gas/ehopt.c

   1 /* ehopt.c--optimize gcc exception frame information.
   2    Copyright (C) 1998-2023 Free Software Foundation, Inc.
   3    Written by Ian Lance Taylor <ian@cygnus.com>.
   4
   5    This file is part of GAS, the GNU Assembler.
   6
   7    GAS is free software; you can redistribute it and/or modify
   8    it under the terms of the GNU General Public License as published by
   9    the Free Software Foundation; either version 3, or (at your option)
  10    any later version.
  11
  12    GAS is distributed in the hope that it will be useful,
  13    but WITHOUT ANY WARRANTY; without even the implied warranty of
  14    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15    GNU General Public License for more details.
  16
  17    You should have received a copy of the GNU General Public License
  18    along with GAS; see the file COPYING.  If not, write to the Free
  19    Software Foundation, 51 Franklin Street - Fifth Floor, Boston, MA
  20    02110-1301, USA.  */
  21
  22 #include "as.h"
  23 #include "subsegs.h"
  24
  25 /* We include this ELF file, even though we may not be assembling for
  26    ELF, since the exception frame information is always in a format
  27    derived from DWARF.  */
  28
  29 #include "dwarf2.h"
  30
  31 /* Try to optimize gcc 2.8 exception frame information.
  32
  33    Exception frame information is emitted for every function in the
  34    .eh_frame or .debug_frame sections.  Simple information for a function
  35    with no exceptions looks like this:
  36
  37 __FRAME_BEGIN__:
  38         .4byte  .LLCIE1 / Length of Common Information Entry
  39 .LSCIE1:
  40 #if .eh_frame
  41         .4byte  0x0     / CIE Identifier Tag
  42 #elif .debug_frame
  43         .4byte  0xffffffff / CIE Identifier Tag
  44 #endif
  45         .byte   0x1     / CIE Version
  46         .byte   0x0     / CIE Augmentation (none)
  47         .byte   0x1     / ULEB128 0x1 (CIE Code Alignment Factor)
  48         .byte   0x7c    / SLEB128 -4 (CIE Data Alignment Factor)
  49         .byte   0x8     / CIE RA Column
  50         .byte   0xc     / DW_CFA_def_cfa
  51         .byte   0x4     / ULEB128 0x4
  52         .byte   0x4     / ULEB128 0x4
  53         .byte   0x88    / DW_CFA_offset, column 0x8
  54         .byte   0x1     / ULEB128 0x1
  55         .align 4
  56 .LECIE1:
  57         .set    .LLCIE1,.LECIE1-.LSCIE1 / CIE Length Symbol
  58         .4byte  .LLFDE1 / FDE Length
  59 .LSFDE1:
  60         .4byte  .LSFDE1-__FRAME_BEGIN__ / FDE CIE offset
  61         .4byte  .LFB1   / FDE initial location
  62         .4byte  .LFE1-.LFB1     / FDE address range
  63         .byte   0x4     / DW_CFA_advance_loc4
  64         .4byte  .LCFI0-.LFB1
  65         .byte   0xe     / DW_CFA_def_cfa_offset
  66         .byte   0x8     / ULEB128 0x8
  67         .byte   0x85    / DW_CFA_offset, column 0x5
  68         .byte   0x2     / ULEB128 0x2
  69         .byte   0x4     / DW_CFA_advance_loc4
  70         .4byte  .LCFI1-.LCFI0
  71         .byte   0xd     / DW_CFA_def_cfa_register
  72         .byte   0x5     / ULEB128 0x5
  73         .byte   0x4     / DW_CFA_advance_loc4
  74         .4byte  .LCFI2-.LCFI1
  75         .byte   0x2e    / DW_CFA_GNU_args_size
  76         .byte   0x4     / ULEB128 0x4
  77         .byte   0x4     / DW_CFA_advance_loc4
  78         .4byte  .LCFI3-.LCFI2
  79         .byte   0x2e    / DW_CFA_GNU_args_size
  80         .byte   0x0     / ULEB128 0x0
  81         .align 4
  82 .LEFDE1:
  83         .set    .LLFDE1,.LEFDE1-.LSFDE1 / FDE Length Symbol
  84
  85    The immediate issue we can address in the assembler is the
  86    DW_CFA_advance_loc4 followed by a four byte value.  The value is
  87    the difference of two addresses in the function.  Since gcc does
  88    not know this value, it always uses four bytes.  We will know the
  89    value at the end of assembly, so we can do better.  */
  90
  91 struct cie_info
  92 {
  93   unsigned code_alignment;
  94   int z_augmentation;
  95 };
  96
