]>
Commit | Line | Data |
---|---|---|
8f90be4c | 1 | /* Output routines for Motorola MCore processor |
d6edb99e | 2 | Copyright (C) 1993, 1999, 2000, 2001 Free Software Foundation, Inc. |
8f90be4c NC |
3 | |
4 | This file is part of GNU CC. | |
5 | ||
6 | GNU CC is free software; you can redistribute it and/or modify | |
7 | it under the terms of the GNU General Public License as published by | |
8 | the Free Software Foundation; either version 2, or (at your option) | |
9 | any later version. | |
10 | ||
11 | GNU CC is distributed in the hope that it will be useful, | |
12 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | GNU General Public License for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GNU CC; see the file COPYING. If not, write to | |
18 | the Free Software Foundation, 675 Mass Ave, Cambridge, MA 02139, USA. */ | |
19 | ||
bc27e96c | 20 | #include "config.h" |
4bd048ef | 21 | #include "system.h" |
4816b8e4 NC |
22 | #include "rtl.h" |
23 | #include "tree.h" | |
24 | #include "tm_p.h" | |
8f90be4c | 25 | #include "assert.h" |
8f90be4c | 26 | #include "mcore.h" |
8f90be4c NC |
27 | #include "regs.h" |
28 | #include "hard-reg-set.h" | |
29 | #include "real.h" | |
30 | #include "insn-config.h" | |
31 | #include "conditions.h" | |
8f90be4c NC |
32 | #include "output.h" |
33 | #include "insn-attr.h" | |
34 | #include "flags.h" | |
35 | #include "obstack.h" | |
36 | #include "expr.h" | |
37 | #include "reload.h" | |
38 | #include "recog.h" | |
39 | #include "function.h" | |
40 | #include "ggc.h" | |
41 | #include "toplev.h" | |
8f90be4c | 42 | |
8f90be4c NC |
43 | /* Maximum size we are allowed to grow the stack in a single operation. |
44 | If we want more, we must do it in increments of at most this size. | |
45 | If this value is 0, we don't check at all. */ | |
46 | const char * mcore_stack_increment_string = 0; | |
47 | int mcore_stack_increment = STACK_UNITS_MAXSTEP; | |
48 | ||
49 | /* For dumping information about frame sizes. */ | |
50 | char * mcore_current_function_name = 0; | |
51 | long mcore_current_compilation_timestamp = 0; | |
52 | ||
53 | /* Global variables for machine-dependent things. */ | |
54 | ||
55 | /* Saved operands from the last compare to use when we generate an scc | |
56 | or bcc insn. */ | |
57 | rtx arch_compare_op0; | |
58 | rtx arch_compare_op1; | |
59 | ||
60 | /* Provides the class number of the smallest class containing | |
61 | reg number. */ | |
62 | int regno_reg_class[FIRST_PSEUDO_REGISTER] = | |
63 | { | |
64 | GENERAL_REGS, ONLYR1_REGS, LRW_REGS, LRW_REGS, | |
65 | LRW_REGS, LRW_REGS, LRW_REGS, LRW_REGS, | |
66 | LRW_REGS, LRW_REGS, LRW_REGS, LRW_REGS, | |
67 | LRW_REGS, LRW_REGS, LRW_REGS, GENERAL_REGS, | |
68 | GENERAL_REGS, C_REGS, NO_REGS, NO_REGS, | |
69 | }; | |
70 | ||
71 | /* Provide reg_class from a letter such as appears in the machine | |
72 | description. */ | |
73 | enum reg_class reg_class_from_letter[] = | |
74 | { | |
75 | /* a */ LRW_REGS, /* b */ ONLYR1_REGS, /* c */ C_REGS, /* d */ NO_REGS, | |
76 | /* e */ NO_REGS, /* f */ NO_REGS, /* g */ NO_REGS, /* h */ NO_REGS, | |
77 | /* i */ NO_REGS, /* j */ NO_REGS, /* k */ NO_REGS, /* l */ NO_REGS, | |
78 | /* m */ NO_REGS, /* n */ NO_REGS, /* o */ NO_REGS, /* p */ NO_REGS, | |
79 | /* q */ NO_REGS, /* r */ GENERAL_REGS, /* s */ NO_REGS, /* t */ NO_REGS, | |
80 | /* u */ NO_REGS, /* v */ NO_REGS, /* w */ NO_REGS, /* x */ ALL_REGS, | |
81 | /* y */ NO_REGS, /* z */ NO_REGS | |
82 | }; | |
83 | ||
f27cd94d NC |
84 | struct mcore_frame |
85 | { | |
86 | int arg_size; /* stdarg spills (bytes) */ | |
87 | int reg_size; /* non-volatile reg saves (bytes) */ | |
88 | int reg_mask; /* non-volatile reg saves */ | |
89 | int local_size; /* locals */ | |
90 | int outbound_size; /* arg overflow on calls out */ | |
91 | int pad_outbound; | |
92 | int pad_local; | |
93 | int pad_reg; | |
94 | /* Describe the steps we'll use to grow it. */ | |
95 | #define MAX_STACK_GROWS 4 /* gives us some spare space */ | |
96 | int growth[MAX_STACK_GROWS]; | |
97 | int arg_offset; | |
98 | int reg_offset; | |
99 | int reg_growth; | |
100 | int local_growth; | |
101 | }; | |
102 | ||
103 | typedef enum | |
104 | { | |
105 | COND_NO, | |
106 | COND_MOV_INSN, | |
107 | COND_CLR_INSN, | |
108 | COND_INC_INSN, | |
109 | COND_DEC_INSN, | |
110 | COND_BRANCH_INSN | |
111 | } | |
112 | cond_type; | |
113 | ||
114 | static void output_stack_adjust PARAMS ((int, int)); | |
115 | static int calc_live_regs PARAMS ((int *)); | |
116 | static int const_ok_for_mcore PARAMS ((int)); | |
117 | static int try_constant_tricks PARAMS ((long, int *, int *)); | |
118 | static const char * output_inline_const PARAMS ((enum machine_mode, rtx *)); | |
119 | static void block_move_sequence PARAMS ((rtx, rtx, rtx, rtx, int, int, int)); | |
120 | static void layout_mcore_frame PARAMS ((struct mcore_frame *)); | |
121 | static cond_type is_cond_candidate PARAMS ((rtx)); | |
122 | static rtx emit_new_cond_insn PARAMS ((rtx, int)); | |
123 | static rtx conditionalize_block PARAMS ((rtx)); | |
124 | static void conditionalize_optimization PARAMS ((rtx)); | |
125 | static void mcore_add_gc_roots PARAMS ((void)); | |
126 | static rtx handle_structs_in_regs PARAMS ((enum machine_mode, tree, int)); | |
127 | static void mcore_mark_dllexport PARAMS ((tree)); | |
128 | static void mcore_mark_dllimport PARAMS ((tree)); | |
129 | static int mcore_dllexport_p PARAMS ((tree)); | |
130 | static int mcore_dllimport_p PARAMS ((tree)); | |
131 | \f | |
8f90be4c NC |
132 | /* Adjust the stack and return the number of bytes taken to do it. */ |
133 | static void | |
134 | output_stack_adjust (direction, size) | |
135 | int direction; | |
136 | int size; | |
137 | { | |
4816b8e4 | 138 | /* If extending stack a lot, we do it incrementally. */ |
8f90be4c NC |
139 | if (direction < 0 && size > mcore_stack_increment && mcore_stack_increment > 0) |
140 | { | |
141 | rtx tmp = gen_rtx (REG, SImode, 1); | |
142 | rtx memref; | |
143 | emit_insn (gen_movsi (tmp, GEN_INT (mcore_stack_increment))); | |
144 | do | |
145 | { | |
146 | emit_insn (gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, tmp)); | |
147 | memref = gen_rtx (MEM, SImode, stack_pointer_rtx); | |
148 | MEM_VOLATILE_P (memref) = 1; | |
149 | emit_insn (gen_movsi (memref, stack_pointer_rtx)); | |
150 | size -= mcore_stack_increment; | |
151 | } | |
152 | while (size > mcore_stack_increment); | |
153 | ||
4816b8e4 NC |
154 | /* SIZE is now the residual for the last adjustment, |
155 | which doesn't require a probe. */ | |
8f90be4c NC |
156 | } |
157 | ||
158 | if (size) | |
159 | { | |
160 | rtx insn; | |
161 | rtx val = GEN_INT (size); | |
162 | ||
163 | if (size > 32) | |
164 | { | |
165 | rtx nval = gen_rtx (REG, SImode, 1); | |
166 | emit_insn (gen_movsi (nval, val)); | |
167 | val = nval; | |
168 | } | |
169 | ||
170 | if (direction > 0) | |
171 | insn = gen_addsi3 (stack_pointer_rtx, stack_pointer_rtx, val); | |
172 | else | |
173 | insn = gen_subsi3 (stack_pointer_rtx, stack_pointer_rtx, val); | |
174 | ||
175 | emit_insn (insn); | |
176 | } | |
177 | } | |
178 | ||
4816b8e4 NC |
179 | /* Work out the registers which need to be saved, |
180 | both as a mask and a count. */ | |
181 | ||
8f90be4c NC |
182 | static int |
183 | calc_live_regs (count) | |
184 | int * count; | |
185 | { | |
186 | int reg; | |
187 | int live_regs_mask = 0; | |
188 | ||
189 | * count = 0; | |
190 | ||
191 | for (reg = 0; reg < FIRST_PSEUDO_REGISTER; reg++) | |
192 | { | |
193 | if (regs_ever_live[reg] && !call_used_regs[reg]) | |
194 | { | |
195 | (*count)++; | |
196 | live_regs_mask |= (1 << reg); | |
197 | } | |
198 | } | |
199 | ||
200 | return live_regs_mask; | |
201 | } | |
202 | ||
203 | /* Print the operand address in x to the stream. */ | |
4816b8e4 | 204 | |
8f90be4c NC |
205 | void |
206 | mcore_print_operand_address (stream, x) | |
207 | FILE * stream; | |
208 | rtx x; | |
209 | { | |
210 | switch (GET_CODE (x)) | |
211 | { | |
212 | case REG: | |
213 | fprintf (stream, "(%s)", reg_names[REGNO (x)]); | |
214 | break; | |
215 | ||
216 | case PLUS: | |
217 | { | |
218 | rtx base = XEXP (x, 0); | |
219 | rtx index = XEXP (x, 1); | |
220 | ||
221 | if (GET_CODE (base) != REG) | |
222 | { | |
223 | /* Ensure that BASE is a register (one of them must be). */ | |
224 | rtx temp = base; | |
225 | base = index; | |
226 | index = temp; | |
227 | } | |
228 | ||
229 | switch (GET_CODE (index)) | |
230 | { | |
231 | case CONST_INT: | |
232 | fprintf (stream, "(%s,%d)", reg_names[REGNO(base)], | |
233 | INTVAL (index)); | |
234 | break; | |
235 | ||
236 | default: | |
237 | debug_rtx (x); | |
238 | ||
239 | abort (); | |
240 | } | |
241 | } | |
242 | ||
243 | break; | |
244 | ||
245 | default: | |
246 | output_addr_const (stream, x); | |
247 | break; | |
248 | } | |
249 | } | |
250 | ||
251 | /* Print operand x (an rtx) in assembler syntax to file stream | |
252 | according to modifier code. | |
253 | ||
254 | 'R' print the next register or memory location along, ie the lsw in | |
255 | a double word value | |
256 | 'O' print a constant without the # | |
257 | 'M' print a constant as its negative | |
258 | 'P' print log2 of a power of two | |
259 | 'Q' print log2 of an inverse of a power of two | |
260 | 'U' print register for ldm/stm instruction | |
4816b8e4 NC |
261 | 'X' print byte number for xtrbN instruction. */ |
262 | ||
8f90be4c NC |
263 | void |
264 | mcore_print_operand (stream, x, code) | |
265 | FILE * stream; | |
266 | rtx x; | |
267 | int code; | |
268 | { | |
269 | switch (code) | |
270 | { | |
271 | case 'N': | |
272 | if (INTVAL(x) == -1) | |
273 | fprintf (asm_out_file, "32"); | |
274 | else | |
275 | fprintf (asm_out_file, "%d", exact_log2 (INTVAL (x) + 1)); | |
276 | break; | |
277 | case 'P': | |
278 | fprintf (asm_out_file, "%d", exact_log2 (INTVAL (x))); | |
279 | break; | |
280 | case 'Q': | |
281 | fprintf (asm_out_file, "%d", exact_log2 (~INTVAL (x))); | |
282 | break; | |
283 | case 'O': | |
284 | fprintf (asm_out_file, "%d", INTVAL (x)); | |
285 | break; | |
286 | case 'M': | |
287 | fprintf (asm_out_file, "%d", - INTVAL (x)); | |
288 | break; | |
289 | case 'R': | |
290 | /* Next location along in memory or register. */ | |
291 | switch (GET_CODE (x)) | |
292 | { | |
293 | case REG: | |
294 | fputs (reg_names[REGNO (x) + 1], (stream)); | |
295 | break; | |
296 | case MEM: | |
297 | mcore_print_operand_address (stream, | |
298 | XEXP (adj_offsettable_operand (x, 4), 0)); | |
299 | break; | |
300 | default: | |
301 | abort (); | |
302 | } | |
303 | break; | |
304 | case 'U': | |
305 | fprintf (asm_out_file, "%s-%s", reg_names[REGNO (x)], | |
306 | reg_names[REGNO (x) + 3]); | |
307 | break; | |
308 | case 'x': | |
309 | fprintf (asm_out_file, "0x%x", INTVAL (x)); | |
310 | break; | |
311 | case 'X': | |
312 | fprintf (asm_out_file, "%d", 3 - INTVAL (x) / 8); | |
313 | break; | |
314 | ||
315 | default: | |
316 | switch (GET_CODE (x)) | |
317 | { | |
318 | case REG: | |
319 | fputs (reg_names[REGNO (x)], (stream)); | |
320 | break; | |
321 | case MEM: | |
322 | output_address (XEXP (x, 0)); | |
323 | break; | |
324 | default: | |
325 | output_addr_const (stream, x); | |
326 | break; | |
327 | } | |
328 | break; | |
329 | } | |
330 | } | |
331 | ||
332 | /* What does a constant cost ? */ | |
4816b8e4 | 333 | |
8f90be4c NC |
334 | int |
335 | mcore_const_costs (exp, code) | |
336 | rtx exp; | |
337 | enum rtx_code code; | |
338 | { | |
339 | ||
340 | int val = INTVAL (exp); | |
341 | ||
342 | /* Easy constants. */ | |
343 | if ( CONST_OK_FOR_I (val) | |
344 | || CONST_OK_FOR_M (val) | |
345 | || CONST_OK_FOR_N (val) | |
346 | || (code == PLUS && CONST_OK_FOR_L (val))) | |
347 | return 1; | |
348 | else if (code == AND | |
349 | && ( CONST_OK_FOR_M (~val) | |
350 | || CONST_OK_FOR_N (~val))) | |
351 | return 2; | |
352 | else if (code == PLUS | |
353 | && ( CONST_OK_FOR_I (-val) | |
354 | || CONST_OK_FOR_M (-val) | |
355 | || CONST_OK_FOR_N (-val))) | |
356 | return 2; | |
357 | ||
358 | return 5; | |
359 | } | |
360 | ||
361 | /* What does an and instruction cost - we do this b/c immediates may | |
362 | have been relaxed. We want to ensure that cse will cse relaxed immeds | |
4816b8e4 NC |
363 | out. Otherwise we'll get bad code (multiple reloads of the same const). */ |
364 | ||
8f90be4c NC |
365 | int |
366 | mcore_and_cost (x) | |
367 | rtx x; | |
368 | { | |
369 | int val; | |
370 | ||
371 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
372 | return 2; | |
373 | ||
374 | val = INTVAL (XEXP (x, 1)); | |
375 | ||
4816b8e4 | 376 | /* Do it directly. */ |
8f90be4c NC |
377 | if (CONST_OK_FOR_K (val) || CONST_OK_FOR_M (~val)) |
378 | return 2; | |
379 | /* Takes one instruction to load. */ | |
380 | else if (const_ok_for_mcore (val)) | |
381 | return 3; | |
382 | /* Takes two instructions to load. */ | |
383 | else if (TARGET_HARDLIT && mcore_const_ok_for_inline (val)) | |
384 | return 4; | |
385 | ||
4816b8e4 | 386 | /* Takes a lrw to load. */ |
8f90be4c NC |
387 | return 5; |
388 | } | |
389 | ||
4816b8e4 NC |
390 | /* What does an or cost - see and_cost(). */ |
391 | ||
8f90be4c NC |
392 | int |
393 | mcore_ior_cost (x) | |
394 | rtx x; | |
395 | { | |
396 | int val; | |
397 | ||
398 | if (GET_CODE (XEXP (x, 1)) != CONST_INT) | |
399 | return 2; | |
400 | ||
401 | val = INTVAL (XEXP (x, 1)); | |
402 | ||
4816b8e4 | 403 | /* Do it directly with bclri. */ |
8f90be4c NC |
404 | if (CONST_OK_FOR_M (val)) |
405 | return 2; | |
4816b8e4 | 406 | /* Takes one instruction to load. */ |
8f90be4c NC |
407 | else if (const_ok_for_mcore (val)) |
408 | return 3; | |
4816b8e4 | 409 | /* Takes two instructions to load. */ |
8f90be4c NC |
410 | else if (TARGET_HARDLIT && mcore_const_ok_for_inline (val)) |
411 | return 4; | |
412 | ||
4816b8e4 | 413 | /* Takes a lrw to load. */ |
8f90be4c NC |
414 | return 5; |
415 | } | |
416 | ||
417 | /* Check to see if a comparison against a constant can be made more efficient | |
418 | by incrementing/decrementing the constant to get one that is more efficient | |
419 | to load. */ | |
4816b8e4 | 420 | |
8f90be4c NC |
421 | int |
422 | mcore_modify_comparison (code) | |
423 | enum rtx_code code; | |
424 | { | |
425 | rtx op1 = arch_compare_op1; | |
426 | ||
427 | if (GET_CODE (op1) == CONST_INT) | |
428 | { | |
429 | int val = INTVAL (op1); | |
430 | ||
431 | switch (code) | |
432 | { | |
433 | case LE: | |
434 | if (CONST_OK_FOR_J (val + 1)) | |
435 | { | |
436 | arch_compare_op1 = GEN_INT (val + 1); | |
437 | return 1; | |
438 | } | |
439 | break; | |
440 | ||
441 | default: | |
442 | break; | |
443 | } | |
444 | } | |
445 | ||
446 | return 0; | |
447 | } | |
448 | ||
449 | /* Prepare the operands for a comparison. */ | |
4816b8e4 | 450 | |
8f90be4c NC |
451 | rtx |
452 | mcore_gen_compare_reg (code) | |
453 | enum rtx_code code; | |
454 | { | |
455 | rtx op0 = arch_compare_op0; | |
456 | rtx op1 = arch_compare_op1; | |
457 | rtx cc_reg = gen_rtx (REG, CCmode, CC_REG); | |
458 | ||
459 | if (CONSTANT_P (op1) && GET_CODE (op1) != CONST_INT) | |
460 | op1 = force_reg (SImode, op1); | |
461 | ||
462 | /* cmpnei: 0-31 (K immediate) | |
4816b8e4 | 463 | cmplti: 1-32 (J immediate, 0 using btsti x,31). */ |
8f90be4c NC |
464 | switch (code) |
465 | { | |
4816b8e4 | 466 | case EQ: /* Use inverted condition, cmpne. */ |
8f90be4c NC |
467 | code = NE; |
468 | /* drop through */ | |
4816b8e4 NC |
469 | |
470 | case NE: /* Use normal condition, cmpne. */ | |
8f90be4c NC |
471 | if (GET_CODE (op1) == CONST_INT && ! CONST_OK_FOR_K (INTVAL (op1))) |
472 | op1 = force_reg (SImode, op1); | |
473 | break; | |
474 | ||
4816b8e4 | 475 | case LE: /* Use inverted condition, reversed cmplt. */ |
8f90be4c NC |
476 | code = GT; |
477 | /* drop through */ | |
4816b8e4 NC |
