]>
Commit | Line | Data |
---|---|---|
9304f876 | 1 | /* Subroutines used for code generation of Andes NDS32 cpu for GNU compiler |
5624e564 | 2 | Copyright (C) 2012-2015 Free Software Foundation, Inc. |
9304f876 CJW |
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 | ||
19ac960a | 21 | /* ------------------------------------------------------------------------ */ |
9304f876 CJW |
22 | |
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
40e23961 MC |
27 | #include "hash-set.h" |
28 | #include "machmode.h" | |
29 | #include "vec.h" | |
30 | #include "double-int.h" | |
31 | #include "input.h" | |
32 | #include "alias.h" | |
33 | #include "symtab.h" | |
34 | #include "wide-int.h" | |
35 | #include "inchash.h" | |
9304f876 | 36 | #include "tree.h" |
d8a2d370 DN |
37 | #include "stor-layout.h" |
38 | #include "varasm.h" | |
39 | #include "calls.h" | |
9304f876 CJW |
40 | #include "rtl.h" |
41 | #include "regs.h" | |
42 | #include "hard-reg-set.h" | |
43 | #include "insn-config.h" /* Required by recog.h. */ | |
44 | #include "conditions.h" | |
45 | #include "output.h" | |
46 | #include "insn-attr.h" /* For DFA state_t. */ | |
47 | #include "insn-codes.h" /* For CODE_FOR_xxx. */ | |
48 | #include "reload.h" /* For push_reload(). */ | |
49 | #include "flags.h" | |
83685514 | 50 | #include "input.h" |
9304f876 CJW |
51 | #include "function.h" |
52 | #include "expr.h" | |
53 | #include "recog.h" | |
54 | #include "diagnostic-core.h" | |
60393bbc AM |
55 | #include "dominance.h" |
56 | #include "cfg.h" | |
57 | #include "cfgrtl.h" | |
58 | #include "cfganal.h" | |
59 | #include "lcm.h" | |
60 | #include "cfgbuild.h" | |
61 | #include "cfgcleanup.h" | |
62 | #include "predict.h" | |
63 | #include "basic-block.h" | |
9304f876 CJW |
64 | #include "df.h" |
65 | #include "tm_p.h" | |
66 | #include "tm-constrs.h" | |
67 | #include "optabs.h" /* For GEN_FCN. */ | |
68 | #include "target.h" | |
69 | #include "target-def.h" | |
70 | #include "langhooks.h" /* For add_builtin_function(). */ | |
71 | #include "ggc.h" | |
9b2b7279 | 72 | #include "builtins.h" |
9304f876 CJW |
73 | |
74 | /* ------------------------------------------------------------------------ */ | |
75 | ||
76 | /* This file is divided into five parts: | |
77 | ||
78 | PART 1: Auxiliary static variable definitions and | |
79 | target hook static variable definitions. | |
80 | ||
81 | PART 2: Auxiliary static function definitions. | |
82 | ||
83 | PART 3: Implement target hook stuff definitions. | |
84 | ||
85 | PART 4: Implemet extern function definitions, | |
86 | the prototype is in nds32-protos.h. | |
87 | ||
88 | PART 5: Initialize target hook structure and definitions. */ | |
89 | ||
90 | /* ------------------------------------------------------------------------ */ | |
91 | ||
92 | /* PART 1: Auxiliary static variable definitions and | |
93 | target hook static variable definitions. */ | |
94 | ||
9304f876 CJW |
95 | /* Define intrinsic register names. |
96 | Please refer to nds32_intrinsic.h file, the index is corresponding to | |
97 | 'enum nds32_intrinsic_registers' data type values. | |
98 | NOTE that the base value starting from 1024. */ | |
99 | static const char * const nds32_intrinsic_register_names[] = | |
100 | { | |
101 | "$PSW", "$IPSW", "$ITYPE", "$IPC" | |
102 | }; | |
103 | ||
104 | /* Defining target-specific uses of __attribute__. */ | |
105 | static const struct attribute_spec nds32_attribute_table[] = | |
106 | { | |
107 | /* Syntax: { name, min_len, max_len, decl_required, type_required, | |
108 | function_type_required, handler, affects_type_identity } */ | |
109 | ||
110 | /* The interrupt vid: [0-63]+ (actual vector number starts from 9 to 72). */ | |
111 | { "interrupt", 1, 64, false, false, false, NULL, false }, | |
112 | /* The exception vid: [1-8]+ (actual vector number starts from 1 to 8). */ | |
113 | { "exception", 1, 8, false, false, false, NULL, false }, | |
114 | /* Argument is user's interrupt numbers. The vector number is always 0. */ | |
115 | { "reset", 1, 1, false, false, false, NULL, false }, | |
116 | ||
117 | /* The attributes describing isr nested type. */ | |
118 | { "nested", 0, 0, false, false, false, NULL, false }, | |
119 | { "not_nested", 0, 0, false, false, false, NULL, false }, | |
120 | { "nested_ready", 0, 0, false, false, false, NULL, false }, | |
121 | ||
122 | /* The attributes describing isr register save scheme. */ | |
123 | { "save_all", 0, 0, false, false, false, NULL, false }, | |
124 | { "partial_save", 0, 0, false, false, false, NULL, false }, | |
125 | ||
126 | /* The attributes used by reset attribute. */ | |
127 | { "nmi", 1, 1, false, false, false, NULL, false }, | |
128 | { "warm", 1, 1, false, false, false, NULL, false }, | |
129 | ||
130 | /* The attribute telling no prologue/epilogue. */ | |
131 | { "naked", 0, 0, false, false, false, NULL, false }, | |
132 | ||
133 | /* The last attribute spec is set to be NULL. */ | |
134 | { NULL, 0, 0, false, false, false, NULL, false } | |
135 | }; | |
136 | ||
137 | ||
138 | /* ------------------------------------------------------------------------ */ | |
139 | ||
140 | /* PART 2: Auxiliary static function definitions. */ | |
141 | ||
142 | /* Function to save and restore machine-specific function data. */ | |
143 | static struct machine_function * | |
144 | nds32_init_machine_status (void) | |
145 | { | |
146 | struct machine_function *machine; | |
766090c2 | 147 | machine = ggc_cleared_alloc<machine_function> (); |
9304f876 CJW |
148 | |
149 | /* Initially assume this function needs prologue/epilogue. */ | |
150 | machine->naked_p = 0; | |
151 | ||
152 | /* Initially assume this function does NOT use fp_as_gp optimization. */ | |
153 | machine->fp_as_gp_p = 0; | |
154 | ||
155 | return machine; | |
156 | } | |
157 | ||
158 | /* Function to compute stack frame size and | |
159 | store into cfun->machine structure. */ | |
160 | static void | |
161 | nds32_compute_stack_frame (void) | |
162 | { | |
163 | int r; | |
164 | int block_size; | |
165 | ||
166 | /* Because nds32_compute_stack_frame() will be called from different place, | |
167 | everytime we enter this function, we have to assume this function | |
168 | needs prologue/epilogue. */ | |
169 | cfun->machine->naked_p = 0; | |
170 | ||
171 | /* Get variadic arguments size to prepare pretend arguments and | |
35da54a6 CJW |
172 | we will push them into stack at prologue by ourself. */ |
173 | cfun->machine->va_args_size = crtl->args.pretend_args_size; | |
174 | if (cfun->machine->va_args_size != 0) | |
175 | { | |
176 | cfun->machine->va_args_first_regno | |
177 | = NDS32_GPR_ARG_FIRST_REGNUM | |
178 | + NDS32_MAX_GPR_REGS_FOR_ARGS | |
179 | - (crtl->args.pretend_args_size / UNITS_PER_WORD); | |
180 | cfun->machine->va_args_last_regno | |
181 | = NDS32_GPR_ARG_FIRST_REGNUM + NDS32_MAX_GPR_REGS_FOR_ARGS - 1; | |
182 | } | |
183 | else | |
184 | { | |
185 | cfun->machine->va_args_first_regno = SP_REGNUM; | |
186 | cfun->machine->va_args_last_regno = SP_REGNUM; | |
187 | } | |
188 | ||
189 | /* Important: We need to make sure that varargs area is 8-byte alignment. */ | |
190 | block_size = cfun->machine->va_args_size; | |
191 | if (!NDS32_DOUBLE_WORD_ALIGN_P (block_size)) | |
192 | { | |
193 | cfun->machine->va_args_area_padding_bytes | |
194 | = NDS32_ROUND_UP_DOUBLE_WORD (block_size) - block_size; | |
195 | } | |
9304f876 CJW |
196 | |
197 | /* Get local variables, incoming variables, and temporary variables size. | |
198 | Note that we need to make sure it is 8-byte alignment because | |
199 | there may be no padding bytes if we are using LRA. */ | |
200 | cfun->machine->local_size = NDS32_ROUND_UP_DOUBLE_WORD (get_frame_size ()); | |
201 | ||
202 | /* Get outgoing arguments size. */ | |
203 | cfun->machine->out_args_size = crtl->outgoing_args_size; | |
204 | ||
205 | /* If $fp value is required to be saved on stack, it needs 4 bytes space. | |
206 | Check whether $fp is ever live. */ | |
207 | cfun->machine->fp_size = (df_regs_ever_live_p (FP_REGNUM)) ? 4 : 0; | |
208 | ||
209 | /* If $gp value is required to be saved on stack, it needs 4 bytes space. | |
210 | Check whether we are using PIC code genration. */ | |
211 | cfun->machine->gp_size = (flag_pic) ? 4 : 0; | |
212 | ||
213 | /* If $lp value is required to be saved on stack, it needs 4 bytes space. | |
214 | Check whether $lp is ever live. */ | |
215 | cfun->machine->lp_size = (df_regs_ever_live_p (LP_REGNUM)) ? 4 : 0; | |
216 | ||
217 | /* Initially there is no padding bytes. */ | |
218 | cfun->machine->callee_saved_area_padding_bytes = 0; | |
219 | ||
220 | /* Calculate the bytes of saving callee-saved registers on stack. */ | |
221 | cfun->machine->callee_saved_regs_size = 0; | |
222 | cfun->machine->callee_saved_regs_first_regno = SP_REGNUM; | |
223 | cfun->machine->callee_saved_regs_last_regno = SP_REGNUM; | |
224 | /* Currently, there is no need to check $r28~$r31 | |
225 | because we will save them in another way. */ | |
226 | for (r = 0; r < 28; r++) | |
227 | { | |
228 | if (NDS32_REQUIRED_CALLEE_SAVED_P (r)) | |
229 | { | |
230 | /* Mark the first required callee-saved register | |
231 | (only need to set it once). | |
232 | If first regno == SP_REGNUM, we can tell that | |
233 | it is the first time to be here. */ | |
234 | if (cfun->machine->callee_saved_regs_first_regno == SP_REGNUM) | |
235 | cfun->machine->callee_saved_regs_first_regno = r; | |
236 | /* Mark the last required callee-saved register. */ | |
237 | cfun->machine->callee_saved_regs_last_regno = r; | |
238 | } | |
239 | } | |
240 | ||
241 | /* Check if this function can omit prologue/epilogue code fragment. | |
242 | If there is 'naked' attribute in this function, | |
243 | we can set 'naked_p' flag to indicate that | |
244 | we do not have to generate prologue/epilogue. | |
245 | Or, if all the following conditions succeed, | |
246 | we can set this function 'naked_p' as well: | |
247 | condition 1: first_regno == last_regno == SP_REGNUM, | |
248 | which means we do not have to save | |
249 | any callee-saved registers. | |
250 | condition 2: Both $lp and $fp are NOT live in this function, | |
2da1e7c0 CJW |
251 | which means we do not need to save them and there |
252 | is no outgoing size. | |
9304f876 CJW |
253 | condition 3: There is no local_size, which means |
254 | we do not need to adjust $sp. */ | |
255 | if (lookup_attribute ("naked", DECL_ATTRIBUTES (current_function_decl)) | |
256 | || (cfun->machine->callee_saved_regs_first_regno == SP_REGNUM | |
257 | && cfun->machine->callee_saved_regs_last_regno == SP_REGNUM | |
258 | && !df_regs_ever_live_p (FP_REGNUM) | |
259 | && !df_regs_ever_live_p (LP_REGNUM) | |
260 | && cfun->machine->local_size == 0)) | |
261 | { | |
2da1e7c0 CJW |
262 | /* Set this function 'naked_p' and other functions can check this flag. |
263 | Note that in nds32 port, the 'naked_p = 1' JUST means there is no | |
264 | callee-saved, local size, and outgoing size. | |
265 | The varargs space and ret instruction may still present in | |
266 | the prologue/epilogue expanding. */ | |
9304f876 CJW |
267 | cfun->machine->naked_p = 1; |
268 | ||
269 | /* No need to save $fp, $gp, and $lp. | |
270 | We should set these value to be zero | |
271 | so that nds32_initial_elimination_offset() can work properly. */ | |
272 | cfun->machine->fp_size = 0; | |
273 | cfun->machine->gp_size = 0; | |
274 | cfun->machine->lp_size = 0; | |
275 | ||
276 | /* If stack usage computation is required, | |
277 | we need to provide the static stack size. */ | |
278 | if (flag_stack_usage_info) | |
279 | current_function_static_stack_size = 0; | |
280 | ||
281 | /* No need to do following adjustment, return immediately. */ | |
282 | return; | |
283 | } | |
284 | ||
285 | /* Adjustment for v3push instructions: | |
286 | If we are using v3push (push25/pop25) instructions, | |
287 | we need to make sure Rb is $r6 and Re is | |
288 | located on $r6, $r8, $r10, or $r14. | |
289 | Some results above will be discarded and recomputed. | |
2da1e7c0 CJW |
290 | Note that it is only available under V3/V3M ISA and we |
291 | DO NOT setup following stuff for isr or variadic function. */ | |
292 | if (TARGET_V3PUSH | |
293 | && !nds32_isr_function_p (current_function_decl) | |
294 | && (cfun->machine->va_args_size == 0)) | |
9304f876 CJW |
295 | { |
296 | /* Recompute: | |
297 | cfun->machine->fp_size | |
298 | cfun->machine->gp_size | |
299 | cfun->machine->lp_size | |
300 | cfun->machine->callee_saved_regs_first_regno | |
301 | cfun->machine->callee_saved_regs_last_regno */ | |
302 | ||
303 | /* For v3push instructions, $fp, $gp, and $lp are always saved. */ | |
304 | cfun->machine->fp_size = 4; | |
305 | cfun->machine->gp_size = 4; | |
306 | cfun->machine->lp_size = 4; | |
307 | ||
308 | /* Remember to set Rb = $r6. */ | |
309 | cfun->machine->callee_saved_regs_first_regno = 6; | |
310 | ||
311 | if (cfun->machine->callee_saved_regs_last_regno <= 6) | |
312 | { | |
313 | /* Re = $r6 */ | |
314 | cfun->machine->callee_saved_regs_last_regno = 6; | |
315 | } | |
316 | else if (cfun->machine->callee_saved_regs_last_regno <= 8) | |
317 | { | |
318 | /* Re = $r8 */ | |
319 | cfun->machine->callee_saved_regs_last_regno = 8; | |
320 | } | |
321 | else if (cfun->machine->callee_saved_regs_last_regno <= 10) | |
322 | { | |
323 | /* Re = $r10 */ | |
324 | cfun->machine->callee_saved_regs_last_regno = 10; | |
325 | } | |
326 | else if (cfun->machine->callee_saved_regs_last_regno <= 14) | |
327 | { | |
328 | /* Re = $r14 */ | |
329 | cfun->machine->callee_saved_regs_last_regno = 14; | |
330 | } | |
331 | else if (cfun->machine->callee_saved_regs_last_regno == SP_REGNUM) | |
332 | { | |
333 | /* If last_regno is SP_REGNUM, which means | |
334 | it is never changed, so set it to Re = $r6. */ | |
335 | cfun->machine->callee_saved_regs_last_regno = 6; | |
336 | } | |
337 | else | |
338 | { | |
339 | /* The program flow should not go here. */ | |
340 | gcc_unreachable (); | |
341 | } | |
342 | } | |
343 | ||
344 | /* We have correctly set callee_saved_regs_first_regno | |
345 | and callee_saved_regs_last_regno. | |
346 | Initially, the callee_saved_regs_size is supposed to be 0. | |
347 | As long as callee_saved_regs_last_regno is not SP_REGNUM, | |
348 | we can update callee_saved_regs_size with new size. */ | |
349 | if (cfun->machine->callee_saved_regs_last_regno != SP_REGNUM) | |
350 | { | |
351 | /* Compute pushed size of callee-saved registers. */ | |
352 | cfun->machine->callee_saved_regs_size | |
353 | = 4 * (cfun->machine->callee_saved_regs_last_regno | |
354 | - cfun->machine->callee_saved_regs_first_regno | |
355 | + 1); | |
356 | } | |
357 | ||
358 | /* Important: We need to make sure that | |
35da54a6 | 359 | (fp_size + gp_size + lp_size + callee_saved_regs_size) |
9304f876 CJW |
360 | is 8-byte alignment. |
361 | If it is not, calculate the padding bytes. */ | |
35da54a6 | 362 | block_size = cfun->machine->fp_size |
9304f876 CJW |
363 | + cfun->machine->gp_size |
364 | + cfun->machine->lp_size | |
365 | + cfun->machine->callee_saved_regs_size; | |
366 | if (!NDS32_DOUBLE_WORD_ALIGN_P (block_size)) | |
367 | { | |
368 | cfun->machine->callee_saved_area_padding_bytes | |
369 | = NDS32_ROUND_UP_DOUBLE_WORD (block_size) - block_size; | |
370 | } | |
371 | ||
372 | /* If stack usage computation is required, | |
373 | we need to provide the static stack size. */ | |
374 | if (flag_stack_usage_info) | |
375 | { | |
376 | current_function_static_stack_size | |
377 | = NDS32_ROUND_UP_DOUBLE_WORD (block_size) | |
378 | + cfun->machine->local_size | |
379 | + cfun->machine->out_args_size; | |
380 | } | |
381 | } | |
382 | ||
383 | /* Function to create a parallel rtx pattern | |
384 | which presents stack push multiple behavior. | |
385 | The overall concept are: | |
386 | "push registers to memory", | |
387 | "adjust stack pointer". */ | |
4e9a2848 | 388 | static void |
47e0e7d2 | 389 | nds32_emit_stack_push_multiple (rtx Rb, rtx Re, rtx En4, bool vaarg_p) |
9304f876 CJW |
390 | { |
391 | int regno; | |
392 | int extra_count; | |
393 | int num_use_regs; | |
394 | int par_index; | |
395 | int offset; | |
4e9a2848 | 396 | int save_fp, save_gp, save_lp; |
9304f876 CJW |
397 | |
398 | rtx reg; | |
399 | rtx mem; | |
400 | rtx push_rtx; | |
401 | rtx adjust_sp_rtx; | |
402 | rtx parallel_insn; | |
47e0e7d2 | 403 | rtx dwarf; |
9304f876 CJW |
404 | |
405 | /* We need to provide a customized rtx which contains | |
406 | necessary information for data analysis, | |
407 | so we create a parallel rtx like this: | |
408 | (parallel [(set (mem (plus (reg:SI SP_REGNUM) (const_int -32))) | |
409 | (reg:SI Rb)) | |
410 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -28))) | |
411 | (reg:SI Rb+1)) | |
412 | ... | |
413 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -16))) | |
414 | (reg:SI Re)) | |
415 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -12))) | |
416 | (reg:SI FP_REGNUM)) | |
417 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -8))) | |
418 | (reg:SI GP_REGNUM)) | |
419 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -4))) | |
420 | (reg:SI LP_REGNUM)) | |
421 | (set (reg:SI SP_REGNUM) | |
422 | (plus (reg:SI SP_REGNUM) (const_int -32)))]) */ | |
423 | ||
4e9a2848 CJW |
424 | /* Determine whether we need to save $fp, $gp, or $lp. */ |
425 | save_fp = INTVAL (En4) & 0x8; | |
426 | save_gp = INTVAL (En4) & 0x4; | |
427 | save_lp = INTVAL (En4) & 0x2; | |
428 | ||
9304f876 CJW |
429 | /* Calculate the number of registers that will be pushed. */ |
430 | extra_count = 0; | |
4e9a2848 | 431 | if (save_fp) |
9304f876 | 432 | extra_count++; |
4e9a2848 | 433 | if (save_gp) |
9304f876 | 434 | extra_count++; |
4e9a2848 | 435 | if (save_lp) |
9304f876 CJW |
436 | extra_count++; |
437 | /* Note that Rb and Re may be SP_REGNUM. DO NOT count it in. */ | |
438 | if (REGNO (Rb) == SP_REGNUM && REGNO (Re) == SP_REGNUM) | |
439 | num_use_regs = extra_count; | |
440 | else | |
441 | num_use_regs = REGNO (Re) - REGNO (Rb) + 1 + extra_count; | |
442 | ||
443 | /* In addition to used registers, | |
444 | we need one more space for (set sp sp-x) rtx. */ | |
445 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
446 | rtvec_alloc (num_use_regs + 1)); | |
447 | par_index = 0; | |
448 | ||
449 | /* Initialize offset and start to create push behavior. */ | |
450 | offset = -(num_use_regs * 4); | |
451 | ||
452 | /* Create (set mem regX) from Rb, Rb+1 up to Re. */ | |
453 | for (regno = REGNO (Rb); regno <= (int) REGNO (Re); regno++) | |
454 | { | |
455 | /* Rb and Re may be SP_REGNUM. | |
456 | We need to break this loop immediately. */ | |
