Reorganize build process.

Drop pre-translated buildvm_*.h.
Rename buildvm_*.dasc to vm_*.dasc.
Move buildvm* to host directory.
Build minilua, unless HOST_LUA is set.
Use HOST_LUA to run DynASM.
Translate only vm_*.dasc for target architecture.

7 files changed, 1852 insertions, 0 deletions

diff --git a/src/host/.gitignore b/src/host/.gitignore
new file mode 100644
index 00000000..762ac2a0
--- /dev/null
+++ b/src/host/.gitignore
@@ -0,0 +1,3 @@
+minilua
+buildvm
+buildvm_arch.h
diff --git a/src/host/buildvm.c b/src/host/buildvm.c
new file mode 100644
index 00000000..7dbf2cae
--- /dev/null
+++ b/src/host/buildvm.c
@@ -0,0 +1,504 @@
+/*
+** LuaJIT VM builder.
+** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
+**
+** This is a tool to build the hand-tuned assembler code required for
+** LuaJIT's bytecode interpreter. It supports a variety of output formats
+** to feed different toolchains (see usage() below).
+**
+** This tool is not particularly optimized because it's only used while
+** _building_ LuaJIT. There's no point in distributing or installing it.
+** Only the object code generated by this tool is linked into LuaJIT.
+**
+** Caveat: some memory is not free'd, error handling is lazy.
+** It's a one-shot tool -- any effort fixing this would be wasted.
+*/
+
+#include "buildvm.h"
+#include "lj_obj.h"
+#include "lj_gc.h"
+#include "lj_bc.h"
+#include "lj_ir.h"
+#include "lj_ircall.h"
+#include "lj_frame.h"
+#include "lj_dispatch.h"
+#if LJ_HASFFI
+#include "lj_ctype.h"
+#include "lj_ccall.h"
+#endif
+#include "luajit.h"
+
+#if defined(_WIN32)
+#include <fcntl.h>
+#include <io.h>
+#endif
+
+/* ------------------------------------------------------------------------ */
+
+/* DynASM glue definitions. */
+#define Dst             ctx
+#define Dst_DECL        BuildCtx *ctx
+#define Dst_REF         (ctx->D)
+#define DASM_CHECKS     1
+
+#include "../dynasm/dasm_proto.h"
+
+/* Glue macros for DynASM. */
+static int collect_reloc(BuildCtx *ctx, uint8_t *addr, int idx, int type);
+
+#define DASM_EXTERN(ctx, addr, idx, type) \
+  collect_reloc(ctx, addr, idx, type)
+
+/* ------------------------------------------------------------------------ */
+
+/* Avoid trouble if cross-compiling for an x86 target. Speed doesn't matter. */
+#define DASM_ALIGNED_WRITES     1
+
+/* Embed architecture-specific DynASM encoder. */
+#if LJ_TARGET_X86ORX64
+#include "../dynasm/dasm_x86.h"
+#elif LJ_TARGET_ARM
+#include "../dynasm/dasm_arm.h"
+#elif LJ_TARGET_PPC
+#include "../dynasm/dasm_ppc.h"
+#elif LJ_TARGET_PPCSPE
+#include "../dynasm/dasm_ppc.h"
+#elif LJ_TARGET_MIPS
+#include "../dynasm/dasm_mips.h"
+#else
+#error "No support for this architecture (yet)"
+#endif
+
+/* Embed generated architecture-specific backend. */
+#include "buildvm_arch.h"
+
+/* ------------------------------------------------------------------------ */
+
+void owrite(BuildCtx *ctx, const void *ptr, size_t sz)
+{
+  if (fwrite(ptr, 1, sz, ctx->fp) != sz) {
+    fprintf(stderr, "Error: cannot write to output file: %s\n",
+            strerror(errno));
+    exit(1);
+  }
+}
+
+/* ------------------------------------------------------------------------ */
+
+/* Emit code as raw bytes. Only used for DynASM debugging. */
+static void emit_raw(BuildCtx *ctx)
+{
+  owrite(ctx, ctx->code, ctx->codesz);
+}
+
+/* -- Build machine code -------------------------------------------------- */
+
+static const char *sym_decorate(BuildCtx *ctx,
+                                const char *prefix, const char *suffix)
+{
+  char name[256];
+  char *p;
+#if LJ_64
+  const char *symprefix = ctx->mode == BUILD_machasm ? "_" : "";
+#else
+  const char *symprefix = ctx->mode != BUILD_elfasm ? "_" : "";
+#endif
+  sprintf(name, "%s%s%s", symprefix, prefix, suffix);
+  p = strchr(name, '@');
+  if (p) {
+    if (!LJ_64 && (ctx->mode == BUILD_coffasm || ctx->mode == BUILD_peobj))
+      name[0] = '@';
+    else
+      *p = '\0';
+  }
+  p = (char *)malloc(strlen(name)+1);  /* MSVC doesn't like strdup. */
+  strcpy(p, name);
+  return p;
+}
+
+#define NRELOCSYM       (sizeof(extnames)/sizeof(extnames[0])-1)
+
+static int relocmap[NRELOCSYM];
+
+/* Collect external relocations. */
+static int collect_reloc(BuildCtx *ctx, uint8_t *addr, int idx, int type)
+{
+  if (ctx->nreloc >= BUILD_MAX_RELOC) {
+    fprintf(stderr, "Error: too many relocations, increase BUILD_MAX_RELOC.\n");
+    exit(1);
+  }
+  if (relocmap[idx] < 0) {
+    relocmap[idx] = ctx->nrelocsym;
+    ctx->relocsym[ctx->nrelocsym] = sym_decorate(ctx, "", extnames[idx]);
+    ctx->nrelocsym++;
+  }
+  ctx->reloc[ctx->nreloc].ofs = (int32_t)(addr - ctx->code);
+  ctx->reloc[ctx->nreloc].sym = relocmap[idx];
+  ctx->reloc[ctx->nreloc].type = type;
+  ctx->nreloc++;
+  return 0;  /* Encode symbol offset of 0. */
+}
+
+/* Naive insertion sort. Performance doesn't matter here. */
+static void sym_insert(BuildCtx *ctx, int32_t ofs,
+                       const char *prefix, const char *suffix)
+{
+  ptrdiff_t i = ctx->nsym++;
+  while (i > 0) {
+    if (ctx->sym[i-1].ofs <= ofs)
+      break;
+    ctx->sym[i] = ctx->sym[i-1];
+    i--;
+  }
+  ctx->sym[i].ofs = ofs;
+  ctx->sym[i].name = sym_decorate(ctx, prefix, suffix);
+}
+
+/* Build the machine code. */
+static int build_code(BuildCtx *ctx)
+{
+  int status;
+  int i;
+
+  /* Initialize DynASM structures. */
+  ctx->nglob = GLOB__MAX;
+  ctx->glob = (void **)malloc(ctx->nglob*sizeof(void *));
+  memset(ctx->glob, 0, ctx->nglob*sizeof(void *));
+  ctx->nreloc = 0;
+
+  ctx->globnames = globnames;
+  ctx->relocsym = (const char **)malloc(NRELOCSYM*sizeof(const char *));
+  ctx->nrelocsym = 0;
+  for (i = 0; i < (int)NRELOCSYM; i++) relocmap[i] = -1;
+
+  ctx->dasm_ident = DASM_IDENT;
+  ctx->dasm_arch = DASM_ARCH;
+
+  dasm_init(Dst, DASM_MAXSECTION);
+  dasm_setupglobal(Dst, ctx->glob, ctx->nglob);
+  dasm_setup(Dst, build_actionlist);
+
+  /* Call arch-specific backend to emit the code. */
+  ctx->npc = build_backend(ctx);
+
+  /* Finalize the code. */
+  (void)dasm_checkstep(Dst, -1);
+  if ((status = dasm_link(Dst, &ctx->codesz))) return status;
+  ctx->code = (uint8_t *)malloc(ctx->codesz);
+  if ((status = dasm_encode(Dst, (void *)ctx->code))) return status;
+
+  /* Allocate symbol table and bytecode offsets. */
+  ctx->beginsym = sym_decorate(ctx, "", LABEL_PREFIX "vm_asm_begin");
+  ctx->sym = (BuildSym *)malloc((ctx->npc+ctx->nglob+1)*sizeof(BuildSym));
+  ctx->nsym = 0;
+  ctx->bc_ofs = (int32_t *)malloc(ctx->npc*sizeof(int32_t));
+
+  /* Collect the opcodes (PC labels). */
+  for (i = 0; i < ctx->npc; i++) {
+    int32_t ofs = dasm_getpclabel(Dst, i);
+    if (ofs < 0) return 0x22000000|i;
+    ctx->bc_ofs[i] = ofs;
+    if ((LJ_HASJIT ||
+         !(i == BC_JFORI || i == BC_JFORL || i == BC_JITERL || i == BC_JLOOP ||
+           i == BC_IFORL || i == BC_IITERL || i == BC_ILOOP)) &&
+        (LJ_HASFFI || i != BC_KCDATA))
+      sym_insert(ctx, ofs, LABEL_PREFIX_BC, bc_names[i]);
+  }
+
+  /* Collect the globals (named labels). */
+  for (i = 0; i < ctx->nglob; i++) {
