guile/libguile/vm.c

/* Copyright 2001,2009-2015,2017-2020,2022-2023
     Free Software Foundation, Inc.

   This file is part of Guile.

   Guile is free software: you can redistribute it and/or modify it
   under the terms of the GNU Lesser General Public License as published
   by the Free Software Foundation, either version 3 of the License, or
   (at your option) any later version.

   Guile is distributed in the hope that it will be useful, but WITHOUT
   ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
   FITNESS FOR A PARTICULAR PURPOSE.  See the GNU Lesser General Public
   License for more details.

   You should have received a copy of the GNU Lesser General Public
   License along with Guile.  If not, see
   <https://www.gnu.org/licenses/>.  */

#if HAVE_CONFIG_H
#  include <config.h>
#endif

#include <alignof.h>
#include <alloca.h>
#include <errno.h>
#include <math.h>
#include <stdint.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <unistd.h>

#ifdef HAVE_SYS_MMAN_H
#include <sys/mman.h>
#endif

#include "alist.h"
#include "async.h"
#include "atomic.h"
#include "atomics-internal.h"
#include "bdw-gc.h"
#include "cache-internal.h"
#include "continuations.h"
#include "control.h"
#include "dynwind.h"
#include "eval.h"
#include "extensions.h"
#include "foreign.h"
#include "frames.h"
#include "gc-inline.h"
#include "gsubr.h"
#include "hooks.h"
#include "instructions.h"
#include "intrinsics.h"
#include "jit.h"
#include "keywords.h"
#include "list.h"
#include "loader.h"
#include "modules.h"
#include "numbers.h"
#include "pairs.h"
#include "ports.h"
#include "procprop.h"
#include "programs.h"
#include "simpos.h"
#include "smob.h"
#include "stackchk.h"
#include "symbols.h"
#include "values.h"
#include "vectors.h"
#include "version.h"
#include "vm-builtins.h"

#include "vm.h"

#include <gc/gc_mark.h>

#if (defined __GNUC__)
# define SCM_NOINLINE __attribute__ ((__noinline__))
#else
# define SCM_NOINLINE /* noinline */
#endif

#include "introspector.h"

static int vm_default_engine = SCM_VM_REGULAR_ENGINE;

/* Unfortunately we can't snarf these: snarfed things are only loaded up from
   (system vm vm), which might not be loaded before an error happens. */
static SCM sym_keyword_argument_error;
static SCM sym_regular;
static SCM sym_debug;

/* The page size.  */
static size_t page_size;

/* The VM has a number of internal assertions that shouldn't normally be
   necessary, but might be if you think you found a bug in the VM. */
/* #define VM_ENABLE_ASSERTIONS */

static void vm_expand_stack (struct scm_vm *vp,
                             union scm_vm_stack_element *new_sp) SCM_NOINLINE;

/* RESTORE is for the case where we know we have done a PUSH of equal or
   greater stack size in the past.  Otherwise PUSH is the thing, which
   may expand the stack.  */
enum vm_increase_sp_kind { VM_SP_PUSH, VM_SP_RESTORE };

static inline void
vm_increase_sp (struct scm_vm *vp, union scm_vm_stack_element *new_sp,
                enum vm_increase_sp_kind kind)
{
  if (kind == VM_SP_PUSH && new_sp < vp->stack_limit)
    vm_expand_stack (vp, new_sp);
  else
    vp->sp = new_sp;
}

static inline void
vm_push_sp (struct scm_vm *vp, union scm_vm_stack_element *new_sp)
{
  vm_increase_sp (vp, new_sp, VM_SP_PUSH);
}

static inline void
vm_restore_sp (struct scm_vm *vp, union scm_vm_stack_element *new_sp)
{
  vm_increase_sp (vp, new_sp, VM_SP_RESTORE);
}


/*
 * VM Continuation
 */

void
scm_i_vm_cont_print (SCM x, SCM port, scm_print_state *pstate)
{
  scm_puts ("#<vm-continuation ", port);
  scm_uintprint (SCM_UNPACK (x), 16, port);
  scm_puts (">", port);
}

int
scm_i_vm_cont_to_frame (SCM cont, struct scm_frame *frame)
{
  struct scm_vm_cont *data = SCM_VM_CONT_DATA (cont);

  frame->stack_holder = data;
  frame->fp_offset = data->fp_offset;
  frame->sp_offset = data->stack_size;
  frame->ip = data->vra;

  return 1;
}

/* Ideally we could avoid copying the C stack if the continuation root
   is inside VM code, and call/cc was invoked within that same call to
   vm_run.  That's currently not implemented.  */
static SCM
capture_stack (union scm_vm_stack_element *stack_top,
               union scm_vm_stack_element *fp,
               union scm_vm_stack_element *sp,
               uint32_t *vra,
               uint8_t *mra,
               scm_t_dynstack *dynstack, uint32_t flags)
{
  struct scm_vm_cont *p;
  size_t stack_size;

  stack_size = stack_top - sp;

  /* Allocate the 'scm_vm_cont' struct and the stack at once.  That way,
     keeping a pointer to 'p->stack_bottom' around won't retain it.
     See <https://bugs.gnu.org/59021>.  */
  p = scm_gc_malloc (sizeof (*p) + stack_size * sizeof (*p->stack_bottom),
                     "capture_vm_cont");

  p->stack_size = stack_size;
  p->stack_bottom = (void *) ((char *) p + sizeof (*p));
  p->vra = vra;
  p->mra = mra;
  p->fp_offset = stack_top - fp;
  memcpy (p->stack_bottom, sp, p->stack_size * sizeof (*p->stack_bottom));
  p->dynstack = dynstack;
  p->flags = flags;
  return scm_cell (scm_tc7_vm_cont, (scm_t_bits) p);
}

SCM
scm_i_capture_current_stack (void)
{
  scm_thread *thread;
  struct scm_vm *vp;

  thread = SCM_I_CURRENT_THREAD;
  vp = &thread->vm;

  return capture_stack (vp->stack_top, vp->fp, vp->sp, vp->ip, NULL,
                        scm_dynstack_capture_all (&thread->dynstack),
                        0);
}

#define FOR_EACH_HOOK(M) \
  M(apply) \
  M(return) \
  M(next) \
  M(abort)

static void
vm_hook_compute_enabled (scm_thread *thread, SCM hook, uint8_t *enabled)
{
  if (thread->vm.trace_level <= 0
      || thread->vm.engine == SCM_VM_REGULAR_ENGINE
      || scm_is_false (hook)
      || scm_is_true (scm_hook_empty_p (hook)))
    *enabled = 0;
  else
    *enabled = 1;
}

static void
vm_recompute_disable_mcode (scm_thread *thread)
{
#if ENABLE_JIT
  /* FIXME: Some of this logic works for ahead-of-time compilation
     too.  */

  uint8_t was_disabled = thread->vm.disable_mcode;
  thread->vm.disable_mcode = 0;

#define DISABLE_MCODE_IF_HOOK_ENABLED(h) \
  if (thread->vm.h##_hook_enabled)       \
    thread->vm.disable_mcode = 1;
  FOR_EACH_HOOK (DISABLE_MCODE_IF_HOOK_ENABLED);
#undef DISABLE_MCODE_IF_HOOK_ENABLED

  if (thread->vm.disable_mcode && !was_disabled)
    scm_jit_clear_mcode_return_addresses (thread);
#endif
}

static int
set_vm_trace_level (scm_thread *thread, int level)
{
  int old_level;
  struct scm_vm *vp = &thread->vm;

  old_level = vp->trace_level;
  vp->trace_level = level;
  vp->disable_mcode = 0;
#define RESET_LEVEL(h) \
  vm_hook_compute_enabled (thread, vp->h##_hook, &vp->h##_hook_enabled);
  FOR_EACH_HOOK (RESET_LEVEL);
#undef RESET_LEVEL
  vm_recompute_disable_mcode (thread);

  return old_level;
}

/* Return the first integer greater than or equal to LEN such that
   LEN % ALIGN == 0.  Return LEN if ALIGN is zero.  */
#define ROUND_UP(len, align)					\
  ((align) ? (((len) - 1UL) | ((align) - 1UL)) + 1UL : (len))

static void
invoke_hook (scm_thread *thread, SCM hook)
{
  struct scm_vm *vp = &thread->vm;
  struct scm_frame c_frame;
  scm_t_cell *frame;
  SCM scm_frame;
  int saved_trace_level;
  uint8_t saved_compare_result;

  if (scm_is_false (hook) || scm_is_null (SCM_HOOK_PROCEDURES (hook)))
    return;

  saved_trace_level = set_vm_trace_level (thread, 0);
  saved_compare_result = vp->compare_result;

  /* Allocate a frame object on the stack.  This is more efficient than calling
     `scm_c_make_frame ()' to allocate on the heap, but it forces hooks to not
     capture frame objects.