  97 static int get_cie_info (struct cie_info *);
  98
  99 /* Extract information from the CIE.  */
 100
 101 static int
 102 get_cie_info (struct cie_info *info)
 103 {
 104   fragS *f;
 105   fixS *fix;
 106   unsigned int offset;
 107   char CIE_id;
 108   char augmentation[10];
 109   int iaug;
 110   int code_alignment = 0;
 111
 112   /* We should find the CIE at the start of the section.  */
 113
 114   f = seg_info (now_seg)->frchainP->frch_root;
 115   fix = seg_info (now_seg)->frchainP->fix_root;
 116
 117   /* Look through the frags of the section to find the code alignment.  */
 118
 119   /* First make sure that the CIE Identifier Tag is 0/-1.  */
 120
 121   if (startswith (segment_name (now_seg), ".debug_frame"))
 122     CIE_id = (char)0xff;
 123   else
 124     CIE_id = 0;
 125
 126   offset = 4;
 127   while (f != NULL && offset >= f->fr_fix)
 128     {
 129       offset -= f->fr_fix;
 130       f = f->fr_next;
 131     }
 132   if (f == NULL
 133       || f->fr_fix - offset < 4
 134       || f->fr_literal[offset] != CIE_id
 135       || f->fr_literal[offset + 1] != CIE_id
 136       || f->fr_literal[offset + 2] != CIE_id
 137       || f->fr_literal[offset + 3] != CIE_id)
 138     return 0;
 139
 140   /* Next make sure the CIE version number is 1.  */
 141
 142   offset += 4;
 143   while (f != NULL && offset >= f->fr_fix)
 144     {
 145       offset -= f->fr_fix;
 146       f = f->fr_next;
 147     }
 148   if (f == NULL
 149       || f->fr_fix - offset < 1
 150       || f->fr_literal[offset] != 1)
 151     return 0;
 152
 153   /* Skip the augmentation (a null terminated string).  */
 154
 155   iaug = 0;
 156   ++offset;
 157   while (1)
 158     {
 159       while (f != NULL && offset >= f->fr_fix)
 160         {
 161           offset -= f->fr_fix;
 162           f = f->fr_next;
 163         }
 164       if (f == NULL)
 165         return 0;
 166
 167       while (offset < f->fr_fix && f->fr_literal[offset] != '\0')
 168         {
 169           if ((size_t) iaug < (sizeof augmentation) - 1)
 170             {
 171               augmentation[iaug] = f->fr_literal[offset];
 172               ++iaug;
 173             }
 174           ++offset;
 175         }
 176       if (offset < f->fr_fix)
 177         break;
 178     }
 179   ++offset;
 180   while (f != NULL && offset >= f->fr_fix)
 181     {
 182       offset -= f->fr_fix;
 183       f = f->fr_next;
 184     }
 185   if (f == NULL)
 186     return 0;
 187
 188   augmentation[iaug] = '\0';
 189   if (augmentation[0] == '\0')
 190     {
 191       /* No augmentation.  */
 192     }
 193   else if (strcmp (augmentation, "eh") == 0)
 194     {
 195       /* We have to skip a pointer.  Unfortunately, we don't know how
 196          large it is.  We find out by looking for a matching fixup.  */
 197       while (fix != NULL
 198              && (fix->fx_frag != f || fix->fx_where != offset))
 199         fix = fix->fx_next;
 200       if (fix == NULL)
 201         offset += 4;
 202       else
 203         offset += fix->fx_size;
 204       while (f != NULL && offset >= f->fr_fix)
 205         {
 206           offset -= f->fr_fix;
 207           f = f->fr_next;
 208         }
 209       if (f == NULL)
 210         return 0;
 211     }
 212   else if (augmentation[0] != 'z')
 213     return 0;
 214
 215   /* We're now at the code alignment factor, which is a ULEB128.  If
 216      it isn't a single byte, forget it.  */
 217
 218   code_alignment = f->fr_literal[offset] & 0xff;
 219   if ((code_alignment & 0x80) != 0)
 220     code_alignment = 0;
 221
 222   info->code_alignment = code_alignment;
 223   info->z_augmentation = (augmentation[0] == 'z');
 224
 225   return 1;
 226 }
 227
 228 enum frame_state
 229 {
 230   state_idle,
 231   state_saw_size,
 232   state_saw_cie_offset,
 233   state_saw_pc_begin,
 234   state_seeing_aug_size,
 235   state_skipping_aug,
 236   state_wait_loc4,
 237   state_saw_loc4,
 238   state_error,
 239 };
 240
 241 /* This function is called from emit_expr.  It looks for cases which
 242    we can optimize.
