OAOO-ify WEAK_POINTER_NWORDS

[sbcl.git] / src / runtime / gencgc.c

/*
 * GENerational Conservative Garbage Collector for SBCL
 */

/*
 * This software is part of the SBCL system. See the README file for
 * more information.
 *
 * This software is derived from the CMU CL system, which was
 * written at Carnegie Mellon University and released into the
 * public domain. The software is in the public domain and is
 * provided with absolutely no warranty. See the COPYING and CREDITS
 * files for more information.
 */

/*
 * For a review of garbage collection techniques (e.g. generational
 * GC) and terminology (e.g. "scavenging") see Paul R. Wilson,
 * "Uniprocessor Garbage Collection Techniques". As of 20000618, this
 * had been accepted for _ACM Computing Surveys_ and was available
 * as a PostScript preprint through
 *   <http://www.cs.utexas.edu/users/oops/papers.html>
 * as
 *   <ftp://ftp.cs.utexas.edu/pub/garbage/bigsurv.ps>.
 */

#include <stdlib.h>
#include <stdio.h>
#include <errno.h>
#include <string.h>
#include "sbcl.h"
#if defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD)
#include "pthreads_win32.h"
#else
#include <signal.h>
#endif
#include "runtime.h"
#include "os.h"
#include "interr.h"
#include "globals.h"
#include "interrupt.h"
#include "validate.h"
#include "lispregs.h"
#include "arch.h"
#include "gc.h"
#include "gc-internal.h"
#include "thread.h"
#include "pseudo-atomic.h"
#include "alloc.h"
#include "genesis/vector.h"
#include "genesis/weak-pointer.h"
#include "genesis/fdefn.h"
#include "genesis/simple-fun.h"
#include "save.h"
#include "genesis/hash-table.h"
#include "genesis/instance.h"
#include "genesis/layout.h"
#include "gencgc.h"
#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
#include "genesis/cons.h"
#endif
#ifdef LISP_FEATURE_X86
#include "forwarding-ptr.h"
#endif

/* forward declarations */
page_index_t  gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t nbytes,
                                    int page_type_flag);

\f
/*
 * GC parameters
 */

/* As usually configured, generations 0-5 are normal collected generations,
   6 is pseudo-static (the objects in which are never moved nor reclaimed),
   and 7 is scratch space used when collecting a generation without promotion,
   wherein it is moved to generation 7 and back again.
 */
enum {
    SCRATCH_GENERATION = PSEUDO_STATIC_GENERATION+1,
    NUM_GENERATIONS
};

/* Should we use page protection to help avoid the scavenging of pages
 * that don't have pointers to younger generations? */
boolean enable_page_protection = 1;

/* Largest allocation seen since last GC. */
os_vm_size_t large_allocation = 0;

\f
/*
 * debugging
 */

/* the verbosity level. All non-error messages are disabled at level 0;
 * and only a few rare messages are printed at level 1. */
#if QSHOW == 2
boolean gencgc_verbose = 1;
#else
boolean gencgc_verbose = 0;
#endif

/* FIXME: At some point enable the various error-checking things below
 * and see what they say. */

/* We hunt for pointers to old-space, when GCing generations >= verify_gen.
 * Set verify_gens to HIGHEST_NORMAL_GENERATION + 1 to disable this kind of
 * check. */
generation_index_t verify_gens = HIGHEST_NORMAL_GENERATION + 1;

/* Should we do a pre-scan verify of generation 0 before it's GCed? */
boolean pre_verify_gen_0 = 0;

/* Should we print a note when code objects are found in the dynamic space
 * during a heap verify? */
boolean verify_dynamic_code_check = 0;

#ifdef LISP_FEATURE_X86
/* Should we check code objects for fixup errors after they are transported? */
boolean check_code_fixups = 0;
#endif

/* Should we check that newly allocated regions are zero filled? */
boolean gencgc_zero_check = 0;

/* Should we check that the free space is zero filled? */
boolean gencgc_enable_verify_zero_fill = 0;

/* When loading a core, don't do a full scan of the memory for the
 * memory region boundaries. (Set to true by coreparse.c if the core
 * contained a pagetable entry).
 */
boolean gencgc_partial_pickup = 0;

/* If defined, free pages are read-protected to ensure that nothing
 * accesses them.
 */

/* #define READ_PROTECT_FREE_PAGES */

\f
/*
 * GC structures and variables
 */

/* the total bytes allocated. These are seen by Lisp DYNAMIC-USAGE. */
os_vm_size_t bytes_allocated = 0;
os_vm_size_t auto_gc_trigger = 0;

/* the source and destination generations. These are set before a GC starts
 * scavenging. */
generation_index_t from_space;
generation_index_t new_space;

/* Set to 1 when in GC */
boolean gc_active_p = 0;

/* should the GC be conservative on stack. If false (only right before
 * saving a core), don't scan the stack / mark pages dont_move. */
static boolean conservative_stack = 1;

/* An array of page structures is allocated on gc initialization.
 * This helps to quickly map between an address and its page structure.
 * page_table_pages is set from the size of the dynamic space. */
page_index_t page_table_pages;
struct page *page_table;

in_use_marker_t *page_table_pinned_dwords;
size_t pins_map_size_in_bytes;

/* In GC cards that have conservative pointers to them, should we wipe out
 * dwords in there that are not used, so that they do not act as false
 * root to other things in the heap from then on? This is a new feature
 * but in testing it is both reliable and no noticeable slowdown. */
int do_wipe_p = 1;

static inline boolean page_allocated_p(page_index_t page) {
    return (page_table[page].allocated != FREE_PAGE_FLAG);
}

static inline boolean page_no_region_p(page_index_t page) {
    return !(page_table[page].allocated & OPEN_REGION_PAGE_FLAG);
}

static inline boolean page_allocated_no_region_p(page_index_t page) {
    return ((page_table[page].allocated & (UNBOXED_PAGE_FLAG | BOXED_PAGE_FLAG))
            && page_no_region_p(page));
}

static inline boolean page_free_p(page_index_t page) {
    return (page_table[page].allocated == FREE_PAGE_FLAG);
}

static inline boolean page_boxed_p(page_index_t page) {
    return (page_table[page].allocated & BOXED_PAGE_FLAG);
}

static inline boolean page_boxed_no_region_p(page_index_t page) {
    return page_boxed_p(page) && page_no_region_p(page);
}

static inline boolean page_unboxed_p(page_index_t page) {
    /* Both flags set == boxed code page */
    return ((page_table[page].allocated & UNBOXED_PAGE_FLAG)
            && !page_boxed_p(page));
}

static inline boolean protect_page_p(page_index_t page, generation_index_t generation) {
    return (page_boxed_no_region_p(page)
            && (page_table[page].bytes_used != 0)
            && !page_table[page].dont_move
            && (page_table[page].gen == generation));
}

/* To map addresses to page structures the address of the first page
 * is needed. */
void *heap_base = NULL;

/* Calculate the start address for the given page number. */
inline void *
page_address(page_index_t page_num)
{
    return (heap_base + (page_num * GENCGC_CARD_BYTES));
}

/* Calculate the address where the allocation region associated with
 * the page starts. */
static inline void *
page_scan_start(page_index_t page_index)
{
    return page_address(page_index)-page_table[page_index].scan_start_offset;
}

/* True if the page starts a contiguous block. */
static inline boolean
page_starts_contiguous_block_p(page_index_t page_index)
{
    return page_table[page_index].scan_start_offset == 0;
}

/* True if the page is the last page in a contiguous block. */
static inline boolean
page_ends_contiguous_block_p(page_index_t page_index, generation_index_t gen)
{
    return (/* page doesn't fill block */
            (page_table[page_index].bytes_used < GENCGC_CARD_BYTES)
            /* page is last allocated page */
            || ((page_index + 1) >= last_free_page)
            /* next page free */
            || page_free_p(page_index + 1)
            /* next page contains no data */
            || (page_table[page_index + 1].bytes_used == 0)
            /* next page is in different generation */
            || (page_table[page_index + 1].gen != gen)
            /* next page starts its own contiguous block */
            || (page_starts_contiguous_block_p(page_index + 1)));
}

/// External function for calling from Lisp.
page_index_t ext_find_page_index(void *addr) { return find_page_index(addr); }

static os_vm_size_t
npage_bytes(page_index_t npages)
{
    gc_assert(npages>=0);
    return ((os_vm_size_t)npages)*GENCGC_CARD_BYTES;
}

/* Check that X is a higher address than Y and return offset from Y to
 * X in bytes. */
static inline os_vm_size_t
void_diff(void *x, void *y)
{
    gc_assert(x >= y);
    return (pointer_sized_uint_t)x - (pointer_sized_uint_t)y;
}

/* a structure to hold the state of a generation
 *
 * CAUTION: If you modify this, make sure to touch up the alien
 * definition in src/code/gc.lisp accordingly. ...or better yes,
 * deal with the FIXME there...
 */
struct generation {

    /* the first page that gc_alloc() checks on its next call */
    page_index_t alloc_start_page;

    /* the first page that gc_alloc_unboxed() checks on its next call */
    page_index_t alloc_unboxed_start_page;

    /* the first page that gc_alloc_large (boxed) considers on its next
     * call. (Although it always allocates after the boxed_region.) */
    page_index_t alloc_large_start_page;

    /* the first page that gc_alloc_large (unboxed) considers on its
     * next call. (Although it always allocates after the
     * current_unboxed_region.) */
    page_index_t alloc_large_unboxed_start_page;

    /* the bytes allocated to this generation */
    os_vm_size_t bytes_allocated;

    /* the number of bytes at which to trigger a GC */
    os_vm_size_t gc_trigger;

    /* to calculate a new level for gc_trigger */
    os_vm_size_t bytes_consed_between_gc;

    /* the number of GCs since the last raise */
    int num_gc;

    /* the number of GCs to run on the generations before raising objects to the
     * next generation */
    int number_of_gcs_before_promotion;

    /* the cumulative sum of the bytes allocated to this generation. It is
     * cleared after a GC on this generations, and update before new
     * objects are added from a GC of a younger generation. Dividing by
     * the bytes_allocated will give the average age of the memory in
     * this generation since its last GC. */
    os_vm_size_t cum_sum_bytes_allocated;