     At the same time, procedures such as `frame-procedure' make sense only
     while the stack frame represented by the frame object is visible, so it
     seems reasonable to limit the lifetime of frame objects.  */

  c_frame.stack_holder = vp;
  c_frame.fp_offset = vp->stack_top - vp->fp;
  c_frame.sp_offset = vp->stack_top - vp->sp;
  c_frame.ip = vp->ip;

  /* Arrange for FRAME to be 8-byte aligned, like any other cell.  */
  frame = alloca (sizeof (*frame) + 8);
  frame = (scm_t_cell *) ROUND_UP ((uintptr_t) frame, 8UL);

  frame->word_0 = SCM_PACK (scm_tc7_frame | (SCM_VM_FRAME_KIND_VM << 8));
  frame->word_1 = SCM_PACK_POINTER (&c_frame);

  scm_frame = SCM_PACK_POINTER (frame);
  scm_c_run_hookn (hook, &scm_frame, 1);

  vp->compare_result = saved_compare_result;
  set_vm_trace_level (thread, saved_trace_level);
}

#define DEFINE_INVOKE_HOOK(h) \
  static void                                             \
  invoke_##h##_hook (scm_thread *thread) SCM_NOINLINE;    \
  static void                                             \
  invoke_##h##_hook (scm_thread *thread)                  \
  {                                                       \
    if (thread->vm.h##_hook_enabled)                      \
      return invoke_hook (thread, thread->vm.h##_hook);   \
  }

FOR_EACH_HOOK (DEFINE_INVOKE_HOOK)

#undef DEFINE_INVOKE_HOOK


/*
 * VM Error Handling
 */


static void vm_error_bad_instruction (uint32_t inst) SCM_NORETURN SCM_NOINLINE;

static void
vm_error_bad_instruction (uint32_t inst)
{
  fprintf (stderr, "VM: Bad instruction: %x\n", inst);
  abort ();
}




static SCM vm_boot_continuation;

#define DECLARE_BUILTIN(builtin, BUILTIN, req, opt, rest)               \
  static SCM vm_builtin_##builtin;                                      \
  static uint32_t *vm_builtin_##builtin##_code;
FOR_EACH_VM_BUILTIN (DECLARE_BUILTIN)
#undef DECLARE_BUILTIN

static const uint32_t vm_boot_continuation_code[] = {
  SCM_PACK_OP_24 (halt, 0)
};

int
scm_i_vm_is_boot_continuation_code (uint32_t *ip)
{
  return ip == vm_boot_continuation_code;
}

SCM
scm_vm_builtin_ref (unsigned idx)
{
  switch (idx)
    {
#define INDEX_TO_NAME(builtin, BUILTIN, req, opt, rest)                 \
      case SCM_VM_BUILTIN_##BUILTIN: return vm_builtin_##builtin;
      FOR_EACH_VM_BUILTIN(INDEX_TO_NAME)
#undef INDEX_TO_NAME
      default: abort();
    }
}

SCM scm_sym_apply;
static SCM scm_sym_values;
static SCM scm_sym_abort_to_prompt;
static SCM scm_sym_call_with_values;
static SCM scm_sym_call_with_current_continuation;

SCM
scm_vm_builtin_name_to_index (SCM name)
#define FUNC_NAME "builtin-name->index"
{
  SCM_VALIDATE_SYMBOL (1, name);

#define NAME_TO_INDEX(builtin, BUILTIN, req, opt, rest) \
  if (scm_is_eq (name, scm_sym_##builtin))              \
    return scm_from_uint (SCM_VM_BUILTIN_##BUILTIN);
  FOR_EACH_VM_BUILTIN(NAME_TO_INDEX)
#undef NAME_TO_INDEX

  return SCM_BOOL_F;
}
#undef FUNC_NAME

SCM
scm_vm_builtin_index_to_name (SCM index)
#define FUNC_NAME "builtin-index->name"
{
  unsigned idx;

  SCM_VALIDATE_UINT_COPY (1, index, idx);

  switch (idx)
    {
#define INDEX_TO_NAME(builtin, BUILTIN, req, opt, rest)         \
      case SCM_VM_BUILTIN_##BUILTIN: return scm_sym_##builtin;
      FOR_EACH_VM_BUILTIN(INDEX_TO_NAME)
#undef INDEX_TO_NAME
      default: return SCM_BOOL_F;
    }
}
#undef FUNC_NAME

static void
scm_init_vm_builtins (void)
{
  scm_c_define_gsubr ("builtin-name->index", 1, 0, 0,
                      scm_vm_builtin_name_to_index);
  scm_c_define_gsubr ("builtin-index->name", 1, 0, 0,
                      scm_vm_builtin_index_to_name);
}

static uint32_t*
instrumented_code (const uint32_t *code, size_t byte_size)
{
  uint32_t *ret, *write;
  ret = scm_i_alloc_primitive_code_with_instrumentation (byte_size / 4, &write);
  memcpy (write, code, byte_size);
  return ret;
}

static void
define_vm_builtins (void)
{
  const uint32_t apply_code[] = {
    SCM_PACK_OP_24 (assert_nargs_ge, 3),
    SCM_PACK_OP_12_12 (shuffle_down, 1, 0),
    SCM_PACK_OP_24 (expand_apply_argument, 0),
    SCM_PACK_OP_24 (tail_call, 0),
  };

  const uint32_t values_code[] = {
    SCM_PACK_OP_12_12 (shuffle_down, 1, 0),
    SCM_PACK_OP_24 (return_values, 0)
  };

  const uint32_t abort_to_prompt_code[] = {
    SCM_PACK_OP_24 (assert_nargs_ge, 2),
    SCM_PACK_OP_24 (abort, 0), /* tag in r1, vals from r2 */
    /* FIXME: Partial continuation should capture caller regs.  */
    SCM_PACK_OP_24 (return_values, 0) /* vals from r0 */
  };

  const uint32_t call_with_values_code[] = {
    SCM_PACK_OP_24 (assert_nargs_ee, 3),
    SCM_PACK_OP_24 (alloc_frame, 6),
    SCM_PACK_OP_12_12 (mov, 0, 4),
    SCM_PACK_OP_12_12 (mov, 4, 3),
    SCM_PACK_OP_24 (call, 5), SCM_PACK_OP_ARG_8_24 (0, 1),
    SCM_PACK_OP_24 (long_fmov, 0), SCM_PACK_OP_ARG_8_24 (0, 1),
    SCM_PACK_OP_12_12 (shuffle_down, 5, 1),
    SCM_PACK_OP_24 (tail_call, 0)
  };

  const uint32_t call_with_current_continuation_code[] = {
    SCM_PACK_OP_24 (assert_nargs_ee, 2),
    SCM_PACK_OP_12_12 (mov, 1, 0),
    SCM_PACK_OP_24 (capture_continuation, 0),
    SCM_PACK_OP_24 (tail_call, 0)
  };

  /* This one isn't exactly a builtin but we still handle it here.  */
  const uint32_t handle_interrupt_code[] = {
    SCM_PACK_OP_24 (alloc_frame, 4),
    SCM_PACK_OP_12_12 (mov, 0, 3),
    SCM_PACK_OP_24 (call, 3), SCM_PACK_OP_ARG_8_24 (0, 1),
    SCM_PACK_OP_24 (return_from_interrupt, 0)
  };

#define DEFINE_BUILTIN(builtin, BUILTIN, req, opt, rest)                \
  {                                                                     \
    size_t sz = sizeof (builtin##_code);                                \
    vm_builtin_##builtin##_code = instrumented_code (builtin##_code, sz); \
    vm_builtin_##builtin =                                              \
      scm_cell (scm_tc7_program | SCM_F_PROGRAM_IS_PRIMITIVE,           \
                (scm_t_bits)vm_builtin_##builtin##_code);               \
  }
  FOR_EACH_VM_BUILTIN (DEFINE_BUILTIN);
#undef INDEX_TO_NAME

  scm_vm_intrinsics.handle_interrupt_code =
    instrumented_code (handle_interrupt_code, sizeof (handle_interrupt_code));
}

