| 1 | /* ehopt.c--optimize gcc exception frame information. |
| 2 | Copyright 1998, 2000, 2001, 2003, 2005, 2007, 2008 |
| 3 | Free Software Foundation, Inc. |
| 4 | Written by Ian Lance Taylor <ian@cygnus.com>. |
| 5 | |
| 6 | This file is part of GAS, the GNU Assembler. |
| 7 | |
| 8 | GAS is free software; you can redistribute it and/or modify |
| 9 | it under the terms of the GNU General Public License as published by |
| 10 | the Free Software Foundation; either version 3, or (at your option) |
| 11 | any later version. |
| 12 | |
| 13 | GAS is distributed in the hope that it will be useful, |
| 14 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 16 | GNU General Public License for more details. |
| 17 | |
| 18 | You should have received a copy of the GNU General Public License |
| 19 | along with GAS; see the file COPYING. If not, write to the Free |
| 20 | Software Foundation, 51 Franklin Street - Fifth Floor, Boston, MA |
| 21 | 02110-1301, USA. */ |
| 22 | |
| 23 | #include "as.h" |
| 24 | #include "subsegs.h" |
| 25 | #include "struc-symbol.h" |
| 26 | |
| 27 | /* We include this ELF file, even though we may not be assembling for |
| 28 | ELF, since the exception frame information is always in a format |
| 29 | derived from DWARF. */ |
| 30 | |
| 31 | #include "elf/dwarf2.h" |
| 32 | |
| 33 | /* Try to optimize gcc 2.8 exception frame information. |
| 34 | |
| 35 | Exception frame information is emitted for every function in the |
| 36 | .eh_frame or .debug_frame sections. Simple information for a function |
| 37 | with no exceptions looks like this: |
| 38 | |
| 39 | __FRAME_BEGIN__: |
| 40 | .4byte .LLCIE1 / Length of Common Information Entry |
| 41 | .LSCIE1: |
| 42 | #if .eh_frame |
| 43 | .4byte 0x0 / CIE Identifier Tag |
| 44 | #elif .debug_frame |
| 45 | .4byte 0xffffffff / CIE Identifier Tag |
| 46 | #endif |
| 47 | .byte 0x1 / CIE Version |
| 48 | .byte 0x0 / CIE Augmentation (none) |
| 49 | .byte 0x1 / ULEB128 0x1 (CIE Code Alignment Factor) |
| 50 | .byte 0x7c / SLEB128 -4 (CIE Data Alignment Factor) |
| 51 | .byte 0x8 / CIE RA Column |
| 52 | .byte 0xc / DW_CFA_def_cfa |
| 53 | .byte 0x4 / ULEB128 0x4 |
| 54 | .byte 0x4 / ULEB128 0x4 |
| 55 | .byte 0x88 / DW_CFA_offset, column 0x8 |
| 56 | .byte 0x1 / ULEB128 0x1 |
| 57 | .align 4 |
| 58 | .LECIE1: |
| 59 | .set .LLCIE1,.LECIE1-.LSCIE1 / CIE Length Symbol |
| 60 | .4byte .LLFDE1 / FDE Length |
| 61 | .LSFDE1: |
| 62 | .4byte .LSFDE1-__FRAME_BEGIN__ / FDE CIE offset |
| 63 | .4byte .LFB1 / FDE initial location |
| 64 | .4byte .LFE1-.LFB1 / FDE address range |
| 65 | .byte 0x4 / DW_CFA_advance_loc4 |
| 66 | .4byte .LCFI0-.LFB1 |
| 67 | .byte 0xe / DW_CFA_def_cfa_offset |
| 68 | .byte 0x8 / ULEB128 0x8 |
| 69 | .byte 0x85 / DW_CFA_offset, column 0x5 |
| 70 | .byte 0x2 / ULEB128 0x2 |
| 71 | .byte 0x4 / DW_CFA_advance_loc4 |
| 72 | .4byte .LCFI1-.LCFI0 |
| 73 | .byte 0xd / DW_CFA_def_cfa_register |
| 74 | .byte 0x5 / ULEB128 0x5 |
| 75 | .byte 0x4 / DW_CFA_advance_loc4 |
| 76 | .4byte .LCFI2-.LCFI1 |
| 77 | .byte 0x2e / DW_CFA_GNU_args_size |
| 78 | .byte 0x4 / ULEB128 0x4 |