+    const char *gl = globnames[i];
+    int len = (int)strlen(gl);
+    if (!ctx->glob[i]) {
+      fprintf(stderr, "Error: undefined global %s\n", gl);
+      exit(2);
+    }
+    /* Skip the _Z symbols. */
+    if (!(len >= 2 && gl[len-2] == '_' && gl[len-1] == 'Z'))
+      sym_insert(ctx, (int32_t)((uint8_t *)(ctx->glob[i]) - ctx->code),
+                 LABEL_PREFIX, globnames[i]);
+  }
+
+  /* Close the address range. */
+  sym_insert(ctx, (int32_t)ctx->codesz, "", "");
+  ctx->nsym--;
+
+  dasm_free(Dst);
+
+  return 0;
+}
+
+/* -- Generate VM enums --------------------------------------------------- */
+
+const char *const bc_names[] = {
+#define BCNAME(name, ma, mb, mc, mt)       #name,
+BCDEF(BCNAME)
+#undef BCNAME
+  NULL
+};
+
+const char *const ir_names[] = {
+#define IRNAME(name, m, m1, m2) #name,
+IRDEF(IRNAME)
+#undef IRNAME
+  NULL
+};
+
+const char *const irt_names[] = {
+#define IRTNAME(name)   #name,
+IRTDEF(IRTNAME)
+#undef IRTNAME
+  NULL
+};
+
+const char *const irfpm_names[] = {
+#define FPMNAME(name)           #name,
+IRFPMDEF(FPMNAME)
+#undef FPMNAME
+  NULL
+};
+
+const char *const irfield_names[] = {
+#define FLNAME(name, ofs)       #name,
+IRFLDEF(FLNAME)
+#undef FLNAME
+  NULL
+};
+
+const char *const ircall_names[] = {
+#define IRCALLNAME(cond, name, nargs, kind, type, flags)        #name,
+IRCALLDEF(IRCALLNAME)
+#undef IRCALLNAME
+  NULL
+};
+
+static const char *const trace_errors[] = {
+#define TREDEF(name, msg)       msg,
+#include "lj_traceerr.h"
+  NULL
+};
+
+static const char *lower(char *buf, const char *s)
+{
+  char *p = buf;
+  while (*s) {
+    *p++ = (*s >= 'A' && *s <= 'Z') ? *s+0x20 : *s;
+    s++;
+  }
+  *p = '\0';
+  return buf;
+}
+
+/* Emit C source code for bytecode-related definitions. */
+static void emit_bcdef(BuildCtx *ctx)
+{
+  int i;
+  fprintf(ctx->fp, "/* This is a generated file. DO NOT EDIT! */\n\n");
+  fprintf(ctx->fp, "LJ_DATADEF const uint16_t lj_bc_ofs[] = {\n");
+  for (i = 0; i < ctx->npc; i++) {
+    if (i != 0)
+      fprintf(ctx->fp, ",\n");
+    fprintf(ctx->fp, "%d", ctx->bc_ofs[i]);
+  }
+}
+
+/* Emit VM definitions as Lua code for debug modules. */
+static void emit_vmdef(BuildCtx *ctx)
+{
+  char buf[80];
+  int i;
+  fprintf(ctx->fp, "-- This is a generated file. DO NOT EDIT!\n\n");
+  fprintf(ctx->fp, "module(...)\n\n");
+
+  fprintf(ctx->fp, "bcnames = \"");
+  for (i = 0; bc_names[i]; i++) fprintf(ctx->fp, "%-6s", bc_names[i]);
+  fprintf(ctx->fp, "\"\n\n");
+
+  fprintf(ctx->fp, "irnames = \"");
+  for (i = 0; ir_names[i]; i++) fprintf(ctx->fp, "%-6s", ir_names[i]);
+  fprintf(ctx->fp, "\"\n\n");
+
+  fprintf(ctx->fp, "irfpm = { [0]=");
+  for (i = 0; irfpm_names[i]; i++)
+    fprintf(ctx->fp, "\"%s\", ", lower(buf, irfpm_names[i]));
+  fprintf(ctx->fp, "}\n\n");
+
+  fprintf(ctx->fp, "irfield = { [0]=");
+  for (i = 0; irfield_names[i]; i++) {
+    char *p;
+    lower(buf, irfield_names[i]);
+    p = strchr(buf, '_');
+    if (p) *p = '.';
+    fprintf(ctx->fp, "\"%s\", ", buf);
+  }
+  fprintf(ctx->fp, "}\n\n");
+
+  fprintf(ctx->fp, "ircall = {\n[0]=");
+  for (i = 0; ircall_names[i]; i++)
+    fprintf(ctx->fp, "\"%s\",\n", ircall_names[i]);
+  fprintf(ctx->fp, "}\n\n");
+
+  fprintf(ctx->fp, "traceerr = {\n[0]=");
+  for (i = 0; trace_errors[i]; i++)
+    fprintf(ctx->fp, "\"%s\",\n", trace_errors[i]);
+  fprintf(ctx->fp, "}\n\n");
+}
+
+/* -- Argument parsing ---------------------------------------------------- */
+
+/* Build mode names. */
+static const char *const modenames[] = {
+#define BUILDNAME(name)         #name,
+BUILDDEF(BUILDNAME)
+#undef BUILDNAME
+  NULL
+};
+
+/* Print usage information and exit. */
+static void usage(void)
+{
+  int i;
+  fprintf(stderr, LUAJIT_VERSION " VM builder.\n");
+  fprintf(stderr, LUAJIT_COPYRIGHT ", " LUAJIT_URL "\n");
+  fprintf(stderr, "Target architecture: " LJ_ARCH_NAME "\n\n");
+  fprintf(stderr, "Usage: buildvm -m mode [-o outfile] [infiles...]\n\n");
+  fprintf(stderr, "Available modes:\n");
+  for (i = 0; i < BUILD__MAX; i++)
+    fprintf(stderr, "  %s\n", modenames[i]);
+  exit(1);
+}
+
+/* Parse the output mode name. */
+static BuildMode parsemode(const char *mode)
+{
+  int i;
+  for (i = 0; modenames[i]; i++)
+    if (!strcmp(mode, modenames[i]))
+      return (BuildMode)i;
+  usage();
+  return (BuildMode)-1;
+}
+
+/* Parse arguments. */
+static void parseargs(BuildCtx *ctx, char **argv)
+{
+  const char *a;
+  int i;
+  ctx->mode = (BuildMode)-1;
+  ctx->outname = "-";
+  for (i = 1; (a = argv[i]) != NULL; i++) {
+    if (a[0] != '-')
+      break;
+    switch (a[1]) {
+    case '-':
+      if (a[2]) goto err;
+      i++;
+      goto ok;
+    case '\0':
+      goto ok;
+    case 'm':
+      i++;
+      if (a[2] || argv[i] == NULL) goto err;
+      ctx->mode = parsemode(argv[i]);
+      break;
+    case 'o':
+      i++;
+      if (a[2] || argv[i] == NULL) goto err;
+      ctx->outname = argv[i];
+      break;
+    default: err:
+      usage();
+      break;
+    }
+  }
+ok:
+  ctx->args = argv+i;
+  if (ctx->mode == (BuildMode)-1) goto err;
+}
+
+int main(int argc, char **argv)
+{
+  BuildCtx ctx_;
+  BuildCtx *ctx = &ctx_;
+  int status, binmode;
+
+  if (sizeof(void *) != 4*LJ_32+8*LJ_64) {
+    fprintf(stderr,"Error: pointer size mismatch in cross-build.\n");
+    fprintf(stderr,"Try: make HOST_CC=\"gcc -m32\" CROSS=... TARGET=...\n\n");
+    return 1;
+  }
+
+  UNUSED(argc);
+  parseargs(ctx, argv);
+
+  if ((status = build_code(ctx))) {
+    fprintf(stderr,"Error: DASM error %08x\n", status);
+    return 1;
+  }
+
+  switch (ctx->mode) {
+  case BUILD_peobj:
+  case BUILD_raw:
+    binmode = 1;
+    break;
+  default:
+    binmode = 0;
+    break;
+  }
+
+  if (ctx->outname[0] == '-' && ctx->outname[1] == '\0') {
+    ctx->fp = stdout;
+#if defined(_WIN32)
+    if (binmode)
+      _setmode(_fileno(stdout), _O_BINARY);  /* Yuck. */
+#endif
+  } else if (!(ctx->fp = fopen(ctx->outname, binmode ? "wb" : "w"))) {
+    fprintf(stderr, "Error: cannot open output file '%s': %s\n",
+            ctx->outname, strerror(errno));
+    exit(1);
+  }
+
+  switch (ctx->mode) {
+  case BUILD_elfasm:
+  case BUILD_coffasm:
+  case BUILD_machasm:
+    emit_asm(ctx);
+    emit_asm_debug(ctx);
+    break;
+  case BUILD_peobj:
+    emit_peobj(ctx);
+    break;
+  case BUILD_raw:
+    emit_raw(ctx);
+    break;
+  case BUILD_bcdef:
+    emit_bcdef(ctx);
+    emit_lib(ctx);
+    break;
+  case BUILD_vmdef:
+    emit_vmdef(ctx);
+    emit_lib(ctx);
+    break;
+  case BUILD_ffdef:
+  case BUILD_libdef:
+  case BUILD_recdef:
+    emit_lib(ctx);
+    break;
+  case BUILD_folddef:
+    emit_fold(ctx);
+    break;
+  default:
+    break;
+  }
+
+  fflush(ctx->fp);
+  if (ferror(ctx->fp)) {
+    fprintf(stderr, "Error: cannot write to output file: %s\n",
+            strerror(errno));
+    exit(1);
+  }
+  fclose(ctx->fp);
+
+  return 0;
+}
+
diff --git a/src/host/buildvm.h b/src/host/buildvm.h
new file mode 100644
index 00000000..2b7168ed
--- /dev/null
+++ b/src/host/buildvm.h
@@ -0,0 +1,104 @@
+/*
+** LuaJIT VM builder.
+** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
+*/
+
+#ifndef _BUILDVM_H
+#define _BUILDVM_H
+
+#include <sys/types.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <errno.h>
+
+#include "lj_def.h"
+#include "lj_arch.h"
+
+/* Hardcoded limits. Increase as needed. */
+#define BUILD_MAX_RELOC         200     /* Max. number of relocations. */
+#define BUILD_MAX_FOLD          4096    /* Max. number of fold rules. */
+
+/* Prefix for scanned library definitions. */
+#define LIBDEF_PREFIX           "LJLIB_"
+
+/* Prefix for scanned fold definitions. */
+#define FOLDDEF_PREFIX          "LJFOLD"
+
+/* Prefixes for generated labels. */
+#define LABEL_PREFIX            "lj_"
+#define LABEL_PREFIX_BC         LABEL_PREFIX "BC_"
+#define LABEL_PREFIX_FF         LABEL_PREFIX "ff_"
+#define LABEL_PREFIX_CF         LABEL_PREFIX "cf_"
+#define LABEL_PREFIX_FFH        LABEL_PREFIX "ffh_"
+#define LABEL_PREFIX_LIBCF      LABEL_PREFIX "lib_cf_"
+#define LABEL_PREFIX_LIBINIT    LABEL_PREFIX "lib_init_"
+
+/* Forward declaration. */
+struct dasm_State;
+
+/* Build modes. */
+#define BUILDDEF(_) \
+  _(elfasm) _(coffasm) _(machasm) _(peobj) _(raw) \
+  _(bcdef) _(ffdef) _(libdef) _(recdef) _(vmdef) \
+  _(folddef)
+
+typedef enum {
+#define BUILDENUM(name)         BUILD_##name,
+BUILDDEF(BUILDENUM)
+#undef BUILDENUM
+  BUILD__MAX
+} BuildMode;
+
+/* Code relocation. */
+typedef struct BuildReloc {
+  int32_t ofs;
+  int sym;
+  int type;
+} BuildReloc;
+
+typedef struct BuildSym {
+  const char *name;
+  int32_t ofs;
+} BuildSym;
+
+/* Build context structure. */
+typedef struct BuildCtx {
+  /* DynASM state pointer. Should be first member. */
+  struct dasm_State *D;
+  /* Parsed command line. */
+  BuildMode mode;
+  FILE *fp;
+  const char *outname;
+  char **args;
+  /* Code and symbols generated by DynASM. */
+  uint8_t *code;
+  size_t codesz;
+  int npc, nglob, nsym, nreloc, nrelocsym;
+  void **glob;
+  BuildSym *sym;
+  const char **relocsym;
+  int32_t *bc_ofs;
+  const char *beginsym;
+  /* Strings generated by DynASM. */
+  const char *const *globnames;
+  const char *dasm_ident;
+  const char *dasm_arch;
+  /* Relocations. */
+  BuildReloc reloc[BUILD_MAX_RELOC];
+} BuildCtx;
+
+extern void owrite(BuildCtx *ctx, const void *ptr, size_t sz);
+extern void emit_asm(BuildCtx *ctx);
+extern void emit_peobj(BuildCtx *ctx);
+extern void emit_lib(BuildCtx *ctx);
+extern void emit_fold(BuildCtx *ctx);
+
+extern const char *const bc_names[];
+extern const char *const ir_names[];
+extern const char *const irt_names[];
+extern const char *const irfpm_names[];
+extern const char *const irfield_names[];
+extern const char *const ircall_names[];
+
+#endif
diff --git a/src/host/buildvm_asm.c b/src/host/buildvm_asm.c
new file mode 100644
index 00000000..f975eadc
--- /dev/null
+++ b/src/host/buildvm_asm.c
@@ -0,0 +1,283 @@
+/*
+** LuaJIT VM builder: Assembler source code emitter.
+** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
+*/
+
+#include "buildvm.h"
+#include "lj_bc.h"
+
+/* ------------------------------------------------------------------------ */
+
+#if LJ_TARGET_X86ORX64
+/* Emit bytes piecewise as assembler text. */
+static void emit_asm_bytes(BuildCtx *ctx, uint8_t *p, int n)
+{
+  int i;
+  for (i = 0; i < n; i++) {
+    if ((i & 15) == 0)
+      fprintf(ctx->fp, "\t.byte %d", p[i]);
+    else
+      fprintf(ctx->fp, ",%d", p[i]);
+    if ((i & 15) == 15) putc('\n', ctx->fp);
+  }
+  if ((n & 15) != 0) putc('\n', ctx->fp);
+}
+
+/* Emit relocation */
+static void emit_asm_reloc(BuildCtx *ctx, int type, const char *sym)
+{
+  switch (ctx->mode) {
+  case BUILD_elfasm:
+    if (type)
+      fprintf(ctx->fp, "\t.long %s-.-4\n", sym);
+    else
+      fprintf(ctx->fp, "\t.long %s\n", sym);
+    break;
+  case BUILD_coffasm:
+    fprintf(ctx->fp, "\t.def %s; .scl 3; .type 32; .endef\n", sym);
+    if (type)
+      fprintf(ctx->fp, "\t.long %s-.-4\n", sym);
+    else
+      fprintf(ctx->fp, "\t.long %s\n", sym);
+    break;
+  default:  /* BUILD_machasm for relative relocations handled below. */
+    fprintf(ctx->fp, "\t.long %s\n", sym);
+    break;
+  }
+}
+
+static const char *const jccnames[] = {
+  "jo", "jno", "jb", "jnb", "jz", "jnz", "jbe", "ja",
+  "js", "jns", "jpe", "jpo", "jl", "jge", "jle", "jg"
+};
+
+/* Emit relocation for the incredibly stupid OSX assembler. */
+static void emit_asm_reloc_mach(BuildCtx *ctx, uint8_t *cp, int n,
+                                const char *sym)
+{
+  const char *opname = NULL;
+  if (--n < 0) goto err;
+  if (cp[n] == 0xe8) {
+    opname = "call";
+  } else if (cp[n] == 0xe9) {
+    opname = "jmp";
+  } else if (cp[n] >= 0x80 && cp[n] <= 0x8f && n > 0 && cp[n-1] == 0x0f) {
+    opname = jccnames[cp[n]-0x80];
+    n--;
+  } else {
+err:
+    fprintf(stderr, "Error: unsupported opcode for %s symbol relocation.\n",
+            sym);
+    exit(1);
+  }
+  emit_asm_bytes(ctx, cp, n);
+  fprintf(ctx->fp, "\t%s %s\n", opname, sym);
+}
+#else
+/* Emit words piecewise as assembler text. */
+static void emit_asm_words(BuildCtx *ctx, uint8_t *p, int n)
+{
+  int i;
+  for (i = 0; i < n; i += 4) {
+    if ((i & 15) == 0)
+      fprintf(ctx->fp, "\t.long 0x%08x", *(uint32_t *)(p+i));
+    else
+      fprintf(ctx->fp, ",0x%08x", *(uint32_t *)(p+i));
+    if ((i & 15) == 12) putc('\n', ctx->fp);
+  }
+  if ((n & 15) != 0) putc('\n', ctx->fp);
+}
+
+/* Emit relocation as part of an instruction. */
+static void emit_asm_wordreloc(BuildCtx *ctx, uint8_t *p, int n,
+                               const char *sym)
+{
+  uint32_t ins;
+  emit_asm_words(ctx, p, n-4);