457 | if (regno == SP_REGNUM) | |
4e9a2848 | 458 | break; |
9304f876 CJW |
459 | |
460 | reg = gen_rtx_REG (SImode, regno); | |
461 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
462 | stack_pointer_rtx, | |
463 | offset)); | |
464 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
465 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
466 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
467 | offset = offset + 4; | |
468 | par_index++; | |
469 | } | |
470 | ||
471 | /* Create (set mem fp), (set mem gp), and (set mem lp) if necessary. */ | |
4e9a2848 | 472 | if (save_fp) |
9304f876 CJW |
473 | { |
474 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
475 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
476 | stack_pointer_rtx, | |
477 | offset)); | |
478 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
479 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
480 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
481 | offset = offset + 4; | |
482 | par_index++; | |
483 | } | |
4e9a2848 | 484 | if (save_gp) |
9304f876 CJW |
485 | { |
486 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
487 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
488 | stack_pointer_rtx, | |
489 | offset)); | |
490 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
491 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
492 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
493 | offset = offset + 4; | |
494 | par_index++; | |
495 | } | |
4e9a2848 | 496 | if (save_lp) |
9304f876 CJW |
497 | { |
498 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
499 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
500 | stack_pointer_rtx, | |
501 | offset)); | |
502 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
503 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
504 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
505 | offset = offset + 4; | |
506 | par_index++; | |
507 | } | |
508 | ||
509 | /* Create (set sp sp-x). */ | |
510 | ||
511 | /* We need to re-calculate the offset value again for adjustment. */ | |
512 | offset = -(num_use_regs * 4); | |
513 | adjust_sp_rtx | |
514 | = gen_rtx_SET (VOIDmode, | |
515 | stack_pointer_rtx, | |
516 | plus_constant (Pmode, stack_pointer_rtx, offset)); | |
517 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
518 | RTX_FRAME_RELATED_P (adjust_sp_rtx) = 1; | |
519 | ||
4e9a2848 CJW |
520 | parallel_insn = emit_insn (parallel_insn); |
521 | ||
522 | /* The insn rtx 'parallel_insn' will change frame layout. | |
523 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
524 | generate CFI (Call Frame Information) stuff. */ | |
525 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
47e0e7d2 CJW |
526 | |
527 | /* Don't use GCC's logic for CFI info if we are generate a push for VAARG | |
528 | since we will not restore those register at epilogue. */ | |
529 | if (vaarg_p) | |
530 | { | |
531 | dwarf = alloc_reg_note (REG_CFA_ADJUST_CFA, | |
532 | copy_rtx (adjust_sp_rtx), NULL_RTX); | |
533 | REG_NOTES (parallel_insn) = dwarf; | |
534 | } | |
9304f876 CJW |
535 | } |
536 | ||
537 | /* Function to create a parallel rtx pattern | |
538 | which presents stack pop multiple behavior. | |
539 | The overall concept are: | |
540 | "pop registers from memory", | |
541 | "adjust stack pointer". */ | |
4e9a2848 CJW |
542 | static void |
543 | nds32_emit_stack_pop_multiple (rtx Rb, rtx Re, rtx En4) | |
9304f876 CJW |
544 | { |
545 | int regno; | |
546 | int extra_count; | |
547 | int num_use_regs; | |
548 | int par_index; | |
549 | int offset; | |
4e9a2848 | 550 | int save_fp, save_gp, save_lp; |
9304f876 CJW |
551 | |
552 | rtx reg; | |
553 | rtx mem; | |
554 | rtx pop_rtx; | |
555 | rtx adjust_sp_rtx; | |
556 | rtx parallel_insn; | |
4e9a2848 | 557 | rtx dwarf = NULL_RTX; |
9304f876 CJW |
558 | |
559 | /* We need to provide a customized rtx which contains | |
560 | necessary information for data analysis, | |
561 | so we create a parallel rtx like this: | |
562 | (parallel [(set (reg:SI Rb) | |
563 | (mem (reg:SI SP_REGNUM))) | |
564 | (set (reg:SI Rb+1) | |
565 | (mem (plus (reg:SI SP_REGNUM) (const_int 4)))) | |
566 | ... | |
567 | (set (reg:SI Re) | |
568 | (mem (plus (reg:SI SP_REGNUM) (const_int 16)))) | |
569 | (set (reg:SI FP_REGNUM) | |
570 | (mem (plus (reg:SI SP_REGNUM) (const_int 20)))) | |
571 | (set (reg:SI GP_REGNUM) | |
572 | (mem (plus (reg:SI SP_REGNUM) (const_int 24)))) | |
573 | (set (reg:SI LP_REGNUM) | |
574 | (mem (plus (reg:SI SP_REGNUM) (const_int 28)))) | |
575 | (set (reg:SI SP_REGNUM) | |
576 | (plus (reg:SI SP_REGNUM) (const_int 32)))]) */ | |
577 | ||
4e9a2848 CJW |
578 | /* Determine whether we need to restore $fp, $gp, or $lp. */ |
579 | save_fp = INTVAL (En4) & 0x8; | |
580 | save_gp = INTVAL (En4) & 0x4; | |
581 | save_lp = INTVAL (En4) & 0x2; | |
582 | ||
9304f876 CJW |
583 | /* Calculate the number of registers that will be poped. */ |
584 | extra_count = 0; | |
4e9a2848 | 585 | if (save_fp) |
9304f876 | 586 | extra_count++; |
4e9a2848 | 587 | if (save_gp) |
9304f876 | 588 | extra_count++; |
4e9a2848 | 589 | if (save_lp) |
9304f876 CJW |
590 | extra_count++; |
591 | /* Note that Rb and Re may be SP_REGNUM. DO NOT count it in. */ | |
592 | if (REGNO (Rb) == SP_REGNUM && REGNO (Re) == SP_REGNUM) | |
593 | num_use_regs = extra_count; | |
594 | else | |
595 | num_use_regs = REGNO (Re) - REGNO (Rb) + 1 + extra_count; | |
596 | ||
597 | /* In addition to used registers, | |
598 | we need one more space for (set sp sp+x) rtx. */ | |
599 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
600 | rtvec_alloc (num_use_regs + 1)); | |
601 | par_index = 0; | |
602 | ||
603 | /* Initialize offset and start to create pop behavior. */ | |
604 | offset = 0; | |
605 | ||
606 | /* Create (set regX mem) from Rb, Rb+1 up to Re. */ | |
607 | for (regno = REGNO (Rb); regno <= (int) REGNO (Re); regno++) | |
608 | { | |
609 | /* Rb and Re may be SP_REGNUM. | |
610 | We need to break this loop immediately. */ | |
611 | if (regno == SP_REGNUM) | |
4e9a2848 | 612 | break; |
9304f876 CJW |
613 | |
614 | reg = gen_rtx_REG (SImode, regno); | |
615 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
616 | stack_pointer_rtx, | |
617 | offset)); | |
618 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
619 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
620 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
621 | offset = offset + 4; | |
622 | par_index++; | |
4e9a2848 CJW |
623 | |
624 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
9304f876 CJW |
625 | } |
626 | ||
627 | /* Create (set fp mem), (set gp mem), and (set lp mem) if necessary. */ | |
4e9a2848 | 628 | if (save_fp) |
9304f876 CJW |
629 | { |
630 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
631 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
632 | stack_pointer_rtx, | |
633 | offset)); | |
634 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
635 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
636 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
637 | offset = offset + 4; | |
638 | par_index++; | |
4e9a2848 CJW |
639 | |
640 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
9304f876 | 641 | } |
4e9a2848 | 642 | if (save_gp) |
9304f876 CJW |
643 | { |
644 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
645 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
646 | stack_pointer_rtx, | |
647 | offset)); | |
648 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
649 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
650 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
651 | offset = offset + 4; | |
652 | par_index++; | |
4e9a2848 CJW |
653 | |
654 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
9304f876 | 655 | } |
4e9a2848 | 656 | if (save_lp) |
9304f876 CJW |
657 | { |
658 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
659 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
660 | stack_pointer_rtx, | |
661 | offset)); | |
662 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
663 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
664 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
665 | offset = offset + 4; | |
666 | par_index++; | |
4e9a2848 CJW |
667 | |
668 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
9304f876 CJW |
669 | } |
670 | ||
671 | /* Create (set sp sp+x). */ | |
672 | ||
673 | /* The offset value is already in place. No need to re-calculate it. */ | |
674 | adjust_sp_rtx | |
675 | = gen_rtx_SET (VOIDmode, | |
676 | stack_pointer_rtx, | |
677 | plus_constant (Pmode, stack_pointer_rtx, offset)); | |
678 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
9304f876 | 679 | |
4e9a2848 CJW |
680 | /* Tell gcc we adjust SP in this insn. */ |
681 | dwarf = alloc_reg_note (REG_CFA_ADJUST_CFA, copy_rtx (adjust_sp_rtx), dwarf); | |
682 | ||
683 | parallel_insn = emit_insn (parallel_insn); | |
684 | ||
685 | /* The insn rtx 'parallel_insn' will change frame layout. | |
686 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
687 | generate CFI (Call Frame Information) stuff. */ | |
688 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
689 | ||
690 | /* Add CFI info by manual. */ | |
691 | REG_NOTES (parallel_insn) = dwarf; | |
9304f876 CJW |
692 | } |
693 | ||
694 | /* Function to create a parallel rtx pattern | |
695 | which presents stack v3push behavior. | |
696 | The overall concept are: | |
697 | "push registers to memory", | |
698 | "adjust stack pointer". */ | |
88437f39 CJW |
699 | static void |
700 | nds32_emit_stack_v3push (rtx Rb, | |
701 | rtx Re, | |
702 | rtx En4 ATTRIBUTE_UNUSED, | |
703 | rtx imm8u) | |
9304f876 CJW |
704 | { |
705 | int regno; | |
706 | int num_use_regs; | |
707 | int par_index; | |
708 | int offset; | |
709 | ||
710 | rtx reg; | |
711 | rtx mem; | |
712 | rtx push_rtx; | |
713 | rtx adjust_sp_rtx; | |
714 | rtx parallel_insn; | |
715 | ||
716 | /* We need to provide a customized rtx which contains | |
717 | necessary information for data analysis, | |
718 | so we create a parallel rtx like this: | |
88437f39 | 719 | (parallel [(set (mem (plus (reg:SI SP_REGNUM) (const_int -32))) |
9304f876 CJW |
720 | (reg:SI Rb)) |
721 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -28))) | |
722 | (reg:SI Rb+1)) | |
723 | ... | |
724 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -16))) | |
725 | (reg:SI Re)) | |
726 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -12))) | |
727 | (reg:SI FP_REGNUM)) | |
728 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -8))) | |
729 | (reg:SI GP_REGNUM)) | |
730 | (set (mem (plus (reg:SI SP_REGNUM) (const_int -4))) | |
731 | (reg:SI LP_REGNUM)) | |
732 | (set (reg:SI SP_REGNUM) | |
733 | (plus (reg:SI SP_REGNUM) (const_int -32-imm8u)))]) */ | |
734 | ||
735 | /* Calculate the number of registers that will be pushed. | |
736 | Since $fp, $gp, and $lp is always pushed with v3push instruction, | |
737 | we need to count these three registers. | |
738 | Under v3push, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
739 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
740 | num_use_regs = REGNO (Re) - REGNO (Rb) + 1 + 3; | |
741 | ||
742 | /* In addition to used registers, | |
743 | we need one more space for (set sp sp-x-imm8u) rtx. */ | |
744 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
745 | rtvec_alloc (num_use_regs + 1)); | |
746 | par_index = 0; | |
747 | ||
748 | /* Initialize offset and start to create push behavior. */ | |
749 | offset = -(num_use_regs * 4); | |
750 | ||
751 | /* Create (set mem regX) from Rb, Rb+1 up to Re. | |
752 | Under v3push, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
753 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
754 | for (regno = REGNO (Rb); regno <= (int) REGNO (Re); regno++) | |
755 | { | |
756 | reg = gen_rtx_REG (SImode, regno); | |
757 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
758 | stack_pointer_rtx, | |
759 | offset)); | |
760 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
761 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
762 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
763 | offset = offset + 4; | |
764 | par_index++; | |
765 | } | |
766 | ||
767 | /* Create (set mem fp). */ | |
768 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
769 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
770 | stack_pointer_rtx, | |
771 | offset)); | |
772 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
773 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
774 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
775 | offset = offset + 4; | |
776 | par_index++; | |
777 | /* Create (set mem gp). */ | |
778 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
779 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
780 | stack_pointer_rtx, | |
781 | offset)); | |
782 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
783 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
784 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
785 | offset = offset + 4; | |
786 | par_index++; | |
787 | /* Create (set mem lp). */ | |
788 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
789 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
790 | stack_pointer_rtx, | |
791 | offset)); | |
792 | push_rtx = gen_rtx_SET (VOIDmode, mem, reg); | |
793 | XVECEXP (parallel_insn, 0, par_index) = push_rtx; | |
794 | RTX_FRAME_RELATED_P (push_rtx) = 1; | |
795 | offset = offset + 4; | |
796 | par_index++; | |
797 | ||
798 | /* Create (set sp sp-x-imm8u). */ | |
799 | ||
800 | /* We need to re-calculate the offset value again for adjustment. */ | |
801 | offset = -(num_use_regs * 4); | |
802 | adjust_sp_rtx | |
803 | = gen_rtx_SET (VOIDmode, | |
804 | stack_pointer_rtx, | |
805 | plus_constant (Pmode, | |
806 | stack_pointer_rtx, | |
807 | offset - INTVAL (imm8u))); | |
808 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
809 | RTX_FRAME_RELATED_P (adjust_sp_rtx) = 1; | |
810 | ||
88437f39 CJW |
811 | parallel_insn = emit_insn (parallel_insn); |
812 | ||
813 | /* The insn rtx 'parallel_insn' will change frame layout. | |
814 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
815 | generate CFI (Call Frame Information) stuff. */ | |
816 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
9304f876 CJW |
817 | } |
818 | ||
819 | /* Function to create a parallel rtx pattern | |
820 | which presents stack v3pop behavior. | |
821 | The overall concept are: | |
822 | "pop registers from memory", | |
823 | "adjust stack pointer". */ | |
88437f39 CJW |
824 | static void |
825 | nds32_emit_stack_v3pop (rtx Rb, | |
826 | rtx Re, | |
827 | rtx En4 ATTRIBUTE_UNUSED, | |
828 | rtx imm8u) | |
9304f876 CJW |
829 | { |
830 | int regno; | |
831 | int num_use_regs; | |
832 | int par_index; | |
833 | int offset; | |
834 | ||
835 | rtx reg; | |
836 | rtx mem; | |
837 | rtx pop_rtx; | |
838 | rtx adjust_sp_rtx; | |
839 | rtx parallel_insn; | |
88437f39 | 840 | rtx dwarf = NULL_RTX; |
9304f876 CJW |
841 | |
842 | /* We need to provide a customized rtx which contains | |
843 | necessary information for data analysis, | |
844 | so we create a parallel rtx like this: | |
845 | (parallel [(set (reg:SI Rb) | |
846 | (mem (reg:SI SP_REGNUM))) | |
847 | (set (reg:SI Rb+1) | |
848 | (mem (plus (reg:SI SP_REGNUM) (const_int 4)))) | |
849 | ... | |
850 | (set (reg:SI Re) | |
851 | (mem (plus (reg:SI SP_REGNUM) (const_int 16)))) | |
852 | (set (reg:SI FP_REGNUM) | |
853 | (mem (plus (reg:SI SP_REGNUM) (const_int 20)))) | |
854 | (set (reg:SI GP_REGNUM) | |
855 | (mem (plus (reg:SI SP_REGNUM) (const_int 24)))) | |
856 | (set (reg:SI LP_REGNUM) | |
857 | (mem (plus (reg:SI SP_REGNUM) (const_int 28)))) | |
858 | (set (reg:SI SP_REGNUM) | |
859 | (plus (reg:SI SP_REGNUM) (const_int 32+imm8u)))]) */ | |
860 | ||
861 | /* Calculate the number of registers that will be poped. | |
862 | Since $fp, $gp, and $lp is always poped with v3pop instruction, | |
863 | we need to count these three registers. | |
864 | Under v3push, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
865 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
866 | num_use_regs = REGNO (Re) - REGNO (Rb) + 1 + 3; | |
867 | ||
868 | /* In addition to used registers, | |
869 | we need one more space for (set sp sp+x+imm8u) rtx. */ | |
870 | parallel_insn = gen_rtx_PARALLEL (VOIDmode, | |
871 | rtvec_alloc (num_use_regs + 1)); | |
872 | par_index = 0; | |
873 | ||
874 | /* Initialize offset and start to create pop behavior. */ | |
875 | offset = 0; | |
876 | ||
877 | /* Create (set regX mem) from Rb, Rb+1 up to Re. | |
878 | Under v3pop, Rb is $r6, while Re is $r6, $r8, $r10, or $r14. | |
879 | So there is no need to worry about Rb=Re=SP_REGNUM case. */ | |
880 | for (regno = REGNO (Rb); regno <= (int) REGNO (Re); regno++) | |
881 | { | |
882 | reg = gen_rtx_REG (SImode, regno); | |
883 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
884 | stack_pointer_rtx, | |
885 | offset)); | |
886 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
887 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
888 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
889 | offset = offset + 4; | |
890 | par_index++; | |
88437f39 CJW |
891 | |
892 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); | |
9304f876 CJW |
893 | } |
894 | ||
895 | /* Create (set fp mem). */ | |
896 | reg = gen_rtx_REG (SImode, FP_REGNUM); | |
897 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
898 | stack_pointer_rtx, | |
899 | offset)); | |
900 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
901 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
902 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
903 | offset = offset + 4; | |
904 | par_index++; | |
88437f39 CJW |
905 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); |
906 | ||
9304f876 CJW |
907 | /* Create (set gp mem). */ |
908 | reg = gen_rtx_REG (SImode, GP_REGNUM); | |
909 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
910 | stack_pointer_rtx, | |
911 | offset)); | |
912 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
913 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
914 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
915 | offset = offset + 4; | |
916 | par_index++; | |
88437f39 CJW |
917 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); |
918 | ||
9304f876 CJW |
919 | /* Create (set lp mem ). */ |
920 | reg = gen_rtx_REG (SImode, LP_REGNUM); | |
921 | mem = gen_frame_mem (SImode, plus_constant (Pmode, | |
922 | stack_pointer_rtx, | |
923 | offset)); | |
924 | pop_rtx = gen_rtx_SET (VOIDmode, reg, mem); | |
