src/buildvm_peobj.c

   1 /*
   2 ** LuaJIT VM builder: PE object emitter.
   3 ** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
   4 **
   5 ** Only used for building on Windows, since we cannot assume the presence
   6 ** of a suitable assembler. The host and target byte order must match.
   7 */
   8
   9 #include "buildvm.h"
  10 #include "lj_bc.h"
  11
  12 #if LJ_TARGET_X86ORX64
  13
  14 /* Context for PE object emitter. */
  15 static char *strtab;
  16 static size_t strtabofs;
  17
  18 /* -- PE object definitions ----------------------------------------------- */
  19
  20 /* PE header. */
  21 typedef struct PEheader {
  22   uint16_t arch;
  23   uint16_t nsects;
  24   uint32_t time;
  25   uint32_t symtabofs;
  26   uint32_t nsyms;
  27   uint16_t opthdrsz;
  28   uint16_t flags;
  29 } PEheader;
  30
  31 /* PE section. */
  32 typedef struct PEsection {
  33   char name[8];
  34   uint32_t vsize;
  35   uint32_t vaddr;
  36   uint32_t size;
  37   uint32_t ofs;
  38   uint32_t relocofs;
  39   uint32_t lineofs;
  40   uint16_t nreloc;
  41   uint16_t nline;
  42   uint32_t flags;
  43 } PEsection;
  44
  45 /* PE relocation. */
  46 typedef struct PEreloc {
  47   uint32_t vaddr;
  48   uint32_t symidx;
  49   uint16_t type;
  50 } PEreloc;
  51
  52 /* Cannot use sizeof, because it pads up to the max. alignment. */
  53 #define PEOBJ_RELOC_SIZE        (4+4+2)
  54
  55 /* PE symbol table entry. */
  56 typedef struct PEsym {
  57   union {
  58     char name[8];
  59     uint32_t nameref[2];
  60   } n;
  61   uint32_t value;
  62   int16_t sect;
  63   uint16_t type;
  64   uint8_t scl;
  65   uint8_t naux;
  66 } PEsym;
  67
  68 /* PE symbol table auxiliary entry for a section. */
  69 typedef struct PEsymaux {
  70   uint32_t size;
  71   uint16_t nreloc;
  72   uint16_t nline;
  73   uint32_t cksum;
  74   uint16_t assoc;
  75   uint8_t comdatsel;
  76   uint8_t unused[3];
  77 } PEsymaux;
  78
  79 /* Cannot use sizeof, because it pads up to the max. alignment. */
  80 #define PEOBJ_SYM_SIZE  (8+4+2+2+1+1)
  81
  82 /* PE object CPU specific defines. */
  83 #if LJ_TARGET_X86
  84 #define PEOBJ_ARCH_TARGET       0x014c
  85 #define PEOBJ_RELOC_REL32       0x14  /* MS: REL32, GNU: DISP32. */
  86 #define PEOBJ_RELOC_DIR32       0x06
  87 #elif LJ_TARGET_X64
  88 #define PEOBJ_ARCH_TARGET       0x8664
  89 #define PEOBJ_RELOC_REL32       0x04  /* MS: REL32, GNU: DISP32. */
  90 #define PEOBJ_RELOC_DIR32       0x02
  91 #define PEOBJ_RELOC_ADDR32NB    0x03
  92 #endif
  93
  94 /* Section numbers (0-based). */
  95 enum {
  96   PEOBJ_SECT_ABS = -2,
  97   PEOBJ_SECT_UNDEF = -1,
  98   PEOBJ_SECT_TEXT,
  99 #if LJ_TARGET_X64
 100   PEOBJ_SECT_PDATA,
 101   PEOBJ_SECT_XDATA,
 102 #endif
 103   PEOBJ_SECT_RDATA_Z,
 104   PEOBJ_NSECTIONS
 105 };
 106
 107 /* Symbol types. */
 108 #define PEOBJ_TYPE_NULL         0
 109 #define PEOBJ_TYPE_FUNC         0x20
 110
 111 /* Symbol storage class. */
 112 #define PEOBJ_SCL_EXTERN        2
 113 #define PEOBJ_SCL_STATIC        3
 114
 115 /* -- PE object emitter --------------------------------------------------- */
 116
 117 /* Emit PE object symbol. */