478 | |
479 | case GT: /* Use normal condition, reversed cmplt. */ | |
8f90be4c NC |
480 | if (GET_CODE (op1) == CONST_INT) |
481 | op1 = force_reg (SImode, op1); | |
482 | break; | |
483 | ||
4816b8e4 | 484 | case GE: /* Use inverted condition, cmplt. */ |
8f90be4c NC |
485 | code = LT; |
486 | /* drop through */ | |
4816b8e4 NC |
487 | |
488 | case LT: /* Use normal condition, cmplt. */ | |
8f90be4c NC |
489 | if (GET_CODE (op1) == CONST_INT && |
490 | /* covered by btsti x,31 */ | |
491 | INTVAL (op1) != 0 && | |
492 | ! CONST_OK_FOR_J (INTVAL (op1))) | |
493 | op1 = force_reg (SImode, op1); | |
494 | break; | |
495 | ||
4816b8e4 | 496 | case GTU: /* Use inverted condition, cmple. */ |
8f90be4c NC |
497 | if (GET_CODE (op1) == CONST_INT && INTVAL (op1) == 0) |
498 | { | |
499 | /* Unsigned > 0 is the same as != 0, but we need | |
500 | to invert the condition, so we want to set | |
501 | code = EQ. This cannot be done however, as the | |
502 | mcore does not support such a test. Instead we | |
503 | cope with this case in the "bgtu" pattern itself | |
504 | so we should never reach this point. */ | |
505 | /* code = EQ; */ | |
506 | abort (); | |
507 | break; | |
508 | } | |
509 | code = LEU; | |
510 | /* drop through */ | |
4816b8e4 NC |
511 | |
512 | case LEU: /* Use normal condition, reversed cmphs. */ | |
8f90be4c NC |
513 | if (GET_CODE (op1) == CONST_INT && INTVAL (op1) != 0) |
514 | op1 = force_reg (SImode, op1); | |
515 | break; | |
516 | ||
4816b8e4 | 517 | case LTU: /* Use inverted condition, cmphs. */ |
8f90be4c NC |
518 | code = GEU; |
519 | /* drop through */ | |
4816b8e4 NC |
520 | |
521 | case GEU: /* Use normal condition, cmphs. */ | |
8f90be4c NC |
522 | if (GET_CODE (op1) == CONST_INT && INTVAL (op1) != 0) |
523 | op1 = force_reg (SImode, op1); | |
524 | break; | |
525 | ||
526 | default: | |
527 | break; | |
528 | } | |
529 | ||
530 | emit_insn (gen_rtx (SET, VOIDmode, cc_reg, gen_rtx (code, CCmode, op0, op1))); | |
531 | ||
532 | return cc_reg; | |
533 | } | |
534 | ||
535 | ||
536 | int | |
537 | mcore_symbolic_address_p (x) | |
538 | rtx x; | |
539 | { | |
540 | switch (GET_CODE (x)) | |
541 | { | |
542 | case SYMBOL_REF: | |
543 | case LABEL_REF: | |
544 | return 1; | |
545 | case CONST: | |
546 | x = XEXP (x, 0); | |
547 | return ( (GET_CODE (XEXP (x, 0)) == SYMBOL_REF | |
548 | || GET_CODE (XEXP (x, 0)) == LABEL_REF) | |
549 | && GET_CODE (XEXP (x, 1)) == CONST_INT); | |
550 | default: | |
551 | return 0; | |
552 | } | |
553 | } | |
554 | ||
555 | int | |
556 | mcore_call_address_operand (x, mode) | |
557 | rtx x; | |
558 | enum machine_mode mode; | |
559 | { | |
560 | return register_operand (x, mode) || CONSTANT_P (x); | |
561 | } | |
562 | ||
563 | /* Functions to output assembly code for a function call. */ | |
f27cd94d | 564 | |
8f90be4c NC |
565 | char * |
566 | mcore_output_call (operands, index) | |
567 | rtx operands[]; | |
568 | int index; | |
569 | { | |
570 | static char buffer[20]; | |
571 | rtx addr = operands [index]; | |
572 | ||
573 | if (REG_P (addr)) | |
574 | { | |
575 | if (TARGET_CG_DATA) | |
576 | { | |
577 | if (mcore_current_function_name == 0) | |
578 | abort (); | |
579 | ||
580 | ASM_OUTPUT_CG_EDGE (asm_out_file, mcore_current_function_name, | |
581 | "unknown", 1); | |
582 | } | |
583 | ||
584 | sprintf (buffer, "jsr\t%%%d", index); | |
585 | } | |
586 | else | |
587 | { | |
588 | if (TARGET_CG_DATA) | |
589 | { | |
590 | if (mcore_current_function_name == 0) | |
591 | abort (); | |
592 | ||
593 | if (GET_CODE (addr) != SYMBOL_REF) | |
594 | abort (); | |
595 | ||
596 | ASM_OUTPUT_CG_EDGE (asm_out_file, mcore_current_function_name, XSTR (addr, 0), 0); | |
597 | } | |
598 | ||
599 | sprintf (buffer, "jbsr\t%%%d", index); | |
600 | } | |
601 | ||
602 | return buffer; | |
603 | } | |
604 | ||
605 | /* Can we load a constant with a single instruction ? */ | |
4816b8e4 | 606 | |
8f90be4c NC |
607 | static int |
608 | const_ok_for_mcore (value) | |
609 | int value; | |
610 | { | |
611 | if (value >= 0 && value <= 127) | |
612 | return 1; | |
613 | ||
614 | /* Try exact power of two. */ | |
615 | if ((value & (value - 1)) == 0) | |
616 | return 1; | |
617 | ||
618 | /* Try exact power of two - 1. */ | |
619 | if ((value & (value + 1)) == 0) | |
620 | return 1; | |
621 | ||
622 | return 0; | |
623 | } | |
624 | ||
625 | /* Can we load a constant inline with up to 2 instructions ? */ | |
4816b8e4 | 626 | |
8f90be4c NC |
627 | int |
628 | mcore_const_ok_for_inline (value) | |
629 | long value; | |
630 | { | |
631 | int x, y; | |
632 | ||
633 | return try_constant_tricks (value, & x, & y) > 0; | |
634 | } | |
635 | ||
636 | /* Are we loading the constant using a not ? */ | |
4816b8e4 | 637 | |
8f90be4c NC |
638 | int |
639 | mcore_const_trick_uses_not (value) | |
640 | long value; | |
641 | { | |
642 | int x, y; | |
643 | ||
644 | return try_constant_tricks (value, & x, & y) == 2; | |
645 | } | |
646 | ||
647 | /* Try tricks to load a constant inline and return the trick number if | |
648 | success (0 is non-inlinable). | |
4816b8e4 NC |
649 | |
650 | 0: not inlinable | |
651 | 1: single instruction (do the usual thing) | |
652 | 2: single insn followed by a 'not' | |
653 | 3: single insn followed by a subi | |
654 | 4: single insn followed by an addi | |
655 | 5: single insn followed by rsubi | |
656 | 6: single insn followed by bseti | |
657 | 7: single insn followed by bclri | |
658 | 8: single insn followed by rotli | |
659 | 9: single insn followed by lsli | |
660 | 10: single insn followed by ixh | |
661 | 11: single insn followed by ixw. */ | |
8f90be4c NC |
662 | |
663 | static int | |
664 | try_constant_tricks (value, x, y) | |
665 | long value; | |
666 | int * x; | |
667 | int * y; | |
668 | { | |
669 | int i; | |
670 | unsigned bit, shf, rot; | |
671 | ||
672 | if (const_ok_for_mcore (value)) | |
4816b8e4 | 673 | return 1; /* Do the usual thing. */ |
8f90be4c NC |
674 | |
675 | if (TARGET_HARDLIT) | |
676 | { | |
677 | if (const_ok_for_mcore (~value)) | |
678 | { | |
679 | *x = ~value; | |
680 | return 2; | |
681 | } | |
682 | ||
683 | for (i = 1; i <= 32; i++) | |
684 | { | |
685 | if (const_ok_for_mcore (value - i)) | |
686 | { | |
687 | *x = value - i; | |
688 | *y = i; | |
689 | ||
690 | return 3; | |
691 | } | |
692 | ||
693 | if (const_ok_for_mcore (value + i)) | |
694 | { | |
695 | *x = value + i; | |
696 | *y = i; | |
697 | ||
698 | return 4; | |
699 | } | |
700 | } | |
701 | ||
702 | bit = 0x80000000UL; | |
703 | ||
704 | for (i = 0; i <= 31; i++) | |
705 | { | |
706 | if (const_ok_for_mcore (i - value)) | |
707 | { | |
708 | *x = i - value; | |
709 | *y = i; | |
710 | ||
711 | return 5; | |
712 | } | |
713 | ||
714 | if (const_ok_for_mcore (value & ~bit)) | |
715 | { | |
716 | *y = bit; | |
717 | *x = value & ~bit; | |
718 | ||
719 | return 6; | |
720 | } | |
721 | ||
722 | if (const_ok_for_mcore (value | bit)) | |
723 | { | |
724 | *y = ~bit; | |
725 | *x = value | bit; | |
726 | ||
727 | return 7; | |
728 | } | |
729 | ||
730 | bit >>= 1; | |
731 | } | |
732 | ||
733 | shf = value; | |
734 | rot = value; | |
735 | ||
736 | for (i = 1; i < 31; i++) | |
737 | { | |
738 | int c; | |
739 | ||
740 | /* MCore has rotate left. */ | |
741 | c = rot << 31; | |
742 | rot >>= 1; | |
743 | rot &= 0x7FFFFFFF; | |
744 | rot |= c; /* Simulate rotate. */ | |
745 | ||
746 | if (const_ok_for_mcore (rot)) | |
747 | { | |
748 | *y = i; | |
749 | *x = rot; | |
750 | ||
751 | return 8; | |
752 | } | |
753 | ||
754 | if (shf & 1) | |
4816b8e4 | 755 | shf = 0; /* Can't use logical shift, low order bit is one. */ |
8f90be4c NC |
756 | |
757 | shf >>= 1; | |
758 | ||
759 | if (shf != 0 && const_ok_for_mcore (shf)) | |
760 | { | |
761 | *y = i; | |
762 | *x = shf; | |
763 | ||
764 | return 9; | |
765 | } | |
766 | } | |
767 | ||
768 | if ((value % 3) == 0 && const_ok_for_mcore (value / 3)) | |
769 | { | |
770 | *x = value / 3; | |
771 | ||
772 | return 10; | |
773 | } | |
774 | ||
775 | if ((value % 5) == 0 && const_ok_for_mcore (value / 5)) | |
776 | { | |
777 | *x = value / 5; | |
778 | ||
779 | return 11; | |
780 | } | |
781 | } | |
782 | ||
783 | return 0; | |
784 | } | |
785 | ||
786 | ||
787 | /* Check whether reg is dead at first. This is done by searching ahead | |
788 | for either the next use (i.e., reg is live), a death note, or a set of | |
789 | reg. Don't just use dead_or_set_p() since reload does not always mark | |
790 | deaths (especially if PRESERVE_DEATH_NOTES_REGNO_P is not defined). We | |
4816b8e4 NC |
791 | can ignore subregs by extracting the actual register. BRC */ |
792 | ||
8f90be4c NC |
793 | int |
794 | mcore_is_dead (first, reg) | |
795 | rtx first; | |
796 | rtx reg; | |
797 | { | |
798 | rtx insn; | |
799 | ||
800 | /* For mcore, subregs can't live independently of their parent regs. */ | |
801 | if (GET_CODE (reg) == SUBREG) | |
802 | reg = SUBREG_REG (reg); | |
803 | ||
804 | /* Dies immediately. */ | |
805 | if (dead_or_set_p (first, reg)) | |
806 | return 1; | |
807 | ||
808 | /* Look for conclusive evidence of live/death, otherwise we have | |
809 | to assume that it is live. */ | |
810 | for (insn = NEXT_INSN (first); insn; insn = NEXT_INSN (insn)) | |
811 | { | |
812 | if (GET_CODE (insn) == JUMP_INSN) | |
813 | return 0; /* We lose track, assume it is alive. */ | |
814 | ||
815 | else if (GET_CODE(insn) == CALL_INSN) | |
816 | { | |
817 | /* Call's might use it for target or register parms. */ | |
818 | if (reg_referenced_p (reg, PATTERN (insn)) | |
819 | || find_reg_fusage (insn, USE, reg)) | |
820 | return 0; | |
821 | else if (dead_or_set_p (insn, reg)) | |
822 | return 1; | |
823 | } | |
824 | else if (GET_CODE (insn) == INSN) | |
825 | { | |
826 | if (reg_referenced_p (reg, PATTERN (insn))) | |
827 | return 0; | |
828 | else if (dead_or_set_p (insn, reg)) | |
829 | return 1; | |
830 | } | |
831 | } | |
832 | ||
833 | /* No conclusive evidence either way, we can not take the chance | |
834 | that control flow hid the use from us -- "I'm not dead yet". */ | |
835 | return 0; | |
836 | } | |
837 | ||
838 | ||
839 | /* Count the number of ones in mask. */ | |
4816b8e4 | 840 | |
8f90be4c NC |
841 | int |
842 | mcore_num_ones (mask) | |
843 | int mask; | |
844 | { | |
4816b8e4 | 845 | /* A trick to count set bits recently posted on comp.compilers. */ |
8f90be4c NC |
846 | mask = (mask >> 1 & 0x55555555) + (mask & 0x55555555); |
847 | mask = ((mask >> 2) & 0x33333333) + (mask & 0x33333333); | |
848 | mask = ((mask >> 4) + mask) & 0x0f0f0f0f; | |
849 | mask = ((mask >> 8) + mask); | |
850 | ||
851 | return (mask + (mask >> 16)) & 0xff; | |
852 | } | |
853 | ||
4816b8e4 NC |
854 | /* Count the number of zeros in mask. */ |
855 | ||
8f90be4c NC |
856 | int |
857 | mcore_num_zeros (mask) | |
858 | int mask; | |
859 | { | |
860 | return 32 - mcore_num_ones (mask); | |
861 | } | |
862 | ||
863 | /* Determine byte being masked. */ | |
4816b8e4 | 864 | |
8f90be4c NC |
865 | int |
866 | mcore_byte_offset (mask) | |
867 | unsigned int mask; | |
868 | { | |
869 | if (mask == 0x00ffffffUL) | |
870 | return 0; | |
871 | else if (mask == 0xff00ffffUL) | |
872 | return 1; | |
873 | else if (mask == 0xffff00ffUL) | |
874 | return 2; | |
875 | else if (mask == 0xffffff00UL) | |
876 | return 3; | |
877 | ||
878 | return -1; | |
879 | } | |
880 | ||
881 | /* Determine halfword being masked. */ | |
4816b8e4 | 882 | |
8f90be4c NC |
883 | int |
884 | mcore_halfword_offset (mask) | |
885 | unsigned int mask; | |
886 | { | |
887 | if (mask == 0x0000ffffL) | |
888 | return 0; | |
889 | else if (mask == 0xffff0000UL) | |
890 | return 1; | |
891 | ||
892 | return -1; | |
893 | } | |
894 | ||
895 | /* Output a series of bseti's corresponding to mask. */ | |
4816b8e4 | 896 | |
f27cd94d | 897 | const char * |
8f90be4c NC |
898 | mcore_output_bseti (dst, mask) |
899 | rtx dst; | |
900 | int mask; | |
901 | { | |
902 | rtx out_operands[2]; | |
903 | int bit; | |
904 | ||
905 | out_operands[0] = dst; | |
906 | ||
907 | for (bit = 0; bit < 32; bit++) | |
908 | { | |
909 | if ((mask & 0x1) == 0x1) | |
910 | { | |
911 | out_operands[1] = GEN_INT (bit); | |
912 | ||
913 | output_asm_insn ("bseti\t%0,%1", out_operands); | |
914 | } | |
915 | mask >>= 1; | |
916 | } | |
917 | ||
918 | return ""; | |
919 | } | |
920 | ||
921 | /* Output a series of bclri's corresponding to mask. */ | |
4816b8e4 | 922 | |
f27cd94d | 923 | const char * |
8f90be4c NC |
924 | mcore_output_bclri (dst, mask) |
925 | rtx dst; | |
926 | int mask; | |
927 | { | |
928 | rtx out_operands[2]; | |
929 | int bit; | |
930 | ||
931 | out_operands[0] = dst; | |
932 | ||
933 | for (bit = 0; bit < 32; bit++) | |
934 | { | |
935 | if ((mask & 0x1) == 0x0) | |
936 | { | |
937 | out_operands[1] = GEN_INT (bit); | |
938 | ||
939 | output_asm_insn ("bclri\t%0,%1", out_operands); | |
940 | } | |
941 | ||
942 | mask >>= 1; | |
943 | } | |
944 | ||
945 | return ""; | |
946 | } | |
947 | ||
948 | /* Output a conditional move of two constants that are +/- 1 within each | |
949 | other. See the "movtK" patterns in mcore.md. I'm not sure this is | |
950 | really worth the effort. */ | |
4816b8e4 | 951 | |
f27cd94d | 952 | const char * |
8f90be4c NC |
953 | mcore_output_cmov (operands, cmp_t, test) |
954 | rtx operands[]; | |
955 | int cmp_t; | |
956 | char * test; | |
957 | { | |
958 | int load_value; | |
959 | int adjust_value; | |
960 | rtx out_operands[4]; | |
961 | ||
962 | out_operands[0] = operands[0]; | |
963 | ||
4816b8e4 | 964 | /* Check to see which constant is loadable. */ |
8f90be4c NC |
965 | if (const_ok_for_mcore (INTVAL (operands[1]))) |
966 | { | |
967 | out_operands[1] = operands[1]; | |
968 | out_operands[2] = operands[2]; | |
969 | } | |
970 | else if (const_ok_for_mcore (INTVAL (operands[2]))) | |
971 | { | |
972 | out_operands[1] = operands[2]; | |
973 | out_operands[2] = operands[1]; | |
974 | ||
4816b8e4 | 975 | /* Complement test since constants are swapped. */ |
8f90be4c NC |
976 | cmp_t = (cmp_t == 0); |
977 | } | |
978 | load_value = INTVAL (out_operands[1]); | |
979 | adjust_value = INTVAL (out_operands[2]); | |
980 | ||
4816b8e4 | 981 | /* First output the test if folded into the pattern. */ |
8f90be4c NC |
982 | |
983 | if (test) | |
984 | output_asm_insn (test, operands); | |
985 | ||
4816b8e4 | 986 | /* Load the constant - for now, only support constants that can be |
8f90be4c NC |
987 | generated with a single instruction. maybe add general inlinable |
988 | constants later (this will increase the # of patterns since the | |
4816b8e4 | 989 | instruction sequence has a different length attribute). */ |
8f90be4c NC |
990 | if (load_value >= 0 && load_value <= 127) |
991 | output_asm_insn ("movi\t%0,%1", out_operands); | |
992 | else if ((load_value & (load_value - 1)) == 0) | |
993 | output_asm_insn ("bgeni\t%0,%P1", out_operands); | |
994 | else if ((load_value & (load_value + 1)) == 0) | |
995 | output_asm_insn ("bmaski\t%0,%N1", out_operands); | |
996 | ||
4816b8e4 | 997 | /* Output the constant adjustment. */ |
8f90be4c NC |
998 | if (load_value > adjust_value) |
999 | { | |
1000 | if (cmp_t) | |
1001 | output_asm_insn ("decf\t%0", out_operands); | |
1002 | else | |
1003 | output_asm_insn ("dect\t%0", out_operands); | |
1004 | } | |
1005 | else | |
1006 | { | |
1007 | if (cmp_t) | |
1008 | output_asm_insn ("incf\t%0", out_operands); | |
1009 | else | |
1010 | output_asm_insn ("inct\t%0", out_operands); | |
1011 | } | |
1012 | ||
1013 | return ""; | |
1014 | } | |
1015 | ||
1016 | /* Outputs the peephole for moving a constant that gets not'ed followed | |
4816b8e4 NC |
1017 | by an and (i.e. combine the not and the and into andn). BRC */ |
1018 | ||
f27cd94d | 1019 | const char * |
8f90be4c NC |
1020 | mcore_output_andn (insn, operands) |
1021 | rtx insn ATTRIBUTE_UNUSED; | |