| 79 | .byte 0x4 / DW_CFA_advance_loc4 |
| 80 | .4byte .LCFI3-.LCFI2 |
| 81 | .byte 0x2e / DW_CFA_GNU_args_size |
| 82 | .byte 0x0 / ULEB128 0x0 |
| 83 | .align 4 |
| 84 | .LEFDE1: |
| 85 | .set .LLFDE1,.LEFDE1-.LSFDE1 / FDE Length Symbol |
| 86 | |
| 87 | The immediate issue we can address in the assembler is the |
| 88 | DW_CFA_advance_loc4 followed by a four byte value. The value is |
| 89 | the difference of two addresses in the function. Since gcc does |
| 90 | not know this value, it always uses four bytes. We will know the |
| 91 | value at the end of assembly, so we can do better. */ |
| 92 | |
| 93 | struct cie_info |
| 94 | { |
| 95 | unsigned code_alignment; |
| 96 | int z_augmentation; |
| 97 | }; |
| 98 | |
| 99 | static int get_cie_info (struct cie_info *); |
| 100 | |
| 101 | /* Extract information from the CIE. */ |
| 102 | |
| 103 | static int |
| 104 | get_cie_info (struct cie_info *info) |
| 105 | { |
| 106 | fragS *f; |
| 107 | fixS *fix; |
| 108 | int offset; |
| 109 | char CIE_id; |
| 110 | char augmentation[10]; |
| 111 | int iaug; |
| 112 | int code_alignment = 0; |
| 113 | |
| 114 | /* We should find the CIE at the start of the section. */ |
| 115 | |
| 116 | f = seg_info (now_seg)->frchainP->frch_root; |
| 117 | fix = seg_info (now_seg)->frchainP->fix_root; |
| 118 | |
| 119 | /* Look through the frags of the section to find the code alignment. */ |
| 120 | |
| 121 | /* First make sure that the CIE Identifier Tag is 0/-1. */ |
| 122 | |
| 123 | if (strcmp (segment_name (now_seg), ".debug_frame") == 0) |
| 124 | CIE_id = (char)0xff; |
| 125 | else |
| 126 | CIE_id = 0; |
| 127 | |
| 128 | offset = 4; |
| 129 | while (f != NULL && offset >= f->fr_fix) |
| 130 | { |
| 131 | offset -= f->fr_fix; |
| 132 | f = f->fr_next; |
| 133 | } |
| 134 | if (f == NULL |
| 135 | || f->fr_fix - offset < 4 |
| 136 | || f->fr_literal[offset] != CIE_id |
| 137 | || f->fr_literal[offset + 1] != CIE_id |
| 138 | || f->fr_literal[offset + 2] != CIE_id |
| 139 | || f->fr_literal[offset + 3] != CIE_id) |
| 140 | return 0; |
| 141 | |
| 142 | /* Next make sure the CIE version number is 1. */ |
| 143 | |
| 144 | offset += 4; |
| 145 | while (f != NULL && offset >= f->fr_fix) |
| 146 | { |
| 147 | offset -= f->fr_fix; |
| 148 | f = f->fr_next; |
| 149 | } |
| 150 | if (f == NULL |
| 151 | || f->fr_fix - offset < 1 |
| 152 | || f->fr_literal[offset] != 1) |
| 153 | return 0; |
| 154 | |
| 155 | /* Skip the augmentation (a null terminated string). */ |
| 156 | |
| 157 | iaug = 0; |
| 158 | ++offset; |
| 159 | while (1) |
| 160 | { |
| 161 | while (f != NULL && offset >= f->fr_fix) |
| 162 | { |
| 163 | offset -= f->fr_fix; |
| 164 | f = f->fr_next; |
| 165 | } |
| 166 | if (f == NULL) |
| 167 | return 0; |
| 168 | |
| 169 | while (offset < f->fr_fix && f->fr_literal[offset] != '\0') |
| 170 | { |
| 171 | if ((size_t) iaug < (sizeof augmentation) - 1) |
| 172 | { |
| 173 | augmentation[iaug] = f->fr_literal[offset]; |
| 174 | ++iaug; |
| 175 | } |
| 176 | ++offset; |
| 177 | } |
| 178 | if (offset < f->fr_fix) |
| 179 | break; |
| 180 | } |
| 181 | ++offset; |
| 182 | while (f != NULL && offset >= f->fr_fix) |