    /* a minimum average memory age before a GC will occur helps
     * prevent a GC when a large number of new live objects have been
     * added, in which case a GC could be a waste of time */
    double minimum_age_before_gc;
};

/* an array of generation structures. There needs to be one more
 * generation structure than actual generations as the oldest
 * generation is temporarily raised then lowered. */
struct generation generations[NUM_GENERATIONS];

/* the oldest generation that is will currently be GCed by default.
 * Valid values are: 0, 1, ... HIGHEST_NORMAL_GENERATION
 *
 * The default of HIGHEST_NORMAL_GENERATION enables GC on all generations.
 *
 * Setting this to 0 effectively disables the generational nature of
 * the GC. In some applications generational GC may not be useful
 * because there are no long-lived objects.
 *
 * An intermediate value could be handy after moving long-lived data
 * into an older generation so an unnecessary GC of this long-lived
 * data can be avoided. */
generation_index_t gencgc_oldest_gen_to_gc = HIGHEST_NORMAL_GENERATION;

/* META: Is nobody aside from me bothered by this especially misleading
 * use of the word "last"?  It could mean either "ultimate" or "prior",
 * but in fact means neither. It is the *FIRST* page that should be grabbed
 * for more space, so it is min free page, or 1+ the max used page. */
/* The maximum free page in the heap is maintained and used to update
 * ALLOCATION_POINTER which is used by the room function to limit its
 * search of the heap. XX Gencgc obviously needs to be better
 * integrated with the Lisp code. */

page_index_t last_free_page;
\f
#ifdef LISP_FEATURE_SB_THREAD
/* This lock is to prevent multiple threads from simultaneously
 * allocating new regions which overlap each other.  Note that the
 * majority of GC is single-threaded, but alloc() may be called from
 * >1 thread at a time and must be thread-safe.  This lock must be
 * seized before all accesses to generations[] or to parts of
 * page_table[] that other threads may want to see */
static pthread_mutex_t free_pages_lock = PTHREAD_MUTEX_INITIALIZER;
/* This lock is used to protect non-thread-local allocation. */
static pthread_mutex_t allocation_lock = PTHREAD_MUTEX_INITIALIZER;
#endif

extern os_vm_size_t gencgc_release_granularity;
os_vm_size_t gencgc_release_granularity = GENCGC_RELEASE_GRANULARITY;

extern os_vm_size_t gencgc_alloc_granularity;
os_vm_size_t gencgc_alloc_granularity = GENCGC_ALLOC_GRANULARITY;

\f
/*
 * miscellaneous heap functions
 */

/* Count the number of pages which are write-protected within the
 * given generation. */
static page_index_t
count_write_protect_generation_pages(generation_index_t generation)
{
    page_index_t i, count = 0;

    for (i = 0; i < last_free_page; i++)
        if (page_allocated_p(i)
            && (page_table[i].gen == generation)
            && (page_table[i].write_protected == 1))
            count++;
    return count;
}

/* Count the number of pages within the given generation. */
static page_index_t
count_generation_pages(generation_index_t generation)
{
    page_index_t i;
    page_index_t count = 0;

    for (i = 0; i < last_free_page; i++)
        if (page_allocated_p(i)
            && (page_table[i].gen == generation))
            count++;
    return count;
}

#if QSHOW
static page_index_t
count_dont_move_pages(void)
{
    page_index_t i;
    page_index_t count = 0;
    for (i = 0; i < last_free_page; i++) {
        if (page_allocated_p(i)
            && (page_table[i].dont_move != 0)) {
            ++count;
        }
    }
    return count;
}
#endif /* QSHOW */

/* Work through the pages and add up the number of bytes used for the
 * given generation. */
static __attribute__((unused)) os_vm_size_t
count_generation_bytes_allocated (generation_index_t gen)
{
    page_index_t i;
    os_vm_size_t result = 0;
    for (i = 0; i < last_free_page; i++) {
        if (page_allocated_p(i)
            && (page_table[i].gen == gen))
            result += page_table[i].bytes_used;
    }
    return result;
}

/* Return the average age of the memory in a generation. */
extern double
generation_average_age(generation_index_t gen)
{
    if (generations[gen].bytes_allocated == 0)
        return 0.0;

    return
        ((double)generations[gen].cum_sum_bytes_allocated)
        / ((double)generations[gen].bytes_allocated);
}

#ifdef LISP_FEATURE_X86
extern void fpu_save(void *);
extern void fpu_restore(void *);
#endif

extern void
write_generation_stats(FILE *file)
{
    generation_index_t i;

#ifdef LISP_FEATURE_X86
    int fpu_state[27];

    /* Can end up here after calling alloc_tramp which doesn't prepare
     * the x87 state, and the C ABI uses a different mode */
    fpu_save(fpu_state);
#endif

    /* Print the heap stats. */
    fprintf(file,
            " Gen StaPg UbSta LaSta LUbSt Boxed Unboxed LB   LUB  !move  Alloc  Waste   Trig    WP  GCs Mem-age\n");

    for (i = 0; i < SCRATCH_GENERATION; i++) {
        page_index_t j;
        page_index_t boxed_cnt = 0;
        page_index_t unboxed_cnt = 0;
        page_index_t large_boxed_cnt = 0;
        page_index_t large_unboxed_cnt = 0;
        page_index_t pinned_cnt=0;

        for (j = 0; j < last_free_page; j++)
            if (page_table[j].gen == i) {

                /* Count the number of boxed pages within the given
                 * generation. */
                if (page_boxed_p(j)) {
                    if (page_table[j].large_object)
                        large_boxed_cnt++;
                    else
                        boxed_cnt++;
                }
                if(page_table[j].dont_move) pinned_cnt++;
                /* Count the number of unboxed pages within the given
                 * generation. */
                if (page_unboxed_p(j)) {
                    if (page_table[j].large_object)
                        large_unboxed_cnt++;
                    else
                        unboxed_cnt++;
                }
            }

        gc_assert(generations[i].bytes_allocated
                  == count_generation_bytes_allocated(i));
        fprintf(file,
                "   %1d: %5ld %5ld %5ld %5ld",
                i,
                (long)generations[i].alloc_start_page,
                (long)generations[i].alloc_unboxed_start_page,
                (long)generations[i].alloc_large_start_page,
                (long)generations[i].alloc_large_unboxed_start_page);
        fprintf(file,
                " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT
                " %5"PAGE_INDEX_FMT" %5"PAGE_INDEX_FMT,
                boxed_cnt, unboxed_cnt, large_boxed_cnt,
                large_unboxed_cnt, pinned_cnt);
        fprintf(file,
                " %8"OS_VM_SIZE_FMT
                " %5"OS_VM_SIZE_FMT
                " %8"OS_VM_SIZE_FMT
                " %4"PAGE_INDEX_FMT" %3d %7.4f\n",
                generations[i].bytes_allocated,
                (npage_bytes(count_generation_pages(i)) - generations[i].bytes_allocated),
                generations[i].gc_trigger,
                count_write_protect_generation_pages(i),
                generations[i].num_gc,
                generation_average_age(i));
    }
    fprintf(file,"   Total bytes allocated    = %"OS_VM_SIZE_FMT"\n", bytes_allocated);
    fprintf(file,"   Dynamic-space-size bytes = %"OS_VM_SIZE_FMT"\n", dynamic_space_size);

#ifdef LISP_FEATURE_X86
    fpu_restore(fpu_state);
#endif
}

extern void
write_heap_exhaustion_report(FILE *file, long available, long requested,
                             struct thread *thread)
{
    fprintf(file,
            "Heap exhausted during %s: %ld bytes available, %ld requested.\n",
            gc_active_p ? "garbage collection" : "allocation",
            available,
            requested);
    write_generation_stats(file);
    fprintf(file, "GC control variables:\n");
    fprintf(file, "   *GC-INHIBIT* = %s\n   *GC-PENDING* = %s\n",
            SymbolValue(GC_INHIBIT,thread)==NIL ? "false" : "true",
            (SymbolValue(GC_PENDING, thread) == T) ?
            "true" : ((SymbolValue(GC_PENDING, thread) == NIL) ?
                      "false" : "in progress"));
#ifdef LISP_FEATURE_SB_THREAD
    fprintf(file, "   *STOP-FOR-GC-PENDING* = %s\n",
            SymbolValue(STOP_FOR_GC_PENDING,thread)==NIL ? "false" : "true");
#endif
}

extern void
print_generation_stats(void)
{
    write_generation_stats(stderr);
}

extern char* gc_logfile;
char * gc_logfile = NULL;

extern void
log_generation_stats(char *logfile, char *header)
{
    if (logfile) {
        FILE * log = fopen(logfile, "a");
        if (log) {
            fprintf(log, "%s\n", header);
            write_generation_stats(log);
            fclose(log);
        } else {
            fprintf(stderr, "Could not open gc logfile: %s\n", logfile);
            fflush(stderr);
        }
    }
}

extern void
report_heap_exhaustion(long available, long requested, struct thread *th)
{
    if (gc_logfile) {
        FILE * log = fopen(gc_logfile, "a");
        if (log) {
            write_heap_exhaustion_report(log, available, requested, th);
            fclose(log);
        } else {
            fprintf(stderr, "Could not open gc logfile: %s\n", gc_logfile);
            fflush(stderr);
        }
    }
    /* Always to stderr as well. */
    write_heap_exhaustion_report(stderr, available, requested, th);
}
\f

#if defined(LISP_FEATURE_X86)
void fast_bzero(void*, size_t); /* in <arch>-assem.S */
#endif

/* Zero the pages from START to END (inclusive), but use mmap/munmap instead
 * if zeroing it ourselves, i.e. in practice give the memory back to the
 * OS. Generally done after a large GC.
 */
void zero_pages_with_mmap(page_index_t start, page_index_t end) {
    page_index_t i;
    void *addr = page_address(start), *new_addr;
    os_vm_size_t length = npage_bytes(1+end-start);

    if (start > end)
      return;

    gc_assert(length >= gencgc_release_granularity);
    gc_assert((length % gencgc_release_granularity) == 0);