 243
 244    Rather than try to parse all this information as we read it, we
 245    look for a single byte DW_CFA_advance_loc4 followed by a 4 byte
 246    difference.  We turn that into a rs_cfa_advance frag, and handle
 247    those frags at the end of the assembly.  If the gcc output changes
 248    somewhat, this optimization may stop working.
 249
 250    This function returns non-zero if it handled the expression and
 251    emit_expr should not do anything, or zero otherwise.  It can also
 252    change *EXP and *PNBYTES.  */
 253
 254 int
 255 check_eh_frame (expressionS *exp, unsigned int *pnbytes)
 256 {
 257   struct frame_data
 258   {
 259     enum frame_state 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 (startswith (segment_name (now_seg), ".eh_frame")
 287       && segment_name (now_seg)[9] != '_')
 288     d = &eh_frame_data;
 289   else if (startswith (segment_name (now_seg), ".debug_frame"))
 290     d = &debug_frame_data;
 291   else
 292     return 0;
 293
 294   if (d->state >= state_saw_size && S_IS_DEFINED (d->size_end_sym))
 295     {
 296       /* We have come to the end of the CIE or FDE.  See below where
 297          we set saw_size.  We must check this first because we may now
 298          be looking at the next size.  */
 299       d->state = state_idle;
 300     }
 301
 302   switch (d->state)
 303     {
 304     case state_idle:
 305       if (*pnbytes == 4)
 306         {
 307           /* This might be the size of the CIE or FDE.  We want to know
 308              the size so that we don't accidentally optimize across an FDE
 309              boundary.  We recognize the size in one of two forms: a
 310              symbol which will later be defined as a difference, or a
 311              subtraction of two symbols.  Either way, we can tell when we
 312              are at the end of the FDE because the symbol becomes defined
 313              (in the case of a subtraction, the end symbol, from which the
 314              start symbol is being subtracted).  Other ways of describing
 315              the size will not be optimized.  */
 316           if ((exp->X_op == O_symbol || exp->X_op == O_subtract)
 317               && ! S_IS_DEFINED (exp->X_add_symbol))
 318             {
 319               d->state = state_saw_size;
 320               d->size_end_sym = exp->X_add_symbol;
 321             }
 322         }
 323       break;
 324
 325     case state_saw_size:
 326     case state_saw_cie_offset:
 327       /* Assume whatever form it appears in, it appears atomically.  */
 328       d->state = (enum frame_state) (d->state + 1);
 329       break;
 330
 331     case state_saw_pc_begin:
 332       /* Decide whether we should see an augmentation.  */
 333       if (! d->cie_info_ok
 334           && ! (d->cie_info_ok = get_cie_info (&d->cie_info)))
 335         d->state = state_error;
 336       else if (d->cie_info.z_augmentation)
 337         {
 338           d->state = state_seeing_aug_size;
 339           d->aug_size = 0;
 340           d->aug_shift = 0;
 341         }
 342       else
 343         d->state = state_wait_loc4;
 344       break;
 345
 346     case state_seeing_aug_size:
 347       /* Bytes == -1 means this comes from an leb128 directive.  */
 348       if ((int)*pnbytes == -1 && exp->X_op == O_constant)
 349         {
 350           d->aug_size = exp->X_add_number;
 351           d->state = state_skipping_aug;
 352         }
 353       else if (*pnbytes == 1 && exp->X_op == O_constant)
 354         {
 355           unsigned char byte = exp->X_add_number;
 356           d->aug_size |= (byte & 0x7f) << d->aug_shift;
 357           d->aug_shift += 7;
 358           if ((byte & 0x80) == 0)
 359             d->state = state_skipping_aug;
 360         }
 361       else