1022 | rtx operands[]; | |
1023 | { | |
1024 | int x, y; | |
1025 | rtx out_operands[3]; | |
f27cd94d | 1026 | const char * load_op; |
8f90be4c NC |
1027 | char buf[256]; |
1028 | ||
1029 | if (try_constant_tricks (INTVAL (operands[1]), &x, &y) != 2) | |
1030 | abort (); | |
1031 | ||
1032 | out_operands[0] = operands[0]; | |
1033 | out_operands[1] = GEN_INT(x); | |
1034 | out_operands[2] = operands[2]; | |
1035 | ||
1036 | if (x >= 0 && x <= 127) | |
1037 | load_op = "movi\t%0,%1"; | |
4816b8e4 NC |
1038 | |
1039 | /* Try exact power of two. */ | |
8f90be4c NC |
1040 | else if ((x & (x - 1)) == 0) |
1041 | load_op = "bgeni\t%0,%P1"; | |
4816b8e4 NC |
1042 | |
1043 | /* Try exact power of two - 1. */ | |
8f90be4c NC |
1044 | else if ((x & (x + 1)) == 0) |
1045 | load_op = "bmaski\t%0,%N1"; | |
4816b8e4 | 1046 | |
8f90be4c NC |
1047 | else |
1048 | load_op = "BADMOVI\t%0,%1"; | |
1049 | ||
1050 | sprintf (buf, "%s\n\tandn\t%%2,%%0", load_op); | |
1051 | output_asm_insn (buf, out_operands); | |
1052 | ||
1053 | return ""; | |
1054 | } | |
1055 | ||
1056 | /* Output an inline constant. */ | |
4816b8e4 | 1057 | |
f27cd94d | 1058 | static const char * |
8f90be4c NC |
1059 | output_inline_const (mode, operands) |
1060 | enum machine_mode mode; | |
1061 | rtx operands[]; | |
1062 | { | |
1063 | int x = 0, y = 0; | |
1064 | int trick_no; | |
1065 | rtx out_operands[3]; | |
1066 | char buf[256]; | |
1067 | char load_op[256]; | |
f27cd94d | 1068 | const char *dst_fmt; |
8f90be4c NC |
1069 | int value; |
1070 | ||
1071 | value = INTVAL (operands[1]); | |
1072 | ||
1073 | if ((trick_no = try_constant_tricks (value, &x, &y)) == 0) | |
1074 | { | |
1075 | /* lrw's are handled separately: Large inlinable constants | |
1076 | never get turned into lrw's. Our caller uses try_constant_tricks | |
1077 | to back off to an lrw rather than calling this routine. */ | |
1078 | abort (); | |
1079 | } | |
1080 | ||
1081 | if (trick_no == 1) | |
1082 | x = value; | |
1083 | ||
4816b8e4 | 1084 | /* operands: 0 = dst, 1 = load immed., 2 = immed. adjustment. */ |
8f90be4c NC |
1085 | out_operands[0] = operands[0]; |
1086 | out_operands[1] = GEN_INT (x); | |
1087 | ||
1088 | if (trick_no > 2) | |
1089 | out_operands[2] = GEN_INT (y); | |
1090 | ||
4816b8e4 | 1091 | /* Select dst format based on mode. */ |
8f90be4c NC |
1092 | if (mode == DImode && (! TARGET_LITTLE_END)) |
1093 | dst_fmt = "%R0"; | |
1094 | else | |
1095 | dst_fmt = "%0"; | |
1096 | ||
1097 | if (x >= 0 && x <= 127) | |
1098 | sprintf (load_op, "movi\t%s,%%1", dst_fmt); | |
4816b8e4 | 1099 | |
8f90be4c NC |
1100 | /* Try exact power of two. */ |
1101 | else if ((x & (x - 1)) == 0) | |
1102 | sprintf (load_op, "bgeni\t%s,%%P1", dst_fmt); | |
4816b8e4 NC |
1103 | |
1104 | /* Try exact power of two - 1. */ | |
8f90be4c NC |
1105 | else if ((x & (x + 1)) == 0) |
1106 | sprintf (load_op, "bmaski\t%s,%%N1", dst_fmt); | |
4816b8e4 | 1107 | |
8f90be4c NC |
1108 | else |
1109 | sprintf (load_op, "BADMOVI\t%s,%%1", dst_fmt); | |
1110 | ||
1111 | switch (trick_no) | |
1112 | { | |
1113 | case 1: | |
1114 | strcpy (buf, load_op); | |
1115 | break; | |
1116 | case 2: /* not */ | |
1117 | sprintf (buf, "%s\n\tnot\t%s\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1118 | break; | |
1119 | case 3: /* add */ | |
1120 | sprintf (buf, "%s\n\taddi\t%s,%%2\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1121 | break; | |
1122 | case 4: /* sub */ | |
1123 | sprintf (buf, "%s\n\tsubi\t%s,%%2\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1124 | break; | |
1125 | case 5: /* rsub */ | |
4816b8e4 | 1126 | /* Never happens unless -mrsubi, see try_constant_tricks(). */ |
8f90be4c NC |
1127 | sprintf (buf, "%s\n\trsubi\t%s,%%2\t// %d 0x%x", load_op, dst_fmt, value, value); |
1128 | break; | |
1129 | case 6: /* bset */ | |
1130 | sprintf (buf, "%s\n\tbseti\t%s,%%P2\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1131 | break; | |
1132 | case 7: /* bclr */ | |
1133 | sprintf (buf, "%s\n\tbclri\t%s,%%Q2\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1134 | break; | |
1135 | case 8: /* rotl */ | |
1136 | sprintf (buf, "%s\n\trotli\t%s,%%2\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1137 | break; | |
1138 | case 9: /* lsl */ | |
1139 | sprintf (buf, "%s\n\tlsli\t%s,%%2\t// %d 0x%x", load_op, dst_fmt, value, value); | |
1140 | break; | |
1141 | case 10: /* ixh */ | |
1142 | sprintf (buf, "%s\n\tixh\t%s,%s\t// %d 0x%x", load_op, dst_fmt, dst_fmt, value, value); | |
1143 | break; | |
1144 | case 11: /* ixw */ | |
1145 | sprintf (buf, "%s\n\tixw\t%s,%s\t// %d 0x%x", load_op, dst_fmt, dst_fmt, value, value); | |
1146 | break; | |
1147 | default: | |
1148 | return ""; | |
1149 | } | |
1150 | ||
1151 | output_asm_insn (buf, out_operands); | |
1152 | ||
1153 | return ""; | |
1154 | } | |
1155 | ||
1156 | /* Output a move of a word or less value. */ | |
4816b8e4 | 1157 | |
f27cd94d | 1158 | const char * |
8f90be4c NC |
1159 | mcore_output_move (insn, operands, mode) |
1160 | rtx insn ATTRIBUTE_UNUSED; | |
1161 | rtx operands[]; | |
1162 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1163 | { | |
1164 | rtx dst = operands[0]; | |
1165 | rtx src = operands[1]; | |
1166 | ||
1167 | if (GET_CODE (dst) == REG) | |
1168 | { | |
1169 | if (GET_CODE (src) == REG) | |
1170 | { | |
1171 | if (REGNO (src) == CC_REG) /* r-c */ | |
1172 | return "mvc\t%0"; | |
1173 | else | |
1174 | return "mov\t%0,%1"; /* r-r*/ | |
1175 | } | |
1176 | else if (GET_CODE (src) == MEM) | |
1177 | { | |
1178 | if (GET_CODE (XEXP (src, 0)) == LABEL_REF) | |
1179 | return "lrw\t%0,[%1]"; /* a-R */ | |
1180 | else | |
1181 | return "ldw\t%0,%1"; /* r-m */ | |
1182 | } | |
1183 | else if (GET_CODE (src) == CONST_INT) | |
1184 | { | |
1185 | int x, y; | |
1186 | ||
1187 | if (CONST_OK_FOR_I (INTVAL (src))) /* r-I */ | |
1188 | return "movi\t%0,%1"; | |
1189 | else if (CONST_OK_FOR_M (INTVAL (src))) /* r-M */ | |
1190 | return "bgeni\t%0,%P1\t// %1 %x1"; | |
1191 | else if (CONST_OK_FOR_N (INTVAL (src))) /* r-N */ | |
1192 | return "bmaski\t%0,%N1\t// %1 %x1"; | |
1193 | else if (try_constant_tricks (INTVAL (src), &x, &y)) /* R-P */ | |
1194 | return output_inline_const (SImode, operands); /* 1-2 insns */ | |
1195 | else | |
4816b8e4 | 1196 | return "lrw\t%0,%x1\t// %1"; /* Get it from literal pool. */ |
8f90be4c NC |
1197 | } |
1198 | else | |
4816b8e4 | 1199 | return "lrw\t%0, %1"; /* Into the literal pool. */ |
8f90be4c NC |
1200 | } |
1201 | else if (GET_CODE (dst) == MEM) /* m-r */ | |
1202 | return "stw\t%1,%0"; | |
1203 | ||
1204 | abort (); | |
1205 | } | |
1206 | ||
1207 | /* Outputs a constant inline -- regardless of the cost. | |
1208 | Useful for things where we've gotten into trouble and think we'd | |
1209 | be doing an lrw into r15 (forbidden). This lets us get out of | |
1210 | that pickle even after register allocation. */ | |
4816b8e4 | 1211 | |
f27cd94d | 1212 | const char * |
8f90be4c NC |
1213 | mcore_output_inline_const_forced (insn, operands, mode) |
1214 | rtx insn ATTRIBUTE_UNUSED; | |
1215 | rtx operands[]; | |
1216 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1217 | { | |
1218 | unsigned long value = INTVAL (operands[1]); | |
1219 | unsigned long ovalue = value; | |
1220 | struct piece | |
1221 | { | |
1222 | int low; | |
1223 | int shift; | |
1224 | } | |
1225 | part[6]; | |
1226 | int i; | |
1227 | ||
1228 | if (mcore_const_ok_for_inline (value)) | |
1229 | return output_inline_const (SImode, operands); | |
1230 | ||
b6a1cbae | 1231 | for (i = 0; (unsigned) i < ARRAY_SIZE (part); i++) |
8f90be4c NC |
1232 | { |
1233 | part[i].shift = 0; | |
1234 | part[i].low = (value & 0x1F); | |
1235 | value -= part[i].low; | |
1236 | ||
1237 | if (mcore_const_ok_for_inline (value)) | |
1238 | break; | |
1239 | else | |
1240 | { | |
1241 | value >>= 5; | |
1242 | part[i].shift = 5; | |
1243 | ||
1244 | while ((value & 1) == 0) | |
1245 | { | |
1246 | part[i].shift++; | |
1247 | value >>= 1; | |
1248 | } | |
1249 | ||
1250 | if (mcore_const_ok_for_inline (value)) | |
1251 | break; | |
1252 | } | |
1253 | } | |
1254 | ||
1255 | /* 5 bits per iteration, a maximum of 5 times == 25 bits and leaves | |
1256 | 7 bits left in the constant -- which we know we can cover with | |
1257 | a movi. The final value can't be zero otherwise we'd have stopped | |
1258 | in the previous iteration. */ | |
1259 | if (value == 0 || ! mcore_const_ok_for_inline (value)) | |
1260 | abort (); | |
1261 | ||
4816b8e4 | 1262 | /* Now, work our way backwards emitting the constant. */ |
8f90be4c NC |
1263 | |
1264 | /* Emit the value that remains -- it will be non-zero. */ | |
1265 | operands[1] = GEN_INT (value); | |
1266 | output_asm_insn (output_inline_const (SImode, operands), operands); | |
1267 | ||
1268 | while (i >= 0) | |
1269 | { | |
1270 | /* Shift anything we've already loaded. */ | |
1271 | if (part[i].shift) | |
1272 | { | |
1273 | operands[2] = GEN_INT (part[i].shift); | |
1274 | output_asm_insn ("lsli %0,%2", operands); | |
1275 | value <<= part[i].shift; | |
1276 | } | |
1277 | ||
1278 | /* Add anything we need into the low 5 bits. */ | |
1279 | if (part[i].low != 0) | |
1280 | { | |
1281 | operands[2] = GEN_INT (part[i].low); | |
1282 | output_asm_insn ("addi %0,%2", operands); | |
1283 | value += part[i].low; | |
1284 | } | |
1285 | ||
1286 | i--; | |
1287 | } | |
1288 | ||
1289 | if (value != ovalue) /* sanity */ | |
1290 | abort (); | |
1291 | ||
4816b8e4 | 1292 | /* We've output all the instructions. */ |
8f90be4c NC |
1293 | return ""; |
1294 | } | |
1295 | ||
1296 | /* Return a sequence of instructions to perform DI or DF move. | |
1297 | Since the MCORE cannot move a DI or DF in one instruction, we have | |
1298 | to take care when we see overlapping source and dest registers. */ | |
4816b8e4 | 1299 | |
f27cd94d | 1300 | const char * |
8f90be4c NC |
1301 | mcore_output_movedouble (operands, mode) |
1302 | rtx operands[]; | |
1303 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1304 | { | |
1305 | rtx dst = operands[0]; | |
1306 | rtx src = operands[1]; | |
1307 | ||
1308 | if (GET_CODE (dst) == REG) | |
1309 | { | |
1310 | if (GET_CODE (src) == REG) | |
1311 | { | |
1312 | int dstreg = REGNO (dst); | |
1313 | int srcreg = REGNO (src); | |
4816b8e4 | 1314 | |
8f90be4c NC |
1315 | /* Ensure the second source not overwritten. */ |
1316 | if (srcreg + 1 == dstreg) | |
1317 | return "mov %R0,%R1\n\tmov %0,%1"; | |
1318 | else | |
1319 | return "mov %0,%1\n\tmov %R0,%R1"; | |
1320 | } | |
1321 | else if (GET_CODE (src) == MEM) | |
1322 | { | |
1323 | rtx memexp = memexp = XEXP (src, 0); | |
1324 | int dstreg = REGNO (dst); | |
1325 | int basereg = -1; | |
1326 | ||
1327 | if (GET_CODE (memexp) == LABEL_REF) | |
1328 | return "lrw\t%0,[%1]\n\tlrw\t%R0,[%R1]"; | |
1329 | else if (GET_CODE (memexp) == REG) | |
1330 | basereg = REGNO (memexp); | |
1331 | else if (GET_CODE (memexp) == PLUS) | |
1332 | { | |
1333 | if (GET_CODE (XEXP (memexp, 0)) == REG) | |
1334 | basereg = REGNO (XEXP (memexp, 0)); | |
1335 | else if (GET_CODE (XEXP (memexp, 1)) == REG) | |
1336 | basereg = REGNO (XEXP (memexp, 1)); | |
1337 | else | |
1338 | abort (); | |
1339 | } | |
1340 | else | |
1341 | abort (); | |
1342 | ||
4816b8e4 | 1343 | /* ??? length attribute is wrong here. */ |
8f90be4c NC |
1344 | if (dstreg == basereg) |
1345 | { | |
4816b8e4 | 1346 | /* Just load them in reverse order. */ |
8f90be4c | 1347 | return "ldw\t%R0,%R1\n\tldw\t%0,%1"; |
4816b8e4 | 1348 | |
8f90be4c | 1349 | /* XXX: alternative: move basereg to basereg+1 |
4816b8e4 | 1350 | and then fall through. */ |
8f90be4c NC |
1351 | } |
1352 | else | |
1353 | return "ldw\t%0,%1\n\tldw\t%R0,%R1"; | |
1354 | } | |
1355 | else if (GET_CODE (src) == CONST_INT) | |
1356 | { | |
1357 | if (TARGET_LITTLE_END) | |
1358 | { | |
1359 | if (CONST_OK_FOR_I (INTVAL (src))) | |
1360 | output_asm_insn ("movi %0,%1", operands); | |
1361 | else if (CONST_OK_FOR_M (INTVAL (src))) | |
1362 | output_asm_insn ("bgeni %0,%P1", operands); | |
1363 | else if (INTVAL (src) == -1) | |
1364 | output_asm_insn ("bmaski %0,32", operands); | |
1365 | else if (CONST_OK_FOR_N (INTVAL (src))) | |
1366 | output_asm_insn ("bmaski %0,%N1", operands); | |
1367 | else | |
1368 | abort (); | |
1369 | ||
1370 | if (INTVAL (src) < 0) | |
1371 | return "bmaski %R0,32"; | |
1372 | else | |
1373 | return "movi %R0,0"; | |
1374 | } | |
1375 | else | |
1376 | { | |
1377 | if (CONST_OK_FOR_I (INTVAL (src))) | |
1378 | output_asm_insn ("movi %R0,%1", operands); | |
1379 | else if (CONST_OK_FOR_M (INTVAL (src))) | |
1380 | output_asm_insn ("bgeni %R0,%P1", operands); | |
1381 | else if (INTVAL (src) == -1) | |
1382 | output_asm_insn ("bmaski %R0,32", operands); | |
1383 | else if (CONST_OK_FOR_N (INTVAL (src))) | |
1384 | output_asm_insn ("bmaski %R0,%N1", operands); | |
1385 | else | |
1386 | abort (); | |
1387 | ||
1388 | if (INTVAL (src) < 0) | |
1389 | return "bmaski %0,32"; | |
1390 | else | |
1391 | return "movi %0,0"; | |
1392 | } | |
1393 | } | |
1394 | else | |
1395 | abort (); | |
1396 | } | |
1397 | else if (GET_CODE (dst) == MEM && GET_CODE (src) == REG) | |
1398 | return "stw\t%1,%0\n\tstw\t%R1,%R0"; | |
1399 | else | |
1400 | abort (); | |
1401 | } | |
1402 | ||
1403 | /* Predicates used by the templates. */ | |
1404 | ||
1405 | /* Non zero if OP can be source of a simple move operation. */ | |
4816b8e4 | 1406 | |
8f90be4c NC |
1407 | int |
1408 | mcore_general_movsrc_operand (op, mode) | |
1409 | rtx op; | |
1410 | enum machine_mode mode; | |
1411 | { | |
1412 | /* Any (MEM LABEL_REF) is OK. That is a pc-relative load. */ | |
1413 | if (GET_CODE (op) == MEM && GET_CODE (XEXP (op, 0)) == LABEL_REF) | |
1414 | return 1; | |
1415 | ||
1416 | return general_operand (op, mode); | |
1417 | } | |
1418 | ||
1419 | /* Non zero if OP can be destination of a simple move operation. */ | |
4816b8e4 | 1420 | |
8f90be4c NC |
1421 | int |
1422 | mcore_general_movdst_operand (op, mode) | |
1423 | rtx op; | |
1424 | enum machine_mode mode; | |
1425 | { | |
1426 | if (GET_CODE (op) == REG && REGNO (op) == CC_REG) | |
1427 | return 0; | |
1428 | ||
1429 | return general_operand (op, mode); | |
1430 | } | |
1431 | ||
1432 | /* Nonzero if OP is a normal arithmetic register. */ | |
4816b8e4 | 1433 | |
8f90be4c NC |
1434 | int |
1435 | mcore_arith_reg_operand (op, mode) | |
1436 | rtx op; | |
1437 | enum machine_mode mode; | |
1438 | { | |
1439 | if (! register_operand (op, mode)) | |
1440 | return 0; | |
1441 | ||
1442 | if (GET_CODE (op) == SUBREG) | |
1443 | op = SUBREG_REG (op); | |
1444 | ||
1445 | if (GET_CODE (op) == REG) | |
1446 | return REGNO (op) != CC_REG; | |
1447 | ||
1448 | return 1; | |
1449 | } | |
1450 | ||
1451 | /* Non zero if OP should be recognized during reload for an ixh/ixw | |
1452 | operand. See the ixh/ixw patterns. */ | |
4816b8e4 | 1453 | |
8f90be4c NC |
1454 | int |
1455 | mcore_reload_operand (op, mode) | |
1456 | rtx op; | |
1457 | enum machine_mode mode; | |
1458 | { | |
1459 | if (mcore_arith_reg_operand (op, mode)) | |
1460 | return 1; | |
1461 | ||
1462 | if (! reload_in_progress) | |
1463 | return 0; | |
1464 | ||
1465 | return GET_CODE (op) == MEM; | |
1466 | } | |
1467 | ||
1468 | /* Nonzero if OP is a valid source operand for an arithmetic insn. */ | |
4816b8e4 | 1469 | |
8f90be4c NC |
1470 | int |
1471 | mcore_arith_J_operand (op, mode) | |
1472 | rtx op; | |
1473 | enum machine_mode mode; | |
1474 | { | |
1475 | if (register_operand (op, mode)) | |
1476 | return 1; | |
1477 | ||
1478 | if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_J (INTVAL (op))) | |
1479 | return 1; | |
1480 | ||
1481 | return 0; | |
1482 | } | |
1483 | ||
1484 | /* Nonzero if OP is a valid source operand for an arithmetic insn. */ | |
4816b8e4 | 1485 | |
8f90be4c NC |