#ifdef LISP_FEATURE_LINUX
    extern os_vm_address_t anon_dynamic_space_start;
    // We use MADV_DONTNEED only on Linux due to differing semantics from BSD.
    // Linux treats it as a demand that the memory be 0-filled, or refreshed
    // from a file that backs the range. BSD takes it as a hint that you don't
    // care if the memory has to brought in from swap when next accessed,
    // i.e. it's not a request to make a user-visible alteration to memory.
    // So in theory this can bring a page in from the core file, if we happen
    // to hit a page that resides in the portion of memory mapped by coreparse.
    // In practice this should not happen because objects from a core file can't
    // become garbage. Except in save-lisp-and-die they can, and we must be
    // cautious not to resurrect bytes that originally came from the file.
    if ((os_vm_address_t)addr >= anon_dynamic_space_start) {
        if (madvise(addr, length, MADV_DONTNEED) != 0)
            lose("madvise failed\n");
    } else
#endif
    {
        os_invalidate(addr, length);
        new_addr = os_validate(addr, length);
        if (new_addr == NULL || new_addr != addr) {
            lose("remap_free_pages: page moved, 0x%08x ==> 0x%08x",
                 start, new_addr);
        }
    }

    for (i = start; i <= end; i++) {
        page_table[i].need_to_zero = 0;
    }
}

/* Zero the pages from START to END (inclusive). Generally done just after
 * a new region has been allocated.
 */
static void
zero_pages(page_index_t start, page_index_t end) {
    if (start > end)
      return;

#if defined(LISP_FEATURE_X86)
    fast_bzero(page_address(start), npage_bytes(1+end-start));
#else
    bzero(page_address(start), npage_bytes(1+end-start));
#endif

}

static void
zero_and_mark_pages(page_index_t start, page_index_t end) {
    page_index_t i;

    zero_pages(start, end);
    for (i = start; i <= end; i++)
        page_table[i].need_to_zero = 0;
}

/* Zero the pages from START to END (inclusive), except for those
 * pages that are known to already zeroed. Mark all pages in the
 * ranges as non-zeroed.
 */
static void
zero_dirty_pages(page_index_t start, page_index_t end) {
    page_index_t i, j;

    for (i = start; i <= end; i++) {
        if (!page_table[i].need_to_zero) continue;
        for (j = i+1; (j <= end) && (page_table[j].need_to_zero); j++);
        zero_pages(i, j-1);
        i = j;
    }

    for (i = start; i <= end; i++) {
        page_table[i].need_to_zero = 1;
    }
}


/*
 * To support quick and inline allocation, regions of memory can be
 * allocated and then allocated from with just a free pointer and a
 * check against an end address.
 *
 * Since objects can be allocated to spaces with different properties
 * e.g. boxed/unboxed, generation, ages; there may need to be many
 * allocation regions.
 *
 * Each allocation region may start within a partly used page. Many
 * features of memory use are noted on a page wise basis, e.g. the
 * generation; so if a region starts within an existing allocated page
 * it must be consistent with this page.
 *
 * During the scavenging of the newspace, objects will be transported
 * into an allocation region, and pointers updated to point to this
 * allocation region. It is possible that these pointers will be
 * scavenged again before the allocation region is closed, e.g. due to
 * trans_list which jumps all over the place to cleanup the list. It
 * is important to be able to determine properties of all objects
 * pointed to when scavenging, e.g to detect pointers to the oldspace.
 * Thus it's important that the allocation regions have the correct
 * properties set when allocated, and not just set when closed. The
 * region allocation routines return regions with the specified
 * properties, and grab all the pages, setting their properties
 * appropriately, except that the amount used is not known.
 *
 * These regions are used to support quicker allocation using just a
 * free pointer. The actual space used by the region is not reflected
 * in the pages tables until it is closed. It can't be scavenged until
 * closed.
 *
 * When finished with the region it should be closed, which will
 * update the page tables for the actual space used returning unused
 * space. Further it may be noted in the new regions which is
 * necessary when scavenging the newspace.
 *
 * Large objects may be allocated directly without an allocation
 * region, the page tables are updated immediately.
 *
 * Unboxed objects don't contain pointers to other objects and so
 * don't need scavenging. Further they can't contain pointers to
 * younger generations so WP is not needed. By allocating pages to
 * unboxed objects the whole page never needs scavenging or
 * write-protecting. */

/* We are only using two regions at present. Both are for the current
 * newspace generation. */
struct alloc_region boxed_region;
struct alloc_region unboxed_region;

/* The generation currently being allocated to. */
static generation_index_t gc_alloc_generation;

static inline page_index_t
generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large)
{
    if (large) {
        if (UNBOXED_PAGE_FLAG == page_type_flag) {
            return generations[generation].alloc_large_unboxed_start_page;
        } else if (BOXED_PAGE_FLAG & page_type_flag) {
            /* Both code and data. */
            return generations[generation].alloc_large_start_page;
        } else {
            lose("bad page type flag: %d", page_type_flag);
        }
    } else {
        if (UNBOXED_PAGE_FLAG == page_type_flag) {
            return generations[generation].alloc_unboxed_start_page;
        } else if (BOXED_PAGE_FLAG & page_type_flag) {
            /* Both code and data. */
            return generations[generation].alloc_start_page;
        } else {
            lose("bad page_type_flag: %d", page_type_flag);
        }
    }
}

static inline void
set_generation_alloc_start_page(generation_index_t generation, int page_type_flag, int large,
                                page_index_t page)
{
    if (large) {
        if (UNBOXED_PAGE_FLAG == page_type_flag) {
            generations[generation].alloc_large_unboxed_start_page = page;
        } else if (BOXED_PAGE_FLAG & page_type_flag) {
            /* Both code and data. */
            generations[generation].alloc_large_start_page = page;
        } else {
            lose("bad page type flag: %d", page_type_flag);
        }
    } else {
        if (UNBOXED_PAGE_FLAG == page_type_flag) {
            generations[generation].alloc_unboxed_start_page = page;
        } else if (BOXED_PAGE_FLAG & page_type_flag) {
            /* Both code and data. */
            generations[generation].alloc_start_page = page;
        } else {
            lose("bad page type flag: %d", page_type_flag);
        }
    }
}

const int n_dwords_in_card = GENCGC_CARD_BYTES / N_WORD_BYTES / 2;
in_use_marker_t *
pinned_dwords(page_index_t page)
{
    if (page_table[page].has_pin_map)
        return &page_table_pinned_dwords[page * (n_dwords_in_card/N_WORD_BITS)];
    return NULL;
}

/* Find a new region with room for at least the given number of bytes.
 *
 * It starts looking at the current generation's alloc_start_page. So
 * may pick up from the previous region if there is enough space. This
 * keeps the allocation contiguous when scavenging the newspace.
 *
 * The alloc_region should have been closed by a call to
 * gc_alloc_update_page_tables(), and will thus be in an empty state.
 *
 * To assist the scavenging functions write-protected pages are not
 * used. Free pages should not be write-protected.
 *
 * It is critical to the conservative GC that the start of regions be
 * known. To help achieve this only small regions are allocated at a
 * time.
 *
 * During scavenging, pointers may be found to within the current
 * region and the page generation must be set so that pointers to the
 * from space can be recognized. Therefore the generation of pages in
 * the region are set to gc_alloc_generation. To prevent another
 * allocation call using the same pages, all the pages in the region
 * are allocated, although they will initially be empty.
 */
static void
gc_alloc_new_region(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
{
    page_index_t first_page;
    page_index_t last_page;
    os_vm_size_t bytes_found;
    page_index_t i;
    int ret;

    /*
    FSHOW((stderr,
           "/alloc_new_region for %d bytes from gen %d\n",
           nbytes, gc_alloc_generation));
    */

    /* Check that the region is in a reset state. */
    gc_assert((alloc_region->first_page == 0)
              && (alloc_region->last_page == -1)
              && (alloc_region->free_pointer == alloc_region->end_addr));
    ret = thread_mutex_lock(&free_pages_lock);
    gc_assert(ret == 0);
    first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0);
    last_page=gc_find_freeish_pages(&first_page, nbytes, page_type_flag);
    bytes_found=(GENCGC_CARD_BYTES - page_table[first_page].bytes_used)
            + npage_bytes(last_page-first_page);

    /* Set up the alloc_region. */
    alloc_region->first_page = first_page;
    alloc_region->last_page = last_page;
    alloc_region->start_addr = page_table[first_page].bytes_used
        + page_address(first_page);
    alloc_region->free_pointer = alloc_region->start_addr;
    alloc_region->end_addr = alloc_region->start_addr + bytes_found;

    /* Set up the pages. */

    /* The first page may have already been in use. */
    if (page_table[first_page].bytes_used == 0) {
        page_table[first_page].allocated = page_type_flag;
        page_table[first_page].gen = gc_alloc_generation;
        page_table[first_page].large_object = 0;
        page_table[first_page].scan_start_offset = 0;
        // wiping should have free()ed and :=NULL
        gc_assert(pinned_dwords(first_page) == NULL);
    }

    gc_assert(page_table[first_page].allocated == page_type_flag);
    page_table[first_page].allocated |= OPEN_REGION_PAGE_FLAG;

    gc_assert(page_table[first_page].gen == gc_alloc_generation);
    gc_assert(page_table[first_page].large_object == 0);

    for (i = first_page+1; i <= last_page; i++) {
        page_table[i].allocated = page_type_flag;
        page_table[i].gen = gc_alloc_generation;
        page_table[i].large_object = 0;
        /* This may not be necessary for unboxed regions (think it was
         * broken before!) */
        page_table[i].scan_start_offset =
            void_diff(page_address(i),alloc_region->start_addr);
        page_table[i].allocated |= OPEN_REGION_PAGE_FLAG ;
    }
    /* Bump up last_free_page. */
    if (last_page+1 > last_free_page) {
        last_free_page = last_page+1;
        /* do we only want to call this on special occasions? like for
         * boxed_region? */
        set_alloc_pointer((lispobj)page_address(last_free_page));
    }
    ret = thread_mutex_unlock(&free_pages_lock);
    gc_assert(ret == 0);

#ifdef READ_PROTECT_FREE_PAGES
    os_protect(page_address(first_page),
               npage_bytes(1+last_page-first_page),
               OS_VM_PROT_ALL);
#endif