SCM
scm_i_call_with_current_continuation (SCM proc)
{
  return scm_call_1 (vm_builtin_call_with_current_continuation, proc);
}


/*
 * VM
 */

#define VM_NAME vm_regular_engine
#define VM_USE_HOOKS 0
#define FUNC_NAME "vm-regular-engine"
#include "vm-engine.c"
#undef FUNC_NAME
#undef VM_USE_HOOKS
#undef VM_NAME

#define VM_NAME vm_debug_engine
#define VM_USE_HOOKS 1
#define FUNC_NAME "vm-debug-engine"
#include "vm-engine.c"
#undef FUNC_NAME
#undef VM_USE_HOOKS
#undef VM_NAME

typedef SCM (*scm_t_vm_engine) (scm_thread *current_thread);

static const scm_t_vm_engine vm_engines[SCM_VM_NUM_ENGINES] =
  { vm_regular_engine, vm_debug_engine };

static union scm_vm_stack_element*
allocate_stack (size_t size)
{
  void *ret;

  if (size >= ((size_t) -1) / sizeof (union scm_vm_stack_element))
    abort ();

  size *= sizeof (union scm_vm_stack_element);

#if HAVE_SYS_MMAN_H
  ret = mmap (NULL, size, PROT_READ | PROT_WRITE,
              MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
  if (ret == NULL)
    /* Shouldn't happen.  */
    abort ();
  if (ret == MAP_FAILED)
    ret = NULL;
#else
  ret = malloc (size);
#endif

  if (!ret)
    perror ("allocate_stack failed");

  return (union scm_vm_stack_element *) ret;
}

static void
free_stack (union scm_vm_stack_element *stack, size_t size)
{
  size *= sizeof (*stack);

#if HAVE_SYS_MMAN_H
  munmap (stack, size);
#else
  free (stack);
#endif
}

/* Ideally what we would like is an mremap or a realloc that grows at
   the bottom, not the top.  Oh well; mmap and memcpy are fast enough,
   considering that they run very infrequently.  */
static union scm_vm_stack_element*
expand_stack (union scm_vm_stack_element *old_bottom, size_t old_size,
              size_t new_size)
#define FUNC_NAME "expand_stack"
{
  union scm_vm_stack_element *new_bottom;
  size_t extension_size;

  if (new_size >= ((size_t) -1) / sizeof (union scm_vm_stack_element))
    abort ();
  if (new_size <= old_size)
    abort ();

  extension_size = new_size - old_size;

  if ((size_t)old_bottom < extension_size * sizeof (union scm_vm_stack_element))
    abort ();

  new_bottom = allocate_stack (new_size);

  if (!new_bottom)
    return NULL;

  memcpy (new_bottom + extension_size,
          old_bottom,
          old_size * sizeof (union scm_vm_stack_element));
  free_stack (old_bottom, old_size);

  return new_bottom;
}
#undef FUNC_NAME

void
scm_i_vm_prepare_stack (struct scm_vm *vp)
{
  /* Not racey, as this will be run the first time a thread enters
     Guile.  */
  if (page_size == 0)
    {
      page_size = getpagesize ();
      /* page_size should be a power of two.  */
      if (page_size & (page_size - 1))
        abort ();
    }

  vp->stack_size = page_size / sizeof (union scm_vm_stack_element);
  vp->stack_bottom = allocate_stack (vp->stack_size);
  if (!vp->stack_bottom)
    /* As in expand_stack, we don't have any way to throw an exception
       if we can't allocate one measely page -- there's no stack to
       handle it.  For now, abort.  */
    abort ();
  vp->stack_top = vp->stack_bottom + vp->stack_size;
  vp->stack_limit = vp->stack_bottom;
  vp->overflow_handler_stack = SCM_EOL;
  vp->ip = NULL;
  vp->sp = vp->stack_top;
  vp->fp = vp->stack_top;
  vp->compare_result = SCM_F_COMPARE_NONE;
  vp->engine = vm_default_engine;
  vp->trace_level = 1;
#define INIT_HOOK(h) vp->h##_hook = SCM_BOOL_F;
  FOR_EACH_HOOK (INIT_HOOK)
#undef INIT_HOOK
}

static void
return_unused_stack_to_os (struct scm_vm *vp)
{
#if HAVE_SYS_MMAN_H
  uintptr_t lo = (uintptr_t) vp->stack_bottom;
  uintptr_t hi = (uintptr_t) vp->sp;

  lo &= ~(page_size - 1U); /* round down */
  hi &= ~(page_size - 1U); /* round down */

  /* Return these pages to the OS.  The next time they are paged in,
     they will be zeroed.  */
  if (lo < hi)
    {
      int ret = 0;

      do
        ret = madvise ((void *) lo, hi - lo, MADV_DONTNEED);
      while (ret && errno == EAGAIN);

      /* If the OS doesn't implement 'madvise' (as is currently the case
         for GNU/Hurd), don't warn the user since there's nothing they
         can do about it.  */
      if (ret && errno != ENOSYS)
        perror ("madvise failed");
    }
#endif
}

#define SLOT_MAP_CACHE_SIZE 32U
struct slot_map_cache_entry
{
  uint32_t *ip;
  const uint8_t *map;
};

struct slot_map_cache
{
  struct slot_map_cache_entry entries[SLOT_MAP_CACHE_SIZE];
};

static const uint8_t *
find_slot_map (uint32_t *ip, struct slot_map_cache *cache)
{
  /* The lower two bits should be zero.  FIXME: Use a better hash
     function; we don't expose scm_raw_hashq currently.  */
  size_t slot = (((uintptr_t) ip) >> 2) % SLOT_MAP_CACHE_SIZE;
  const uint8_t *map;

  if (cache->entries[slot].ip == ip)
    map = cache->entries[slot].map;
  else
    {
      map = scm_find_slot_map_unlocked (ip);
      cache->entries[slot].ip = ip;
      cache->entries[slot].map = map;
    }

  return map;
}

enum slot_desc
  {
    SLOT_DESC_DEAD = 0,
    SLOT_DESC_LIVE_RAW = 1,
    SLOT_DESC_LIVE_GC = 2,
    SLOT_DESC_UNUSED = 3
  };

/* Mark the active VM stack region.  */
struct GC_ms_entry *
scm_i_vm_mark_stack (struct scm_vm *vp, struct GC_ms_entry *mark_stack_ptr,
                     struct GC_ms_entry *mark_stack_limit)
{
  union scm_vm_stack_element *sp, *fp;
  /* The first frame will be marked conservatively (without a slot map).
     This is because GC can happen at any point within the hottest
     activation, due to multiple threads or per-instruction hooks, and
     providing slot maps for all points in a program would take a
     prohibitive amount of space.  */
  const uint8_t *slot_map = NULL;
  void *upper = (void *) GC_greatest_plausible_heap_addr;
  void *lower = (void *) GC_least_plausible_heap_addr;
  struct slot_map_cache cache;

  memset (&cache, 0, sizeof (cache));

  for (fp = vp->fp, sp = vp->sp;
       fp < vp->stack_top;
       fp = SCM_FRAME_DYNAMIC_LINK (fp))
    {
      ptrdiff_t nlocals = SCM_FRAME_NUM_LOCALS (fp, sp);
      size_t slot = nlocals - 1;
      for (slot = nlocals - 1; sp < fp; sp++, slot--)
        {
          enum slot_desc desc = SLOT_DESC_LIVE_GC;

          if (slot_map)
            desc = (slot_map[slot / 4U] >> ((slot % 4U) * 2)) & 3U;

          switch (desc)
            {
            case SLOT_DESC_LIVE_RAW:
              break;
            case SLOT_DESC_UNUSED:
            case SLOT_DESC_LIVE_GC:
              if (SCM_NIMP (sp->as_scm) &&
                  sp->as_ptr >= lower && sp->as_ptr <= upper)
                mark_stack_ptr = GC_mark_and_push (sp->as_ptr,
                                                   mark_stack_ptr,
                                                   mark_stack_limit,
                                                   NULL);
              break;
            case SLOT_DESC_DEAD:
              /* This value may become dead as a result of GC,
                 so we can't just leave it on the stack.  */
              sp->as_scm = SCM_UNSPECIFIED;
              break;
            }
        }
      sp = SCM_FRAME_PREVIOUS_SP (fp);
      /* Inner frames may have a dead slots map for precise marking.
         Note that there may be other reasons to not have a dead slots
         map, e.g. if all of the frame's slots below the callee frame
         are live.  */
      slot_map = find_slot_map (SCM_FRAME_VIRTUAL_RETURN_ADDRESS (fp), &cache);
    }