+  ins = *(uint32_t *)(p+n-4);
+#if LJ_TARGET_ARM
+  if ((ins & 0xff000000u) == 0xfa000000u) {
+    fprintf(ctx->fp, "\tblx %s\n", sym);
+  } else if ((ins & 0x0e000000u) == 0x0a000000u) {
+    fprintf(ctx->fp, "\t%s%.2s %s\n", (ins & 0x01000000u) ? "bl" : "b",
+            "eqnecsccmiplvsvchilsgeltgtle" + 2*(ins >> 28), sym);
+  } else {
+    fprintf(stderr,
+            "Error: unsupported opcode %08x for %s symbol relocation.\n",
+            ins, sym);
+    exit(1);
+  }
+#elif LJ_TARGET_PPC || LJ_TARGET_PPCSPE
+  if ((ins >> 26) == 16) {
+    fprintf(ctx->fp, "\t%s %d, %d, %s\n",
+            (ins & 1) ? "bcl" : "bc", (ins >> 21) & 31, (ins >> 16) & 31, sym);
+  } else if ((ins >> 26) == 18) {
+    fprintf(ctx->fp, "\t%s %s\n", (ins & 1) ? "bl" : "b", sym);
+  } else {
+    fprintf(stderr,
+            "Error: unsupported opcode %08x for %s symbol relocation.\n",
+            ins, sym);
+    exit(1);
+  }
+#elif LJ_TARGET_MIPS
+  fprintf(stderr,
+          "Error: unsupported opcode %08x for %s symbol relocation.\n",
+          ins, sym);
+  exit(1);
+#else
+#error "missing relocation support for this architecture"
+#endif
+}
+#endif
+
+#if LJ_TARGET_ARM
+#define ELFASM_PX       "%%"
+#else
+#define ELFASM_PX       "@"
+#endif
+
+/* Emit an assembler label. */
+static void emit_asm_label(BuildCtx *ctx, const char *name, int size, int isfunc)
+{
+  switch (ctx->mode) {
+  case BUILD_elfasm:
+    fprintf(ctx->fp,
+      "\n\t.globl %s\n"
+      "\t.hidden %s\n"
+      "\t.type %s, " ELFASM_PX "%s\n"
+      "\t.size %s, %d\n"
+      "%s:\n",
+      name, name, name, isfunc ? "function" : "object", name, size, name);
+    break;
+  case BUILD_coffasm:
+    fprintf(ctx->fp, "\n\t.globl %s\n", name);
+    if (isfunc)
+      fprintf(ctx->fp, "\t.def %s; .scl 3; .type 32; .endef\n", name);
+    fprintf(ctx->fp, "%s:\n", name);
+    break;
+  case BUILD_machasm:
+    fprintf(ctx->fp,
+      "\n\t.private_extern %s\n"
+      "%s:\n", name, name);
+    break;
+  default:
+    break;
+  }
+}
+
+/* Emit alignment. */
+static void emit_asm_align(BuildCtx *ctx, int bits)
+{
+  switch (ctx->mode) {
+  case BUILD_elfasm:
+  case BUILD_coffasm:
+    fprintf(ctx->fp, "\t.p2align %d\n", bits);
+    break;
+  case BUILD_machasm:
+    fprintf(ctx->fp, "\t.align %d\n", bits);
+    break;
+  default:
+    break;
+  }
+}
+
+/* ------------------------------------------------------------------------ */
+
+/* Emit assembler source code. */
+void emit_asm(BuildCtx *ctx)
+{
+  int i, rel;
+
+  fprintf(ctx->fp, "\t.file \"buildvm_%s.dasc\"\n", ctx->dasm_arch);
+  fprintf(ctx->fp, "\t.text\n");
+  emit_asm_align(ctx, 4);
+
+  emit_asm_label(ctx, ctx->beginsym, 0, 0);
+  if (ctx->mode != BUILD_machasm)
+    fprintf(ctx->fp, ".Lbegin:\n");
+
+#if LJ_TARGET_ARM && defined(__GNUC__) && !defined(LUAJIT_NO_UNWIND)
+  /* This should really be moved into buildvm_arm.dasc. */
+  fprintf(ctx->fp,
+          ".fnstart\n"
+          ".save {r4, r5, r6, r7, r8, r9, r10, r11, lr}\n"
+          ".pad #28\n");
+#endif
+#if LJ_TARGET_MIPS
+  fprintf(ctx->fp, ".set nomips16\n.abicalls\n.set noreorder\n.set nomacro\n");
+#endif
+
+  for (i = rel = 0; i < ctx->nsym; i++) {
+    int32_t ofs = ctx->sym[i].ofs;
+    int32_t next = ctx->sym[i+1].ofs;
+#if LJ_TARGET_ARM && defined(__GNUC__) && !defined(LUAJIT_NO_UNWIND) && \
+    LJ_HASFFI
+    if (!strcmp(ctx->sym[i].name, "lj_vm_ffi_call"))
+      fprintf(ctx->fp,
+              ".globl lj_err_unwind_arm\n"
+              ".personality lj_err_unwind_arm\n"
+              ".fnend\n"
+              ".fnstart\n"
+              ".save {r4, r5, r11, lr}\n"
+              ".setfp r11, sp\n");
+#endif
+    emit_asm_label(ctx, ctx->sym[i].name, next - ofs, 1);
+    while (rel < ctx->nreloc && ctx->reloc[rel].ofs <= next) {
+      BuildReloc *r = &ctx->reloc[rel];
+      int n = r->ofs - ofs;
+#if LJ_TARGET_X86ORX64
+      if (ctx->mode == BUILD_machasm && r->type != 0) {
+        emit_asm_reloc_mach(ctx, ctx->code+ofs, n, ctx->relocsym[r->sym]);
+      } else {
+        emit_asm_bytes(ctx, ctx->code+ofs, n);
+        emit_asm_reloc(ctx, r->type, ctx->relocsym[r->sym]);
+      }
+      ofs += n+4;
+#else
+      emit_asm_wordreloc(ctx, ctx->code+ofs, n, ctx->relocsym[r->sym]);
+      ofs += n;
+#endif
+      rel++;
+    }
+#if LJ_TARGET_X86ORX64
+    emit_asm_bytes(ctx, ctx->code+ofs, next-ofs);
+#else
+    emit_asm_words(ctx, ctx->code+ofs, next-ofs);
+#endif
+  }
+
+#if LJ_TARGET_ARM && defined(__GNUC__) && !defined(LUAJIT_NO_UNWIND)
+  fprintf(ctx->fp,
+#if !LJ_HASFFI
+          ".globl lj_err_unwind_arm\n"
+          ".personality lj_err_unwind_arm\n"
+#endif
+          ".fnend\n");
+#endif
+
+  fprintf(ctx->fp, "\n");
+  switch (ctx->mode) {
+  case BUILD_elfasm:
+    fprintf(ctx->fp, "\t.section .note.GNU-stack,\"\"," ELFASM_PX "progbits\n");
+#if LJ_TARGET_PPCSPE
+    /* Soft-float ABI + SPE. */
+    fprintf(ctx->fp, "\t.gnu_attribute 4, 2\n\t.gnu_attribute 8, 3\n");
+#elif LJ_TARGET_PPC
+    /* Hard-float ABI. */
+    fprintf(ctx->fp, "\t.gnu_attribute 4, 1\n");
+#endif
+    /* fallthrough */
+  case BUILD_coffasm:
+    fprintf(ctx->fp, "\t.ident \"%s\"\n", ctx->dasm_ident);
+    break;
+  case BUILD_machasm:
+    fprintf(ctx->fp,
+      "\t.cstring\n"
+      "\t.ascii \"%s\\0\"\n", ctx->dasm_ident);
+    break;
+  default:
+    break;
+  }
+  fprintf(ctx->fp, "\n");
+}
+
diff --git a/src/host/buildvm_fold.c b/src/host/buildvm_fold.c
new file mode 100644
index 00000000..73f4f80a
--- /dev/null
+++ b/src/host/buildvm_fold.c
@@ -0,0 +1,229 @@
+/*
+** LuaJIT VM builder: IR folding hash table generator.
+** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
+*/
+
+#include "buildvm.h"
+#include "lj_obj.h"
+#include "lj_ir.h"
+
+/* Context for the folding hash table generator. */
+static int lineno;
+static int funcidx;
+static uint32_t foldkeys[BUILD_MAX_FOLD];
+static uint32_t nkeys;
+
+/* Try to fill the hash table with keys using the hash parameters. */
+static int tryhash(uint32_t *htab, uint32_t sz, uint32_t r, int dorol)
+{
+  uint32_t i;
+  if (dorol && ((r & 31) == 0 || (r>>5) == 0))
+    return 0;  /* Avoid zero rotates. */
+  memset(htab, 0xff, (sz+1)*sizeof(uint32_t));
+  for (i = 0; i < nkeys; i++) {
+    uint32_t key = foldkeys[i];
+    uint32_t k = key & 0xffffff;
+    uint32_t h = (dorol ? lj_rol(lj_rol(k, r>>5) - k, r&31) :
+                          (((k << (r>>5)) - k) << (r&31))) % sz;
+    if (htab[h] != 0xffffffff) {  /* Collision on primary slot. */
+      if (htab[h+1] != 0xffffffff) {  /* Collision on secondary slot. */
+        /* Try to move the colliding key, if possible. */
+        if (h < sz-1 && htab[h+2] == 0xffffffff) {
+          uint32_t k2 = htab[h+1] & 0xffffff;
+          uint32_t h2 = (dorol ? lj_rol(lj_rol(k2, r>>5) - k2, r&31) :
+                                 (((k2 << (r>>5)) - k2) << (r&31))) % sz;
+          if (h2 != h+1) return 0;  /* Cannot resolve collision. */
+          htab[h+2] = htab[h+1];  /* Move colliding key to secondary slot. */
+        } else {
+          return 0;  /* Collision. */
+        }
+      }
+      htab[h+1] = key;
+    } else {
+      htab[h] = key;
+    }
+  }
+  return 1;  /* Success, all keys could be stored. */
+}
+
+/* Print the generated hash table. */
+static void printhash(BuildCtx *ctx, uint32_t *htab, uint32_t sz)
+{
+  uint32_t i;
+  fprintf(ctx->fp, "static const uint32_t fold_hash[%d] = {\n0x%08x",
+          sz+1, htab[0]);
+  for (i = 1; i < sz+1; i++)
+    fprintf(ctx->fp, ",\n0x%08x", htab[i]);
+  fprintf(ctx->fp, "\n};\n\n");
+}
+
+/* Exhaustive search for the shortest semi-perfect hash table. */
+static void makehash(BuildCtx *ctx)
+{
+  uint32_t htab[BUILD_MAX_FOLD*2+1];
+  uint32_t sz, r;
+  /* Search for the smallest hash table with an odd size. */
+  for (sz = (nkeys|1); sz < BUILD_MAX_FOLD*2; sz += 2) {
+    /* First try all shift hash combinations. */
+    for (r = 0; r < 32*32; r++) {
+      if (tryhash(htab, sz, r, 0)) {
+        printhash(ctx, htab, sz);
+        fprintf(ctx->fp,
+                "#define fold_hashkey(k)\t(((((k)<<%u)-(k))<<%u)%%%u)\n\n",
+                r>>5, r&31, sz);
+        return;
+      }
+    }
+    /* Then try all rotate hash combinations. */
+    for (r = 0; r < 32*32; r++) {
+      if (tryhash(htab, sz, r, 1)) {
+        printhash(ctx, htab, sz);
+        fprintf(ctx->fp,
+          "#define fold_hashkey(k)\t(lj_rol(lj_rol((k),%u)-(k),%u)%%%u)\n\n",
+                r>>5, r&31, sz);
+        return;
+      }
+    }
+  }
+  fprintf(stderr, "Error: search for perfect hash failed\n");
+  exit(1);
+}
+
+/* Parse one token of a fold rule. */
+static uint32_t nexttoken(char **pp, int allowlit, int allowany)
+{
+  char *p = *pp;
+  if (p) {
+    uint32_t i;
+    char *q = strchr(p, ' ');
+    if (q) *q++ = '\0';
+    *pp = q;
+    if (allowlit && !strncmp(p, "IRFPM_", 6)) {
+      for (i = 0; irfpm_names[i]; i++)
+        if (!strcmp(irfpm_names[i], p+6))
+          return i;
+    } else if (allowlit && !strncmp(p, "IRFL_", 5)) {
+      for (i = 0; irfield_names[i]; i++)
+        if (!strcmp(irfield_names[i], p+5))
+          return i;
+    } else if (allowlit && !strncmp(p, "IRCALL_", 7)) {
+      for (i = 0; ircall_names[i]; i++)
+        if (!strcmp(ircall_names[i], p+7))
+          return i;
+    } else if (allowlit && !strncmp(p, "IRCONV_", 7)) {
+      for (i = 0; irt_names[i]; i++) {
+        const char *r = strchr(p+7, '_');
+        if (r && !strncmp(irt_names[i], p+7, r-(p+7))) {
+          uint32_t j;
+          for (j = 0; irt_names[j]; j++)
+            if (!strcmp(irt_names[j], r+1))
+              return (i << 5) + j;
+        }
+      }
+    } else if (allowlit && *p >= '0' && *p <= '9') {
+      for (i = 0; *p >= '0' && *p <= '9'; p++)
+        i = i*10 + (*p - '0');
+      if (*p == '\0')
+        return i;
+    } else if (allowany && !strcmp("any", p)) {
+      return allowany;
+    } else {
+      for (i = 0; ir_names[i]; i++)
+        if (!strcmp(ir_names[i], p))
+          return i;
+    }
+    fprintf(stderr, "Error: bad fold definition token \"%s\" at line %d\n", p, lineno);
+    exit(1);
+  }
+  return 0;
+}
+
+/* Parse a fold rule. */
+static void foldrule(char *p)
+{
+  uint32_t op = nexttoken(&p, 0, 0);
+  uint32_t left = nexttoken(&p, 0, 0x7f);
+  uint32_t right = nexttoken(&p, 1, 0x3ff);
+  uint32_t key = (funcidx << 24) | (op << 17) | (left << 10) | right;
+  uint32_t i;
+  if (nkeys >= BUILD_MAX_FOLD) {
+    fprintf(stderr, "Error: too many fold rules, increase BUILD_MAX_FOLD.\n");
+    exit(1);
+  }
+  /* Simple insertion sort to detect duplicates. */
+  for (i = nkeys; i > 0; i--) {
+    if ((foldkeys[i-1]&0xffffff) < (key & 0xffffff))
+      break;
+    if ((foldkeys[i-1]&0xffffff) == (key & 0xffffff)) {
+      fprintf(stderr, "Error: duplicate fold definition at line %d\n", lineno);
+      exit(1);
+    }
+    foldkeys[i] = foldkeys[i-1];
+  }
+  foldkeys[i] = key;
+  nkeys++;
+}
+
+/* Emit C source code for IR folding hash table. */
+void emit_fold(BuildCtx *ctx)
+{
+  char buf[256];  /* We don't care about analyzing lines longer than that. */
+  const char *fname = ctx->args[0];
+  FILE *fp;
+
+  if (fname == NULL) {
+    fprintf(stderr, "Error: missing input filename\n");
+    exit(1);
+  }
+
+  if (fname[0] == '-' && fname[1] == '\0') {
+    fp = stdin;
+  } else {
+    fp = fopen(fname, "r");
+    if (!fp) {
+      fprintf(stderr, "Error: cannot open input file '%s': %s\n",
+              fname, strerror(errno));
+      exit(1);
+    }
+  }
+
+  fprintf(ctx->fp, "/* This is a generated file. DO NOT EDIT! */\n\n");
+  fprintf(ctx->fp, "static const FoldFunc fold_func[] = {\n");
+
+  lineno = 0;
+  funcidx = 0;
+  nkeys = 0;
+  while (fgets(buf, sizeof(buf), fp) != NULL) {
+    lineno++;
+    /* The prefix must be at the start of a line, otherwise it's ignored. */
+    if (!strncmp(buf, FOLDDEF_PREFIX, sizeof(FOLDDEF_PREFIX)-1)) {
+      char *p = buf+sizeof(FOLDDEF_PREFIX)-1;
+      char *q = strchr(p, ')');
+      if (p[0] == '(' && q) {
+        p++;
+        *q = '\0';
+        foldrule(p);
+      } else if ((p[0] == 'F' || p[0] == 'X') && p[1] == '(' && q) {
+        p += 2;
+        *q = '\0';
+        if (funcidx)
+          fprintf(ctx->fp, ",\n");
+        if (p[-2] == 'X')
+          fprintf(ctx->fp, "  %s", p);
+        else
+          fprintf(ctx->fp, "  fold_%s", p);
+        funcidx++;
+      } else {
+        buf[strlen(buf)-1] = '\0';
+        fprintf(stderr, "Error: unknown fold definition tag %s%s at line %d\n",
+                FOLDDEF_PREFIX, p, lineno);
+        exit(1);
+      }
+    }
+  }
+  fclose(fp);
+  fprintf(ctx->fp, "\n};\n\n");
+
+  makehash(ctx);
+}
+
diff --git a/src/host/buildvm_lib.c b/src/host/buildvm_lib.c
new file mode 100644
index 00000000..3231d3ad
--- /dev/null
+++ b/src/host/buildvm_lib.c
@@ -0,0 +1,377 @@
+/*
+** LuaJIT VM builder: library definition compiler.
+** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
+*/
+
+#include "buildvm.h"
+#include "lj_obj.h"
+#include "lj_lib.h"
+
+/* Context for library definitions. */
+static uint8_t obuf[8192];
+static uint8_t *optr;
+static char modname[80];
+static size_t modnamelen;
+static char funcname[80];
+static int modstate, regfunc;
+static int ffid, recffid, ffasmfunc;
+
+enum {
+  REGFUNC_OK,
+  REGFUNC_NOREG,
+  REGFUNC_NOREGUV
+};
+
+static void libdef_name(const char *p, int kind)
+{
+  size_t n = strlen(p);
+  if (kind != LIBINIT_STRING) {
+    if (n > modnamelen && p[modnamelen] == '_' &&
+        !strncmp(p, modname, modnamelen)) {
+      p += modnamelen+1;
+      n -= modnamelen+1;
+    }
+  }
+  if (n > LIBINIT_MAXSTR) {
+    fprintf(stderr, "Error: string too long: '%s'\n",  p);
+    exit(1);
+  }
+  if (optr+1+n+2 > obuf+sizeof(obuf)) {  /* +2 for caller. */
+    fprintf(stderr, "Error: output buffer overflow\n");
+    exit(1);
+  }
+  *optr++ = (uint8_t)(n | kind);
+  memcpy(optr, p, n);
+  optr += n;
+}
+
+static void libdef_endmodule(BuildCtx *ctx)
+{
+  if (modstate != 0) {
+    char line[80];
+    const uint8_t *p;
+    int n;
+    if (modstate == 1)
+      fprintf(ctx->fp, "  (lua_CFunction)0");
+    fprintf(ctx->fp, "\n};\n");
+    fprintf(ctx->fp, "static const uint8_t %s%s[] = {\n",
+            LABEL_PREFIX_LIBINIT, modname);
+    line[0] = '\0';
+    for (n = 0, p = obuf; p < optr; p++) {
+      n += sprintf(line+n, "%d,", *p);
+      if (n >= 75) {
+        fprintf(ctx->fp, "%s\n", line);
+        n = 0;
+        line[0] = '\0';
+      }
+    }
+    fprintf(ctx->fp, "%s%d\n};\n#endif\n\n", line, LIBINIT_END);
+  }
+}
+
+static void libdef_module(BuildCtx *ctx, char *p, int arg)
+{
+  UNUSED(arg);
+  if (ctx->mode == BUILD_libdef) {
+    libdef_endmodule(ctx);
+    optr = obuf;
+    *optr++ = (uint8_t)ffid;
+    *optr++ = (uint8_t)ffasmfunc;
+    *optr++ = 0;  /* Hash table size. */
+    modstate = 1;
+    fprintf(ctx->fp, "#ifdef %sMODULE_%s\n", LIBDEF_PREFIX, p);
+    fprintf(ctx->fp, "#undef %sMODULE_%s\n", LIBDEF_PREFIX, p);
+    fprintf(ctx->fp, "static const lua_CFunction %s%s[] = {\n",
+            LABEL_PREFIX_LIBCF, p);
+  }
+  modnamelen = strlen(p);
+  if (modnamelen > sizeof(modname)-1) {
+    fprintf(stderr, "Error: module name too long: '%s'\n", p);
+    exit(1);
+  }
+  strcpy(modname, p);
+}
+
+static int find_ffofs(BuildCtx *ctx, const char *name)
+{
+  int i;
+  for (i = 0; i < ctx->nglob; i++) {
+    const char *gl = ctx->globnames[i];
+    if (gl[0] == 'f' && gl[1] == 'f' && gl[2] == '_' && !strcmp(gl+3, name)) {
+      return (int)((uint8_t *)ctx->glob[i] - ctx->code);
+    }
+  }
+  fprintf(stderr, "Error: undefined fast function %s%s\n",
+          LABEL_PREFIX_FF, name);
+  exit(1);
+}
+
+static void libdef_func(BuildCtx *ctx, char *p, int arg)
+{
+  if (arg != LIBINIT_CF)
+    ffasmfunc++;
+  if (ctx->mode == BUILD_libdef) {
+    if (modstate == 0) {
+      fprintf(stderr, "Error: no module for function definition %s\n", p);
+      exit(1);
+    }
+    if (regfunc == REGFUNC_NOREG) {
+      if (optr+1 > obuf+sizeof(obuf)) {
+        fprintf(stderr, "Error: output buffer overflow\n");
+        exit(1);
+      }
+      *optr++ = LIBINIT_FFID;
+    } else {
+      if (arg != LIBINIT_ASM_) {
+        if (modstate != 1) fprintf(ctx->fp, ",\n");
+        modstate = 2;
+        fprintf(ctx->fp, "  %s%s", arg ? LABEL_PREFIX_FFH : LABEL_PREFIX_CF, p);
+      }
+      if (regfunc != REGFUNC_NOREGUV) obuf[2]++;  /* Bump hash table size. */
+      libdef_name(regfunc == REGFUNC_NOREGUV ? "" : p, arg);
+    }
+  } else if (ctx->mode == BUILD_ffdef) {
+    fprintf(ctx->fp, "FFDEF(%s)\n", p);
+  } else if (ctx->mode == BUILD_recdef) {
+    if (strlen(p) > sizeof(funcname)-1) {
+      fprintf(stderr, "Error: function name too long: '%s'\n", p);
+      exit(1);
+    }
+    strcpy(funcname, p);
+  } else if (ctx->mode == BUILD_vmdef) {
+    int i;
+    for (i = 1; p[i] && modname[i-1]; i++)
+      if (p[i] == '_') p[i] = '.';
+    fprintf(ctx->fp, "\"%s\",\n", p);
+  } else if (ctx->mode == BUILD_bcdef) {
+    if (arg != LIBINIT_CF)
+      fprintf(ctx->fp, ",\n%d", find_ffofs(ctx, p));
+  }
+  ffid++;
+  regfunc = REGFUNC_OK;
+}
+
+static uint32_t find_rec(char *name)
+{
+  char *p = (char *)obuf;
+  uint32_t n;
+  for (n = 2; *p; n++) {
+    if (strcmp(p, name) == 0)
+      return n;
+    p += strlen(p)+1;
+  }
+  if (p+strlen(name)+1 >= (char *)obuf+sizeof(obuf)) {
+    fprintf(stderr, "Error: output buffer overflow\n");
+    exit(1);
+  }
+  strcpy(p, name);
+  return n;
+}
+
+static void libdef_rec(BuildCtx *ctx, char *p, int arg)
+{
+  UNUSED(arg);
+  if (ctx->mode == BUILD_recdef) {
+    char *q;
+    uint32_t n;
+    for (; recffid+1 < ffid; recffid++)
+      fprintf(ctx->fp, ",\n0");
+    recffid = ffid;
+    if (*p == '.') p = funcname;
+    q = strchr(p, ' ');
+    if (q) *q++ = '\0';
+    n = find_rec(p);
+    if (q)
+      fprintf(ctx->fp, ",\n0x%02x00+(%s)", n, q);
+    else
+      fprintf(ctx->fp, ",\n0x%02x00", n);
+  }
+}
+
+static void memcpy_endian(void *dst, void *src, size_t n)
+{
+  union { uint8_t b; uint32_t u; } host_endian;
+  host_endian.u = 1;
+  if (host_endian.b == LJ_ENDIAN_SELECT(1, 0)) {
+    memcpy(dst, src, n);
+  } else {
+    size_t i;
+    for (i = 0; i < n; i++)
+      ((uint8_t *)dst)[i] = ((uint8_t *)src)[n-i-1];
+  }
+}
+
+static void libdef_push(BuildCtx *ctx, char *p, int arg)
+{
+  UNUSED(arg);
+  if (ctx->mode == BUILD_libdef) {
+    int len = (int)strlen(p);
+    if (*p == '"') {
+      if (len > 1 && p[len-1] == '"') {
+        p[len-1] = '\0';
+        libdef_name(p+1, LIBINIT_STRING);
+        return;
+      }
+    } else if (*p >= '0' && *p <= '9') {
+      char *ep;
+      double d = strtod(p, &ep);
+      if (*ep == '\0') {
+        if (optr+1+sizeof(double) > obuf+sizeof(obuf)) {
+          fprintf(stderr, "Error: output buffer overflow\n");
+          exit(1);
+        }
+        *optr++ = LIBINIT_NUMBER;
+        memcpy_endian(optr, &d, sizeof(double));
+        optr += sizeof(double);
+        return;
+      }
+    } else if (!strcmp(p, "lastcl")) {
+      if (optr+1 > obuf+sizeof(obuf)) {
+        fprintf(stderr, "Error: output buffer overflow\n");
+        exit(1);
+      }
+      *optr++ = LIBINIT_LASTCL;
+      return;
+    } else if (len > 4 && !strncmp(p, "top-", 4)) {
+      if (optr+2 > obuf+sizeof(obuf)) {
+        fprintf(stderr, "Error: output buffer overflow\n");
+        exit(1);
+      }
+      *optr++ = LIBINIT_COPY;
+      *optr++ = (uint8_t)atoi(p+4);
+      return;
+    }
+    fprintf(stderr, "Error: bad value for %sPUSH(%s)\n", LIBDEF_PREFIX, p);
+    exit(1);
+  }
+}
+
+static void libdef_set(BuildCtx *ctx, char *p, int arg)
+{
+  UNUSED(arg);
+  if (ctx->mode == BUILD_libdef) {
+    if (p[0] == '!' && p[1] == '\0') p[0] = '\0';  /* Set env. */
+    libdef_name(p, LIBINIT_STRING);
+    *optr++ = LIBINIT_SET;
+    obuf[2]++;  /* Bump hash table size. */
+  }
+}
+
+static void libdef_regfunc(BuildCtx *ctx, char *p, int arg)
+{
+  UNUSED(ctx); UNUSED(p);
+  regfunc = arg;
+}
+
+typedef void (*LibDefFunc)(BuildCtx *ctx, char *p, int arg);
+
+typedef struct LibDefHandler {
+  const char *suffix;
+  const char *stop;
+  const LibDefFunc func;
+  const int arg;
+} LibDefHandler;
+
+static const LibDefHandler libdef_handlers[] = {
+  { "MODULE_",  " \t\r\n",      libdef_module,          0 },