    /* If the first page was only partial, don't check whether it's
     * zeroed (it won't be) and don't zero it (since the parts that
     * we're interested in are guaranteed to be zeroed).
     */
    if (page_table[first_page].bytes_used) {
        first_page++;
    }

    zero_dirty_pages(first_page, last_page);

    /* we can do this after releasing free_pages_lock */
    if (gencgc_zero_check) {
        word_t *p;
        for (p = (word_t *)alloc_region->start_addr;
             p < (word_t *)alloc_region->end_addr; p++) {
            if (*p != 0) {
                lose("The new region is not zero at %p (start=%p, end=%p).\n",
                     p, alloc_region->start_addr, alloc_region->end_addr);
            }
        }
    }
}

/* If the record_new_objects flag is 2 then all new regions created
 * are recorded.
 *
 * If it's 1 then then it is only recorded if the first page of the
 * current region is <= new_areas_ignore_page. This helps avoid
 * unnecessary recording when doing full scavenge pass.
 *
 * The new_object structure holds the page, byte offset, and size of
 * new regions of objects. Each new area is placed in the array of
 * these structures pointer to by new_areas. new_areas_index holds the
 * offset into new_areas.
 *
 * If new_area overflows NUM_NEW_AREAS then it stops adding them. The
 * later code must detect this and handle it, probably by doing a full
 * scavenge of a generation. */
#define NUM_NEW_AREAS 512
static int record_new_objects = 0;
static page_index_t new_areas_ignore_page;
struct new_area {
    page_index_t page;
    size_t offset;
    size_t size;
};
static struct new_area (*new_areas)[];
static size_t new_areas_index;
size_t max_new_areas;

/* Add a new area to new_areas. */
static void
add_new_area(page_index_t first_page, size_t offset, size_t size)
{
    size_t new_area_start, c;
    ssize_t i;

    /* Ignore if full. */
    if (new_areas_index >= NUM_NEW_AREAS)
        return;

    switch (record_new_objects) {
    case 0:
        return;
    case 1:
        if (first_page > new_areas_ignore_page)
            return;
        break;
    case 2:
        break;
    default:
        gc_abort();
    }

    new_area_start = npage_bytes(first_page) + offset;

    /* Search backwards for a prior area that this follows from. If
       found this will save adding a new area. */
    for (i = new_areas_index-1, c = 0; (i >= 0) && (c < 8); i--, c++) {
        size_t area_end =
            npage_bytes((*new_areas)[i].page)
            + (*new_areas)[i].offset
            + (*new_areas)[i].size;
        /*FSHOW((stderr,
               "/add_new_area S1 %d %d %d %d\n",
               i, c, new_area_start, area_end));*/
        if (new_area_start == area_end) {
            /*FSHOW((stderr,
                   "/adding to [%d] %d %d %d with %d %d %d:\n",
                   i,
                   (*new_areas)[i].page,
                   (*new_areas)[i].offset,
                   (*new_areas)[i].size,
                   first_page,
                   offset,
                    size);*/
            (*new_areas)[i].size += size;
            return;
        }
    }

    (*new_areas)[new_areas_index].page = first_page;
    (*new_areas)[new_areas_index].offset = offset;
    (*new_areas)[new_areas_index].size = size;
    /*FSHOW((stderr,
           "/new_area %d page %d offset %d size %d\n",
           new_areas_index, first_page, offset, size));*/
    new_areas_index++;

    /* Note the max new_areas used. */
    if (new_areas_index > max_new_areas)
        max_new_areas = new_areas_index;
}

/* Update the tables for the alloc_region. The region may be added to
 * the new_areas.
 *
 * When done the alloc_region is set up so that the next quick alloc
 * will fail safely and thus a new region will be allocated. Further
 * it is safe to try to re-update the page table of this reset
 * alloc_region. */
void
gc_alloc_update_page_tables(int page_type_flag, struct alloc_region *alloc_region)
{
    boolean more;
    page_index_t first_page;
    page_index_t next_page;
    os_vm_size_t bytes_used;
    os_vm_size_t region_size;
    os_vm_size_t byte_cnt;
    page_bytes_t orig_first_page_bytes_used;
    int ret;


    first_page = alloc_region->first_page;

    /* Catch an unused alloc_region. */
    if ((first_page == 0) && (alloc_region->last_page == -1))
        return;

    next_page = first_page+1;

    ret = thread_mutex_lock(&free_pages_lock);
    gc_assert(ret == 0);
    if (alloc_region->free_pointer != alloc_region->start_addr) {
        /* some bytes were allocated in the region */
        orig_first_page_bytes_used = page_table[first_page].bytes_used;

        gc_assert(alloc_region->start_addr ==
                  (page_address(first_page)
                   + page_table[first_page].bytes_used));

        /* All the pages used need to be updated */

        /* Update the first page. */

        /* If the page was free then set up the gen, and
         * scan_start_offset. */
        if (page_table[first_page].bytes_used == 0)
            gc_assert(page_starts_contiguous_block_p(first_page));
        page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);

        gc_assert(page_table[first_page].allocated & page_type_flag);
        gc_assert(page_table[first_page].gen == gc_alloc_generation);
        gc_assert(page_table[first_page].large_object == 0);

        byte_cnt = 0;

        /* Calculate the number of bytes used in this page. This is not
         * always the number of new bytes, unless it was free. */
        more = 0;
        if ((bytes_used = void_diff(alloc_region->free_pointer,
                                    page_address(first_page)))
            >GENCGC_CARD_BYTES) {
            bytes_used = GENCGC_CARD_BYTES;
            more = 1;
        }
        page_table[first_page].bytes_used = bytes_used;
        byte_cnt += bytes_used;


        /* All the rest of the pages should be free. We need to set
         * their scan_start_offset pointer to the start of the
         * region, and set the bytes_used. */
        while (more) {
            page_table[next_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
            gc_assert(page_table[next_page].allocated & page_type_flag);
            gc_assert(page_table[next_page].bytes_used == 0);
            gc_assert(page_table[next_page].gen == gc_alloc_generation);
            gc_assert(page_table[next_page].large_object == 0);

            gc_assert(page_table[next_page].scan_start_offset ==
                      void_diff(page_address(next_page),
                                alloc_region->start_addr));

            /* Calculate the number of bytes used in this page. */
            more = 0;
            if ((bytes_used = void_diff(alloc_region->free_pointer,
                                        page_address(next_page)))>GENCGC_CARD_BYTES) {
                bytes_used = GENCGC_CARD_BYTES;
                more = 1;
            }
            page_table[next_page].bytes_used = bytes_used;
            byte_cnt += bytes_used;

            next_page++;
        }

        region_size = void_diff(alloc_region->free_pointer,
                                alloc_region->start_addr);
        bytes_allocated += region_size;
        generations[gc_alloc_generation].bytes_allocated += region_size;

        gc_assert((byte_cnt- orig_first_page_bytes_used) == region_size);

        /* Set the generations alloc restart page to the last page of
         * the region. */
        set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 0, next_page-1);

        /* Add the region to the new_areas if requested. */
        if (BOXED_PAGE_FLAG & page_type_flag)
            add_new_area(first_page,orig_first_page_bytes_used, region_size);

        /*
        FSHOW((stderr,
               "/gc_alloc_update_page_tables update %d bytes to gen %d\n",
               region_size,
               gc_alloc_generation));
        */
    } else {
        /* There are no bytes allocated. Unallocate the first_page if
         * there are 0 bytes_used. */
        page_table[first_page].allocated &= ~(OPEN_REGION_PAGE_FLAG);
        if (page_table[first_page].bytes_used == 0)
            page_table[first_page].allocated = FREE_PAGE_FLAG;
    }

    /* Unallocate any unused pages. */
    while (next_page <= alloc_region->last_page) {
        gc_assert(page_table[next_page].bytes_used == 0);
        page_table[next_page].allocated = FREE_PAGE_FLAG;
        next_page++;
    }
    ret = thread_mutex_unlock(&free_pages_lock);
    gc_assert(ret == 0);

    /* alloc_region is per-thread, we're ok to do this unlocked */
    gc_set_region_empty(alloc_region);
}

/* Allocate a possibly large object. */
void *
gc_alloc_large(sword_t nbytes, int page_type_flag, struct alloc_region *alloc_region)
{
    boolean more;
    page_index_t first_page, next_page, last_page;
    page_bytes_t orig_first_page_bytes_used;
    os_vm_size_t byte_cnt;
    os_vm_size_t bytes_used;
    int ret;

    ret = thread_mutex_lock(&free_pages_lock);
    gc_assert(ret == 0);

    first_page = generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1);
    if (first_page <= alloc_region->last_page) {
        first_page = alloc_region->last_page+1;
    }

    last_page=gc_find_freeish_pages(&first_page,nbytes, page_type_flag);

    gc_assert(first_page > alloc_region->last_page);

    set_generation_alloc_start_page(gc_alloc_generation, page_type_flag, 1, last_page);

    /* Set up the pages. */
    orig_first_page_bytes_used = page_table[first_page].bytes_used;

    /* If the first page was free then set up the gen, and
     * scan_start_offset. */
    if (page_table[first_page].bytes_used == 0) {
        page_table[first_page].allocated = page_type_flag;
        page_table[first_page].gen = gc_alloc_generation;
        page_table[first_page].scan_start_offset = 0;
        page_table[first_page].large_object = 1;
    }

    gc_assert(page_table[first_page].allocated == page_type_flag);
    gc_assert(page_table[first_page].gen == gc_alloc_generation);
    gc_assert(page_table[first_page].large_object == 1);

    byte_cnt = 0;

    /* Calc. the number of bytes used in this page. This is not
     * always the number of new bytes, unless it was free. */
    more = 0;
    if ((bytes_used = nbytes+orig_first_page_bytes_used) > GENCGC_CARD_BYTES) {
        bytes_used = GENCGC_CARD_BYTES;
        more = 1;
    }
    page_table[first_page].bytes_used = bytes_used;
    byte_cnt += bytes_used;

    next_page = first_page+1;