  return_unused_stack_to_os (vp);

  return mark_stack_ptr;
}

/* Free the VM stack, as this thread is exiting.  */
void
scm_i_vm_free_stack (struct scm_vm *vp)
{
  free_stack (vp->stack_bottom, vp->stack_size);
  /* Not strictly necessary, but good to avoid confusion when debugging
     thread-related GC issues.  */
  memset (vp, 0, sizeof (*vp));
}

struct vm_expand_stack_data
{
  struct scm_vm *vp;
  size_t stack_size;
  union scm_vm_stack_element *new_sp;
};

static void *
vm_expand_stack_inner (void *data_ptr)
{
  struct vm_expand_stack_data *data = data_ptr;

  struct scm_vm *vp = data->vp;
  union scm_vm_stack_element *old_top, *new_bottom;
  size_t new_size;
  ptrdiff_t reloc;

  old_top = vp->stack_top;
  new_size = vp->stack_size;
  while (new_size < data->stack_size)
    new_size *= 2;

  new_bottom = expand_stack (vp->stack_bottom, vp->stack_size, new_size);
  if (!new_bottom)
    return NULL;

  vp->stack_bottom = new_bottom;
  vp->stack_size = new_size;
  vp->stack_top = vp->stack_bottom + new_size;
  vp->stack_limit = vp->stack_bottom;
  reloc = vp->stack_top - old_top;

  if (vp->fp)
    vp->fp += reloc;
  data->new_sp += reloc;

  return new_bottom;
}

static ptrdiff_t
current_overflow_size (struct scm_vm *vp)
{
  if (scm_is_pair (vp->overflow_handler_stack))
    return scm_to_ptrdiff_t (scm_caar (vp->overflow_handler_stack));
  return -1;
}

static int
should_handle_stack_overflow (struct scm_vm *vp, ptrdiff_t stack_size)
{
  ptrdiff_t overflow_size = current_overflow_size (vp);
  return overflow_size >= 0 && stack_size >= overflow_size;
}

static void
reset_stack_limit (struct scm_vm *vp)
{
  if (should_handle_stack_overflow (vp, vp->stack_size))
    vp->stack_limit = vp->stack_top - current_overflow_size (vp);
  else
    vp->stack_limit = vp->stack_bottom;
}

struct overflow_handler_data
{
  struct scm_vm *vp;
  SCM overflow_handler_stack;
};

static void
wind_overflow_handler (void *ptr)
{
  struct overflow_handler_data *data = ptr;

  data->vp->overflow_handler_stack = data->overflow_handler_stack;

  reset_stack_limit (data->vp);
}

static void
unwind_overflow_handler (void *ptr)
{
  struct overflow_handler_data *data = ptr;

  data->vp->overflow_handler_stack = scm_cdr (data->overflow_handler_stack);

  reset_stack_limit (data->vp);
}

static void
vm_expand_stack (struct scm_vm *vp, union scm_vm_stack_element *new_sp)
{
  ptrdiff_t stack_size = vp->stack_top - new_sp;

  if (stack_size > vp->stack_size)
    {
      struct vm_expand_stack_data data;

      data.vp = vp;
      data.stack_size = stack_size;
      data.new_sp = new_sp;
      
      if (!GC_call_with_alloc_lock (vm_expand_stack_inner, &data))
        /* Throw an unwind-only exception.  */
        scm_report_stack_overflow ();

      new_sp = data.new_sp;
    }

  vp->sp = new_sp;

  if (should_handle_stack_overflow (vp, stack_size))
    {
      SCM more_stack, new_limit;

      struct overflow_handler_data data;
      data.vp = vp;
      data.overflow_handler_stack = vp->overflow_handler_stack;

      scm_dynwind_begin (SCM_F_DYNWIND_REWINDABLE);

      scm_dynwind_rewind_handler (unwind_overflow_handler, &data,
                                  SCM_F_WIND_EXPLICITLY);
      scm_dynwind_unwind_handler (wind_overflow_handler, &data,
                                  SCM_F_WIND_EXPLICITLY);

      /* Call the overflow handler.  */
      more_stack = scm_call_0 (scm_cdar (data.overflow_handler_stack));

      /* If the overflow handler returns, its return value should be an
         integral number of words from the outer stack limit to transfer
         to the inner limit.  */
      if (scm_to_ptrdiff_t (more_stack) <= 0)
        scm_out_of_range (NULL, more_stack);
      new_limit = scm_sum (scm_caar (data.overflow_handler_stack), more_stack);
      if (scm_is_pair (scm_cdr (data.overflow_handler_stack)))
        new_limit = scm_min (new_limit,
                             scm_caadr (data.overflow_handler_stack));

      /* Ensure the new limit is in range.  */
      scm_to_ptrdiff_t (new_limit);

      /* Increase the limit that we will restore.  */
      scm_set_car_x (scm_car (data.overflow_handler_stack), new_limit);

      scm_dynwind_end ();

      /* Recurse.  */
      return vm_expand_stack (vp, new_sp);
    }
}

static uint32_t
frame_locals_count (scm_thread *thread)
{
  return SCM_FRAME_NUM_LOCALS (thread->vm.fp, thread->vm.sp);
}

static void
thread_expand_stack (scm_thread *thread, union scm_vm_stack_element *new_sp)
{
  vm_expand_stack (&thread->vm, new_sp);
}

/* This duplicates the inlined "ALLOC_FRAME" macro from vm-engine.c, but
   it seems to be necessary for perf; the inlined version avoids the
   needs to flush IP in the common case.  */
static void
alloc_frame (scm_thread *thread, uint32_t nlocals)
{
  union scm_vm_stack_element *sp = thread->vm.fp - nlocals;

  if (SCM_UNLIKELY (sp < thread->vm.stack_limit))
    thread_expand_stack (thread, sp);
  else
    thread->vm.sp = sp;
}

static uint32_t
compute_kwargs_npositional (scm_thread *thread, uint32_t nreq, uint32_t nopt)
{
  uint32_t npositional, nargs;

  nargs = frame_locals_count (thread);

  /* look in optionals for first keyword or last positional */
  /* starting after the last required positional arg */
  npositional = nreq;
  while (/* while we have args */
         npositional < nargs
         /* and we still have positionals to fill */
         && npositional < nreq + nopt
         /* and we haven't reached a keyword yet */
         && !scm_is_keyword (SCM_FRAME_LOCAL (thread->vm.fp, npositional)))
    /* bind this optional arg (by leaving it in place) */
    npositional++;

  return npositional;
}

static void
bind_kwargs (scm_thread *thread, uint32_t npositional, uint32_t nlocals,
             SCM kwargs, uint8_t strict, uint8_t allow_other_keys)
{
  uint32_t nargs, nkw, n;
  union scm_vm_stack_element *fp;

  nargs = frame_locals_count (thread);
  nkw = nargs - npositional;

  /* shuffle non-positional arguments above nlocals */
  alloc_frame (thread, nlocals + nkw);

  fp = thread->vm.fp;
  n = nkw;
  while (n--)
    SCM_FRAME_LOCAL (fp, nlocals + n) = SCM_FRAME_LOCAL (fp, npositional + n);

  /* Fill optionals & keyword args with SCM_UNDEFINED */
  n = npositional;
  while (n < nlocals)
    SCM_FRAME_LOCAL (fp, n++) = SCM_UNDEFINED;

