is to handle named argument. | |
1718 | Note that the strategies of TARGET_HARD_FLOAT and !TARGET_HARD_FLOAT | |
1719 | are different. */ | |
1720 | if (TARGET_HARD_FLOAT) | |
1721 | { | |
1722 | /* For TARGET_HARD_FLOAT calling convention, we use GPR and FPR | |
1723 | to pass argument. We have to further check TYPE and MODE so | |
1724 | that we can determine which kind of register we shall use. */ | |
1725 | ||
1726 | /* Note that we need to pass argument entirely in registers under | |
1727 | hard float abi. */ | |
1728 | if (GET_MODE_CLASS (mode) == MODE_FLOAT | |
1729 | && NDS32_ARG_ENTIRE_IN_FPR_REG_P (cum->fpr_offset, mode, type)) | |
1730 | { | |
1731 | /* Pick up the next available FPR register number. */ | |
1732 | regno | |
1733 | = NDS32_AVAILABLE_REGNUM_FOR_FPR_ARG (cum->fpr_offset, mode, type); | |
1734 | return gen_rtx_REG (mode, regno); | |
1735 | } | |
1736 | else if (GET_MODE_CLASS (mode) != MODE_FLOAT | |
1737 | && NDS32_ARG_ENTIRE_IN_GPR_REG_P (cum->gpr_offset, mode, type)) | |
1738 | { | |
1739 | /* Pick up the next available GPR register number. */ | |
1740 | regno | |
1741 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type); | |
1742 | return gen_rtx_REG (mode, regno); | |
1743 | } | |
1744 | } | |
1745 | else | |
1746 | { | |
1747 | /* For !TARGET_HARD_FLOAT calling convention, we always use GPR to pass | |
1748 | argument. Since we allow to pass argument partially in registers, | |
1749 | we can just return it if there are still registers available. */ | |
1750 | if (NDS32_ARG_PARTIAL_IN_GPR_REG_P (cum->gpr_offset, mode, type)) | |
1751 | { | |
1752 | /* Pick up the next available register number. */ | |
1753 | regno | |
1754 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type); | |
1755 | return gen_rtx_REG (mode, regno); | |
1756 | } | |
1757 | ||
1758 | } | |
1759 | ||
1760 | /* No register available, return NULL_RTX. | |
1761 | The compiler will use stack to pass argument instead. */ | |
1762 | return NULL_RTX; | |
1763 | } | |
1764 | ||
1765 | static bool | |
1766 | nds32_must_pass_in_stack (machine_mode mode, const_tree type) | |
1767 | { | |
1768 | /* Return true if a type must be passed in memory. | |
1769 | If it is NOT using hard float abi, small aggregates can be | |
1770 | passed in a register even we are calling a variadic function. | |
1771 | So there is no need to take padding into consideration. */ | |
1772 | if (TARGET_HARD_FLOAT) | |
1773 | return must_pass_in_stack_var_size_or_pad (mode, type); | |
1774 | else | |
1775 | return must_pass_in_stack_var_size (mode, type); | |
1776 | } | |
1777 | ||
1778 | static int | |
1779 | nds32_arg_partial_bytes (cumulative_args_t ca, machine_mode mode, | |
1780 | tree type, bool named ATTRIBUTE_UNUSED) | |
1781 | { | |
1782 | /* Returns the number of bytes at the beginning of an argument that | |
1783 | must be put in registers. The value must be zero for arguments that are | |
1784 | passed entirely in registers or that are entirely pushed on the stack. | |
1785 | Besides, TARGET_FUNCTION_ARG for these arguments should return the | |
1786 | first register to be used by the caller for this argument. */ | |
1787 | unsigned int needed_reg_count; | |
1788 | unsigned int remaining_reg_count; | |
1789 | CUMULATIVE_ARGS *cum; | |
1790 | ||
1791 | cum = get_cumulative_args (ca); | |
1792 | ||
1793 | /* Under hard float abi, we better have argument entirely passed in | |
1794 | registers or pushed on the stack so that we can reduce the complexity | |
1795 | of dealing with cum->gpr_offset and cum->fpr_offset. */ | |
1796 | if (TARGET_HARD_FLOAT) | |
1797 | return 0; | |
1798 | ||
1799 | /* If we have already runned out of argument registers, return zero | |
1800 | so that the argument will be entirely pushed on the stack. */ | |
1801 | if (NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1802 | >= NDS32_GPR_ARG_FIRST_REGNUM + NDS32_MAX_GPR_REGS_FOR_ARGS) | |
1803 | return 0; | |
1804 | ||
1805 | /* Calculate how many registers do we need for this argument. */ | |
1806 | needed_reg_count = NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1807 | ||
1808 | /* Calculate how many argument registers have left for passing argument. | |
1809 | Note that we should count it from next available register number. */ | |
1810 | remaining_reg_count | |
1811 | = NDS32_MAX_GPR_REGS_FOR_ARGS | |
1812 | - (NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1813 | - NDS32_GPR_ARG_FIRST_REGNUM); | |
1814 | ||
1815 | /* Note that we have to return the nubmer of bytes, not registers count. */ | |
1816 | if (needed_reg_count > remaining_reg_count) | |
1817 | return remaining_reg_count * UNITS_PER_WORD; | |
1818 | ||
1819 | return 0; | |
1820 | } | |
1821 | ||
1822 | static void | |
1823 | nds32_function_arg_advance (cumulative_args_t ca, machine_mode mode, | |
1824 | const_tree type, bool named) | |
1825 | { | |
1826 | CUMULATIVE_ARGS *cum = get_cumulative_args (ca); | |
1827 | ||
1828 | if (named) | |
1829 | { | |
1830 | /* We need to further check TYPE and MODE so that we can determine | |
1831 | which kind of register we shall advance. */ | |
1832 | ||
1833 | /* Under hard float abi, we may advance FPR registers. */ | |
1834 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (mode) == MODE_FLOAT) | |
1835 | { | |
1836 | cum->fpr_offset | |
1837 | = NDS32_AVAILABLE_REGNUM_FOR_FPR_ARG (cum->fpr_offset, mode, type) | |
1838 | - NDS32_FPR_ARG_FIRST_REGNUM | |
1839 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1840 | } | |
1841 | else | |
1842 | { | |
1843 | cum->gpr_offset | |
1844 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1845 | - NDS32_GPR_ARG_FIRST_REGNUM | |
1846 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1847 | } | |
1848 | } | |
1849 | else | |
1850 | { | |
1851 | /* If this nameless argument is NOT under TARGET_HARD_FLOAT, | |
1852 | we can advance next register as well so that caller is | |
1853 | able to pass arguments in registers and callee must be | |
1854 | in charge of pushing all of them into stack. */ | |
1855 | if (!TARGET_HARD_FLOAT) | |
1856 | { | |
1857 | cum->gpr_offset | |
1858 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1859 | - NDS32_GPR_ARG_FIRST_REGNUM | |
1860 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1861 | } | |
1862 | } | |
1863 | } | |
1864 | ||
1865 | static unsigned int | |
1866 | nds32_function_arg_boundary (machine_mode mode, const_tree type) | |
1867 | { | |
1868 | return (nds32_needs_double_word_align (mode, type) | |
1869 | ? NDS32_DOUBLE_WORD_ALIGNMENT | |
1870 | : PARM_BOUNDARY); | |
1871 | } | |
1872 | ||
1873 | /* -- How Scalar Function Values Are Returned. */ | |
1874 | ||
1875 | static rtx | |
1876 | nds32_function_value (const_tree ret_type, | |
1877 | const_tree fn_decl_or_type ATTRIBUTE_UNUSED, | |
1878 | bool outgoing ATTRIBUTE_UNUSED) | |
1879 | { | |
1880 | machine_mode mode; | |
1881 | int unsignedp; | |
1882 | ||
1883 | mode = TYPE_MODE (ret_type); | |
1884 | unsignedp = TYPE_UNSIGNED (ret_type); | |
1885 | ||
1886 | if (INTEGRAL_TYPE_P (ret_type)) | |
1887 | mode = promote_mode (ret_type, mode, &unsignedp); | |
1888 | ||
1889 | if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode)) | |
1890 | return gen_rtx_REG (mode, NDS32_FPR_RET_FIRST_REGNUM); | |
1891 | else | |
1892 | return gen_rtx_REG (mode, NDS32_GPR_RET_FIRST_REGNUM); | |
1893 | } | |
1894 | ||
1895 | static rtx | |
1896 | nds32_libcall_value (machine_mode mode, | |
1897 | const_rtx fun ATTRIBUTE_UNUSED) | |
1898 | { | |
1899 | if (TARGET_HARD_FLOAT && (mode == SFmode || mode == DFmode)) | |
1900 | return gen_rtx_REG (mode, NDS32_FPR_RET_FIRST_REGNUM); | |
1901 | ||
1902 | return gen_rtx_REG (mode, NDS32_GPR_RET_FIRST_REGNUM); | |
1903 | } | |
1904 | ||
1905 | static bool | |
1906 | nds32_function_value_regno_p (const unsigned int regno) | |
1907 | { | |
1908 | if (regno == NDS32_GPR_RET_FIRST_REGNUM | |
1909 | || (TARGET_HARD_FLOAT | |
1910 | && regno == NDS32_FPR_RET_FIRST_REGNUM)) | |
1911 | return true; | |
1912 | ||
1913 | return false; | |
1914 | } | |
1915 | ||
1916 | /* -- How Large Values Are Returned. */ | |
1917 | ||
1918 | static bool | |
1919 | nds32_return_in_memory (const_tree type, | |
1920 | const_tree fntype ATTRIBUTE_UNUSED) | |
1921 | { | |
1922 | /* Note that int_size_in_bytes can return -1 if the size can vary | |
1923 | or is larger than an integer. */ | |
1924 | HOST_WIDE_INT size = int_size_in_bytes (type); | |
1925 | ||
1926 | /* For COMPLEX_TYPE, if the total size cannot be hold within two registers, | |
1927 | the return value is supposed to be in memory. We need to be aware of | |
1928 | that the size may be -1. */ | |
1929 | if (TREE_CODE (type) == COMPLEX_TYPE) | |
1930 | if (size < 0 || size > 2 * UNITS_PER_WORD) | |
1931 | return true; | |
1932 | ||
1933 | /* If it is BLKmode and the total size cannot be hold within two registers, | |
1934 | the return value is supposed to be in memory. We need to be aware of | |
1935 | that the size may be -1. */ | |
1936 | if (TYPE_MODE (type) == BLKmode) | |
1937 | if (size < 0 || size > 2 * UNITS_PER_WORD) | |
1938 | return true; | |
1939 | ||
1940 | /* For other cases, having result in memory is unnecessary. */ | |
1941 | return false; | |
1942 | } | |
1943 | ||
1944 | /* -- Function Entry and Exit. */ | |
1945 | ||
1946 | /* The content produced from this function | |
1947 | will be placed before prologue body. */ | |
1948 | static void | |
1949 | nds32_asm_function_prologue (FILE *file) | |
1950 | { | |
1951 | int r; | |
1952 | const char *func_name; | |
1953 | tree attrs; | |
1954 | tree name; | |
1955 | ||
1956 | /* All stack frame information is supposed to be | |
1957 | already computed when expanding prologue. | |
1958 | The result is in cfun->machine. | |
1959 | DO NOT call nds32_compute_stack_frame() here | |
1960 | because it may corrupt the essential information. */ | |
1961 | ||
1962 | fprintf (file, "\t! BEGIN PROLOGUE\n"); | |
1963 | fprintf (file, "\t! fp needed: %d\n", frame_pointer_needed); | |
1964 | fprintf (file, "\t! pretend_args: %d\n", cfun->machine->va_args_size); | |
1965 | fprintf (file, "\t! local_size: %d\n", cfun->machine->local_size); | |
1966 | fprintf (file, "\t! out_args_size: %d\n", cfun->machine->out_args_size); | |
1967 | ||
1968 | /* Use df_regs_ever_live_p() to detect if the register | |
1969 | is ever used in the current function. */ | |
1970 | fprintf (file, "\t! registers ever_live: "); | |
1971 | for (r = 0; r < 65; r++) | |
1972 | { | |
1973 | if (df_regs_ever_live_p (r)) | |
1974 | fprintf (file, "%s, ", reg_names[r]); | |
1975 | } | |
1976 | fputc ('\n', file); | |
1977 | ||
1978 | /* Display the attributes of this function. */ | |
1979 | fprintf (file, "\t! function attributes: "); | |
1980 | /* Get the attributes tree list. | |
1981 | Note that GCC builds attributes list with reverse order. */ | |
1982 | attrs = DECL_ATTRIBUTES (current_function_decl); | |
1983 | ||
1984 | /* If there is no any attribute, print out "None". */ | |
1985 | if (!attrs) | |
1986 | fprintf (file, "None"); | |
1987 | ||
1988 | /* If there are some attributes, try if we need to | |
1989 | construct isr vector information. */ | |
1990 | func_name = IDENTIFIER_POINTER (DECL_NAME (current_function_decl)); | |
1991 | nds32_construct_isr_vectors_information (attrs, func_name); | |
1992 | ||
1993 | /* Display all attributes of this function. */ | |
1994 | while (attrs) | |
1995 | { | |
1996 | name = TREE_PURPOSE (attrs); | |
1997 | fprintf (file, "%s ", IDENTIFIER_POINTER (name)); | |
1998 | ||
1999 | /* Pick up the next attribute. */ | |
2000 | attrs = TREE_CHAIN (attrs); | |
2001 | } | |
2002 | fputc ('\n', file); | |
2003 | } | |
2004 | ||
2005 | /* After rtl prologue has been expanded, this function is used. */ | |
2006 | static void | |
2007 | nds32_asm_function_end_prologue (FILE *file) | |
2008 | { | |
2009 | fprintf (file, "\t! END PROLOGUE\n"); | |
2010 | ||
2011 | /* If frame pointer is NOT needed and -mfp-as-gp is issued, | |
2012 | we can generate special directive: ".omit_fp_begin" | |
2013 | to guide linker doing fp-as-gp optimization. | |
2014 | However, for a naked function, which means | |
2015 | it should not have prologue/epilogue, | |
2016 | using fp-as-gp still requires saving $fp by push/pop behavior and | |
2017 | there is no benefit to use fp-as-gp on such small function. | |
2018 | So we need to make sure this function is NOT naked as well. */ | |
2019 | if (!frame_pointer_needed | |
2020 | && !cfun->machine->naked_p | |
2021 | && cfun->machine->fp_as_gp_p) | |
2022 | { | |
2023 | fprintf (file, "\t! ----------------------------------------\n"); | |
2024 | fprintf (file, "\t! Guide linker to do " | |
2025 | "link time optimization: fp-as-gp\n"); | |
2026 | fprintf (file, "\t! We add one more instruction to " | |
2027 | "initialize $fp near to $gp location.\n"); | |
2028 | fprintf (file, "\t! If linker fails to use fp-as-gp transformation,\n"); | |
2029 | fprintf (file, "\t! this extra instruction should be " | |
2030 | "eliminated at link stage.\n"); | |
2031 | fprintf (file, "\t.omit_fp_begin\n"); | |
2032 | fprintf (file, "\tla\t$fp,_FP_BASE_\n"); | |
2033 | fprintf (file, "\t! ----------------------------------------\n"); | |
2034 | } | |
2035 | } | |
2036 | ||
2037 | /* Before rtl epilogue has been expanded, this function is used. */ | |
2038 | static void | |
2039 | nds32_asm_function_begin_epilogue (FILE *file) | |
2040 | { | |
2041 | /* If frame pointer is NOT needed and -mfp-as-gp is issued, | |
2042 | we can generate special directive: ".omit_fp_end" | |
2043 | to claim fp-as-gp optimization range. | |
2044 | However, for a naked function, | |
2045 | which means it should not have prologue/epilogue, | |
2046 | using fp-as-gp still requires saving $fp by push/pop behavior and | |
2047 | there is no benefit to use fp-as-gp on such small function. | |
2048 | So we need to make sure this function is NOT naked as well. */ | |
2049 | if (!frame_pointer_needed | |
2050 | && !cfun->machine->naked_p | |
2051 | && cfun->machine->fp_as_gp_p) | |
2052 | { | |
2053 | fprintf (file, "\t! ----------------------------------------\n"); | |
2054 | fprintf (file, "\t! Claim the range of fp-as-gp " | |
2055 | "link time optimization\n"); | |
2056 | fprintf (file, "\t.omit_fp_end\n"); | |
2057 | fprintf (file, "\t! ----------------------------------------\n"); | |
2058 | } | |
2059 | ||
2060 | fprintf (file, "\t! BEGIN EPILOGUE\n"); | |
2061 | } | |
2062 | ||
2063 | /* The content produced from this function | |
2064 | will be placed after epilogue body. */ | |
2065 | static void | |
2066 | nds32_asm_function_epilogue (FILE *file) | |
2067 | { | |
2068 | fprintf (file, "\t! END EPILOGUE\n"); | |
2069 | } | |
2070 | ||
2071 | static void | |
2072 | nds32_asm_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
2073 | HOST_WIDE_INT delta, | |
2074 | HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED, | |
2075 | tree function) | |
2076 | { | |
2077 | int this_regno; | |
2078 | ||
2079 | /* Make sure unwind info is emitted for the thunk if needed. */ | |
2080 | final_start_function (emit_barrier (), file, 1); | |
2081 | ||
2082 | this_regno = (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function) | |
2083 | ? 1 | |