+  { "CF(",      ")",            libdef_func,            LIBINIT_CF },
+  { "ASM(",     ")",            libdef_func,            LIBINIT_ASM },
+  { "ASM_(",    ")",            libdef_func,            LIBINIT_ASM_ },
+  { "REC(",     ")",            libdef_rec,             0 },
+  { "PUSH(",    ")",            libdef_push,            0 },
+  { "SET(",     ")",            libdef_set,             0 },
+  { "NOREGUV",  NULL,           libdef_regfunc,         REGFUNC_NOREGUV },
+  { "NOREG",    NULL,           libdef_regfunc,         REGFUNC_NOREG },
+  { NULL,       NULL,           (LibDefFunc)0,          0 }
+};
+
+/* Emit C source code for library function definitions. */
+void emit_lib(BuildCtx *ctx)
+{
+  const char *fname;
+
+  if (ctx->mode == BUILD_ffdef || ctx->mode == BUILD_libdef ||
+      ctx->mode == BUILD_recdef)
+    fprintf(ctx->fp, "/* This is a generated file. DO NOT EDIT! */\n\n");
+  else if (ctx->mode == BUILD_vmdef)
+    fprintf(ctx->fp, "ffnames = {\n[0]=\"Lua\",\n\"C\",\n");
+  if (ctx->mode == BUILD_recdef)
+    fprintf(ctx->fp, "static const uint16_t recff_idmap[] = {\n0,\n0x0100");
+  recffid = ffid = FF_C+1;
+  ffasmfunc = 0;
+
+  while ((fname = *ctx->args++)) {
+    char buf[256];  /* We don't care about analyzing lines longer than that. */
+    FILE *fp;
+    if (fname[0] == '-' && fname[1] == '\0') {
+      fp = stdin;
+    } else {
+      fp = fopen(fname, "r");
+      if (!fp) {
+        fprintf(stderr, "Error: cannot open input file '%s': %s\n",
+                fname, strerror(errno));
+        exit(1);
+      }
+    }
+    modstate = 0;
+    regfunc = REGFUNC_OK;
+    while (fgets(buf, sizeof(buf), fp) != NULL) {
+      char *p;
+      for (p = buf; (p = strstr(p, LIBDEF_PREFIX)) != NULL; ) {
+        const LibDefHandler *ldh;
+        p += sizeof(LIBDEF_PREFIX)-1;
+        for (ldh = libdef_handlers; ldh->suffix != NULL; ldh++) {
+          size_t n, len = strlen(ldh->suffix);
+          if (!strncmp(p, ldh->suffix, len)) {
+            p += len;
+            n = ldh->stop ? strcspn(p, ldh->stop) : 0;
+            if (!p[n]) break;
+            p[n] = '\0';
+            ldh->func(ctx, p, ldh->arg);
+            p += n+1;
+            break;
+          }
+        }
+        if (ldh->suffix == NULL) {
+          buf[strlen(buf)-1] = '\0';
+          fprintf(stderr, "Error: unknown library definition tag %s%s\n",
+                  LIBDEF_PREFIX, p);
+          exit(1);
+        }
+      }
+    }
+    fclose(fp);
+    if (ctx->mode == BUILD_libdef) {
+      libdef_endmodule(ctx);
+    }
+  }
+
+  if (ctx->mode == BUILD_ffdef) {
+    fprintf(ctx->fp, "\n#undef FFDEF\n\n");
+    fprintf(ctx->fp,
+      "#ifndef FF_NUM_ASMFUNC\n#define FF_NUM_ASMFUNC %d\n#endif\n\n",
+      ffasmfunc);
+  } else if (ctx->mode == BUILD_vmdef) {
+    fprintf(ctx->fp, "}\n\n");
+  } else if (ctx->mode == BUILD_bcdef) {
+    int i;
+    fprintf(ctx->fp, "\n};\n\n");
+    fprintf(ctx->fp, "LJ_DATADEF const uint16_t lj_bc_mode[] = {\n");
+    fprintf(ctx->fp, "BCDEF(BCMODE)\n");
+    for (i = ffasmfunc-1; i > 0; i--)
+      fprintf(ctx->fp, "BCMODE_FF,\n");
+    fprintf(ctx->fp, "BCMODE_FF\n};\n\n");
+  } else if (ctx->mode == BUILD_recdef) {
+    char *p = (char *)obuf;
+    fprintf(ctx->fp, "\n};\n\n");
+    fprintf(ctx->fp, "static const RecordFunc recff_func[] = {\n"
+            "recff_nyi,\n"
+            "recff_c");
+    while (*p) {
+      fprintf(ctx->fp, ",\nrecff_%s", p);
+      p += strlen(p)+1;
+    }
+    fprintf(ctx->fp, "\n};\n\n");
+  }
+}
+
diff --git a/src/host/buildvm_peobj.c b/src/host/buildvm_peobj.c
new file mode 100644
index 00000000..17b3293a
--- /dev/null
+++ b/src/host/buildvm_peobj.c
@@ -0,0 +1,352 @@
+/*
+** LuaJIT VM builder: PE object emitter.
+** Copyright (C) 2005-2012 Mike Pall. See Copyright Notice in luajit.h
+**
+** Only used for building on Windows, since we cannot assume the presence
+** of a suitable assembler. The host and target byte order must match.
+*/
+
+#include "buildvm.h"
+#include "lj_bc.h"
+
+#if LJ_TARGET_X86ORX64
+
+/* Context for PE object emitter. */
+static char *strtab;
+static size_t strtabofs;
+
+/* -- PE object definitions ----------------------------------------------- */
+
+/* PE header. */
+typedef struct PEheader {
+  uint16_t arch;
+  uint16_t nsects;
+  uint32_t time;
+  uint32_t symtabofs;
+  uint32_t nsyms;
+  uint16_t opthdrsz;
+  uint16_t flags;
+} PEheader;
+
+/* PE section. */
+typedef struct PEsection {
+  char name[8];
+  uint32_t vsize;
+  uint32_t vaddr;
+  uint32_t size;
+  uint32_t ofs;
+  uint32_t relocofs;
+  uint32_t lineofs;
+  uint16_t nreloc;
+  uint16_t nline;
+  uint32_t flags;
+} PEsection;
+
+/* PE relocation. */
+typedef struct PEreloc {
+  uint32_t vaddr;
+  uint32_t symidx;
+  uint16_t type;
+} PEreloc;
+
+/* Cannot use sizeof, because it pads up to the max. alignment. */
+#define PEOBJ_RELOC_SIZE        (4+4+2)
+
+/* PE symbol table entry. */
+typedef struct PEsym {
+  union {
+    char name[8];
+    uint32_t nameref[2];
+  } n;
+  uint32_t value;
+  int16_t sect;
+  uint16_t type;
+  uint8_t scl;
+  uint8_t naux;
+} PEsym;
+
+/* PE symbol table auxiliary entry for a section. */
+typedef struct PEsymaux {
+  uint32_t size;
+  uint16_t nreloc;
+  uint16_t nline;
+  uint32_t cksum;
+  uint16_t assoc;
+  uint8_t comdatsel;
+  uint8_t unused[3];
+} PEsymaux;
+
+/* Cannot use sizeof, because it pads up to the max. alignment. */
+#define PEOBJ_SYM_SIZE  (8+4+2+2+1+1)
+
+/* PE object CPU specific defines. */
+#if LJ_TARGET_X86
+#define PEOBJ_ARCH_TARGET       0x014c
+#define PEOBJ_RELOC_REL32       0x14  /* MS: REL32, GNU: DISP32. */
+#define PEOBJ_RELOC_DIR32       0x06
+#elif LJ_TARGET_X64
+#define PEOBJ_ARCH_TARGET       0x8664
+#define PEOBJ_RELOC_REL32       0x04  /* MS: REL32, GNU: DISP32. */
+#define PEOBJ_RELOC_DIR32       0x02
+#define PEOBJ_RELOC_ADDR32NB    0x03
+#endif
+
+/* Section numbers (0-based). */
+enum {
+  PEOBJ_SECT_ABS = -2,
+  PEOBJ_SECT_UNDEF = -1,
+  PEOBJ_SECT_TEXT,
+#if LJ_TARGET_X64
+  PEOBJ_SECT_PDATA,
+  PEOBJ_SECT_XDATA,
+#endif
+  PEOBJ_SECT_RDATA_Z,
+  PEOBJ_NSECTIONS
+};
+
+/* Symbol types. */
+#define PEOBJ_TYPE_NULL         0
+#define PEOBJ_TYPE_FUNC         0x20
+
+/* Symbol storage class. */
+#define PEOBJ_SCL_EXTERN        2
+#define PEOBJ_SCL_STATIC        3
+
+/* -- PE object emitter --------------------------------------------------- */
+
+/* Emit PE object symbol. */
+static void emit_peobj_sym(BuildCtx *ctx, const char *name, uint32_t value,
+                           int sect, int type, int scl)
+{
+  PEsym sym;
+  size_t len = strlen(name);
+  if (!strtab) {  /* Pass 1: only calculate string table length. */
+    if (len > 8) strtabofs += len+1;
+    return;
+  }
+  if (len <= 8) {
+    memcpy(sym.n.name, name, len);
+    memset(sym.n.name+len, 0, 8-len);
+  } else {
+    sym.n.nameref[0] = 0;
+    sym.n.nameref[1] = (uint32_t)strtabofs;
+    memcpy(strtab + strtabofs, name, len);
+    strtab[strtabofs+len] = 0;
+    strtabofs += len+1;
+  }
+  sym.value = value;
+  sym.sect = (int16_t)(sect+1);  /* 1-based section number. */
+  sym.type = (uint16_t)type;
+  sym.scl = (uint8_t)scl;
+  sym.naux = 0;
+  owrite(ctx, &sym, PEOBJ_SYM_SIZE);
+}
+
+/* Emit PE object section symbol. */
+static void emit_peobj_sym_sect(BuildCtx *ctx, PEsection *pesect, int sect)
+{
+  PEsym sym;
+  PEsymaux aux;
+  if (!strtab) return;  /* Pass 1: no output. */