1486 | int |
1487 | mcore_arith_K_operand (op, mode) | |
1488 | rtx op; | |
1489 | enum machine_mode mode; | |
1490 | { | |
1491 | if (register_operand (op, mode)) | |
1492 | return 1; | |
1493 | ||
1494 | if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_K (INTVAL (op))) | |
1495 | return 1; | |
1496 | ||
1497 | return 0; | |
1498 | } | |
1499 | ||
1500 | /* Nonzero if OP is a valid source operand for a shift or rotate insn. */ | |
4816b8e4 | 1501 | |
8f90be4c NC |
1502 | int |
1503 | mcore_arith_K_operand_not_0 (op, mode) | |
1504 | rtx op; | |
1505 | enum machine_mode mode; | |
1506 | { | |
1507 | if (register_operand (op, mode)) | |
1508 | return 1; | |
1509 | ||
1510 | if ( GET_CODE (op) == CONST_INT | |
1511 | && CONST_OK_FOR_K (INTVAL (op)) | |
1512 | && INTVAL (op) != 0) | |
1513 | return 1; | |
1514 | ||
1515 | return 0; | |
1516 | } | |
1517 | ||
1518 | int | |
1519 | mcore_arith_K_S_operand (op, mode) | |
1520 | rtx op; | |
1521 | enum machine_mode mode; | |
1522 | { | |
1523 | if (register_operand (op, mode)) | |
1524 | return 1; | |
1525 | ||
1526 | if (GET_CODE (op) == CONST_INT) | |
1527 | { | |
1528 | if (CONST_OK_FOR_K (INTVAL (op)) || CONST_OK_FOR_M (~INTVAL (op))) | |
1529 | return 1; | |
1530 | } | |
1531 | ||
1532 | return 0; | |
1533 | } | |
1534 | ||
1535 | int | |
1536 | mcore_arith_S_operand (op) | |
1537 | rtx op; | |
1538 | { | |
1539 | if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_M (~INTVAL (op))) | |
1540 | return 1; | |
1541 | ||
1542 | return 0; | |
1543 | } | |
1544 | ||
1545 | int | |
1546 | mcore_arith_M_operand (op, mode) | |
1547 | rtx op; | |
1548 | enum machine_mode mode; | |
1549 | { | |
1550 | if (register_operand (op, mode)) | |
1551 | return 1; | |
1552 | ||
1553 | if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_M (INTVAL (op))) | |
1554 | return 1; | |
1555 | ||
1556 | return 0; | |
1557 | } | |
1558 | ||
4816b8e4 NC |
1559 | /* Nonzero if OP is a valid source operand for loading. */ |
1560 | ||
8f90be4c NC |
1561 | int |
1562 | mcore_arith_imm_operand (op, mode) | |
1563 | rtx op; | |
1564 | enum machine_mode mode; | |
1565 | { | |
1566 | if (register_operand (op, mode)) | |
1567 | return 1; | |
1568 | ||
1569 | if (GET_CODE (op) == CONST_INT && const_ok_for_mcore (INTVAL (op))) | |
1570 | return 1; | |
1571 | ||
1572 | return 0; | |
1573 | } | |
1574 | ||
1575 | int | |
1576 | mcore_arith_any_imm_operand (op, mode) | |
1577 | rtx op; | |
1578 | enum machine_mode mode; | |
1579 | { | |
1580 | if (register_operand (op, mode)) | |
1581 | return 1; | |
1582 | ||
1583 | if (GET_CODE (op) == CONST_INT) | |
1584 | return 1; | |
1585 | ||
1586 | return 0; | |
1587 | } | |
1588 | ||
4816b8e4 NC |
1589 | /* Nonzero if OP is a valid source operand for a cmov with two consts +/- 1. */ |
1590 | ||
8f90be4c NC |
1591 | int |
1592 | mcore_arith_O_operand (op, mode) | |
1593 | rtx op; | |
1594 | enum machine_mode mode; | |
1595 | { | |
1596 | if (register_operand (op, mode)) | |
1597 | return 1; | |
1598 | ||
1599 | if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_O (INTVAL (op))) | |
1600 | return 1; | |
1601 | ||
1602 | return 0; | |
1603 | } | |
1604 | ||
1605 | /* Nonzero if OP is a valid source operand for a btsti. */ | |
4816b8e4 | 1606 | |
8f90be4c NC |
1607 | int |
1608 | mcore_literal_K_operand (op, mode) | |
1609 | rtx op; | |
1610 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1611 | { | |
1612 | if (GET_CODE (op) == CONST_INT && CONST_OK_FOR_K (INTVAL (op))) | |
1613 | return 1; | |
1614 | ||
1615 | return 0; | |
1616 | } | |
1617 | ||
1618 | /* Nonzero if OP is a valid source operand for an add/sub insn. */ | |
4816b8e4 | 1619 | |
8f90be4c NC |
1620 | int |
1621 | mcore_addsub_operand (op, mode) | |
1622 | rtx op; | |
1623 | enum machine_mode mode; | |
1624 | { | |
1625 | if (register_operand (op, mode)) | |
1626 | return 1; | |
1627 | ||
1628 | if (GET_CODE (op) == CONST_INT) | |
1629 | { | |
1630 | return 1; | |
1631 | ||
1632 | /* The following is removed because it precludes large constants from being | |
1633 | returned as valid source operands for and add/sub insn. While large | |
1634 | constants may not directly be used in an add/sub, they may if first loaded | |
1635 | into a register. Thus, this predicate should indicate that they are valid, | |
1636 | and the constraint in mcore.md should control whether an additional load to | |
4816b8e4 | 1637 | register is needed. (see mcore.md, addsi). -- DAC 4/2/1998 */ |
8f90be4c NC |
1638 | /* |
1639 | if (CONST_OK_FOR_J(INTVAL(op)) || CONST_OK_FOR_L(INTVAL(op))) | |
1640 | return 1; | |
1641 | */ | |
1642 | } | |
1643 | ||
1644 | return 0; | |
1645 | } | |
1646 | ||
1647 | /* Nonzero if OP is a valid source operand for a compare operation. */ | |
4816b8e4 | 1648 | |
8f90be4c NC |
1649 | int |
1650 | mcore_compare_operand (op, mode) | |
1651 | rtx op; | |
1652 | enum machine_mode mode; | |
1653 | { | |
1654 | if (register_operand (op, mode)) | |
1655 | return 1; | |
1656 | ||
1657 | if (GET_CODE (op) == CONST_INT && INTVAL (op) == 0) | |
1658 | return 1; | |
1659 | ||
1660 | return 0; | |
1661 | } | |
1662 | ||
4816b8e4 NC |
1663 | /* Expand insert bit field. BRC */ |
1664 | ||
8f90be4c NC |
1665 | int |
1666 | mcore_expand_insv (operands) | |
1667 | rtx operands[]; | |
1668 | { | |
1669 | int width = INTVAL (operands[1]); | |
1670 | int posn = INTVAL (operands[2]); | |
1671 | int mask; | |
1672 | rtx mreg, sreg, ereg; | |
1673 | ||
1674 | /* To get width 1 insv, the test in store_bit_field() (expmed.c, line 191) | |
1675 | for width==1 must be removed. Look around line 368. This is something | |
4816b8e4 | 1676 | we really want the md part to do. */ |
8f90be4c NC |
1677 | if (width == 1 && GET_CODE (operands[3]) == CONST_INT) |
1678 | { | |
4816b8e4 NC |
1679 | /* Do directly with bseti or bclri. */ |
1680 | /* RBE: 2/97 consider only low bit of constant. */ | |
8f90be4c NC |
1681 | if ((INTVAL(operands[3])&1) == 0) |
1682 | { | |
1683 | mask = ~(1 << posn); | |
1684 | emit_insn (gen_rtx (SET, SImode, operands[0], | |
1685 | gen_rtx (AND, SImode, operands[0], GEN_INT (mask)))); | |
1686 | } | |
1687 | else | |
1688 | { | |
1689 | mask = 1 << posn; | |
1690 | emit_insn (gen_rtx (SET, SImode, operands[0], | |
1691 | gen_rtx (IOR, SImode, operands[0], GEN_INT (mask)))); | |
1692 | } | |
1693 | ||
1694 | return 1; | |
1695 | } | |
1696 | ||
1697 | /* Look at some bitfield placements that we aren't interested | |
4816b8e4 | 1698 | in handling ourselves, unless specifically directed to do so. */ |
8f90be4c NC |
1699 | if (! TARGET_W_FIELD) |
1700 | return 0; /* Generally, give up about now. */ | |
1701 | ||
1702 | if (width == 8 && posn % 8 == 0) | |
1703 | /* Byte sized and aligned; let caller break it up. */ | |
1704 | return 0; | |
1705 | ||
1706 | if (width == 16 && posn % 16 == 0) | |
1707 | /* Short sized and aligned; let caller break it up. */ | |
1708 | return 0; | |
1709 | ||
1710 | /* The general case - we can do this a little bit better than what the | |
1711 | machine independent part tries. This will get rid of all the subregs | |
1712 | that mess up constant folding in combine when working with relaxed | |
4816b8e4 | 1713 | immediates. */ |
8f90be4c NC |
1714 | |
1715 | /* If setting the entire field, do it directly. */ | |
1716 | if (GET_CODE (operands[3]) == CONST_INT && | |
1717 | INTVAL (operands[3]) == ((1 << width) - 1)) | |
1718 | { | |
1719 | mreg = force_reg (SImode, GEN_INT (INTVAL (operands[3]) << posn)); | |
1720 | emit_insn (gen_rtx (SET, SImode, operands[0], | |
1721 | gen_rtx (IOR, SImode, operands[0], mreg))); | |
1722 | return 1; | |
1723 | } | |
1724 | ||
1725 | /* Generate the clear mask. */ | |
1726 | mreg = force_reg (SImode, GEN_INT (~(((1 << width) - 1) << posn))); | |
1727 | ||
1728 | /* Clear the field, to overlay it later with the source. */ | |
1729 | emit_insn (gen_rtx (SET, SImode, operands[0], | |
1730 | gen_rtx (AND, SImode, operands[0], mreg))); | |
1731 | ||
1732 | /* If the source is constant 0, we've nothing to add back. */ | |
1733 | if (GET_CODE (operands[3]) == CONST_INT && INTVAL (operands[3]) == 0) | |
1734 | return 1; | |
1735 | ||
1736 | /* XXX: Should we worry about more games with constant values? | |
1737 | We've covered the high profile: set/clear single-bit and many-bit | |
1738 | fields. How often do we see "arbitrary bit pattern" constants? */ | |
1739 | sreg = copy_to_mode_reg (SImode, operands[3]); | |
1740 | ||
1741 | /* Extract src as same width as dst (needed for signed values). We | |
1742 | always have to do this since we widen everything to SImode. | |
1743 | We don't have to mask if we're shifting this up against the | |
1744 | MSB of the register (e.g., the shift will push out any hi-order | |
4816b8e4 | 1745 | bits. */ |
f27cd94d | 1746 | if (width + posn != (int) GET_MODE_SIZE (SImode)) |
8f90be4c NC |
1747 | { |
1748 | ereg = force_reg (SImode, GEN_INT ((1 << width) - 1)); | |
1749 | emit_insn (gen_rtx (SET, SImode, sreg, | |
1750 | gen_rtx (AND, SImode, sreg, ereg))); | |
1751 | } | |
1752 | ||
4816b8e4 | 1753 | /* Insert source value in dest. */ |
8f90be4c NC |
1754 | if (posn != 0) |
1755 | emit_insn (gen_rtx (SET, SImode, sreg, | |
1756 | gen_rtx (ASHIFT, SImode, sreg, GEN_INT (posn)))); | |
1757 | ||
1758 | emit_insn (gen_rtx (SET, SImode, operands[0], | |
1759 | gen_rtx (IOR, SImode, operands[0], sreg))); | |
1760 | ||
1761 | return 1; | |
1762 | } | |
1763 | ||
1764 | /* Return 1 if OP is a load multiple operation. It is known to be a | |
1765 | PARALLEL and the first section will be tested. */ | |
1766 | int | |
1767 | mcore_load_multiple_operation (op, mode) | |
1768 | rtx op; | |
1769 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1770 | { | |
1771 | int count = XVECLEN (op, 0); | |
1772 | int dest_regno; | |
1773 | rtx src_addr; | |
1774 | int i; | |
1775 | ||
1776 | /* Perform a quick check so we don't blow up below. */ | |
1777 | if (count <= 1 | |
1778 | || GET_CODE (XVECEXP (op, 0, 0)) != SET | |
1779 | || GET_CODE (SET_DEST (XVECEXP (op, 0, 0))) != REG | |
1780 | || GET_CODE (SET_SRC (XVECEXP (op, 0, 0))) != MEM) | |
1781 | return 0; | |
1782 | ||
1783 | dest_regno = REGNO (SET_DEST (XVECEXP (op, 0, 0))); | |
1784 | src_addr = XEXP (SET_SRC (XVECEXP (op, 0, 0)), 0); | |
1785 | ||
1786 | for (i = 1; i < count; i++) | |
1787 | { | |
1788 | rtx elt = XVECEXP (op, 0, i); | |
1789 | ||
1790 | if (GET_CODE (elt) != SET | |
1791 | || GET_CODE (SET_DEST (elt)) != REG | |
1792 | || GET_MODE (SET_DEST (elt)) != SImode | |
f27cd94d NC |
1793 | || REGNO (SET_DEST (elt)) != (unsigned) (dest_regno + i) |
1794 | || GET_CODE (SET_SRC (elt)) != MEM | |
1795 | || GET_MODE (SET_SRC (elt)) != SImode | |
8f90be4c NC |
1796 | || GET_CODE (XEXP (SET_SRC (elt), 0)) != PLUS |
1797 | || ! rtx_equal_p (XEXP (XEXP (SET_SRC (elt), 0), 0), src_addr) | |
1798 | || GET_CODE (XEXP (XEXP (SET_SRC (elt), 0), 1)) != CONST_INT | |
1799 | || INTVAL (XEXP (XEXP (SET_SRC (elt), 0), 1)) != i * 4) | |
1800 | return 0; | |
1801 | } | |
1802 | ||
1803 | return 1; | |
1804 | } | |
1805 | ||
1806 | /* Similar, but tests for store multiple. */ | |
4816b8e4 | 1807 | |
8f90be4c NC |
1808 | int |
1809 | mcore_store_multiple_operation (op, mode) | |
1810 | rtx op; | |
1811 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
1812 | { | |
1813 | int count = XVECLEN (op, 0); | |
1814 | int src_regno; | |
1815 | rtx dest_addr; | |
1816 | int i; | |
1817 | ||
1818 | /* Perform a quick check so we don't blow up below. */ | |
1819 | if (count <= 1 | |
1820 | || GET_CODE (XVECEXP (op, 0, 0)) != SET | |
1821 | || GET_CODE (SET_DEST (XVECEXP (op, 0, 0))) != MEM | |
1822 | || GET_CODE (SET_SRC (XVECEXP (op, 0, 0))) != REG) | |
1823 | return 0; | |
1824 | ||
1825 | src_regno = REGNO (SET_SRC (XVECEXP (op, 0, 0))); | |
1826 | dest_addr = XEXP (SET_DEST (XVECEXP (op, 0, 0)), 0); | |
1827 | ||
1828 | for (i = 1; i < count; i++) | |
1829 | { | |
1830 | rtx elt = XVECEXP (op, 0, i); | |
1831 | ||
1832 | if (GET_CODE (elt) != SET | |
1833 | || GET_CODE (SET_SRC (elt)) != REG | |
1834 | || GET_MODE (SET_SRC (elt)) != SImode | |
f27cd94d | 1835 | || REGNO (SET_SRC (elt)) != (unsigned) (src_regno + i) |
8f90be4c NC |
1836 | || GET_CODE (SET_DEST (elt)) != MEM |
1837 | || GET_MODE (SET_DEST (elt)) != SImode | |
1838 | || GET_CODE (XEXP (SET_DEST (elt), 0)) != PLUS | |
1839 | || ! rtx_equal_p (XEXP (XEXP (SET_DEST (elt), 0), 0), dest_addr) | |
1840 | || GET_CODE (XEXP (XEXP (SET_DEST (elt), 0), 1)) != CONST_INT | |
1841 | || INTVAL (XEXP (XEXP (SET_DEST (elt), 0), 1)) != i * 4) | |
1842 | return 0; | |
1843 | } | |
1844 | ||
1845 | return 1; | |
1846 | } | |
1847 | \f | |
1848 | /* ??? Block move stuff stolen from m88k. This code has not been | |
1849 | verified for correctness. */ | |
1850 | ||
1851 | /* Emit code to perform a block move. Choose the best method. | |
1852 | ||
1853 | OPERANDS[0] is the destination. | |
1854 | OPERANDS[1] is the source. | |
1855 | OPERANDS[2] is the size. | |
1856 | OPERANDS[3] is the alignment safe to use. */ | |
1857 | ||
1858 | /* Emit code to perform a block move with an offset sequence of ldw/st | |
1859 | instructions (..., ldw 0, stw 1, ldw 1, stw 0, ...). SIZE and ALIGN are | |
1860 | known constants. DEST and SRC are registers. OFFSET is the known | |
1861 | starting point for the output pattern. */ | |
1862 | ||
1863 | static enum machine_mode mode_from_align[] = | |
1864 | { | |
1865 | VOIDmode, QImode, HImode, VOIDmode, SImode, | |
1866 | VOIDmode, VOIDmode, VOIDmode, DImode | |
1867 | }; | |
1868 | ||
1869 | static void | |
1870 | block_move_sequence (dest, dst_mem, src, src_mem, size, align, offset) | |
1871 | rtx dest, dst_mem; | |
1872 | rtx src, src_mem; | |
1873 | int size; | |
1874 | int align; | |
1875 | int offset; | |
1876 | { | |
1877 | rtx temp[2]; | |
1878 | enum machine_mode mode[2]; | |
1879 | int amount[2]; | |
1880 | int active[2]; | |
1881 | int phase = 0; | |
1882 | int next; | |
1883 | int offset_ld = offset; | |
1884 | int offset_st = offset; | |
1885 | ||
1886 | active[0] = active[1] = FALSE; | |
1887 | ||
1888 | /* Establish parameters for the first load and for the second load if | |
1889 | it is known to be the same mode as the first. */ | |
1890 | amount[0] = amount[1] = align; | |
1891 | ||
1892 | mode[0] = mode_from_align[align]; | |
1893 | ||
1894 | temp[0] = gen_reg_rtx (mode[0]); | |
1895 | ||
1896 | if (size >= 2 * align) | |
1897 | { | |
1898 | mode[1] = mode[0]; | |
1899 | temp[1] = gen_reg_rtx (mode[1]); | |
1900 | } | |
1901 | ||
1902 | do | |
1903 | { | |
1904 | rtx srcp, dstp; | |
1905 | ||
1906 | next = phase; | |
1907 | phase = !phase; | |
1908 | ||
1909 | if (size > 0) | |
1910 | { | |
1911 | /* Change modes as the sequence tails off. */ | |
1912 | if (size < amount[next]) | |
1913 | { | |
1914 | amount[next] = (size >= 4 ? 4 : (size >= 2 ? 2 : 1)); | |
1915 | mode[next] = mode_from_align[amount[next]]; | |
1916 | temp[next] = gen_reg_rtx (mode[next]); | |
1917 | } | |
1918 | ||
1919 | size -= amount[next]; | |
1920 | srcp = gen_rtx (MEM, | |
1921 | #if 0 | |
1922 | MEM_IN_STRUCT_P (src_mem) ? mode[next] : BLKmode, | |
1923 | #else | |
1924 | mode[next], | |
1925 | #endif | |
1926 | gen_rtx (PLUS, Pmode, src, | |
1927 | gen_rtx (CONST_INT, SImode, offset_ld))); | |
1928 | ||
1929 | RTX_UNCHANGING_P (srcp) = RTX_UNCHANGING_P (src_mem); | |
1930 | MEM_VOLATILE_P (srcp) = MEM_VOLATILE_P (src_mem); | |
1931 | MEM_IN_STRUCT_P (srcp) = 1; | |
1932 | emit_insn (gen_rtx (SET, VOIDmode, temp[next], srcp)); | |
1933 | offset_ld += amount[next]; | |
1934 | active[next] = TRUE; | |
1935 | } | |
1936 | ||
1937 | if (active[phase]) | |
1938 | { | |
1939 | active[phase] = FALSE; | |
1940 | ||
1941 | dstp = gen_rtx (MEM, | |
1942 | #if 0 | |