2084 | : 0); | |
2085 | ||
2086 | if (delta != 0) | |
2087 | { | |
2088 | if (satisfies_constraint_Is15 (GEN_INT (delta))) | |
2089 | { | |
2090 | fprintf (file, "\taddi\t$r%d, $r%d, " HOST_WIDE_INT_PRINT_DEC "\n", | |
2091 | this_regno, this_regno, delta); | |
2092 | } | |
2093 | else if (satisfies_constraint_Is20 (GEN_INT (delta))) | |
2094 | { | |
2095 | fprintf (file, "\tmovi\t$ta, " HOST_WIDE_INT_PRINT_DEC "\n", delta); | |
2096 | fprintf (file, "\tadd\t$r%d, $r%d, $ta\n", this_regno, this_regno); | |
2097 | } | |
2098 | else | |
2099 | { | |
2100 | fprintf (file, | |
2101 | "\tsethi\t$ta, hi20(" HOST_WIDE_INT_PRINT_DEC ")\n", | |
2102 | delta); | |
2103 | fprintf (file, | |
2104 | "\tori\t$ta, $ta, lo12(" HOST_WIDE_INT_PRINT_DEC ")\n", | |
2105 | delta); | |
2106 | fprintf (file, "\tadd\t$r%d, $r%d, $ta\n", this_regno, this_regno); | |
2107 | } | |
2108 | } | |
2109 | ||
2110 | fprintf (file, "\tb\t"); | |
2111 | assemble_name (file, XSTR (XEXP (DECL_RTL (function), 0), 0)); | |
2112 | fprintf (file, "\n"); | |
2113 | ||
2114 | final_end_function (); | |
2115 | } | |
2116 | ||
2117 | /* -- Permitting tail calls. */ | |
2118 | ||
2119 | /* Return true if it is ok to do sibling call optimization. */ | |
2120 | static bool | |
2121 | nds32_function_ok_for_sibcall (tree decl, | |
2122 | tree exp ATTRIBUTE_UNUSED) | |
2123 | { | |
2124 | /* The DECL is NULL if it is an indirect call. */ | |
2125 | ||
2126 | /* 1. Do not apply sibling call if -mv3push is enabled, | |
2127 | because pop25 instruction also represents return behavior. | |
2128 | 2. If this function is a variadic function, do not apply sibling call | |
2129 | because the stack layout may be a mess. | |
2130 | 3. We don't want to apply sibling call optimization for indirect | |
2131 | sibcall because the pop behavior in epilogue may pollute the | |
2132 | content of caller-saved regsiter when the register is used for | |
2133 | indirect sibcall. */ | |
2134 | return (!TARGET_V3PUSH | |
2135 | && (cfun->machine->va_args_size == 0) | |
2136 | && decl); | |
2137 | } | |
2138 | ||
2139 | /* Determine whether we need to enable warning for function return check. */ | |
2140 | static bool | |
2141 | nds32_warn_func_return (tree decl) | |
2142 | { | |
2143 | /* Naked functions are implemented entirely in assembly, including the | |
2144 | return sequence, so suppress warnings about this. */ | |
2145 | return !nds32_naked_function_p (decl); | |
2146 | } | |
2147 | ||
2148 | \f | |
2149 | /* Implementing the Varargs Macros. */ | |
2150 | ||
2151 | static void | |
2152 | nds32_setup_incoming_varargs (cumulative_args_t ca, | |
2153 | machine_mode mode, | |
2154 | tree type, | |
2155 | int *pretend_args_size, | |
2156 | int second_time ATTRIBUTE_UNUSED) | |
2157 | { | |
2158 | unsigned int total_args_regs; | |
2159 | unsigned int num_of_used_regs; | |
2160 | unsigned int remaining_reg_count; | |
2161 | CUMULATIVE_ARGS *cum; | |
2162 | ||
2163 | /* If we are under hard float abi, we do not need to set *pretend_args_size. | |
2164 | So that all nameless arguments are pushed by caller and all situation | |
2165 | can be handled by GCC itself. */ | |
2166 | if (TARGET_HARD_FLOAT) | |
2167 | return; | |
2168 | ||
2169 | /* We are using NDS32_MAX_GPR_REGS_FOR_ARGS registers, | |
2170 | counting from NDS32_GPR_ARG_FIRST_REGNUM, for saving incoming arguments. | |
2171 | However, for nameless(anonymous) arguments, we should push them on the | |
2172 | stack so that all the nameless arguments appear to have been passed | |
2173 | consecutively in the memory for accessing. Hence, we need to check and | |
2174 | exclude the registers that are used for named arguments. */ | |
2175 | ||
2176 | cum = get_cumulative_args (ca); | |
2177 | ||
2178 | /* The MODE and TYPE describe the last argument. | |
2179 | We need those information to determine the remaining registers | |
2180 | for varargs. */ | |
2181 | total_args_regs | |
2182 | = NDS32_MAX_GPR_REGS_FOR_ARGS + NDS32_GPR_ARG_FIRST_REGNUM; | |
2183 | num_of_used_regs | |
2184 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
2185 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
2186 | ||
2187 | remaining_reg_count = total_args_regs - num_of_used_regs; | |
2188 | *pretend_args_size = remaining_reg_count * UNITS_PER_WORD; | |
2189 | ||
2190 | return; | |
2191 | } | |
2192 | ||
2193 | static bool | |
2194 | nds32_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED) | |
2195 | { | |
2196 | /* If this hook returns true, the named argument of FUNCTION_ARG is always | |
2197 | true for named arguments, and false for unnamed arguments. */ | |
2198 | return true; | |
2199 | } | |
2200 | ||
2201 | \f | |
2202 | /* Trampolines for Nested Functions. */ | |
2203 | ||
2204 | static void | |
2205 | nds32_asm_trampoline_template (FILE *f) | |
2206 | { | |
2207 | if (TARGET_REDUCED_REGS) | |
2208 | { | |
2209 | /* Trampoline is not supported on reduced-set registers yet. */ | |
2210 | sorry ("a nested function is not supported for reduced registers"); | |
2211 | } | |
2212 | else | |
2213 | { | |
2214 | asm_fprintf (f, "\t! Trampoline code template\n"); | |
2215 | asm_fprintf (f, "\t! This code fragment will be copied " | |
2216 | "into stack on demand\n"); | |
2217 | ||
2218 | asm_fprintf (f, "\tmfusr\t$r16,$pc\n"); | |
2219 | asm_fprintf (f, "\tlwi\t$r15,[$r16 + 20] " | |
2220 | "! load nested function address\n"); | |
2221 | asm_fprintf (f, "\tlwi\t$r16,[$r16 + 16] " | |
2222 | "! load chain_value\n"); | |
2223 | asm_fprintf (f, "\tjr\t$r15\n"); | |
2224 | } | |
2225 | ||
2226 | /* Preserve space ($pc + 16) for saving chain_value, | |
2227 | nds32_trampoline_init will fill the value in this slot. */ | |
2228 | asm_fprintf (f, "\t! space for saving chain_value\n"); | |
2229 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
2230 | ||
2231 | /* Preserve space ($pc + 20) for saving nested function address, | |
2232 | nds32_trampoline_init will fill the value in this slot. */ | |
2233 | asm_fprintf (f, "\t! space for saving nested function address\n"); | |
2234 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
2235 | } | |
2236 | ||
2237 | /* Emit RTL insns to initialize the variable parts of a trampoline. */ | |
2238 | static void | |
2239 | nds32_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
2240 | { | |
2241 | int i; | |
2242 | ||
2243 | /* Nested function address. */ | |
2244 | rtx fnaddr; | |
2245 | /* The memory rtx that is going to | |
2246 | be filled with chain_value. */ | |
2247 | rtx chain_value_mem; | |
2248 | /* The memory rtx that is going to | |
2249 | be filled with nested function address. */ | |
2250 | rtx nested_func_mem; | |
2251 | ||
2252 | /* Start address of trampoline code in stack, for doing cache sync. */ | |
2253 | rtx sync_cache_addr; | |
2254 | /* Temporary register for sync instruction. */ | |
2255 | rtx tmp_reg; | |
2256 | /* Instruction-cache sync instruction, | |
2257 | requesting an argument as starting address. */ | |
2258 | rtx isync_insn; | |
2259 | /* For convenience reason of doing comparison. */ | |
2260 | int tramp_align_in_bytes; | |
2261 | ||
2262 | /* Trampoline is not supported on reduced-set registers yet. */ | |
2263 | if (TARGET_REDUCED_REGS) | |
2264 | sorry ("a nested function is not supported for reduced registers"); | |
2265 | ||
2266 | /* STEP 1: Copy trampoline code template into stack, | |
2267 | fill up essential data into stack. */ | |
2268 | ||
2269 | /* Extract nested function address rtx. */ | |
2270 | fnaddr = XEXP (DECL_RTL (fndecl), 0); | |
2271 | ||
2272 | /* m_tramp is memory rtx that is going to be filled with trampoline code. | |
2273 | We have nds32_asm_trampoline_template() to emit template pattern. */ | |
2274 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
2275 | GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL); | |
2276 | ||
2277 | /* After copying trampoline code into stack, | |
2278 | fill chain_value into stack. */ | |
2279 | chain_value_mem = adjust_address (m_tramp, SImode, 16); | |
2280 | emit_move_insn (chain_value_mem, chain_value); | |
2281 | /* After copying trampoline code int stack, | |
2282 | fill nested function address into stack. */ | |
2283 | nested_func_mem = adjust_address (m_tramp, SImode, 20); | |
2284 | emit_move_insn (nested_func_mem, fnaddr); | |
2285 | ||
2286 | /* STEP 2: Sync instruction-cache. */ | |
2287 | ||
2288 | /* We have successfully filled trampoline code into stack. | |
2289 | However, in order to execute code in stack correctly, | |
2290 | we must sync instruction cache. */ | |
2291 | sync_cache_addr = XEXP (m_tramp, 0); | |
2292 | tmp_reg = gen_reg_rtx (SImode); | |
2293 | isync_insn = gen_unspec_volatile_isync (tmp_reg); | |
2294 | ||
2295 | /* Because nds32_cache_block_size is in bytes, | |
2296 | we get trampoline alignment in bytes for convenient comparison. */ | |
2297 | tramp_align_in_bytes = TRAMPOLINE_ALIGNMENT / BITS_PER_UNIT; | |
2298 | ||
2299 | if (tramp_align_in_bytes >= nds32_cache_block_size | |
2300 | && (tramp_align_in_bytes % nds32_cache_block_size) == 0) | |
2301 | { | |
2302 | /* Under this condition, the starting address of trampoline | |
2303 | must be aligned to the starting address of each cache block | |
2304 | and we do not have to worry about cross-boundary issue. */ | |
2305 | for (i = 0; | |
2306 | i < (TRAMPOLINE_SIZE + nds32_cache_block_size - 1) | |
2307 | / nds32_cache_block_size; | |
2308 | i++) | |
2309 | { | |
2310 | emit_move_insn (tmp_reg, | |
2311 | plus_constant (Pmode, sync_cache_addr, | |
2312 | nds32_cache_block_size * i)); | |
2313 | emit_insn (isync_insn); | |
2314 | } | |
2315 | } | |
2316 | else if (TRAMPOLINE_SIZE > nds32_cache_block_size) | |
2317 | { | |
2318 | /* The starting address of trampoline code | |
2319 | may not be aligned to the cache block, | |
2320 | so the trampoline code may be across two cache block. | |
2321 | We need to sync the last element, which is 4-byte size, | |
2322 | of trampoline template. */ | |
2323 | for (i = 0; | |
2324 | i < (TRAMPOLINE_SIZE + nds32_cache_block_size - 1) | |
2325 | / nds32_cache_block_size; | |
2326 | i++) | |
2327 | { | |
2328 | emit_move_insn (tmp_reg, | |
2329 | plus_constant (Pmode, sync_cache_addr, | |
2330 | nds32_cache_block_size * i)); | |
2331 | emit_insn (isync_insn); | |
2332 | } | |
2333 | ||
2334 | /* The last element of trampoline template is 4-byte size. */ | |
2335 | emit_move_insn (tmp_reg, | |
2336 | plus_constant (Pmode, sync_cache_addr, | |
2337 | TRAMPOLINE_SIZE - 4)); | |
2338 | emit_insn (isync_insn); | |
2339 | } | |
2340 | else | |
2341 | { | |
2342 | /* This is the simplest case. | |
2343 | Because TRAMPOLINE_SIZE is less than or | |
2344 | equal to nds32_cache_block_size, | |
2345 | we can just sync start address and | |
2346 | the last element of trampoline code. */ | |
2347 | ||
2348 | /* Sync starting address of tampoline code. */ | |
2349 | emit_move_insn (tmp_reg, sync_cache_addr); | |
2350 | emit_insn (isync_insn); | |
2351 | /* Sync the last element, which is 4-byte size, | |
2352 | of trampoline template. */ | |
2353 | emit_move_insn (tmp_reg, | |
2354 | plus_constant (Pmode, sync_cache_addr, | |
2355 | TRAMPOLINE_SIZE - 4)); | |
2356 | emit_insn (isync_insn); | |
2357 | } | |
2358 | ||
2359 | /* Set instruction serialization barrier | |
2360 | to guarantee the correct operations. */ | |
2361 | emit_insn (gen_unspec_volatile_isb ()); | |
2362 | } | |
2363 | ||
2364 | \f | |
2365 | /* Addressing Modes. */ | |
2366 | ||
2367 | static bool | |
2368 | nds32_legitimate_address_p (machine_mode mode, rtx x, bool strict) | |
2369 | { | |
2370 | if (TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE) | |
2371 | { | |
2372 | /* When using floating-point instructions, | |
2373 | we don't allow 'addr' to be [symbol_ref], [CONST] pattern. */ | |
2374 | if ((mode == DFmode || mode == SFmode) | |
2375 | && (GET_CODE (x) == SYMBOL_REF | |
2376 | || GET_CODE(x) == CONST)) | |
2377 | return false; | |
2378 | ||
2379 | /* Allow [post_modify] addressing mode, when using FPU instructions. */ | |
2380 | if (GET_CODE (x) == POST_MODIFY | |
2381 | && mode == DFmode) | |
2382 | { | |
2383 | if (GET_CODE (XEXP (x, 0)) == REG | |
2384 | && GET_CODE (XEXP (x, 1)) == PLUS) | |
2385 | { | |
2386 | rtx plus_op = XEXP (x, 1); | |
2387 | rtx op0 = XEXP (plus_op, 0); | |
2388 | rtx op1 = XEXP (plus_op, 1); | |
2389 | ||
2390 | if (nds32_address_register_rtx_p (op0, strict) | |
2391 | && CONST_INT_P (op1)) | |
2392 | { | |
2393 | if (satisfies_constraint_Is14 (op1)) | |
2394 | { | |
2395 | /* Make sure address is word alignment. | |
2396 | Currently we do not have 64-bit load/store yet, | |
2397 | so we will use two 32-bit load/store instructions to do | |
2398 | memory access and they are single word alignment. */ | |
2399 | if (NDS32_SINGLE_WORD_ALIGN_P (INTVAL (op1))) | |
2400 | return true; | |
2401 | } | |
2402 | } | |
2403 | } | |
2404 | } | |
2405 | } | |
2406 | ||
2407 | /* For (mem:DI addr) or (mem:DF addr) case, | |
2408 | we only allow 'addr' to be [reg], [symbol_ref], | |
2409 | [const], or [reg + const_int] pattern. */ | |
2410 | if (mode == DImode || mode == DFmode) | |
2411 | { | |
2412 | /* Allow [Reg + const_int] addressing mode. */ | |
2413 | if (GET_CODE (x) == PLUS) | |
2414 | { | |
2415 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict) | |
2416 | && nds32_legitimate_index_p (mode, XEXP (x, 1), strict) | |
2417 | && CONST_INT_P (XEXP (x, 1))) | |
2418 | return true; | |
2419 | else if (nds32_address_register_rtx_p (XEXP (x, 1), strict) | |
2420 | && nds32_legitimate_index_p (mode, XEXP (x, 0), strict) | |
2421 | && CONST_INT_P (XEXP (x, 0))) | |
2422 | return true; | |
2423 | } | |
2424 | ||
2425 | /* Allow [post_inc] and [post_dec] addressing mode. */ | |
2426 | if (GET_CODE (x) == POST_INC || GET_CODE (x) == POST_DEC) | |
2427 | { | |
2428 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict)) | |
2429 | return true; | |
2430 | } | |
2431 | ||
2432 | /* Now check [reg], [symbol_ref], and [const]. */ | |
2433 | if (GET_CODE (x) != REG | |
2434 | && GET_CODE (x) != SYMBOL_REF | |
2435 | && GET_CODE (x) != CONST) | |
2436 | return false; | |
2437 | } | |
2438 | ||
2439 | /* Check if 'x' is a valid address. */ | |
2440 | switch (GET_CODE (x)) | |
2441 | { | |
2442 | case REG: | |
2443 | /* (mem (reg A)) => [Ra] */ | |
2444 | return nds32_address_register_rtx_p (x, strict); | |
2445 | ||
2446 | case SYMBOL_REF: | |
2447 | /* (mem (symbol_ref A)) => [symbol_ref] */ | |
2448 | /* If -mcmodel=large, the 'symbol_ref' is not a valid address | |
2449 | during or after LRA/reload phase. */ | |
2450 | if (TARGET_CMODEL_LARGE | |
2451 | && (reload_completed | |
2452 | || reload_in_progress | |
2453 | || lra_in_progress)) | |
2454 | return false; | |
2455 | /* If -mcmodel=medium and the symbol references to rodata section, | |
2456 | the 'symbol_ref' is not a valid address during or after | |
2457 | LRA/reload phase. */ | |
2458 | if (TARGET_CMODEL_MEDIUM | |
2459 | && NDS32_SYMBOL_REF_RODATA_P (x) | |
2460 | && (reload_completed | |
2461 | || reload_in_progress | |
2462 | || lra_in_progress)) | |
2463 | return false; | |
2464 | ||
2465 | return true; | |
2466 | ||
2467 | case CONST: | |
2468 | /* (mem (const (...))) | |