    /* All the rest of the pages should be free. We need to set their
     * scan_start_offset pointer to the start of the region, and set
     * the bytes_used. */
    while (more) {
        gc_assert(page_free_p(next_page));
        gc_assert(page_table[next_page].bytes_used == 0);
        page_table[next_page].allocated = page_type_flag;
        page_table[next_page].gen = gc_alloc_generation;
        page_table[next_page].large_object = 1;

        page_table[next_page].scan_start_offset =
            npage_bytes(next_page-first_page) - orig_first_page_bytes_used;

        /* Calculate the number of bytes used in this page. */
        more = 0;
        bytes_used=(nbytes+orig_first_page_bytes_used)-byte_cnt;
        if (bytes_used > GENCGC_CARD_BYTES) {
            bytes_used = GENCGC_CARD_BYTES;
            more = 1;
        }
        page_table[next_page].bytes_used = bytes_used;
        page_table[next_page].write_protected=0;
        page_table[next_page].dont_move=0;
        byte_cnt += bytes_used;
        next_page++;
    }

    gc_assert((byte_cnt-orig_first_page_bytes_used) == (size_t)nbytes);

    bytes_allocated += nbytes;
    generations[gc_alloc_generation].bytes_allocated += nbytes;

    /* Add the region to the new_areas if requested. */
    if (BOXED_PAGE_FLAG & page_type_flag)
        add_new_area(first_page,orig_first_page_bytes_used,nbytes);

    /* Bump up last_free_page */
    if (last_page+1 > last_free_page) {
        last_free_page = last_page+1;
        set_alloc_pointer((lispobj)(page_address(last_free_page)));
    }
    ret = thread_mutex_unlock(&free_pages_lock);
    gc_assert(ret == 0);

#ifdef READ_PROTECT_FREE_PAGES
    os_protect(page_address(first_page),
               npage_bytes(1+last_page-first_page),
               OS_VM_PROT_ALL);
#endif

    zero_dirty_pages(first_page, last_page);

    return page_address(first_page);
}

static page_index_t gencgc_alloc_start_page = -1;

void
gc_heap_exhausted_error_or_lose (sword_t available, sword_t requested)
{
    struct thread *thread = arch_os_get_current_thread();
    /* Write basic information before doing anything else: if we don't
     * call to lisp this is a must, and even if we do there is always
     * the danger that we bounce back here before the error has been
     * handled, or indeed even printed.
     */
    report_heap_exhaustion(available, requested, thread);
    if (gc_active_p || (available == 0)) {
        /* If we are in GC, or totally out of memory there is no way
         * to sanely transfer control to the lisp-side of things.
         */
        lose("Heap exhausted, game over.");
    }
    else {
        /* FIXME: assert free_pages_lock held */
        (void)thread_mutex_unlock(&free_pages_lock);
#if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
        gc_assert(get_pseudo_atomic_atomic(thread));
        clear_pseudo_atomic_atomic(thread);
        if (get_pseudo_atomic_interrupted(thread))
            do_pending_interrupt();
#endif
        /* Another issue is that signalling HEAP-EXHAUSTED error leads
         * to running user code at arbitrary places, even in a
         * WITHOUT-INTERRUPTS which may lead to a deadlock without
         * running out of the heap. So at this point all bets are
         * off. */
        if (SymbolValue(INTERRUPTS_ENABLED,thread) == NIL)
            corruption_warning_and_maybe_lose
                ("Signalling HEAP-EXHAUSTED in a WITHOUT-INTERRUPTS.");
        /* available and requested should be double word aligned, thus
           they can passed as fixnums and shifted later. */
        funcall2(StaticSymbolFunction(HEAP_EXHAUSTED_ERROR), available, requested);
        lose("HEAP-EXHAUSTED-ERROR fell through");
    }
}

page_index_t
gc_find_freeish_pages(page_index_t *restart_page_ptr, sword_t bytes,
                      int page_type_flag)
{
    page_index_t most_bytes_found_from = 0, most_bytes_found_to = 0;
    page_index_t first_page, last_page, restart_page = *restart_page_ptr;
    os_vm_size_t nbytes = bytes;
    os_vm_size_t nbytes_goal = nbytes;
    os_vm_size_t bytes_found = 0;
    os_vm_size_t most_bytes_found = 0;
    boolean small_object = nbytes < GENCGC_CARD_BYTES;
    /* FIXME: assert(free_pages_lock is held); */

    if (nbytes_goal < gencgc_alloc_granularity)
        nbytes_goal = gencgc_alloc_granularity;

    /* Toggled by gc_and_save for heap compaction, normally -1. */
    if (gencgc_alloc_start_page != -1) {
        restart_page = gencgc_alloc_start_page;
    }

    /* FIXME: This is on bytes instead of nbytes pending cleanup of
     * long from the interface. */
    gc_assert(bytes>=0);
    /* Search for a page with at least nbytes of space. We prefer
     * not to split small objects on multiple pages, to reduce the
     * number of contiguous allocation regions spaning multiple
     * pages: this helps avoid excessive conservativism.
     *
     * For other objects, we guarantee that they start on their own
     * page boundary.
     */
    first_page = restart_page;
    while (first_page < page_table_pages) {
        bytes_found = 0;
        if (page_free_p(first_page)) {
                gc_assert(0 == page_table[first_page].bytes_used);
                bytes_found = GENCGC_CARD_BYTES;
        } else if (small_object &&
                   (page_table[first_page].allocated == page_type_flag) &&
                   (page_table[first_page].large_object == 0) &&
                   (page_table[first_page].gen == gc_alloc_generation) &&
                   (page_table[first_page].write_protected == 0) &&
                   (page_table[first_page].dont_move == 0)) {
            bytes_found = GENCGC_CARD_BYTES - page_table[first_page].bytes_used;
            if (bytes_found < nbytes) {
                if (bytes_found > most_bytes_found)
                    most_bytes_found = bytes_found;
                first_page++;
                continue;
            }
        } else {
            first_page++;
            continue;
        }

        gc_assert(page_table[first_page].write_protected == 0);
        for (last_page = first_page+1;
             ((last_page < page_table_pages) &&
              page_free_p(last_page) &&
              (bytes_found < nbytes_goal));
             last_page++) {
            bytes_found += GENCGC_CARD_BYTES;
            gc_assert(0 == page_table[last_page].bytes_used);
            gc_assert(0 == page_table[last_page].write_protected);
        }

        if (bytes_found > most_bytes_found) {
            most_bytes_found = bytes_found;
            most_bytes_found_from = first_page;
            most_bytes_found_to = last_page;
        }
        if (bytes_found >= nbytes_goal)
            break;

        first_page = last_page;
    }

    bytes_found = most_bytes_found;
    restart_page = first_page + 1;

    /* Check for a failure */
    if (bytes_found < nbytes) {
        gc_assert(restart_page >= page_table_pages);
        gc_heap_exhausted_error_or_lose(most_bytes_found, nbytes);
    }

    gc_assert(most_bytes_found_to);
    *restart_page_ptr = most_bytes_found_from;
    return most_bytes_found_to-1;
}

/* Allocate bytes.  All the rest of the special-purpose allocation
 * functions will eventually call this  */

void *
gc_alloc_with_region(sword_t nbytes,int page_type_flag, struct alloc_region *my_region,
                     int quick_p)
{
    void *new_free_pointer;

    if (nbytes>=LARGE_OBJECT_SIZE)
        return gc_alloc_large(nbytes, page_type_flag, my_region);

    /* Check whether there is room in the current alloc region. */
    new_free_pointer = my_region->free_pointer + nbytes;

    /* fprintf(stderr, "alloc %d bytes from %p to %p\n", nbytes,
       my_region->free_pointer, new_free_pointer); */

    if (new_free_pointer <= my_region->end_addr) {
        /* If so then allocate from the current alloc region. */
        void *new_obj = my_region->free_pointer;
        my_region->free_pointer = new_free_pointer;

        /* Unless a `quick' alloc was requested, check whether the
           alloc region is almost empty. */
        if (!quick_p &&
            void_diff(my_region->end_addr,my_region->free_pointer) <= 32) {
            /* If so, finished with the current region. */
            gc_alloc_update_page_tables(page_type_flag, my_region);
            /* Set up a new region. */
            gc_alloc_new_region(32 /*bytes*/, page_type_flag, my_region);
        }

        return((void *)new_obj);
    }

    /* Else not enough free space in the current region: retry with a
     * new region. */

    gc_alloc_update_page_tables(page_type_flag, my_region);
    gc_alloc_new_region(nbytes, page_type_flag, my_region);
    return gc_alloc_with_region(nbytes, page_type_flag, my_region,0);
}
\f
/* Copy a large object. If the object is in a large object region then
 * it is simply promoted, else it is copied. If it's large enough then
 * it's copied to a large object region.
 *
 * Bignums and vectors may have shrunk. If the object is not copied
 * the space needs to be reclaimed, and the page_tables corrected. */
static lispobj
general_copy_large_object(lispobj object, word_t nwords, boolean boxedp)
{
    lispobj *new;
    page_index_t first_page;

    CHECK_COPY_PRECONDITIONS(object, nwords);

    if ((nwords > 1024*1024) && gencgc_verbose) {
        FSHOW((stderr, "/general_copy_large_object: %d bytes\n",
               nwords*N_WORD_BYTES));
    }

    /* Check whether it's a large object. */
    first_page = find_page_index((void *)object);
    gc_assert(first_page >= 0);

    if (page_table[first_page].large_object) {
        /* Promote the object. Note: Unboxed objects may have been
         * allocated to a BOXED region so it may be necessary to
         * change the region to UNBOXED. */
        os_vm_size_t remaining_bytes;
        os_vm_size_t bytes_freed;
        page_index_t next_page;
        page_bytes_t old_bytes_used;