1943 | MEM_IN_STRUCT_P (dst_mem) ? mode[phase] : BLKmode, | |
1944 | #else | |
1945 | mode[phase], | |
1946 | #endif | |
1947 | gen_rtx (PLUS, Pmode, dest, | |
1948 | gen_rtx (CONST_INT, SImode, offset_st))); | |
1949 | ||
1950 | RTX_UNCHANGING_P (dstp) = RTX_UNCHANGING_P (dst_mem); | |
1951 | MEM_VOLATILE_P (dstp) = MEM_VOLATILE_P (dst_mem); | |
1952 | MEM_IN_STRUCT_P (dstp) = 1; | |
1953 | emit_insn (gen_rtx (SET, VOIDmode, dstp, temp[phase])); | |
1954 | offset_st += amount[phase]; | |
1955 | } | |
1956 | } | |
1957 | while (active[next]); | |
1958 | } | |
1959 | ||
1960 | void | |
1961 | mcore_expand_block_move (dst_mem, src_mem, operands) | |
1962 | rtx dst_mem; | |
1963 | rtx src_mem; | |
1964 | rtx * operands; | |
1965 | { | |
1966 | int align = INTVAL (operands[3]); | |
1967 | int bytes; | |
1968 | ||
1969 | if (GET_CODE (operands[2]) == CONST_INT) | |
1970 | { | |
1971 | bytes = INTVAL (operands[2]); | |
1972 | ||
1973 | if (bytes <= 0) | |
1974 | return; | |
1975 | if (align > 4) | |
1976 | align = 4; | |
1977 | ||
1978 | /* RBE: bumped 1 and 2 byte align from 1 and 2 to 4 and 8 bytes before | |
4816b8e4 | 1979 | we give up and go to memcpy. */ |
8f90be4c NC |
1980 | if ((align == 4 && (bytes <= 4*4 |
1981 | || ((bytes & 01) == 0 && bytes <= 8*4) | |
1982 | || ((bytes & 03) == 0 && bytes <= 16*4))) | |
1983 | || (align == 2 && bytes <= 4*2) | |
1984 | || (align == 1 && bytes <= 4*1)) | |
1985 | { | |
1986 | block_move_sequence (operands[0], dst_mem, operands[1], src_mem, | |
1987 | bytes, align, 0); | |
1988 | return; | |
1989 | } | |
1990 | } | |
1991 | ||
1992 | /* If we get here, just use the library routine. */ | |
1993 | emit_library_call (gen_rtx (SYMBOL_REF, Pmode, "memcpy"), 0, VOIDmode, 3, | |
1994 | operands[0], Pmode, operands[1], Pmode, operands[2], | |
1995 | SImode); | |
1996 | } | |
1997 | \f | |
1998 | ||
1999 | /* Code to generate prologue and epilogue sequences. */ | |
2000 | static int number_of_regs_before_varargs; | |
4816b8e4 | 2001 | |
8f90be4c NC |
2002 | /* Set by SETUP_INCOMING_VARARGS to indicate to prolog that this is |
2003 | for a varargs function. */ | |
2004 | static int current_function_anonymous_args; | |
2005 | ||
8f90be4c NC |
2006 | #define STACK_BYTES (STACK_BOUNDARY/BITS_PER_UNIT) |
2007 | #define STORE_REACH (64) /* Maximum displace of word store + 4. */ | |
4816b8e4 | 2008 | #define ADDI_REACH (32) /* Maximum addi operand. */ |
8f90be4c | 2009 | |
8f90be4c NC |
2010 | static void |
2011 | layout_mcore_frame (infp) | |
2012 | struct mcore_frame * infp; | |
2013 | { | |
2014 | int n; | |
2015 | unsigned int i; | |
2016 | int nbytes; | |
2017 | int regarg; | |
2018 | int localregarg; | |
2019 | int localreg; | |
2020 | int outbounds; | |
2021 | unsigned int growths; | |
2022 | int step; | |
2023 | ||
2024 | /* Might have to spill bytes to re-assemble a big argument that | |
4816b8e4 | 2025 | was passed partially in registers and partially on the stack. */ |
8f90be4c NC |
2026 | nbytes = current_function_pretend_args_size; |
2027 | ||
2028 | /* Determine how much space for spilled anonymous args (e.g., stdarg). */ | |
2029 | if (current_function_anonymous_args) | |
2030 | nbytes += (NPARM_REGS - number_of_regs_before_varargs) * UNITS_PER_WORD; | |
2031 | ||
2032 | infp->arg_size = nbytes; | |
2033 | ||
2034 | /* How much space to save non-volatile registers we stomp. */ | |
2035 | infp->reg_mask = calc_live_regs (& n); | |
2036 | infp->reg_size = n * 4; | |
2037 | ||
2038 | /* And the rest of it... locals and space for overflowed outbounds. */ | |
2039 | infp->local_size = get_frame_size (); | |
2040 | infp->outbound_size = current_function_outgoing_args_size; | |
2041 | ||
2042 | /* Make sure we have a whole number of words for the locals. */ | |
2043 | if (infp->local_size % STACK_BYTES) | |
2044 | infp->local_size = (infp->local_size + STACK_BYTES - 1) & ~ (STACK_BYTES -1); | |
2045 | ||
2046 | /* Only thing we know we have to pad is the outbound space, since | |
2047 | we've aligned our locals assuming that base of locals is aligned. */ | |
2048 | infp->pad_local = 0; | |
2049 | infp->pad_reg = 0; | |
2050 | infp->pad_outbound = 0; | |
2051 | if (infp->outbound_size % STACK_BYTES) | |
2052 | infp->pad_outbound = STACK_BYTES - (infp->outbound_size % STACK_BYTES); | |
2053 | ||
2054 | /* Now we see how we want to stage the prologue so that it does | |
2055 | the most appropriate stack growth and register saves to either: | |
2056 | (1) run fast, | |
2057 | (2) reduce instruction space, or | |
2058 | (3) reduce stack space. */ | |
b6a1cbae | 2059 | for (i = 0; i < ARRAY_SIZE (infp->growth); i++) |
8f90be4c NC |
2060 | infp->growth[i] = 0; |
2061 | ||
2062 | regarg = infp->reg_size + infp->arg_size; | |
2063 | localregarg = infp->local_size + regarg; | |
2064 | localreg = infp->local_size + infp->reg_size; | |
2065 | outbounds = infp->outbound_size + infp->pad_outbound; | |
2066 | growths = 0; | |
2067 | ||
2068 | /* XXX: Consider one where we consider localregarg + outbound too! */ | |
2069 | ||
2070 | /* Frame of <= 32 bytes and using stm would get <= 2 registers. | |
2071 | use stw's with offsets and buy the frame in one shot. */ | |
2072 | if (localregarg <= ADDI_REACH | |
2073 | && (infp->reg_size <= 8 || (infp->reg_mask & 0xc000) != 0xc000)) | |
2074 | { | |
2075 | /* Make sure we'll be aligned. */ | |
2076 | if (localregarg % STACK_BYTES) | |
2077 | infp->pad_reg = STACK_BYTES - (localregarg % STACK_BYTES); | |
2078 | ||
2079 | step = localregarg + infp->pad_reg; | |
2080 | infp->reg_offset = infp->local_size; | |
2081 | ||
2082 | if (outbounds + step <= ADDI_REACH && !frame_pointer_needed) | |
2083 | { | |
2084 | step += outbounds; | |
2085 | infp->reg_offset += outbounds; | |
2086 | outbounds = 0; | |
2087 | } | |
2088 | ||
2089 | infp->arg_offset = step - 4; | |
2090 | infp->growth[growths++] = step; | |
2091 | infp->reg_growth = growths; | |
2092 | infp->local_growth = growths; | |
2093 | ||
4816b8e4 | 2094 | /* If we haven't already folded it in. */ |
8f90be4c NC |
2095 | if (outbounds) |
2096 | infp->growth[growths++] = outbounds; | |
2097 | ||
2098 | goto finish; | |
2099 | } | |
2100 | ||
2101 | /* Frame can't be done with a single subi, but can be done with 2 | |
2102 | insns. If the 'stm' is getting <= 2 registers, we use stw's and | |
2103 | shift some of the stack purchase into the first subi, so both are | |
2104 | single instructions. */ | |
2105 | if (localregarg <= STORE_REACH | |
2106 | && (infp->local_size > ADDI_REACH) | |
2107 | && (infp->reg_size <= 8 || (infp->reg_mask & 0xc000) != 0xc000)) | |
2108 | { | |
2109 | int all; | |
2110 | ||
2111 | /* Make sure we'll be aligned; use either pad_reg or pad_local. */ | |
2112 | if (localregarg % STACK_BYTES) | |
2113 | infp->pad_reg = STACK_BYTES - (localregarg % STACK_BYTES); | |
2114 | ||
2115 | all = localregarg + infp->pad_reg + infp->pad_local; | |
2116 | step = ADDI_REACH; /* As much up front as we can. */ | |
2117 | if (step > all) | |
2118 | step = all; | |
2119 | ||
2120 | /* XXX: Consider whether step will still be aligned; we believe so. */ | |
2121 | infp->arg_offset = step - 4; | |
2122 | infp->growth[growths++] = step; | |
2123 | infp->reg_growth = growths; | |
2124 | infp->reg_offset = step - infp->pad_reg - infp->reg_size; | |
2125 | all -= step; | |
2126 | ||
4816b8e4 | 2127 | /* Can we fold in any space required for outbounds? */ |
8f90be4c NC |
2128 | if (outbounds + all <= ADDI_REACH && !frame_pointer_needed) |
2129 | { | |
2130 | all += outbounds; | |
2131 | outbounds = 0; | |
2132 | } | |
2133 | ||
4816b8e4 | 2134 | /* Get the rest of the locals in place. */ |
8f90be4c NC |
2135 | step = all; |
2136 | infp->growth[growths++] = step; | |
2137 | infp->local_growth = growths; | |
2138 | all -= step; | |
2139 | ||
2140 | assert (all == 0); | |
2141 | ||
4816b8e4 | 2142 | /* Finish off if we need to do so. */ |
8f90be4c NC |
2143 | if (outbounds) |
2144 | infp->growth[growths++] = outbounds; | |
2145 | ||
2146 | goto finish; | |
2147 | } | |
2148 | ||
2149 | /* Registers + args is nicely aligned, so we'll buy that in one shot. | |
2150 | Then we buy the rest of the frame in 1 or 2 steps depending on | |
2151 | whether we need a frame pointer. */ | |
2152 | if ((regarg % STACK_BYTES) == 0) | |
2153 | { | |
2154 | infp->growth[growths++] = regarg; | |
2155 | infp->reg_growth = growths; | |
2156 | infp->arg_offset = regarg - 4; | |
2157 | infp->reg_offset = 0; | |
2158 | ||
2159 | if (infp->local_size % STACK_BYTES) | |
2160 | infp->pad_local = STACK_BYTES - (infp->local_size % STACK_BYTES); | |
2161 | ||
2162 | step = infp->local_size + infp->pad_local; | |
2163 | ||
2164 | if (!frame_pointer_needed) | |
2165 | { | |
2166 | step += outbounds; | |
2167 | outbounds = 0; | |
2168 | } | |
2169 | ||
2170 | infp->growth[growths++] = step; | |
2171 | infp->local_growth = growths; | |
2172 | ||
4816b8e4 | 2173 | /* If there's any left to be done. */ |
8f90be4c NC |
2174 | if (outbounds) |
2175 | infp->growth[growths++] = outbounds; | |
2176 | ||
2177 | goto finish; | |
2178 | } | |
2179 | ||
2180 | /* XXX: optimizations that we'll want to play with.... | |
4816b8e4 NC |
2181 | -- regarg is not aligned, but it's a small number of registers; |
2182 | use some of localsize so that regarg is aligned and then | |
2183 | save the registers. */ | |
8f90be4c NC |
2184 | |
2185 | /* Simple encoding; plods down the stack buying the pieces as it goes. | |
4816b8e4 NC |
2186 | -- does not optimize space consumption. |
2187 | -- does not attempt to optimize instruction counts. | |
2188 | -- but it is safe for all alignments. */ | |
8f90be4c NC |
2189 | if (regarg % STACK_BYTES != 0) |
2190 | infp->pad_reg = STACK_BYTES - (regarg % STACK_BYTES); | |
2191 | ||
2192 | infp->growth[growths++] = infp->arg_size + infp->reg_size + infp->pad_reg; | |
2193 | infp->reg_growth = growths; | |
2194 | infp->arg_offset = infp->growth[0] - 4; | |
2195 | infp->reg_offset = 0; | |
2196 | ||
2197 | if (frame_pointer_needed) | |
2198 | { | |
2199 | if (infp->local_size % STACK_BYTES != 0) | |
2200 | infp->pad_local = STACK_BYTES - (infp->local_size % STACK_BYTES); | |
2201 | ||
2202 | infp->growth[growths++] = infp->local_size + infp->pad_local; | |
2203 | infp->local_growth = growths; | |
2204 | ||
2205 | infp->growth[growths++] = outbounds; | |
2206 | } | |
2207 | else | |
2208 | { | |
2209 | if ((infp->local_size + outbounds) % STACK_BYTES != 0) | |
2210 | infp->pad_local = STACK_BYTES - ((infp->local_size + outbounds) % STACK_BYTES); | |
2211 | ||
2212 | infp->growth[growths++] = infp->local_size + infp->pad_local + outbounds; | |
2213 | infp->local_growth = growths; | |
2214 | } | |
2215 | ||
f27cd94d | 2216 | /* Anything else that we've forgotten?, plus a few consistency checks. */ |
8f90be4c NC |
2217 | finish: |
2218 | assert (infp->reg_offset >= 0); | |
2219 | assert (growths <= MAX_STACK_GROWS); | |
2220 | ||
2221 | for (i = 0; i < growths; i++) | |
2222 | { | |
2223 | if (infp->growth[i] % STACK_BYTES) | |
2224 | { | |
2225 | fprintf (stderr,"stack growth of %d is not %d aligned\n", | |
2226 | infp->growth[i], STACK_BYTES); | |
2227 | abort (); | |
2228 | } | |
2229 | } | |
2230 | } | |
2231 | ||
2232 | /* Define the offset between two registers, one to be eliminated, and | |
2233 | the other its replacement, at the start of a routine. */ | |
4816b8e4 | 2234 | |
8f90be4c NC |
2235 | int |
2236 | mcore_initial_elimination_offset (from, to) | |
2237 | int from; | |
2238 | int to; | |
2239 | { | |
2240 | int above_frame; | |
2241 | int below_frame; | |
2242 | struct mcore_frame fi; | |
2243 | ||
2244 | layout_mcore_frame (& fi); | |
2245 | ||
2246 | /* fp to ap */ | |
2247 | above_frame = fi.local_size + fi.pad_local + fi.reg_size + fi.pad_reg; | |
2248 | /* sp to fp */ | |
2249 | below_frame = fi.outbound_size + fi.pad_outbound; | |
2250 | ||
2251 | if (from == ARG_POINTER_REGNUM && to == FRAME_POINTER_REGNUM) | |
2252 | return above_frame; | |
2253 | ||
2254 | if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
2255 | return above_frame + below_frame; | |
2256 | ||
2257 | if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM) | |
2258 | return below_frame; | |
2259 | ||
2260 | abort (); | |
2261 | ||
2262 | return 0; | |
2263 | } | |
2264 | ||
4816b8e4 NC |
2265 | /* Keep track of some information about varargs for the prolog. */ |
2266 | ||
8f90be4c NC |
2267 | void |
2268 | mcore_setup_incoming_varargs (args_so_far, mode, type, ptr_pretend_size) | |
2269 | CUMULATIVE_ARGS args_so_far; | |
2270 | enum machine_mode mode; | |
2271 | tree type; | |
2272 | int * ptr_pretend_size ATTRIBUTE_UNUSED; | |
2273 | { | |
2274 | current_function_anonymous_args = 1; | |
2275 | ||
2276 | /* We need to know how many argument registers are used before | |
2277 | the varargs start, so that we can push the remaining argument | |
2278 | registers during the prologue. */ | |
2279 | number_of_regs_before_varargs = args_so_far + mcore_num_arg_regs (mode, type); | |
2280 | ||
2281 | /* There is a bug somwehere in the arg handling code. | |
2282 | Until I can find it this workaround always pushes the | |
2283 | last named argument onto the stack. */ | |
2284 | number_of_regs_before_varargs = args_so_far; | |
2285 | ||
2286 | /* The last named argument may be split between argument registers | |
2287 | and the stack. Allow for this here. */ | |
2288 | if (number_of_regs_before_varargs > NPARM_REGS) | |
2289 | number_of_regs_before_varargs = NPARM_REGS; | |
2290 | } | |
2291 | ||
2292 | void | |
2293 | mcore_expand_prolog () | |
2294 | { | |
2295 | struct mcore_frame fi; | |
2296 | int space_allocated = 0; | |
2297 | int growth = 0; | |
2298 | ||
2299 | /* Find out what we're doing. */ | |
2300 | layout_mcore_frame (&fi); | |
2301 | ||
2302 | space_allocated = fi.arg_size + fi.reg_size + fi.local_size + | |
2303 | fi.outbound_size + fi.pad_outbound + fi.pad_local + fi.pad_reg; | |
2304 | ||
2305 | if (TARGET_CG_DATA) | |
2306 | { | |
2307 | /* Emit a symbol for this routine's frame size. */ | |
2308 | rtx x; | |
2309 | int len; | |
2310 | ||
2311 | x = DECL_RTL (current_function_decl); | |
2312 | ||
2313 | if (GET_CODE (x) != MEM) | |
2314 | abort (); | |
2315 | ||
2316 | x = XEXP (x, 0); | |
2317 | ||
2318 | if (GET_CODE (x) != SYMBOL_REF) | |
2319 | abort (); | |
2320 | ||
2321 | if (mcore_current_function_name) | |
2322 | free (mcore_current_function_name); | |
2323 | ||
2324 | len = strlen (XSTR (x, 0)) + 1; | |
dd3b81b4 | 2325 | mcore_current_function_name = (char *) xmalloc (len); |
8f90be4c NC |
2326 | |
2327 | memcpy (mcore_current_function_name, XSTR (x, 0), len); | |
2328 | ||
2329 | ASM_OUTPUT_CG_NODE (asm_out_file, mcore_current_function_name, space_allocated); | |
2330 | ||
2331 | if (current_function_calls_alloca) | |
2332 | ASM_OUTPUT_CG_EDGE (asm_out_file, mcore_current_function_name, "alloca", 1); | |
2333 | ||
2334 | /* 970425: RBE: | |
2335 | We're looking at how the 8byte alignment affects stack layout | |
2336 | and where we had to pad things. This emits information we can | |
2337 | extract which tells us about frame sizes and the like. */ | |
2338 | fprintf (asm_out_file, | |
2339 | "\t.equ\t__$frame$info$_%s_$_%d_%d_x%x_%d_%d_%d,0\n", | |
2340 | mcore_current_function_name, | |
2341 | fi.arg_size, fi.reg_size, fi.reg_mask, | |
2342 | fi.local_size, fi.outbound_size, | |
2343 | frame_pointer_needed); | |
2344 | } | |
2345 | ||
2346 | if (mcore_naked_function_p ()) | |
2347 | return; | |
2348 | ||
2349 | /* Handle stdarg+regsaves in one shot: can't be more than 64 bytes. */ | |
2350 | output_stack_adjust (-1, fi.growth[growth++]); /* grows it */ | |
2351 | ||
2352 | /* If we have a parameter passed partially in regs and partially in memory, | |
2353 | the registers will have been stored to memory already in function.c. So | |