925 | XVECEXP (parallel_insn, 0, par_index) = pop_rtx; | |
926 | RTX_FRAME_RELATED_P (pop_rtx) = 1; | |
927 | offset = offset + 4; | |
928 | par_index++; | |
88437f39 | 929 | dwarf = alloc_reg_note (REG_CFA_RESTORE, reg, dwarf); |
9304f876 CJW |
930 | |
931 | /* Create (set sp sp+x+imm8u). */ | |
932 | ||
933 | /* The offset value is already in place. No need to re-calculate it. */ | |
934 | adjust_sp_rtx | |
935 | = gen_rtx_SET (VOIDmode, | |
936 | stack_pointer_rtx, | |
937 | plus_constant (Pmode, | |
938 | stack_pointer_rtx, | |
939 | offset + INTVAL (imm8u))); | |
940 | XVECEXP (parallel_insn, 0, par_index) = adjust_sp_rtx; | |
9304f876 | 941 | |
88437f39 CJW |
942 | /* Tell gcc we adjust SP in this insn. */ |
943 | dwarf = alloc_reg_note (REG_CFA_ADJUST_CFA, copy_rtx (adjust_sp_rtx), dwarf); | |
944 | ||
945 | parallel_insn = emit_insn (parallel_insn); | |
946 | ||
947 | /* The insn rtx 'parallel_insn' will change frame layout. | |
948 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
949 | generate CFI (Call Frame Information) stuff. */ | |
950 | RTX_FRAME_RELATED_P (parallel_insn) = 1; | |
951 | ||
952 | /* Add CFI info by manual. */ | |
953 | REG_NOTES (parallel_insn) = dwarf; | |
9304f876 CJW |
954 | } |
955 | ||
9304f876 CJW |
956 | /* Function that may creates more instructions |
957 | for large value on adjusting stack pointer. | |
958 | ||
959 | In nds32 target, 'addi' can be used for stack pointer | |
960 | adjustment in prologue/epilogue stage. | |
961 | However, sometimes there are too many local variables so that | |
962 | the adjustment value is not able to be fit in the 'addi' instruction. | |
963 | One solution is to move value into a register | |
964 | and then use 'add' instruction. | |
965 | In practice, we use TA_REGNUM ($r15) to accomplish this purpose. | |
966 | Also, we need to return zero for sp adjustment so that | |
967 | proglogue/epilogue knows there is no need to create 'addi' instruction. */ | |
968 | static int | |
969 | nds32_force_addi_stack_int (int full_value) | |
970 | { | |
971 | int adjust_value; | |
972 | ||
973 | rtx tmp_reg; | |
974 | rtx sp_adjust_insn; | |
975 | ||
976 | if (!satisfies_constraint_Is15 (GEN_INT (full_value))) | |
977 | { | |
978 | /* The value is not able to fit in single addi instruction. | |
979 | Create more instructions of moving value into a register | |
980 | and then add stack pointer with it. */ | |
981 | ||
982 | /* $r15 is going to be temporary register to hold the value. */ | |
983 | tmp_reg = gen_rtx_REG (SImode, TA_REGNUM); | |
984 | ||
985 | /* Create one more instruction to move value | |
986 | into the temporary register. */ | |
987 | emit_move_insn (tmp_reg, GEN_INT (full_value)); | |
988 | ||
989 | /* Create new 'add' rtx. */ | |
990 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
991 | stack_pointer_rtx, | |
992 | tmp_reg); | |
993 | /* Emit rtx into insn list and receive its transformed insn rtx. */ | |
994 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
995 | ||
996 | /* At prologue, we need to tell GCC that this is frame related insn, | |
997 | so that we can consider this instruction to output debug information. | |
998 | If full_value is NEGATIVE, it means this function | |
999 | is invoked by expand_prologue. */ | |
1000 | if (full_value < 0) | |
1001 | { | |
1002 | /* Because (tmp_reg <- full_value) may be split into two | |
1003 | rtl patterns, we can not set its RTX_FRAME_RELATED_P. | |
1004 | We need to construct another (sp <- sp + full_value) | |
1005 | and then insert it into sp_adjust_insn's reg note to | |
1006 | represent a frame related expression. | |
1007 | GCC knows how to refer it and output debug information. */ | |
1008 | ||
1009 | rtx plus_rtx; | |
1010 | rtx set_rtx; | |
1011 | ||
1012 | plus_rtx = plus_constant (Pmode, stack_pointer_rtx, full_value); | |
1013 | set_rtx = gen_rtx_SET (VOIDmode, stack_pointer_rtx, plus_rtx); | |
1014 | add_reg_note (sp_adjust_insn, REG_FRAME_RELATED_EXPR, set_rtx); | |
1015 | ||
1016 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
1017 | } | |
1018 | ||
1019 | /* We have used alternative way to adjust stack pointer value. | |
1020 | Return zero so that prologue/epilogue | |
1021 | will not generate other instructions. */ | |
1022 | return 0; | |
1023 | } | |
1024 | else | |
1025 | { | |
1026 | /* The value is able to fit in addi instruction. | |
1027 | However, remember to make it to be positive value | |
1028 | because we want to return 'adjustment' result. */ | |
1029 | adjust_value = (full_value < 0) ? (-full_value) : (full_value); | |
1030 | ||
1031 | return adjust_value; | |
1032 | } | |
1033 | } | |
1034 | ||
1035 | /* Return true if MODE/TYPE need double word alignment. */ | |
1036 | static bool | |
ef4bddc2 | 1037 | nds32_needs_double_word_align (machine_mode mode, const_tree type) |
9304f876 CJW |
1038 | { |
1039 | unsigned int align; | |
1040 | ||
634bdae9 CJW |
1041 | /* Pick up the alignment according to the mode or type. */ |
1042 | align = NDS32_MODE_TYPE_ALIGN (mode, type); | |
9304f876 CJW |
1043 | |
1044 | return (align > PARM_BOUNDARY); | |
1045 | } | |
1046 | ||
1047 | /* Return true if FUNC is a naked function. */ | |
810f736f CJW |
1048 | static bool |
1049 | nds32_naked_function_p (tree func) | |
9304f876 CJW |
1050 | { |
1051 | tree t; | |
1052 | ||
1053 | if (TREE_CODE (func) != FUNCTION_DECL) | |
1054 | abort (); | |
1055 | ||
1056 | t = lookup_attribute ("naked", DECL_ATTRIBUTES (func)); | |
1057 | ||
1058 | return (t != NULL_TREE); | |
1059 | } | |
1060 | ||
1061 | /* Function that check if 'X' is a valid address register. | |
1062 | The variable 'STRICT' is very important to | |
1063 | make decision for register number. | |
1064 | ||
1065 | STRICT : true | |
1066 | => We are in reload pass or after reload pass. | |
1067 | The register number should be strictly limited in general registers. | |
1068 | ||
1069 | STRICT : false | |
1070 | => Before reload pass, we are free to use any register number. */ | |
1071 | static bool | |
1072 | nds32_address_register_rtx_p (rtx x, bool strict) | |
1073 | { | |
1074 | int regno; | |
1075 | ||
1076 | if (GET_CODE (x) != REG) | |
1077 | return false; | |
1078 | ||
1079 | regno = REGNO (x); | |
1080 | ||
1081 | if (strict) | |
1082 | return REGNO_OK_FOR_BASE_P (regno); | |
1083 | else | |
1084 | return true; | |
1085 | } | |
1086 | ||
1087 | /* Function that check if 'INDEX' is valid to be a index rtx for address. | |
1088 | ||
1089 | OUTER_MODE : Machine mode of outer address rtx. | |
1090 | INDEX : Check if this rtx is valid to be a index for address. | |
1091 | STRICT : If it is true, we are in reload pass or after reload pass. */ | |
1092 | static bool | |
ef4bddc2 | 1093 | nds32_legitimate_index_p (machine_mode outer_mode, |
9304f876 CJW |
1094 | rtx index, |
1095 | bool strict) | |
1096 | { | |
1097 | int regno; | |
1098 | rtx op0; | |
1099 | rtx op1; | |
1100 | ||
1101 | switch (GET_CODE (index)) | |
1102 | { | |
1103 | case REG: | |
1104 | regno = REGNO (index); | |
1105 | /* If we are in reload pass or after reload pass, | |
1106 | we need to limit it to general register. */ | |
1107 | if (strict) | |
1108 | return REGNO_OK_FOR_INDEX_P (regno); | |
1109 | else | |
1110 | return true; | |
1111 | ||
1112 | case CONST_INT: | |
1113 | /* The alignment of the integer value is determined by 'outer_mode'. */ | |
1114 | if (GET_MODE_SIZE (outer_mode) == 1) | |
1115 | { | |
1116 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1117 | if (!satisfies_constraint_Is15 (index)) | |
1118 | return false; | |
1119 | ||
1120 | /* Pass all test, the value is valid, return true. */ | |
1121 | return true; | |
1122 | } | |
1123 | if (GET_MODE_SIZE (outer_mode) == 2 | |
1124 | && NDS32_HALF_WORD_ALIGN_P (INTVAL (index))) | |
1125 | { | |
1126 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1127 | if (!satisfies_constraint_Is16 (index)) | |
1128 | return false; | |
1129 | ||
1130 | /* Pass all test, the value is valid, return true. */ | |
1131 | return true; | |
1132 | } | |
1133 | if (GET_MODE_SIZE (outer_mode) == 4 | |
1134 | && NDS32_SINGLE_WORD_ALIGN_P (INTVAL (index))) | |
1135 | { | |
1136 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1137 | if (!satisfies_constraint_Is17 (index)) | |
1138 | return false; | |
1139 | ||
1140 | /* Pass all test, the value is valid, return true. */ | |
1141 | return true; | |
1142 | } | |
1143 | if (GET_MODE_SIZE (outer_mode) == 8 | |
1144 | && NDS32_SINGLE_WORD_ALIGN_P (INTVAL (index))) | |
1145 | { | |
1146 | /* Further check if the value is legal for the 'outer_mode'. */ | |
1147 | if (!satisfies_constraint_Is17 (gen_int_mode (INTVAL (index) + 4, | |
1148 | SImode))) | |
1149 | return false; | |
1150 | ||
1151 | /* Pass all test, the value is valid, return true. */ | |
1152 | return true; | |
1153 | } | |
1154 | ||
1155 | return false; | |
1156 | ||
1157 | case MULT: | |
1158 | op0 = XEXP (index, 0); | |
1159 | op1 = XEXP (index, 1); | |
1160 | ||
1161 | if (REG_P (op0) && CONST_INT_P (op1)) | |
1162 | { | |
1163 | int multiplier; | |
1164 | multiplier = INTVAL (op1); | |
1165 | ||
1166 | /* We only allow (mult reg const_int_1) | |
1167 | or (mult reg const_int_2) or (mult reg const_int_4). */ | |
1168 | if (multiplier != 1 && multiplier != 2 && multiplier != 4) | |
1169 | return false; | |
1170 | ||
1171 | regno = REGNO (op0); | |
1172 | /* Limit it in general registers if we are | |
1173 | in reload pass or after reload pass. */ | |
1174 | if(strict) | |
1175 | return REGNO_OK_FOR_INDEX_P (regno); | |
1176 | else | |
1177 | return true; | |
1178 | } | |
1179 | ||
1180 | return false; | |
1181 | ||
1182 | case ASHIFT: | |
1183 | op0 = XEXP (index, 0); | |
1184 | op1 = XEXP (index, 1); | |
1185 | ||
1186 | if (REG_P (op0) && CONST_INT_P (op1)) | |
1187 | { | |
1188 | int sv; | |
1189 | /* op1 is already the sv value for use to do left shift. */ | |
1190 | sv = INTVAL (op1); | |
1191 | ||
1192 | /* We only allow (ashift reg const_int_0) | |
1193 | or (ashift reg const_int_1) or (ashift reg const_int_2). */ | |
1194 | if (sv != 0 && sv != 1 && sv !=2) | |
1195 | return false; | |
1196 | ||
1197 | regno = REGNO (op0); | |
1198 | /* Limit it in general registers if we are | |
1199 | in reload pass or after reload pass. */ | |
1200 | if(strict) | |
1201 | return REGNO_OK_FOR_INDEX_P (regno); | |
1202 | else | |
1203 | return true; | |
1204 | } | |
1205 | ||
1206 | return false; | |
1207 | ||
1208 | default: | |
1209 | return false; | |
1210 | } | |
1211 | } | |
1212 | ||
9304f876 CJW |
1213 | /* ------------------------------------------------------------------------ */ |
1214 | ||
1215 | /* PART 3: Implement target hook stuff definitions. */ | |
1216 | \f | |
1217 | /* Register Classes. */ | |
1218 | ||
1219 | static unsigned char | |
1220 | nds32_class_max_nregs (reg_class_t rclass ATTRIBUTE_UNUSED, | |
ef4bddc2 | 1221 | machine_mode mode) |
9304f876 CJW |
1222 | { |
1223 | /* Return the maximum number of consecutive registers | |
1224 | needed to represent "mode" in a register of "rclass". */ | |
1225 | return ((GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD); | |
1226 | } | |
1227 | ||
1228 | static int | |
1229 | nds32_register_priority (int hard_regno) | |
1230 | { | |
1231 | /* Encourage to use r0-r7 for LRA when optimize for size. */ | |
1232 | if (optimize_size && hard_regno < 8) | |
1233 | return 4; | |
1234 | return 3; | |
1235 | } | |
1236 | ||
1237 | \f | |
1238 | /* Stack Layout and Calling Conventions. */ | |
1239 | ||
1240 | /* There are three kinds of pointer concepts using in GCC compiler: | |
1241 | ||
1242 | frame pointer: A pointer to the first location of local variables. | |
1243 | stack pointer: A pointer to the top of a stack frame. | |
1244 | argument pointer: A pointer to the incoming arguments. | |
1245 | ||
1246 | In nds32 target calling convention, we are using 8-byte alignment. | |
1247 | Besides, we would like to have each stack frame of a function includes: | |
1248 | ||
1249 | [Block A] | |
1250 | 1. previous hard frame pointer | |
1251 | 2. return address | |
1252 | 3. callee-saved registers | |
1253 | 4. <padding bytes> (we will calculte in nds32_compute_stack_frame() | |
1254 | and save it at | |
1255 | cfun->machine->callee_saved_area_padding_bytes) | |
1256 | ||
1257 | [Block B] | |
1258 | 1. local variables | |
1259 | 2. spilling location | |
1260 | 3. <padding bytes> (it will be calculated by GCC itself) | |
1261 | 4. incoming arguments | |
1262 | 5. <padding bytes> (it will be calculated by GCC itself) | |
1263 | ||
1264 | [Block C] | |
1265 | 1. <padding bytes> (it will be calculated by GCC itself) | |
1266 | 2. outgoing arguments | |
1267 | ||
1268 | We 'wrap' these blocks together with | |
1269 | hard frame pointer ($r28) and stack pointer ($r31). | |
1270 | By applying the basic frame/stack/argument pointers concept, | |
1271 | the layout of a stack frame shoule be like this: | |
1272 | ||
1273 | | | | |
1274 | old stack pointer -> ---- | |
1275 | | | \ | |
1276 | | | saved arguments for | |
1277 | | | vararg functions | |
1278 | | | / | |
1279 | hard frame pointer -> -- | |
1280 | & argument pointer | | \ | |
1281 | | | previous hardware frame pointer | |
1282 | | | return address | |
1283 | | | callee-saved registers | |
1284 | | | / | |
1285 | frame pointer -> -- | |
1286 | | | \ | |
1287 | | | local variables | |
1288 | | | and incoming arguments | |
1289 | | | / | |
1290 | -- | |
1291 | | | \ | |
1292 | | | outgoing | |
1293 | | | arguments | |
1294 | | | / | |
1295 | stack pointer -> ---- | |
1296 | ||
1297 | $SFP and $AP are used to represent frame pointer and arguments pointer, | |
1298 | which will be both eliminated as hard frame pointer. */ | |
1299 | ||
1300 | /* -- Eliminating Frame Pointer and Arg Pointer. */ | |
1301 | ||
19ac960a CJW |
1302 | static bool |
1303 | nds32_can_eliminate (const int from_reg, const int to_reg) | |
9304f876 CJW |
1304 | { |
1305 | if (from_reg == ARG_POINTER_REGNUM && to_reg == STACK_POINTER_REGNUM) | |
1306 | return true; | |
1307 | ||
1308 | if (from_reg == ARG_POINTER_REGNUM && to_reg == HARD_FRAME_POINTER_REGNUM) | |
1309 | return true; | |
1310 | ||
1311 | if (from_reg == FRAME_POINTER_REGNUM && to_reg == STACK_POINTER_REGNUM) | |
1312 | return true; | |
1313 | ||
1314 | if (from_reg == FRAME_POINTER_REGNUM && to_reg == HARD_FRAME_POINTER_REGNUM) | |
1315 | return true; | |
1316 | ||
1317 | return false; | |
1318 | } | |
1319 | ||
1320 | /* -- Passing Arguments in Registers. */ | |
1321 | ||
1322 | static rtx | |
ef4bddc2 | 1323 | nds32_function_arg (cumulative_args_t ca, machine_mode mode, |
9304f876 CJW |
1324 | const_tree type, bool named) |
1325 | { | |
7f6cd86b | 1326 | unsigned int regno; |
9304f876 CJW |
1327 | CUMULATIVE_ARGS *cum = get_cumulative_args (ca); |
1328 | ||
1329 | /* The last time this hook is called, | |
1330 | it is called with MODE == VOIDmode. */ | |
1331 | if (mode == VOIDmode) | |
1332 | return NULL_RTX; | |
1333 | ||
7f6cd86b | 1334 | /* For nameless arguments, we need to take care it individually. */ |
9304f876 | 1335 | if (!named) |
9304f876 | 1336 | { |
7f6cd86b CJW |
1337 | /* If we are under hard float abi, we have arguments passed on the |
1338 | stack and all situation can be handled by GCC itself. */ | |
1339 | if (TARGET_HARD_FLOAT) | |
1340 | return NULL_RTX; | |
1341 | ||
1342 | if (NDS32_ARG_PARTIAL_IN_GPR_REG_P (cum->gpr_offset, mode, type)) | |
1343 | { | |
1344 | /* If we still have enough registers to pass argument, pick up | |
1345 | next available register number. */ | |
1346 | regno | |
1347 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type); | |
1348 | return gen_rtx_REG (mode, regno); | |
1349 | } | |
634bdae9 | 1350 | |
9304f876 CJW |
1351 | /* No register available, return NULL_RTX. |
1352 | The compiler will use stack to pass argument instead. */ | |
1353 | return NULL_RTX; | |
1354 | } | |
7f6cd86b CJW |
1355 | |
1356 | /* The following is to handle named argument. | |
1357 | Note that the strategies of TARGET_HARD_FLOAT and !TARGET_HARD_FLOAT | |
1358 | are different. */ | |
1359 | if (TARGET_HARD_FLOAT) | |
1360 | { | |
1361 | /* Currently we have not implemented hard float yet. */ | |
1362 | gcc_unreachable (); | |
1363 | } | |
1364 | else | |
1365 | { | |
1366 | /* For !TARGET_HARD_FLOAT calling convention, we always use GPR to pass | |
1367 | argument. Since we allow to pass argument partially in registers, | |
1368 | we can just return it if there are still registers available. */ | |
1369 | if (NDS32_ARG_PARTIAL_IN_GPR_REG_P (cum->gpr_offset, mode, type)) | |
1370 | { | |
1371 | /* Pick up the next available register number. */ | |
1372 | regno | |
1373 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type); | |
1374 | return gen_rtx_REG (mode, regno); | |
1375 | } | |
1376 | ||
1377 | } | |
1378 | ||
1379 | /* No register available, return NULL_RTX. | |
1380 | The compiler will use stack to pass argument instead. */ | |
1381 | return NULL_RTX; | |
9304f876 CJW |
1382 | } |
1383 | ||
d40f3c40 | 1384 | static bool |
ef4bddc2 | 1385 | nds32_must_pass_in_stack (machine_mode mode, const_tree type) |
d40f3c40 CJW |
1386 | { |
1387 | /* Return true if a type must be passed in memory. | |
1388 | If it is NOT using hard float abi, small aggregates can be | |
1389 | passed in a register even we are calling a variadic function. | |
1390 | So there is no need to take padding into consideration. */ | |
1391 | if (TARGET_HARD_FLOAT) | |
1392 | return must_pass_in_stack_var_size_or_pad (mode, type); | |
1393 | else | |
1394 | return must_pass_in_stack_var_size (mode, type); | |
1395 | } | |
1396 | ||
650fc469 | 1397 | static int |
ef4bddc2 | 1398 | nds32_arg_partial_bytes (cumulative_args_t ca, machine_mode mode, |
650fc469 CJW |
1399 | tree type, bool named ATTRIBUTE_UNUSED) |
1400 | { | |
1401 | /* Returns the number of bytes at the beginning of an argument that | |
1402 | must be put in registers. The value must be zero for arguments that are | |
1403 | passed entirely in registers or that are entirely pushed on the stack. | |
1404 | Besides, TARGET_FUNCTION_ARG for these arguments should return the | |
1405 | first register to be used by the caller for this argument. */ | |
1406 | unsigned int needed_reg_count; | |
1407 | unsigned int remaining_reg_count; | |
1408 | CUMULATIVE_ARGS *cum; | |
1409 | ||
1410 | cum = get_cumulative_args (ca); | |
1411 | ||
1412 | /* Under hard float abi, we better have argument entirely passed in | |
1413 | registers or pushed on the stack so that we can reduce the complexity | |
1414 | of dealing with cum->gpr_offset and cum->fpr_offset. */ | |