| 183 | { |
| 184 | offset -= f->fr_fix; |
| 185 | f = f->fr_next; |
| 186 | } |
| 187 | if (f == NULL) |
| 188 | return 0; |
| 189 | |
| 190 | augmentation[iaug] = '\0'; |
| 191 | if (augmentation[0] == '\0') |
| 192 | { |
| 193 | /* No augmentation. */ |
| 194 | } |
| 195 | else if (strcmp (augmentation, "eh") == 0) |
| 196 | { |
| 197 | /* We have to skip a pointer. Unfortunately, we don't know how |
| 198 | large it is. We find out by looking for a matching fixup. */ |
| 199 | while (fix != NULL |
| 200 | && (fix->fx_frag != f || fix->fx_where != offset)) |
| 201 | fix = fix->fx_next; |
| 202 | if (fix == NULL) |
| 203 | offset += 4; |
| 204 | else |
| 205 | offset += fix->fx_size; |
| 206 | while (f != NULL && offset >= f->fr_fix) |
| 207 | { |
| 208 | offset -= f->fr_fix; |
| 209 | f = f->fr_next; |
| 210 | } |
| 211 | if (f == NULL) |
| 212 | return 0; |
| 213 | } |
| 214 | else if (augmentation[0] != 'z') |
| 215 | return 0; |
| 216 | |
| 217 | /* We're now at the code alignment factor, which is a ULEB128. If |
| 218 | it isn't a single byte, forget it. */ |
| 219 | |
| 220 | code_alignment = f->fr_literal[offset] & 0xff; |
| 221 | if ((code_alignment & 0x80) != 0) |
| 222 | code_alignment = 0; |
| 223 | |
| 224 | info->code_alignment = code_alignment; |
| 225 | info->z_augmentation = (augmentation[0] == 'z'); |
| 226 | |
| 227 | return 1; |
| 228 | } |
| 229 | |
| 230 | /* This function is called from emit_expr. It looks for cases which |
| 231 | we can optimize. |
| 232 | |
| 233 | Rather than try to parse all this information as we read it, we |
| 234 | look for a single byte DW_CFA_advance_loc4 followed by a 4 byte |
| 235 | difference. We turn that into a rs_cfa_advance frag, and handle |
| 236 | those frags at the end of the assembly. If the gcc output changes |
| 237 | somewhat, this optimization may stop working. |
| 238 | |
| 239 | This function returns non-zero if it handled the expression and |
| 240 | emit_expr should not do anything, or zero otherwise. It can also |
| 241 | change *EXP and *PNBYTES. */ |
| 242 | |
| 243 | int |
| 244 | check_eh_frame (expressionS *exp, unsigned int *pnbytes) |
| 245 | { |
| 246 | struct frame_data |
| 247 | { |
| 248 | enum frame_state |
| 249 | { |
| 250 | state_idle, |
| 251 | state_saw_size, |
| 252 | state_saw_cie_offset, |
| 253 | state_saw_pc_begin, |
| 254 | state_seeing_aug_size, |
| 255 | state_skipping_aug, |
| 256 | state_wait_loc4, |
| 257 | state_saw_loc4, |
| 258 | state_error, |
| 259 | } state; |
| 260 | |
| 261 | int cie_info_ok; |
| 262 | struct cie_info cie_info; |
| 263 | |
| 264 | symbolS *size_end_sym; |
| 265 | fragS *loc4_frag; |
| 266 | int loc4_fix; |
| 267 | |
| 268 | int aug_size; |
| 269 | int aug_shift; |
| 270 | }; |
| 271 | |
| 272 | static struct frame_data eh_frame_data; |
| 273 | static struct frame_data debug_frame_data; |
| 274 | struct frame_data *d; |
| 275 | |
| 276 | /* Don't optimize. */ |
| 277 | if (flag_traditional_format) |
| 278 | return 0; |
| 279 | |
| 280 | #ifdef md_allow_eh_opt |
| 281 | if (! md_allow_eh_opt) |
| 282 | return 0; |