 362         d->state = state_error;
 363       if (d->state == state_skipping_aug && d->aug_size == 0)
 364         d->state = state_wait_loc4;
 365       break;
 366
 367     case state_skipping_aug:
 368       if ((int)*pnbytes < 0)
 369         d->state = state_error;
 370       else
 371         {
 372           int left = (d->aug_size -= *pnbytes);
 373           if (left == 0)
 374             d->state = state_wait_loc4;
 375           else if (left < 0)
 376             d->state = state_error;
 377         }
 378       break;
 379
 380     case state_wait_loc4:
 381       if (*pnbytes == 1
 382           && exp->X_op == O_constant
 383           && exp->X_add_number == DW_CFA_advance_loc4)
 384         {
 385           /* This might be a DW_CFA_advance_loc4.  Record the frag and the
 386              position within the frag, so that we can change it later.  */
 387           frag_grow (1);
 388           d->state = state_saw_loc4;
 389           d->loc4_frag = frag_now;
 390           d->loc4_fix = frag_now_fix ();
 391         }
 392       break;
 393
 394     case state_saw_loc4:
 395       d->state = state_wait_loc4;
 396       if (*pnbytes != 4)
 397         break;
 398       if (exp->X_op == O_constant)
 399         {
 400           /* This is a case which we can optimize.  The two symbols being
 401              subtracted were in the same frag and the expression was
 402              reduced to a constant.  We can do the optimization entirely
 403              in this function.  */
 404           if (exp->X_add_number < 0x40)
 405             {
 406               d->loc4_frag->fr_literal[d->loc4_fix]
 407                 = DW_CFA_advance_loc | exp->X_add_number;
 408               /* No more bytes needed.  */
 409               return 1;
 410             }
 411           else if (exp->X_add_number < 0x100)
 412             {
 413               d->loc4_frag->fr_literal[d->loc4_fix] = DW_CFA_advance_loc1;
 414               *pnbytes = 1;
 415             }
 416           else if (exp->X_add_number < 0x10000)
 417             {
 418               d->loc4_frag->fr_literal[d->loc4_fix] = DW_CFA_advance_loc2;
 419               *pnbytes = 2;
 420             }
 421         }
 422       else if (exp->X_op == O_subtract && d->cie_info.code_alignment == 1)
 423         {
 424           /* This is a case we can optimize.  The expression was not
 425              reduced, so we can not finish the optimization until the end
 426              of the assembly.  We set up a variant frag which we handle
 427              later.  */
 428           frag_var (rs_cfa, 4, 0, 1 << 3, make_expr_symbol (exp),
 429                     d->loc4_fix, (char *) d->loc4_frag);
 430           return 1;
 431         }
 432       else if ((exp->X_op == O_divide
 433                 || exp->X_op == O_right_shift)
 434                && d->cie_info.code_alignment > 1)
 435         {
 436           if (symbol_symbolS (exp->X_add_symbol)
 437               && symbol_constant_p (exp->X_op_symbol)
 438               && S_GET_SEGMENT (exp->X_op_symbol) == absolute_section
 439               && ((exp->X_op == O_divide
 440                    ? *symbol_X_add_number (exp->X_op_symbol)
 441                    : (offsetT) 1 << *symbol_X_add_number (exp->X_op_symbol))
 442                   == (offsetT) d->cie_info.code_alignment))
 443             {
 444               expressionS *symval;
 445
 446               symval = symbol_get_value_expression (exp->X_add_symbol);
 447               if (symval->X_op == O_subtract)
 448                 {
 449                   /* This is a case we can optimize as well.  The
 450                      expression was not reduced, so we can not finish
 451                      the optimization until the end of the assembly.