2354 | we only need to do something here for varargs functions. */ | |
2355 | if (fi.arg_size != 0 && current_function_pretend_args_size == 0) | |
2356 | { | |
2357 | int offset; | |
2358 | int rn = FIRST_PARM_REG + NPARM_REGS - 1; | |
2359 | int remaining = fi.arg_size; | |
2360 | ||
2361 | for (offset = fi.arg_offset; remaining >= 4; offset -= 4, rn--, remaining -= 4) | |
2362 | { | |
2363 | emit_insn (gen_movsi | |
2364 | (gen_rtx (MEM, SImode, | |
2365 | plus_constant (stack_pointer_rtx, offset)), | |
2366 | gen_rtx (REG, SImode, rn))); | |
2367 | } | |
2368 | } | |
2369 | ||
4816b8e4 | 2370 | /* Do we need another stack adjustment before we do the register saves? */ |
8f90be4c NC |
2371 | if (growth < fi.reg_growth) |
2372 | output_stack_adjust (-1, fi.growth[growth++]); /* grows it */ | |
2373 | ||
2374 | if (fi.reg_size != 0) | |
2375 | { | |
2376 | int i; | |
2377 | int offs = fi.reg_offset; | |
2378 | ||
2379 | for (i = 15; i >= 0; i--) | |
2380 | { | |
2381 | if (offs == 0 && i == 15 && ((fi.reg_mask & 0xc000) == 0xc000)) | |
2382 | { | |
2383 | int first_reg = 15; | |
2384 | ||
2385 | while (fi.reg_mask & (1 << first_reg)) | |
2386 | first_reg--; | |
2387 | first_reg++; | |
2388 | ||
2389 | emit_insn (gen_store_multiple (gen_rtx (MEM, SImode, stack_pointer_rtx), | |
2390 | gen_rtx (REG, SImode, first_reg), | |
2391 | GEN_INT (16 - first_reg))); | |
2392 | ||
2393 | i -= (15 - first_reg); | |
2394 | offs += (16 - first_reg) * 4; | |
2395 | } | |
2396 | else if (fi.reg_mask & (1 << i)) | |
2397 | { | |
2398 | emit_insn (gen_movsi | |
2399 | (gen_rtx (MEM, SImode, | |
2400 | plus_constant (stack_pointer_rtx, offs)), | |
2401 | gen_rtx (REG, SImode, i))); | |
2402 | offs += 4; | |
2403 | } | |
2404 | } | |
2405 | } | |
2406 | ||
2407 | /* Figure the locals + outbounds. */ | |
2408 | if (frame_pointer_needed) | |
2409 | { | |
2410 | /* If we haven't already purchased to 'fp'. */ | |
2411 | if (growth < fi.local_growth) | |
2412 | output_stack_adjust (-1, fi.growth[growth++]); /* grows it */ | |
2413 | ||
2414 | emit_insn (gen_movsi (frame_pointer_rtx, stack_pointer_rtx)); | |
2415 | ||
4816b8e4 | 2416 | /* ... and then go any remaining distance for outbounds, etc. */ |
8f90be4c NC |
2417 | if (fi.growth[growth]) |
2418 | output_stack_adjust (-1, fi.growth[growth++]); | |
2419 | } | |
2420 | else | |
2421 | { | |
2422 | if (growth < fi.local_growth) | |
2423 | output_stack_adjust (-1, fi.growth[growth++]); /* grows it */ | |
2424 | if (fi.growth[growth]) | |
2425 | output_stack_adjust (-1, fi.growth[growth++]); | |
2426 | } | |
2427 | } | |
2428 | ||
2429 | void | |
2430 | mcore_expand_epilog () | |
2431 | { | |
2432 | struct mcore_frame fi; | |
2433 | int i; | |
2434 | int offs; | |
2435 | int growth = MAX_STACK_GROWS - 1 ; | |
2436 | ||
f27cd94d | 2437 | |
8f90be4c NC |
2438 | /* Find out what we're doing. */ |
2439 | layout_mcore_frame(&fi); | |
2440 | ||
2441 | if (mcore_naked_function_p ()) | |
2442 | return; | |
f27cd94d | 2443 | |
8f90be4c NC |
2444 | /* If we had a frame pointer, restore the sp from that. */ |
2445 | if (frame_pointer_needed) | |
2446 | { | |
2447 | emit_insn (gen_movsi (stack_pointer_rtx, frame_pointer_rtx)); | |
2448 | growth = fi.local_growth - 1; | |
2449 | } | |
2450 | else | |
2451 | { | |
2452 | /* XXX: while loop should accumulate and do a single sell. */ | |
2453 | while (growth >= fi.local_growth) | |
2454 | { | |
2455 | if (fi.growth[growth] != 0) | |
2456 | output_stack_adjust (1, fi.growth[growth]); | |
2457 | growth--; | |
2458 | } | |
2459 | } | |
2460 | ||
2461 | /* Make sure we've shrunk stack back to the point where the registers | |
2462 | were laid down. This is typically 0/1 iterations. Then pull the | |
4816b8e4 | 2463 | register save information back off the stack. */ |
8f90be4c NC |
2464 | while (growth >= fi.reg_growth) |
2465 | output_stack_adjust ( 1, fi.growth[growth--]); | |
2466 | ||
2467 | offs = fi.reg_offset; | |
2468 | ||
2469 | for (i = 15; i >= 0; i--) | |
2470 | { | |
2471 | if (offs == 0 && i == 15 && ((fi.reg_mask & 0xc000) == 0xc000)) | |
2472 | { | |
2473 | int first_reg; | |
2474 | ||
2475 | /* Find the starting register. */ | |
2476 | first_reg = 15; | |
2477 | ||
2478 | while (fi.reg_mask & (1 << first_reg)) | |
2479 | first_reg--; | |
2480 | ||
2481 | first_reg++; | |
2482 | ||
2483 | emit_insn (gen_load_multiple (gen_rtx (REG, SImode, first_reg), | |
2484 | gen_rtx (MEM, SImode, stack_pointer_rtx), | |
2485 | GEN_INT (16 - first_reg))); | |
2486 | ||
2487 | i -= (15 - first_reg); | |
2488 | offs += (16 - first_reg) * 4; | |
2489 | } | |
2490 | else if (fi.reg_mask & (1 << i)) | |
2491 | { | |
2492 | emit_insn (gen_movsi | |
2493 | (gen_rtx (REG, SImode, i), | |
2494 | gen_rtx (MEM, SImode, | |
2495 | plus_constant (stack_pointer_rtx, offs)))); | |
2496 | offs += 4; | |
2497 | } | |
2498 | } | |
2499 | ||
2500 | /* Give back anything else. */ | |
4816b8e4 | 2501 | /* XXX: Should accumuate total and then give it back. */ |
8f90be4c NC |
2502 | while (growth >= 0) |
2503 | output_stack_adjust ( 1, fi.growth[growth--]); | |
2504 | } | |
2505 | \f | |
2506 | /* This code is borrowed from the SH port. */ | |
2507 | ||
2508 | /* The MCORE cannot load a large constant into a register, constants have to | |
2509 | come from a pc relative load. The reference of a pc relative load | |
2510 | instruction must be less than 1k infront of the instruction. This | |
2511 | means that we often have to dump a constant inside a function, and | |
2512 | generate code to branch around it. | |
2513 | ||
2514 | It is important to minimize this, since the branches will slow things | |
2515 | down and make things bigger. | |
2516 | ||
2517 | Worst case code looks like: | |
2518 | ||
2519 | lrw L1,r0 | |
2520 | br L2 | |
2521 | align | |
2522 | L1: .long value | |
2523 | L2: | |
2524 | .. | |
2525 | ||
2526 | lrw L3,r0 | |
2527 | br L4 | |
2528 | align | |
2529 | L3: .long value | |
2530 | L4: | |
2531 | .. | |
2532 | ||
2533 | We fix this by performing a scan before scheduling, which notices which | |
2534 | instructions need to have their operands fetched from the constant table | |
2535 | and builds the table. | |
2536 | ||
2537 | The algorithm is: | |
2538 | ||
2539 | scan, find an instruction which needs a pcrel move. Look forward, find the | |
2540 | last barrier which is within MAX_COUNT bytes of the requirement. | |
2541 | If there isn't one, make one. Process all the instructions between | |
2542 | the find and the barrier. | |
2543 | ||
2544 | In the above example, we can tell that L3 is within 1k of L1, so | |
2545 | the first move can be shrunk from the 2 insn+constant sequence into | |
2546 | just 1 insn, and the constant moved to L3 to make: | |
2547 | ||
2548 | lrw L1,r0 | |
2549 | .. | |
2550 | lrw L3,r0 | |
2551 | bra L4 | |
2552 | align | |
2553 | L3:.long value | |
2554 | L4:.long value | |
2555 | ||
2556 | Then the second move becomes the target for the shortening process. */ | |
2557 | ||
2558 | typedef struct | |
2559 | { | |
2560 | rtx value; /* Value in table. */ | |
2561 | rtx label; /* Label of value. */ | |
2562 | } pool_node; | |
2563 | ||
2564 | /* The maximum number of constants that can fit into one pool, since | |
2565 | the pc relative range is 0...1020 bytes and constants are at least 4 | |
2566 | bytes long. We subtact 4 from the range to allow for the case where | |
2567 | we need to add a branch/align before the constant pool. */ | |
2568 | ||
2569 | #define MAX_COUNT 1016 | |
2570 | #define MAX_POOL_SIZE (MAX_COUNT/4) | |
2571 | static pool_node pool_vector[MAX_POOL_SIZE]; | |
2572 | static int pool_size; | |
2573 | ||
2574 | /* Dump out any constants accumulated in the final pass. These | |
2575 | will only be labels. */ | |
4816b8e4 | 2576 | |
f27cd94d | 2577 | const char * |
8f90be4c NC |
2578 | mcore_output_jump_label_table () |
2579 | { | |
2580 | int i; | |
2581 | ||
2582 | if (pool_size) | |
2583 | { | |
2584 | fprintf (asm_out_file, "\t.align 2\n"); | |
2585 | ||
2586 | for (i = 0; i < pool_size; i++) | |
2587 | { | |
2588 | pool_node * p = pool_vector + i; | |
2589 | ||
2590 | ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L", CODE_LABEL_NUMBER (p->label)); | |
2591 | ||
2592 | output_asm_insn (".long %0", &p->value); | |
2593 | } | |
2594 | ||
2595 | pool_size = 0; | |
2596 | } | |
2597 | ||
2598 | return ""; | |
2599 | } | |
2600 | ||
2601 | #if 0 /* XXX temporarily suppressed until I have time to look at what this code does. */ | |
2602 | ||
2603 | /* We need these below. They use information stored in tables to figure out | |
2604 | what values are in what registers, etc. This is okay, since these tables | |
2605 | are valid at the time mcore_dependent_simplify_rtx() is invoked. Don't | |
4816b8e4 | 2606 | use them anywhere else. BRC */ |
8f90be4c NC |
2607 | |
2608 | extern unsigned HOST_WIDE_INT nonzero_bits PARAMS ((rtx, enum machine_mode)); | |
2609 | extern int num_sign_bit_copies PARAMS ((Rtx, enum machine_mode)); | |
2610 | ||
2611 | /* Do machine dependent simplifications: see simplify_rtx() in combine.c. | |
2612 | GENERAL_SIMPLIFY controls whether general machine independent | |
2613 | simplifications should be tried after machine dependent ones. Thus, | |
2614 | we can filter out certain simplifications and keep the simplify_rtx() | |
2615 | from changing things that we just simplified in a machine dependent | |
4816b8e4 | 2616 | fashion. This is experimental. BRC */ |
8f90be4c NC |
2617 | rtx |
2618 | mcore_dependent_simplify_rtx (x, int_op0_mode, last, in_dest, general_simplify) | |
2619 | rtx x; | |
2620 | int int_op0_mode; | |
2621 | int last; | |
2622 | int in_dest; | |
2623 | int * general_simplify; | |
2624 | { | |
2625 | enum machine_mode mode = GET_MODE (x); | |
2626 | enum rtx_code code = GET_CODE (x); | |
2627 | ||
4816b8e4 | 2628 | /* Always simplify unless explicitly asked not to. */ |
8f90be4c NC |
2629 | * general_simplify = 1; |
2630 | ||
2631 | if (code == IF_THEN_ELSE) | |
2632 | { | |
2633 | int i; | |
2634 | rtx cond = XEXP(x, 0); | |
d6edb99e ZW |
2635 | rtx true_rtx = XEXP(x, 1); |
2636 | rtx false_rtx = XEXP(x, 2); | |
8f90be4c NC |
2637 | enum rtx_code true_code = GET_CODE (cond); |
2638 | ||
2639 | /* On the mcore, when doing -mcmov-one, we don't want to simplify: | |
2640 | ||
2641 | (if_then_else (ne A 0) C1 0) | |
2642 | ||
2643 | if it would be turned into a shift by simplify_if_then_else(). | |
2644 | instead, leave it alone so that it will collapse into a conditional | |
2645 | move. besides, at least for the mcore, doing this simplification does | |
4816b8e4 | 2646 | not typically help. see combine.c, line 4217. BRC */ |
8f90be4c NC |
2647 | |
2648 | if (true_code == NE && XEXP (cond, 1) == const0_rtx | |
d6edb99e | 2649 | && false_rtx == const0_rtx && GET_CODE (true_rtx) == CONST_INT |
8f90be4c | 2650 | && ((1 == nonzero_bits (XEXP (cond, 0), mode) |
d6edb99e | 2651 | && (i = exact_log2 (INTVAL (true_rtx))) >= 0) |
8f90be4c NC |
2652 | || ((num_sign_bit_copies (XEXP (cond, 0), mode) |
2653 | == GET_MODE_BITSIZE (mode)) | |
d6edb99e | 2654 | && (i = exact_log2 (- INTVAL (true_rtx))) >= 0))) |
8f90be4c NC |
2655 | { |
2656 | *general_simplify = 0; | |
2657 | return x; | |
2658 | } | |
2659 | } | |
2660 | ||
2661 | return x; | |
2662 | } | |
2663 | #endif | |
2664 | ||
8f90be4c | 2665 | /* Check whether insn is a candidate for a conditional. */ |
4816b8e4 | 2666 | |
8f90be4c NC |
2667 | static cond_type |
2668 | is_cond_candidate (insn) | |
2669 | rtx insn; | |
2670 | { | |
2671 | /* The only things we conditionalize are those that can be directly | |
2672 | changed into a conditional. Only bother with SImode items. If | |
2673 | we wanted to be a little more aggressive, we could also do other | |
4816b8e4 | 2674 | modes such as DImode with reg-reg move or load 0. */ |
8f90be4c NC |
2675 | if (GET_CODE (insn) == INSN) |
2676 | { | |
2677 | rtx pat = PATTERN (insn); | |
2678 | rtx src, dst; | |
2679 | ||
2680 | if (GET_CODE (pat) != SET) | |
2681 | return COND_NO; | |
2682 | ||
2683 | dst = XEXP (pat, 0); | |
2684 | ||
2685 | if ((GET_CODE (dst) != REG && | |
2686 | GET_CODE (dst) != SUBREG) || | |
2687 | GET_MODE (dst) != SImode) | |
2688 | return COND_NO; | |
2689 | ||
2690 | src = XEXP (pat, 1); | |
2691 | ||
2692 | if ((GET_CODE (src) == REG || | |
2693 | (GET_CODE (src) == SUBREG && | |
2694 | GET_CODE (SUBREG_REG (src)) == REG)) && | |
2695 | GET_MODE (src) == SImode) | |
2696 | return COND_MOV_INSN; | |
2697 | else if (GET_CODE (src) == CONST_INT && | |
2698 | INTVAL (src) == 0) | |
2699 | return COND_CLR_INSN; | |
2700 | else if (GET_CODE (src) == PLUS && | |
2701 | (GET_CODE (XEXP (src, 0)) == REG || | |
2702 | (GET_CODE (XEXP (src, 0)) == SUBREG && | |
2703 | GET_CODE (SUBREG_REG (XEXP (src, 0))) == REG)) && | |
2704 | GET_MODE (XEXP (src, 0)) == SImode && | |
2705 | GET_CODE (XEXP (src, 1)) == CONST_INT && | |
2706 | INTVAL (XEXP (src, 1)) == 1) | |
2707 | return COND_INC_INSN; | |
2708 | else if (((GET_CODE (src) == MINUS && | |
2709 | GET_CODE (XEXP (src, 1)) == CONST_INT && | |
2710 | INTVAL( XEXP (src, 1)) == 1) || | |
2711 | (GET_CODE (src) == PLUS && | |
2712 | GET_CODE (XEXP (src, 1)) == CONST_INT && | |
2713 | INTVAL (XEXP (src, 1)) == -1)) && | |
2714 | (GET_CODE (XEXP (src, 0)) == REG || | |
2715 | (GET_CODE (XEXP (src, 0)) == SUBREG && | |
2716 | GET_CODE (SUBREG_REG (XEXP (src, 0))) == REG)) && | |
2717 | GET_MODE (XEXP (src, 0)) == SImode) | |
2718 | return COND_DEC_INSN; | |
2719 | ||
2720 | /* some insns that we don't bother with: | |
2721 | (set (rx:DI) (ry:DI)) | |
2722 | (set (rx:DI) (const_int 0)) | |
2723 | */ | |
2724 | ||
2725 | } | |
2726 | else if (GET_CODE (insn) == JUMP_INSN && | |
2727 | GET_CODE (PATTERN (insn)) == SET && | |
2728 | GET_CODE (XEXP (PATTERN (insn), 1)) == LABEL_REF) | |
2729 | return COND_BRANCH_INSN; | |
2730 | ||
2731 | return COND_NO; | |
2732 | } | |
2733 | ||
2734 | /* Emit a conditional version of insn and replace the old insn with the | |
2735 | new one. Return the new insn if emitted. */ | |
4816b8e4 | 2736 | |
8f90be4c NC |
2737 | static rtx |
2738 | emit_new_cond_insn (insn, cond) | |
2739 | rtx insn; | |
2740 | int cond; | |
2741 | { | |
2742 | rtx c_insn = 0; | |
2743 | rtx pat, dst, src; | |
2744 | cond_type num; | |
2745 | ||
2746 | if ((num = is_cond_candidate (insn)) == COND_NO) | |
2747 | return NULL; | |
2748 | ||
2749 | pat = PATTERN (insn); | |
2750 | ||
2751 | if (GET_CODE (insn) == INSN) | |
2752 | { | |
2753 | dst = SET_DEST (pat); | |
2754 | src = SET_SRC (pat); | |
2755 | } | |
2756 | else | |
2757 | dst = JUMP_LABEL (insn); | |
2758 | ||
2759 | switch (num) | |
2760 | { | |
2761 | case COND_MOV_INSN: | |
2762 | case COND_CLR_INSN: | |
2763 | if (cond) | |
2764 | c_insn = gen_movt0 (dst, src, dst); | |
2765 | else | |
2766 | c_insn = gen_movt0 (dst, dst, src); | |
2767 | break; | |
2768 | ||
2769 | case COND_INC_INSN: | |
2770 | if (cond) | |
2771 | c_insn = gen_incscc (dst, dst); | |
2772 | else | |
2773 | c_insn = gen_incscc_false (dst, dst); | |
2774 | break; | |
2775 | ||
2776 | case COND_DEC_INSN: | |
2777 | if (cond) | |
2778 | c_insn = gen_decscc (dst, dst); | |
2779 | else | |
2780 | c_insn = gen_decscc_false (dst, dst); | |
2781 | break; | |
2782 | ||
2783 | case COND_BRANCH_INSN: | |
2784 | if (cond) | |
2785 | c_insn = gen_branch_true (dst); | |
2786 | else | |
2787 | c_insn = gen_branch_false (dst); | |
2788 | break; | |
2789 | ||
2790 | default: | |
2791 | return NULL; | |
2792 | } | |
2793 | ||
2794 | /* Only copy the notes if they exist. */ | |
2795 | if (rtx_length [GET_CODE (c_insn)] >= 7 && rtx_length [GET_CODE (insn)] >= 7) | |
2796 | { | |
2797 | /* We really don't need to bother with the notes and links at this | |