1415 | if (TARGET_HARD_FLOAT) | |
1416 | return 0; | |
1417 | ||
1418 | /* If we have already runned out of argument registers, return zero | |
1419 | so that the argument will be entirely pushed on the stack. */ | |
1420 | if (NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1421 | >= NDS32_GPR_ARG_FIRST_REGNUM + NDS32_MAX_GPR_REGS_FOR_ARGS) | |
1422 | return 0; | |
1423 | ||
1424 | /* Calculate how many registers do we need for this argument. */ | |
1425 | needed_reg_count = NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1426 | ||
1427 | /* Calculate how many argument registers have left for passing argument. | |
1428 | Note that we should count it from next available register number. */ | |
1429 | remaining_reg_count | |
1430 | = NDS32_MAX_GPR_REGS_FOR_ARGS | |
1431 | - (NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1432 | - NDS32_GPR_ARG_FIRST_REGNUM); | |
1433 | ||
1434 | /* Note that we have to return the nubmer of bytes, not registers count. */ | |
1435 | if (needed_reg_count > remaining_reg_count) | |
1436 | return remaining_reg_count * UNITS_PER_WORD; | |
1437 | ||
1438 | return 0; | |
1439 | } | |
1440 | ||
9304f876 | 1441 | static void |
ef4bddc2 | 1442 | nds32_function_arg_advance (cumulative_args_t ca, machine_mode mode, |
9304f876 CJW |
1443 | const_tree type, bool named) |
1444 | { | |
ef4bddc2 | 1445 | machine_mode sub_mode; |
9304f876 CJW |
1446 | CUMULATIVE_ARGS *cum = get_cumulative_args (ca); |
1447 | ||
9304f876 CJW |
1448 | if (named) |
1449 | { | |
7f6cd86b CJW |
1450 | /* We need to further check TYPE and MODE so that we can determine |
1451 | which kind of register we shall advance. */ | |
1452 | if (type && TREE_CODE (type) == COMPLEX_TYPE) | |
1453 | sub_mode = TYPE_MODE (TREE_TYPE (type)); | |
1454 | else | |
1455 | sub_mode = mode; | |
1456 | ||
1457 | /* Under hard float abi, we may advance FPR registers. */ | |
1458 | if (TARGET_HARD_FLOAT && GET_MODE_CLASS (sub_mode) == MODE_FLOAT) | |
1459 | { | |
1460 | /* Currently we have not implemented hard float yet. */ | |
1461 | gcc_unreachable (); | |
1462 | } | |
1463 | else | |
1464 | { | |
1465 | cum->gpr_offset | |
1466 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1467 | - NDS32_GPR_ARG_FIRST_REGNUM | |
1468 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1469 | } | |
1470 | } | |
1471 | else | |
1472 | { | |
1473 | /* If this nameless argument is NOT under TARGET_HARD_FLOAT, | |
1474 | we can advance next register as well so that caller is | |
1475 | able to pass arguments in registers and callee must be | |
1476 | in charge of pushing all of them into stack. */ | |
1477 | if (!TARGET_HARD_FLOAT) | |
1478 | { | |
1479 | cum->gpr_offset | |
1480 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1481 | - NDS32_GPR_ARG_FIRST_REGNUM | |
1482 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1483 | } | |
9304f876 CJW |
1484 | } |
1485 | } | |
1486 | ||
1487 | static unsigned int | |
ef4bddc2 | 1488 | nds32_function_arg_boundary (machine_mode mode, const_tree type) |
9304f876 CJW |
1489 | { |
1490 | return (nds32_needs_double_word_align (mode, type) | |
1491 | ? NDS32_DOUBLE_WORD_ALIGNMENT | |
1492 | : PARM_BOUNDARY); | |
1493 | } | |
1494 | ||
1495 | /* -- How Scalar Function Values Are Returned. */ | |
1496 | ||
1497 | static rtx | |
1498 | nds32_function_value (const_tree ret_type, | |
1499 | const_tree fn_decl_or_type ATTRIBUTE_UNUSED, | |
1500 | bool outgoing ATTRIBUTE_UNUSED) | |
1501 | { | |
ef4bddc2 | 1502 | machine_mode mode; |
9304f876 CJW |
1503 | int unsignedp; |
1504 | ||
1505 | mode = TYPE_MODE (ret_type); | |
1506 | unsignedp = TYPE_UNSIGNED (ret_type); | |
1507 | ||
1508 | mode = promote_mode (ret_type, mode, &unsignedp); | |
1509 | ||
1510 | return gen_rtx_REG (mode, NDS32_GPR_RET_FIRST_REGNUM); | |
1511 | } | |
1512 | ||
1513 | static rtx | |
ef4bddc2 | 1514 | nds32_libcall_value (machine_mode mode, |
9304f876 CJW |
1515 | const_rtx fun ATTRIBUTE_UNUSED) |
1516 | { | |
1517 | return gen_rtx_REG (mode, NDS32_GPR_RET_FIRST_REGNUM); | |
1518 | } | |
1519 | ||
1520 | static bool | |
1521 | nds32_function_value_regno_p (const unsigned int regno) | |
1522 | { | |
1523 | return (regno == NDS32_GPR_RET_FIRST_REGNUM); | |
1524 | } | |
1525 | ||
1526 | /* -- Function Entry and Exit. */ | |
1527 | ||
1528 | /* The content produced from this function | |
1529 | will be placed before prologue body. */ | |
1530 | static void | |
1531 | nds32_asm_function_prologue (FILE *file, | |
1532 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
1533 | { | |
1534 | int r; | |
1535 | const char *func_name; | |
1536 | tree attrs; | |
1537 | tree name; | |
1538 | ||
1539 | /* All stack frame information is supposed to be | |
1540 | already computed when expanding prologue. | |
1541 | The result is in cfun->machine. | |
1542 | DO NOT call nds32_compute_stack_frame() here | |
1543 | because it may corrupt the essential information. */ | |
1544 | ||
1545 | fprintf (file, "\t! BEGIN PROLOGUE\n"); | |
1546 | fprintf (file, "\t! fp needed: %d\n", frame_pointer_needed); | |
1547 | fprintf (file, "\t! pretend_args: %d\n", cfun->machine->va_args_size); | |
1548 | fprintf (file, "\t! local_size: %d\n", cfun->machine->local_size); | |
1549 | fprintf (file, "\t! out_args_size: %d\n", cfun->machine->out_args_size); | |
1550 | ||
1551 | /* Use df_regs_ever_live_p() to detect if the register | |
1552 | is ever used in the current function. */ | |
1553 | fprintf (file, "\t! registers ever_live: "); | |
1554 | for (r = 0; r < 32; r++) | |
1555 | { | |
1556 | if (df_regs_ever_live_p (r)) | |
1557 | fprintf (file, "%s, ", reg_names[r]); | |
1558 | } | |
1559 | fputc ('\n', file); | |
1560 | ||
1561 | /* Display the attributes of this function. */ | |
1562 | fprintf (file, "\t! function attributes: "); | |
f2dafb91 CJW |
1563 | /* Get the attributes tree list. |
1564 | Note that GCC builds attributes list with reverse order. */ | |
1565 | attrs = DECL_ATTRIBUTES (current_function_decl); | |
9304f876 CJW |
1566 | |
1567 | /* If there is no any attribute, print out "None". */ | |
1568 | if (!attrs) | |
1569 | fprintf (file, "None"); | |
1570 | ||
1571 | /* If there are some attributes, try if we need to | |
1572 | construct isr vector information. */ | |
1573 | func_name = IDENTIFIER_POINTER (DECL_NAME (current_function_decl)); | |
1574 | nds32_construct_isr_vectors_information (attrs, func_name); | |
1575 | ||
1576 | /* Display all attributes of this function. */ | |
1577 | while (attrs) | |
1578 | { | |
1579 | name = TREE_PURPOSE (attrs); | |
1580 | fprintf (file, "%s ", IDENTIFIER_POINTER (name)); | |
1581 | ||
1582 | /* Pick up the next attribute. */ | |
1583 | attrs = TREE_CHAIN (attrs); | |
1584 | } | |
1585 | fputc ('\n', file); | |
1586 | } | |
1587 | ||
1588 | /* After rtl prologue has been expanded, this function is used. */ | |
1589 | static void | |
1590 | nds32_asm_function_end_prologue (FILE *file) | |
1591 | { | |
1592 | fprintf (file, "\t! END PROLOGUE\n"); | |
1593 | ||
1594 | /* If frame pointer is NOT needed and -mfp-as-gp is issued, | |
1595 | we can generate special directive: ".omit_fp_begin" | |
1596 | to guide linker doing fp-as-gp optimization. | |
1597 | However, for a naked function, which means | |
1598 | it should not have prologue/epilogue, | |
1599 | using fp-as-gp still requires saving $fp by push/pop behavior and | |
1600 | there is no benefit to use fp-as-gp on such small function. | |
1601 | So we need to make sure this function is NOT naked as well. */ | |
1602 | if (!frame_pointer_needed | |
1603 | && !cfun->machine->naked_p | |
1604 | && cfun->machine->fp_as_gp_p) | |
1605 | { | |
1606 | fprintf (file, "\t! ----------------------------------------\n"); | |
1607 | fprintf (file, "\t! Guide linker to do " | |
1608 | "link time optimization: fp-as-gp\n"); | |
1609 | fprintf (file, "\t! We add one more instruction to " | |
1610 | "initialize $fp near to $gp location.\n"); | |
1611 | fprintf (file, "\t! If linker fails to use fp-as-gp transformation,\n"); | |
1612 | fprintf (file, "\t! this extra instruction should be " | |
1613 | "eliminated at link stage.\n"); | |
1614 | fprintf (file, "\t.omit_fp_begin\n"); | |
1615 | fprintf (file, "\tla\t$fp,_FP_BASE_\n"); | |
1616 | fprintf (file, "\t! ----------------------------------------\n"); | |
1617 | } | |
1618 | } | |
1619 | ||
1620 | /* Before rtl epilogue has been expanded, this function is used. */ | |
1621 | static void | |
1622 | nds32_asm_function_begin_epilogue (FILE *file) | |
1623 | { | |
1624 | /* If frame pointer is NOT needed and -mfp-as-gp is issued, | |
1625 | we can generate special directive: ".omit_fp_end" | |
1626 | to claim fp-as-gp optimization range. | |
1627 | However, for a naked function, | |
1628 | which means it should not have prologue/epilogue, | |
1629 | using fp-as-gp still requires saving $fp by push/pop behavior and | |
1630 | there is no benefit to use fp-as-gp on such small function. | |
1631 | So we need to make sure this function is NOT naked as well. */ | |
1632 | if (!frame_pointer_needed | |
1633 | && !cfun->machine->naked_p | |
1634 | && cfun->machine->fp_as_gp_p) | |
1635 | { | |
1636 | fprintf (file, "\t! ----------------------------------------\n"); | |
1637 | fprintf (file, "\t! Claim the range of fp-as-gp " | |
1638 | "link time optimization\n"); | |
1639 | fprintf (file, "\t.omit_fp_end\n"); | |
1640 | fprintf (file, "\t! ----------------------------------------\n"); | |
1641 | } | |
1642 | ||
1643 | fprintf (file, "\t! BEGIN EPILOGUE\n"); | |
1644 | } | |
1645 | ||
1646 | /* The content produced from this function | |
1647 | will be placed after epilogue body. */ | |
1648 | static void | |
1649 | nds32_asm_function_epilogue (FILE *file, | |
1650 | HOST_WIDE_INT size ATTRIBUTE_UNUSED) | |
1651 | { | |
1652 | fprintf (file, "\t! END EPILOGUE\n"); | |
1653 | } | |
1654 | ||
1655 | static void | |
1656 | nds32_asm_output_mi_thunk (FILE *file, tree thunk ATTRIBUTE_UNUSED, | |
1657 | HOST_WIDE_INT delta, | |
1658 | HOST_WIDE_INT vcall_offset ATTRIBUTE_UNUSED, | |
1659 | tree function) | |
1660 | { | |
1661 | int this_regno; | |
1662 | ||
1663 | /* Make sure unwind info is emitted for the thunk if needed. */ | |
1664 | final_start_function (emit_barrier (), file, 1); | |
1665 | ||
1666 | this_regno = (aggregate_value_p (TREE_TYPE (TREE_TYPE (function)), function) | |
1667 | ? 1 | |
1668 | : 0); | |
1669 | ||
1670 | if (delta != 0) | |
1671 | { | |
1672 | if (satisfies_constraint_Is15 (GEN_INT (delta))) | |
1673 | { | |
1674 | fprintf (file, "\taddi\t$r%d, $r%d, %ld\n", | |
1675 | this_regno, this_regno, delta); | |
1676 | } | |
1677 | else if (satisfies_constraint_Is20 (GEN_INT (delta))) | |
1678 | { | |
1679 | fprintf (file, "\tmovi\t$ta, %ld\n", delta); | |
1680 | fprintf (file, "\tadd\t$r%d, $r%d, $ta\n", this_regno, this_regno); | |
1681 | } | |
1682 | else | |
1683 | { | |
1684 | fprintf (file, "\tsethi\t$ta, hi20(%ld)\n", delta); | |
1685 | fprintf (file, "\tori\t$ta, $ta, lo12(%ld)\n", delta); | |
1686 | fprintf (file, "\tadd\t$r%d, $r%d, $ta\n", this_regno, this_regno); | |
1687 | } | |
1688 | } | |
1689 | ||
1690 | fprintf (file, "\tb\t"); | |
1691 | assemble_name (file, XSTR (XEXP (DECL_RTL (function), 0), 0)); | |
1692 | fprintf (file, "\n"); | |
1693 | ||
1694 | final_end_function (); | |
1695 | } | |
1696 | ||
1697 | /* -- Permitting tail calls. */ | |
1698 | ||
1699 | /* Determine whether we need to enable warning for function return check. */ | |
1700 | static bool | |
1701 | nds32_warn_func_return (tree decl) | |
1702 | { | |
1703 | /* Naked functions are implemented entirely in assembly, including the | |
1704 | return sequence, so suppress warnings about this. */ | |
1705 | return !nds32_naked_function_p (decl); | |
1706 | } | |
1707 | ||
1708 | \f | |
1709 | /* Implementing the Varargs Macros. */ | |
1710 | ||
d4a6a4d9 CJW |
1711 | static void |
1712 | nds32_setup_incoming_varargs (cumulative_args_t ca, | |
ef4bddc2 | 1713 | machine_mode mode, |
d4a6a4d9 CJW |
1714 | tree type, |
1715 | int *pretend_args_size, | |
1716 | int second_time ATTRIBUTE_UNUSED) | |
1717 | { | |
1718 | unsigned int total_args_regs; | |
1719 | unsigned int num_of_used_regs; | |
1720 | unsigned int remaining_reg_count; | |
1721 | CUMULATIVE_ARGS *cum; | |
1722 | ||
1723 | /* If we are under hard float abi, we do not need to set *pretend_args_size. | |
1724 | So that all nameless arguments are pushed by caller and all situation | |
1725 | can be handled by GCC itself. */ | |
1726 | if (TARGET_HARD_FLOAT) | |
1727 | return; | |
1728 | ||
1729 | /* We are using NDS32_MAX_GPR_REGS_FOR_ARGS registers, | |
1730 | counting from NDS32_GPR_ARG_FIRST_REGNUM, for saving incoming arguments. | |
1731 | However, for nameless(anonymous) arguments, we should push them on the | |
1732 | stack so that all the nameless arguments appear to have been passed | |
1733 | consecutively in the memory for accessing. Hence, we need to check and | |
1734 | exclude the registers that are used for named arguments. */ | |
1735 | ||
1736 | cum = get_cumulative_args (ca); | |
1737 | ||
1738 | /* The MODE and TYPE describe the last argument. | |
1739 | We need those information to determine the remaining registers | |
1740 | for varargs. */ | |
1741 | total_args_regs | |
1742 | = NDS32_MAX_GPR_REGS_FOR_ARGS + NDS32_GPR_ARG_FIRST_REGNUM; | |
1743 | num_of_used_regs | |
1744 | = NDS32_AVAILABLE_REGNUM_FOR_GPR_ARG (cum->gpr_offset, mode, type) | |
1745 | + NDS32_NEED_N_REGS_FOR_ARG (mode, type); | |
1746 | ||
1747 | remaining_reg_count = total_args_regs - num_of_used_regs; | |
1748 | *pretend_args_size = remaining_reg_count * UNITS_PER_WORD; | |
1749 | ||
1750 | return; | |
1751 | } | |
1752 | ||
9304f876 CJW |
1753 | static bool |
1754 | nds32_strict_argument_naming (cumulative_args_t ca ATTRIBUTE_UNUSED) | |
1755 | { | |
d4a6a4d9 CJW |
1756 | /* If this hook returns true, the named argument of FUNCTION_ARG is always |
1757 | true for named arguments, and false for unnamed arguments. */ | |
9304f876 CJW |
1758 | return true; |
1759 | } | |
1760 | ||
1761 | \f | |
1762 | /* Trampolines for Nested Functions. */ | |
1763 | ||
1764 | static void | |
1765 | nds32_asm_trampoline_template (FILE *f) | |
1766 | { | |
1767 | if (TARGET_REDUCED_REGS) | |
1768 | { | |
1769 | /* Trampoline is not supported on reduced-set registers yet. */ | |
1770 | sorry ("a nested function is not supported for reduced registers"); | |
1771 | } | |
1772 | else | |
1773 | { | |
1774 | asm_fprintf (f, "\t! Trampoline code template\n"); | |
1775 | asm_fprintf (f, "\t! This code fragment will be copied " | |
1776 | "into stack on demand\n"); | |
1777 | ||
1778 | asm_fprintf (f, "\tmfusr\t$r16,$pc\n"); | |
1779 | asm_fprintf (f, "\tlwi\t$r15,[$r16 + 20] " | |
1780 | "! load nested function address\n"); | |
1781 | asm_fprintf (f, "\tlwi\t$r16,[$r16 + 16] " | |
1782 | "! load chain_value\n"); | |
1783 | asm_fprintf (f, "\tjr\t$r15\n"); | |
1784 | } | |
1785 | ||
1786 | /* Preserve space ($pc + 16) for saving chain_value, | |
1787 | nds32_trampoline_init will fill the value in this slot. */ | |
1788 | asm_fprintf (f, "\t! space for saving chain_value\n"); | |
1789 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
1790 | ||
1791 | /* Preserve space ($pc + 20) for saving nested function address, | |
1792 | nds32_trampoline_init will fill the value in this slot. */ | |
1793 | asm_fprintf (f, "\t! space for saving nested function address\n"); | |
1794 | assemble_aligned_integer (UNITS_PER_WORD, const0_rtx); | |
1795 | } | |
1796 | ||
1797 | /* Emit RTL insns to initialize the variable parts of a trampoline. */ | |
1798 | static void | |
1799 | nds32_trampoline_init (rtx m_tramp, tree fndecl, rtx chain_value) | |
1800 | { | |
1801 | int i; | |
1802 | ||
1803 | /* Nested function address. */ | |
1804 | rtx fnaddr; | |
1805 | /* The memory rtx that is going to | |
1806 | be filled with chain_value. */ | |
1807 | rtx chain_value_mem; | |
1808 | /* The memory rtx that is going to | |
1809 | be filled with nested function address. */ | |
1810 | rtx nested_func_mem; | |
1811 | ||
1812 | /* Start address of trampoline code in stack, for doing cache sync. */ | |
1813 | rtx sync_cache_addr; | |
1814 | /* Temporary register for sync instruction. */ | |
1815 | rtx tmp_reg; | |
1816 | /* Instruction-cache sync instruction, | |
1817 | requesting an argument as starting address. */ | |
1818 | rtx isync_insn; | |
1819 | /* For convenience reason of doing comparison. */ | |
1820 | int tramp_align_in_bytes; | |
1821 | ||
1822 | /* Trampoline is not supported on reduced-set registers yet. */ | |
1823 | if (TARGET_REDUCED_REGS) | |
1824 | sorry ("a nested function is not supported for reduced registers"); | |
1825 | ||
1826 | /* STEP 1: Copy trampoline code template into stack, | |
1827 | fill up essential data into stack. */ | |
1828 | ||
1829 | /* Extract nested function address rtx. */ | |
1830 | fnaddr = XEXP (DECL_RTL (fndecl), 0); | |
1831 | ||
1832 | /* m_tramp is memory rtx that is going to be filled with trampoline code. | |
1833 | We have nds32_asm_trampoline_template() to emit template pattern. */ | |
1834 | emit_block_move (m_tramp, assemble_trampoline_template (), | |
1835 | GEN_INT (TRAMPOLINE_SIZE), BLOCK_OP_NORMAL); | |
1836 | ||
1837 | /* After copying trampoline code into stack, | |
1838 | fill chain_value into stack. */ | |
1839 | chain_value_mem = adjust_address (m_tramp, SImode, 16); | |
1840 | emit_move_insn (chain_value_mem, chain_value); | |
1841 | /* After copying trampoline code int stack, | |
1842 | fill nested function address into stack. */ | |
1843 | nested_func_mem = adjust_address (m_tramp, SImode, 20); | |
1844 | emit_move_insn (nested_func_mem, fnaddr); | |
1845 | ||
1846 | /* STEP 2: Sync instruction-cache. */ | |
1847 | ||
1848 | /* We have successfully filled trampoline code into stack. | |
1849 | However, in order to execute code in stack correctly, | |
1850 | we must sync instruction cache. */ | |
1851 | sync_cache_addr = XEXP (m_tramp, 0); | |
1852 | tmp_reg = gen_reg_rtx (SImode); | |
1853 | isync_insn = gen_unspec_volatile_isync (tmp_reg); | |
1854 | ||
1855 | /* Because nds32_cache_block_size is in bytes, | |
1856 | we get trampoline alignment in bytes for convenient comparison. */ | |
1857 | tramp_align_in_bytes = TRAMPOLINE_ALIGNMENT / BITS_PER_UNIT; | |
1858 | ||
1859 | if (tramp_align_in_bytes >= nds32_cache_block_size | |
1860 | && (tramp_align_in_bytes % nds32_cache_block_size) == 0) | |
1861 | { | |
1862 | /* Under this condition, the starting address of trampoline | |