        /* FIXME: This comment is somewhat stale.
         *
         * Note: Any page write-protection must be removed, else a
         * later scavenge_newspace may incorrectly not scavenge these
         * pages. This would not be necessary if they are added to the
         * new areas, but let's do it for them all (they'll probably
         * be written anyway?). */

        gc_assert(page_starts_contiguous_block_p(first_page));
        next_page = first_page;
        remaining_bytes = nwords*N_WORD_BYTES;

        while (remaining_bytes > GENCGC_CARD_BYTES) {
            gc_assert(page_table[next_page].gen == from_space);
            gc_assert(page_table[next_page].large_object);
            gc_assert(page_table[next_page].scan_start_offset ==
                      npage_bytes(next_page-first_page));
            gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);
            /* Should have been unprotected by unprotect_oldspace()
             * for boxed objects, and after promotion unboxed ones
             * should not be on protected pages at all. */
            gc_assert(!page_table[next_page].write_protected);

            if (boxedp)
                gc_assert(page_boxed_p(next_page));
            else {
                gc_assert(page_allocated_no_region_p(next_page));
                page_table[next_page].allocated = UNBOXED_PAGE_FLAG;
            }
            page_table[next_page].gen = new_space;

            remaining_bytes -= GENCGC_CARD_BYTES;
            next_page++;
        }

        /* Now only one page remains, but the object may have shrunk so
         * there may be more unused pages which will be freed. */

        /* Object may have shrunk but shouldn't have grown - check. */
        gc_assert(page_table[next_page].bytes_used >= remaining_bytes);

        page_table[next_page].gen = new_space;

        if (boxedp)
            gc_assert(page_boxed_p(next_page));
        else
            page_table[next_page].allocated = UNBOXED_PAGE_FLAG;

        /* Adjust the bytes_used. */
        old_bytes_used = page_table[next_page].bytes_used;
        page_table[next_page].bytes_used = remaining_bytes;

        bytes_freed = old_bytes_used - remaining_bytes;

        /* Free any remaining pages; needs care. */
        next_page++;
        while ((old_bytes_used == GENCGC_CARD_BYTES) &&
               (page_table[next_page].gen == from_space) &&
               /* FIXME: It is not obvious to me why this is necessary
                * as a loop condition: it seems to me that the
                * scan_start_offset test should be sufficient, but
                * experimentally that is not the case. --NS
                * 2011-11-28 */
               (boxedp ?
                page_boxed_p(next_page) :
                page_allocated_no_region_p(next_page)) &&
               page_table[next_page].large_object &&
               (page_table[next_page].scan_start_offset ==
                npage_bytes(next_page - first_page))) {
            /* Checks out OK, free the page. Don't need to both zeroing
             * pages as this should have been done before shrinking the
             * object. These pages shouldn't be write-protected, even if
             * boxed they should be zero filled. */
            gc_assert(page_table[next_page].write_protected == 0);

            old_bytes_used = page_table[next_page].bytes_used;
            page_table[next_page].allocated = FREE_PAGE_FLAG;
            page_table[next_page].bytes_used = 0;
            bytes_freed += old_bytes_used;
            next_page++;
        }

        if ((bytes_freed > 0) && gencgc_verbose) {
            FSHOW((stderr,
                   "/general_copy_large_object bytes_freed=%"OS_VM_SIZE_FMT"\n",
                   bytes_freed));
        }

        generations[from_space].bytes_allocated -= nwords*N_WORD_BYTES
            + bytes_freed;
        generations[new_space].bytes_allocated += nwords*N_WORD_BYTES;
        bytes_allocated -= bytes_freed;

        /* Add the region to the new_areas if requested. */
        if (boxedp)
            add_new_area(first_page,0,nwords*N_WORD_BYTES);

        return(object);

    } else {
        /* Allocate space. */
        new = gc_general_alloc(nwords*N_WORD_BYTES,
                               (boxedp ? BOXED_PAGE_FLAG : UNBOXED_PAGE_FLAG),
                               ALLOC_QUICK);

        /* Copy the object. */
        memcpy(new,native_pointer(object),nwords*N_WORD_BYTES);

        /* Return Lisp pointer of new object. */
        return make_lispobj(new, lowtag_of(object));
    }
}

lispobj
copy_large_object(lispobj object, sword_t nwords)
{
    return general_copy_large_object(object, nwords, 1);
}

lispobj
copy_large_unboxed_object(lispobj object, sword_t nwords)
{
    return general_copy_large_object(object, nwords, 0);
}

/* to copy unboxed objects */
lispobj
copy_unboxed_object(lispobj object, sword_t nwords)
{
    return gc_general_copy_object(object, nwords, UNBOXED_PAGE_FLAG);
}
\f

/*
 * code and code-related objects
 */
/*
static lispobj trans_fun_header(lispobj object);
static lispobj trans_boxed(lispobj object);
*/

/* Scan a x86 compiled code object, looking for possible fixups that
 * have been missed after a move.
 *
 * Two types of fixups are needed:
 * 1. Absolute fixups to within the code object.
 * 2. Relative fixups to outside the code object.
 *
 * Currently only absolute fixups to the constant vector, or to the
 * code area are checked. */
#ifdef LISP_FEATURE_X86
void
sniff_code_object(struct code *code, os_vm_size_t displacement)
{
    sword_t nheader_words, ncode_words, nwords;
    os_vm_address_t constants_start_addr = NULL, constants_end_addr, p;
    os_vm_address_t code_start_addr, code_end_addr;
    os_vm_address_t code_addr = (os_vm_address_t)code;
    int fixup_found = 0;

    if (!check_code_fixups)
        return;

    FSHOW((stderr, "/sniffing code: %p, %lu\n", code, displacement));

    ncode_words = code_instruction_words(code->code_size);
    nheader_words = code_header_words(*(lispobj *)code);
    nwords = ncode_words + nheader_words;

    constants_start_addr = code_addr + 5*N_WORD_BYTES;
    constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
    code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
    code_end_addr = code_addr + nwords*N_WORD_BYTES;

    /* Work through the unboxed code. */
    for (p = code_start_addr; p < code_end_addr; p++) {
        void *data = *(void **)p;
        unsigned d1 = *((unsigned char *)p - 1);
        unsigned d2 = *((unsigned char *)p - 2);
        unsigned d3 = *((unsigned char *)p - 3);
        unsigned d4 = *((unsigned char *)p - 4);
#if QSHOW
        unsigned d5 = *((unsigned char *)p - 5);
        unsigned d6 = *((unsigned char *)p - 6);
#endif

        /* Check for code references. */
        /* Check for a 32 bit word that looks like an absolute
           reference to within the code adea of the code object. */
        if ((data >= (void*)(code_start_addr-displacement))
            && (data < (void*)(code_end_addr-displacement))) {
            /* function header */
            if ((d4 == 0x5e)
                && (((unsigned)p - 4 - 4*HeaderValue(*((unsigned *)p-1))) ==
                    (unsigned)code)) {
                /* Skip the function header */
                p += 6*4 - 4 - 1;
                continue;
            }
            /* the case of PUSH imm32 */
            if (d1 == 0x68) {
                fixup_found = 1;
                FSHOW((stderr,
                       "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                       p, d6, d5, d4, d3, d2, d1, data));
                FSHOW((stderr, "/PUSH $0x%.8x\n", data));
            }
            /* the case of MOV [reg-8],imm32 */
            if ((d3 == 0xc7)
                && (d2==0x40 || d2==0x41 || d2==0x42 || d2==0x43
                    || d2==0x45 || d2==0x46 || d2==0x47)
                && (d1 == 0xf8)) {
                fixup_found = 1;
                FSHOW((stderr,
                       "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                       p, d6, d5, d4, d3, d2, d1, data));
                FSHOW((stderr, "/MOV [reg-8],$0x%.8x\n", data));
            }
            /* the case of LEA reg,[disp32] */
            if ((d2 == 0x8d) && ((d1 & 0xc7) == 5)) {
                fixup_found = 1;
                FSHOW((stderr,
                       "/code ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                       p, d6, d5, d4, d3, d2, d1, data));
                FSHOW((stderr,"/LEA reg,[$0x%.8x]\n", data));
            }
        }

        /* Check for constant references. */
        /* Check for a 32 bit word that looks like an absolute
           reference to within the constant vector. Constant references
           will be aligned. */
        if ((data >= (void*)(constants_start_addr-displacement))
            && (data < (void*)(constants_end_addr-displacement))
            && (((unsigned)data & 0x3) == 0)) {
            /*  Mov eax,m32 */
            if (d1 == 0xa1) {
                fixup_found = 1;
                FSHOW((stderr,
                       "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                       p, d6, d5, d4, d3, d2, d1, data));
                FSHOW((stderr,"/MOV eax,0x%.8x\n", data));
            }

            /*  the case of MOV m32,EAX */
            if (d1 == 0xa3) {
                fixup_found = 1;
                FSHOW((stderr,
                       "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                       p, d6, d5, d4, d3, d2, d1, data));
                FSHOW((stderr, "/MOV 0x%.8x,eax\n", data));
            }

            /* the case of CMP m32,imm32 */
            if ((d1 == 0x3d) && (d2 == 0x81)) {
                fixup_found = 1;
                FSHOW((stderr,
                       "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                       p, d6, d5, d4, d3, d2, d1, data));
                /* XX Check this */
                FSHOW((stderr, "/CMP 0x%.8x,immed32\n", data));
            }

            /* Check for a mod=00, r/m=101 byte. */
            if ((d1 & 0xc7) == 5) {
                /* Cmp m32,reg */
                if (d2 == 0x39) {
                    fixup_found = 1;
                    FSHOW((stderr,
                           "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                           p, d6, d5, d4, d3, d2, d1, data));
                    FSHOW((stderr,"/CMP 0x%.8x,reg\n", data));
                }
                /* the case of CMP reg32,m32 */
                if (d2 == 0x3b) {
                    fixup_found = 1;
                    FSHOW((stderr,
                           "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                           p, d6, d5, d4, d3, d2, d1, data));
                    FSHOW((stderr, "/CMP reg32,0x%.8x\n", data));
                }
                /* the case of MOV m32,reg32 */
                if (d2 == 0x89) {
                    fixup_found = 1;
                    FSHOW((stderr,
                           "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                           p, d6, d5, d4, d3, d2, d1, data));
                    FSHOW((stderr, "/MOV 0x%.8x,reg32\n", data));
                }
                /* the case of MOV reg32,m32 */
                if (d2 == 0x8b) {
                    fixup_found = 1;
                    FSHOW((stderr,
                           "/abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                           p, d6, d5, d4, d3, d2, d1, data));
                    FSHOW((stderr, "/MOV reg32,0x%.8x\n", data));
                }
                /* the case of LEA reg32,m32 */
                if (d2 == 0x8d) {
                    fixup_found = 1;
                    FSHOW((stderr,
                           "abs const ref @%x: %.2x %.2x %.2x %.2x %.2x %.2x (%.8x)\n",
                           p, d6, d5, d4, d3, d2, d1, data));
                    FSHOW((stderr, "/LEA reg32,0x%.8x\n", data));
                }
            }
        }
    }