  /* Now bind keywords, in the order given.  */
  for (n = 0; n < nkw; n++)
    {
      SCM kw = SCM_FRAME_LOCAL (fp, nlocals + n);

      if (scm_is_keyword (kw))
        {
          SCM walk;
          for (walk = kwargs; scm_is_pair (walk); walk = SCM_CDR (walk))
            if (scm_is_eq (SCM_CAAR (walk), kw))
              {
                SCM si = SCM_CDAR (walk);
                if (n + 1 < nkw)
                  SCM_FRAME_LOCAL (fp, scm_to_uint32 (si)) =
                    SCM_FRAME_LOCAL (fp, nlocals + n + 1);
                else
                  scm_error_scm (sym_keyword_argument_error, SCM_BOOL_F,
                                 scm_from_latin1_string
                                 ("Keyword argument has no value"),
                                 SCM_EOL, scm_list_1 (kw));
                break;
              }
          if (!allow_other_keys && !scm_is_pair (walk))
            scm_error_scm (sym_keyword_argument_error, SCM_BOOL_F,
                           scm_from_latin1_string ("Unrecognized keyword"),
                           SCM_EOL, scm_list_1 (kw));
          n++;
        }
      else if (strict)
        {
          scm_error_scm (sym_keyword_argument_error, SCM_BOOL_F,
                         scm_from_latin1_string ("Invalid keyword"),
                         SCM_EOL, scm_list_1 (kw));
        }
      else
        {
          /* Ignore this argument.  It might get consed onto a rest list.  */
        }
    }
}

static SCM
cons_rest (scm_thread *thread, uint32_t base)
{
  SCM rest = SCM_EOL;
  uint32_t n = frame_locals_count (thread) - base;

  while (n--)
    rest = scm_inline_cons (thread, SCM_FRAME_LOCAL (thread->vm.fp, base + n),
                            rest);

  return rest;
}

static void
push_interrupt_frame (scm_thread *thread, uint8_t *mra)
{
  union scm_vm_stack_element *old_fp, *new_fp;
  size_t frame_overhead = 3;
  size_t old_frame_size = frame_locals_count (thread);
  SCM proc = scm_i_async_pop (thread);

#if ENABLE_JIT
  if (!mra)
    mra = scm_jit_return_to_interpreter_trampoline;
#endif

  /* Reserve space for frame and callee.  */
  alloc_frame (thread, old_frame_size + frame_overhead + 1);

  old_fp = thread->vm.fp;
  new_fp = SCM_FRAME_SLOT (old_fp, old_frame_size + frame_overhead - 1);
  SCM_FRAME_SET_DYNAMIC_LINK (new_fp, old_fp);
  /* Arrange to return to the same handle-interrupts opcode to handle
     any additional interrupts.  */
  SCM_FRAME_SET_VIRTUAL_RETURN_ADDRESS (new_fp, thread->vm.ip);
  SCM_FRAME_SET_MACHINE_RETURN_ADDRESS (new_fp, mra);
  SCM_FRAME_LOCAL (new_fp, 0) = proc;

  thread->vm.fp = new_fp;
}

struct return_to_continuation_data
{
  struct scm_vm_cont *cp;
  struct scm_vm *vp;
};

/* Called with the GC lock to prevent the stack marker from traversing a
   stack in an inconsistent state.  */
static void *
vm_return_to_continuation_inner (void *data_ptr)
{
  struct return_to_continuation_data *data = data_ptr;
  struct scm_vm *vp = data->vp;
  struct scm_vm_cont *cp = data->cp;

  /* We know that there is enough space for the continuation, because we
     captured it in the past.  However there may have been an expansion
     since the capture, so we may have to re-link the frame
     pointers.  */
  memcpy (vp->stack_top - cp->stack_size,
          cp->stack_bottom,
          cp->stack_size * sizeof (*cp->stack_bottom));
  vp->fp = vp->stack_top - cp->fp_offset;
  vm_restore_sp (vp, vp->stack_top - cp->stack_size);

  return NULL;
}

static void reinstate_continuation_x (scm_thread *thread, SCM cont) SCM_NORETURN;

static void
reinstate_continuation_x (scm_thread *thread, SCM cont)
{
  scm_t_contregs *continuation = scm_i_contregs (cont);
  struct scm_vm *vp = &thread->vm;
  struct scm_vm_cont *cp;
  size_t n, i, frame_overhead = 3;
  union scm_vm_stack_element *argv;
  struct return_to_continuation_data data;

  if (!scm_is_eq (continuation->root, thread->continuation_root))
    scm_misc_error
      ("%continuation-call",
       "invoking continuation would cross continuation barrier: ~A",
       scm_list_1 (cont));

  n = frame_locals_count (thread) - 1;
  argv = alloca (n * sizeof (*argv));
  memcpy (argv, vp->sp, n * sizeof (*argv));

  cp = SCM_VM_CONT_DATA (continuation->vm_cont);

  data.cp = cp;
  data.vp = vp;
  GC_call_with_alloc_lock (vm_return_to_continuation_inner, &data);

  /* Now we have the continuation properly copied over.  We just need to
     copy on an empty frame and the return values, as the continuation
     expects.  */
  vm_push_sp (vp, vp->sp - frame_overhead - n);
  for (i = 0; i < frame_overhead; i++)
    vp->sp[n+i].as_scm = SCM_BOOL_F;
  memcpy(vp->sp, argv, n * sizeof (union scm_vm_stack_element));

  vp->ip = cp->vra;

  scm_i_reinstate_continuation (cont, cp->mra);
}

static SCM
capture_continuation (scm_thread *thread)
{
  struct scm_vm *vp = &thread->vm;
  void *mra = SCM_FRAME_MACHINE_RETURN_ADDRESS (vp->fp);
#if ENABLE_JIT
  if (mra == scm_jit_return_to_interpreter_trampoline)
    mra = NULL;
#endif
  SCM vm_cont = capture_stack (vp->stack_top,
                               SCM_FRAME_DYNAMIC_LINK (vp->fp),
                               SCM_FRAME_PREVIOUS_SP (vp->fp),
                               SCM_FRAME_VIRTUAL_RETURN_ADDRESS (vp->fp),
                               mra,
                               scm_dynstack_capture_all (&thread->dynstack),
                               0);
  return scm_i_make_continuation (thread, vm_cont);
}

struct compose_continuation_data
{
  struct scm_vm *vp;
  struct scm_vm_cont *cp;
};

static void *
compose_continuation_inner (void *data_ptr)
{
  struct compose_continuation_data *data = data_ptr;
  struct scm_vm *vp = data->vp;
  struct scm_vm_cont *cp = data->cp;

  memcpy (vp->fp - cp->stack_size,
          cp->stack_bottom,
          cp->stack_size * sizeof (*cp->stack_bottom));

  vp->fp -= cp->fp_offset;
  vp->ip = cp->vra;

  return cp->mra;
}

static uint8_t*
compose_continuation (scm_thread *thread, SCM cont)
{
  struct scm_vm *vp = &thread->vm;
  size_t nargs;
  struct compose_continuation_data data;
  struct scm_vm_cont *cp;
  union scm_vm_stack_element *args;
  ptrdiff_t old_fp_offset;
  uint8_t *mra;

  if (SCM_UNLIKELY (! SCM_VM_CONT_REWINDABLE_P (cont)))
    scm_wrong_type_arg_msg (NULL, 0, cont, "resumable continuation");

#if ENABLE_JIT
  if (!SCM_FRAME_MACHINE_RETURN_ADDRESS (vp->fp))
    SCM_FRAME_SET_MACHINE_RETURN_ADDRESS
      (vp->fp, scm_jit_return_to_interpreter_trampoline);
#endif

  nargs = frame_locals_count (thread) - 1;
  args = alloca (nargs * sizeof (*args));
  memcpy (args, vp->sp, nargs * sizeof (*args));

  cp = SCM_VM_CONT_DATA (cont);

  old_fp_offset = vp->stack_top - vp->fp;

  vm_push_sp (vp, vp->fp - (cp->stack_size + nargs));

  data.vp = vp;
  data.cp = cp;
  mra = GC_call_with_alloc_lock (compose_continuation_inner, &data);

  /* The resumed continuation will expect ARGS on the stack as if from a
     multiple-value return.  */
  memcpy (vp->sp, args, nargs * sizeof (*args));

  /* The prompt captured a slice of the dynamic stack.  Here we wind
     those entries onto the current thread's stack.  We also have to
     relocate any prompts that we see along the way.  */
  {
    scm_t_bits *walk;

    for (walk = SCM_DYNSTACK_FIRST (cp->dynstack);
         SCM_DYNSTACK_TAG (walk);
         walk = SCM_DYNSTACK_NEXT (walk))
      {
        scm_t_bits tag = SCM_DYNSTACK_TAG (walk);

        if (SCM_DYNSTACK_TAG_TYPE (tag) == SCM_DYNSTACK_TYPE_PROMPT)
          scm_dynstack_wind_prompt (&thread->dynstack, walk, old_fp_offset,
                                    thread->vm.registers);
        else
          scm_dynstack_wind_1 (&thread->dynstack, walk);
      }
  }

  return mra;
}

static void
expand_apply_argument (scm_thread *thread)
{
  SCM x = thread->vm.sp[0].as_scm;
  int len = scm_ilength (x);

  if (SCM_UNLIKELY (len < 0))
    scm_error (scm_arg_type_key, "apply", "Apply to non-list: ~S",
               scm_list_1 (x), scm_list_1 (x));

  alloc_frame (thread, frame_locals_count (thread) - 1 + len);

  while (len--)
    {
      thread->vm.sp[len].as_scm = SCM_CAR (x);
      x = SCM_CDR (x);
    }
}

