gas/ehopt.c

   1 /* ehopt.c--optimize gcc exception frame information.
   2    Copyright 1998, 2000, 2001, 2003 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 2, 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, 59 Temple Place - Suite 330, Boston, MA
  20 02111-1307, 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 "elf/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   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 #if defined (BFD_ASSEMBLER) || defined (MANY_SEGMENTS)
 115   f = seg_info (now_seg)->frchainP->frch_root;
 116 #else
 117   f = frchain_now->frch_root;
 118 #endif
 119 #ifdef BFD_ASSEMBLER
 120   fix = seg_info (now_seg)->frchainP->fix_root;
 121 #else
 122   fix = *seg_fix_rootP;
 123 #endif
 124
 125   /* Look through the frags of the section to find the code alignment.  */
 126
 127   /* First make sure that the CIE Identifier Tag is 0/-1.  */
 128
 129   if (strcmp (segment_name (now_seg), ".debug_frame") == 0)
 130     CIE_id = (char)0xff;
 131   else
 132     CIE_id = 0;
 133
 134   offset = 4;
 135   while (f != NULL && offset >= f->fr_fix)
 136     {
 137       offset -= f->fr_fix;
 138       f = f->fr_next;
 139     }
 140   if (f == NULL
 141       || f->fr_fix - offset < 4
 142       || f->fr_literal[offset] != CIE_id
 143       || f->fr_literal[offset + 1] != CIE_id
 144       || f->fr_literal[offset + 2] != CIE_id
 145       || f->fr_literal[offset + 3] != CIE_id)
 146     return 0;
 147
 148   /* Next make sure the CIE version number is 1.  */
 149
 150   offset += 4;
 151   while (f != NULL && offset >= f->fr_fix)
 152     {
 153       offset -= f->fr_fix;
 154       f = f->fr_next;
 155     }
 156   if (f == NULL
 157       || f->fr_fix - offset < 1
 158       || f->fr_literal[offset] != 1)
 159     return 0;
 160
 161   /* Skip the augmentation (a null terminated string).  */
 162
 163   iaug = 0;
 164   ++offset;
 165   while (1)
 166     {
 167       while (f != NULL && offset >= f->fr_fix)
 168         {
 169           offset -= f->fr_fix;
 170           f = f->fr_next;
 171         }
 172       if (f == NULL)
 173         return 0;
 174
 175       while (offset < f->fr_fix && f->fr_literal[offset] != '\0')
 176         {
 177           if ((size_t) iaug < (sizeof augmentation) - 1)
 178             {
 179               augmentation[iaug] = f->fr_literal[offset];
 180               ++iaug;
 181             }
 182           ++offset;
 183         }
 184       if (offset < f->fr_fix)
 185         break;
 186     }
 187   ++offset;
 188   while (f != NULL && offset >= f->fr_fix)
 189     {
 190       offset -= f->fr_fix;
 191       f = f->fr_next;
 192     }
 193   if (f == NULL)
 194     return 0;
 195
 196   augmentation[iaug] = '\0';
 197   if (augmentation[0] == '\0')
 198     {
 199       /* No augmentation.  */
 200     }
 201   else if (strcmp (augmentation, "eh") == 0)
 202     {
 203       /* We have to skip a pointer.  Unfortunately, we don't know how
 204          large it is.  We find out by looking for a matching fixup.  */
 205       while (fix != NULL
 206              && (fix->fx_frag != f || fix->fx_where != offset))
 207         fix = fix->fx_next;
 208       if (fix == NULL)
 209         offset += 4;
 210       else
 211         offset += fix->fx_size;
 212       while (f != NULL && offset >= f->fr_fix)
 213         {
 214           offset -= f->fr_fix;
 215           f = f->fr_next;
 216         }
 217       if (f == NULL)
 218         return 0;
 219     }
 220   else if (augmentation[0] != 'z')
 221     return 0;
 222
 223   /* We're now at the code alignment factor, which is a ULEB128.  If
 224      it isn't a single byte, forget it.  */
 225
 226   code_alignment = f->fr_literal[offset] & 0xff;
 227   if ((code_alignment & 0x80) != 0)
 228     code_alignment = 0;
 229
 230   info->code_alignment = code_alignment;
 231   info->z_augmentation = (augmentation[0] == 'z');
 232
 233   return 1;
 234 }
 235
 236 /* This function is called from emit_expr.  It looks for cases which
 237    we can optimize.
 238
 239    Rather than try to parse all this information as we read it, we
 240    look for a single byte DW_CFA_advance_loc4 followed by a 4 byte
 241    difference.  We turn that into a rs_cfa_advance frag, and handle
 242    those frags at the end of the assembly.  If the gcc output changes
 243    somewhat, this optimization may stop working.