+  memcpy(sym.n.name, pesect[sect].name, 8);
+  sym.value = 0;
+  sym.sect = (int16_t)(sect+1);  /* 1-based section number. */
+  sym.type = PEOBJ_TYPE_NULL;
+  sym.scl = PEOBJ_SCL_STATIC;
+  sym.naux = 1;
+  owrite(ctx, &sym, PEOBJ_SYM_SIZE);
+  memset(&aux, 0, sizeof(PEsymaux));
+  aux.size = pesect[sect].size;
+  aux.nreloc = pesect[sect].nreloc;
+  owrite(ctx, &aux, PEOBJ_SYM_SIZE);
+}
+
+/* Emit Windows PE object file. */
+void emit_peobj(BuildCtx *ctx)
+{
+  PEheader pehdr;
+  PEsection pesect[PEOBJ_NSECTIONS];
+  uint32_t sofs;
+  int i, nrsym;
+  union { uint8_t b; uint32_t u; } host_endian;
+
+  host_endian.u = 1;
+  if (host_endian.b != LJ_ENDIAN_SELECT(1, 0)) {
+    fprintf(stderr, "Error: different byte order for host and target\n");
+    exit(1);
+  }
+
+  sofs = sizeof(PEheader) + PEOBJ_NSECTIONS*sizeof(PEsection);
+
+  /* Fill in PE sections. */
+  memset(&pesect, 0, PEOBJ_NSECTIONS*sizeof(PEsection));
+  memcpy(pesect[PEOBJ_SECT_TEXT].name, ".text", sizeof(".text")-1);
+  pesect[PEOBJ_SECT_TEXT].ofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_TEXT].size = (uint32_t)ctx->codesz);
+  pesect[PEOBJ_SECT_TEXT].relocofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_TEXT].nreloc = (uint16_t)ctx->nreloc) * PEOBJ_RELOC_SIZE;
+  /* Flags: 60 = read+execute, 50 = align16, 20 = code. */
+  pesect[PEOBJ_SECT_TEXT].flags = 0x60500020;
+
+#if LJ_TARGET_X64
+  memcpy(pesect[PEOBJ_SECT_PDATA].name, ".pdata", sizeof(".pdata")-1);
+  pesect[PEOBJ_SECT_PDATA].ofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_PDATA].size = 6*4);
+  pesect[PEOBJ_SECT_PDATA].relocofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_PDATA].nreloc = 6) * PEOBJ_RELOC_SIZE;
+  /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
+  pesect[PEOBJ_SECT_PDATA].flags = 0x40300040;
+
+  memcpy(pesect[PEOBJ_SECT_XDATA].name, ".xdata", sizeof(".xdata")-1);
+  pesect[PEOBJ_SECT_XDATA].ofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_XDATA].size = 8*2+4+6*2);  /* See below. */
+  pesect[PEOBJ_SECT_XDATA].relocofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_XDATA].nreloc = 1) * PEOBJ_RELOC_SIZE;
+  /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
+  pesect[PEOBJ_SECT_XDATA].flags = 0x40300040;
+#endif
+
+  memcpy(pesect[PEOBJ_SECT_RDATA_Z].name, ".rdata$Z", sizeof(".rdata$Z")-1);
+  pesect[PEOBJ_SECT_RDATA_Z].ofs = sofs;
+  sofs += (pesect[PEOBJ_SECT_RDATA_Z].size = (uint32_t)strlen(ctx->dasm_ident)+1);
+  /* Flags: 40 = read, 30 = align4, 40 = initialized data. */
+  pesect[PEOBJ_SECT_RDATA_Z].flags = 0x40300040;
+
+  /* Fill in PE header. */
+  pehdr.arch = PEOBJ_ARCH_TARGET;
+  pehdr.nsects = PEOBJ_NSECTIONS;
+  pehdr.time = 0;  /* Timestamp is optional. */
+  pehdr.symtabofs = sofs;
+  pehdr.opthdrsz = 0;
+  pehdr.flags = 0;
+
+  /* Compute the size of the symbol table:
+  ** @feat.00 + nsections*2
+  ** + asm_start + nsym
+  ** + nrsym
+  */
+  nrsym = ctx->nrelocsym;
+  pehdr.nsyms = 1+PEOBJ_NSECTIONS*2 + 1+ctx->nsym + nrsym;
+#if LJ_TARGET_X64
+  pehdr.nsyms += 1;  /* Symbol for lj_err_unwind_win64. */
+#endif
+
+  /* Write PE object header and all sections. */
+  owrite(ctx, &pehdr, sizeof(PEheader));
+  owrite(ctx, &pesect, sizeof(PEsection)*PEOBJ_NSECTIONS);
+
+  /* Write .text section. */
+  owrite(ctx, ctx->code, ctx->codesz);
+  for (i = 0; i < ctx->nreloc; i++) {
+    PEreloc reloc;
+    reloc.vaddr = (uint32_t)ctx->reloc[i].ofs;
+    reloc.symidx = 1+2+ctx->reloc[i].sym;  /* Reloc syms are after .text sym. */
+    reloc.type = ctx->reloc[i].type ? PEOBJ_RELOC_REL32 : PEOBJ_RELOC_DIR32;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+  }
+
+#if LJ_TARGET_X64
+  { /* Write .pdata section. */
+    uint32_t fcofs = (uint32_t)ctx->sym[ctx->nsym-1].ofs;
+    uint32_t pdata[3];  /* Start of .text, end of .text and .xdata. */
+    PEreloc reloc;
+    pdata[0] = 0; pdata[1] = fcofs; pdata[2] = 0;
+    owrite(ctx, &pdata, sizeof(pdata));
+    pdata[0] = fcofs; pdata[1] = (uint32_t)ctx->codesz; pdata[2] = 20;
+    owrite(ctx, &pdata, sizeof(pdata));
+    reloc.vaddr = 0; reloc.symidx = 1+2+nrsym+2+2+1;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 4; reloc.symidx = 1+2+nrsym+2+2+1;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 8; reloc.symidx = 1+2+nrsym+2;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 12; reloc.symidx = 1+2+nrsym+2+2+1;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 16; reloc.symidx = 1+2+nrsym+2+2+1;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+    reloc.vaddr = 20; reloc.symidx = 1+2+nrsym+2;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+  }
+  { /* Write .xdata section. */
+    uint16_t xdata[8+2+6];
+    PEreloc reloc;
+    xdata[0] = 0x01|0x08|0x10;  /* Ver. 1, uhandler/ehandler, prolog size 0. */
+    xdata[1] = 0x0005;  /* Number of unwind codes, no frame pointer. */
+    xdata[2] = 0x4200;  /* Stack offset 4*8+8 = aword*5. */
+    xdata[3] = 0x3000;  /* Push rbx. */
+    xdata[4] = 0x6000;  /* Push rsi. */
+    xdata[5] = 0x7000;  /* Push rdi. */
+    xdata[6] = 0x5000;  /* Push rbp. */
+    xdata[7] = 0;  /* Alignment. */
+    xdata[8] = xdata[9] = 0;  /* Relocated address of exception handler. */
+    xdata[10] = 0x01;  /* Ver. 1, no handler, prolog size 0. */
+    xdata[11] = 0x1504;  /* Number of unwind codes, fp = rbp, fpofs = 16. */
+    xdata[12] = 0x0300;  /* set_fpreg. */
+    xdata[13] = 0x0200;  /* stack offset 0*8+8 = aword*1. */
+    xdata[14] = 0x3000;  /* Push rbx. */
+    xdata[15] = 0x5000;  /* Push rbp. */
+    owrite(ctx, &xdata, sizeof(xdata));
+    reloc.vaddr = 2*8; reloc.symidx = 1+2+nrsym+2+2;
+    reloc.type = PEOBJ_RELOC_ADDR32NB;
+    owrite(ctx, &reloc, PEOBJ_RELOC_SIZE);
+  }
+#endif
+
+  /* Write .rdata$Z section. */
+  owrite(ctx, ctx->dasm_ident, strlen(ctx->dasm_ident)+1);
+
+  /* Write symbol table. */
+  strtab = NULL;  /* 1st pass: collect string sizes. */
+  for (;;) {
+    strtabofs = 4;
+    /* Mark as SafeSEH compliant. */
+    emit_peobj_sym(ctx, "@feat.00", 1,
+                   PEOBJ_SECT_ABS, PEOBJ_TYPE_NULL, PEOBJ_SCL_STATIC);
+
+    emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_TEXT);
+    for (i = 0; i < nrsym; i++)
+      emit_peobj_sym(ctx, ctx->relocsym[i], 0,
+                     PEOBJ_SECT_UNDEF, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
+
+#if LJ_TARGET_X64
+    emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_PDATA);
+    emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_XDATA);
+    emit_peobj_sym(ctx, "lj_err_unwind_win64", 0,
+                   PEOBJ_SECT_UNDEF, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
+#endif
+
+    emit_peobj_sym(ctx, ctx->beginsym, 0,
+                   PEOBJ_SECT_TEXT, PEOBJ_TYPE_NULL, PEOBJ_SCL_EXTERN);
+    for (i = 0; i < ctx->nsym; i++)
+      emit_peobj_sym(ctx, ctx->sym[i].name, (uint32_t)ctx->sym[i].ofs,
+                     PEOBJ_SECT_TEXT, PEOBJ_TYPE_FUNC, PEOBJ_SCL_EXTERN);
+
+    emit_peobj_sym_sect(ctx, pesect, PEOBJ_SECT_RDATA_Z);
+
+    if (strtab)
+      break;
+    /* 2nd pass: alloc strtab, write syms and copy strings. */
+    strtab = (char *)malloc(strtabofs);
+    *(uint32_t *)strtab = (uint32_t)strtabofs;
+  }
+
+  /* Write string table. */
+  owrite(ctx, strtab, strtabofs);
+}
+
+#else
+
+void emit_peobj(BuildCtx *ctx)
+{
+  UNUSED(ctx);
+  fprintf(stderr, "Error: no PE object support for this target\n");
+  exit(1);
+}
+
+#endif