 118 static void emit_peobj_sym(BuildCtx *ctx, const char *name, uint32_t value,
 119                            int sect, int type, int scl)
 120 {
 121   PEsym sym;
 122   size_t len = strlen(name);
 123   if (!strtab) {  /* Pass 1: only calculate string table length. */
 124     if (len > 8) strtabofs += len+1;
 125     return;
 126   }
 127   if (len <= 8) {
 128     memcpy(sym.n.name, name, len);
 129     memset(sym.n.name+len, 0, 8-len);
 130   } else {
 131     sym.n.nameref[0] = 0;
 132     sym.n.nameref[1] = (uint32_t)strtabofs;
 133     memcpy(strtab + strtabofs, name, len);
 134     strtab[strtabofs+len] = 0;
 135     strtabofs += len+1;
 136   }
 137   sym.value = value;
 138   sym.sect = (int16_t)(sect+1);  /* 1-based section number. */
 139   sym.type = (uint16_t)type;
 140   sym.scl = (uint8_t)scl;
 141   sym.naux = 0;
 142   owrite(ctx, &sym, PEOBJ_SYM_SIZE);
 143 }
 144
 145 /* Emit PE object section symbol. */
 146 static void emit_peobj_sym_sect(BuildCtx *ctx, PEsection *pesect, int sect)
 147 {
 148   PEsym sym;
 149   PEsymaux aux;
 150   if (!strtab) return;  /* Pass 1: no output. */
 151   memcpy(sym.n.name, pesect[sect].name, 8);
 152   sym.value = 0;
 153   sym.sect = (int16_t)(sect+1);  /* 1-based section number. */
 154   sym.type = PEOBJ_TYPE_NULL;
 155   sym.scl = PEOBJ_SCL_STATIC;
 156   sym.naux = 1;
 157   owrite(ctx, &sym, PEOBJ_SYM_SIZE);
 158   memset(&aux, 0, sizeof(PEsymaux));
 159   aux.size = pesect[sect].size;
 160   aux.nreloc = pesect[sect].nreloc;
 161   owrite(ctx, &aux, PEOBJ_SYM_SIZE);
 162 }
 163
 164 /* Emit Windows PE object file. */
 165 void emit_peobj(BuildCtx *ctx)
 166 {
 167   PEheader pehdr;
 168   PEsection pesect[PEOBJ_NSECTIONS];
 169   uint32_t sofs;
 170   int i, nrsym;
 171   union { uint8_t b; uint32_t u; } host_endian;
 172
 173   host_endian.u = 1;
 174   if (host_endian.b != LJ_ENDIAN_SELECT(1, 0)) {
 175     fprintf(stderr, "Error: different byte order for host and target\n");
 176     exit(1);
 177   }
 178
 179   sofs = sizeof(PEheader) + PEOBJ_NSECTIONS*sizeof(PEsection);
 180
 181   /* Fill in PE sections. */
 182   memset(&pesect, 0, PEOBJ_NSECTIONS*sizeof(PEsection));
 183   memcpy(pesect[PEOBJ_SECT_TEXT].name, ".text", sizeof(".text")-1);
 184   pesect[PEOBJ_SECT_TEXT].ofs = sofs;
 185   sofs += (pesect[PEOBJ_SECT_TEXT].size = (uint32_t)ctx->codesz);
 186   pesect[PEOBJ_SECT_TEXT].relocofs = sofs;
 187   sofs += (pesect[PEOBJ_SECT_TEXT].nreloc = (uint16_t)ctx->nreloc) * PEOBJ_RELOC_SIZE;
 188   /* Flags: 60 = read+execute, 50 = align16, 20 = code. */
 189   pesect[PEOBJ_SECT_TEXT].flags = 0x60500020;
 190
 191 #if LJ_TARGET_X64
 192   memcpy(pesect[PEOBJ_SECT_PDATA].name, ".pdata", sizeof(".pdata")-1);
 193   pesect[PEOBJ_SECT_PDATA].ofs = sofs;
 194   sofs += (pesect[PEOBJ_SECT_PDATA].size = 6*4);
 195   pesect[PEOBJ_SECT_PDATA].relocofs = sofs;
 196   sofs += (pesect[PEOBJ_SECT_PDATA].nreloc = 6) * PEOBJ_RELOC_SIZE;
 197   /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
 198   pesect[PEOBJ_SECT_PDATA].flags = 0x40300040;
 199
 200   memcpy(pesect[PEOBJ_SECT_XDATA].name, ".xdata", sizeof(".xdata")-1);
 201   pesect[PEOBJ_SECT_XDATA].ofs = sofs;