1863 | must be aligned to the starting address of each cache block | |
1864 | and we do not have to worry about cross-boundary issue. */ | |
1865 | for (i = 0; | |
1866 | i < (TRAMPOLINE_SIZE + nds32_cache_block_size - 1) | |
1867 | / nds32_cache_block_size; | |
1868 | i++) | |
1869 | { | |
1870 | emit_move_insn (tmp_reg, | |
1871 | plus_constant (Pmode, sync_cache_addr, | |
1872 | nds32_cache_block_size * i)); | |
1873 | emit_insn (isync_insn); | |
1874 | } | |
1875 | } | |
1876 | else if (TRAMPOLINE_SIZE > nds32_cache_block_size) | |
1877 | { | |
1878 | /* The starting address of trampoline code | |
1879 | may not be aligned to the cache block, | |
1880 | so the trampoline code may be across two cache block. | |
1881 | We need to sync the last element, which is 4-byte size, | |
1882 | of trampoline template. */ | |
1883 | for (i = 0; | |
1884 | i < (TRAMPOLINE_SIZE + nds32_cache_block_size - 1) | |
1885 | / nds32_cache_block_size; | |
1886 | i++) | |
1887 | { | |
1888 | emit_move_insn (tmp_reg, | |
1889 | plus_constant (Pmode, sync_cache_addr, | |
1890 | nds32_cache_block_size * i)); | |
1891 | emit_insn (isync_insn); | |
1892 | } | |
1893 | ||
1894 | /* The last element of trampoline template is 4-byte size. */ | |
1895 | emit_move_insn (tmp_reg, | |
1896 | plus_constant (Pmode, sync_cache_addr, | |
1897 | TRAMPOLINE_SIZE - 4)); | |
1898 | emit_insn (isync_insn); | |
1899 | } | |
1900 | else | |
1901 | { | |
1902 | /* This is the simplest case. | |
1903 | Because TRAMPOLINE_SIZE is less than or | |
1904 | equal to nds32_cache_block_size, | |
1905 | we can just sync start address and | |
1906 | the last element of trampoline code. */ | |
1907 | ||
1908 | /* Sync starting address of tampoline code. */ | |
1909 | emit_move_insn (tmp_reg, sync_cache_addr); | |
1910 | emit_insn (isync_insn); | |
1911 | /* Sync the last element, which is 4-byte size, | |
1912 | of trampoline template. */ | |
1913 | emit_move_insn (tmp_reg, | |
1914 | plus_constant (Pmode, sync_cache_addr, | |
1915 | TRAMPOLINE_SIZE - 4)); | |
1916 | emit_insn (isync_insn); | |
1917 | } | |
1918 | ||
1919 | /* Set instruction serialization barrier | |
1920 | to guarantee the correct operations. */ | |
1921 | emit_insn (gen_unspec_volatile_isb ()); | |
1922 | } | |
1923 | ||
1924 | \f | |
1925 | /* Addressing Modes. */ | |
1926 | ||
1927 | static bool | |
ef4bddc2 | 1928 | nds32_legitimate_address_p (machine_mode mode, rtx x, bool strict) |
9304f876 CJW |
1929 | { |
1930 | /* For (mem:DI addr) or (mem:DF addr) case, | |
1931 | we only allow 'addr' to be [reg], [symbol_ref], | |
1932 | [const], or [reg + const_int] pattern. */ | |
1933 | if (mode == DImode || mode == DFmode) | |
1934 | { | |
1935 | /* Allow [Reg + const_int] addressing mode. */ | |
1936 | if (GET_CODE (x) == PLUS) | |
1937 | { | |
1938 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict) | |
1939 | && nds32_legitimate_index_p (mode, XEXP (x, 1), strict) | |
1940 | && CONST_INT_P (XEXP (x, 1))) | |
1941 | return true; | |
1942 | ||
1943 | else if (nds32_address_register_rtx_p (XEXP (x, 1), strict) | |
1944 | && nds32_legitimate_index_p (mode, XEXP (x, 0), strict) | |
1945 | && CONST_INT_P (XEXP (x, 0))) | |
1946 | return true; | |
1947 | } | |
1948 | ||
1949 | /* Now check [reg], [symbol_ref], and [const]. */ | |
1950 | if (GET_CODE (x) != REG | |
1951 | && GET_CODE (x) != SYMBOL_REF | |
1952 | && GET_CODE (x) != CONST) | |
1953 | return false; | |
1954 | } | |
1955 | ||
1956 | /* Check if 'x' is a valid address. */ | |
1957 | switch (GET_CODE (x)) | |
1958 | { | |
1959 | case REG: | |
1960 | /* (mem (reg A)) => [Ra] */ | |
1961 | return nds32_address_register_rtx_p (x, strict); | |
1962 | ||
1963 | case SYMBOL_REF: | |
1964 | ||
1965 | if (!TARGET_GP_DIRECT | |
1966 | && (reload_completed | |
1967 | || reload_in_progress | |
1968 | || lra_in_progress)) | |
1969 | return false; | |
1970 | ||
1971 | /* (mem (symbol_ref A)) => [symbol_ref] */ | |
1972 | return !currently_expanding_to_rtl; | |
1973 | ||
1974 | case CONST: | |
1975 | ||
1976 | if (!TARGET_GP_DIRECT | |
1977 | && (reload_completed | |
1978 | || reload_in_progress | |
1979 | || lra_in_progress)) | |
1980 | return false; | |
1981 | ||
1982 | /* (mem (const (...))) | |
1983 | => [ + const_addr ], where const_addr = symbol_ref + const_int */ | |
1984 | if (GET_CODE (XEXP (x, 0)) == PLUS) | |
1985 | { | |
1986 | rtx plus_op = XEXP (x, 0); | |
1987 | ||
1988 | rtx op0 = XEXP (plus_op, 0); | |
1989 | rtx op1 = XEXP (plus_op, 1); | |
1990 | ||
1991 | if (GET_CODE (op0) == SYMBOL_REF && CONST_INT_P (op1)) | |
1992 | return true; | |
1993 | else | |
1994 | return false; | |
1995 | } | |
1996 | ||
1997 | return false; | |
1998 | ||
1999 | case POST_MODIFY: | |
2000 | /* (mem (post_modify (reg) (plus (reg) (reg)))) | |
2001 | => [Ra], Rb */ | |
2002 | /* (mem (post_modify (reg) (plus (reg) (const_int)))) | |
2003 | => [Ra], const_int */ | |
2004 | if (GET_CODE (XEXP (x, 0)) == REG | |
2005 | && GET_CODE (XEXP (x, 1)) == PLUS) | |
2006 | { | |
2007 | rtx plus_op = XEXP (x, 1); | |
2008 | ||
2009 | rtx op0 = XEXP (plus_op, 0); | |
2010 | rtx op1 = XEXP (plus_op, 1); | |
2011 | ||
2012 | if (nds32_address_register_rtx_p (op0, strict) | |
2013 | && nds32_legitimate_index_p (mode, op1, strict)) | |
2014 | return true; | |
2015 | else | |
2016 | return false; | |
2017 | } | |
2018 | ||
2019 | return false; | |
2020 | ||
2021 | case POST_INC: | |
2022 | case POST_DEC: | |
2023 | /* (mem (post_inc reg)) => [Ra], 1/2/4 */ | |
2024 | /* (mem (post_dec reg)) => [Ra], -1/-2/-4 */ | |
2025 | /* The 1/2/4 or -1/-2/-4 have been displayed in nds32.md. | |
2026 | We only need to deal with register Ra. */ | |
2027 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict)) | |
2028 | return true; | |
2029 | else | |
2030 | return false; | |
2031 | ||
2032 | case PLUS: | |
2033 | /* (mem (plus reg const_int)) | |
2034 | => [Ra + imm] */ | |
2035 | /* (mem (plus reg reg)) | |
2036 | => [Ra + Rb] */ | |
2037 | /* (mem (plus (mult reg const_int) reg)) | |
2038 | => [Ra + Rb << sv] */ | |
2039 | if (nds32_address_register_rtx_p (XEXP (x, 0), strict) | |
2040 | && nds32_legitimate_index_p (mode, XEXP (x, 1), strict)) | |
2041 | return true; | |
2042 | else if (nds32_address_register_rtx_p (XEXP (x, 1), strict) | |
2043 | && nds32_legitimate_index_p (mode, XEXP (x, 0), strict)) | |
2044 | return true; | |
2045 | else | |
2046 | return false; | |
2047 | ||
2048 | case LO_SUM: | |
34425025 CJW |
2049 | /* (mem (lo_sum (reg) (symbol_ref))) */ |
2050 | /* (mem (lo_sum (reg) (const))) */ | |
2051 | gcc_assert (REG_P (XEXP (x, 0))); | |
2052 | if (GET_CODE (XEXP (x, 1)) == SYMBOL_REF | |
2053 | || GET_CODE (XEXP (x, 1)) == CONST) | |
2054 | return nds32_legitimate_address_p (mode, XEXP (x, 1), strict); | |
2055 | else | |
2056 | return false; | |
9304f876 CJW |
2057 | |
2058 | default: | |
2059 | return false; | |
2060 | } | |
2061 | } | |
2062 | ||
2063 | \f | |
2064 | /* Describing Relative Costs of Operations. */ | |
2065 | ||
19ac960a | 2066 | static int |
ef4bddc2 | 2067 | nds32_register_move_cost (machine_mode mode ATTRIBUTE_UNUSED, |
19ac960a CJW |
2068 | reg_class_t from, |
2069 | reg_class_t to) | |
9304f876 CJW |
2070 | { |
2071 | if (from == HIGH_REGS || to == HIGH_REGS) | |
2072 | return 6; | |
2073 | ||
2074 | return 2; | |
2075 | } | |
2076 | ||
19ac960a | 2077 | static int |
ef4bddc2 | 2078 | nds32_memory_move_cost (machine_mode mode ATTRIBUTE_UNUSED, |
19ac960a CJW |
2079 | reg_class_t rclass ATTRIBUTE_UNUSED, |
2080 | bool in ATTRIBUTE_UNUSED) | |
9304f876 CJW |
2081 | { |
2082 | return 8; | |
2083 | } | |
2084 | ||
2085 | /* This target hook describes the relative costs of RTL expressions. | |
2086 | Return 'true' when all subexpressions of x have been processed. | |
2087 | Return 'false' to sum the costs of sub-rtx, plus cost of this operation. | |
2088 | Refer to gcc/rtlanal.c for more information. */ | |
2089 | static bool | |
2090 | nds32_rtx_costs (rtx x, | |
2091 | int code, | |
2092 | int outer_code, | |
89a4b547 | 2093 | int opno, |
9304f876 CJW |
2094 | int *total, |
2095 | bool speed) | |
2096 | { | |
89a4b547 | 2097 | return nds32_rtx_costs_impl (x, code, outer_code, opno, total, speed); |
9304f876 CJW |
2098 | } |
2099 | ||
19ac960a CJW |
2100 | static int |
2101 | nds32_address_cost (rtx address, | |
ef4bddc2 | 2102 | machine_mode mode, |
19ac960a CJW |
2103 | addr_space_t as, |
2104 | bool speed) | |
9304f876 | 2105 | { |
89a4b547 | 2106 | return nds32_address_cost_impl (address, mode, as, speed); |
9304f876 CJW |
2107 | } |
2108 | ||
511a41d7 CJW |
2109 | \f |
2110 | /* Dividing the Output into Sections (Texts, Data, . . . ). */ | |
2111 | ||
2112 | /* If references to a symbol or a constant must be treated differently | |
2113 | depending on something about the variable or function named by the symbol | |
2114 | (such as what section it is in), we use this hook to store flags | |
2115 | in symbol_ref rtx. */ | |
2116 | static void | |
2117 | nds32_encode_section_info (tree decl, rtx rtl, int new_decl_p) | |
2118 | { | |
2119 | default_encode_section_info (decl, rtl, new_decl_p); | |
2120 | ||
2121 | /* For the memory rtx, if it references to rodata section, we can store | |
2122 | NDS32_SYMBOL_FLAG_RODATA flag into symbol_ref rtx so that the | |
2123 | nds32_legitimate_address_p() can determine how to treat such symbol_ref | |
2124 | based on -mcmodel=X and this information. */ | |
2125 | if (MEM_P (rtl) && MEM_READONLY_P (rtl)) | |
2126 | { | |
2127 | rtx addr = XEXP (rtl, 0); | |
2128 | ||
2129 | if (GET_CODE (addr) == SYMBOL_REF) | |
2130 | { | |
2131 | /* For (mem (symbol_ref X)) case. */ | |
2132 | SYMBOL_REF_FLAGS (addr) |= NDS32_SYMBOL_FLAG_RODATA; | |
2133 | } | |
2134 | else if (GET_CODE (addr) == CONST | |
2135 | && GET_CODE (XEXP (addr, 0)) == PLUS) | |
2136 | { | |
2137 | /* For (mem (const (plus (symbol_ref X) (const_int N)))) case. */ | |
2138 | rtx plus_op = XEXP (addr, 0); | |
2139 | rtx op0 = XEXP (plus_op, 0); | |
2140 | rtx op1 = XEXP (plus_op, 1); | |
2141 | ||
2142 | if (GET_CODE (op0) == SYMBOL_REF && CONST_INT_P (op1)) | |
2143 | SYMBOL_REF_FLAGS (op0) |= NDS32_SYMBOL_FLAG_RODATA; | |
2144 | } | |
2145 | } | |
2146 | } | |
2147 | ||
9304f876 CJW |
2148 | \f |
2149 | /* Defining the Output Assembler Language. */ | |
2150 | ||
2151 | /* -- The Overall Framework of an Assembler File. */ | |
2152 | ||
2153 | static void | |
2154 | nds32_asm_file_start (void) | |
2155 | { | |
9304f876 CJW |
2156 | default_file_start (); |
2157 | ||
2158 | /* Tell assembler which ABI we are using. */ | |
2159 | fprintf (asm_out_file, "\t! ABI version\n"); | |
2160 | fprintf (asm_out_file, "\t.abi_2\n"); | |
2161 | ||
2162 | /* Tell assembler that this asm code is generated by compiler. */ | |
2163 | fprintf (asm_out_file, "\t! This asm file is generated by compiler\n"); | |
2164 | fprintf (asm_out_file, "\t.flag\tverbatim\n"); | |
2165 | /* Give assembler the size of each vector for interrupt handler. */ | |
2166 | fprintf (asm_out_file, "\t! This vector size directive is required " | |
2167 | "for checking inconsistency on interrupt handler\n"); | |
2168 | fprintf (asm_out_file, "\t.vec_size\t%d\n", nds32_isr_vector_size); | |
2169 | ||
2170 | /* If user enables '-mforce-fp-as-gp' or compiles programs with -Os, | |
2171 | the compiler may produce 'la $fp,_FP_BASE_' instruction | |
2172 | at prologue for fp-as-gp optimization. | |
2173 | We should emit weak reference of _FP_BASE_ to avoid undefined reference | |
2174 | in case user does not pass '--relax' option to linker. */ | |
2175 | if (TARGET_FORCE_FP_AS_GP || optimize_size) | |
2176 | { | |
2177 | fprintf (asm_out_file, "\t! This weak reference is required to do " | |
2178 | "fp-as-gp link time optimization\n"); | |
2179 | fprintf (asm_out_file, "\t.weak\t_FP_BASE_\n"); | |
2180 | } | |
2181 | /* If user enables '-mex9', we should emit relaxation directive | |
2182 | to tell linker that this file is allowed to do ex9 optimization. */ | |
2183 | if (TARGET_EX9) | |
2184 | { | |
2185 | fprintf (asm_out_file, "\t! This relaxation directive is required " | |
2186 | "to do ex9 link time optimization\n"); | |
2187 | fprintf (asm_out_file, "\t.relax\tex9\n"); | |
2188 | } | |
2189 | ||
2190 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2191 | ||
2192 | if (TARGET_ISA_V2) | |
2193 | fprintf (asm_out_file, "\t! ISA family\t\t: %s\n", "V2"); | |
2194 | if (TARGET_ISA_V3) | |
2195 | fprintf (asm_out_file, "\t! ISA family\t\t: %s\n", "V3"); | |
2196 | if (TARGET_ISA_V3M) | |
2197 | fprintf (asm_out_file, "\t! ISA family\t\t: %s\n", "V3M"); | |
2198 | ||
77b7a1ca CJW |
2199 | if (TARGET_CMODEL_SMALL) |
2200 | fprintf (asm_out_file, "\t! Code model\t\t: %s\n", "SMALL"); | |
2201 | if (TARGET_CMODEL_MEDIUM) | |
2202 | fprintf (asm_out_file, "\t! Code model\t\t: %s\n", "MEDIUM"); | |
2203 | if (TARGET_CMODEL_LARGE) | |
2204 | fprintf (asm_out_file, "\t! Code model\t\t: %s\n", "LARGE"); | |
2205 | ||
9304f876 CJW |
2206 | fprintf (asm_out_file, "\t! Endian setting\t: %s\n", |
2207 | ((TARGET_BIG_ENDIAN) ? "big-endian" | |
2208 | : "little-endian")); | |
2209 | ||
2210 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2211 | ||
2212 | fprintf (asm_out_file, "\t! Use conditional move\t\t: %s\n", | |
2213 | ((TARGET_CMOV) ? "Yes" | |
2214 | : "No")); | |
2215 | fprintf (asm_out_file, "\t! Use performance extension\t: %s\n", | |
2216 | ((TARGET_PERF_EXT) ? "Yes" | |
2217 | : "No")); | |
2218 | ||
2219 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2220 | ||
2221 | fprintf (asm_out_file, "\t! V3PUSH instructions\t: %s\n", | |
2222 | ((TARGET_V3PUSH) ? "Yes" | |
2223 | : "No")); | |
2224 | fprintf (asm_out_file, "\t! 16-bit instructions\t: %s\n", | |
2225 | ((TARGET_16_BIT) ? "Yes" | |
2226 | : "No")); | |
9304f876 CJW |
2227 | fprintf (asm_out_file, "\t! Reduced registers set\t: %s\n", |
2228 | ((TARGET_REDUCED_REGS) ? "Yes" | |
2229 | : "No")); | |
2230 | ||
2231 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2232 | ||
2233 | if (optimize_size) | |
2234 | fprintf (asm_out_file, "\t! Optimization level\t: -Os\n"); | |
2235 | else | |
2236 | fprintf (asm_out_file, "\t! Optimization level\t: -O%d\n", optimize); | |
2237 | ||
2238 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2239 | ||
2240 | fprintf (asm_out_file, "\t! Cache block size\t: %d\n", | |
2241 | nds32_cache_block_size); | |
2242 | ||
2243 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2244 | ||
c23a919b | 2245 | nds32_asm_file_start_for_isr (); |
9304f876 CJW |
2246 | } |
2247 | ||
2248 | static void | |
2249 | nds32_asm_file_end (void) | |
2250 | { | |
c23a919b | 2251 | nds32_asm_file_end_for_isr (); |
9304f876 CJW |
2252 | |
2253 | fprintf (asm_out_file, "\t! ------------------------------------\n"); | |
2254 | } | |
2255 | ||
2256 | /* -- Output and Generation of Labels. */ | |
2257 | ||
2258 | static void | |
2259 | nds32_asm_globalize_label (FILE *stream, const char *name) | |
2260 | { | |
2261 | fputs ("\t.global\t", stream); | |
2262 | assemble_name (stream, name); | |
2263 | fputs ("\n", stream); | |
2264 | } | |
2265 | ||
2266 | /* -- Output of Assembler Instructions. */ | |
2267 | ||
2268 | static void | |
2269 | nds32_print_operand (FILE *stream, rtx x, int code) | |
2270 | { | |
2271 | int op_value; | |
2272 | ||
2273 | switch (code) | |
2274 | { | |
2275 | case 0 : | |
2276 | /* Do nothing special. */ | |
2277 | break; | |
2278 | ||
2279 | case 'V': | |
2280 | /* 'x' is supposed to be CONST_INT, get the value. */ | |
2281 | gcc_assert (CONST_INT_P (x)); | |
2282 | op_value = INTVAL (x); | |
2283 | ||
2284 | /* According to the Andes architecture, | |
2285 | the system/user register index range is 0 ~ 1023. | |
2286 | In order to avoid conflict between user-specified-integer value | |
2287 | and enum-specified-register value, | |
2288 | the 'enum nds32_intrinsic_registers' value | |
2289 | in nds32_intrinsic.h starts from 1024. */ | |
2290 | if (op_value < 1024 && op_value >= 0) | |
2291 | { | |
2292 | /* If user gives integer value directly (0~1023), | |
2293 | we just print out the value. */ | |
2294 | fprintf (stream, "%d", op_value); | |
2295 | } | |
2296 | else if (op_value < 0 | |
2297 | || op_value >= ((int) ARRAY_SIZE (nds32_intrinsic_register_names) | |
2298 | + 1024)) | |
2299 | { | |
2300 | /* The enum index value for array size is out of range. */ | |
2301 | error ("intrinsic register index is out of range"); | |
2302 | } | |
2303 | else | |
2304 | { | |
2305 | /* If user applies normal way with __NDS32_REG_XXX__ enum data, | |
2306 | we can print out register name. Remember to substract 1024. */ | |
2307 | fprintf (stream, "%s", | |
2308 | nds32_intrinsic_register_names[op_value - 1024]); | |
2309 | } | |
2310 | ||
2311 | /* No need to handle following process, so return immediately. */ | |
2312 | return; | |
2313 | ||
2314 | default : | |
2315 | /* Unknown flag. */ | |
2316 | output_operand_lossage ("invalid operand output code"); | |
2317 | break; | |
2318 | } | |
2319 | ||
2320 | switch (GET_CODE (x)) | |
2321 | { | |
2322 | case LABEL_REF: | |
2323 | case SYMBOL_REF: | |
2324 | output_addr_const (stream, x); | |
2325 | break; | |
2326 | ||
2327 | case REG: | |
2328 | /* Forbid using static chain register ($r16) | |
2329 | on reduced-set registers configuration. */ | |
2330 | if (TARGET_REDUCED_REGS | |
2331 | && REGNO (x) == STATIC_CHAIN_REGNUM) | |
2332 | sorry ("a nested function is not supported for reduced registers"); | |
2333 | ||
2334 | /* Normal cases, print out register name. */ | |
2335 | fputs (reg_names[REGNO (x)], stream); | |
2336 | break; | |
2337 | ||
2338 | case MEM: | |
2339 | output_address (XEXP (x, 0)); | |
2340 | break; | |
2341 | ||
2342 | case CODE_LABEL: | |
2343 | case CONST_INT: | |
2344 | case CONST: | |
2345 | output_addr_const (stream, x); | |
2346 | break; | |
2347 | ||
2348 | default: | |
2349 | /* Generally, output_addr_const () is able to handle most cases. | |
2350 | We want to see what CODE could appear, | |
2351 | so we use gcc_unreachable() to stop it. */ | |
2352 | debug_rtx (x); | |
2353 | gcc_unreachable (); | |
2354 | break; | |
2355 | } | |
2356 | } | |
2357 | ||
2358 | static void | |
2359 | nds32_print_operand_address (FILE *stream, rtx x) | |
2360 | { | |
2361 | rtx op0, op1; | |
2362 | ||
2363 | switch (GET_CODE (x)) | |
2364 | { | |
2365 | case SYMBOL_REF: | |
2366 | case CONST: | |
2367 | /* [ + symbol_ref] */ | |
2368 | /* [ + const_addr], where const_addr = symbol_ref + const_int */ | |
2369 | fputs ("[ + ", stream); | |
2370 | output_addr_const (stream, x); | |
2371 | fputs ("]", stream); | |
2372 | break; | |