    /* If anything was found, print some information on the code
     * object. */
    if (fixup_found) {
        FSHOW((stderr,
               "/compiled code object at %x: header words = %d, code words = %d\n",
               code, nheader_words, ncode_words));
        FSHOW((stderr,
               "/const start = %x, end = %x\n",
               constants_start_addr, constants_end_addr));
        FSHOW((stderr,
               "/code start = %x, end = %x\n",
               code_start_addr, code_end_addr));
    }
}
#endif

#ifdef LISP_FEATURE_X86
void
gencgc_apply_code_fixups(struct code *old_code, struct code *new_code)
{
    sword_t nheader_words, ncode_words, nwords;
    os_vm_address_t __attribute__((unused)) constants_start_addr, constants_end_addr;
    os_vm_address_t __attribute__((unused)) code_start_addr, code_end_addr;
    os_vm_address_t code_addr = (os_vm_address_t)new_code;
    os_vm_address_t old_addr = (os_vm_address_t)old_code;
    os_vm_size_t displacement = code_addr - old_addr;
    lispobj fixups = NIL;
    struct vector *fixups_vector;

    ncode_words = code_instruction_words(new_code->code_size);
    nheader_words = code_header_words(*(lispobj *)new_code);
    nwords = ncode_words + nheader_words;
    /* FSHOW((stderr,
             "/compiled code object at %x: header words = %d, code words = %d\n",
             new_code, nheader_words, ncode_words)); */
    constants_start_addr = code_addr + 5*N_WORD_BYTES;
    constants_end_addr = code_addr + nheader_words*N_WORD_BYTES;
    code_start_addr = code_addr + nheader_words*N_WORD_BYTES;
    code_end_addr = code_addr + nwords*N_WORD_BYTES;
    /*
    FSHOW((stderr,
           "/const start = %x, end = %x\n",
           constants_start_addr,constants_end_addr));
    FSHOW((stderr,
           "/code start = %x; end = %x\n",
           code_start_addr,code_end_addr));
    */

    fixups = new_code->fixups;
    /* It will be a Lisp vector if valid, or 0 if there are no fixups */
    if (fixups == 0 || !is_lisp_pointer(fixups)) {
        /* Check for possible errors. */
        if (check_code_fixups)
            sniff_code_object(new_code, displacement);

        return;
    }

    fixups_vector = (struct vector *)native_pointer(fixups);

    /* Could be pointing to a forwarding pointer. */
    /* This is extremely unlikely, because the only referent of the fixups
       is usually the code itself; so scavenging the vector won't occur
       until after the code object is known to be live. As we're just now
       enlivening the code, the fixups shouldn't have been forwarded.
       Maybe the vector is on the special binding stack though ... */
    if (is_lisp_pointer(fixups) &&
        (find_page_index((void*)fixups_vector) != -1) &&
        forwarding_pointer_p((lispobj*)fixups_vector))  {
        /* If so, then follow it. */
        /*SHOW("following pointer to a forwarding pointer");*/
        fixups_vector = (struct vector *)
          native_pointer(forwarding_pointer_value((lispobj*)fixups_vector));
    }

    /*SHOW("got fixups");*/

    if (widetag_of(fixups_vector->header) == SIMPLE_ARRAY_WORD_WIDETAG) {
        /* Got the fixups for the code block. Now work through the vector,
           and apply a fixup at each address. */
        sword_t length = fixnum_value(fixups_vector->length);
        sword_t i;
        for (i = 0; i < length; i++) {
            long offset = fixups_vector->data[i];
            /* Now check the current value of offset. */
            os_vm_address_t old_value = *(os_vm_address_t *)(code_start_addr + offset);

            /* If it's within the old_code object then it must be an
             * absolute fixup (relative ones are not saved) */
            if ((old_value >= old_addr)
                && (old_value < (old_addr + nwords*N_WORD_BYTES)))
                /* So add the dispacement. */
                *(os_vm_address_t *)(code_start_addr + offset) =
                    old_value + displacement;
            else
                /* It is outside the old code object so it must be a
                 * relative fixup (absolute fixups are not saved). So
                 * subtract the displacement. */
                *(os_vm_address_t *)(code_start_addr + offset) =
                    old_value - displacement;
        }
    } else {
        /* This used to just print a note to stderr, but a bogus fixup seems to
         * indicate real heap corruption, so a hard hailure is in order. */
        lose("fixup vector %p has a bad widetag: %d\n",
             fixups_vector, widetag_of(fixups_vector->header));
    }

    /* Check for possible errors. */
    if (check_code_fixups) {
        sniff_code_object(new_code,displacement);
    }
}
#endif

static lispobj
trans_boxed_large(lispobj object)
{
    gc_assert(is_lisp_pointer(object));
    return copy_large_object(object,
                             (HeaderValue(*native_pointer(object)) | 1) + 1);
}
\f
/*
 * weak pointers
 */

/* XX This is a hack adapted from cgc.c. These don't work too
 * efficiently with the gencgc as a list of the weak pointers is
 * maintained within the objects which causes writes to the pages. A
 * limited attempt is made to avoid unnecessary writes, but this needs
 * a re-think. */
static sword_t
scav_weak_pointer(lispobj *where, lispobj object)
{
    /* Since we overwrite the 'next' field, we have to make
     * sure not to do so for pointers already in the list.
     * Instead of searching the list of weak_pointers each
     * time, we ensure that next is always NULL when the weak
     * pointer isn't in the list, and not NULL otherwise.
     * Since we can't use NULL to denote end of list, we
     * use a pointer back to the same weak_pointer.
     */
    struct weak_pointer * wp = (struct weak_pointer*)where;

    if (NULL == wp->next) {
        wp->next = weak_pointers;
        weak_pointers = wp;
        if (NULL == wp->next)
            wp->next = wp;
    }

    /* Do not let GC scavenge the value slot of the weak pointer.
     * (That is why it is a weak pointer.) */

    return WEAK_POINTER_NWORDS;
}

\f
lispobj *
search_read_only_space(void *pointer)
{
    lispobj *start = (lispobj *) READ_ONLY_SPACE_START;
    lispobj *end = (lispobj *) SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0);
    if ((pointer < (void *)start) || (pointer >= (void *)end))
        return NULL;
    return gc_search_space(start, pointer);
}

lispobj *
search_static_space(void *pointer)
{
    lispobj *start = (lispobj *)STATIC_SPACE_START;
    lispobj *end = (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0);
    if ((pointer < (void *)start) || (pointer >= (void *)end))
        return NULL;
    return gc_search_space(start, pointer);
}

/* a faster version for searching the dynamic space. This will work even
 * if the object is in a current allocation region. */
lispobj *
search_dynamic_space(void *pointer)
{
    page_index_t page_index = find_page_index(pointer);
    lispobj *start;

    /* The address may be invalid, so do some checks. */
    if ((page_index == -1) || page_free_p(page_index))
        return NULL;
    start = (lispobj *)page_scan_start(page_index);
    return gc_search_space(start, pointer);
}

// Return the starting address of the object containing 'addr'
// if and only if the object is one which would be evacuated from 'from_space'
// were it allowed to be either discarded as garbage or moved.
// 'addr_page_index' is the page containing 'addr' and must not be -1.
// Return 0 if there is no such object - that is, if addr is past the
// end of the used bytes, or its pages are not in 'from_space' etc.
static lispobj*
conservative_root_p(void *addr, page_index_t addr_page_index)
{
#ifdef GENCGC_IS_PRECISE
    /* If we're in precise gencgc (non-x86oid as of this writing) then
     * we are only called on valid object pointers in the first place,
     * so we just have to do a bounds-check against the heap, a
     * generation check, and the already-pinned check. */
    if ((page_table[addr_page_index].gen != from_space)
        || (page_table[addr_page_index].dont_move != 0))
        return 0;
    return (lispobj*)1;
#else
    /* quick check 1: Address is quite likely to have been invalid. */
    if (page_free_p(addr_page_index)
        || (page_table[addr_page_index].bytes_used == 0)
        || (page_table[addr_page_index].gen != from_space))
        return 0;
    gc_assert(!(page_table[addr_page_index].allocated&OPEN_REGION_PAGE_FLAG));

    /* quick check 2: Check the offset within the page.
     *
     */
    if (((uword_t)addr & (GENCGC_CARD_BYTES - 1)) >
        page_table[addr_page_index].bytes_used)
        return 0;

    /* Filter out anything which can't be a pointer to a Lisp object
     * (or, as a special case which also requires dont_move, a return
     * address referring to something in a CodeObject). This is
     * expensive but important, since it vastly reduces the
     * probability that random garbage will be bogusly interpreted as
     * a pointer which prevents a page from moving. */
    lispobj* object_start = search_dynamic_space(addr);
    if (!object_start) return 0;

    /* If the containing object is a code object and 'addr' points
     * anywhere beyond the boxed words,
     * presume it to be a valid unboxed return address. */
    if (instruction_ptr_p(addr, object_start))
        return object_start;

    /* Large object pages only contain ONE object, and it will never
     * be a CONS.  However, arrays and bignums can be allocated larger
     * than necessary and then shrunk to fit, leaving what look like
     * (0 . 0) CONSes at the end.  These appear valid to
     * properly_tagged_descriptor_p(), so pick them off here. */
    if (((lowtag_of((lispobj)addr) == LIST_POINTER_LOWTAG) &&
         page_table[addr_page_index].large_object)
        || !properly_tagged_descriptor_p(addr, object_start))
        return 0;

    return object_start;
#endif
}

/* Adjust large bignum and vector objects. This will adjust the
 * allocated region if the size has shrunk, and move unboxed objects
 * into unboxed pages. The pages are not promoted here, and the
 * promoted region is not added to the new_regions; this is really
 * only designed to be called from preserve_pointer(). Shouldn't fail
 * if this is missed, just may delay the moving of objects to unboxed
 * pages, and the freeing of pages. */
static void
maybe_adjust_large_object(lispobj *where)
{
    page_index_t first_page;
    page_index_t next_page;
    sword_t nwords;

    uword_t remaining_bytes;
    uword_t bytes_freed;
    uword_t old_bytes_used;

    int boxed;