2469 | => [ + const_addr ], where const_addr = symbol_ref + const_int */ | |
2470 | if (GET_CODE (XEXP (x, 0)) == PLUS) | |
2471 | { | |
2472 | rtx plus_op = XEXP (x, 0); | |
2473 | ||
2474 | rtx op0 = XEXP (plus_op, 0); | |
2475 | rtx op1 = XEXP (plus_op, 1); | |
2476 | ||
2477 | if (GET_CODE (op0) == SYMBOL_REF && CONST_INT_P (op1)) | |
2478 | { | |
2479 | /* Now we see the [ + const_addr ] pattern, but we need | |
2480 | some further checking. */ | |
2481 | /* If -mcmodel=large, the 'const_addr' is not a valid address | |
2482 | during or after LRA/reload phase. */ | |
2483 | if (TARGET_CMODEL_LARGE | |
2484 | && (reload_completed | |
2485 | || reload_in_progress | |
2486 | || lra_in_progress)) | |
2487 | return false; | |
2488 | /* If -mcmodel=medium and the symbol references to rodata section, | |
2489 | the 'const_addr' is not a valid address during or after | |
2490 | LRA/reload phase. */ | |
2491 | if (TARGET_CMODEL_MEDIUM | |
2492 | && NDS32_SYMBOL_REF_RODATA_P (op0) | |
2493 | && (reload_completed | |
2494 | || reload_in_progress | |
2495 | || lra_in_progress)) | |
2496 | return false; | |
2497 | ||
2498 | /* At this point we can make sure 'const_addr' is a | |
2499 | valid address. */ | |
2500 | return true; | |
2501 | } | |
2502 | } | |
2503 | ||
2504 | return false; | |
2505 | ||
2506 | case POST_MODIFY: | |
2507 | /* (mem (post_modify (reg) (plus (reg) (reg)))) | |
2508 | => [Ra], Rb */ | |
2509 | /* (mem (post_modify (reg) (plus (reg) (const_int)))) | |
2510 | => [Ra], const_int */ | |
2511 | if (GET_CODE (XEXP (x, 0)) == REG | |
2512 | && GET_CODE (XEXP (x, 1)) == PLUS) | |
2513 | { | |
2514 | rtx plus_op = XEXP (x, 1); | |
2515 | ||
2516 | rtx op0 = XEXP (plus_op, 0); | |
2517 | rtx op1 = XEXP (plus_op, 1); | |
2518 | ||
2519 | if (nds32_address_register_rtx_p (op0, strict) | |
2520 | && nds32_legitimate_index_p (mode, op1, strict)) | |
2521 | return true; | |
2522 | else | |
2523 | return false; | |
2524 | } | |
2525 | ||
2526 | return false; | |
2527 | ||
2528 | case POST_INC: | |
2529 | case POST_DEC: | |
2530 | /* (mem (post_inc reg)) => [Ra], 1/2/4 */ | |
2531 | /* (mem (post_dec reg)) => [Ra], -1/-2/-4 */ | |
2532 | /* The 1/2/4 or -1/-2/-4 have been displayed in nds32.md. | |
2533 | We only need to deal with register Ra. */ | |
2534 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict)) | |
2535 | return true; | |
2536 | else | |
2537 | return false; | |
2538 | ||
2539 | case PLUS: | |
2540 | /* (mem (plus reg const_int)) | |
2541 | => [Ra + imm] */ | |
2542 | /* (mem (plus reg reg)) | |
2543 | => [Ra + Rb] */ | |
2544 | /* (mem (plus (mult reg const_int) reg)) | |
2545 | => [Ra + Rb << sv] */ | |
2546 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict) | |
2547 | && nds32_legitimate_index_p (mode, XEXP (x, 1), strict)) | |
2548 | return true; | |
2549 | else if (nds32_address_register_rtx_p (XEXP (x, 1), strict) | |
2550 | && nds32_legitimate_index_p (mode, XEXP (x, 0), strict)) | |
2551 | return true; | |
2552 | else | |
2553 | return false; | |
2554 | ||
2555 | case LO_SUM: | |
2556 | /* (mem (lo_sum (reg) (symbol_ref))) */ | |
2557 | /* (mem (lo_sum (reg) (const))) */ | |
2558 | gcc_assert (REG_P (XEXP (x, 0))); | |
2559 | if (GET_CODE (XEXP (x, 1)) == SYMBOL_REF | |
2560 | || GET_CODE (XEXP (x, 1)) == CONST) | |
2561 | return nds32_legitimate_address_p (mode, XEXP (x, 1), strict); | |
2562 | else | |
2563 | return false; | |
2564 | ||
2565 | default: | |
2566 | return false; | |
2567 | } | |
2568 | } | |
2569 | ||
2570 | \f | |
2571 | /* Condition Code Status. */ | |
2572 | ||
2573 | /* -- Representation of condition codes using registers. */ | |
2574 | ||
2575 | static void | |
2576 | nds32_canonicalize_comparison (int *code, | |
2577 | rtx *op0 ATTRIBUTE_UNUSED, | |
2578 | rtx *op1, | |
2579 | bool op0_preserve_value ATTRIBUTE_UNUSED) | |
2580 | { | |
2581 | /* When the instruction combination pass tries to combine a comparison insn | |
2582 | with its previous insns, it also transforms the operator in order to | |
2583 | minimize its constant field. For example, it tries to transform a | |
2584 | comparison insn from | |
2585 | (set (reg:SI 54) | |
2586 | (ltu:SI (reg:SI 52) | |
2587 | (const_int 10 [0xa]))) | |
2588 | to | |
2589 | (set (reg:SI 54) | |
2590 | (leu:SI (reg:SI 52) | |
2591 | (const_int 9 [0x9]))) | |
2592 | ||
2593 | However, the nds32 target only provides instructions supporting the LTU | |
2594 | operation directly, and the implementation of the pattern "cbranchsi4" | |
2595 | only expands the LTU form. In order to handle the non-LTU operations | |
2596 | generated from passes other than the RTL expansion pass, we have to | |
2597 | implement this hook to revert those changes. Since we only expand the LTU | |
2598 | operator in the RTL expansion pass, we might only need to handle the LEU | |
2599 | case, unless we find other optimization passes perform more aggressive | |
2600 | transformations. */ | |
2601 | ||
2602 | if (*code == LEU && CONST_INT_P (*op1)) | |
2603 | { | |
2604 | *op1 = gen_int_mode (INTVAL (*op1) + 1, SImode); | |
2605 | *code = LTU; | |
2606 | } | |
2607 | } | |
2608 | ||
2609 | \f | |
2610 | /* Describing Relative Costs of Operations. */ | |
2611 | ||
2612 | static int | |
2613 | nds32_register_move_cost (machine_mode mode ATTRIBUTE_UNUSED, | |
2614 | reg_class_t from, | |
2615 | reg_class_t to) | |
2616 | { | |
2617 | if ((from == FP_REGS && to != FP_REGS) | |
2618 | || (from != FP_REGS && to == FP_REGS)) | |
2619 | return 9; | |
2620 | else if (from == HIGH_REGS || to == HIGH_REGS) | |
2621 | return optimize_size ? 6 : 2; | |
2622 | else | |
2623 | return 2; | |
2624 | } | |
2625 | ||
2626 | static int | |
2627 | nds32_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED, | |
2628 | reg_class_t rclass ATTRIBUTE_UNUSED, | |
2629 | bool in ATTRIBUTE_UNUSED) | |
2630 | { | |
2631 | return 8; | |
2632 | } | |
2633 | ||
2634 | /* This target hook describes the relative costs of RTL expressions. | |
2635 | Return 'true' when all subexpressions of x have been processed. | |
2636 | Return 'false' to sum the costs of sub-rtx, plus cost of this operation. | |
2637 | Refer to gcc/rtlanal.c for more information. */ | |
2638 | static bool | |
2639 | nds32_rtx_costs (rtx x, | |
2640 | machine_mode mode, | |
2641 | int outer_code, | |
2642 | int opno, | |
2643 | int *total, | |
2644 | bool speed) | |
2645 | { | |
2646 | return nds32_rtx_costs_impl (x, mode, outer_code, opno, total, speed); | |
2647 | } | |
2648 | ||
2649 | static int | |
2650 | nds32_address_cost (rtx address, | |
2651 | machine_mode mode, | |
2652 | addr_space_t as, | |
2653 | bool speed) | |
2654 | { | |
2655 | return nds32_address_cost_impl (address, mode, as, speed); | |
2656 | } | |
2657 | ||
2658 | \f | |
2659 | /* Dividing the Output into Sections (Texts, Data, . . . ). */ | |
2660 | ||
2661 | /* If references to a symbol or a constant must be treated differently | |
2662 | depending on something about the variable or function named by the symbol | |
2663 | (such as what section it is in), we use this hook to store flags | |
2664 | in symbol_ref rtx. */ | |
2665 | static void | |
2666 | nds32_encode_section_info (tree decl, rtx rtl, int new_decl_p) | |
2667 | { | |
2668 | default_encode_section_info (decl, rtl, new_decl_p); | |
2669 | ||
2670 | /* For the memory rtx, if it references to rodata section, we can store | |
2671 | NDS32_SYMBOL_FLAG_RODATA flag into symbol_ref rtx so that the | |
2672 | nds32_legitimate_address_p() can determine how to treat such symbol_ref | |
2673 | based on -mcmodel=X and this information. */ | |
2674 | if (MEM_P (rtl) && MEM_READONLY_P (rtl)) | |
2675 | { | |
2676 | rtx addr = XEXP (rtl, 0); | |
2677 | ||
2678 | if (GET_CODE (addr) == SYMBOL_REF) | |
2679 | { | |
2680 | /* For (mem (symbol_ref X)) case. */ | |
2681 | SYMBOL_REF_FLAGS (addr) |= NDS32_SYMBOL_FLAG_RODATA; | |
2682 | } | |
2683 | else if (GET_CODE (addr) == CONST | |
2684 | && GET_CODE (XEXP (addr, 0)) == PLUS) | |
2685 | { | |
2686 | /* For (mem (const (plus (symbol_ref X) (const_int N)))) case. */ | |
2687 | rtx plus_op = XEXP (addr, 0); | |
2688 | rtx op0 = XEXP (plus_op, 0); | |
2689 | rtx op1 = XEXP (plus_op, 1); | |
2690 | ||
2691 | if (GET_CODE (op0) == SYMBOL_REF && CONST_INT_P (op1)) | |
2692 | SYMBOL_REF_FLAGS (op0) |= NDS32_SYMBOL_FLAG_RODATA; | |
2693 | } | |
2694 | } | |
2695 | } | |
2696 | ||
2697 | \f | |
2698 | /* Defining the Output Assembler Language. */ | |
2699 | ||
2700 | /* -- The Overall Framework of an Assembler File. */ | |
2701 | ||
2702 | static void | |
2703 | nds32_asm_file_start (void) | |
2704 | { | |
2705 | default_file_start (); | |
2706 | ||
2707 | /* Tell assembler which ABI we are using. */ | |
2708 | fprintf (asm_out_file, "\t! ABI version\n"); | |
2709 | if (TARGET_HARD_FLOAT) | |
2710 | fprintf (asm_out_file, "\t.abi_2fp_plus\n"); | |
2711 | else | |
2712 | fprintf (asm_out_file, "\t.abi_2\n"); | |
2713 | ||
2714 | /* Tell assembler that this asm code is generated by compiler. */ | |
2715 | fprintf (asm_out_file, "\t! This asm file is generated by compiler\n"); | |
2716 | fprintf (asm_out_file, "\t.flag\tverbatim\n"); | |
2717 | /* Give assembler the size of each vector for interrupt handler. */ | |
2718 | fprintf (asm_out_file, "\t! This vector size directive is required " | |
2719 | "for checking inconsistency on interrupt handler\n"); | |
2720 | fprintf (asm_out_file, "\t.vec_size\t%d\n", nds32_isr_vector_size); | |
2721 | ||
2722 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2723 | ||
2724 | if (TARGET_ISA_V2) | |
2725 | fprintf (asm_out_file, "\t! ISA family\t\t: %s\n", "V2"); | |
2726 | if (TARGET_ISA_V3) | |
2727 | fprintf (asm_out_file, "\t! ISA family\t\t: %s\n", "V3"); | |
2728 | if (TARGET_ISA_V3M) | |
2729 | fprintf (asm_out_file, "\t! ISA family\t\t: %s\n", "V3M"); | |
2730 | ||
2731 | if (TARGET_CMODEL_SMALL) | |
2732 | fprintf (asm_out_file, "\t! Code model\t\t: %s\n", "SMALL"); | |
2733 | if (TARGET_CMODEL_MEDIUM) | |
2734 | fprintf (asm_out_file, "\t! Code model\t\t: %s\n", "MEDIUM"); | |
2735 | if (TARGET_CMODEL_LARGE) | |
2736 | fprintf (asm_out_file, "\t! Code model\t\t: %s\n", "LARGE"); | |
2737 | ||
2738 | fprintf (asm_out_file, "\t! Endian setting\t: %s\n", | |
2739 | ((TARGET_BIG_ENDIAN) ? "big-endian" | |
2740 | : "little-endian")); | |
2741 | fprintf (asm_out_file, "\t! Use SP floating-point instruction\t: %s\n", | |
2742 | ((TARGET_FPU_SINGLE) ? "Yes" | |
2743 | : "No")); | |
2744 | fprintf (asm_out_file, "\t! Use DP floating-point instruction\t: %s\n", | |
2745 | ((TARGET_FPU_DOUBLE) ? "Yes" | |
2746 | : "No")); | |
2747 | fprintf (asm_out_file, "\t! ABI version\t\t: %s\n", | |
2748 | ((TARGET_HARD_FLOAT) ? "ABI2FP+" | |
2749 | : "ABI2")); | |
2750 | ||
2751 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2752 | ||
2753 | fprintf (asm_out_file, "\t! Use conditional move\t\t: %s\n", | |
2754 | ((TARGET_CMOV) ? "Yes" | |
2755 | : "No")); | |
2756 | fprintf (asm_out_file, "\t! Use performance extension\t: %s\n", | |
2757 | ((TARGET_EXT_PERF) ? "Yes" | |
2758 | : "No")); | |
2759 | fprintf (asm_out_file, "\t! Use performance extension 2\t: %s\n", | |
2760 | ((TARGET_EXT_PERF2) ? "Yes" | |
2761 | : "No")); | |
2762 | fprintf (asm_out_file, "\t! Use string extension\t\t: %s\n", | |
2763 | ((TARGET_EXT_STRING) ? "Yes" | |
2764 | : "No")); | |
2765 | ||
2766 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2767 | ||
2768 | fprintf (asm_out_file, "\t! V3PUSH instructions\t: %s\n", | |
2769 | ((TARGET_V3PUSH) ? "Yes" | |
2770 | : "No")); | |
2771 | fprintf (asm_out_file, "\t! 16-bit instructions\t: %s\n", | |
2772 | ((TARGET_16_BIT) ? "Yes" | |
2773 | : "No")); | |
2774 | fprintf (asm_out_file, "\t! Reduced registers set\t: %s\n", | |
2775 | ((TARGET_REDUCED_REGS) ? "Yes" | |
2776 | : "No")); | |
2777 | ||
2778 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2779 | ||
2780 | if (optimize_size) | |
2781 | fprintf (asm_out_file, "\t! Optimization level\t: -Os\n"); | |
2782 | else | |
2783 | fprintf (asm_out_file, "\t! Optimization level\t: -O%d\n", optimize); | |
2784 | ||
2785 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2786 | ||
2787 | fprintf (asm_out_file, "\t! Cache block size\t: %d\n", | |
2788 | nds32_cache_block_size); | |
2789 | ||
2790 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2791 | ||
2792 | nds32_asm_file_start_for_isr (); | |
2793 | } | |
2794 | ||
2795 | static void | |
2796 | nds32_asm_file_end (void) | |
2797 | { | |
2798 | nds32_asm_file_end_for_isr (); | |
2799 | ||
2800 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2801 | } | |
2802 | ||
2803 | /* -- Output and Generation of Labels. */ | |
2804 | ||
2805 | static void | |
2806 | nds32_asm_globalize_label (FILE *stream, const char *name) | |
2807 | { | |
2808 | fputs ("\t.global\t", stream); | |
2809 | assemble_name (stream, name); | |
2810 | fputs ("\n", stream); | |
2811 | } | |
2812 | ||
2813 | /* -- Output of Assembler Instructions. */ | |
2814 | ||
2815 | static void | |
2816 | nds32_print_operand (FILE *stream, rtx x, int code) | |
2817 | { | |
2818 | HOST_WIDE_INT one_position; | |
2819 | HOST_WIDE_INT zero_position; | |
2820 | bool pick_lsb_p = false; | |
2821 | bool pick_msb_p = false; | |
2822 | int regno; | |
2823 | ||
2824 | int op_value; | |
2825 | ||
2826 | switch (code) | |
2827 | { | |
2828 | case 0 : | |
2829 | /* Do nothing special. */ | |
2830 | break; | |
2831 | ||
2832 | case 'b': | |
2833 | /* Use exact_log2() to search the 0-bit position. */ | |
2834 | gcc_assert (CONST_INT_P (x)); | |
2835 | zero_position = exact_log2 (~UINTVAL (x) & GET_MODE_MASK (SImode)); | |
2836 | gcc_assert (zero_position != -1); | |
2837 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, zero_position); | |
2838 | ||
2839 | /* No need to handle following process, so return immediately. */ | |
2840 | return; | |
2841 | ||
2842 | case 'e': | |
2843 | gcc_assert (MEM_P (x) | |
2844 | && GET_CODE (XEXP (x, 0)) == PLUS | |
2845 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT); | |
2846 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, INTVAL (XEXP (XEXP (x, 0), 1))); | |
2847 | ||
2848 | /* No need to handle following process, so return immediately. */ | |
2849 | return; | |
2850 | case 'B': | |
2851 | /* Use exact_log2() to search the 1-bit position. */ | |
2852 | gcc_assert (CONST_INT_P (x)); | |
2853 | one_position = exact_log2 (UINTVAL (x) & GET_MODE_MASK (SImode)); | |
2854 | gcc_assert (one_position != -1); | |
2855 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, one_position); | |
2856 | ||
2857 | /* No need to handle following process, so return immediately. */ | |
2858 | return; | |
2859 | ||
2860 | case 'L': | |
2861 | /* X is supposed to be REG rtx. */ | |
2862 | gcc_assert (REG_P (x)); | |
2863 | /* Claim that we are going to pick LSB part of X. */ | |
2864 | pick_lsb_p = true; | |
2865 | break; | |
2866 | ||
2867 | case 'H': | |
2868 | /* X is supposed to be REG rtx. */ | |
2869 | gcc_assert (REG_P (x)); | |
2870 | /* Claim that we are going to pick MSB part of X. */ | |
2871 | pick_msb_p = true; | |
2872 | break; | |
2873 | ||
2874 | case 'V': | |
2875 | /* 'x' is supposed to be CONST_INT, get the value. */ | |