2373 | ||
2374 | case REG: | |
2375 | /* Forbid using static chain register ($r16) | |
2376 | on reduced-set registers configuration. */ | |
2377 | if (TARGET_REDUCED_REGS | |
2378 | && REGNO (x) == STATIC_CHAIN_REGNUM) | |
2379 | sorry ("a nested function is not supported for reduced registers"); | |
2380 | ||
2381 | /* [Ra] */ | |
2382 | fprintf (stream, "[%s]", reg_names[REGNO (x)]); | |
2383 | break; | |
2384 | ||
2385 | case PLUS: | |
2386 | op0 = XEXP (x, 0); | |
2387 | op1 = XEXP (x, 1); | |
2388 | ||
2389 | /* Checking op0, forbid using static chain register ($r16) | |
2390 | on reduced-set registers configuration. */ | |
2391 | if (TARGET_REDUCED_REGS | |
2392 | && REG_P (op0) | |
2393 | && REGNO (op0) == STATIC_CHAIN_REGNUM) | |
2394 | sorry ("a nested function is not supported for reduced registers"); | |
2395 | /* Checking op1, forbid using static chain register ($r16) | |
2396 | on reduced-set registers configuration. */ | |
2397 | if (TARGET_REDUCED_REGS | |
2398 | && REG_P (op1) | |
2399 | && REGNO (op1) == STATIC_CHAIN_REGNUM) | |
2400 | sorry ("a nested function is not supported for reduced registers"); | |
2401 | ||
2402 | if (REG_P (op0) && CONST_INT_P (op1)) | |
2403 | { | |
2404 | /* [Ra + imm] */ | |
2405 | fprintf (stream, "[%s + (%d)]", | |
2406 | reg_names[REGNO (op0)], (int)INTVAL (op1)); | |
2407 | } | |
2408 | else if (REG_P (op0) && REG_P (op1)) | |
2409 | { | |
2410 | /* [Ra + Rb] */ | |
2411 | fprintf (stream, "[%s + %s]", | |
2412 | reg_names[REGNO (op0)], reg_names[REGNO (op1)]); | |
2413 | } | |
2414 | else if (GET_CODE (op0) == MULT && REG_P (op1)) | |
2415 | { | |
2416 | /* [Ra + Rb << sv] | |
2417 | From observation, the pattern looks like: | |
2418 | (plus:SI (mult:SI (reg:SI 58) | |
2419 | (const_int 4 [0x4])) | |
2420 | (reg/f:SI 57)) */ | |
2421 | int sv; | |
2422 | ||
2423 | /* We need to set sv to output shift value. */ | |
2424 | if (INTVAL (XEXP (op0, 1)) == 1) | |
2425 | sv = 0; | |
2426 | else if (INTVAL (XEXP (op0, 1)) == 2) | |
2427 | sv = 1; | |
2428 | else if (INTVAL (XEXP (op0, 1)) == 4) | |
2429 | sv = 2; | |
2430 | else | |
2431 | gcc_unreachable (); | |
2432 | ||
2433 | fprintf (stream, "[%s + %s << %d]", | |
2434 | reg_names[REGNO (op1)], | |
2435 | reg_names[REGNO (XEXP (op0, 0))], | |
2436 | sv); | |
2437 | } | |
2438 | else | |
2439 | { | |
2440 | /* The control flow is not supposed to be here. */ | |
2441 | debug_rtx (x); | |
2442 | gcc_unreachable (); | |
2443 | } | |
2444 | ||
2445 | break; | |
2446 | ||
2447 | case POST_MODIFY: | |
2448 | /* (post_modify (regA) (plus (regA) (regB))) | |
2449 | (post_modify (regA) (plus (regA) (const_int))) | |
2450 | We would like to extract | |
2451 | regA and regB (or const_int) from plus rtx. */ | |
2452 | op0 = XEXP (XEXP (x, 1), 0); | |
2453 | op1 = XEXP (XEXP (x, 1), 1); | |
2454 | ||
2455 | /* Checking op0, forbid using static chain register ($r16) | |
2456 | on reduced-set registers configuration. */ | |
2457 | if (TARGET_REDUCED_REGS | |
2458 | && REG_P (op0) | |
2459 | && REGNO (op0) == STATIC_CHAIN_REGNUM) | |
2460 | sorry ("a nested function is not supported for reduced registers"); | |
2461 | /* Checking op1, forbid using static chain register ($r16) | |
2462 | on reduced-set registers configuration. */ | |
2463 | if (TARGET_REDUCED_REGS | |
2464 | && REG_P (op1) | |
2465 | && REGNO (op1) == STATIC_CHAIN_REGNUM) | |
2466 | sorry ("a nested function is not supported for reduced registers"); | |
2467 | ||
2468 | if (REG_P (op0) && REG_P (op1)) | |
2469 | { | |
2470 | /* [Ra], Rb */ | |
2471 | fprintf (stream, "[%s], %s", | |
2472 | reg_names[REGNO (op0)], reg_names[REGNO (op1)]); | |
2473 | } | |
2474 | else if (REG_P (op0) && CONST_INT_P (op1)) | |
2475 | { | |
2476 | /* [Ra], imm */ | |
2477 | fprintf (stream, "[%s], %d", | |
2478 | reg_names[REGNO (op0)], (int)INTVAL (op1)); | |
2479 | } | |
2480 | else | |
2481 | { | |
2482 | /* The control flow is not supposed to be here. */ | |
2483 | debug_rtx (x); | |
2484 | gcc_unreachable (); | |
2485 | } | |
2486 | ||
2487 | break; | |
2488 | ||
2489 | case POST_INC: | |
2490 | case POST_DEC: | |
2491 | op0 = XEXP (x, 0); | |
2492 | ||
2493 | /* Checking op0, forbid using static chain register ($r16) | |
2494 | on reduced-set registers configuration. */ | |
2495 | if (TARGET_REDUCED_REGS | |
2496 | && REG_P (op0) | |
2497 | && REGNO (op0) == STATIC_CHAIN_REGNUM) | |
2498 | sorry ("a nested function is not supported for reduced registers"); | |
2499 | ||
2500 | if (REG_P (op0)) | |
2501 | { | |
2502 | /* "[Ra], 1/2/4" or "[Ra], -1/-2/-4" | |
2503 | The 1/2/4 or -1/-2/-4 have been displayed in nds32.md. | |
2504 | We only need to deal with register Ra. */ | |
2505 | fprintf (stream, "[%s]", reg_names[REGNO (op0)]); | |
2506 | } | |
2507 | else | |
2508 | { | |
2509 | /* The control flow is not supposed to be here. */ | |
2510 | debug_rtx (x); | |
2511 | gcc_unreachable (); | |
2512 | } | |
2513 | ||
2514 | break; | |
2515 | ||
2516 | default : | |
2517 | /* Generally, output_addr_const () is able to handle most cases. | |
2518 | We want to see what CODE could appear, | |
2519 | so we use gcc_unreachable() to stop it. */ | |
2520 | debug_rtx (x); | |
2521 | gcc_unreachable (); | |
2522 | break; | |
2523 | } | |
2524 | } | |
2525 | ||
2526 | \f | |
2527 | /* Defining target-specific uses of __attribute__. */ | |
2528 | ||
2529 | /* Add some checking after merging attributes. */ | |
2530 | static tree | |
2531 | nds32_merge_decl_attributes (tree olddecl, tree newdecl) | |
2532 | { | |
2533 | tree combined_attrs; | |
2534 | ||
2535 | /* Create combined attributes. */ | |
2536 | combined_attrs = merge_attributes (DECL_ATTRIBUTES (olddecl), | |
2537 | DECL_ATTRIBUTES (newdecl)); | |
2538 | ||
59043e75 | 2539 | /* Since newdecl is acutally a duplicate of olddecl, |
9304f876 CJW |
2540 | we can take olddecl for some operations. */ |
2541 | if (TREE_CODE (olddecl) == FUNCTION_DECL) | |
2542 | { | |
2543 | /* Check isr-specific attributes conflict. */ | |
2544 | nds32_check_isr_attrs_conflict (olddecl, combined_attrs); | |
2545 | } | |
2546 | ||
2547 | return combined_attrs; | |
2548 | } | |
2549 | ||
2550 | /* Add some checking when inserting attributes. */ | |
2551 | static void | |
2552 | nds32_insert_attributes (tree decl, tree *attributes) | |
2553 | { | |
2554 | /* For function declaration, we need to check isr-specific attributes: | |
2555 | 1. Call nds32_check_isr_attrs_conflict() to check any conflict. | |
2556 | 2. Check valid integer value for interrupt/exception. | |
2557 | 3. Check valid integer value for reset. | |
2558 | 4. Check valid function for nmi/warm. */ | |
2559 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
2560 | { | |
2561 | tree func_attrs; | |
2562 | tree intr, excp, reset; | |
2563 | ||
2564 | /* Pick up function attributes. */ | |
2565 | func_attrs = *attributes; | |
2566 | ||
2567 | /* 1. Call nds32_check_isr_attrs_conflict() to check any conflict. */ | |
2568 | nds32_check_isr_attrs_conflict (decl, func_attrs); | |
2569 | ||
2570 | /* Now we are starting to check valid id value | |
2571 | for interrupt/exception/reset. | |
2572 | Note that we ONLY check its validity here. | |
2573 | To construct isr vector information, it is still performed | |
2574 | by nds32_construct_isr_vectors_information(). */ | |
2575 | intr = lookup_attribute ("interrupt", func_attrs); | |
2576 | excp = lookup_attribute ("exception", func_attrs); | |
2577 | reset = lookup_attribute ("reset", func_attrs); | |
2578 | ||
2579 | if (intr || excp) | |
2580 | { | |
2581 | /* Deal with interrupt/exception. */ | |
2582 | tree id_list; | |
2583 | unsigned int lower_bound, upper_bound; | |
2584 | ||
2585 | /* The way to handle interrupt or exception is the same, | |
2586 | we just need to take care of actual vector number. | |
2587 | For interrupt(0..63), the actual vector number is (9..72). | |
2588 | For exception(1..8), the actual vector number is (1..8). */ | |
2589 | lower_bound = (intr) ? (0) : (1); | |
2590 | upper_bound = (intr) ? (63) : (8); | |
2591 | ||
2592 | /* Prepare id list so that we can traverse id value. */ | |
2593 | id_list = (intr) ? (TREE_VALUE (intr)) : (TREE_VALUE (excp)); | |
2594 | ||
2595 | /* 2. Check valid integer value for interrupt/exception. */ | |
2596 | while (id_list) | |
2597 | { | |
2598 | tree id; | |
2599 | ||
2600 | /* Pick up each vector id value. */ | |
2601 | id = TREE_VALUE (id_list); | |
2602 | /* Issue error if it is not a valid integer value. */ | |
2603 | if (TREE_CODE (id) != INTEGER_CST | |
807e902e KZ |
2604 | || wi::ltu_p (id, lower_bound) |
2605 | || wi::gtu_p (id, upper_bound)) | |
9304f876 CJW |
2606 | error ("invalid id value for interrupt/exception attribute"); |
2607 | ||
2608 | /* Advance to next id. */ | |
2609 | id_list = TREE_CHAIN (id_list); | |
2610 | } | |
2611 | } | |
2612 | else if (reset) | |
2613 | { | |
2614 | /* Deal with reset. */ | |
2615 | tree id_list; | |
2616 | tree id; | |
2617 | tree nmi, warm; | |
2618 | unsigned int lower_bound; | |
2619 | unsigned int upper_bound; | |
2620 | ||
2621 | /* Prepare id_list and identify id value so that | |
2622 | we can check if total number of vectors is valid. */ | |
2623 | id_list = TREE_VALUE (reset); | |
2624 | id = TREE_VALUE (id_list); | |
2625 | ||
2626 | /* The maximum numbers for user's interrupt is 64. */ | |
2627 | lower_bound = 0; | |
2628 | upper_bound = 64; | |
2629 | ||
2630 | /* 3. Check valid integer value for reset. */ | |
2631 | if (TREE_CODE (id) != INTEGER_CST | |
807e902e KZ |
2632 | || wi::ltu_p (id, lower_bound) |
2633 | || wi::gtu_p (id, upper_bound)) | |
9304f876 CJW |
2634 | error ("invalid id value for reset attribute"); |
2635 | ||
2636 | /* 4. Check valid function for nmi/warm. */ | |
2637 | nmi = lookup_attribute ("nmi", func_attrs); | |
2638 | warm = lookup_attribute ("warm", func_attrs); | |
2639 | ||
2640 | if (nmi != NULL_TREE) | |
2641 | { | |
2642 | tree nmi_func_list; | |
2643 | tree nmi_func; | |
2644 | ||
2645 | nmi_func_list = TREE_VALUE (nmi); | |
2646 | nmi_func = TREE_VALUE (nmi_func_list); | |
2647 | ||
2648 | /* Issue error if it is not a valid nmi function. */ | |
2649 | if (TREE_CODE (nmi_func) != IDENTIFIER_NODE) | |
2650 | error ("invalid nmi function for reset attribute"); | |
2651 | } | |
2652 | ||
2653 | if (warm != NULL_TREE) | |
2654 | { | |
2655 | tree warm_func_list; | |
2656 | tree warm_func; | |
2657 | ||
2658 | warm_func_list = TREE_VALUE (warm); | |
2659 | warm_func = TREE_VALUE (warm_func_list); | |
2660 | ||
2661 | /* Issue error if it is not a valid warm function. */ | |
2662 | if (TREE_CODE (warm_func) != IDENTIFIER_NODE) | |
2663 | error ("invalid warm function for reset attribute"); | |
2664 | } | |
2665 | } | |
2666 | else | |
2667 | { | |
2668 | /* No interrupt, exception, or reset attribute is set. */ | |
2669 | return; | |
2670 | } | |
2671 | } | |
2672 | } | |
2673 | ||
2674 | static bool | |
2675 | nds32_option_pragma_parse (tree args ATTRIBUTE_UNUSED, | |
2676 | tree pop_target ATTRIBUTE_UNUSED) | |
2677 | { | |
2678 | /* Currently, we do not parse any pragma target by ourself, | |
2679 | so just simply return false. */ | |
2680 | return false; | |
2681 | } | |
2682 | ||
2683 | static void | |
2684 | nds32_option_override (void) | |
2685 | { | |
2686 | /* After all the command options have been parsed, | |
2687 | we shall deal with some flags for changing compiler settings. */ | |
2688 | ||
2689 | /* At first, we check if we have to strictly | |
2690 | set some flags based on ISA family. */ | |
2691 | if (TARGET_ISA_V2) | |
2692 | { | |
2693 | /* Under V2 ISA, we need to strictly disable TARGET_V3PUSH. */ | |
2694 | target_flags &= ~MASK_V3PUSH; | |
2695 | } | |
2696 | if (TARGET_ISA_V3) | |
2697 | { | |
2698 | /* Under V3 ISA, currently nothing should be strictly set. */ | |
2699 | } | |
2700 | if (TARGET_ISA_V3M) | |
2701 | { | |
2702 | /* Under V3M ISA, we need to strictly enable TARGET_REDUCED_REGS. */ | |
2703 | target_flags |= MASK_REDUCED_REGS; | |
2704 | /* Under V3M ISA, we need to strictly disable TARGET_PERF_EXT. */ | |
2705 | target_flags &= ~MASK_PERF_EXT; | |
2706 | } | |
2707 | ||
2708 | /* See if we are using reduced-set registers: | |
2709 | $r0~$r5, $r6~$r10, $r15, $r28, $r29, $r30, $r31 | |
2710 | If so, we must forbid using $r11~$r14, $r16~$r27. */ | |
2711 | if (TARGET_REDUCED_REGS) | |
2712 | { | |
2713 | int r; | |
2714 | ||
2715 | /* Prevent register allocator from | |
2716 | choosing it as doing register allocation. */ | |
2717 | for (r = 11; r <= 14; r++) | |
2718 | fixed_regs[r] = call_used_regs[r] = 1; | |
2719 | for (r = 16; r <= 27; r++) | |
2720 | fixed_regs[r] = call_used_regs[r] = 1; | |
2721 | } | |
2722 | ||
2723 | /* See if user explicitly would like to use fp-as-gp optimization. | |
2724 | If so, we must prevent $fp from being allocated | |
2725 | during register allocation. */ | |
2726 | if (TARGET_FORCE_FP_AS_GP) | |
2727 | fixed_regs[FP_REGNUM] = call_used_regs[FP_REGNUM] = 1; | |
2728 | ||
2729 | if (!TARGET_16_BIT) | |
2730 | { | |
2731 | /* Under no 16 bit ISA, we need to strictly disable TARGET_V3PUSH. */ | |
2732 | target_flags &= ~MASK_V3PUSH; | |
2733 | } | |
2734 | ||
2735 | /* Currently, we don't support PIC code generation yet. */ | |
2736 | if (flag_pic) | |
2737 | sorry ("not support -fpic"); | |
2738 | } | |
2739 | ||
2740 | \f | |
2741 | /* Miscellaneous Parameters. */ | |
2742 | ||
2743 | static void | |
2744 | nds32_init_builtins (void) | |
2745 | { | |
aaa44d2d | 2746 | nds32_init_builtins_impl (); |
9304f876 CJW |
2747 | } |
2748 | ||
2749 | static rtx | |
2750 | nds32_expand_builtin (tree exp, | |
2751 | rtx target, | |
aaa44d2d | 2752 | rtx subtarget, |
ef4bddc2 | 2753 | machine_mode mode, |
aaa44d2d | 2754 | int ignore) |
9304f876 | 2755 | { |
aaa44d2d | 2756 | return nds32_expand_builtin_impl (exp, target, subtarget, mode, ignore); |
9304f876 CJW |
2757 | } |
2758 | ||
2759 | ||
2760 | /* ------------------------------------------------------------------------ */ | |
2761 | ||
2762 | /* PART 4: Implemet extern function definitions, | |
2763 | the prototype is in nds32-protos.h. */ | |
2764 | \f | |
2765 | /* Defining Data Structures for Per-function Information. */ | |
2766 | ||
2767 | void | |
2768 | nds32_init_expanders (void) | |
2769 | { | |
2770 | /* Arrange to initialize and mark the machine per-function status. */ | |
2771 | init_machine_status = nds32_init_machine_status; | |
2772 | } | |
2773 | ||
2774 | \f | |
2775 | /* Register Usage. */ | |
2776 | ||
2777 | /* -- How Values Fit in Registers. */ | |
2778 | ||
2779 | int | |
2780 | nds32_hard_regno_nregs (int regno ATTRIBUTE_UNUSED, | |
ef4bddc2 | 2781 | machine_mode mode) |
9304f876 CJW |
2782 | { |
2783 | return ((GET_MODE_SIZE (mode) + UNITS_PER_WORD - 1) / UNITS_PER_WORD); | |
2784 | } | |
2785 | ||
2786 | int | |
ef4bddc2 | 2787 | nds32_hard_regno_mode_ok (int regno, machine_mode mode) |
9304f876 CJW |
2788 | { |
2789 | /* Restrict double-word quantities to even register pairs. */ | |
2790 | if (HARD_REGNO_NREGS (regno, mode) == 1 | |
2791 | || !((regno) & 1)) | |
2792 | return 1; | |
2793 | ||
2794 | return 0; | |
2795 | } | |
2796 | ||
2797 | \f | |
2798 | /* Register Classes. */ | |
2799 | ||
2800 | enum reg_class | |
2801 | nds32_regno_reg_class (int regno) | |
2802 | { | |
2803 | /* Refer to nds32.h for more register class details. */ | |
2804 | ||
2805 | if (regno >= 0 && regno <= 7) | |
2806 | return LOW_REGS; | |
2807 | else if (regno >= 8 && regno <= 11) | |
2808 | return MIDDLE_REGS; | |
2809 | else if (regno >= 12 && regno <= 14) | |
2810 | return HIGH_REGS; | |
2811 | else if (regno == 15) | |
2812 | return R15_TA_REG; | |
2813 | else if (regno >= 16 && regno <= 19) | |
2814 | return MIDDLE_REGS; | |
2815 | else if (regno >= 20 && regno <= 31) | |
2816 | return HIGH_REGS; | |
2817 | else if (regno == 32 || regno == 33) | |
2818 | return FRAME_REGS; | |
2819 | else | |
2820 | return NO_REGS; | |
2821 | } | |
2822 | ||
2823 | \f | |
2824 | /* Stack Layout and Calling Conventions. */ | |
2825 | ||
2826 | /* -- Basic Stack Layout. */ | |
2827 | ||
2828 | rtx | |
2829 | nds32_return_addr_rtx (int count, | |
2830 | rtx frameaddr ATTRIBUTE_UNUSED) | |
2831 | { | |
2832 | /* There is no way to determine the return address | |
2833 | if frameaddr is the frame that has 'count' steps | |
2834 | up from current frame. */ | |
2835 | if (count != 0) | |
2836 | return NULL_RTX; | |
2837 | ||
2838 | /* If count == 0, it means we are at current frame, | |
2839 | the return address is $r30 ($lp). */ | |
2840 | return get_hard_reg_initial_val (Pmode, LP_REGNUM); | |
2841 | } | |
2842 | ||
2843 | /* -- Eliminating Frame Pointer and Arg Pointer. */ | |
2844 | ||
2845 | HOST_WIDE_INT | |
2846 | nds32_initial_elimination_offset (unsigned int from_reg, unsigned int to_reg) | |
2847 | { | |
2848 | HOST_WIDE_INT offset; | |
2849 | ||
2850 | /* Compute and setup stack frame size. | |
2851 | The result will be in cfun->machine. */ | |
2852 | nds32_compute_stack_frame (); | |
2853 | ||
2854 | /* Remember to consider | |
2855 | cfun->machine->callee_saved_area_padding_bytes | |
2856 | when calculating offset. */ | |
2857 | if (from_reg == ARG_POINTER_REGNUM && to_reg == STACK_POINTER_REGNUM) | |
2858 | { | |
2859 | offset = (cfun->machine->fp_size | |
2860 | + cfun->machine->gp_size | |
2861 | + cfun->machine->lp_size | |
2862 | + cfun->machine->callee_saved_regs_size | |
2863 | + cfun->machine->callee_saved_area_padding_bytes | |
2864 | + cfun->machine->local_size | |
2865 | + cfun->machine->out_args_size); | |
2866 | } | |
2867 | else if (from_reg == ARG_POINTER_REGNUM | |
2868 | && to_reg == HARD_FRAME_POINTER_REGNUM) | |
2869 | { | |
2870 | offset = 0; | |
2871 | } | |
2872 | else if (from_reg == FRAME_POINTER_REGNUM | |
2873 | && to_reg == STACK_POINTER_REGNUM) | |
2874 | { | |
2875 | offset = (cfun->machine->local_size + cfun->machine->out_args_size); | |
2876 | } | |
2877 | else if (from_reg == FRAME_POINTER_REGNUM | |
2878 | && to_reg == HARD_FRAME_POINTER_REGNUM) | |
2879 | { | |
2880 | offset = (-1) * (cfun->machine->fp_size | |
2881 | + cfun->machine->gp_size | |
2882 | + cfun->machine->lp_size | |
2883 | + cfun->machine->callee_saved_regs_size | |
2884 | + cfun->machine->callee_saved_area_padding_bytes); | |
2885 | } | |
2886 | else | |
2887 | { | |
2888 | gcc_unreachable (); | |
2889 | } | |
2890 | ||
2891 | return offset; | |
2892 | } | |
2893 | ||
2894 | /* -- Passing Arguments in Registers. */ | |
2895 | ||
2896 | void | |
2897 | nds32_init_cumulative_args (CUMULATIVE_ARGS *cum, | |
2898 | tree fntype ATTRIBUTE_UNUSED, | |