| 283 | #endif |
| 284 | |
| 285 | /* Select the proper section data. */ |
| 286 | if (strcmp (segment_name (now_seg), ".eh_frame") == 0) |
| 287 | d = &eh_frame_data; |
| 288 | else if (strcmp (segment_name (now_seg), ".debug_frame") == 0) |
| 289 | d = &debug_frame_data; |
| 290 | else |
| 291 | return 0; |
| 292 | |
| 293 | if (d->state >= state_saw_size && S_IS_DEFINED (d->size_end_sym)) |
| 294 | { |
| 295 | /* We have come to the end of the CIE or FDE. See below where |
| 296 | we set saw_size. We must check this first because we may now |
| 297 | be looking at the next size. */ |
| 298 | d->state = state_idle; |
| 299 | } |
| 300 | |
| 301 | switch (d->state) |
| 302 | { |
| 303 | case state_idle: |
| 304 | if (*pnbytes == 4) |
| 305 | { |
| 306 | /* This might be the size of the CIE or FDE. We want to know |
| 307 | the size so that we don't accidentally optimize across an FDE |
| 308 | boundary. We recognize the size in one of two forms: a |
| 309 | symbol which will later be defined as a difference, or a |
| 310 | subtraction of two symbols. Either way, we can tell when we |
| 311 | are at the end of the FDE because the symbol becomes defined |
| 312 | (in the case of a subtraction, the end symbol, from which the |
| 313 | start symbol is being subtracted). Other ways of describing |
| 314 | the size will not be optimized. */ |
| 315 | if ((exp->X_op == O_symbol || exp->X_op == O_subtract) |
| 316 | && ! S_IS_DEFINED (exp->X_add_symbol)) |
| 317 | { |
| 318 | d->state = state_saw_size; |
| 319 | d->size_end_sym = exp->X_add_symbol; |
| 320 | } |
| 321 | } |
| 322 | break; |
| 323 | |
| 324 | case state_saw_size: |
| 325 | case state_saw_cie_offset: |
| 326 | /* Assume whatever form it appears in, it appears atomically. */ |
| 327 | d->state += 1; |
| 328 | break; |
| 329 | |
| 330 | case state_saw_pc_begin: |
| 331 | /* Decide whether we should see an augmentation. */ |
| 332 | if (! d->cie_info_ok |
| 333 | && ! (d->cie_info_ok = get_cie_info (&d->cie_info))) |
| 334 | d->state = state_error; |
| 335 | else if (d->cie_info.z_augmentation) |
| 336 | { |
| 337 | d->state = state_seeing_aug_size; |
| 338 | d->aug_size = 0; |
| 339 | d->aug_shift = 0; |
| 340 | } |
| 341 | else |
| 342 | d->state = state_wait_loc4; |
| 343 | break; |
| 344 | |
| 345 | case state_seeing_aug_size: |
| 346 | /* Bytes == -1 means this comes from an leb128 directive. */ |
| 347 | if ((int)*pnbytes == -1 && exp->X_op == O_constant) |
| 348 | { |
| 349 | d->aug_size = exp->X_add_number; |
| 350 | d->state = state_skipping_aug; |
| 351 | } |
| 352 | else if (*pnbytes == 1 && exp->X_op == O_constant) |
| 353 | { |
| 354 | unsigned char byte = exp->X_add_number; |
| 355 | d->aug_size |= (byte & 0x7f) << d->aug_shift; |
| 356 | d->aug_shift += 7; |
| 357 | if ((byte & 0x80) == 0) |
| 358 | d->state = state_skipping_aug; |
| 359 | } |
| 360 | else |
| 361 | d->state = state_error; |
| 362 | if (d->state == state_skipping_aug && d->aug_size == 0) |
| 363 | d->state = state_wait_loc4; |
| 364 | break; |
| 365 | |
| 366 | case state_skipping_aug: |
| 367 | if ((int)*pnbytes < 0) |