 452                      We set up a variant frag which we handle later.  */
 453                   frag_var (rs_cfa, 4, 0, d->cie_info.code_alignment << 3,
 454                             make_expr_symbol (symval),
 455                             d->loc4_fix, (char *) d->loc4_frag);
 456                   return 1;
 457                 }
 458             }
 459         }
 460       break;
 461
 462     case state_error:
 463       /* Just skipping everything.  */
 464       break;
 465     }
 466
 467   return 0;
 468 }
 469
 470 /* The function estimates the size of a rs_cfa variant frag based on
 471    the current values of the symbols.  It is called before the
 472    relaxation loop.  We set fr_subtype{0:2} to the expected length.  */
 473
 474 int
 475 eh_frame_estimate_size_before_relax (fragS *frag)
 476 {
 477   offsetT diff;
 478   int ca = frag->fr_subtype >> 3;
 479   int ret;
 480
 481   diff = resolve_symbol_value (frag->fr_symbol);
 482
 483   gas_assert (ca > 0);
 484   diff /= ca;
 485   if (diff == 0)
 486     ret = -1;
 487   else if (diff < 0x40)
 488     ret = 0;
 489   else if (diff < 0x100)
 490     ret = 1;
 491   else if (diff < 0x10000)
 492     ret = 2;
 493   else
 494     ret = 4;
 495
 496   frag->fr_subtype = (frag->fr_subtype & ~7) | (ret & 7);
 497
 498   return ret;
 499 }
 500
 501 /* This function relaxes a rs_cfa variant frag based on the current
 502    values of the symbols.  fr_subtype{0:2} is the current length of
 503    the frag.  This returns the change in frag length.  */
 504
 505 int
 506 eh_frame_relax_frag (fragS *frag)
 507 {
 508   int oldsize, newsize;
 509
 510   oldsize = frag->fr_subtype & 7;
 511   if (oldsize == 7)
 512     oldsize = -1;
 513   newsize = eh_frame_estimate_size_before_relax (frag);
 514   return newsize - oldsize;
 515 }
 516
 517 /* This function converts a rs_cfa variant frag into a normal fill
 518    frag.  This is called after all relaxation has been done.
 519    fr_subtype{0:2} will be the desired length of the frag.  */
 520
 521 void
 522 eh_frame_convert_frag (fragS *frag)
 523 {
 524   offsetT diff;
 525   fragS *loc4_frag;
 526   int loc4_fix, ca;
 527
 528   loc4_frag = (fragS *) frag->fr_opcode;
 529   loc4_fix = (int) frag->fr_offset;
 530
 531   diff = resolve_symbol_value (frag->fr_symbol);
 532
 533   ca = frag->fr_subtype >> 3;
 534   gas_assert (ca > 0);
 535   diff /= ca;
 536   switch (frag->fr_subtype & 7)
 537     {
 538     case 0:
 539       gas_assert (diff < 0x40);
 540       loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc | diff;
 541       break;
 542
 543     case 1:
 544       gas_assert (diff < 0x100);
 545       loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc1;
 546       frag->fr_literal[frag->fr_fix] = diff;
 547       break;
 548
 549     case 2:
 550       gas_assert (diff < 0x10000);
 551       loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc2;
 552       md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 2);
 553       break;
 554
 555     case 4:
 556       md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 4);
 557       break;
 558
 559     case 7:
 560       gas_assert (diff == 0);
 561       frag->fr_fix -= 8;
 562       break;
 563
 564     default:
 565       abort ();
 566     }
 567
 568   frag->fr_fix += frag->fr_subtype & 7;
 569   frag->fr_type = rs_fill;
 570   frag->fr_subtype = 0;
 571   frag->fr_offset = 0;
 572 }