2899 | rtx libname ATTRIBUTE_UNUSED, | |
2900 | tree fndecl ATTRIBUTE_UNUSED, | |
2901 | int n_named_args ATTRIBUTE_UNUSED) | |
2902 | { | |
2903 | /* Initial available registers | |
2904 | (in offset, corresponding to NDS32_GPR_ARG_FIRST_REGNUM) | |
2905 | for passing arguments. */ | |
9d93cc24 | 2906 | cum->gpr_offset = 0; |
9304f876 CJW |
2907 | } |
2908 | ||
2909 | /* -- Function Entry and Exit. */ | |
2910 | ||
2911 | /* Function for normal multiple push prologue. */ | |
2912 | void | |
2913 | nds32_expand_prologue (void) | |
2914 | { | |
2915 | int fp_adjust; | |
2916 | int sp_adjust; | |
2917 | int en4_const; | |
2918 | ||
2919 | rtx Rb, Re; | |
9304f876 CJW |
2920 | rtx fp_adjust_insn, sp_adjust_insn; |
2921 | ||
9304f876 CJW |
2922 | /* Compute and setup stack frame size. |
2923 | The result will be in cfun->machine. */ | |
2924 | nds32_compute_stack_frame (); | |
2925 | ||
557430f1 CJW |
2926 | /* If this is a variadic function, first we need to push argument |
2927 | registers that hold the unnamed argument value. */ | |
2928 | if (cfun->machine->va_args_size != 0) | |
2929 | { | |
2930 | Rb = gen_rtx_REG (SImode, cfun->machine->va_args_first_regno); | |
2931 | Re = gen_rtx_REG (SImode, cfun->machine->va_args_last_regno); | |
2932 | /* No need to push $fp, $gp, or $lp, so use GEN_INT(0). */ | |
47e0e7d2 | 2933 | nds32_emit_stack_push_multiple (Rb, Re, GEN_INT (0), true); |
557430f1 CJW |
2934 | |
2935 | /* We may also need to adjust stack pointer for padding bytes | |
2936 | because varargs may cause $sp not 8-byte aligned. */ | |
2937 | if (cfun->machine->va_args_area_padding_bytes) | |
2938 | { | |
2939 | /* Generate sp adjustment instruction. */ | |
2940 | sp_adjust = cfun->machine->va_args_area_padding_bytes; | |
2941 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
2942 | stack_pointer_rtx, | |
2943 | GEN_INT (-1 * sp_adjust)); | |
2944 | ||
2945 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
2946 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
2947 | ||
2948 | /* The insn rtx 'sp_adjust_insn' will change frame layout. | |
2949 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
2950 | generate CFI (Call Frame Information) stuff. */ | |
2951 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
2952 | } | |
2953 | } | |
2954 | ||
9304f876 CJW |
2955 | /* If the function is 'naked', |
2956 | we do not have to generate prologue code fragment. */ | |
2957 | if (cfun->machine->naked_p) | |
2958 | return; | |
2959 | ||
2960 | /* Get callee_first_regno and callee_last_regno. */ | |
2961 | Rb = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_first_regno); | |
2962 | Re = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_last_regno); | |
2963 | ||
4e9a2848 | 2964 | /* nds32_emit_stack_push_multiple(first_regno, last_regno), |
9304f876 CJW |
2965 | the pattern 'stack_push_multiple' is implemented in nds32.md. |
2966 | For En4 field, we have to calculate its constant value. | |
2967 | Refer to Andes ISA for more information. */ | |
2968 | en4_const = 0; | |
2969 | if (cfun->machine->fp_size) | |
2970 | en4_const += 8; | |
2971 | if (cfun->machine->gp_size) | |
2972 | en4_const += 4; | |
2973 | if (cfun->machine->lp_size) | |
2974 | en4_const += 2; | |
2975 | ||
2976 | /* If $fp, $gp, $lp, and all callee-save registers are NOT required | |
2977 | to be saved, we don't have to create multiple push instruction. | |
2978 | Otherwise, a multiple push instruction is needed. */ | |
2979 | if (!(REGNO (Rb) == SP_REGNUM && REGNO (Re) == SP_REGNUM && en4_const == 0)) | |
2980 | { | |
2981 | /* Create multiple push instruction rtx. */ | |
47e0e7d2 | 2982 | nds32_emit_stack_push_multiple (Rb, Re, GEN_INT (en4_const), false); |
9304f876 CJW |
2983 | } |
2984 | ||
2985 | /* Check frame_pointer_needed to see | |
2986 | if we shall emit fp adjustment instruction. */ | |
2987 | if (frame_pointer_needed) | |
2988 | { | |
2989 | /* adjust $fp = $sp + ($fp size) + ($gp size) + ($lp size) | |
2990 | + (4 * callee-saved-registers) | |
2991 | Note: No need to adjust | |
2992 | cfun->machine->callee_saved_area_padding_bytes, | |
2993 | because, at this point, stack pointer is just | |
2994 | at the position after push instruction. */ | |
2995 | fp_adjust = cfun->machine->fp_size | |
2996 | + cfun->machine->gp_size | |
2997 | + cfun->machine->lp_size | |
2998 | + cfun->machine->callee_saved_regs_size; | |
2999 | fp_adjust_insn = gen_addsi3 (hard_frame_pointer_rtx, | |
3000 | stack_pointer_rtx, | |
3001 | GEN_INT (fp_adjust)); | |
3002 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3003 | fp_adjust_insn = emit_insn (fp_adjust_insn); | |
7064dcad CJW |
3004 | |
3005 | /* The insn rtx 'fp_adjust_insn' will change frame layout. */ | |
3006 | RTX_FRAME_RELATED_P (fp_adjust_insn) = 1; | |
9304f876 CJW |
3007 | } |
3008 | ||
3009 | /* Adjust $sp = $sp - local_size - out_args_size | |
3010 | - callee_saved_area_padding_bytes. */ | |
3011 | sp_adjust = cfun->machine->local_size | |
3012 | + cfun->machine->out_args_size | |
3013 | + cfun->machine->callee_saved_area_padding_bytes; | |
3014 | /* sp_adjust value may be out of range of the addi instruction, | |
3015 | create alternative add behavior with TA_REGNUM if necessary, | |
3016 | using NEGATIVE value to tell that we are decreasing address. */ | |
3017 | sp_adjust = nds32_force_addi_stack_int ( (-1) * sp_adjust); | |
3018 | if (sp_adjust) | |
3019 | { | |
3020 | /* Generate sp adjustment instruction if and only if sp_adjust != 0. */ | |
3021 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3022 | stack_pointer_rtx, | |
3023 | GEN_INT (-1 * sp_adjust)); | |
3024 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3025 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
3026 | ||
3027 | /* The insn rtx 'sp_adjust_insn' will change frame layout. | |
3028 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
3029 | generate CFI (Call Frame Information) stuff. */ | |
3030 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
3031 | } | |
3032 | ||
3033 | /* Prevent the instruction scheduler from | |
3034 | moving instructions across the boundary. */ | |
3035 | emit_insn (gen_blockage ()); | |
3036 | } | |
3037 | ||
3038 | /* Function for normal multiple pop epilogue. */ | |
3039 | void | |
3040 | nds32_expand_epilogue (void) | |
3041 | { | |
3042 | int sp_adjust; | |
3043 | int en4_const; | |
3044 | ||
3045 | rtx Rb, Re; | |
9304f876 CJW |
3046 | rtx sp_adjust_insn; |
3047 | ||
3048 | /* Compute and setup stack frame size. | |
3049 | The result will be in cfun->machine. */ | |
3050 | nds32_compute_stack_frame (); | |
3051 | ||
3052 | /* Prevent the instruction scheduler from | |
3053 | moving instructions across the boundary. */ | |
3054 | emit_insn (gen_blockage ()); | |
3055 | ||
3056 | /* If the function is 'naked', we do not have to generate | |
557430f1 CJW |
3057 | epilogue code fragment BUT 'ret' instruction. |
3058 | However, if this function is also a variadic function, | |
3059 | we need to create adjust stack pointer before 'ret' instruction. */ | |
9304f876 CJW |
3060 | if (cfun->machine->naked_p) |
3061 | { | |
557430f1 CJW |
3062 | /* If this is a variadic function, we do not have to restore argument |
3063 | registers but need to adjust stack pointer back to previous stack | |
3064 | frame location before return. */ | |
3065 | if (cfun->machine->va_args_size != 0) | |
3066 | { | |
3067 | /* Generate sp adjustment instruction. | |
3068 | We need to consider padding bytes here. */ | |
3069 | sp_adjust = cfun->machine->va_args_size | |
3070 | + cfun->machine->va_args_area_padding_bytes; | |
3071 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3072 | stack_pointer_rtx, | |
3073 | GEN_INT (sp_adjust)); | |
3074 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3075 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
3076 | ||
3077 | /* The insn rtx 'sp_adjust_insn' will change frame layout. | |
3078 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
3079 | generate CFI (Call Frame Information) stuff. */ | |
3080 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
3081 | } | |
3082 | ||
9304f876 CJW |
3083 | /* Generate return instruction by using |
3084 | unspec_volatile_func_return pattern. | |
3085 | Make sure this instruction is after gen_blockage(). | |
3086 | NOTE that $lp will become 'live' | |
3087 | after this instruction has been emitted. */ | |
3088 | emit_insn (gen_unspec_volatile_func_return ()); | |
3089 | return; | |
3090 | } | |
3091 | ||
3092 | if (frame_pointer_needed) | |
3093 | { | |
3094 | /* adjust $sp = $fp - ($fp size) - ($gp size) - ($lp size) | |
3095 | - (4 * callee-saved-registers) | |
3096 | Note: No need to adjust | |
3097 | cfun->machine->callee_saved_area_padding_bytes, | |
3098 | because we want to adjust stack pointer | |
3099 | to the position for pop instruction. */ | |
3100 | sp_adjust = cfun->machine->fp_size | |
3101 | + cfun->machine->gp_size | |
3102 | + cfun->machine->lp_size | |
3103 | + cfun->machine->callee_saved_regs_size; | |
3104 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3105 | hard_frame_pointer_rtx, | |
3106 | GEN_INT (-1 * sp_adjust)); | |
3107 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3108 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
557430f1 CJW |
3109 | |
3110 | /* The insn rtx 'sp_adjust_insn' will change frame layout. */ | |
3111 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
9304f876 CJW |
3112 | } |
3113 | else | |
3114 | { | |
3115 | /* If frame pointer is NOT needed, | |
3116 | we cannot calculate the sp adjustment from frame pointer. | |
3117 | Instead, we calculate the adjustment by local_size, | |
3118 | out_args_size, and callee_saved_area_padding_bytes. | |
3119 | Notice that such sp adjustment value may be out of range, | |
3120 | so we have to deal with it as well. */ | |
3121 | ||
3122 | /* Adjust $sp = $sp + local_size + out_args_size | |
3123 | + callee_saved_area_padding_bytes. */ | |
3124 | sp_adjust = cfun->machine->local_size | |
3125 | + cfun->machine->out_args_size | |
3126 | + cfun->machine->callee_saved_area_padding_bytes; | |
3127 | /* sp_adjust value may be out of range of the addi instruction, | |
3128 | create alternative add behavior with TA_REGNUM if necessary, | |
3129 | using POSITIVE value to tell that we are increasing address. */ | |
3130 | sp_adjust = nds32_force_addi_stack_int (sp_adjust); | |
3131 | if (sp_adjust) | |
3132 | { | |
3133 | /* Generate sp adjustment instruction | |
3134 | if and only if sp_adjust != 0. */ | |
3135 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3136 | stack_pointer_rtx, | |
3137 | GEN_INT (sp_adjust)); | |
3138 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3139 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
557430f1 CJW |
3140 | |
3141 | /* The insn rtx 'sp_adjust_insn' will change frame layout. */ | |
3142 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
9304f876 CJW |
3143 | } |
3144 | } | |
3145 | ||
3146 | /* Get callee_first_regno and callee_last_regno. */ | |
3147 | Rb = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_first_regno); | |
3148 | Re = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_last_regno); | |
3149 | ||
4e9a2848 | 3150 | /* nds32_emit_stack_pop_multiple(first_regno, last_regno), |
9304f876 CJW |
3151 | the pattern 'stack_pop_multiple' is implementad in nds32.md. |
3152 | For En4 field, we have to calculate its constant value. | |
3153 | Refer to Andes ISA for more information. */ | |
3154 | en4_const = 0; | |
3155 | if (cfun->machine->fp_size) | |
3156 | en4_const += 8; | |
3157 | if (cfun->machine->gp_size) | |
3158 | en4_const += 4; | |
3159 | if (cfun->machine->lp_size) | |
3160 | en4_const += 2; | |
3161 | ||
3162 | /* If $fp, $gp, $lp, and all callee-save registers are NOT required | |
3163 | to be saved, we don't have to create multiple pop instruction. | |
3164 | Otherwise, a multiple pop instruction is needed. */ | |
3165 | if (!(REGNO (Rb) == SP_REGNUM && REGNO (Re) == SP_REGNUM && en4_const == 0)) | |
3166 | { | |
3167 | /* Create multiple pop instruction rtx. */ | |
4e9a2848 | 3168 | nds32_emit_stack_pop_multiple (Rb, Re, GEN_INT (en4_const)); |
9304f876 CJW |
3169 | } |
3170 | ||
557430f1 CJW |
3171 | /* If this is a variadic function, we do not have to restore argument |
3172 | registers but need to adjust stack pointer back to previous stack | |
3173 | frame location before return. */ | |
3174 | if (cfun->machine->va_args_size != 0) | |
3175 | { | |
3176 | /* Generate sp adjustment instruction. | |
3177 | We need to consider padding bytes here. */ | |
3178 | sp_adjust = cfun->machine->va_args_size | |
3179 | + cfun->machine->va_args_area_padding_bytes; | |
3180 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3181 | stack_pointer_rtx, | |
3182 | GEN_INT (sp_adjust)); | |
3183 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3184 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
3185 | ||
3186 | /* The insn rtx 'sp_adjust_insn' will change frame layout. | |
3187 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
3188 | generate CFI (Call Frame Information) stuff. */ | |
3189 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
3190 | } | |
3191 | ||
9304f876 CJW |
3192 | /* Generate return instruction by using |
3193 | unspec_volatile_func_return pattern. */ | |
3194 | emit_insn (gen_unspec_volatile_func_return ()); | |
3195 | } | |
3196 | ||
3197 | /* Function for v3push prologue. */ | |
3198 | void | |
3199 | nds32_expand_prologue_v3push (void) | |
3200 | { | |
3201 | int fp_adjust; | |
3202 | int sp_adjust; | |
3203 | ||
3204 | rtx Rb, Re; | |
9304f876 CJW |
3205 | rtx fp_adjust_insn, sp_adjust_insn; |
3206 | ||
9304f876 CJW |
3207 | /* Compute and setup stack frame size. |
3208 | The result will be in cfun->machine. */ | |
3209 | nds32_compute_stack_frame (); | |
3210 | ||
3211 | /* If the function is 'naked', | |
3212 | we do not have to generate prologue code fragment. */ | |
3213 | if (cfun->machine->naked_p) | |
3214 | return; | |
3215 | ||
3216 | /* Get callee_first_regno and callee_last_regno. */ | |
3217 | Rb = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_first_regno); | |
3218 | Re = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_last_regno); | |
3219 | ||
3220 | /* Calculate sp_adjust first to test if 'push25 Re,imm8u' is available, | |
3221 | where imm8u has to be 8-byte alignment. */ | |
3222 | sp_adjust = cfun->machine->local_size | |
3223 | + cfun->machine->out_args_size | |
3224 | + cfun->machine->callee_saved_area_padding_bytes; | |
3225 | ||
3226 | if (satisfies_constraint_Iu08 (GEN_INT (sp_adjust)) | |
3227 | && NDS32_DOUBLE_WORD_ALIGN_P (sp_adjust)) | |
3228 | { | |
3229 | /* We can use 'push25 Re,imm8u'. */ | |
3230 | ||
88437f39 | 3231 | /* nds32_emit_stack_v3push(last_regno, sp_adjust), |
9304f876 CJW |
3232 | the pattern 'stack_v3push' is implemented in nds32.md. |
3233 | The (const_int 14) means v3push always push { $fp $gp $lp }. */ | |
88437f39 CJW |
3234 | nds32_emit_stack_v3push (Rb, Re, |
3235 | GEN_INT (14), GEN_INT (sp_adjust)); | |
9304f876 CJW |
3236 | |
3237 | /* Check frame_pointer_needed to see | |
3238 | if we shall emit fp adjustment instruction. */ | |
3239 | if (frame_pointer_needed) | |
3240 | { | |
3241 | /* adjust $fp = $sp + 4 ($fp size) | |
3242 | + 4 ($gp size) | |
3243 | + 4 ($lp size) | |
3244 | + (4 * n) (callee-saved registers) | |
3245 | + sp_adjust ('push25 Re,imm8u') | |
3246 | Note: Since we use 'push25 Re,imm8u', | |
3247 | the position of stack pointer is further | |
3248 | changed after push instruction. | |
3249 | Hence, we need to take sp_adjust value | |
3250 | into consideration. */ | |
3251 | fp_adjust = cfun->machine->fp_size | |
3252 | + cfun->machine->gp_size | |
3253 | + cfun->machine->lp_size | |
3254 | + cfun->machine->callee_saved_regs_size | |
3255 | + sp_adjust; | |
3256 | fp_adjust_insn = gen_addsi3 (hard_frame_pointer_rtx, | |
3257 | stack_pointer_rtx, | |
3258 | GEN_INT (fp_adjust)); | |
3259 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3260 | fp_adjust_insn = emit_insn (fp_adjust_insn); | |
3261 | } | |
3262 | } | |
3263 | else | |
3264 | { | |
3265 | /* We have to use 'push25 Re,0' and | |
3266 | expand one more instruction to adjust $sp later. */ | |
3267 | ||
88437f39 | 3268 | /* nds32_emit_stack_v3push(last_regno, sp_adjust), |
9304f876 CJW |
3269 | the pattern 'stack_v3push' is implemented in nds32.md. |
3270 | The (const_int 14) means v3push always push { $fp $gp $lp }. */ | |
88437f39 CJW |
3271 | nds32_emit_stack_v3push (Rb, Re, |
3272 | GEN_INT (14), GEN_INT (0)); | |
9304f876 CJW |
3273 | |
3274 | /* Check frame_pointer_needed to see | |
3275 | if we shall emit fp adjustment instruction. */ | |
3276 | if (frame_pointer_needed) | |
3277 | { | |
3278 | /* adjust $fp = $sp + 4 ($fp size) | |
3279 | + 4 ($gp size) | |
3280 | + 4 ($lp size) | |
3281 | + (4 * n) (callee-saved registers) | |
3282 | Note: Since we use 'push25 Re,0', | |
3283 | the stack pointer is just at the position | |
3284 | after push instruction. | |
3285 | No need to take sp_adjust into consideration. */ | |
3286 | fp_adjust = cfun->machine->fp_size | |
3287 | + cfun->machine->gp_size | |
3288 | + cfun->machine->lp_size | |
3289 | + cfun->machine->callee_saved_regs_size; | |
3290 | fp_adjust_insn = gen_addsi3 (hard_frame_pointer_rtx, | |
3291 | stack_pointer_rtx, | |
3292 | GEN_INT (fp_adjust)); | |
3293 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3294 | fp_adjust_insn = emit_insn (fp_adjust_insn); | |
3295 | } | |
3296 | ||
3297 | /* Because we use 'push25 Re,0', | |
3298 | we need to expand one more instruction to adjust $sp. | |
3299 | However, sp_adjust value may be out of range of the addi instruction, | |
3300 | create alternative add behavior with TA_REGNUM if necessary, | |
3301 | using NEGATIVE value to tell that we are decreasing address. */ | |
3302 | sp_adjust = nds32_force_addi_stack_int ( (-1) * sp_adjust); | |
3303 | if (sp_adjust) | |
3304 | { | |
3305 | /* Generate sp adjustment instruction | |
3306 | if and only if sp_adjust != 0. */ | |
3307 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3308 | stack_pointer_rtx, | |
3309 | GEN_INT (-1 * sp_adjust)); | |
3310 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3311 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
3312 | ||
3313 | /* The insn rtx 'sp_adjust_insn' will change frame layout. | |
3314 | We need to use RTX_FRAME_RELATED_P so that GCC is able to | |
3315 | generate CFI (Call Frame Information) stuff. */ | |
3316 | RTX_FRAME_RELATED_P (sp_adjust_insn) = 1; | |