2876 | gcc_assert (CONST_INT_P (x)); | |
2877 | op_value = INTVAL (x); | |
2878 | ||
2879 | /* According to the Andes architecture, | |
2880 | the system/user register index range is 0 ~ 1023. | |
2881 | In order to avoid conflict between user-specified-integer value | |
2882 | and enum-specified-register value, | |
2883 | the 'enum nds32_intrinsic_registers' value | |
2884 | in nds32_intrinsic.h starts from 1024. */ | |
2885 | if (op_value < 1024 && op_value >= 0) | |
2886 | { | |
2887 | /* If user gives integer value directly (0~1023), | |
2888 | we just print out the value. */ | |
2889 | fprintf (stream, HOST_WIDE_INT_PRINT_DEC, op_value); | |
2890 | } | |
2891 | else if (op_value < 0 | |
2892 | || op_value >= ((int) ARRAY_SIZE (nds32_intrinsic_register_names) | |
2893 | + 1024)) | |
2894 | { | |
2895 | /* The enum index value for array size is out of range. */ | |
2896 | error ("intrinsic register index is out of range"); | |
2897 | } | |
2898 | else | |
2899 | { | |
2900 | /* If user applies normal way with __NDS32_REG_XXX__ enum data, | |
2901 | we can print out register name. Remember to substract 1024. */ | |
2902 | fprintf (stream, "%s", | |
2903 | nds32_intrinsic_register_names[op_value - 1024]); | |
2904 | } | |
2905 | ||
2906 | /* No need to handle following process, so return immediately. */ | |
2907 | return; | |
2908 | ||
2909 | default : | |
2910 | /* Unknown flag. */ | |
2911 | output_operand_lossage ("invalid operand output code"); | |
2912 | break; | |
2913 | } | |
2914 | ||
2915 | switch (GET_CODE (x)) | |
2916 | { | |
2917 | case LABEL_REF: | |
2918 | case SYMBOL_REF: | |
2919 | output_addr_const (stream, x); | |
2920 | break; | |
2921 | ||
2922 | case REG: | |
2923 | /* Print a Double-precision register name. */ | |
2924 | if ((GET_MODE (x) == DImode || GET_MODE (x) == DFmode) | |
2925 | && NDS32_IS_FPR_REGNUM (REGNO (x))) | |
2926 | { | |
2927 | regno = REGNO (x); | |
2928 | if (!NDS32_FPR_REGNO_OK_FOR_DOUBLE (regno)) | |
2929 | { | |
2930 | output_operand_lossage ("invalid operand for code '%c'", code); | |
2931 | break; | |
2932 | } | |
2933 | fprintf (stream, "$fd%d", (regno - NDS32_FIRST_FPR_REGNUM) >> 1); | |
2934 | break; | |
2935 | } | |
2936 | ||
2937 | /* Print LSB or MSB part of register pair if the | |
2938 | constraint modifier 'L' or 'H' is specified. */ | |
2939 | if ((GET_MODE (x) == DImode || GET_MODE (x) == DFmode) | |
2940 | && NDS32_IS_GPR_REGNUM (REGNO (x))) | |
2941 | { | |
2942 | if ((pick_lsb_p && WORDS_BIG_ENDIAN) | |
2943 | || (pick_msb_p && !WORDS_BIG_ENDIAN)) | |
2944 | { | |
2945 | /* If we would like to print out LSB register under big-endian, | |
2946 | or print out MSB register under little-endian, we need to | |
2947 | increase register number. */ | |
2948 | regno = REGNO (x); | |
2949 | regno++; | |
2950 | fputs (reg_names[regno], stream); | |
2951 | break; | |
2952 | } | |
2953 | } | |
2954 | ||
2955 | /* Forbid using static chain register ($r16) | |
2956 | on reduced-set registers configuration. */ | |
2957 | if (TARGET_REDUCED_REGS | |
2958 | && REGNO (x) == STATIC_CHAIN_REGNUM) | |
2959 | sorry ("a nested function is not supported for reduced registers"); | |
2960 | ||
2961 | /* Normal cases, print out register name. */ | |
2962 | fputs (reg_names[REGNO (x)], stream); | |
2963 | break; | |
2964 | ||
2965 | case MEM: | |
2966 | output_address (GET_MODE (x), XEXP (x, 0)); | |
2967 | break; | |
2968 | ||
2969 | case HIGH: | |
2970 | if (GET_CODE (XEXP (x, 0)) == CONST_DOUBLE) | |
2971 | { | |
2972 | const REAL_VALUE_TYPE *rv; | |
2973 | long val; | |
2974 | gcc_assert (GET_MODE (x) == SFmode); | |
2975 | ||
2976 | rv = CONST_DOUBLE_REAL_VALUE (XEXP (x, 0)); | |
2977 | REAL_VALUE_TO_TARGET_SINGLE (*rv, val); | |
2978 | ||
2979 | fprintf (stream, "hi20(0x%lx)", val); | |
2980 | } | |
2981 | else | |
2982 | gcc_unreachable (); | |
2983 | break; | |
2984 | ||
2985 | case CONST_DOUBLE: | |
2986 | const REAL_VALUE_TYPE *rv; | |
2987 | long val; | |
2988 | gcc_assert (GET_MODE (x) == SFmode); | |
2989 | ||
2990 | rv = CONST_DOUBLE_REAL_VALUE (x); | |
2991 | REAL_VALUE_TO_TARGET_SINGLE (*rv, val); | |
2992 | ||
2993 | fprintf (stream, "0x%lx", val); | |
2994 | break; | |
2995 | ||
2996 | case CODE_LABEL: | |
2997 | case CONST_INT: | |
2998 | case CONST: | |
2999 | output_addr_const (stream, x); | |
3000 | break; | |
3001 | ||
3002 | default: | |
3003 | /* Generally, output_addr_const () is able to handle most cases. | |
3004 | We want to see what CODE could appear, | |
3005 | so we use gcc_unreachable() to stop it. */ | |
3006 | debug_rtx (x); | |
3007 | gcc_unreachable (); | |
3008 | break; | |
3009 | } | |
3010 | } | |
3011 | ||
3012 | static void | |
3013 | nds32_print_operand_address (FILE *stream, machine_mode /*mode*/, rtx x) | |
3014 | { | |
3015 | rtx op0, op1; | |
3016 | ||
3017 | switch (GET_CODE (x)) | |
3018 | { | |
3019 | case SYMBOL_REF: | |
3020 | case CONST: | |
3021 | /* [ + symbol_ref] */ | |
3022 | /* [ + const_addr], where const_addr = symbol_ref + const_int */ | |
3023 | fputs ("[ + ", stream); | |
3024 | output_addr_const (stream, x); | |
3025 | fputs ("]", stream); | |
3026 | break; | |
3027 | ||
3028 | case REG: | |
3029 | /* Forbid using static chain register ($r16) | |
3030 | on reduced-set registers configuration. */ | |
3031 | if (TARGET_REDUCED_REGS | |
3032 | && REGNO (x) == STATIC_CHAIN_REGNUM) | |
3033 | sorry ("a nested function is not supported for reduced registers"); | |
3034 | ||
3035 | /* [Ra] */ | |
3036 | fprintf (stream, "[%s]", reg_names[REGNO (x)]); | |
3037 | break; | |
3038 | ||
3039 | case PLUS: | |
3040 | op0 = XEXP (x, 0); | |
3041 | op1 = XEXP (x, 1); | |
3042 | ||
3043 | /* Checking op0, forbid using static chain register ($r16) | |
3044 | on reduced-set registers configuration. */ | |
3045 | if (TARGET_REDUCED_REGS | |
3046 | && REG_P (op0) | |
3047 | && REGNO (op0) == STATIC_CHAIN_REGNUM) | |
3048 | sorry ("a nested function is not supported for reduced registers"); | |
3049 | /* Checking op1, forbid using static chain register ($r16) | |
3050 | on reduced-set registers configuration. */ | |
3051 | if (TARGET_REDUCED_REGS | |
3052 | && REG_P (op1) | |
3053 | && REGNO (op1) == STATIC_CHAIN_REGNUM) | |
3054 | sorry ("a nested function is not supported for reduced registers"); | |
3055 | ||
3056 | if (REG_P (op0) && CONST_INT_P (op1)) | |
3057 | { | |
3058 | /* [Ra + imm] */ | |
3059 | fprintf (stream, "[%s + (" HOST_WIDE_INT_PRINT_DEC ")]", | |
3060 | reg_names[REGNO (op0)], INTVAL (op1)); | |
3061 | } | |
3062 | else if (REG_P (op0) && REG_P (op1)) | |
3063 | { | |
3064 | /* [Ra + Rb] */ | |
3065 | fprintf (stream, "[%s + %s]", | |
3066 | reg_names[REGNO (op0)], reg_names[REGNO (op1)]); | |
3067 | } | |
3068 | else if (GET_CODE (op0) == MULT && REG_P (op1)) | |
3069 | { | |
3070 | /* [Ra + Rb << sv] | |
3071 | From observation, the pattern looks like: | |
3072 | (plus:SI (mult:SI (reg:SI 58) | |
3073 | (const_int 4 [0x4])) | |
3074 | (reg/f:SI 57)) */ | |
3075 | int sv; | |
3076 | ||
3077 | /* We need to set sv to output shift value. */ | |
3078 | if (INTVAL (XEXP (op0, 1)) == 1) | |
3079 | sv = 0; | |
3080 | else if (INTVAL (XEXP (op0, 1)) == 2) | |
3081 | sv = 1; | |
3082 | else if (INTVAL (XEXP (op0, 1)) == 4) | |
3083 | sv = 2; | |
3084 | else if (INTVAL (XEXP (op0, 1)) == 8) | |
3085 | sv = 3; | |
3086 | else | |
3087 | gcc_unreachable (); | |
3088 | ||
3089 | fprintf (stream, "[%s + %s << %d]", | |
3090 | reg_names[REGNO (op1)], | |
3091 | reg_names[REGNO (XEXP (op0, 0))], | |
3092 | sv); | |
3093 | } | |
3094 | else | |
3095 | { | |
3096 | /* The control flow is not supposed to be here. */ | |
3097 | debug_rtx (x); | |
3098 | gcc_unreachable (); | |
3099 | } | |
3100 | ||
3101 | break; | |
3102 | ||
3103 | case POST_MODIFY: | |
3104 | /* (post_modify (regA) (plus (regA) (regB))) | |
3105 | (post_modify (regA) (plus (regA) (const_int))) | |
3106 | We would like to extract | |
3107 | regA and regB (or const_int) from plus rtx. */ | |
3108 | op0 = XEXP (XEXP (x, 1), 0); | |
3109 | op1 = XEXP (XEXP (x, 1), 1); | |
3110 | ||
3111 | /* Checking op0, forbid using static chain register ($r16) | |
3112 | on reduced-set registers configuration. */ | |
3113 | if (TARGET_REDUCED_REGS | |
3114 | && REG_P (op0) | |
3115 | && REGNO (op0) == STATIC_CHAIN_REGNUM) | |
3116 | sorry ("a nested function is not supported for reduced registers"); | |
3117 | /* Checking op1, forbid using static chain register ($r16) | |
3118 | on reduced-set registers configuration. */ | |
3119 | if (TARGET_REDUCED_REGS | |
3120 | && REG_P (op1) | |
3121 | && REGNO (op1) == STATIC_CHAIN_REGNUM) | |
3122 | sorry ("a nested function is not supported for reduced registers"); | |
3123 | ||
3124 | if (REG_P (op0) && REG_P (op1)) | |
3125 | { | |
3126 | /* [Ra], Rb */ | |
3127 | fprintf (stream, "[%s], %s", | |
3128 | reg_names[REGNO (op0)], reg_names[REGNO (op1)]); | |
3129 | } | |
3130 | else if (REG_P (op0) && CONST_INT_P (op1)) | |
3131 | { | |
3132 | /* [Ra], imm */ | |
3133 | fprintf (stream, "[%s], " HOST_WIDE_INT_PRINT_DEC, | |
3134 | reg_names[REGNO (op0)], INTVAL (op1)); | |
3135 | } | |
3136 | else | |
3137 | { | |
3138 | /* The control flow is not supposed to be here. */ | |
3139 | debug_rtx (x); | |
3140 | gcc_unreachable (); | |
3141 | } | |
3142 | ||
3143 | break; | |
3144 | ||
3145 | case POST_INC: | |
3146 | case POST_DEC: | |
3147 | op0 = XEXP (x, 0); | |
3148 | ||
3149 | /* Checking op0, forbid using static chain register ($r16) | |
3150 | on reduced-set registers configuration. */ | |
3151 | if (TARGET_REDUCED_REGS | |
3152 | && REG_P (op0) | |
3153 | && REGNO (op0) == STATIC_CHAIN_REGNUM) | |
3154 | sorry ("a nested function is not supported for reduced registers"); | |
3155 | ||
3156 | if (REG_P (op0)) | |
3157 | { | |
3158 | /* "[Ra], 1/2/4" or "[Ra], -1/-2/-4" | |
3159 | The 1/2/4 or -1/-2/-4 have been displayed in nds32.md. | |
3160 | We only need to deal with register Ra. */ | |
3161 | fprintf (stream, "[%s]", reg_names[REGNO (op0)]); | |
3162 | } | |
3163 | else | |
3164 | { | |
3165 | /* The control flow is not supposed to be here. */ | |
3166 | debug_rtx (x); | |
3167 | gcc_unreachable (); | |
3168 | } | |
3169 | ||
3170 | break; | |
3171 | ||
3172 | default : | |
3173 | /* Generally, output_addr_const () is able to handle most cases. | |
3174 | We want to see what CODE could appear, | |
3175 | so we use gcc_unreachable() to stop it. */ | |
3176 | debug_rtx (x); | |
3177 | gcc_unreachable (); | |
3178 | break; | |
3179 | } | |
3180 | } | |
3181 | ||
3182 | /* -- Assembler Commands for Exception Regions. */ | |
3183 | ||
3184 | static rtx | |
3185 | nds32_dwarf_register_span (rtx reg) | |
3186 | { | |
3187 | rtx dwarf_high, dwarf_low; | |
3188 | rtx dwarf_single; | |
3189 | machine_mode mode; | |
3190 | int regno; | |
3191 | ||
3192 | mode = GET_MODE (reg); | |
3193 | regno = REGNO (reg); | |
3194 | ||
3195 | /* We need to adjust dwarf register information for floating-point registers | |
3196 | rather than using default register number mapping. */ | |
3197 | if (regno >= NDS32_FIRST_FPR_REGNUM | |
3198 | && regno <= NDS32_LAST_FPR_REGNUM) | |
3199 | { | |
3200 | if (mode == DFmode || mode == SCmode) | |
3201 | { | |
3202 | /* By default, GCC maps increasing register numbers to increasing | |
3203 | memory locations, but paired FPRs in NDS32 target are always | |
3204 | big-endian, i.e.: | |
3205 | ||
3206 | fd0 : fs0 fs1 | |
3207 | (MSB) (LSB) | |
3208 | ||
3209 | We must return parallel rtx to represent such layout. */ | |
3210 | dwarf_high = gen_rtx_REG (word_mode, regno); | |
3211 | dwarf_low = gen_rtx_REG (word_mode, regno + 1); | |
3212 | return gen_rtx_PARALLEL (VOIDmode, | |
3213 | gen_rtvec (2, dwarf_low, dwarf_high)); | |
3214 | } | |
3215 | else if (mode == DCmode) | |
3216 | { | |
3217 | rtx dwarf_high_re = gen_rtx_REG (word_mode, regno); | |
3218 | rtx dwarf_low_re = gen_rtx_REG (word_mode, regno + 1); | |
3219 | rtx dwarf_high_im = gen_rtx_REG (word_mode, regno); | |
3220 | rtx dwarf_low_im = gen_rtx_REG (word_mode, regno + 1); | |
3221 | return gen_rtx_PARALLEL (VOIDmode, | |
3222 | gen_rtvec (4, dwarf_low_re, dwarf_high_re, | |
3223 | dwarf_high_im, dwarf_low_im)); | |
3224 | } | |
3225 | else if (mode == SFmode || mode == SImode) | |
3226 | { | |
3227 | /* Create new dwarf information with adjusted register number. */ | |
3228 | dwarf_single = gen_rtx_REG (word_mode, regno); | |
3229 | return gen_rtx_PARALLEL (VOIDmode, gen_rtvec (1, dwarf_single)); | |
3230 | } | |
3231 | else | |
3232 | { | |
3233 | /* We should not be here. */ | |
3234 | gcc_unreachable (); | |
3235 | } | |
3236 | } | |
3237 | ||
3238 | return NULL_RTX; | |
3239 | } | |
3240 | ||
3241 | /* Map internal gcc register numbers to DWARF2 register numbers. */ | |
3242 | ||
3243 | unsigned int | |
3244 | nds32_dbx_register_number (unsigned int regno) | |
3245 | { | |
3246 | /* The nds32 port in GDB maintains a mapping between dwarf register | |
3247 | number and displayed register name. For backward compatibility to | |
3248 | previous toolchain, currently our gdb still has four registers | |
3249 | (d0.l, d0.h, d1.l, and d1.h) between GPR and FPR while compiler | |
3250 | does not count those four registers in its register number table. | |
3251 | So we have to add 4 on its register number and then create new | |
3252 | dwarf information. Hopefully we can discard such workaround | |
3253 | in the future. */ | |
3254 | if (NDS32_IS_FPR_REGNUM (regno)) | |
3255 | return regno + 4; | |
3256 | ||
3257 | return regno; | |
3258 | } | |
3259 | ||
3260 | \f | |
3261 | /* Defining target-specific uses of __attribute__. */ | |
3262 | ||
3263 | /* Add some checking after merging attributes. */ | |
3264 | static tree | |
3265 | nds32_merge_decl_attributes (tree olddecl, tree newdecl) | |
3266 | { | |
3267 | tree combined_attrs; | |
3268 | ||
3269 | /* Create combined attributes. */ | |
3270 | combined_attrs = merge_attributes (DECL_ATTRIBUTES (olddecl), | |
3271 | DECL_ATTRIBUTES (newdecl)); | |
3272 | ||
3273 | /* Since newdecl is acutally a duplicate of olddecl, | |
3274 | we can take olddecl for some operations. */ | |
3275 | if (TREE_CODE (olddecl) == FUNCTION_DECL) | |
3276 | { | |
3277 | /* Check isr-specific attributes conflict. */ | |
3278 | nds32_check_isr_attrs_conflict (olddecl, combined_attrs); | |
3279 | } | |
3280 | ||
3281 | return combined_attrs; | |
3282 | } | |
3283 | ||
3284 | /* Add some checking when inserting attributes. */ | |
3285 | static void | |
3286 | nds32_insert_attributes (tree decl, tree *attributes) | |
3287 | { | |
3288 | /* For function declaration, we need to check isr-specific attributes: | |
3289 | 1. Call nds32_check_isr_attrs_conflict() to check any conflict. | |
3290 | 2. Check valid integer value for interrupt/exception. | |
3291 | 3. Check valid integer value for reset. | |
3292 | 4. Check valid function for nmi/warm. */ | |
3293 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
3294 | { | |
3295 | tree func_attrs; | |
3296 | tree intr, excp, reset; | |
3297 | ||
3298 | /* Pick up function attributes. */ | |
3299 | func_attrs = *attributes; | |
3300 | ||
3301 | /* 1. Call nds32_check_isr_attrs_conflict() to check any conflict. */ | |
3302 | nds32_check_isr_attrs_conflict (decl, func_attrs); | |