2798 | point, but go ahead and save the notes. This will help is_dead() | |
2799 | when applying peepholes (links don't matter since they are not | |
2800 | used any more beyond this point for the mcore). */ | |
2801 | REG_NOTES (c_insn) = REG_NOTES (insn); | |
2802 | } | |
2803 | ||
2804 | if (num == COND_BRANCH_INSN) | |
2805 | { | |
2806 | /* For jumps, we need to be a little bit careful and emit the new jump | |
2807 | before the old one and to update the use count for the target label. | |
2808 | This way, the barrier following the old (uncond) jump will get | |
2809 | deleted, but the label won't. */ | |
2810 | c_insn = emit_jump_insn_before (c_insn, insn); | |
2811 | ||
2812 | ++ LABEL_NUSES (dst); | |
2813 | ||
2814 | JUMP_LABEL (c_insn) = dst; | |
2815 | } | |
2816 | else | |
2817 | c_insn = emit_insn_after (c_insn, insn); | |
2818 | ||
2819 | delete_insn (insn); | |
2820 | ||
2821 | return c_insn; | |
2822 | } | |
2823 | ||
2824 | /* Attempt to change a basic block into a series of conditional insns. This | |
2825 | works by taking the branch at the end of the 1st block and scanning for the | |
2826 | end of the 2nd block. If all instructions in the 2nd block have cond. | |
2827 | versions and the label at the start of block 3 is the same as the target | |
2828 | from the branch at block 1, then conditionalize all insn in block 2 using | |
2829 | the inverse condition of the branch at block 1. (Note I'm bending the | |
2830 | definition of basic block here.) | |
2831 | ||
2832 | e.g., change: | |
2833 | ||
2834 | bt L2 <-- end of block 1 (delete) | |
2835 | mov r7,r8 | |
2836 | addu r7,1 | |
2837 | br L3 <-- end of block 2 | |
2838 | ||
2839 | L2: ... <-- start of block 3 (NUSES==1) | |
2840 | L3: ... | |
2841 | ||
2842 | to: | |
2843 | ||
2844 | movf r7,r8 | |
2845 | incf r7 | |
2846 | bf L3 | |
2847 | ||
2848 | L3: ... | |
2849 | ||
2850 | we can delete the L2 label if NUSES==1 and re-apply the optimization | |
2851 | starting at the last instruction of block 2. This may allow an entire | |
4816b8e4 | 2852 | if-then-else statement to be conditionalized. BRC */ |
8f90be4c NC |
2853 | static rtx |
2854 | conditionalize_block (first) | |
2855 | rtx first; | |
2856 | { | |
2857 | rtx insn; | |
2858 | rtx br_pat; | |
2859 | rtx end_blk_1_br = 0; | |
2860 | rtx end_blk_2_insn = 0; | |
2861 | rtx start_blk_3_lab = 0; | |
2862 | int cond; | |
2863 | int br_lab_num; | |
2864 | int blk_size = 0; | |
2865 | ||
2866 | ||
2867 | /* Check that the first insn is a candidate conditional jump. This is | |
2868 | the one that we'll eliminate. If not, advance to the next insn to | |
2869 | try. */ | |
2870 | if (GET_CODE (first) != JUMP_INSN || | |
2871 | GET_CODE (PATTERN (first)) != SET || | |
2872 | GET_CODE (XEXP (PATTERN (first), 1)) != IF_THEN_ELSE) | |
2873 | return NEXT_INSN (first); | |
2874 | ||
2875 | /* Extract some information we need. */ | |
2876 | end_blk_1_br = first; | |
2877 | br_pat = PATTERN (end_blk_1_br); | |
2878 | ||
2879 | /* Complement the condition since we use the reverse cond. for the insns. */ | |
2880 | cond = (GET_CODE (XEXP (XEXP (br_pat, 1), 0)) == EQ); | |
2881 | ||
2882 | /* Determine what kind of branch we have. */ | |
2883 | if (GET_CODE (XEXP (XEXP (br_pat, 1), 1)) == LABEL_REF) | |
2884 | { | |
2885 | /* A normal branch, so extract label out of first arm. */ | |
2886 | br_lab_num = CODE_LABEL_NUMBER (XEXP (XEXP (XEXP (br_pat, 1), 1), 0)); | |
2887 | } | |
2888 | else | |
2889 | { | |
2890 | /* An inverse branch, so extract the label out of the 2nd arm | |
2891 | and complement the condition. */ | |
2892 | cond = (cond == 0); | |
2893 | br_lab_num = CODE_LABEL_NUMBER (XEXP (XEXP (XEXP (br_pat, 1), 2), 0)); | |
2894 | } | |
2895 | ||
2896 | /* Scan forward for the start of block 2: it must start with a | |
2897 | label and that label must be the same as the branch target | |
2898 | label from block 1. We don't care about whether block 2 actually | |
2899 | ends with a branch or a label (an uncond. branch is | |
2900 | conditionalizable). */ | |
2901 | for (insn = NEXT_INSN (first); insn; insn = NEXT_INSN (insn)) | |
2902 | { | |
2903 | enum rtx_code code; | |
2904 | ||
2905 | code = GET_CODE (insn); | |
2906 | ||
2907 | /* Look for the label at the start of block 3. */ | |
2908 | if (code == CODE_LABEL && CODE_LABEL_NUMBER (insn) == br_lab_num) | |
2909 | break; | |
2910 | ||
2911 | /* Skip barriers, notes, and conditionalizable insns. If the | |
2912 | insn is not conditionalizable or makes this optimization fail, | |
2913 | just return the next insn so we can start over from that point. */ | |
2914 | if (code != BARRIER && code != NOTE && !is_cond_candidate (insn)) | |
2915 | return NEXT_INSN (insn); | |
2916 | ||
2917 | /* Remember the last real insn before the label (ie end of block 2). */ | |
2918 | if (code == JUMP_INSN || code == INSN) | |
2919 | { | |
2920 | blk_size ++; | |
2921 | end_blk_2_insn = insn; | |
2922 | } | |
2923 | } | |
2924 | ||
2925 | if (!insn) | |
2926 | return insn; | |
2927 | ||
2928 | /* It is possible for this optimization to slow performance if the blocks | |
2929 | are long. This really depends upon whether the branch is likely taken | |
2930 | or not. If the branch is taken, we slow performance in many cases. But, | |
2931 | if the branch is not taken, we always help performance (for a single | |
2932 | block, but for a double block (i.e. when the optimization is re-applied) | |
2933 | this is not true since the 'right thing' depends on the overall length of | |
2934 | the collapsed block). As a compromise, don't apply this optimization on | |
2935 | blocks larger than size 2 (unlikely for the mcore) when speed is important. | |
2936 | the best threshold depends on the latencies of the instructions (i.e., | |
2937 | the branch penalty). */ | |
2938 | if (optimize > 1 && blk_size > 2) | |
2939 | return insn; | |
2940 | ||
2941 | /* At this point, we've found the start of block 3 and we know that | |
2942 | it is the destination of the branch from block 1. Also, all | |
2943 | instructions in the block 2 are conditionalizable. So, apply the | |
2944 | conditionalization and delete the branch. */ | |
2945 | start_blk_3_lab = insn; | |
2946 | ||
2947 | for (insn = NEXT_INSN (end_blk_1_br); insn != start_blk_3_lab; | |
2948 | insn = NEXT_INSN (insn)) | |
2949 | { | |
2950 | rtx newinsn; | |
2951 | ||
2952 | if (INSN_DELETED_P (insn)) | |
2953 | continue; | |
2954 | ||
2955 | /* Try to form a conditional variant of the instruction and emit it. */ | |
2956 | if ((newinsn = emit_new_cond_insn (insn, cond))) | |
2957 | { | |
2958 | if (end_blk_2_insn == insn) | |
2959 | end_blk_2_insn = newinsn; | |
2960 | ||
2961 | insn = newinsn; | |
2962 | } | |
2963 | } | |
2964 | ||
2965 | /* Note whether we will delete the label starting blk 3 when the jump | |
2966 | gets deleted. If so, we want to re-apply this optimization at the | |
2967 | last real instruction right before the label. */ | |
2968 | if (LABEL_NUSES (start_blk_3_lab) == 1) | |
2969 | { | |
2970 | start_blk_3_lab = 0; | |
2971 | } | |
2972 | ||
2973 | /* ??? we probably should redistribute the death notes for this insn, esp. | |
2974 | the death of cc, but it doesn't really matter this late in the game. | |
2975 | The peepholes all use is_dead() which will find the correct death | |
2976 | regardless of whether there is a note. */ | |
2977 | delete_insn (end_blk_1_br); | |
2978 | ||
2979 | if (! start_blk_3_lab) | |
2980 | return end_blk_2_insn; | |
2981 | ||
4816b8e4 | 2982 | /* Return the insn right after the label at the start of block 3. */ |
8f90be4c NC |
2983 | return NEXT_INSN (start_blk_3_lab); |
2984 | } | |
2985 | ||
2986 | /* Apply the conditionalization of blocks optimization. This is the | |
2987 | outer loop that traverses through the insns scanning for a branch | |
2988 | that signifies an opportunity to apply the optimization. Note that | |
2989 | this optimization is applied late. If we could apply it earlier, | |
2990 | say before cse 2, it may expose more optimization opportunities. | |
2991 | but, the pay back probably isn't really worth the effort (we'd have | |
2992 | to update all reg/flow/notes/links/etc to make it work - and stick it | |
4816b8e4 NC |
2993 | in before cse 2). */ |
2994 | ||
8f90be4c NC |
2995 | static void |
2996 | conditionalize_optimization (first) | |
2997 | rtx first; | |
2998 | { | |
2999 | rtx insn; | |
3000 | ||
3001 | for (insn = first; insn; insn = conditionalize_block (insn)) | |
3002 | continue; | |
3003 | } | |
3004 | ||
3005 | static int saved_warn_return_type = -1; | |
3006 | static int saved_warn_return_type_count = 0; | |
3007 | ||
4816b8e4 NC |
3008 | /* This function is called from toplev.c before reorg. */ |
3009 | ||
8f90be4c NC |
3010 | void |
3011 | mcore_dependent_reorg (first) | |
3012 | rtx first; | |
3013 | { | |
3014 | /* Reset this variable. */ | |
3015 | current_function_anonymous_args = 0; | |
3016 | ||
4816b8e4 | 3017 | /* Restore the warn_return_type if it has been altered. */ |
8f90be4c NC |
3018 | if (saved_warn_return_type != -1) |
3019 | { | |
3020 | /* Only restore the value if we have reached another function. | |
3021 | The test of warn_return_type occurs in final_function () in | |
3022 | c-decl.c a long time after the code for the function is generated, | |
3023 | so we need a counter to tell us when we have finished parsing that | |
3024 | function and can restore the flag. */ | |
3025 | if (--saved_warn_return_type_count == 0) | |
3026 | { | |
3027 | warn_return_type = saved_warn_return_type; | |
3028 | saved_warn_return_type = -1; | |
3029 | } | |
3030 | } | |
3031 | ||
3032 | if (optimize == 0) | |
3033 | return; | |
3034 | ||
3035 | /* Conditionalize blocks where we can. */ | |
3036 | conditionalize_optimization (first); | |
3037 | ||
3038 | /* Literal pool generation is now pushed off until the assembler. */ | |
3039 | } | |
3040 | ||
3041 | \f | |
3042 | /* Return the reg_class to use when reloading the rtx X into the class | |
3043 | CLASS. */ | |
3044 | ||
3045 | /* If the input is (PLUS REG CONSTANT) representing a stack slot address, | |
3046 | then we want to restrict the class to LRW_REGS since that ensures that | |
3047 | will be able to safely load the constant. | |
3048 | ||
3049 | If the input is a constant that should be loaded with mvir1, then use | |
3050 | ONLYR1_REGS. | |
3051 | ||
3052 | ??? We don't handle the case where we have (PLUS REG CONSTANT) and | |
3053 | the constant should be loaded with mvir1, because that can lead to cases | |
3054 | where an instruction needs two ONLYR1_REGS reloads. */ | |
3055 | enum reg_class | |
3056 | mcore_reload_class (x, class) | |
3057 | rtx x; | |
3058 | enum reg_class class; | |
3059 | { | |
3060 | enum reg_class new_class; | |
3061 | ||
3062 | if (class == GENERAL_REGS && CONSTANT_P (x) | |
3063 | && (GET_CODE (x) != CONST_INT | |
3064 | || ( ! CONST_OK_FOR_I (INTVAL (x)) | |
3065 | && ! CONST_OK_FOR_M (INTVAL (x)) | |
3066 | && ! CONST_OK_FOR_N (INTVAL (x))))) | |
3067 | new_class = LRW_REGS; | |
3068 | else | |
3069 | new_class = class; | |
3070 | ||
3071 | return new_class; | |
3072 | } | |
3073 | ||
3074 | /* Tell me if a pair of reg/subreg rtx's actually refer to the same | |
3075 | register. Note that the current version doesn't worry about whether | |
3076 | they are the same mode or note (e.g., a QImode in r2 matches an HImode | |
3077 | in r2 matches an SImode in r2. Might think in the future about whether | |
3078 | we want to be able to say something about modes. */ | |
3079 | int | |
3080 | mcore_is_same_reg (x, y) | |
3081 | rtx x; | |
3082 | rtx y; | |
3083 | { | |
3084 | /* Strip any and all of the subreg wrappers. */ | |
3085 | while (GET_CODE (x) == SUBREG) | |
3086 | x = SUBREG_REG (x); | |
3087 | ||
3088 | while (GET_CODE (y) == SUBREG) | |
3089 | y = SUBREG_REG (y); | |
3090 | ||
3091 | if (GET_CODE(x) == REG && GET_CODE(y) == REG && REGNO(x) == REGNO(y)) | |
3092 | return 1; | |
3093 | ||
3094 | return 0; | |
3095 | } | |
3096 | ||
3097 | /* Called to register all of our global variables with the garbage | |
3098 | collector. */ | |
3099 | static void | |
3100 | mcore_add_gc_roots () | |
3101 | { | |
3102 | ggc_add_rtx_root (&arch_compare_op0, 1); | |
3103 | ggc_add_rtx_root (&arch_compare_op1, 1); | |
3104 | } | |
3105 | ||
3106 | void | |
3107 | mcore_override_options () | |
3108 | { | |
3109 | if (mcore_stack_increment_string) | |
3110 | { | |
3111 | mcore_stack_increment = atoi (mcore_stack_increment_string); | |
3112 | ||
3113 | if (mcore_stack_increment < 0 | |
3114 | || (mcore_stack_increment == 0 | |
3115 | && (mcore_stack_increment_string[0] != '0' | |
3116 | || mcore_stack_increment_string[1] != 0))) | |
3117 | error ("Invalid option `-mstack-increment=%s'", | |
3118 | mcore_stack_increment_string); | |
3119 | } | |
3120 | ||
3121 | /* Only the m340 supports little endian code. */ | |
3122 | if (TARGET_LITTLE_END && ! TARGET_M340) | |
3123 | target_flags |= M340_BIT; | |
3124 | ||
3125 | mcore_add_gc_roots (); | |
3126 | } | |
3127 | \f | |
3128 | int | |
3129 | mcore_must_pass_on_stack (mode, type) | |
3130 | enum machine_mode mode ATTRIBUTE_UNUSED; | |
3131 | tree type; | |
3132 | { | |
3133 | if (type == NULL) | |
3134 | return 0; | |
3135 | ||
3136 | /* If the argugment can have its address taken, it must | |
3137 | be placed on the stack. */ | |
3138 | if (TREE_ADDRESSABLE (type)) | |
3139 | return 1; | |
3140 | ||
3141 | return 0; | |
3142 | } | |
3143 | ||
3144 | /* Compute the number of word sized registers needed to | |
3145 | hold a function argument of mode MODE and type TYPE. */ | |
3146 | int | |
3147 | mcore_num_arg_regs (mode, type) | |
3148 | enum machine_mode mode; | |
3149 | tree type; | |
3150 | { | |
3151 | int size; | |
3152 | ||
3153 | if (MUST_PASS_IN_STACK (mode, type)) | |
3154 | return 0; | |
3155 | ||
3156 | if (type && mode == BLKmode) | |
3157 | size = int_size_in_bytes (type); | |
3158 | else | |
3159 | size = GET_MODE_SIZE (mode); | |
3160 | ||
3161 | return ROUND_ADVANCE (size); | |
3162 | } | |
3163 | ||
3164 | static rtx | |
3165 | handle_structs_in_regs (mode, type, reg) | |
3166 | enum machine_mode mode; | |
3167 | tree type; | |
3168 | int reg; | |
3169 | { | |
3170 | int size; | |
3171 | ||
3172 | /* The MCore ABI defines that a structure whoes size is not a whole multiple | |
3173 | of bytes is passed packed into registers (or spilled onto the stack if | |
3174 | not enough registers are available) with the last few bytes of the | |
3175 | structure being packed, left-justified, into the last register/stack slot. | |
3176 | GCC handles this correctly if the last word is in a stack slot, but we | |
3177 | have to generate a special, PARALLEL RTX if the last word is in an | |
3178 | argument register. */ | |
3179 | if (type | |
3180 | && TYPE_MODE (type) == BLKmode | |
3181 | && TREE_CODE (TYPE_SIZE (type)) == INTEGER_CST | |
3182 | && (size = int_size_in_bytes (type)) > UNITS_PER_WORD | |
3183 | && (size % UNITS_PER_WORD != 0) | |
3184 | && (reg + mcore_num_arg_regs (mode, type) <= (FIRST_PARM_REG + NPARM_REGS))) | |
3185 | { | |
3186 | rtx arg_regs [NPARM_REGS]; | |
3187 | int nregs; | |
3188 | rtx result; | |
3189 | rtvec rtvec; | |
3190 | ||
3191 | for (nregs = 0; size > 0; size -= UNITS_PER_WORD) | |
3192 | { | |
3193 | arg_regs [nregs] = | |
3194 | gen_rtx_EXPR_LIST (SImode, gen_rtx_REG (SImode, reg ++), | |
3195 | GEN_INT (nregs * UNITS_PER_WORD)); | |
3196 | nregs ++; | |
3197 | } | |
3198 | ||
3199 | /* We assume here that NPARM_REGS == 6. The assert checks this. */ | |
b6a1cbae | 3200 | assert (ARRAY_SIZE (arg_regs) == 6); |
8f90be4c NC |
3201 | rtvec = gen_rtvec (nregs, arg_regs[0], arg_regs[1], arg_regs[2], |
3202 | arg_regs[3], arg_regs[4], arg_regs[5]); | |
3203 | ||