 244
 245    This function returns non-zero if it handled the expression and
 246    emit_expr should not do anything, or zero otherwise.  It can also
 247    change *EXP and *PNBYTES.  */
 248
 249 int
 250 check_eh_frame (expressionS *exp, unsigned int *pnbytes)
 251 {
 252   struct frame_data
 253   {
 254     enum frame_state
 255     {
 256       state_idle,
 257       state_saw_size,
 258       state_saw_cie_offset,
 259       state_saw_pc_begin,
 260       state_seeing_aug_size,
 261       state_skipping_aug,
 262       state_wait_loc4,
 263       state_saw_loc4,
 264       state_error,
 265     } state;
 266
 267     int cie_info_ok;
 268     struct cie_info cie_info;
 269
 270     symbolS *size_end_sym;
 271     fragS *loc4_frag;
 272     int loc4_fix;
 273
 274     int aug_size;
 275     int aug_shift;
 276   };
 277
 278   static struct frame_data eh_frame_data;
 279   static struct frame_data debug_frame_data;
 280   struct frame_data *d;
 281
 282   /* Don't optimize.  */
 283   if (flag_traditional_format)
 284     return 0;
 285
 286   /* Select the proper section data.  */
 287   if (strcmp (segment_name (now_seg), ".eh_frame") == 0)
 288     d = &eh_frame_data;
 289   else if (strcmp (segment_name (now_seg), ".debug_frame") == 0)
 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 += 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 (d->cie_info.code_alignment > 0
 405               && exp->X_add_number % d->cie_info.code_alignment == 0
 406               && exp->X_add_number / d->cie_info.code_alignment < 0x40)
 407             {
 408               d->loc4_frag->fr_literal[d->loc4_fix]
 409                 = DW_CFA_advance_loc
 410                   | (exp->X_add_number / d->cie_info.code_alignment);
 411               /* No more bytes needed.  */
 412               return 1;
 413             }
 414           else if (exp->X_add_number < 0x100)
 415             {
 416               d->loc4_frag->fr_literal[d->loc4_fix] = DW_CFA_advance_loc1;
 417               *pnbytes = 1;
 418             }
 419           else if (exp->X_add_number < 0x10000)
 420             {
 421               d->loc4_frag->fr_literal[d->loc4_fix] = DW_CFA_advance_loc2;
 422               *pnbytes = 2;
 423             }
 424         }
 425       else if (exp->X_op == O_subtract)
 426         {
 427           /* This is a case we can optimize.  The expression was not
 428              reduced, so we can not finish the optimization until the end
 429              of the assembly.  We set up a variant frag which we handle
 430              later.  */
 431           int fr_subtype;
 432
 433           if (d->cie_info.code_alignment > 0)
 434             fr_subtype = d->cie_info.code_alignment << 3;
 435           else
 436             fr_subtype = 0;
 437
 438           frag_var (rs_cfa, 4, 0, fr_subtype, make_expr_symbol (exp),
 439                     d->loc4_fix, (char *) d->loc4_frag);
 440           return 1;
 441         }
 442       break;
 443
 444     case state_error:
 445       /* Just skipping everything.  */
 446       break;
 447     }
 448
 449   return 0;
 450 }
 451
 452 /* The function estimates the size of a rs_cfa variant frag based on
 453    the current values of the symbols.  It is called before the
 454    relaxation loop.  We set fr_subtype{0:2} to the expected length.  */
 455
 456 int
 457 eh_frame_estimate_size_before_relax (fragS *frag)
 458 {
 459   offsetT diff;
 460   int ca = frag->fr_subtype >> 3;
 461   int ret;
 462
 463   diff = resolve_symbol_value (frag->fr_symbol);
 464
 465   if (ca > 0 && diff % ca == 0 && diff / ca < 0x40)
 466     ret = 0;
 467   else if (diff < 0x100)
 468     ret = 1;
 469   else if (diff < 0x10000)
 470     ret = 2;
 471   else
 472     ret = 4;
 473
 474   frag->fr_subtype = (frag->fr_subtype & ~7) | ret;
 475
 476   return ret;
 477 }
 478
 479 /* This function relaxes a rs_cfa variant frag based on the current
 480    values of the symbols.  fr_subtype{0:2} is the current length of
 481    the frag.  This returns the change in frag length.  */
 482
 483 int
 484 eh_frame_relax_frag (fragS *frag)
 485 {
 486   int oldsize, newsize;
 487
 488   oldsize = frag->fr_subtype & 7;
 489   newsize = eh_frame_estimate_size_before_relax (frag);
 490   return newsize - oldsize;
 491 }
 492
 493 /* This function converts a rs_cfa variant frag into a normal fill
 494    frag.  This is called after all relaxation has been done.
 495    fr_subtype{0:2} will be the desired length of the frag.  */
 496
 497 void
 498 eh_frame_convert_frag (fragS *frag)
 499 {
 500   offsetT diff;
 501   fragS *loc4_frag;
 502   int loc4_fix;
 503
 504   loc4_frag = (fragS *) frag->fr_opcode;
 505   loc4_fix = (int) frag->fr_offset;
 506
 507   diff = resolve_symbol_value (frag->fr_symbol);
 508
 509   switch (frag->fr_subtype & 7)
 510     {
 511     case 0:
 512       {
 513         int ca = frag->fr_subtype >> 3;
 514         assert (ca > 0 && diff % ca == 0 && diff / ca < 0x40);
 515         loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc | (diff / ca);
 516       }
 517       break;
 518
 519     case 1:
 520       assert (diff < 0x100);
 521       loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc1;
 522       frag->fr_literal[frag->fr_fix] = diff;
 523       break;
 524
 525     case 2:
 526       assert (diff < 0x10000);
 527       loc4_frag->fr_literal[loc4_fix] = DW_CFA_advance_loc2;
 528       md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 2);
 529       break;
 530
 531     default:
 532       md_number_to_chars (frag->fr_literal + frag->fr_fix, diff, 4);
 533       break;
 534     }
 535
 536   frag->fr_fix += frag->fr_subtype & 7;
 537   frag->fr_type = rs_fill;
 538   frag->fr_subtype = 0;
 539   frag->fr_offset = 0;
 540 }