3303 | ||
3304 | /* Now we are starting to check valid id value | |
3305 | for interrupt/exception/reset. | |
3306 | Note that we ONLY check its validity here. | |
3307 | To construct isr vector information, it is still performed | |
3308 | by nds32_construct_isr_vectors_information(). */ | |
3309 | intr = lookup_attribute ("interrupt", func_attrs); | |
3310 | excp = lookup_attribute ("exception", func_attrs); | |
3311 | reset = lookup_attribute ("reset", func_attrs); | |
3312 | ||
3313 | if (intr || excp) | |
3314 | { | |
3315 | /* Deal with interrupt/exception. */ | |
3316 | tree id_list; | |
3317 | unsigned int lower_bound, upper_bound; | |
3318 | ||
3319 | /* The way to handle interrupt or exception is the same, | |
3320 | we just need to take care of actual vector number. | |
3321 | For interrupt(0..63), the actual vector number is (9..72). | |
3322 | For exception(1..8), the actual vector number is (1..8). */ | |
3323 | lower_bound = (intr) ? (0) : (1); | |
3324 | upper_bound = (intr) ? (63) : (8); | |
3325 | ||
3326 | /* Prepare id list so that we can traverse id value. */ | |
3327 | id_list = (intr) ? (TREE_VALUE (intr)) : (TREE_VALUE (excp)); | |
3328 | ||
3329 | /* 2. Check valid integer value for interrupt/exception. */ | |
3330 | while (id_list) | |
3331 | { | |
3332 | tree id; | |
3333 | ||
3334 | /* Pick up each vector id value. */ | |
3335 | id = TREE_VALUE (id_list); | |
3336 | /* Issue error if it is not a valid integer value. */ | |
3337 | if (TREE_CODE (id) != INTEGER_CST | |
3338 | || wi::ltu_p (wi::to_wide (id), lower_bound) | |
3339 | || wi::gtu_p (wi::to_wide (id), upper_bound)) | |
3340 | error ("invalid id value for interrupt/exception attribute"); | |
3341 | ||
3342 | /* Advance to next id. */ | |
3343 | id_list = TREE_CHAIN (id_list); | |
3344 | } | |
3345 | } | |
3346 | else if (reset) | |
3347 | { | |
3348 | /* Deal with reset. */ | |
3349 | tree id_list; | |
3350 | tree id; | |
3351 | tree nmi, warm; | |
3352 | unsigned int lower_bound; | |
3353 | unsigned int upper_bound; | |
3354 | ||
3355 | /* Prepare id_list and identify id value so that | |
3356 | we can check if total number of vectors is valid. */ | |
3357 | id_list = TREE_VALUE (reset); | |
3358 | id = TREE_VALUE (id_list); | |
3359 | ||
3360 | /* The maximum numbers for user's interrupt is 64. */ | |
3361 | lower_bound = 0; | |
3362 | upper_bound = 64; | |
3363 | ||
3364 | /* 3. Check valid integer value for reset. */ | |
3365 | if (TREE_CODE (id) != INTEGER_CST | |
3366 | || wi::ltu_p (wi::to_wide (id), lower_bound) | |
3367 | || wi::gtu_p (wi::to_wide (id), upper_bound)) | |
3368 | error ("invalid id value for reset attribute"); | |
3369 | ||
3370 | /* 4. Check valid function for nmi/warm. */ | |
3371 | nmi = lookup_attribute ("nmi", func_attrs); | |
3372 | warm = lookup_attribute ("warm", func_attrs); | |
3373 | ||
3374 | if (nmi != NULL_TREE) | |
3375 | { | |
3376 | tree nmi_func_list; | |
3377 | tree nmi_func; | |
3378 | ||
3379 | nmi_func_list = TREE_VALUE (nmi); | |
3380 | nmi_func = TREE_VALUE (nmi_func_list); | |
3381 | ||
3382 | /* Issue error if it is not a valid nmi function. */ | |
3383 | if (TREE_CODE (nmi_func) != IDENTIFIER_NODE) | |
3384 | error ("invalid nmi function for reset attribute"); | |
3385 | } | |
3386 | ||
3387 | if (warm != NULL_TREE) | |
3388 | { | |
3389 | tree warm_func_list; | |
3390 | tree warm_func; | |
3391 | ||
3392 | warm_func_list = TREE_VALUE (warm); | |
3393 | warm_func = TREE_VALUE (warm_func_list); | |
3394 | ||
3395 | /* Issue error if it is not a valid warm function. */ | |
3396 | if (TREE_CODE (warm_func) != IDENTIFIER_NODE) | |
3397 | error ("invalid warm function for reset attribute"); | |
3398 | } | |
3399 | } | |
3400 | else | |
3401 | { | |
3402 | /* No interrupt, exception, or reset attribute is set. */ | |
3403 | return; | |
3404 | } | |
3405 | } | |
3406 | } | |
3407 | ||
3408 | static bool | |
3409 | nds32_option_pragma_parse (tree args ATTRIBUTE_UNUSED, | |
3410 | tree pop_target ATTRIBUTE_UNUSED) | |
3411 | { | |
3412 | /* Currently, we do not parse any pragma target by ourself, | |
3413 | so just simply return false. */ | |
3414 | return false; | |
3415 | } | |
3416 | ||
3417 | static void | |
3418 | nds32_option_override (void) | |
3419 | { | |
3420 | /* After all the command options have been parsed, | |
3421 | we shall deal with some flags for changing compiler settings. */ | |
3422 | ||
3423 | /* At first, we check if we have to strictly | |
3424 | set some flags based on ISA family. */ | |
3425 | if (TARGET_ISA_V2) | |
3426 | { | |
3427 | /* Under V2 ISA, we need to strictly disable TARGET_V3PUSH. */ | |
3428 | target_flags &= ~MASK_V3PUSH; | |
3429 | } | |
3430 | if (TARGET_ISA_V3) | |
3431 | { | |
3432 | /* Under V3 ISA, currently nothing should be strictly set. */ | |
3433 | } | |
3434 | if (TARGET_ISA_V3M) | |
3435 | { | |
3436 | /* Under V3M ISA, we need to strictly enable TARGET_REDUCED_REGS. */ | |
3437 | target_flags |= MASK_REDUCED_REGS; | |
3438 | /* Under V3M ISA, we need to strictly disable TARGET_EXT_PERF. */ | |
3439 | target_flags &= ~MASK_EXT_PERF; | |
3440 | /* Under V3M ISA, we need to strictly disable TARGET_EXT_PERF2. */ | |
3441 | target_flags &= ~MASK_EXT_PERF2; | |
3442 | /* Under V3M ISA, we need to strictly disable TARGET_EXT_STRING. */ | |
3443 | target_flags &= ~MASK_EXT_STRING; | |
3444 | } | |
3445 | ||
3446 | /* See if we are using reduced-set registers: | |
3447 | $r0~$r5, $r6~$r10, $r15, $r28, $r29, $r30, $r31 | |
3448 | If so, we must forbid using $r11~$r14, $r16~$r27. */ | |
3449 | if (TARGET_REDUCED_REGS) | |
3450 | { | |
3451 | int r; | |
3452 | ||
3453 | /* Prevent register allocator from | |
3454 | choosing it as doing register allocation. */ | |
3455 | for (r = 11; r <= 14; r++) | |
3456 | fixed_regs[r] = call_used_regs[r] = 1; | |
3457 | for (r = 16; r <= 27; r++) | |
3458 | fixed_regs[r] = call_used_regs[r] = 1; | |
3459 | } | |
3460 | ||
3461 | if (!TARGET_16_BIT) | |
3462 | { | |
3463 | /* Under no 16 bit ISA, we need to strictly disable TARGET_V3PUSH. */ | |
3464 | target_flags &= ~MASK_V3PUSH; | |
3465 | } | |
3466 | ||
3467 | if (TARGET_HARD_FLOAT && !(TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE)) | |
3468 | { | |
3469 | if (nds32_arch_option == ARCH_V3S || nds32_arch_option == ARCH_V3F) | |
3470 | error ("Disable FPU ISA, " | |
3471 | "the ABI option must be enable '-mfloat-abi=soft'"); | |
3472 | else | |
3473 | error ("'-mabi=2fp+' option only support when FPU available, " | |
3474 | "must be enable '-mext-fpu-sp' or '-mext-fpu-dp'"); | |
3475 | } | |
3476 | ||
3477 | /* Currently, we don't support PIC code generation yet. */ | |
3478 | if (flag_pic) | |
3479 | sorry ("position-independent code not supported"); | |
3480 | ||
3481 | nds32_register_passes (); | |
3482 | } | |
3483 | ||
3484 | \f | |
3485 | /* Miscellaneous Parameters. */ | |
3486 | ||
3487 | static rtx_insn * | |
3488 | nds32_md_asm_adjust (vec<rtx> &outputs ATTRIBUTE_UNUSED, | |
3489 | vec<rtx> &inputs ATTRIBUTE_UNUSED, | |
3490 | vec<const char *> &constraints ATTRIBUTE_UNUSED, | |
3491 | vec<rtx> &clobbers, HARD_REG_SET &clobbered_regs) | |
3492 | { | |
3493 | clobbers.safe_push (gen_rtx_REG (SImode, TA_REGNUM)); | |
3494 | SET_HARD_REG_BIT (clobbered_regs, TA_REGNUM); | |
3495 | return NULL; | |
3496 | } | |
3497 | ||
3498 | static void | |
3499 | nds32_init_builtins (void) | |
3500 | { | |
3501 | nds32_init_builtins_impl (); | |
3502 | } | |
3503 | ||
3504 | static tree | |
3505 | nds32_builtin_decl (unsigned code, bool initialize_p) | |
3506 | { | |
3507 | /* Implement in nds32-intrinsic.c. */ | |
3508 | return nds32_builtin_decl_impl (code, initialize_p); | |
3509 | } | |
3510 | ||
3511 | static rtx | |
3512 | nds32_expand_builtin (tree exp, | |
3513 | rtx target, | |
3514 | rtx subtarget, | |
3515 | machine_mode mode, | |
3516 | int ignore) | |
3517 | { | |
3518 | return nds32_expand_builtin_impl (exp, target, subtarget, mode, ignore); | |
3519 | } | |
3520 | ||
3521 | ||
3522 | /* ------------------------------------------------------------------------ */ | |
3523 | ||
3524 | /* PART 4: Implemet extern function definitions, | |
3525 | the prototype is in nds32-protos.h. */ | |
3526 | \f | |
3527 | /* Run-time Target Specification. */ | |
3528 | ||
3529 | void | |
3530 | nds32_cpu_cpp_builtins(struct cpp_reader *pfile) | |
3531 | { | |
3532 | #define builtin_define(TXT) cpp_define (pfile, TXT) | |
3533 | #define builtin_assert(TXT) cpp_assert (pfile, TXT) | |
3534 | builtin_define ("__nds32__"); | |
3535 | builtin_define ("__NDS32__"); | |
3536 | ||
3537 | if (TARGET_HARD_FLOAT) | |
3538 | builtin_define ("__NDS32_ABI_2FP_PLUS__"); | |
3539 | else | |
3540 | builtin_define ("__NDS32_ABI_2__"); | |
3541 | ||
3542 | if (TARGET_ISA_V2) | |
3543 | builtin_define ("__NDS32_ISA_V2__"); | |
3544 | if (TARGET_ISA_V3) | |
3545 | builtin_define ("__NDS32_ISA_V3__"); | |
3546 | if (TARGET_ISA_V3M) | |
3547 | builtin_define ("__NDS32_ISA_V3M__"); | |
3548 | ||
3549 | if (TARGET_FPU_SINGLE) | |
3550 | builtin_define ("__NDS32_EXT_FPU_SP__"); | |
3551 | if (TARGET_FPU_DOUBLE) | |
3552 | builtin_define ("__NDS32_EXT_FPU_DP__"); | |
3553 | ||
3554 | if (TARGET_EXT_FPU_FMA) | |
3555 | builtin_define ("__NDS32_EXT_FPU_FMA__"); | |
3556 | if (NDS32_EXT_FPU_DOT_E) | |
3557 | builtin_define ("__NDS32_EXT_FPU_DOT_E__"); | |
3558 | if (TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE) | |
3559 | { | |
3560 | switch (nds32_fp_regnum) | |
3561 | { | |
3562 | case 0: | |
3563 | case 4: | |
3564 | builtin_define ("__NDS32_EXT_FPU_CONFIG_0__"); | |
3565 | break; | |
3566 | case 1: | |
3567 | case 5: | |
3568 | builtin_define ("__NDS32_EXT_FPU_CONFIG_1__"); | |
3569 | break; | |
3570 | case 2: | |
3571 | case 6: | |
3572 | builtin_define ("__NDS32_EXT_FPU_CONFIG_2__"); | |
3573 | break; | |
3574 | case 3: | |
3575 | case 7: | |
3576 | builtin_define ("__NDS32_EXT_FPU_CONFIG_3__"); | |
3577 | break; | |
3578 | default: | |
3579 | abort (); | |
3580 | } | |
3581 | } | |
3582 | ||
3583 | if (TARGET_BIG_ENDIAN) | |
3584 | builtin_define ("__NDS32_EB__"); | |
3585 | else | |
3586 | builtin_define ("__NDS32_EL__"); | |
3587 | ||
3588 | if (TARGET_REDUCED_REGS) | |
3589 | builtin_define ("__NDS32_REDUCED_REGS__"); | |
3590 | if (TARGET_CMOV) | |
3591 | builtin_define ("__NDS32_CMOV__"); | |
3592 | if (TARGET_EXT_PERF) | |
3593 | builtin_define ("__NDS32_EXT_PERF__"); | |
3594 | if (TARGET_EXT_PERF2) | |
3595 | builtin_define ("__NDS32_EXT_PERF2__"); | |
3596 | if (TARGET_EXT_STRING) | |
3597 | builtin_define ("__NDS32_EXT_STRING__"); | |
3598 | if (TARGET_16_BIT) | |
3599 | builtin_define ("__NDS32_16_BIT__"); | |
3600 | if (TARGET_GP_DIRECT) | |
3601 | builtin_define ("__NDS32_GP_DIRECT__"); | |
3602 | if (TARGET_VH) | |
3603 | builtin_define ("__NDS32_VH__"); | |
3604 | ||
3605 | if (TARGET_BIG_ENDIAN) | |
3606 | builtin_define ("__big_endian__"); | |
3607 | ||
3608 | builtin_assert ("cpu=nds32"); | |
3609 | builtin_assert ("machine=nds32"); | |
3610 | ||
3611 | if (TARGET_HARD_FLOAT) | |
3612 | builtin_define ("__NDS32_ABI_2FP_PLUS"); | |
3613 | else | |
3614 | builtin_define ("__NDS32_ABI_2"); | |
3615 | ||
3616 | #undef builtin_define | |
3617 | #undef builtin_assert | |
3618 | } | |
3619 | ||
3620 | \f | |
3621 | /* Defining Data Structures for Per-function Information. */ | |
3622 | ||
3623 | void | |
3624 | nds32_init_expanders (void) | |
3625 | { | |
3626 | /* Arrange to initialize and mark the machine per-function status. */ | |
3627 | init_machine_status = nds32_init_machine_status; | |
3628 | } | |
3629 | ||
3630 | \f | |
3631 | /* Register Usage. */ | |
3632 | ||
3633 | /* -- Order of Allocation of Registers. */ | |
3634 | ||
3635 | void | |
3636 | nds32_adjust_reg_alloc_order (void) | |
3637 | { | |
3638 | const int nds32_reg_alloc_order[] = REG_ALLOC_ORDER; | |
3639 | ||
3640 | /* Copy the default register allocation order, which is designed | |
3641 | to optimize for code size. */ | |
3642 | memcpy(reg_alloc_order, nds32_reg_alloc_order, sizeof (reg_alloc_order)); | |
3643 | ||
3644 | /* Adjust few register allocation order when optimizing for speed. */ | |
3645 | if (!optimize_size) | |
3646 | { | |
3647 | memcpy (reg_alloc_order, nds32_reg_alloc_order_for_speed, | |
3648 | sizeof (nds32_reg_alloc_order_for_speed)); | |
3649 | } | |
3650 | } | |
3651 | ||
3652 | /* -- How Values Fit in Registers. */ | |
3653 | ||
3654 | static unsigned | |
3655 | nds32_hard_regno_nregs (unsigned regno ATTRIBUTE_UNUSED, | |
3656 | machine_mode mode) | |
3657 | { | |
3658 | return ((GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD); | |
3659 | } | |
3660 | ||
3661 | /* Implement TARGET_HARD_REGNO_MODE_OK. */ | |
3662 | ||
3663 | static bool | |
3664 | nds32_hard_regno_mode_ok (unsigned int regno, machine_mode mode) | |
3665 | { | |
3666 | if (regno > FIRST_PSEUDO_REGISTER) | |
3667 | return true; | |
3668 | ||
3669 | if ((TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE) && NDS32_IS_FPR_REGNUM (regno)) | |
3670 | { | |
3671 | if (NDS32_IS_EXT_FPR_REGNUM(regno)) | |
3672 | return (NDS32_FPR_REGNO_OK_FOR_DOUBLE(regno) && (mode == DFmode)); | |
3673 | else if (mode == SFmode || mode == SImode) | |
3674 | return NDS32_FPR_REGNO_OK_FOR_SINGLE (regno); | |
3675 | else if (mode == DFmode) | |
3676 | return NDS32_FPR_REGNO_OK_FOR_DOUBLE (regno); | |
3677 | ||
3678 | return false; | |
3679 | } | |
3680 | ||
3681 | /* Restrict double-word quantities to even register pairs. */ | |
3682 | if (regno <= NDS32_LAST_GPR_REGNUM) | |
3683 | return (targetm.hard_regno_nregs (regno, mode) == 1 | |
3684 | || !((regno) & 1)); | |
3685 | ||
3686 | return false; | |
3687 | } | |
3688 | ||
3689 | #undef TARGET_HARD_REGNO_MODE_OK | |
3690 | #define TARGET_HARD_REGNO_MODE_OK nds32_hard_regno_mode_ok | |
3691 | ||
3692 | /* Implement TARGET_MODES_TIEABLE_P. We can use general registers to | |
3693 | tie QI/HI/SI modes together. */ | |
3694 | ||
3695 | static bool | |
3696 | nds32_modes_tieable_p (machine_mode mode1, machine_mode mode2) | |
3697 | { | |
3698 | if ((GET_MODE_CLASS (mode1) == MODE_INT | |
3699 | && GET_MODE_CLASS (mode2) == MODE_INT) | |
3700 | && GET_MODE_SIZE (mode1) <= UNITS_PER_WORD | |
3701 | && GET_MODE_SIZE (mode2) <= UNITS_PER_WORD) | |
3702 | return true; | |
3703 | ||
3704 | if (GET_MODE_SIZE (mode1) == GET_MODE_SIZE (mode2)) | |
3705 | { | |
3706 | if ((TARGET_FPU_SINGLE && !TARGET_FPU_DOUBLE) | |
3707 | && (mode1 == DFmode || mode2 == DFmode)) | |
3708 | return false; | |
3709 | else | |
3710 | return true; | |
3711 | } | |
3712 | ||
3713 | return false; | |
3714 | } | |
3715 | ||
3716 | #undef TARGET_MODES_TIEABLE_P | |
3717 | #define TARGET_MODES_TIEABLE_P nds32_modes_tieable_p | |
3718 | \f | |
3719 | /* Register Classes. */ | |
3720 | ||
3721 | enum reg_class | |
3722 | nds32_regno_reg_class (int regno) | |
3723 | { | |
3724 | /* Refer to nds32.h for more register class details. */ | |
3725 | ||
3726 | if (regno >= 0 && regno <= 7) | |
3727 | return LOW_REGS; | |
3728 | else if (regno >= 8 && regno <= 11) | |
3729 | return MIDDLE_REGS; | |
3730 | else if (regno >= 12 && regno <= 14) | |
3731 | return HIGH_REGS; | |
3732 | else if (regno == 15) | |
3733 | return R15_TA_REG; | |