3204 | result = gen_rtx_PARALLEL (mode, rtvec); | |
3205 | return result; | |
3206 | } | |
3207 | ||
3208 | return gen_rtx_REG (mode, reg); | |
3209 | } | |
3210 | ||
3211 | rtx | |
3212 | mcore_function_value (valtype, func) | |
3213 | tree valtype; | |
3214 | tree func ATTRIBUTE_UNUSED; | |
3215 | { | |
3216 | enum machine_mode mode; | |
3217 | int unsigned_p; | |
3218 | ||
3219 | mode = TYPE_MODE (valtype); | |
3220 | ||
3221 | PROMOTE_MODE (mode, unsigned_p, NULL); | |
3222 | ||
3223 | return handle_structs_in_regs (mode, valtype, FIRST_RET_REG); | |
3224 | } | |
3225 | ||
3226 | /* Define where to put the arguments to a function. | |
3227 | Value is zero to push the argument on the stack, | |
3228 | or a hard register in which to store the argument. | |
3229 | ||
3230 | MODE is the argument's machine mode. | |
3231 | TYPE is the data type of the argument (as a tree). | |
3232 | This is null for libcalls where that information may | |
3233 | not be available. | |
3234 | CUM is a variable of type CUMULATIVE_ARGS which gives info about | |
3235 | the preceding args and about the function being called. | |
3236 | NAMED is nonzero if this argument is a named parameter | |
3237 | (otherwise it is an extra parameter matching an ellipsis). | |
3238 | ||
3239 | On MCore the first args are normally in registers | |
3240 | and the rest are pushed. Any arg that starts within the first | |
3241 | NPARM_REGS words is at least partially passed in a register unless | |
3242 | its data type forbids. */ | |
3243 | rtx | |
3244 | mcore_function_arg (cum, mode, type, named) | |
3245 | CUMULATIVE_ARGS cum; | |
3246 | enum machine_mode mode; | |
3247 | tree type; | |
3248 | int named; | |
3249 | { | |
3250 | int arg_reg; | |
3251 | ||
3252 | if (! named) | |
3253 | return 0; | |
3254 | ||
3255 | if (MUST_PASS_IN_STACK (mode, type)) | |
3256 | return 0; | |
3257 | ||
3258 | arg_reg = ROUND_REG (cum, mode); | |
3259 | ||
3260 | if (arg_reg < NPARM_REGS) | |
3261 | return handle_structs_in_regs (mode, type, FIRST_PARM_REG + arg_reg); | |
3262 | ||
3263 | return 0; | |
3264 | } | |
3265 | ||
3266 | /* Implements the FUNCTION_ARG_PARTIAL_NREGS macro. | |
3267 | Returns the number of argument registers required to hold *part* of | |
3268 | a parameter of machine mode MODE and type TYPE (which may be NULL if | |
3269 | the type is not known). If the argument fits entirly in the argument | |
3270 | registers, or entirely on the stack, then 0 is returned. CUM is the | |
3271 | number of argument registers already used by earlier parameters to | |
3272 | the function. */ | |
3273 | int | |
3274 | mcore_function_arg_partial_nregs (cum, mode, type, named) | |
3275 | CUMULATIVE_ARGS cum; | |
3276 | enum machine_mode mode; | |
3277 | tree type; | |
3278 | int named; | |
3279 | { | |
3280 | int reg = ROUND_REG (cum, mode); | |
3281 | ||
3282 | if (named == 0) | |
3283 | return 0; | |
3284 | ||
3285 | if (MUST_PASS_IN_STACK (mode, type)) | |
3286 | return 0; | |
3287 | ||
3288 | /* REG is not the *hardware* register number of the register that holds | |
3289 | the argument, it is the *argument* register number. So for example, | |
3290 | the first argument to a function goes in argument register 0, which | |
3291 | translates (for the MCore) into hardware register 2. The second | |
3292 | argument goes into argument register 1, which translates into hardware | |
3293 | register 3, and so on. NPARM_REGS is the number of argument registers | |
3294 | supported by the target, not the maximum hardware register number of | |
3295 | the target. */ | |
3296 | if (reg >= NPARM_REGS) | |
3297 | return 0; | |
3298 | ||
3299 | /* If the argument fits entirely in registers, return 0. */ | |
3300 | if (reg + mcore_num_arg_regs (mode, type) <= NPARM_REGS) | |
3301 | return 0; | |
3302 | ||
3303 | /* The argument overflows the number of available argument registers. | |
3304 | Compute how many argument registers have not yet been assigned to | |
3305 | hold an argument. */ | |
3306 | reg = NPARM_REGS - reg; | |
3307 | ||
3308 | /* Return partially in registers and partially on the stack. */ | |
3309 | return reg; | |
3310 | } | |
3311 | \f | |
3312 | /* Return non-zero if SYMBOL is marked as being dllexport'd. */ | |
3313 | int | |
3314 | mcore_dllexport_name_p (symbol) | |
cbd3488b | 3315 | const char * symbol; |
8f90be4c NC |
3316 | { |
3317 | return symbol[0] == '@' && symbol[1] == 'e' && symbol[2] == '.'; | |
3318 | } | |
3319 | ||
3320 | /* Return non-zero if SYMBOL is marked as being dllimport'd. */ | |
3321 | int | |
3322 | mcore_dllimport_name_p (symbol) | |
cbd3488b | 3323 | const char * symbol; |
8f90be4c NC |
3324 | { |
3325 | return symbol[0] == '@' && symbol[1] == 'i' && symbol[2] == '.'; | |
3326 | } | |
3327 | ||
3328 | /* Mark a DECL as being dllexport'd. */ | |
3329 | static void | |
3330 | mcore_mark_dllexport (decl) | |
3331 | tree decl; | |
3332 | { | |
cbd3488b | 3333 | const char * oldname; |
8f90be4c NC |
3334 | char * newname; |
3335 | rtx rtlname; | |
3336 | tree idp; | |
3337 | ||
3338 | rtlname = XEXP (DECL_RTL (decl), 0); | |
3339 | ||
3340 | if (GET_CODE (rtlname) == SYMBOL_REF) | |
3341 | oldname = XSTR (rtlname, 0); | |
3342 | else if ( GET_CODE (rtlname) == MEM | |
3343 | && GET_CODE (XEXP (rtlname, 0)) == SYMBOL_REF) | |
3344 | oldname = XSTR (XEXP (rtlname, 0), 0); | |
3345 | else | |
3346 | abort (); | |
3347 | ||
3348 | if (mcore_dllexport_name_p (oldname)) | |
3349 | return; /* Already done. */ | |
3350 | ||
3351 | newname = alloca (strlen (oldname) + 4); | |
3352 | sprintf (newname, "@e.%s", oldname); | |
3353 | ||
3354 | /* We pass newname through get_identifier to ensure it has a unique | |
3355 | address. RTL processing can sometimes peek inside the symbol ref | |
3356 | and compare the string's addresses to see if two symbols are | |
3357 | identical. */ | |
3358 | /* ??? At least I think that's why we do this. */ | |
3359 | idp = get_identifier (newname); | |
3360 | ||
3361 | XEXP (DECL_RTL (decl), 0) = | |
3362 | gen_rtx (SYMBOL_REF, Pmode, IDENTIFIER_POINTER (idp)); | |
3363 | } | |
3364 | ||
3365 | /* Mark a DECL as being dllimport'd. */ | |
3366 | static void | |
3367 | mcore_mark_dllimport (decl) | |
3368 | tree decl; | |
3369 | { | |
cbd3488b | 3370 | const char * oldname; |
8f90be4c NC |
3371 | char * newname; |
3372 | tree idp; | |
3373 | rtx rtlname; | |
3374 | rtx newrtl; | |
3375 | ||
3376 | rtlname = XEXP (DECL_RTL (decl), 0); | |
3377 | ||
3378 | if (GET_CODE (rtlname) == SYMBOL_REF) | |
3379 | oldname = XSTR (rtlname, 0); | |
3380 | else if ( GET_CODE (rtlname) == MEM | |
3381 | && GET_CODE (XEXP (rtlname, 0)) == SYMBOL_REF) | |
3382 | oldname = XSTR (XEXP (rtlname, 0), 0); | |
3383 | else | |
3384 | abort (); | |
3385 | ||
3386 | if (mcore_dllexport_name_p (oldname)) | |
3387 | abort (); /* This shouldn't happen. */ | |
3388 | else if (mcore_dllimport_name_p (oldname)) | |
3389 | return; /* Already done. */ | |
3390 | ||
3391 | /* ??? One can well ask why we're making these checks here, | |
3392 | and that would be a good question. */ | |
3393 | ||
3394 | /* Imported variables can't be initialized. */ | |
3395 | if (TREE_CODE (decl) == VAR_DECL | |
3396 | && !DECL_VIRTUAL_P (decl) | |
3397 | && DECL_INITIAL (decl)) | |
3398 | { | |
3399 | error_with_decl (decl, "initialized variable `%s' is marked dllimport"); | |
3400 | return; | |
3401 | } | |
3402 | ||
3403 | /* `extern' needn't be specified with dllimport. | |
3404 | Specify `extern' now and hope for the best. Sigh. */ | |
3405 | if (TREE_CODE (decl) == VAR_DECL | |
3406 | /* ??? Is this test for vtables needed? */ | |
3407 | && !DECL_VIRTUAL_P (decl)) | |
3408 | { | |
3409 | DECL_EXTERNAL (decl) = 1; | |
3410 | TREE_PUBLIC (decl) = 1; | |
3411 | } | |
3412 | ||
3413 | newname = alloca (strlen (oldname) + 11); | |
3414 | sprintf (newname, "@i.__imp_%s", oldname); | |
3415 | ||
3416 | /* We pass newname through get_identifier to ensure it has a unique | |
3417 | address. RTL processing can sometimes peek inside the symbol ref | |
3418 | and compare the string's addresses to see if two symbols are | |
3419 | identical. */ | |
3420 | /* ??? At least I think that's why we do this. */ | |
3421 | idp = get_identifier (newname); | |
3422 | ||
3423 | newrtl = gen_rtx (MEM, Pmode, | |
3424 | gen_rtx (SYMBOL_REF, Pmode, | |
3425 | IDENTIFIER_POINTER (idp))); | |
3426 | XEXP (DECL_RTL (decl), 0) = newrtl; | |
3427 | } | |
3428 | ||
3429 | static int | |
3430 | mcore_dllexport_p (decl) | |
3431 | tree decl; | |
3432 | { | |
3433 | if ( TREE_CODE (decl) != VAR_DECL | |
3434 | && TREE_CODE (decl) != FUNCTION_DECL) | |
3435 | return 0; | |
3436 | ||
3437 | return lookup_attribute ("dllexport", DECL_MACHINE_ATTRIBUTES (decl)) != 0; | |
3438 | } | |
3439 | ||
3440 | static int | |
3441 | mcore_dllimport_p (decl) | |
3442 | tree decl; | |
3443 | { | |
3444 | if ( TREE_CODE (decl) != VAR_DECL | |
3445 | && TREE_CODE (decl) != FUNCTION_DECL) | |
3446 | return 0; | |
3447 | ||
3448 | return lookup_attribute ("dllimport", DECL_MACHINE_ATTRIBUTES (decl)) != 0; | |
3449 | } | |
3450 | ||
3451 | /* Cover function to implement ENCODE_SECTION_INFO. */ | |
3452 | void | |
3453 | mcore_encode_section_info (decl) | |
3454 | tree decl; | |
3455 | { | |
3456 | /* This bit is copied from arm.h. */ | |
3457 | if (optimize > 0 | |
3458 | && TREE_CONSTANT (decl) | |
3459 | && (!flag_writable_strings || TREE_CODE (decl) != STRING_CST)) | |
3460 | { | |
3461 | rtx rtl = (TREE_CODE_CLASS (TREE_CODE (decl)) != 'd' | |
3462 | ? TREE_CST_RTL (decl) : DECL_RTL (decl)); | |
3463 | SYMBOL_REF_FLAG (XEXP (rtl, 0)) = 1; | |
3464 | } | |
3465 | ||
3466 | /* Mark the decl so we can tell from the rtl whether the object is | |
3467 | dllexport'd or dllimport'd. */ | |
3468 | if (mcore_dllexport_p (decl)) | |
3469 | mcore_mark_dllexport (decl); | |
3470 | else if (mcore_dllimport_p (decl)) | |
3471 | mcore_mark_dllimport (decl); | |
3472 | ||
3473 | /* It might be that DECL has already been marked as dllimport, but | |
3474 | a subsequent definition nullified that. The attribute is gone | |
3475 | but DECL_RTL still has @i.__imp_foo. We need to remove that. */ | |
3476 | else if ((TREE_CODE (decl) == FUNCTION_DECL | |
3477 | || TREE_CODE (decl) == VAR_DECL) | |
3478 | && DECL_RTL (decl) != NULL_RTX | |
3479 | && GET_CODE (DECL_RTL (decl)) == MEM | |
3480 | && GET_CODE (XEXP (DECL_RTL (decl), 0)) == MEM | |
3481 | && GET_CODE (XEXP (XEXP (DECL_RTL (decl), 0), 0)) == SYMBOL_REF | |
3482 | && mcore_dllimport_name_p (XSTR (XEXP (XEXP (DECL_RTL (decl), 0), 0), 0))) | |
3483 | { | |
3cce094d | 3484 | const char * oldname = XSTR (XEXP (XEXP (DECL_RTL (decl), 0), 0), 0); |
8f90be4c NC |
3485 | tree idp = get_identifier (oldname + 9); |
3486 | rtx newrtl = gen_rtx (SYMBOL_REF, Pmode, IDENTIFIER_POINTER (idp)); | |
3487 | ||
3488 | XEXP (DECL_RTL (decl), 0) = newrtl; | |
3489 | ||
3490 | /* We previously set TREE_PUBLIC and DECL_EXTERNAL. | |
3491 | ??? We leave these alone for now. */ | |
3492 | } | |
3493 | } | |
3494 | ||
3495 | /* MCore specific attribute support. | |
3496 | dllexport - for exporting a function/variable that will live in a dll | |
3497 | dllimport - for importing a function/variable from a dll | |
3498 | naked - do not create a function prologue/epilogue. */ | |
3499 | int | |
3500 | mcore_valid_machine_decl_attribute (decl, attributes, attr, args) | |
3501 | tree decl; | |
3502 | tree attributes ATTRIBUTE_UNUSED; | |
3503 | tree attr; | |
3504 | tree args; | |
3505 | { | |
3506 | if (args != NULL_TREE) | |
3507 | return 0; | |
3508 | ||
3509 | if (is_attribute_p ("dllexport", attr)) | |
3510 | return 1; | |
3511 | ||
3512 | if (is_attribute_p ("dllimport", attr)) | |
3513 | return 1; | |
3514 | ||
3515 | if (is_attribute_p ("naked", attr) && | |
3516 | TREE_CODE (decl) == FUNCTION_DECL) | |
3517 | { | |
3518 | /* PR14310 - don't complain about lack of return statement | |
3519 | in naked functions. The solution here is a gross hack | |
3520 | but this is the only way to solve the problem without | |
3521 | adding a new feature to GCC. I did try submitting a patch | |
3522 | that would add such a new feature, but it was (rightfully) | |
3523 | rejected on the grounds that it was creeping featurism, | |
3524 | so hence this code. */ | |
3525 | if (warn_return_type) | |
3526 | { | |
3527 | saved_warn_return_type = warn_return_type; | |
3528 | warn_return_type = 0; | |
3529 | saved_warn_return_type_count = 2; | |
3530 | } | |
3531 | else if (saved_warn_return_type_count) | |
3532 | saved_warn_return_type_count = 2; | |
3533 | ||
3534 | return 1; | |
3535 | } | |
3536 | ||
3537 | return 0; | |
3538 | } | |
3539 | ||
3540 | /* Merge attributes in decls OLD and NEW. | |
3541 | This handles the following situation: | |
3542 | ||
3543 | __declspec (dllimport) int foo; | |
3544 | int foo; | |
3545 | ||
3546 | The second instance of `foo' nullifies the dllimport. */ | |
3547 | tree | |
3548 | mcore_merge_machine_decl_attributes (old, new) | |
3549 | tree old; | |
3550 | tree new; | |
3551 | { | |
3552 | tree a; | |
3553 | int delete_dllimport_p; | |
3554 | ||
3555 | old = DECL_MACHINE_ATTRIBUTES (old); | |
3556 | new = DECL_MACHINE_ATTRIBUTES (new); | |
3557 | ||
3558 | /* What we need to do here is remove from `old' dllimport if it doesn't | |
3559 | appear in `new'. dllimport behaves like extern: if a declaration is | |
3560 | marked dllimport and a definition appears later, then the object | |
3561 | is not dllimport'd. */ | |
3562 | if ( lookup_attribute ("dllimport", old) != NULL_TREE | |
3563 | && lookup_attribute ("dllimport", new) == NULL_TREE) | |
3564 | delete_dllimport_p = 1; | |
3565 | else | |
3566 | delete_dllimport_p = 0; | |
3567 | ||
3568 | a = merge_attributes (old, new); | |
3569 | ||
3570 | if (delete_dllimport_p) | |
3571 | { | |
3572 | tree prev,t; | |
3573 | ||
3574 | /* Scan the list for dllimport and delete it. */ | |
3575 | for (prev = NULL_TREE, t = a; t; prev = t, t = TREE_CHAIN (t)) | |
3576 | { | |
3577 | if (is_attribute_p ("dllimport", TREE_PURPOSE (t))) | |
3578 | { | |
3579 | if (prev == NULL_TREE) | |
3580 | a = TREE_CHAIN (a); | |
3581 | else | |
3582 | TREE_CHAIN (prev) = TREE_CHAIN (t); | |
3583 | break; | |
3584 | } | |
3585 | } | |
3586 | } | |
3587 | ||
3588 | return a; | |
3589 | } | |
3590 | ||
3591 | /* Cover function for UNIQUE_SECTION. */ | |
3592 | ||
3593 | void | |
3594 | mcore_unique_section (decl, reloc) | |
3595 | tree decl; | |
3596 | int reloc ATTRIBUTE_UNUSED; | |
3597 | { | |
3598 | int len; | |
3599 | char * name; | |
3600 | char * string; | |
f27cd94d | 3601 | const char * prefix; |
8f90be4c NC |
3602 | |
3603 | name = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl)); | |
3604 | ||
3605 | /* Strip off any encoding in name. */ | |
3606 | STRIP_NAME_ENCODING (name, name); | |
3607 | ||
3608 | /* The object is put in, for example, section .text$foo. | |
3609 | The linker will then ultimately place them in .text | |
3610 | (everything from the $ on is stripped). */ | |
3611 | if (TREE_CODE (decl) == FUNCTION_DECL) | |
3612 | prefix = ".text$"; | |
3613 | /* For compatability with EPOC, we ignore the fact that the | |
3614 | section might have relocs against it. */ | |
3615 | else if (DECL_READONLY_SECTION (decl, 0)) | |
3616 | prefix = ".rdata$"; | |
3617 | else | |
3618 | prefix = ".data$"; | |
3619 | ||
3620 | len = strlen (name) + strlen (prefix); | |
3621 | string = alloca (len + 1); | |
3622 | ||
3623 | sprintf (string, "%s%s", prefix, name); | |
3624 | ||
3625 | DECL_SECTION_NAME (decl) = build_string (len, string); | |
3626 | } | |
3627 | ||
3628 | int | |
3629 | mcore_naked_function_p () | |
3630 | { | |
3631 | return lookup_attribute ("naked", DECL_MACHINE_ATTRIBUTES (current_function_decl)) != NULL_TREE; | |
3632 | } |