/* This is here to avoid putting the code for "alloc-frame" in subr
   calls. */
static void
unpack_values_object (scm_thread *thread, SCM obj)
{
  size_t n, nvals = scm_i_nvalues (obj);
  alloc_frame (thread, nvals);
  for (n = 0; n < nvals; n++)
    SCM_FRAME_LOCAL (thread->vm.fp, n) = scm_i_value_ref (obj, n);
}

static void
foreign_call (scm_thread *thread, SCM cif, SCM pointer)
{
  SCM ret;
  int err = 0;

  ret = scm_i_foreign_call (cif, pointer, &err, thread->vm.sp);

  alloc_frame (thread, 2);
  SCM_FRAME_LOCAL (thread->vm.fp, 0) = ret;
  SCM_FRAME_LOCAL (thread->vm.fp, 1) = scm_from_int (err);
}

static SCM
capture_delimited_continuation (struct scm_vm *vp,
                                union scm_vm_stack_element *saved_fp,
                                uint8_t *saved_mra,
                                jmp_buf *saved_registers,
                                scm_t_dynstack *dynstack,
                                jmp_buf *current_registers)
{
  SCM vm_cont;
  uint32_t flags;
  union scm_vm_stack_element *base_fp;

  flags = SCM_F_VM_CONT_PARTIAL;
  /* If we are aborting to a prompt that has the same registers as those
     of the abort, it means there are no intervening C frames on the
     stack, and so the continuation can be relocated elsewhere on the
     stack: it is rewindable.  */
  if (saved_registers && saved_registers == current_registers)
    flags |= SCM_F_VM_CONT_REWINDABLE;

  /* Walk the stack until we find the first frame newer than saved_fp.
     We will save the stack until that frame.  It used to be that we
     could determine the stack base in O(1) time, but that's no longer
     the case, since the thunk application doesn't occur where the
     prompt is saved.  */
  for (base_fp = vp->fp;
       SCM_FRAME_DYNAMIC_LINK (base_fp) < saved_fp;
       base_fp = SCM_FRAME_DYNAMIC_LINK (base_fp));

  if (SCM_FRAME_DYNAMIC_LINK (base_fp) != saved_fp)
    abort();

  scm_dynstack_relocate_prompts (dynstack, vp->stack_top - base_fp);

  /* Capture from the base_fp to the top thunk application frame.  Don't
     capture values from the most recent frame, as they are the abort
     args.  */
  vm_cont = capture_stack (base_fp, vp->fp, vp->fp, vp->ip,
                           saved_mra, dynstack, flags);

  return scm_i_make_composable_continuation (vm_cont);
}

void
scm_i_vm_abort (SCM *tag_and_argv, size_t n)
{
  scm_call_n (vm_builtin_abort_to_prompt, tag_and_argv, n);
  /* Unreachable.  */
  abort ();
}

/* The same as scm_i_vm_abort(), but possibly called in response to
   resource allocation failures, so we might not be able to make a
   call, as that might require stack expansion.  Grrr.  */
void
scm_i_vm_emergency_abort (SCM *tag_and_argv, size_t n)
{
  scm_thread *thread = SCM_I_CURRENT_THREAD;
  struct scm_vm *vp = &thread->vm;
  scm_t_dynstack *dynstack = &thread->dynstack;
  SCM tag, cont;
  size_t nargs;
  scm_t_bits *prompt;
  scm_t_dynstack_prompt_flags flags;
  ptrdiff_t fp_offset, sp_offset;
  union scm_vm_stack_element *fp, *sp;
  SCM *argv;
  uint32_t *vra;
  uint8_t *mra;
  jmp_buf *registers;

  tag = tag_and_argv[0];
  argv = tag_and_argv + 1;
  nargs = n - 1;

  prompt = scm_dynstack_find_prompt (dynstack, tag,
                                     &flags, &fp_offset, &sp_offset,
                                     &vra, &mra, &registers);

  if (!prompt)
    {
      fprintf (stderr, "guile: fatal: emergency abort to unknown prompt\n");
      abort ();
    }

  if (!(flags & SCM_F_DYNSTACK_PROMPT_ESCAPE_ONLY))
    {
      fprintf (stderr, "guile: fatal: emergency abort to non-linear prompt\n");
      abort ();
    }

  cont = SCM_BOOL_F;

  /* Unwind.  */
  scm_dynstack_unwind (dynstack, prompt);

  fp = vp->stack_top - fp_offset;
  sp = vp->stack_top - sp_offset;

  /* Restore FP first so that a concurrent 'scm_i_vm_mark_stack' does
     not overwrite the 'abort' arguments assigned below (see
     <https://bugs.gnu.org/28211>).  */
  vp->fp = fp;

  /* Continuation gets nargs+1 values: the one more is for the cont.  */
  sp = sp - nargs - 1;

  /* Shuffle abort arguments down to the prompt continuation.  We have
     to be jumping to an older part of the stack.  */
  if (sp < vp->sp)
    abort ();
  sp[nargs].as_scm = cont;

  while (nargs--)
    sp[nargs].as_scm = *argv++;

  /* Restore VM regs */
  vp->sp = sp;
  vp->ip = vra;

  /* Jump! */
  vp->mra_after_abort = mra;
  longjmp (*registers, 1);
}

static uint8_t *
abort_to_prompt (scm_thread *thread, uint8_t *saved_mra)
{
  struct scm_vm *vp = &thread->vm;
  scm_t_dynstack *dynstack = &thread->dynstack;
  SCM tag, cont;
  size_t nargs;
  scm_t_bits *prompt;
  scm_t_dynstack_prompt_flags flags;
  ptrdiff_t fp_offset, sp_offset;
  union scm_vm_stack_element *fp, *sp;
  uint32_t *vra;
  uint8_t *mra;
  jmp_buf *registers;

  tag = SCM_FRAME_LOCAL (vp->fp, 1);
  nargs = frame_locals_count (thread) - 2;

  prompt = scm_dynstack_find_prompt (dynstack, tag,
                                     &flags, &fp_offset, &sp_offset,
                                     &vra, &mra, &registers);

  if (!prompt)
    scm_misc_error ("abort", "Abort to unknown prompt", scm_list_1 (tag));

  fp = vp->stack_top - fp_offset;

  /* Only reify if the continuation referenced in the handler. */
  if (flags & SCM_F_DYNSTACK_PROMPT_ESCAPE_ONLY)
    cont = SCM_BOOL_F;
  else
    {
      scm_t_dynstack *captured;

      captured = scm_dynstack_capture (dynstack, SCM_DYNSTACK_NEXT (prompt));
      cont = capture_delimited_continuation (vp, fp, saved_mra, registers,
                                             captured, thread->vm.registers);
    }

  /* Unwind.  */
  scm_dynstack_unwind (dynstack, prompt);

  /* Recompute FP, as scm_dynstack_unwind may have expanded the stack.  */
  fp = vp->stack_top - fp_offset;
  sp = vp->stack_top - sp_offset;

  /* Continuation gets nargs+1 values: the one more is for the cont.  */
  sp = sp - nargs - 1;

  /* Restore FP first so that a concurrent 'scm_i_vm_mark_stack' does
     not overwrite the 'abort' arguments assigned below (see
     <https://bugs.gnu.org/28211>).  */
  vp->fp = fp;

  /* Shuffle abort arguments down to the prompt continuation.  We have
     to be jumping to an older part of the stack.  */
  if (sp < vp->sp)
    abort ();
  sp[nargs].as_scm = cont;
  while (nargs--)
    sp[nargs] = vp->sp[nargs];

  /* Restore VM regs */
  vp->sp = sp;
  vp->ip = vra;