3734 | else if (regno >= 16 && regno <= 19) | |
3735 | return MIDDLE_REGS; | |
3736 | else if (regno >= 20 && regno <= 31) | |
3737 | return HIGH_REGS; | |
3738 | else if (regno == 32 || regno == 33) | |
3739 | { | |
3740 | /* $SFP and $AP is FRAME_REGS in fact, However prevent IRA don't | |
3741 | know how to allocate register for $SFP and $AP, just tell IRA they | |
3742 | are GENERAL_REGS, and ARM do this hack too. */ | |
3743 | return GENERAL_REGS; | |
3744 | } | |
3745 | else if (regno >= 34 && regno <= 97) | |
3746 | return FP_REGS; | |
3747 | else | |
3748 | return NO_REGS; | |
3749 | } | |
3750 | ||
3751 | \f | |
3752 | /* Stack Layout and Calling Conventions. */ | |
3753 | ||
3754 | /* -- Basic Stack Layout. */ | |
3755 | ||
3756 | rtx | |
3757 | nds32_return_addr_rtx (int count, | |
3758 | rtx frameaddr ATTRIBUTE_UNUSED) | |
3759 | { | |
3760 | /* There is no way to determine the return address | |
3761 | if frameaddr is the frame that has 'count' steps | |
3762 | up from current frame. */ | |
3763 | if (count != 0) | |
3764 | return NULL_RTX; | |
3765 | ||
3766 | /* If count == 0, it means we are at current frame, | |
3767 | the return address is $r30 ($lp). */ | |
3768 | return get_hard_reg_initial_val (Pmode, LP_REGNUM); | |
3769 | } | |
3770 | ||
3771 | /* -- Eliminating Frame Pointer and Arg Pointer. */ | |
3772 | ||
3773 | HOST_WIDE_INT | |
3774 | nds32_initial_elimination_offset (unsigned int from_reg, unsigned int to_reg) | |
3775 | { | |
3776 | HOST_WIDE_INT offset; | |
3777 | ||
3778 | /* Compute and setup stack frame size. | |
3779 | The result will be in cfun->machine. */ | |
3780 | nds32_compute_stack_frame (); | |
3781 | ||
3782 | /* Remember to consider | |
3783 | cfun->machine->callee_saved_area_gpr_padding_bytes | |
3784 | when calculating offset. */ | |
3785 | if (from_reg == ARG_POINTER_REGNUM && to_reg == STACK_POINTER_REGNUM) | |
3786 | { | |
3787 | offset = (cfun->machine->fp_size | |
3788 | + cfun->machine->gp_size | |
3789 | + cfun->machine->lp_size | |
3790 | + cfun->machine->callee_saved_gpr_regs_size | |
3791 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
3792 | + cfun->machine->callee_saved_fpr_regs_size | |
3793 | + cfun->machine->local_size | |
3794 | + cfun->machine->out_args_size); | |
3795 | } | |
3796 | else if (from_reg == ARG_POINTER_REGNUM | |
3797 | && to_reg == HARD_FRAME_POINTER_REGNUM) | |
3798 | { | |
3799 | offset = 0; | |
3800 | } | |
3801 | else if (from_reg == FRAME_POINTER_REGNUM | |
3802 | && to_reg == STACK_POINTER_REGNUM) | |
3803 | { | |
3804 | offset = (cfun->machine->local_size + cfun->machine->out_args_size); | |
3805 | } | |
3806 | else if (from_reg == FRAME_POINTER_REGNUM | |
3807 | && to_reg == HARD_FRAME_POINTER_REGNUM) | |
3808 | { | |
3809 | offset = (-1) * (cfun->machine->fp_size | |
3810 | + cfun->machine->gp_size | |
3811 | + cfun->machine->lp_size | |
3812 | + cfun->machine->callee_saved_gpr_regs_size | |
3813 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
3814 | + cfun->machine->callee_saved_fpr_regs_size); | |
3815 | } | |
3816 | else | |
3817 | { | |
3818 | gcc_unreachable (); | |
3819 | } | |
3820 | ||
3821 | return offset; | |
3822 | } | |
3823 | ||
3824 | /* -- Passing Arguments in Registers. */ | |
3825 | ||
3826 | void | |
3827 | nds32_init_cumulative_args (CUMULATIVE_ARGS *cum, | |
3828 | tree fntype ATTRIBUTE_UNUSED, | |
3829 | rtx libname ATTRIBUTE_UNUSED, | |
3830 | tree fndecl ATTRIBUTE_UNUSED, | |
3831 | int n_named_args ATTRIBUTE_UNUSED) | |
3832 | { | |
3833 | /* Initial available registers. The values are offset against | |
3834 | NDS32_GPR_ARG_FIRST_REGNUM and NDS32_FPR_ARG_FIRST_REGNUM | |
3835 | for passing arguments. */ | |
3836 | cum->gpr_offset = 0; | |
3837 | cum->fpr_offset = 0; | |
3838 | } | |
3839 | ||
3840 | /* -- Function Entry and Exit. */ | |
3841 | ||
3842 | /* Function for normal multiple push prologue. */ | |
3843 | void | |
3844 | nds32_expand_prologue (void) | |
3845 | { | |
3846 | int fp_adjust; | |
3847 | int sp_adjust; | |
3848 | unsigned Rb, Re; | |
3849 | ||
3850 | /* Compute and setup stack frame size. | |
3851 | The result will be in cfun->machine. */ | |
3852 | nds32_compute_stack_frame (); | |
3853 | ||
3854 | /* If this is a variadic function, first we need to push argument | |
3855 | registers that hold the unnamed argument value. */ | |
3856 | if (cfun->machine->va_args_size != 0) | |
3857 | { | |
3858 | Rb = cfun->machine->va_args_first_regno; | |
3859 | Re = cfun->machine->va_args_last_regno; | |
3860 | /* No need to push $fp, $gp, or $lp. */ | |
3861 | nds32_emit_stack_push_multiple (Rb, Re, false, false, false, true); | |
3862 | ||
3863 | /* We may also need to adjust stack pointer for padding bytes | |
3864 | because varargs may cause $sp not 8-byte aligned. */ | |
3865 | if (cfun->machine->va_args_area_padding_bytes) | |
3866 | { | |
3867 | /* Generate sp adjustment instruction. */ | |
3868 | sp_adjust = cfun->machine->va_args_area_padding_bytes; | |
3869 | ||
3870 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
3871 | stack_pointer_rtx, | |
3872 | -1 * sp_adjust); | |
3873 | } | |
3874 | } | |
3875 | ||
3876 | /* If the function is 'naked', | |
3877 | we do not have to generate prologue code fragment. */ | |
3878 | if (cfun->machine->naked_p) | |
3879 | return; | |
3880 | ||
3881 | /* Get callee_first_regno and callee_last_regno. */ | |
3882 | Rb = cfun->machine->callee_saved_first_gpr_regno; | |
3883 | Re = cfun->machine->callee_saved_last_gpr_regno; | |
3884 | ||
3885 | /* If $fp, $gp, $lp, and all callee-save registers are NOT required | |
3886 | to be saved, we don't have to create multiple push instruction. | |
3887 | Otherwise, a multiple push instruction is needed. */ | |
3888 | if (!(Rb == SP_REGNUM && Re == SP_REGNUM | |
3889 | && cfun->machine->fp_size == 0 | |
3890 | && cfun->machine->gp_size == 0 | |
3891 | && cfun->machine->lp_size == 0)) | |
3892 | { | |
3893 | /* Create multiple push instruction rtx. */ | |
3894 | nds32_emit_stack_push_multiple ( | |
3895 | Rb, Re, | |
3896 | cfun->machine->fp_size, cfun->machine->gp_size, cfun->machine->lp_size, | |
3897 | false); | |
3898 | } | |
3899 | ||
3900 | /* Check frame_pointer_needed to see | |
3901 | if we shall emit fp adjustment instruction. */ | |
3902 | if (frame_pointer_needed) | |
3903 | { | |
3904 | /* adjust $fp = $sp + ($fp size) + ($gp size) + ($lp size) | |
3905 | + (4 * callee-saved-registers) | |
3906 | Note: No need to adjust | |
3907 | cfun->machine->callee_saved_area_gpr_padding_bytes, | |
3908 | because, at this point, stack pointer is just | |
3909 | at the position after push instruction. */ | |
3910 | fp_adjust = cfun->machine->fp_size | |
3911 | + cfun->machine->gp_size | |
3912 | + cfun->machine->lp_size | |
3913 | + cfun->machine->callee_saved_gpr_regs_size; | |
3914 | ||
3915 | nds32_emit_adjust_frame (hard_frame_pointer_rtx, | |
3916 | stack_pointer_rtx, | |
3917 | fp_adjust); | |
3918 | } | |
3919 | ||
3920 | /* Save fpu registers. */ | |
3921 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
3922 | { | |
3923 | /* When $sp moved to bottom of stack, we need to check whether | |
3924 | the range of offset in the FPU instruction. */ | |
3925 | int fpr_offset = cfun->machine->local_size | |
3926 | + cfun->machine->out_args_size | |
3927 | + cfun->machine->callee_saved_fpr_regs_size; | |
3928 | ||
3929 | /* Check FPU instruction offset imm14s. */ | |
3930 | if (!satisfies_constraint_Is14 (GEN_INT (fpr_offset))) | |
3931 | { | |
3932 | int fpr_space = cfun->machine->callee_saved_area_gpr_padding_bytes | |
3933 | + cfun->machine->callee_saved_fpr_regs_size; | |
3934 | ||
3935 | /* Save fpu registers, need to allocate stack space | |
3936 | for fpu callee registers. And now $sp position | |
3937 | on callee saved fpr registers. */ | |
3938 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
3939 | stack_pointer_rtx, | |
3940 | -1 * fpr_space); | |
3941 | ||
3942 | /* Emit fpu store instruction, using [$sp + offset] store | |
3943 | fpu registers. */ | |
3944 | nds32_emit_push_fpr_callee_saved (0); | |
3945 | ||
3946 | /* Adjust $sp = $sp - local_size - out_args_size. */ | |
3947 | sp_adjust = cfun->machine->local_size | |
3948 | + cfun->machine->out_args_size; | |
3949 | ||
3950 | /* Allocate stack space for local size and out args size. */ | |
3951 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
3952 | stack_pointer_rtx, | |
3953 | -1 * sp_adjust); | |
3954 | } | |
3955 | else | |
3956 | { | |
3957 | /* Offset range in Is14, so $sp moved to bottom of stack. */ | |
3958 | ||
3959 | /* Adjust $sp = $sp - local_size - out_args_size | |
3960 | - callee_saved_area_gpr_padding_bytes | |
3961 | - callee_saved_fpr_regs_size. */ | |
3962 | sp_adjust = cfun->machine->local_size | |
3963 | + cfun->machine->out_args_size | |
3964 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
3965 | + cfun->machine->callee_saved_fpr_regs_size; | |
3966 | ||
3967 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
3968 | stack_pointer_rtx, | |
3969 | -1 * sp_adjust); | |
3970 | ||
3971 | /* Emit fpu store instruction, using [$sp + offset] store | |
3972 | fpu registers. */ | |
3973 | int fpr_position = cfun->machine->out_args_size | |
3974 | + cfun->machine->local_size; | |
3975 | nds32_emit_push_fpr_callee_saved (fpr_position); | |
3976 | } | |
3977 | } | |
3978 | else | |
3979 | { | |
3980 | /* Adjust $sp = $sp - local_size - out_args_size | |
3981 | - callee_saved_area_gpr_padding_bytes. */ | |
3982 | sp_adjust = cfun->machine->local_size | |
3983 | + cfun->machine->out_args_size | |
3984 | + cfun->machine->callee_saved_area_gpr_padding_bytes; | |
3985 | ||
3986 | /* sp_adjust value may be out of range of the addi instruction, | |
3987 | create alternative add behavior with TA_REGNUM if necessary, | |
3988 | using NEGATIVE value to tell that we are decreasing address. */ | |
3989 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
3990 | stack_pointer_rtx, | |
3991 | -1 * sp_adjust); | |
3992 | } | |
3993 | ||
3994 | /* Prevent the instruction scheduler from | |
3995 | moving instructions across the boundary. */ | |
3996 | emit_insn (gen_blockage ()); | |
3997 | } | |
3998 | ||
3999 | /* Function for normal multiple pop epilogue. */ | |
4000 | void | |
4001 | nds32_expand_epilogue (bool sibcall_p) | |
4002 | { | |
4003 | int sp_adjust; | |
4004 | unsigned Rb, Re; | |
4005 | ||
4006 | /* Compute and setup stack frame size. | |
4007 | The result will be in cfun->machine. */ | |
4008 | nds32_compute_stack_frame (); | |
4009 | ||
4010 | /* Prevent the instruction scheduler from | |
4011 | moving instructions across the boundary. */ | |
4012 | emit_insn (gen_blockage ()); | |
4013 | ||
4014 | /* If the function is 'naked', we do not have to generate | |
4015 | epilogue code fragment BUT 'ret' instruction. | |
4016 | However, if this function is also a variadic function, | |
4017 | we need to create adjust stack pointer before 'ret' instruction. */ | |
4018 | if (cfun->machine->naked_p) | |
4019 | { | |
4020 | /* If this is a variadic function, we do not have to restore argument | |
4021 | registers but need to adjust stack pointer back to previous stack | |
4022 | frame location before return. */ | |
4023 | if (cfun->machine->va_args_size != 0) | |
4024 | { | |
4025 | /* Generate sp adjustment instruction. | |
4026 | We need to consider padding bytes here. */ | |
4027 | sp_adjust = cfun->machine->va_args_size | |
4028 | + cfun->machine->va_args_area_padding_bytes; | |
4029 | ||
4030 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4031 | stack_pointer_rtx, | |
4032 | sp_adjust); | |
4033 | } | |
4034 | ||
4035 | /* Generate return instruction by using 'return_internal' pattern. | |
4036 | Make sure this instruction is after gen_blockage(). */ | |
4037 | if (!sibcall_p) | |
4038 | emit_jump_insn (gen_return_internal ()); | |
4039 | return; | |
4040 | } | |
4041 | ||
4042 | if (frame_pointer_needed) | |
4043 | { | |
4044 | /* Restore fpu registers. */ | |
4045 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4046 | { | |
4047 | int gpr_padding = cfun->machine->callee_saved_area_gpr_padding_bytes; | |
4048 | ||
4049 | /* adjust $sp = $fp - ($fp size) - ($gp size) - ($lp size) | |
4050 | - (4 * callee-saved-registers) | |
4051 | - (4 * exception-handling-data-registers) | |
4052 | - (4 * callee-saved-gpr-registers padding byte) | |
4053 | - (4 * callee-saved-fpr-registers) | |
4054 | Note: we want to adjust stack pointer | |
4055 | to the position for callee-saved fpr register, | |
4056 | And restore fpu register use .bi instruction to adjust $sp | |
4057 | from callee-saved fpr register to pop instruction. */ | |
4058 | sp_adjust = cfun->machine->fp_size | |
4059 | + cfun->machine->gp_size | |
4060 | + cfun->machine->lp_size | |
4061 | + cfun->machine->callee_saved_gpr_regs_size | |
4062 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
4063 | + cfun->machine->callee_saved_fpr_regs_size; | |
4064 | ||
4065 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4066 | hard_frame_pointer_rtx, | |
4067 | -1 * sp_adjust); | |
4068 | ||
4069 | /* Emit fpu load instruction, using .bi instruction | |
4070 | load fpu registers. */ | |
4071 | nds32_emit_pop_fpr_callee_saved (gpr_padding); | |
4072 | } | |
4073 | else | |
4074 | { | |
4075 | /* adjust $sp = $fp - ($fp size) - ($gp size) - ($lp size) | |
4076 | - (4 * callee-saved-registers) | |
4077 | - (4 * exception-handling-data-registers) | |
4078 | Note: No need to adjust | |
4079 | cfun->machine->callee_saved_area_gpr_padding_bytes, | |
4080 | because we want to adjust stack pointer | |
4081 | to the position for pop instruction. */ | |
4082 | sp_adjust = cfun->machine->fp_size | |
4083 | + cfun->machine->gp_size | |
4084 | + cfun->machine->lp_size | |
4085 | + cfun->machine->callee_saved_gpr_regs_size; | |
4086 | ||
4087 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4088 | hard_frame_pointer_rtx, | |
4089 | -1 * sp_adjust); | |
4090 | } | |
4091 | } | |
4092 | else | |
4093 | { | |
4094 | /* Restore fpu registers. */ | |
4095 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4096 | { | |
4097 | int gpr_padding = cfun->machine->callee_saved_area_gpr_padding_bytes; | |
4098 | ||
4099 | /* Adjust $sp = $sp + local_size + out_args_size. */ | |
4100 | sp_adjust = cfun->machine->local_size | |
4101 | + cfun->machine->out_args_size; | |
4102 | ||
4103 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4104 | stack_pointer_rtx, | |
4105 | sp_adjust); | |
4106 | ||
4107 | /* Emit fpu load instruction, using .bi instruction | |
4108 | load fpu registers, and adjust $sp from callee-saved fpr register | |
4109 | to callee-saved gpr register. */ | |
4110 | nds32_emit_pop_fpr_callee_saved (gpr_padding); | |
4111 | } | |
4112 | else | |
4113 | { | |
4114 | /* If frame pointer is NOT needed, | |
4115 | we cannot calculate the sp adjustment from frame pointer. | |
4116 | Instead, we calculate the adjustment by local_size, | |
4117 | out_args_size, and callee_saved_area_gpr_padding_bytes. | |
4118 | Notice that such sp adjustment value may be out of range, | |
4119 | so we have to deal with it as well. */ | |
4120 | ||
4121 | /* Adjust $sp = $sp + local_size + out_args_size | |