| 368 | d->state = state_error; |
| 369 | else |
| 370 | { |
| 371 | int left = (d->aug_size -= *pnbytes); |
| 372 | if (left == 0) |
| 373 | d->state = state_wait_loc4; |
| 374 | else if (left < 0) |
| 375 | d->state = state_error; |
| 376 | } |
| 377 | break; |
| 378 | |
| 379 | case state_wait_loc4: |
| 380 | if (*pnbytes == 1 |
| 381 | && exp->X_op == O_constant |
| 382 | && exp->X_add_number == DW_CFA_advance_loc4) |
| 383 | { |
| 384 | /* This might be a DW_CFA_advance_loc4. Record the frag and the |
| 385 | position within the frag, so that we can change it later. */ |
| 386 | frag_grow (1); |
| 387 | d->state = state_saw_loc4; |
| 388 | d->loc4_frag = frag_now; |
| 389 | d->loc4_fix = frag_now_fix (); |
| 390 | } |
| 391 | break; |
| 392 | |
| 393 | case state_saw_loc4: |
| 394 | d->state = state_wait_loc4; |
| 395 | if (*pnbytes != 4) |
| 396 | break; |
| 397 | if (exp->X_op == O_constant) |
| 398 | { |
| 399 | /* This is a case which we can optimize. The two symbols being |
| 400 | subtracted were in the same frag and the expression was |
| 401 | reduced to a constant. We can do the optimization entirely |
| 402 | in this function. */ |
| 403 | if (exp->X_add_number < 0x40) |
| 404 | { |
| 405 | d->loc4_frag->fr_literal[d->loc4_fix] |
| 406 | = DW_CFA_advance_loc | exp->X_add_number; |
| 407 | /* No more bytes needed. */ |
| 408 | return 1; |
| 409 | } |
| 410 | else if (exp->X_add_number < 0x100) |
| 411 | { |
| 412 | d->loc4_frag->fr_literal[d->loc4_fix] = DW_CFA_advance_loc1; |
| 413 | *pnbytes = 1; |
| 414 | } |
| 415 | else if (exp->X_add_number < 0x10000) |
| 416 | { |
| 417 | d->loc4_frag->fr_literal[d->loc4_fix] = DW_CFA_advance_loc2; |
| 418 | *pnbytes = 2; |
| 419 | } |
| 420 | } |
| 421 | else if (exp->X_op == O_subtract && d->cie_info.code_alignment == 1) |
| 422 | { |
| 423 | /* This is a case we can optimize. The expression was not |
| 424 | reduced, so we can not finish the optimization until the end |
| 425 | of the assembly. We set up a variant frag which we handle |
| 426 | later. */ |
| 427 | frag_var (rs_cfa, 4, 0, 1 << 3, make_expr_symbol (exp), |
| 428 | d->loc4_fix, (char *) d->loc4_frag); |
| 429 | return 1; |
| 430 | } |
| 431 | else if ((exp->X_op == O_divide |
| 432 | || exp->X_op == O_right_shift) |
| 433 | && d->cie_info.code_alignment > 1) |
| 434 | { |
| 435 | if (exp->X_add_symbol->bsym |
| 436 | && exp->X_op_symbol->bsym |
| 437 | && exp->X_add_symbol->sy_value.X_op == O_subtract |
| 438 | && exp->X_op_symbol->sy_value.X_op == O_constant |
| 439 | && ((exp->X_op == O_divide |
| 440 | ? exp->X_op_symbol->sy_value.X_add_number |
| 441 | : (offsetT) 1 << exp->X_op_symbol->sy_value.X_add_number) |
| 442 | == (offsetT) d->cie_info.code_alignment)) |
| 443 | { |
| 444 | /* This is a case we can optimize as well. The expression was |
| 445 | not reduced, so we can not finish the optimization until the |
| 446 | end of the assembly. We set up a variant frag which we |
| 447 | handle later. */ |
| 448 | frag_var (rs_cfa, 4, 0, d->cie_info.code_alignment << 3, |
| 449 | make_expr_symbol (&exp->X_add_symbol->sy_value), |
| 450 | d->loc4_fix, (char *) d->loc4_frag); |