 202   sofs += (pesect[PEOBJ_SECT_XDATA].size = 8*2+4+6*2);  /* See below. */
 203   pesect[PEOBJ_SECT_XDATA].relocofs = sofs;
 204   sofs += (pesect[PEOBJ_SECT_XDATA].nreloc = 1) * PEOBJ_RELOC_SIZE;
 205   /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
 206   pesect[PEOBJ_SECT_XDATA].flags = 0x40300040;
 207 #endif
 208
 209   memcpy(pesect[PEOBJ_SECT_RDATA_Z].name, ".rdata$Z", sizeof(".rdata$Z")-1);
 210   pesect[PEOBJ_SECT_RDATA_Z].ofs = sofs;
 211   sofs += (pesect[PEOBJ_SECT_RDATA_Z].size = (uint32_t)strlen(ctx->dasm_ident)+1);
 212   /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
 213   pesect[PEOBJ_SECT_RDATA_Z].flags = 0x40300040;
 214
 215   /* Fill in PE header. */
 216   pehdr.arch = PEOBJ_ARCH_TARGET;
 217   pehdr.nsects = PEOBJ_NSECTIONS;
 218   pehdr.time = 0;  /* Timestamp is optional. */
 219   pehdr.symtabofs = sofs;
 220   pehdr.opthdrsz = 0;
 221   pehdr.flags = 0;
 222
 223   /* Compute the size of the symbol table:
 224   ** @feat.00 + nsections*2
 225   ** + asm_start + nsym
 226   ** + nrsym
 227   */
 228   nrsym = ctx->nrelocsym;
 229   pehdr.nsyms = 1+PEOBJ_NSECTIONS*2 + 1+ctx->nsym + nrsym;
 230 #if LJ_TARGET_X64
 231   pehdr.nsyms += 1;  /* Symbol for lj_err_unwind_win64. */
 232 #endif
 233
 234   /* Write PE object header and all sections. */
 235   owrite(ctx, &pehdr, sizeof(PEheader));
 236   owrite(ctx, &pesect, sizeof(PEsection)*PEOBJ_NSECTIONS);
 237
 238   /* Write .text section. */
 239   owrite(ctx, ctx->code, ctx->codesz);
 240   for (i = 0; i < ctx->nreloc; i++) {
 241     PEreloc reloc;
 242     reloc.vaddr = (uint32_t)ctx->reloc[i].ofs;
 243     reloc.symidx = 1+2+ctx->reloc[i].sym;  /* Reloc syms are after .text sym. */
 244     reloc.type = ctx->reloc[i].type ? PEOBJ_RELOC_REL32 : PEOBJ_RELOC_DIR32;
 245     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 246   }
 247
 248 #if LJ_TARGET_X64
 249   { /* Write .pdata section. */
 250     uint32_t fcofs = (uint32_t)ctx->sym[ctx->nsym-1].ofs;
 251     uint32_t pdata[3];  /* Start of .text, end of .text and .xdata. */
 252     PEreloc reloc;
 253     pdata[0] = 0; pdata[1] = fcofs; pdata[2] = 0;
 254     owrite(ctx, &pdata, sizeof(pdata));
 255     pdata[0] = fcofs; pdata[1] = (uint32_t)ctx->codesz; pdata[2] = 20;
 256     owrite(ctx, &pdata, sizeof(pdata));
 257     reloc.vaddr = 0; reloc.symidx = 1+2+nrsym+2+2+1;
 258     reloc.type = PEOBJ_RELOC_ADDR32NB;
 259     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 260     reloc.vaddr = 4; reloc.symidx = 1+2+nrsym+2+2+1;
 261     reloc.type = PEOBJ_RELOC_ADDR32NB;
 262     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 263     reloc.vaddr = 8; reloc.symidx = 1+2+nrsym+2;
 264     reloc.type = PEOBJ_RELOC_ADDR32NB;
 265     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 266     reloc.vaddr = 12; reloc.symidx = 1+2+nrsym+2+2+1;
 267     reloc.type = PEOBJ_RELOC_ADDR32NB;
 268     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 269     reloc.vaddr = 16; reloc.symidx = 1+2+nrsym+2+2+1;
 270     reloc.type = PEOBJ_RELOC_ADDR32NB;
 271     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 272     reloc.vaddr = 20; reloc.symidx = 1+2+nrsym+2;
 273     reloc.type = PEOBJ_RELOC_ADDR32NB;