4122 | + callee_saved_area_gpr_padding_bytes. */ | |
4123 | sp_adjust = cfun->machine->local_size | |
4124 | + cfun->machine->out_args_size | |
4125 | + cfun->machine->callee_saved_area_gpr_padding_bytes; | |
4126 | ||
4127 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4128 | stack_pointer_rtx, | |
4129 | sp_adjust); | |
4130 | } | |
4131 | } | |
4132 | ||
4133 | /* Get callee_first_regno and callee_last_regno. */ | |
4134 | Rb = cfun->machine->callee_saved_first_gpr_regno; | |
4135 | Re = cfun->machine->callee_saved_last_gpr_regno; | |
4136 | ||
4137 | /* If $fp, $gp, $lp, and all callee-save registers are NOT required | |
4138 | to be saved, we don't have to create multiple pop instruction. | |
4139 | Otherwise, a multiple pop instruction is needed. */ | |
4140 | if (!(Rb == SP_REGNUM && Re == SP_REGNUM | |
4141 | && cfun->machine->fp_size == 0 | |
4142 | && cfun->machine->gp_size == 0 | |
4143 | && cfun->machine->lp_size == 0)) | |
4144 | { | |
4145 | /* Create multiple pop instruction rtx. */ | |
4146 | nds32_emit_stack_pop_multiple ( | |
4147 | Rb, Re, | |
4148 | cfun->machine->fp_size, cfun->machine->gp_size, cfun->machine->lp_size); | |
4149 | } | |
4150 | ||
4151 | /* If this is a variadic function, we do not have to restore argument | |
4152 | registers but need to adjust stack pointer back to previous stack | |
4153 | frame location before return. */ | |
4154 | if (cfun->machine->va_args_size != 0) | |
4155 | { | |
4156 | /* Generate sp adjustment instruction. | |
4157 | We need to consider padding bytes here. */ | |
4158 | sp_adjust = cfun->machine->va_args_size | |
4159 | + cfun->machine->va_args_area_padding_bytes; | |
4160 | ||
4161 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4162 | stack_pointer_rtx, | |
4163 | sp_adjust); | |
4164 | } | |
4165 | ||
4166 | /* Generate return instruction. */ | |
4167 | if (!sibcall_p) | |
4168 | emit_jump_insn (gen_return_internal ()); | |
4169 | } | |
4170 | ||
4171 | /* Function for v3push prologue. */ | |
4172 | void | |
4173 | nds32_expand_prologue_v3push (void) | |
4174 | { | |
4175 | int fp_adjust; | |
4176 | int sp_adjust; | |
4177 | int fpr_space = 0; | |
4178 | unsigned Rb, Re; | |
4179 | ||
4180 | /* Compute and setup stack frame size. | |
4181 | The result will be in cfun->machine. */ | |
4182 | nds32_compute_stack_frame (); | |
4183 | ||
4184 | if (cfun->machine->callee_saved_gpr_regs_size > 0) | |
4185 | df_set_regs_ever_live (FP_REGNUM, 1); | |
4186 | ||
4187 | /* If the function is 'naked', | |
4188 | we do not have to generate prologue code fragment. */ | |
4189 | if (cfun->machine->naked_p) | |
4190 | return; | |
4191 | ||
4192 | /* Get callee_first_regno and callee_last_regno. */ | |
4193 | Rb = cfun->machine->callee_saved_first_gpr_regno; | |
4194 | Re = cfun->machine->callee_saved_last_gpr_regno; | |
4195 | ||
4196 | /* Calculate sp_adjust first to test if 'push25 Re,imm8u' is available, | |
4197 | where imm8u has to be 8-byte alignment. */ | |
4198 | sp_adjust = cfun->machine->local_size | |
4199 | + cfun->machine->out_args_size | |
4200 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
4201 | + cfun->machine->callee_saved_fpr_regs_size; | |
4202 | ||
4203 | if (satisfies_constraint_Iu08 (GEN_INT (sp_adjust)) | |
4204 | && NDS32_DOUBLE_WORD_ALIGN_P (sp_adjust)) | |
4205 | { | |
4206 | /* We can use 'push25 Re,imm8u'. */ | |
4207 | ||
4208 | /* nds32_emit_stack_v3push(last_regno, sp_adjust), | |
4209 | the pattern 'stack_v3push' is implemented in nds32.md. */ | |
4210 | nds32_emit_stack_v3push (Rb, Re, sp_adjust); | |
4211 | ||
4212 | /* Save fpu registers. */ | |
4213 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4214 | { | |
4215 | /* Calculate fpr position. */ | |
4216 | int fpr_position = cfun->machine->local_size | |
4217 | + cfun->machine->out_args_size; | |
4218 | /* Emit fpu store instruction, using [$sp + offset] store | |
4219 | fpu registers. */ | |
4220 | nds32_emit_push_fpr_callee_saved (fpr_position); | |
4221 | } | |
4222 | ||
4223 | /* Check frame_pointer_needed to see | |
4224 | if we shall emit fp adjustment instruction. */ | |
4225 | if (frame_pointer_needed) | |
4226 | { | |
4227 | /* adjust $fp = $sp + 4 ($fp size) | |
4228 | + 4 ($gp size) | |
4229 | + 4 ($lp size) | |
4230 | + (4 * n) (callee-saved registers) | |
4231 | + sp_adjust ('push25 Re,imm8u') | |
4232 | Note: Since we use 'push25 Re,imm8u', | |
4233 | the position of stack pointer is further | |
4234 | changed after push instruction. | |
4235 | Hence, we need to take sp_adjust value | |
4236 | into consideration. */ | |
4237 | fp_adjust = cfun->machine->fp_size | |
4238 | + cfun->machine->gp_size | |
4239 | + cfun->machine->lp_size | |
4240 | + cfun->machine->callee_saved_gpr_regs_size | |
4241 | + sp_adjust; | |
4242 | ||
4243 | nds32_emit_adjust_frame (hard_frame_pointer_rtx, | |
4244 | stack_pointer_rtx, | |
4245 | fp_adjust); | |
4246 | } | |
4247 | } | |
4248 | else | |
4249 | { | |
4250 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4251 | { | |
4252 | /* Calculate fpr space. */ | |
4253 | fpr_space = cfun->machine->callee_saved_area_gpr_padding_bytes | |
4254 | + cfun->machine->callee_saved_fpr_regs_size; | |
4255 | ||
4256 | /* We have to use 'push25 Re, fpr_space', to pre-allocate | |
4257 | callee saved fpr registers space. */ | |
4258 | nds32_emit_stack_v3push (Rb, Re, fpr_space); | |
4259 | nds32_emit_push_fpr_callee_saved (0); | |
4260 | } | |
4261 | else | |
4262 | { | |
4263 | /* We have to use 'push25 Re,0' and | |
4264 | expand one more instruction to adjust $sp later. */ | |
4265 | ||
4266 | /* nds32_emit_stack_v3push(last_regno, sp_adjust), | |
4267 | the pattern 'stack_v3push' is implemented in nds32.md. */ | |
4268 | nds32_emit_stack_v3push (Rb, Re, 0); | |
4269 | } | |
4270 | ||
4271 | /* Check frame_pointer_needed to see | |
4272 | if we shall emit fp adjustment instruction. */ | |
4273 | if (frame_pointer_needed) | |
4274 | { | |
4275 | /* adjust $fp = $sp + 4 ($fp size) | |
4276 | + 4 ($gp size) | |
4277 | + 4 ($lp size) | |
4278 | + (4 * n) (callee-saved registers) | |
4279 | Note: Since we use 'push25 Re,0', | |
4280 | the stack pointer is just at the position | |
4281 | after push instruction. | |
4282 | No need to take sp_adjust into consideration. */ | |
4283 | fp_adjust = cfun->machine->fp_size | |
4284 | + cfun->machine->gp_size | |
4285 | + cfun->machine->lp_size | |
4286 | + cfun->machine->callee_saved_gpr_regs_size; | |
4287 | ||
4288 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4289 | { | |
4290 | /* We use 'push25 Re, fpr_space', the $sp is | |
4291 | on callee saved fpr position, so need to consider | |
4292 | fpr space. */ | |
4293 | fp_adjust = fp_adjust + fpr_space; | |
4294 | } | |
4295 | ||
4296 | nds32_emit_adjust_frame (hard_frame_pointer_rtx, | |
4297 | stack_pointer_rtx, | |
4298 | fp_adjust); | |
4299 | } | |
4300 | ||
4301 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4302 | { | |
4303 | /* We use 'push25 Re, fpr_space', | |
4304 | the $sp is on callee saved fpr position, | |
4305 | no need to consider fpr space. */ | |
4306 | sp_adjust = sp_adjust - fpr_space; | |
4307 | } | |
4308 | ||
4309 | /* Because we use 'push25 Re,0', | |
4310 | we need to expand one more instruction to adjust $sp. | |
4311 | using NEGATIVE value to tell that we are decreasing address. */ | |
4312 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4313 | stack_pointer_rtx, | |
4314 | -1 * sp_adjust); | |
4315 | } | |
4316 | ||
4317 | /* Prevent the instruction scheduler from | |
4318 | moving instructions across the boundary. */ | |
4319 | emit_insn (gen_blockage ()); | |
4320 | } | |
4321 | ||
4322 | /* Function for v3pop epilogue. */ | |
4323 | void | |
4324 | nds32_expand_epilogue_v3pop (bool sibcall_p) | |
4325 | { | |
4326 | int sp_adjust; | |
4327 | unsigned Rb, Re; | |
4328 | ||
4329 | /* Compute and setup stack frame size. | |
4330 | The result will be in cfun->machine. */ | |
4331 | nds32_compute_stack_frame (); | |
4332 | ||
4333 | /* Prevent the instruction scheduler from | |
4334 | moving instructions across the boundary. */ | |
4335 | emit_insn (gen_blockage ()); | |
4336 | ||
4337 | /* If the function is 'naked', we do not have to generate | |
4338 | epilogue code fragment BUT 'ret' instruction. */ | |
4339 | if (cfun->machine->naked_p) | |
4340 | { | |
4341 | /* Generate return instruction by using 'return_internal' pattern. | |
4342 | Make sure this instruction is after gen_blockage(). */ | |
4343 | if (!sibcall_p) | |
4344 | emit_jump_insn (gen_return_internal ()); | |
4345 | return; | |
4346 | } | |
4347 | ||
4348 | /* Get callee_first_regno and callee_last_regno. */ | |
4349 | Rb = cfun->machine->callee_saved_first_gpr_regno; | |
4350 | Re = cfun->machine->callee_saved_last_gpr_regno; | |
4351 | ||
4352 | /* Calculate sp_adjust first to test if 'pop25 Re,imm8u' is available, | |
4353 | where imm8u has to be 8-byte alignment. */ | |
4354 | sp_adjust = cfun->machine->local_size | |
4355 | + cfun->machine->out_args_size | |
4356 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
4357 | + cfun->machine->callee_saved_fpr_regs_size; | |
4358 | ||
4359 | /* We have to consider alloca issue as well. | |
4360 | If the function does call alloca(), the stack pointer is not fixed. | |
4361 | In that case, we cannot use 'pop25 Re,imm8u' directly. | |
4362 | We have to caculate stack pointer from frame pointer | |
4363 | and then use 'pop25 Re,0'. | |
4364 | Of course, the frame_pointer_needed should be nonzero | |
4365 | if the function calls alloca(). */ | |
4366 | if (satisfies_constraint_Iu08 (GEN_INT (sp_adjust)) | |
4367 | && NDS32_DOUBLE_WORD_ALIGN_P (sp_adjust) | |
4368 | && !cfun->calls_alloca) | |
4369 | { | |
4370 | /* Restore fpu registers. */ | |
4371 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4372 | { | |
4373 | int fpr_position = cfun->machine->local_size | |
4374 | + cfun->machine->out_args_size; | |
4375 | /* Emit fpu load instruction, using [$sp + offset] restore | |
4376 | fpu registers. */ | |
4377 | nds32_emit_v3pop_fpr_callee_saved (fpr_position); | |
4378 | } | |
4379 | ||
4380 | /* We can use 'pop25 Re,imm8u'. */ | |
4381 | ||
4382 | /* nds32_emit_stack_v3pop(last_regno, sp_adjust), | |
4383 | the pattern 'stack_v3pop' is implementad in nds32.md. */ | |
4384 | nds32_emit_stack_v3pop (Rb, Re, sp_adjust); | |
4385 | } | |
4386 | else | |
4387 | { | |
4388 | /* We have to use 'pop25 Re,0', and prior to it, | |
4389 | we must expand one more instruction to adjust $sp. */ | |
4390 | ||
4391 | if (frame_pointer_needed) | |
4392 | { | |
4393 | /* adjust $sp = $fp - 4 ($fp size) | |
4394 | - 4 ($gp size) | |
4395 | - 4 ($lp size) | |
4396 | - (4 * n) (callee-saved registers) | |
4397 | Note: No need to adjust | |
4398 | cfun->machine->callee_saved_area_gpr_padding_bytes, | |
4399 | because we want to adjust stack pointer | |
4400 | to the position for pop instruction. */ | |
4401 | sp_adjust = cfun->machine->fp_size | |
4402 | + cfun->machine->gp_size | |
4403 | + cfun->machine->lp_size | |
4404 | + cfun->machine->callee_saved_gpr_regs_size; | |
4405 | ||
4406 | /* Restore fpu registers. */ | |
4407 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4408 | { | |
4409 | /* Set $sp to callee saved fpr position, we need to restore | |
4410 | fpr registers. */ | |
4411 | sp_adjust = sp_adjust | |
4412 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
4413 | + cfun->machine->callee_saved_fpr_regs_size; | |
4414 | ||
4415 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4416 | hard_frame_pointer_rtx, | |
4417 | -1 * sp_adjust); | |
4418 | ||
4419 | /* Emit fpu load instruction, using [$sp + offset] restore | |
4420 | fpu registers. */ | |
4421 | nds32_emit_v3pop_fpr_callee_saved (0); | |
4422 | } | |
4423 | else | |
4424 | { | |
4425 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4426 | hard_frame_pointer_rtx, | |
4427 | -1 * sp_adjust); | |
4428 | } | |
4429 | } | |
4430 | else | |
4431 | { | |
4432 | /* If frame pointer is NOT needed, | |
4433 | we cannot calculate the sp adjustment from frame pointer. | |
4434 | Instead, we calculate the adjustment by local_size, | |
4435 | out_args_size, and callee_saved_area_padding_bytes. | |
4436 | Notice that such sp adjustment value may be out of range, | |
4437 | so we have to deal with it as well. */ | |
4438 | ||
4439 | /* Adjust $sp = $sp + local_size + out_args_size | |
4440 | + callee_saved_area_gpr_padding_bytes | |
4441 | + callee_saved_fpr_regs_size. */ | |
4442 | sp_adjust = cfun->machine->local_size | |
4443 | + cfun->machine->out_args_size | |
4444 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
4445 | + cfun->machine->callee_saved_fpr_regs_size; | |
4446 | ||
4447 | /* Restore fpu registers. */ | |
4448 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4449 | { | |
4450 | /* Set $sp to callee saved fpr position, we need to restore | |
4451 | fpr registers. */ | |
4452 | sp_adjust = sp_adjust | |
4453 | - cfun->machine->callee_saved_area_gpr_padding_bytes | |
4454 | - cfun->machine->callee_saved_fpr_regs_size; | |
4455 | ||
4456 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4457 | stack_pointer_rtx, | |
4458 | sp_adjust); | |
4459 | ||
4460 | /* Emit fpu load instruction, using [$sp + offset] restore | |
4461 | fpu registers. */ | |
4462 | nds32_emit_v3pop_fpr_callee_saved (0); | |
4463 | } | |
4464 | else | |
4465 | { | |
4466 | /* sp_adjust value may be out of range of the addi instruction, | |
4467 | create alternative add behavior with TA_REGNUM if necessary, | |
4468 | using POSITIVE value to tell that we are increasing | |
4469 | address. */ | |
4470 | nds32_emit_adjust_frame (stack_pointer_rtx, | |
4471 | stack_pointer_rtx, | |
4472 | sp_adjust); | |
4473 | } | |
4474 | } | |
4475 | ||
4476 | if (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM) | |
4477 | { | |
4478 | /* We have fpr need to restore, so $sp is set on callee saved fpr | |
4479 | position. And we use 'pop25 Re, fpr_space' to adjust $sp. */ | |
4480 | int fpr_space = cfun->machine->callee_saved_area_gpr_padding_bytes | |
4481 | + cfun->machine->callee_saved_fpr_regs_size; | |
4482 | nds32_emit_stack_v3pop (Rb, Re, fpr_space); | |
4483 | } | |
4484 | else | |
4485 | { | |
4486 | /* nds32_emit_stack_v3pop(last_regno, sp_adjust), | |
4487 | the pattern 'stack_v3pop' is implementad in nds32.md. */ | |
4488 | nds32_emit_stack_v3pop (Rb, Re, 0); | |
4489 | } | |
4490 | } | |
4491 | /* Generate return instruction. */ | |
4492 | emit_jump_insn (gen_pop25return ()); | |
4493 | } | |
4494 | ||
4495 | /* Return nonzero if this function is known to have a null epilogue. | |
4496 | This allows the optimizer to omit jumps to jumps if no stack | |
4497 | was created. */ | |
4498 | int | |
4499 | nds32_can_use_return_insn (void) | |
4500 | { | |
4501 | int sp_adjust; | |
4502 | ||
4503 | /* Prior to reloading, we can't tell how many registers must be saved. | |
4504 | Thus we can not determine whether this function has null epilogue. */ | |
4505 | if (!reload_completed) | |
4506 | return 0; | |
4507 | ||
4508 | sp_adjust = cfun->machine->local_size | |
4509 | + cfun->machine->out_args_size | |
4510 | + cfun->machine->callee_saved_area_gpr_padding_bytes | |
4511 | + cfun->machine->callee_saved_fpr_regs_size; | |