| 451 | return 1; |
| 452 | } |
| 453 | } |
| 454 | break; |
| 455 | |
| 456 | case state_error: |
| 457 | /* Just skipping everything. */ |
| 458 | break; |
| 459 | } |
| 460 | |
| 461 | return 0; |
| 462 | } |
| 463 | |
| 464 | /* The function estimates the size of a rs_cfa variant frag based on |
| 465 | the current values of the symbols. It is called before the |
| 466 | relaxation loop. We set fr_subtype{0:2} to the expected length. */ |
| 467 | |
| 468 | int |
| 469 | eh_frame_estimate_size_before_relax (fragS *frag) |
| 470 | { |
| 471 | offsetT diff; |
| 472 | int ca = frag->fr_subtype >> 3; |
| 473 | int ret; |
| 474 | |
| 475 | diff = resolve_symbol_value (frag->fr_symbol); |
| 476 | |
| 477 | assert (ca > 0); |
| 478 | diff /= ca; |
| 479 | if (diff < 0x40) |
| 480 | ret = 0; |
| 481 | else if (diff < 0x100) |
| 482 | ret = 1; |
| 483 | else if (diff < 0x10000) |
| 484 | ret = 2; |
| 485 | else |
| 486 | ret = 4; |
| 487 | |
| 488 | frag->fr_subtype = (frag->fr_subtype & ~7) | ret; |
| 489 | |
| 490 | return ret; |
| 491 | } |
| 492 | |
| 493 | /* This function relaxes a rs_cfa variant frag based on the current |
| 494 | values of the symbols. fr_subtype{0:2} is the current length of |
| 495 | the frag. This returns the change in frag length. */ |
| 496 | |
| 497 | int |
| 498 | eh_frame_relax_frag (fragS *frag) |
| 499 | { |
| 500 | int oldsize, newsize; |
| 501 | |
| 502 | oldsize = frag->fr_subtype & 7; |
| 503 | newsize = eh_frame_estimate_size_before_relax (frag); |
| 504 | return newsize - oldsize; |
| 505 | } |
| 506 | |
| 507 | /* This function converts a rs_cfa variant frag into a normal fill |
| 508 | frag. This is called after all relaxation has been done. |
| 509 | fr_subtype{0:2} will be the desired length of the frag. */ |
| 510 | |
| 511 | void |
| 512 | eh_frame_convert_frag (fragS *frag) |
| 513 | { |
| 514 | offsetT diff; |
| 515 | fragS *loc4_frag; |
| 516 | int loc4_fix, ca; |
| 517 | |
| 518 | loc4_frag = (fragS *) frag->fr_opcode; |
| 519 | loc4_fix = (int) frag->fr_offset; |
| 520 | |
| 521 | diff = resolve_symbol_value (frag->fr_symbol); |
| 522 | |
| 523 | ca = frag->fr_subtype >> 3; |
| 524 | assert (ca > 0); |
| 525 | diff /= ca; |
| 526 | switch (frag->fr_subtype & 7) |
| 527 | { |
| 528 | case 0: |
| 529 | assert (diff < 0x40); |
| 530 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc | diff; |
| 531 | break; |
| 532 | |
| 533 | case 1: |
| 534 | assert (diff < 0x100); |
| 535 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc1; |
| 536 | frag->fr_literal[frag->fr_fix] = diff; |
| 537 | break; |
| 538 | |
| 539 | case 2: |
| 540 | assert (diff < 0x10000); |
| 541 | loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc2; |
| 542 | md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 2); |
| 543 | break; |
| 544 | |
| 545 | default: |
| 546 | md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 4); |
| 547 | break; |
| 548 | } |
| 549 | |
| 550 | frag->fr_fix += frag->fr_subtype & 7; |
| 551 | frag->fr_type = rs_fill; |
| 552 | frag->fr_subtype = 0; |
| 553 | frag->fr_offset = 0; |
| 554 | } |