 274     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 275   }
 276   { /* Write .xdata section. */
 277     uint16_t xdata[8+2+6];
 278     PEreloc reloc;
 279     xdata[0] = 0x01|0x08|0x10;  /* Ver. 1, uhandler/ehandler, prolog size 0. */
 280     xdata[1] = 0x0005;  /* Number of unwind codes, no frame pointer. */
 281     xdata[2] = 0x4200;  /* Stack offset 4*8+8 = aword*5. */
 282     xdata[3] = 0x3000;  /* Push rbx. */
 283     xdata[4] = 0x6000;  /* Push rsi. */
 284     xdata[5] = 0x7000;  /* Push rdi. */
 285     xdata[6] = 0x5000;  /* Push rbp. */
 286     xdata[7] = 0;  /* Alignment. */
 287     xdata[8] = xdata[9] = 0;  /* Relocated address of exception handler. */
 288     xdata[10] = 0x01;  /* Ver. 1, no handler, prolog size 0. */
 289     xdata[11] = 0x1504;  /* Number of unwind codes, fp = rbp, fpofs = 16. */
 290     xdata[12] = 0x0300;  /* set_fpreg. */
 291     xdata[13] = 0x0200;  /* stack offset 0*8+8 = aword*1. */
 292     xdata[14] = 0x3000;  /* Push rbx. */
 293     xdata[15] = 0x5000;  /* Push rbp. */
 294     owrite(ctx, &xdata, sizeof(xdata));
 295     reloc.vaddr = 2*8; reloc.symidx = 1+2+nrsym+2+2;
 296     reloc.type = PEOBJ_RELOC_ADDR32NB;
 297     owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
 298   }
 299 #endif
 300
 301   /* Write .rdata$Z section. */
 302   owrite(ctx, ctx->dasm_ident, strlen(ctx->dasm_ident)+1);
 303
 304   /* Write symbol table. */
 305   strtab = NULL;  /* 1st pass: collect string sizes. */
 306   for (;;) {
 307     strtabofs = 4;
 308     /* Mark as SafeSEH compliant. */
 309     emit_peobj_sym(ctx, "@feat.00", 1,
 310                    PEOBJ_SECT_ABS, PEOBJ_TYPE_NULL, PEOBJ_SCL_STATIC);
 311
 312     emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_TEXT);
 313     for (i = 0; i < nrsym; i++)
 314       emit_peobj_sym(ctx, ctx->relocsym[i], 0,
 315                      PEOBJ_SECT_UNDEF, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
 316
 317 #if LJ_TARGET_X64
 318     emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_PDATA);
 319     emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_XDATA);
 320     emit_peobj_sym(ctx, "lj_err_unwind_win64", 0,
 321                    PEOBJ_SECT_UNDEF, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
 322 #endif
 323
 324     emit_peobj_sym(ctx, ctx->beginsym, 0,
 325                    PEOBJ_SECT_TEXT, PEOBJ_TYPE_NULL, PEOBJ_SCL_EXTERN);
 326     for (i = 0; i < ctx->nsym; i++)
 327       emit_peobj_sym(ctx, ctx->sym[i].name, (uint32_t)ctx->sym[i].ofs,
 328                      PEOBJ_SECT_TEXT, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
 329
 330     emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_RDATA_Z);
 331
 332     if (strtab)
 333       break;
 334     /* 2nd pass: alloc strtab, write syms and copy strings. */
 335     strtab = (char *)malloc(strtabofs);
 336     *(uint32_t *)strtab = (uint32_t)strtabofs;
 337   }
 338
 339   /* Write string table. */
 340   owrite(ctx, strtab, strtabofs);
 341 }
 342
 343 #else
 344
 345 void emit_peobj(BuildCtx *ctx)
 346 {
 347   UNUSED(ctx);
 348   fprintf(stderr, "Error: no PE object support for this target\n");
 349   exit(1);
 350 }
 351
 352 #endif