3317 | } | |
3318 | } | |
3319 | ||
3320 | /* Prevent the instruction scheduler from | |
3321 | moving instructions across the boundary. */ | |
3322 | emit_insn (gen_blockage ()); | |
3323 | } | |
3324 | ||
3325 | /* Function for v3pop epilogue. */ | |
3326 | void | |
3327 | nds32_expand_epilogue_v3pop (void) | |
3328 | { | |
3329 | int sp_adjust; | |
3330 | ||
3331 | rtx Rb, Re; | |
9304f876 CJW |
3332 | rtx sp_adjust_insn; |
3333 | ||
3334 | /* Compute and setup stack frame size. | |
3335 | The result will be in cfun->machine. */ | |
3336 | nds32_compute_stack_frame (); | |
3337 | ||
3338 | /* Prevent the instruction scheduler from | |
3339 | moving instructions across the boundary. */ | |
3340 | emit_insn (gen_blockage ()); | |
3341 | ||
3342 | /* If the function is 'naked', we do not have to generate | |
3343 | epilogue code fragment BUT 'ret' instruction. */ | |
3344 | if (cfun->machine->naked_p) | |
3345 | { | |
3346 | /* Generate return instruction by using | |
3347 | unspec_volatile_func_return pattern. | |
3348 | Make sure this instruction is after gen_blockage(). | |
3349 | NOTE that $lp will become 'live' | |
3350 | after this instruction has been emitted. */ | |
3351 | emit_insn (gen_unspec_volatile_func_return ()); | |
3352 | return; | |
3353 | } | |
3354 | ||
3355 | /* Get callee_first_regno and callee_last_regno. */ | |
3356 | Rb = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_first_regno); | |
3357 | Re = gen_rtx_REG (SImode, cfun->machine->callee_saved_regs_last_regno); | |
3358 | ||
3359 | /* Calculate sp_adjust first to test if 'pop25 Re,imm8u' is available, | |
3360 | where imm8u has to be 8-byte alignment. */ | |
3361 | sp_adjust = cfun->machine->local_size | |
3362 | + cfun->machine->out_args_size | |
3363 | + cfun->machine->callee_saved_area_padding_bytes; | |
3364 | ||
3365 | /* We have to consider alloca issue as well. | |
3366 | If the function does call alloca(), the stack pointer is not fixed. | |
3367 | In that case, we cannot use 'pop25 Re,imm8u' directly. | |
3368 | We have to caculate stack pointer from frame pointer | |
3369 | and then use 'pop25 Re,0'. | |
3370 | Of course, the frame_pointer_needed should be nonzero | |
3371 | if the function calls alloca(). */ | |
3372 | if (satisfies_constraint_Iu08 (GEN_INT (sp_adjust)) | |
3373 | && NDS32_DOUBLE_WORD_ALIGN_P (sp_adjust) | |
3374 | && !cfun->calls_alloca) | |
3375 | { | |
3376 | /* We can use 'pop25 Re,imm8u'. */ | |
3377 | ||
88437f39 | 3378 | /* nds32_emit_stack_v3pop(last_regno, sp_adjust), |
9304f876 CJW |
3379 | the pattern 'stack_v3pop' is implementad in nds32.md. |
3380 | The (const_int 14) means v3pop always pop { $fp $gp $lp }. */ | |
88437f39 CJW |
3381 | nds32_emit_stack_v3pop (Rb, Re, |
3382 | GEN_INT (14), GEN_INT (sp_adjust)); | |
9304f876 CJW |
3383 | } |
3384 | else | |
3385 | { | |
3386 | /* We have to use 'pop25 Re,0', and prior to it, | |
3387 | we must expand one more instruction to adjust $sp. */ | |
3388 | ||
3389 | if (frame_pointer_needed) | |
3390 | { | |
3391 | /* adjust $sp = $fp - 4 ($fp size) | |
3392 | - 4 ($gp size) | |
3393 | - 4 ($lp size) | |
3394 | - (4 * n) (callee-saved registers) | |
3395 | Note: No need to adjust | |
3396 | cfun->machine->callee_saved_area_padding_bytes, | |
3397 | because we want to adjust stack pointer | |
3398 | to the position for pop instruction. */ | |
3399 | sp_adjust = cfun->machine->fp_size | |
3400 | + cfun->machine->gp_size | |
3401 | + cfun->machine->lp_size | |
3402 | + cfun->machine->callee_saved_regs_size; | |
3403 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3404 | hard_frame_pointer_rtx, | |
3405 | GEN_INT (-1 * sp_adjust)); | |
3406 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3407 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
3408 | } | |
3409 | else | |
3410 | { | |
3411 | /* If frame pointer is NOT needed, | |
3412 | we cannot calculate the sp adjustment from frame pointer. | |
3413 | Instead, we calculate the adjustment by local_size, | |
3414 | out_args_size, and callee_saved_area_padding_bytes. | |
3415 | Notice that such sp adjustment value may be out of range, | |
3416 | so we have to deal with it as well. */ | |
3417 | ||
3418 | /* Adjust $sp = $sp + local_size + out_args_size | |
3419 | + callee_saved_area_padding_bytes. */ | |
3420 | sp_adjust = cfun->machine->local_size | |
3421 | + cfun->machine->out_args_size | |
3422 | + cfun->machine->callee_saved_area_padding_bytes; | |
3423 | /* sp_adjust value may be out of range of the addi instruction, | |
3424 | create alternative add behavior with TA_REGNUM if necessary, | |
3425 | using POSITIVE value to tell that we are increasing address. */ | |
3426 | sp_adjust = nds32_force_addi_stack_int (sp_adjust); | |
3427 | if (sp_adjust) | |
3428 | { | |
3429 | /* Generate sp adjustment instruction | |
3430 | if and only if sp_adjust != 0. */ | |
3431 | sp_adjust_insn = gen_addsi3 (stack_pointer_rtx, | |
3432 | stack_pointer_rtx, | |
3433 | GEN_INT (sp_adjust)); | |
3434 | /* Emit rtx into instructions list and receive INSN rtx form. */ | |
3435 | sp_adjust_insn = emit_insn (sp_adjust_insn); | |
3436 | } | |
3437 | } | |
3438 | ||
88437f39 | 3439 | /* nds32_emit_stack_v3pop(last_regno, sp_adjust), |
9304f876 CJW |
3440 | the pattern 'stack_v3pop' is implementad in nds32.md. */ |
3441 | /* The (const_int 14) means v3pop always pop { $fp $gp $lp }. */ | |
88437f39 CJW |
3442 | nds32_emit_stack_v3pop (Rb, Re, |
3443 | GEN_INT (14), GEN_INT (0)); | |
9304f876 CJW |
3444 | } |
3445 | } | |
3446 | ||
3447 | /* ------------------------------------------------------------------------ */ | |
3448 | ||
3449 | /* Function to test 333-form for load/store instructions. | |
3450 | This is auxiliary extern function for auxiliary macro in nds32.h. | |
3451 | Because it is a little complicated, we use function instead of macro. */ | |
3452 | bool | |
ef4bddc2 | 3453 | nds32_ls_333_p (rtx rt, rtx ra, rtx imm, machine_mode mode) |
9304f876 CJW |
3454 | { |
3455 | if (REGNO_REG_CLASS (REGNO (rt)) == LOW_REGS | |
3456 | && REGNO_REG_CLASS (REGNO (ra)) == LOW_REGS) | |
3457 | { | |
3458 | if (GET_MODE_SIZE (mode) == 4) | |
3459 | return satisfies_constraint_Iu05 (imm); | |
3460 | ||
3461 | if (GET_MODE_SIZE (mode) == 2) | |
3462 | return satisfies_constraint_Iu04 (imm); | |
3463 | ||
3464 | if (GET_MODE_SIZE (mode) == 1) | |
3465 | return satisfies_constraint_Iu03 (imm); | |
3466 | } | |
3467 | ||
3468 | return false; | |
3469 | } | |
3470 | ||
3471 | ||
9304f876 CJW |
3472 | /* Computing the Length of an Insn. |
3473 | Modifies the length assigned to instruction INSN. | |
3474 | LEN is the initially computed length of the insn. */ | |
3475 | int | |
ca009aee | 3476 | nds32_adjust_insn_length (rtx_insn *insn, int length) |
9304f876 CJW |
3477 | { |
3478 | rtx src, dst; | |
3479 | ||
3480 | switch (recog_memoized (insn)) | |
3481 | { | |
3482 | case CODE_FOR_move_df: | |
3483 | case CODE_FOR_move_di: | |
3484 | /* Adjust length of movd44 to 2. */ | |
3485 | src = XEXP (PATTERN (insn), 1); | |
3486 | dst = XEXP (PATTERN (insn), 0); | |
3487 | ||
3488 | if (REG_P (src) | |
3489 | && REG_P (dst) | |
3490 | && (REGNO (src) % 2) == 0 | |
3491 | && (REGNO (dst) % 2) == 0) | |
3492 | length = 2; | |
3493 | break; | |
3494 | ||
3495 | default: | |
3496 | break; | |
3497 | } | |
3498 | ||
3499 | return length; | |
3500 | } | |
3501 | ||
3502 | ||
9304f876 CJW |
3503 | /* Return align 2 (log base 2) if the next instruction of LABEL is 4 byte. */ |
3504 | int | |
3505 | nds32_target_alignment (rtx label) | |
3506 | { | |
84034c69 | 3507 | rtx_insn *insn; |
9304f876 CJW |
3508 | |
3509 | if (optimize_size) | |
3510 | return 0; | |
3511 | ||
3512 | insn = next_active_insn (label); | |
3513 | ||
3514 | if (insn == 0) | |
3515 | return 0; | |
3516 | else if ((get_attr_length (insn) % 4) == 0) | |
3517 | return 2; | |
3518 | else | |
3519 | return 0; | |
3520 | } | |
3521 | ||
3522 | /* ------------------------------------------------------------------------ */ | |
3523 | ||
3524 | /* PART 5: Initialize target hook structure and definitions. */ | |
3525 | \f | |
3526 | /* Controlling the Compilation Driver. */ | |
3527 | ||
3528 | \f | |
3529 | /* Run-time Target Specification. */ | |
3530 | ||
3531 | \f | |
3532 | /* Defining Data Structures for Per-function Information. */ | |
3533 | ||
3534 | \f | |
3535 | /* Storage Layout. */ | |
3536 | ||
3537 | #undef TARGET_PROMOTE_FUNCTION_MODE | |
3538 | #define TARGET_PROMOTE_FUNCTION_MODE \ | |
3539 | default_promote_function_mode_always_promote | |
3540 | ||
3541 | \f | |
3542 | /* Layout of Source Language Data Types. */ | |
3543 | ||
3544 | \f | |
3545 | /* Register Usage. */ | |
3546 | ||
3547 | /* -- Basic Characteristics of Registers. */ | |
3548 | ||
3549 | /* -- Order of Allocation of Registers. */ | |
3550 | ||
3551 | /* -- How Values Fit in Registers. */ | |
3552 | ||
3553 | /* -- Handling Leaf Functions. */ | |
3554 | ||
3555 | /* -- Registers That Form a Stack. */ | |
3556 | ||
3557 | \f | |
3558 | /* Register Classes. */ | |
3559 | ||
3560 | #undef TARGET_CLASS_MAX_NREGS | |
3561 | #define TARGET_CLASS_MAX_NREGS nds32_class_max_nregs | |
3562 | ||
3563 | #undef TARGET_LRA_P | |
3564 | #define TARGET_LRA_P hook_bool_void_true | |
3565 | ||
3566 | #undef TARGET_REGISTER_PRIORITY | |
3567 | #define TARGET_REGISTER_PRIORITY nds32_register_priority | |
3568 | ||
3569 | \f | |
3570 | /* Obsolete Macros for Defining Constraints. */ | |
3571 | ||
3572 | \f | |
3573 | /* Stack Layout and Calling Conventions. */ | |
3574 | ||
3575 | /* -- Basic Stack Layout. */ | |
3576 | ||
3577 | /* -- Exception Handling Support. */ | |
3578 | ||
3579 | /* -- Specifying How Stack Checking is Done. */ | |
3580 | ||
3581 | /* -- Registers That Address the Stack Frame. */ | |
3582 | ||
3583 | /* -- Eliminating Frame Pointer and Arg Pointer. */ | |
3584 | ||
3585 | #undef TARGET_CAN_ELIMINATE | |
3586 | #define TARGET_CAN_ELIMINATE nds32_can_eliminate | |
3587 | ||
3588 | /* -- Passing Function Arguments on the Stack. */ | |
3589 | ||
3590 | /* -- Passing Arguments in Registers. */ | |
3591 | ||
3592 | #undef TARGET_FUNCTION_ARG | |
3593 | #define TARGET_FUNCTION_ARG nds32_function_arg | |
3594 | ||
d40f3c40 CJW |
3595 | #undef TARGET_MUST_PASS_IN_STACK |
3596 | #define TARGET_MUST_PASS_IN_STACK nds32_must_pass_in_stack | |
3597 | ||
650fc469 CJW |
3598 | #undef TARGET_ARG_PARTIAL_BYTES |
3599 | #define TARGET_ARG_PARTIAL_BYTES nds32_arg_partial_bytes | |
3600 | ||
9304f876 CJW |
3601 | #undef TARGET_FUNCTION_ARG_ADVANCE |
3602 | #define TARGET_FUNCTION_ARG_ADVANCE nds32_function_arg_advance | |
3603 | ||
3604 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
3605 | #define TARGET_FUNCTION_ARG_BOUNDARY nds32_function_arg_boundary | |
3606 | ||
3607 | /* -- How Scalar Function Values Are Returned. */ | |
3608 | ||
3609 | #undef TARGET_FUNCTION_VALUE | |
3610 | #define TARGET_FUNCTION_VALUE nds32_function_value | |
3611 | ||
3612 | #undef TARGET_LIBCALL_VALUE | |
3613 | #define TARGET_LIBCALL_VALUE nds32_libcall_value | |
3614 | ||
3615 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
3616 | #define TARGET_FUNCTION_VALUE_REGNO_P nds32_function_value_regno_p | |
3617 | ||
3618 | /* -- How Large Values Are Returned. */ | |
3619 | ||
3620 | /* -- Caller-Saves Register Allocation. */ | |
3621 | ||
3622 | /* -- Function Entry and Exit. */ | |
3623 | ||
3624 | #undef TARGET_ASM_FUNCTION_PROLOGUE | |
3625 | #define TARGET_ASM_FUNCTION_PROLOGUE nds32_asm_function_prologue | |
3626 | ||
3627 | #undef TARGET_ASM_FUNCTION_END_PROLOGUE | |
3628 | #define TARGET_ASM_FUNCTION_END_PROLOGUE nds32_asm_function_end_prologue | |
3629 | ||
3630 | #undef TARGET_ASM_FUNCTION_BEGIN_EPILOGUE | |
3631 | #define TARGET_ASM_FUNCTION_BEGIN_EPILOGUE nds32_asm_function_begin_epilogue | |
3632 | ||
3633 | #undef TARGET_ASM_FUNCTION_EPILOGUE | |
3634 | #define TARGET_ASM_FUNCTION_EPILOGUE nds32_asm_function_epilogue | |
3635 | ||
3636 | #undef TARGET_ASM_OUTPUT_MI_THUNK | |
3637 | #define TARGET_ASM_OUTPUT_MI_THUNK nds32_asm_output_mi_thunk | |
3638 | ||
3639 | #undef TARGET_ASM_CAN_OUTPUT_MI_THUNK | |
3640 | #define TARGET_ASM_CAN_OUTPUT_MI_THUNK default_can_output_mi_thunk_no_vcall | |
3641 | ||
3642 | /* -- Generating Code for Profiling. */ | |
3643 | ||
3644 | /* -- Permitting tail calls. */ | |
3645 | ||
3646 | #undef TARGET_WARN_FUNC_RETURN | |
3647 | #define TARGET_WARN_FUNC_RETURN nds32_warn_func_return | |
3648 | ||
3649 | /* Stack smashing protection. */ | |
3650 | ||
3651 | \f | |
3652 | /* Implementing the Varargs Macros. */ | |
3653 | ||
d4a6a4d9 CJW |
3654 | #undef TARGET_SETUP_INCOMING_VARARGS |
3655 | #define TARGET_SETUP_INCOMING_VARARGS nds32_setup_incoming_varargs | |
3656 | ||
9304f876 CJW |
3657 | #undef TARGET_STRICT_ARGUMENT_NAMING |
3658 | #define TARGET_STRICT_ARGUMENT_NAMING nds32_strict_argument_naming | |
3659 | ||
3660 | \f | |
3661 | /* Trampolines for Nested Functions. */ | |
3662 | ||
3663 | #undef TARGET_ASM_TRAMPOLINE_TEMPLATE | |
3664 | #define TARGET_ASM_TRAMPOLINE_TEMPLATE nds32_asm_trampoline_template | |
3665 | ||
3666 | #undef TARGET_TRAMPOLINE_INIT | |
3667 | #define TARGET_TRAMPOLINE_INIT nds32_trampoline_init | |
3668 | ||
3669 | \f | |
3670 | /* Implicit Calls to Library Routines. */ | |
3671 | ||
3672 | \f | |
3673 | /* Addressing Modes. */ | |
3674 | ||
3675 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
3676 | #define TARGET_LEGITIMATE_ADDRESS_P nds32_legitimate_address_p | |
3677 | ||
3678 | \f | |
3679 | /* Anchored Addresses. */ | |
3680 | ||
3681 | \f | |
3682 | /* Condition Code Status. */ | |
3683 | ||
3684 | /* -- Representation of condition codes using (cc0). */ | |
3685 | ||
3686 | /* -- Representation of condition codes using registers. */ | |
3687 | ||
3688 | /* -- Macros to control conditional execution. */ | |
3689 | ||
3690 | \f | |
3691 | /* Describing Relative Costs of Operations. */ | |
3692 | ||
3693 | #undef TARGET_REGISTER_MOVE_COST | |
3694 | #define TARGET_REGISTER_MOVE_COST nds32_register_move_cost | |
3695 | ||
3696 | #undef TARGET_MEMORY_MOVE_COST | |
3697 | #define TARGET_MEMORY_MOVE_COST nds32_memory_move_cost | |
3698 | ||
3699 | #undef TARGET_RTX_COSTS | |
3700 | #define TARGET_RTX_COSTS nds32_rtx_costs | |
3701 | ||
3702 | #undef TARGET_ADDRESS_COST | |
3703 | #define TARGET_ADDRESS_COST nds32_address_cost | |
3704 | ||
3705 | \f | |
3706 | /* Adjusting the Instruction Scheduler. */ | |
3707 | ||
3708 | \f | |
3709 | /* Dividing the Output into Sections (Texts, Data, . . . ). */ | |
3710 | ||
511a41d7 CJW |
3711 | #undef TARGET_ENCODE_SECTION_INFO |
3712 | #define TARGET_ENCODE_SECTION_INFO nds32_encode_section_info | |
3713 | ||
9304f876 CJW |
3714 | \f |
3715 | /* Position Independent Code. */ | |
3716 | ||
3717 | \f | |
3718 | /* Defining the Output Assembler Language. */ | |
3719 | ||
3720 | /* -- The Overall Framework of an Assembler File. */ | |
3721 | ||
3722 | #undef TARGET_ASM_FILE_START | |
3723 | #define TARGET_ASM_FILE_START nds32_asm_file_start | |
3724 | #undef TARGET_ASM_FILE_END | |
3725 | #define TARGET_ASM_FILE_END nds32_asm_file_end | |
3726 | ||
3727 | /* -- Output of Data. */ | |
3728 | ||
3729 | #undef TARGET_ASM_ALIGNED_HI_OP | |
3730 | #define TARGET_ASM_ALIGNED_HI_OP "\t.hword\t" | |
3731 | ||
3732 | #undef TARGET_ASM_ALIGNED_SI_OP | |
3733 | #define TARGET_ASM_ALIGNED_SI_OP "\t.word\t" | |
3734 | ||
3735 | /* -- Output of Uninitialized Variables. */ | |
3736 | ||
3737 | /* -- Output and Generation of Labels. */ | |
3738 | ||
3739 | #undef TARGET_ASM_GLOBALIZE_LABEL | |
3740 | #define TARGET_ASM_GLOBALIZE_LABEL nds32_asm_globalize_label | |
3741 | ||
3742 | /* -- How Initialization Functions Are Handled. */ | |
3743 | ||
3744 | /* -- Macros Controlling Initialization Routines. */ | |
3745 | ||
3746 | /* -- Output of Assembler Instructions. */ | |
3747 | ||
3748 | #undef TARGET_PRINT_OPERAND | |
3749 | #define TARGET_PRINT_OPERAND nds32_print_operand | |
3750 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
3751 | #define TARGET_PRINT_OPERAND_ADDRESS nds32_print_operand_address | |
3752 | ||
3753 | /* -- Output of Dispatch Tables. */ | |
3754 | ||
3755 | /* -- Assembler Commands for Exception Regions. */ | |
3756 | ||
3757 | /* -- Assembler Commands for Alignment. */ | |
3758 | ||
3759 | \f | |
3760 | /* Controlling Debugging Information Format. */ | |
3761 | ||
3762 | /* -- Macros Affecting All Debugging Formats. */ | |
3763 | ||
3764 | /* -- Specific Options for DBX Output. */ | |
3765 | ||
3766 | /* -- Open-Ended Hooks for DBX Format. */ | |
3767 | ||
3768 | /* -- File Names in DBX Format. */ | |
3769 | ||
3770 | /* -- Macros for SDB and DWARF Output. */ | |
3771 | ||
3772 | /* -- Macros for VMS Debug Format. */ | |
3773 | ||
3774 | \f | |
3775 | /* Cross Compilation and Floating Point. */ | |
3776 | ||
3777 | \f | |
3778 | /* Mode Switching Instructions. */ | |
3779 | ||
3780 | \f | |
3781 | /* Defining target-specific uses of __attribute__. */ | |
3782 | ||
3783 | #undef TARGET_ATTRIBUTE_TABLE | |
3784 | #define TARGET_ATTRIBUTE_TABLE nds32_attribute_table | |
3785 | ||
3786 | #undef TARGET_MERGE_DECL_ATTRIBUTES | |
3787 | #define TARGET_MERGE_DECL_ATTRIBUTES nds32_merge_decl_attributes | |
3788 | ||
3789 | #undef TARGET_INSERT_ATTRIBUTES | |
3790 | #define TARGET_INSERT_ATTRIBUTES nds32_insert_attributes | |
3791 | ||
3792 | #undef TARGET_OPTION_PRAGMA_PARSE | |
3793 | #define TARGET_OPTION_PRAGMA_PARSE nds32_option_pragma_parse | |
3794 | ||
3795 | #undef TARGET_OPTION_OVERRIDE | |
3796 | #define TARGET_OPTION_OVERRIDE nds32_option_override | |
3797 | ||
3798 | \f | |
3799 | /* Emulating TLS. */ | |
3800 | ||
3801 | \f | |
3802 | /* Defining coprocessor specifics for MIPS targets. */ | |
3803 | ||
3804 | \f | |
3805 | /* Parameters for Precompiled Header Validity Checking. */ | |
3806 | ||
3807 | \f | |
3808 | /* C++ ABI parameters. */ | |
3809 | ||
3810 | \f | |
3811 | /* Adding support for named address spaces. */ | |
3812 | ||
3813 | \f | |
3814 | /* Miscellaneous Parameters. */ | |
3815 | ||
3816 | #undef TARGET_INIT_BUILTINS | |
3817 | #define TARGET_INIT_BUILTINS nds32_init_builtins | |
3818 | ||
3819 | #undef TARGET_EXPAND_BUILTIN | |
3820 | #define TARGET_EXPAND_BUILTIN nds32_expand_builtin | |
3821 | ||
3822 | \f | |
3823 | /* ------------------------------------------------------------------------ */ | |
3824 | ||
3825 | /* Initialize the GCC target structure. */ | |
3826 | ||
3827 | struct gcc_target targetm = TARGET_INITIALIZER; | |
3828 | ||
3829 | /* ------------------------------------------------------------------------ */ |