4512 | if (!cfun->machine->fp_as_gp_p | |
4513 | && satisfies_constraint_Iu08 (GEN_INT (sp_adjust)) | |
4514 | && NDS32_DOUBLE_WORD_ALIGN_P (sp_adjust) | |
4515 | && !cfun->calls_alloca | |
4516 | && NDS32_V3PUSH_AVAILABLE_P | |
4517 | && !(TARGET_HARD_FLOAT | |
4518 | && (cfun->machine->callee_saved_first_fpr_regno != SP_REGNUM))) | |
4519 | return 1; | |
4520 | ||
4521 | /* If no stack was created, two conditions must be satisfied: | |
4522 | 1. This is a naked function. | |
4523 | So there is no callee-saved, local size, or outgoing size. | |
4524 | 2. This is NOT a variadic function. | |
4525 | So there is no pushing arguement registers into the stack. */ | |
4526 | return (cfun->machine->naked_p && (cfun->machine->va_args_size == 0)); | |
4527 | } | |
4528 | ||
4529 | /* ------------------------------------------------------------------------ */ | |
4530 | ||
4531 | /* Function to test 333-form for load/store instructions. | |
4532 | This is auxiliary extern function for auxiliary macro in nds32.h. | |
4533 | Because it is a little complicated, we use function instead of macro. */ | |
4534 | bool | |
4535 | nds32_ls_333_p (rtx rt, rtx ra, rtx imm, machine_mode mode) | |
4536 | { | |
4537 | if (REGNO_REG_CLASS (REGNO (rt)) == LOW_REGS | |
4538 | && REGNO_REG_CLASS (REGNO (ra)) == LOW_REGS) | |
4539 | { | |
4540 | if (GET_MODE_SIZE (mode) == 4) | |
4541 | return satisfies_constraint_Iu05 (imm); | |
4542 | ||
4543 | if (GET_MODE_SIZE (mode) == 2) | |
4544 | return satisfies_constraint_Iu04 (imm); | |
4545 | ||
4546 | if (GET_MODE_SIZE (mode) == 1) | |
4547 | return satisfies_constraint_Iu03 (imm); | |
4548 | } | |
4549 | ||
4550 | return false; | |
4551 | } | |
4552 | ||
4553 | bool | |
4554 | nds32_split_double_word_load_store_p(rtx *operands, bool load_p) | |
4555 | { | |
4556 | rtx mem = load_p ? operands[1] : operands[0]; | |
4557 | /* Do split at split2 if -O0 or schedule 2 not enable. */ | |
4558 | if (optimize == 0 || !flag_schedule_insns_after_reload) | |
4559 | return !satisfies_constraint_Da (mem) || MEM_VOLATILE_P (mem); | |
4560 | ||
4561 | /* Split double word load store after copy propgation. */ | |
4562 | if (current_pass == NULL) | |
4563 | return false; | |
4564 | ||
4565 | const char *pass_name = current_pass->name; | |
4566 | if (pass_name && ((strcmp (pass_name, "split4") == 0) | |
4567 | || (strcmp (pass_name, "split5") == 0))) | |
4568 | return !satisfies_constraint_Da (mem) || MEM_VOLATILE_P (mem); | |
4569 | ||
4570 | return false; | |
4571 | } | |
4572 | ||
4573 | static bool | |
4574 | nds32_use_blocks_for_constant_p (machine_mode mode, | |
4575 | const_rtx x ATTRIBUTE_UNUSED) | |
4576 | { | |
4577 | if ((TARGET_FPU_SINGLE || TARGET_FPU_DOUBLE) | |
4578 | && (mode == DFmode || mode == SFmode)) | |
4579 | return true; | |
4580 | else | |
4581 | return false; | |
4582 | } | |
4583 | ||
4584 | /* Return align 2 (log base 2) if the next instruction of LABEL is 4 byte. */ | |
4585 | int | |
4586 | nds32_target_alignment (rtx_insn *label) | |
4587 | { | |
4588 | rtx_insn *insn; | |
4589 | ||
4590 | if (optimize_size) | |
4591 | return 0; | |
4592 | ||
4593 | insn = next_active_insn (label); | |
4594 | ||
4595 | if (insn == 0) | |
4596 | return 0; | |
4597 | else if ((get_attr_length (insn) % 4) == 0) | |
4598 | return 2; | |
4599 | else | |
4600 | return 0; | |
4601 | } | |
4602 | ||
4603 | /* ------------------------------------------------------------------------ */ | |
4604 | ||
4605 | /* PART 5: Initialize target hook structure and definitions. */ | |
4606 | \f | |
4607 | /* Controlling the Compilation Driver. */ | |
4608 | ||
4609 | \f | |
4610 | /* Run-time Target Specification. */ | |
4611 | ||
4612 | \f | |
4613 | /* Defining Data Structures for Per-function Information. */ | |
4614 | ||
4615 | \f | |
4616 | /* Storage Layout. */ | |
4617 | ||
4618 | #undef TARGET_PROMOTE_FUNCTION_MODE | |
4619 | #define TARGET_PROMOTE_FUNCTION_MODE \ | |
4620 | default_promote_function_mode_always_promote | |
4621 | ||
4622 | \f | |
4623 | /* Layout of Source Language Data Types. */ | |
4624 | ||
4625 | \f | |
4626 | /* Register Usage. */ | |
4627 | ||
4628 | /* -- Basic Characteristics of Registers. */ | |
4629 | ||
4630 | #undef TARGET_CONDITIONAL_REGISTER_USAGE | |
4631 | #define TARGET_CONDITIONAL_REGISTER_USAGE nds32_conditional_register_usage | |
4632 | ||
4633 | /* -- Order of Allocation of Registers. */ | |
4634 | ||
4635 | /* -- How Values Fit in Registers. */ | |
4636 | ||
4637 | #undef TARGET_HARD_REGNO_NREGS | |
4638 | #define TARGET_HARD_REGNO_NREGS nds32_hard_regno_nregs | |
4639 | ||
4640 | /* -- Handling Leaf Functions. */ | |
4641 | ||
4642 | /* -- Registers That Form a Stack. */ | |
4643 | ||
4644 | \f | |
4645 | /* Register Classes. */ | |
4646 | ||
4647 | #undef TARGET_CLASS_MAX_NREGS | |
4648 | #define TARGET_CLASS_MAX_NREGS nds32_class_max_nregs | |
4649 | ||
4650 | #undef TARGET_REGISTER_PRIORITY | |
4651 | #define TARGET_REGISTER_PRIORITY nds32_register_priority | |
4652 | ||
4653 | #undef TARGET_CAN_CHANGE_MODE_CLASS | |
4654 | #define TARGET_CAN_CHANGE_MODE_CLASS nds32_can_change_mode_class | |
4655 | ||
4656 | \f | |
4657 | /* Obsolete Macros for Defining Constraints. */ | |
4658 | ||
4659 | \f | |
4660 | /* Stack Layout and Calling Conventions. */ | |
4661 | ||
4662 | /* -- Basic Stack Layout. */ | |
4663 | ||
4664 | /* -- Exception Handling Support. */ | |
4665 | ||
4666 | /* -- Specifying How Stack Checking is Done. */ | |
4667 | ||
4668 | /* -- Registers That Address the Stack Frame. */ | |
4669 | ||
4670 | /* -- Eliminating Frame Pointer and Arg Pointer. */ | |
4671 | ||
4672 | #undef TARGET_CAN_ELIMINATE | |
4673 | #define TARGET_CAN_ELIMINATE nds32_can_eliminate | |
4674 | ||
4675 | /* -- Passing Function Arguments on the Stack. */ | |
4676 | ||
4677 | /* -- Passing Arguments in Registers. */ | |
4678 | ||
4679 | #undef TARGET_FUNCTION_ARG | |
4680 | #define TARGET_FUNCTION_ARG nds32_function_arg | |
4681 | ||
4682 | #undef TARGET_MUST_PASS_IN_STACK | |
4683 | #define TARGET_MUST_PASS_IN_STACK nds32_must_pass_in_stack | |
4684 | ||
4685 | #undef TARGET_ARG_PARTIAL_BYTES | |
4686 | #define TARGET_ARG_PARTIAL_BYTES nds32_arg_partial_bytes | |
4687 | ||
4688 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
4689 | #define TARGET_FUNCTION_ARG_ADVANCE nds32_function_arg_advance | |
4690 | ||
4691 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
4692 | #define TARGET_FUNCTION_ARG_BOUNDARY nds32_function_arg_boundary | |
4693 | ||
4694 | /* -- How Scalar Function Values Are Returned. */ | |
4695 | ||
4696 | #undef TARGET_FUNCTION_VALUE | |
4697 | #define TARGET_FUNCTION_VALUE nds32_function_value | |
4698 | ||
4699 | #undef TARGET_LIBCALL_VALUE | |
4700 | #define TARGET_LIBCALL_VALUE nds32_libcall_value | |
4701 | ||
4702 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
4703 | #define TARGET_FUNCTION_VALUE_REGNO_P nds32_function_value_regno_p | |
4704 | ||
4705 | /* -- How Large Values Are Returned. */ | |
4706 | ||
4707 | #undef TARGET_RETURN_IN_MEMORY | |
4708 | #define TARGET_RETURN_IN_MEMORY nds32_return_in_memory | |
4709 | ||
4710 | /* -- Caller-Saves Register Allocation. */ | |
4711 | ||
4712 | /* -- Function Entry and Exit. */ | |
4713 | ||
4714 | #undef TARGET_ASM_FUNCTION_PROLOGUE | |
4715 | #define TARGET_ASM_FUNCTION_PROLOGUE nds32_asm_function_prologue | |
4716 | ||
4717 | #undef TARGET_ASM_FUNCTION_END_PROLOGUE | |
4718 | #define TARGET_ASM_FUNCTION_END_PROLOGUE nds32_asm_function_end_prologue | |
4719 | ||
4720 | #undef TARGET_ASM_FUNCTION_BEGIN_EPILOGUE | |
4721 | #define TARGET_ASM_FUNCTION_BEGIN_EPILOGUE nds32_asm_function_begin_epilogue | |
4722 | ||
4723 | #undef TARGET_ASM_FUNCTION_EPILOGUE | |
4724 | #define TARGET_ASM_FUNCTION_EPILOGUE nds32_asm_function_epilogue | |
4725 | ||
4726 | #undef TARGET_ASM_OUTPUT_MI_THUNK | |
4727 | #define TARGET_ASM_OUTPUT_MI_THUNK nds32_asm_output_mi_thunk | |
4728 | ||
4729 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
4730 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall | |
4731 | ||
4732 | /* -- Generating Code for Profiling. */ | |
4733 | ||
4734 | /* -- Permitting tail calls. */ | |
4735 | ||
4736 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
4737 | #define TARGET_FUNCTION_OK_FOR_SIBCALL nds32_function_ok_for_sibcall | |
4738 | ||
4739 | #undef TARGET_WARN_FUNC_RETURN | |
4740 | #define TARGET_WARN_FUNC_RETURN nds32_warn_func_return | |
4741 | ||
4742 | /* Stack smashing protection. */ | |
4743 | ||
4744 | \f | |
4745 | /* Implementing the Varargs Macros. */ | |
4746 | ||
4747 | #undef TARGET_SETUP_INCOMING_VARARGS | |
4748 | #define TARGET_SETUP_INCOMING_VARARGS nds32_setup_incoming_varargs | |
4749 | ||
4750 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
4751 | #define TARGET_STRICT_ARGUMENT_NAMING nds32_strict_argument_naming | |
4752 | ||
4753 | \f | |
4754 | /* Trampolines for Nested Functions. */ | |
4755 | ||
4756 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
4757 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE nds32_asm_trampoline_template | |
4758 | ||
4759 | #undef TARGET_TRAMPOLINE_INIT | |
4760 | #define TARGET_TRAMPOLINE_INIT nds32_trampoline_init | |
4761 | ||
4762 | \f | |
4763 | /* Implicit Calls to Library Routines. */ | |
4764 | ||
4765 | \f | |
4766 | /* Addressing Modes. */ | |
4767 | ||
4768 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
4769 | #define TARGET_LEGITIMATE_ADDRESS_P nds32_legitimate_address_p | |
4770 | ||
4771 | \f | |
4772 | /* Anchored Addresses. */ | |
4773 | ||
4774 | \f | |
4775 | /* Condition Code Status. */ | |
4776 | ||
4777 | /* -- Representation of condition codes using (cc0). */ | |
4778 | ||
4779 | /* -- Representation of condition codes using registers. */ | |
4780 | ||
4781 | #undef TARGET_CANONICALIZE_COMPARISON | |
4782 | #define TARGET_CANONICALIZE_COMPARISON nds32_canonicalize_comparison | |
4783 | ||
4784 | /* -- Macros to control conditional execution. */ | |
4785 | ||
4786 | \f | |
4787 | /* Describing Relative Costs of Operations. */ | |
4788 | ||
4789 | #undef TARGET_REGISTER_MOVE_COST | |
4790 | #define TARGET_REGISTER_MOVE_COST nds32_register_move_cost | |
4791 | ||
4792 | #undef TARGET_MEMORY_MOVE_COST | |
4793 | #define TARGET_MEMORY_MOVE_COST nds32_memory_move_cost | |
4794 | ||
4795 | #undef TARGET_RTX_COSTS | |
4796 | #define TARGET_RTX_COSTS nds32_rtx_costs | |
4797 | ||
4798 | #undef TARGET_ADDRESS_COST | |
4799 | #define TARGET_ADDRESS_COST nds32_address_cost | |
4800 | ||
4801 | \f | |
4802 | /* Adjusting the Instruction Scheduler. */ | |
4803 | ||
4804 | \f | |
4805 | /* Dividing the Output into Sections (Texts, Data, . . . ). */ | |
4806 | ||
4807 | #undef TARGET_ENCODE_SECTION_INFO | |
4808 | #define TARGET_ENCODE_SECTION_INFO nds32_encode_section_info | |
4809 | ||
4810 | \f | |
4811 | /* Position Independent Code. */ | |
4812 | ||
4813 | \f | |
4814 | /* Defining the Output Assembler Language. */ | |
4815 | ||
4816 | /* -- The Overall Framework of an Assembler File. */ | |
4817 | ||
4818 | #undef TARGET_ASM_FILE_START | |
4819 | #define TARGET_ASM_FILE_START nds32_asm_file_start | |
4820 | #undef TARGET_ASM_FILE_END | |
4821 | #define TARGET_ASM_FILE_END nds32_asm_file_end | |
4822 | ||
4823 | /* -- Output of Data. */ | |
4824 | ||
4825 | #undef TARGET_ASM_ALIGNED_HI_OP | |
4826 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
4827 | ||
4828 | #undef TARGET_ASM_ALIGNED_SI_OP | |
4829 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
4830 | ||
4831 | /* -- Output of Uninitialized Variables. */ | |
4832 | ||
4833 | /* -- Output and Generation of Labels. */ | |
4834 | ||
4835 | #undef TARGET_ASM_GLOBALIZE_LABEL | |
4836 | #define TARGET_ASM_GLOBALIZE_LABEL nds32_asm_globalize_label | |
4837 | ||
4838 | /* -- How Initialization Functions Are Handled. */ | |
4839 | ||
4840 | /* -- Macros Controlling Initialization Routines. */ | |
4841 | ||
4842 | /* -- Output of Assembler Instructions. */ | |
4843 | ||
4844 | #undef TARGET_PRINT_OPERAND | |
4845 | #define TARGET_PRINT_OPERAND nds32_print_operand | |
4846 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
4847 | #define TARGET_PRINT_OPERAND_ADDRESS nds32_print_operand_address | |
4848 | ||
4849 | /* -- Output of Dispatch Tables. */ | |
4850 | ||
4851 | /* -- Assembler Commands for Exception Regions. */ | |
4852 | ||
4853 | #undef TARGET_DWARF_REGISTER_SPAN | |
4854 | #define TARGET_DWARF_REGISTER_SPAN nds32_dwarf_register_span | |
4855 | ||
4856 | /* -- Assembler Commands for Alignment. */ | |
4857 | ||
4858 | \f | |
4859 | /* Controlling Debugging Information Format. */ | |
4860 | ||
4861 | /* -- Macros Affecting All Debugging Formats. */ | |
4862 | ||
4863 | /* -- Specific Options for DBX Output. */ | |
4864 | ||
4865 | /* -- Open-Ended Hooks for DBX Format. */ | |
4866 | ||
4867 | /* -- File Names in DBX Format. */ | |
4868 | ||
4869 | /* -- Macros for DWARF Output. */ | |
4870 | ||
4871 | /* -- Macros for VMS Debug Format. */ | |
4872 | ||
4873 | \f | |
4874 | /* Cross Compilation and Floating Point. */ | |
4875 | ||
4876 | \f | |
4877 | /* Mode Switching Instructions. */ | |
4878 | ||
4879 | \f | |
4880 | /* Defining target-specific uses of __attribute__. */ | |
4881 | ||
4882 | #undef TARGET_ATTRIBUTE_TABLE | |
4883 | #define TARGET_ATTRIBUTE_TABLE nds32_attribute_table | |
4884 | ||
4885 | #undef TARGET_MERGE_DECL_ATTRIBUTES | |
4886 | #define TARGET_MERGE_DECL_ATTRIBUTES nds32_merge_decl_attributes | |
4887 | ||
4888 | #undef TARGET_INSERT_ATTRIBUTES | |
4889 | #define TARGET_INSERT_ATTRIBUTES nds32_insert_attributes | |
4890 | ||
4891 | #undef TARGET_OPTION_PRAGMA_PARSE | |
4892 | #define TARGET_OPTION_PRAGMA_PARSE nds32_option_pragma_parse | |
4893 | ||
4894 | #undef TARGET_OPTION_OVERRIDE | |
4895 | #define TARGET_OPTION_OVERRIDE nds32_option_override | |
4896 | ||
4897 | \f | |
4898 | /* Emulating TLS. */ | |
4899 | ||
4900 | \f | |
4901 | /* Defining coprocessor specifics for MIPS targets. */ | |
4902 | ||
4903 | \f | |
4904 | /* Parameters for Precompiled Header Validity Checking. */ | |
4905 | ||
4906 | \f | |
4907 | /* C++ ABI parameters. */ | |
4908 | ||
4909 | \f | |
4910 | /* Adding support for named address spaces. */ | |
4911 | ||
4912 | \f | |
4913 | /* Miscellaneous Parameters. */ | |
4914 | ||
4915 | #undef TARGET_MD_ASM_ADJUST | |
4916 | #define TARGET_MD_ASM_ADJUST nds32_md_asm_adjust | |
4917 | ||
4918 | #undef TARGET_INIT_BUILTINS | |
4919 | #define TARGET_INIT_BUILTINS nds32_init_builtins | |
4920 | ||
4921 | #undef TARGET_BUILTIN_DECL | |
4922 | #define TARGET_BUILTIN_DECL nds32_builtin_decl | |
4923 | ||
4924 | #undef TARGET_EXPAND_BUILTIN | |
4925 | #define TARGET_EXPAND_BUILTIN nds32_expand_builtin | |
4926 | ||
4927 | ||
4928 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
4929 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P nds32_use_blocks_for_constant_p | |
4930 | ||
4931 | \f | |
4932 | /* ------------------------------------------------------------------------ */ | |
4933 | ||
4934 | /* Initialize the GCC target structure. */ | |
4935 | ||
4936 | struct gcc_target targetm = TARGET_INITIALIZER; | |
4937 | ||
4938 | /* ------------------------------------------------------------------------ */ |