  /* If there are intervening C frames, then jump over them, making a
     nonlocal exit.  Otherwise fall through and let the VM pick up where
     it left off.  */
  if (thread->vm.registers != registers)
    {
      vp->mra_after_abort = mra;
      longjmp (*registers, 1);
    }

  return mra;
}

static uint32_t *
get_callee_vcode (scm_thread *thread)
{
  struct scm_vm *vp = &thread->vm;

  SCM proc = SCM_FRAME_LOCAL (vp->fp, 0);

  if (SCM_LIKELY (SCM_PROGRAM_P (proc)))
    return SCM_PROGRAM_CODE (proc);

  while (SCM_STRUCTP (proc) && SCM_STRUCT_APPLICABLE_P (proc))
    {
      proc = SCM_STRUCT_PROCEDURE (proc);
      SCM_FRAME_LOCAL (vp->fp, 0) = proc;

      if (SCM_PROGRAM_P (proc))
        return SCM_PROGRAM_CODE (proc);
    }

  if (SCM_HAS_TYP7 (proc, scm_tc7_smob) && SCM_SMOB_APPLICABLE_P (proc))
    {
      uint32_t n = frame_locals_count (thread);

      alloc_frame (thread, n + 1);

      /* Although we could make VM modifications to avoid this shuffle,
         it's easier to piggy-back on the subr arg parsing machinery.
         Hopefully applicable smobs will go away in the mid-term.  */
      while (n--)
        SCM_FRAME_LOCAL (vp->fp, n + 1) = SCM_FRAME_LOCAL (vp->fp, n);

      proc = SCM_SMOB_DESCRIPTOR (proc).apply_trampoline;
      SCM_FRAME_LOCAL (vp->fp, 0) = proc;
      return SCM_PROGRAM_CODE (proc);
    }

  vp->ip = SCM_FRAME_VIRTUAL_RETURN_ADDRESS (vp->fp);

  scm_error (scm_arg_type_key, NULL, "Wrong type to apply: ~S",
             scm_list_1 (proc), scm_list_1 (proc));
}

SCM
scm_call_n (SCM proc, SCM *argv, size_t nargs)
{
  scm_thread *thread;
  struct scm_vm *vp;
  union scm_vm_stack_element *return_fp, *call_fp;
  /* Since nargs can only describe the length of a valid argv array in
     elements and each element is at least 4 bytes, nargs will not be
     greater than INTMAX/2 and therefore we don't have to check for
     overflow here or below.  */
  size_t return_nlocals = 0, call_nlocals = nargs + 1, frame_size = 3;
  ptrdiff_t stack_reserve_words;
  size_t i;

  thread = SCM_I_CURRENT_THREAD;
  vp = &thread->vm;

  SCM_CHECK_STACK;

  /* It's not valid for argv to point into the stack already.  */
  if ((void *) argv < (void *) vp->stack_top &&
      (void *) argv >= (void *) vp->sp)
    abort();

  /* Check that we have enough space for the two stack frames: the
     innermost one that makes the call, and its continuation which
     receives the resulting value(s) and returns from the engine
     call.  */
  stack_reserve_words = call_nlocals + frame_size + return_nlocals + frame_size;
  vm_push_sp (vp, vp->sp - stack_reserve_words);

  call_fp = vp->sp + call_nlocals;
  return_fp = call_fp + frame_size + return_nlocals;

  SCM_FRAME_SET_VIRTUAL_RETURN_ADDRESS (return_fp, vp->ip);
  SCM_FRAME_SET_MACHINE_RETURN_ADDRESS (return_fp, 0);
  SCM_FRAME_SET_DYNAMIC_LINK (return_fp, vp->fp);

  vp->ip = (uint32_t *) vm_boot_continuation_code;

  SCM_FRAME_SET_VIRTUAL_RETURN_ADDRESS (call_fp, vp->ip);
  SCM_FRAME_SET_MACHINE_RETURN_ADDRESS (call_fp, 0);
  SCM_FRAME_SET_DYNAMIC_LINK (call_fp, return_fp);
  SCM_FRAME_LOCAL (call_fp, 0) = proc;
  for (i = 0; i < nargs; i++)
    SCM_FRAME_LOCAL (call_fp, i + 1) = argv[i];

  vp->fp = call_fp;

  {
    jmp_buf registers;
    int resume;
    jmp_buf *prev_registers = thread->vm.registers;
    SCM ret;

    resume = setjmp (registers);

    thread->vm.registers = &registers;

    if (SCM_UNLIKELY (resume))
      {
        uint8_t *mcode = vp->mra_after_abort;
        scm_gc_after_nonlocal_exit ();
        /* Non-local return.  */
        if (vp->abort_hook_enabled)
          spct_invoke_abort_hook (thread);
#if ENABLE_JIT
        if (mcode && !vp->disable_mcode)
          scm_jit_enter_mcode (thread, mcode);
#endif
      }
    else
      vp->ip = get_callee_vcode (thread);

    ret = vm_engines[vp->engine](thread);
    thread->vm.registers = prev_registers;

    return ret;
  }
}

/* Scheme interface */

#define VM_ADD_HOOK(h, f)                                               \
  {                                                                     \
    scm_thread *t = SCM_I_CURRENT_THREAD;                               \
    SCM hook = t->vm.h##_hook;                                          \
    if (scm_is_false (hook))                                            \
      hook = t->vm.h##_hook = scm_make_hook (SCM_I_MAKINUM (1));        \
    scm_add_hook_x (hook, f, SCM_UNDEFINED);                            \
    vm_hook_compute_enabled (t, hook, &t->vm.h##_hook_enabled);         \
    vm_recompute_disable_mcode (t);                                     \
    return SCM_UNSPECIFIED;                                             \
  }

#define VM_REMOVE_HOOK(h, f)                                            \
  {                                                                     \
    scm_thread *t = SCM_I_CURRENT_THREAD;                               \
    SCM hook = t->vm.h##_hook;                                          \
    if (scm_is_true (hook))                                             \
      scm_remove_hook_x (hook, f);                                      \
    vm_hook_compute_enabled (t, hook, &t->vm.h##_hook_enabled);         \
    vm_recompute_disable_mcode (t);                                     \
    return SCM_UNSPECIFIED;                                             \
  }

SCM_DEFINE (scm_vm_add_apply_hook_x, "vm-add-apply-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_add_apply_hook_x
{
  VM_ADD_HOOK (apply, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_remove_apply_hook_x, "vm-remove-apply-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_remove_apply_hook_x
{
  VM_REMOVE_HOOK (apply, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_add_return_hook_x, "vm-add-return-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_add_return_hook_x
{
  VM_ADD_HOOK (return, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_remove_return_hook_x, "vm-remove-return-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_remove_return_hook_x
{
  VM_REMOVE_HOOK (return, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_add_next_hook_x, "vm-add-next-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_add_next_hook_x
{
  VM_ADD_HOOK (next, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_remove_next_hook_x, "vm-remove-next-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_remove_next_hook_x
{
  VM_REMOVE_HOOK (next, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_add_abort_hook_x, "vm-add-abort-hook!", 1, 0, 0,
	    (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_add_abort_hook_x
{
  VM_ADD_HOOK (abort, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_remove_abort_hook_x, "vm-remove-abort-hook!", 1, 0, 0,
            (SCM f),
	    "")
#define FUNC_NAME s_scm_vm_remove_abort_hook_x
{
  VM_REMOVE_HOOK (abort, f);
}
#undef FUNC_NAME

SCM_DEFINE (scm_vm_trace_level, "vm-trace-level", 0, 0, 0,
	    (void),
	    "")
#define FUNC_NAME s_scm_vm_trace_level
{
  return scm_from_int (SCM_I_CURRENT_THREAD->vm.trace_level);
}
#undef FUNC_NAME

SCM_DEFINE (scm_set_vm_trace_level_x, "set-vm-trace-level!", 1, 0, 0,
	    (SCM level),
	    "")
#define FUNC_NAME s_scm_set_vm_trace_level_x
{
  scm_thread *thread = SCM_I_CURRENT_THREAD;
  return scm_from_int (set_vm_trace_level (thread, scm_to_int (level)));
}
#undef FUNC_NAME


/*
 * VM engines
 */

static int
symbol_to_vm_engine (SCM engine, const char *FUNC_NAME)
{
  if (scm_is_eq (engine, sym_regular))
    return SCM_VM_REGULAR_ENGINE;
  else if (scm_is_eq (engine, sym_debug))
    return SCM_VM_DEBUG_ENGINE;
  else
    SCM_MISC_ERROR ("Unknown VM engine: ~a", scm_list_1 (engine));
}
  
static SCM
vm_engine_to_symbol (int engine, const char *FUNC_NAME)
{
  switch (engine)
    {
    case SCM_VM_REGULAR_ENGINE:
      return sym_regular;
    case SCM_VM_DEBUG_ENGINE:
      return sym_debug;
    default:
      /* ? */
      SCM_MISC_ERROR ("Unknown VM engine: ~a",
                      scm_list_1 (scm_from_int (engine)));
    }
}
  
SCM_DEFINE (scm_vm_engine, "vm-engine", 0, 0, 0,
	    (void),
	    "")
#define FUNC_NAME s_scm_vm_engine
{
  return vm_engine_to_symbol (SCM_I_CURRENT_THREAD->vm.engine, FUNC_NAME);
}
#undef FUNC_NAME

void
scm_c_set_vm_engine_x (int engine)
#define FUNC_NAME "set-vm-engine!"
{
  scm_thread *thread = SCM_I_CURRENT_THREAD;

  if (engine < 0 || engine >= SCM_VM_NUM_ENGINES)
    SCM_MISC_ERROR ("Unknown VM engine: ~a",
                    scm_list_1 (scm_from_int (engine)));
    
  thread->vm.engine = engine;
  /* Trigger update of the various hook_enabled flags.  */
  set_vm_trace_level (thread, thread->vm.trace_level);
}
#undef FUNC_NAME

SCM_DEFINE (scm_set_vm_engine_x, "set-vm-engine!", 1, 0, 0,
	    (SCM engine),
	    "")
#define FUNC_NAME s_scm_set_vm_engine_x
{
  scm_c_set_vm_engine_x (symbol_to_vm_engine (engine, FUNC_NAME));
  return SCM_UNSPECIFIED;
}
#undef FUNC_NAME

void
scm_c_set_default_vm_engine_x (int engine)
#define FUNC_NAME "set-default-vm-engine!"
{
  if (engine < 0 || engine >= SCM_VM_NUM_ENGINES)
    SCM_MISC_ERROR ("Unknown VM engine: ~a",
                    scm_list_1 (scm_from_int (engine)));
    
  vm_default_engine = engine;
}
#undef FUNC_NAME

SCM_DEFINE (scm_set_default_vm_engine_x, "set-default-vm-engine!", 1, 0, 0,
	    (SCM engine),
	    "")
#define FUNC_NAME s_scm_set_default_vm_engine_x
{
  scm_c_set_default_vm_engine_x (symbol_to_vm_engine (engine, FUNC_NAME));
  return SCM_UNSPECIFIED;
}
#undef FUNC_NAME

/* FIXME: This function makes no sense, but we keep it to make sure we
   have a way of switching to the debug or regular VM.  */
SCM_DEFINE (scm_call_with_vm, "call-with-vm", 1, 0, 1,
	    (SCM proc, SCM args),
	    "Apply @var{proc} to @var{args} in a dynamic extent in which\n"
            "@var{vm} is the current VM.")
#define FUNC_NAME s_scm_call_with_vm
{
  return scm_apply_0 (proc, args);
}
#undef FUNC_NAME

SCM_DEFINE (scm_call_with_stack_overflow_handler,
            "call-with-stack-overflow-handler", 3, 0, 0,
	    (SCM limit, SCM thunk, SCM handler),
	    "Call @var{thunk} in an environment in which the stack limit has\n"
            "been reduced to @var{limit} additional words.  If the limit is\n"
            "reached, @var{handler} (a thunk) will be invoked in the dynamic\n"
            "environment of the error.  For the extent of the call to\n"
            "@var{handler}, the stack limit and handler are restored to the\n"
            "values that were in place when\n"
            "@code{call-with-stack-overflow-handler} was called.")
#define FUNC_NAME s_scm_call_with_stack_overflow_handler
{
  struct scm_thread *t = SCM_I_CURRENT_THREAD;
  ptrdiff_t c_limit, stack_size;
  struct overflow_handler_data data;
  SCM new_limit, ret;

  stack_size = t->vm.stack_top - t->vm.sp;

  c_limit = scm_to_ptrdiff_t (limit);
  if (c_limit <= 0)
    scm_out_of_range (FUNC_NAME, limit);

  new_limit = scm_sum (scm_from_ptrdiff_t (stack_size), limit);
  if (scm_is_pair (t->vm.overflow_handler_stack))
    new_limit = scm_min (new_limit, scm_caar (t->vm.overflow_handler_stack));

  /* Hacky check that the current stack depth plus the limit is within
     the range of a ptrdiff_t.  */
  scm_to_ptrdiff_t (new_limit);

  data.vp = &t->vm;
  data.overflow_handler_stack =
    scm_acons (limit, handler, t->vm.overflow_handler_stack);

  scm_dynwind_begin (SCM_F_DYNWIND_REWINDABLE);

  scm_dynwind_rewind_handler (wind_overflow_handler, &data,
                              SCM_F_WIND_EXPLICITLY);
  scm_dynwind_unwind_handler (unwind_overflow_handler, &data,
                              SCM_F_WIND_EXPLICITLY);

  ret = scm_call_0 (thunk);

  scm_dynwind_end ();

  return ret;
}
#undef FUNC_NAME


/*
 * Initialize
 */

SCM
scm_load_compiled_with_vm (SCM file)
{
  return scm_call_0 (scm_load_thunk_from_file (file));
}

  
void
scm_init_vm_builtin_properties (void)
{
  /* FIXME: Seems hacky to do this here, but oh well :/ */
  scm_sym_apply = scm_from_utf8_symbol ("apply");
  scm_sym_values = scm_from_utf8_symbol ("values");
  scm_sym_abort_to_prompt = scm_from_utf8_symbol ("abort-to-prompt");
  scm_sym_call_with_values = scm_from_utf8_symbol ("call-with-values");
  scm_sym_call_with_current_continuation =
    scm_from_utf8_symbol ("call-with-current-continuation");

#define INIT_BUILTIN(builtin, BUILTIN, req, opt, rest)                  \
  scm_set_procedure_property_x (vm_builtin_##builtin, scm_sym_name,     \
                                scm_sym_##builtin);
  FOR_EACH_VM_BUILTIN (INIT_BUILTIN);
#undef INIT_BUILTIN
}

void
scm_bootstrap_vm (void)
{
  scm_c_register_extension ("libguile-" SCM_EFFECTIVE_VERSION,
                            "scm_init_vm",
                            (scm_t_extension_init_func)scm_init_vm, NULL);
  scm_c_register_extension ("libguile-" SCM_EFFECTIVE_VERSION,
                            "scm_init_vm_builtins",
                            (scm_t_extension_init_func)scm_init_vm_builtins,
                            NULL);

  scm_vm_intrinsics.expand_stack = thread_expand_stack;
  scm_vm_intrinsics.cons_rest = cons_rest;
  scm_vm_intrinsics.compute_kwargs_npositional = compute_kwargs_npositional;
  scm_vm_intrinsics.bind_kwargs = bind_kwargs;
  scm_vm_intrinsics.push_interrupt_frame = push_interrupt_frame;
  scm_vm_intrinsics.reinstate_continuation_x = reinstate_continuation_x;
  scm_vm_intrinsics.capture_continuation = capture_continuation;
  scm_vm_intrinsics.compose_continuation = compose_continuation;
  scm_vm_intrinsics.expand_apply_argument = expand_apply_argument;
  scm_vm_intrinsics.abort_to_prompt = abort_to_prompt;
  scm_vm_intrinsics.get_callee_vcode = get_callee_vcode;
  scm_vm_intrinsics.unpack_values_object = unpack_values_object;
  scm_vm_intrinsics.foreign_call = foreign_call;

  sym_keyword_argument_error = scm_from_latin1_symbol ("keyword-argument-error");
  sym_regular = scm_from_latin1_symbol ("regular");
  sym_debug = scm_from_latin1_symbol ("debug");

  vm_boot_continuation = scm_i_make_program (vm_boot_continuation_code);
  SCM_SET_CELL_WORD_0 (vm_boot_continuation,
                       (SCM_CELL_WORD_0 (vm_boot_continuation)
                        | SCM_F_PROGRAM_IS_BOOT));

  define_vm_builtins ();
}

void
scm_init_vm (void)
{
#ifndef SCM_MAGIC_SNARFER
#include "vm.x"
#endif
}