    /* Check whether it's a vector or bignum object. */
    switch (widetag_of(where[0])) {
    case SIMPLE_VECTOR_WIDETAG:
        boxed = BOXED_PAGE_FLAG;
        break;
    case BIGNUM_WIDETAG:
#include "genesis/specialized-vectors.inc"
        boxed = UNBOXED_PAGE_FLAG;
        break;
    default:
        return;
    }

    /* Find its current size. */
    nwords = (sizetab[widetag_of(where[0])])(where);

    first_page = find_page_index((void *)where);
    gc_assert(first_page >= 0);

    /* Note: Any page write-protection must be removed, else a later
     * scavenge_newspace may incorrectly not scavenge these pages.
     * This would not be necessary if they are added to the new areas,
     * but lets do it for them all (they'll probably be written
     * anyway?). */

    gc_assert(page_starts_contiguous_block_p(first_page));

    next_page = first_page;
    remaining_bytes = nwords*N_WORD_BYTES;
    while (remaining_bytes > GENCGC_CARD_BYTES) {
        gc_assert(page_table[next_page].gen == from_space);
        gc_assert(page_allocated_no_region_p(next_page));
        gc_assert(page_table[next_page].large_object);
        gc_assert(page_table[next_page].scan_start_offset ==
                  npage_bytes(next_page-first_page));
        gc_assert(page_table[next_page].bytes_used == GENCGC_CARD_BYTES);

        page_table[next_page].allocated = boxed;

        /* Shouldn't be write-protected at this stage. Essential that the
         * pages aren't. */
        gc_assert(!page_table[next_page].write_protected);
        remaining_bytes -= GENCGC_CARD_BYTES;
        next_page++;
    }

    /* Now only one page remains, but the object may have shrunk so
     * there may be more unused pages which will be freed. */

    /* Object may have shrunk but shouldn't have grown - check. */
    gc_assert(page_table[next_page].bytes_used >= remaining_bytes);

    page_table[next_page].allocated = boxed;
    gc_assert(page_table[next_page].allocated ==
              page_table[first_page].allocated);

    /* Adjust the bytes_used. */
    old_bytes_used = page_table[next_page].bytes_used;
    page_table[next_page].bytes_used = remaining_bytes;

    bytes_freed = old_bytes_used - remaining_bytes;

    /* Free any remaining pages; needs care. */
    next_page++;
    while ((old_bytes_used == GENCGC_CARD_BYTES) &&
           (page_table[next_page].gen == from_space) &&
           page_allocated_no_region_p(next_page) &&
           page_table[next_page].large_object &&
           (page_table[next_page].scan_start_offset ==
            npage_bytes(next_page - first_page))) {
        /* It checks out OK, free the page. We don't need to both zeroing
         * pages as this should have been done before shrinking the
         * object. These pages shouldn't be write protected as they
         * should be zero filled. */
        gc_assert(page_table[next_page].write_protected == 0);

        old_bytes_used = page_table[next_page].bytes_used;
        page_table[next_page].allocated = FREE_PAGE_FLAG;
        page_table[next_page].bytes_used = 0;
        bytes_freed += old_bytes_used;
        next_page++;
    }

    if ((bytes_freed > 0) && gencgc_verbose) {
        FSHOW((stderr,
               "/maybe_adjust_large_object() freed %d\n",
               bytes_freed));
    }

    generations[from_space].bytes_allocated -= bytes_freed;
    bytes_allocated -= bytes_freed;

    return;
}

/*
 * Why is this restricted to protected objects only?
 * Because the rest of the page has been scavenged already,
 * and since that leaves forwarding pointers in the unprotected
 * areas you cannot scavenge it again until those are gone.
 */
static void
scavenge_pinned_range(void* page_base, int start, int count)
{
    // 'start' and 'count' are expressed in units of dwords
    scavenge((lispobj*)page_base + 2*start, 2*count);
}

static void
scavenge_pinned_ranges()
{
    page_index_t page;
    for (page = 0; page < last_free_page; page++) {
        in_use_marker_t* bitmap = pinned_dwords(page);
        if (bitmap)
            bitmap_scan(bitmap,
                        GENCGC_CARD_BYTES / (2*N_WORD_BYTES) / N_WORD_BITS,
                        0, scavenge_pinned_range, page_address(page));
    }
}

static void wipe_range(void* page_base, int start, int count)
{
    bzero((lispobj*)page_base + 2*start, count*2*N_WORD_BYTES);
}

static void
wipe_nonpinned_words()
{
    page_index_t i;
    in_use_marker_t* bitmap;

    for (i = 0; i < last_free_page; i++) {
        if (page_table[i].dont_move && (bitmap = pinned_dwords(i)) != 0) {
            bitmap_scan(bitmap,
                        GENCGC_CARD_BYTES / (2*N_WORD_BYTES) / N_WORD_BITS,
                        BIT_SCAN_INVERT | BIT_SCAN_CLEAR,
                        wipe_range, page_address(i));
            page_table[i].has_pin_map = 0;
            // move the page to newspace
            generations[new_space].bytes_allocated += page_table[i].bytes_used;
            generations[page_table[i].gen].bytes_allocated -= page_table[i].bytes_used;
            page_table[i].gen = new_space;
        }
    }
#ifndef LISP_FEATURE_WIN32
    madvise(page_table_pinned_dwords, pins_map_size_in_bytes, MADV_DONTNEED);
#endif
}

static void __attribute__((unused))
pin_words(page_index_t pageindex, lispobj *mark_which_pointer)
{
    struct page *page = &page_table[pageindex];
    gc_assert(mark_which_pointer);
    if (!page->has_pin_map) {
        page->has_pin_map = 1;
#ifdef DEBUG
        {
          int i;
          in_use_marker_t* map = pinned_dwords(pageindex);
          for (i=0; i<n_dwords_in_card/N_WORD_BITS; ++i)
            gc_assert(map[i] == 0);
        }
#endif
    }
    lispobj *page_base = page_address(pageindex);
    unsigned int begin_dword_index = (mark_which_pointer - page_base) / 2;
    in_use_marker_t *bitmap = pinned_dwords(pageindex);
    if (bitmap[begin_dword_index/N_WORD_BITS]
        & ((uword_t)1 << (begin_dword_index % N_WORD_BITS)))
      return; // already seen this object

    lispobj header = *mark_which_pointer;
    int size = 2;
    // Don't bother calling a sizing function for cons cells.
    if (!is_cons_half(header))
        size = (sizetab[widetag_of(header)])(mark_which_pointer);
    gc_assert(size % 2 == 0);
    unsigned int end_dword_index = begin_dword_index + size / 2;
    unsigned int index;
    for (index = begin_dword_index; index < end_dword_index; index++)
        bitmap[index/N_WORD_BITS] |= (uword_t)1 << (index % N_WORD_BITS);
}

/* Take a possible pointer to a Lisp object and mark its page in the
 * page_table so that it will not be relocated during a GC.
 *
 * This involves locating the page it points to, then backing up to
 * the start of its region, then marking all pages dont_move from there
 * up to the first page that's not full or has a different generation
 *
 * It is assumed that all the page static flags have been cleared at
 * the start of a GC.
 *
 * It is also assumed that the current gc_alloc() region has been
 * flushed and the tables updated. */

// TODO: there's probably a way to be a little more efficient here.
// As things are, we start by finding the object that encloses 'addr',
// then we see if 'addr' was a "valid" Lisp pointer to that object
// - meaning we expect the correct lowtag on the pointer - except
// that for code objects we don't require a correct lowtag
// and we allow a pointer to anywhere in the object.
//
// It should be possible to avoid calling search_dynamic_space
// more of the time. First, check if the page pointed to might hold code.
// If it does, then we continue regardless of the pointer's lowtag
// (because of the special allowance). If the page definitely does *not*
// hold code, then we require up front that the lowtake make sense,
// by doing the same checks that are in properly_tagged_descriptor_p.
//
// Problem: when code is allocated from a per-thread region,
// does it ensure that the occupied pages are flagged as having code?

static void
preserve_pointer(void *addr)
{
#ifdef LISP_FEATURE_IMMOBILE_SPACE
  /* Immobile space MUST be lower than dynamic space,
     or else this test needs to be revised */
    if (addr < (void*)IMMOBILE_SPACE_END) {
        extern void immobile_space_preserve_pointer(void*);
        immobile_space_preserve_pointer(addr);
        return;
    }
#endif
    page_index_t addr_page_index = find_page_index(addr);
    lispobj *object_start;

    if (addr_page_index == -1
        || (object_start = conservative_root_p(addr, addr_page_index)) == 0)
        return;

    /* (Now that we know that addr_page_index is in range, it's
     * safe to index into page_table[] with it.) */
    unsigned int region_allocation = page_table[addr_page_index].allocated;

    /* Find the beginning of the region.  Note that there may be
     * objects in the region preceding the one that we were passed a
     * pointer to: if this is the case, we will write-protect all the
     * previous objects' pages too.     */

#if 0
    /* I think this'd work just as well, but without the assertions.
     * -dan 2004.01.01 */
    page_index_t first_page = find_page_index(page_scan_start(addr_page_index))
#else
    page_index_t first_page = addr_page_index;
    while (!page_starts_contiguous_block_p(first_page)) {
        --first_page;
        /* Do some checks. */
        gc_assert(page_table[first_page].bytes_used == GENCGC_CARD_BYTES);
        gc_assert(page_table[first_page].gen == from_space);
        gc_assert(page_table[first_page].allocated == region_allocation);
    }
#endif

    /* Adjust any large objects before promotion as they won't be
     * copied after promotion. */
    if (page_table[first_page].large_object) {
        maybe_adjust_large_object(page_address(first_page));
        /* It may have moved to unboxed pages. */
        region_allocation = page_table[first_page].allocated;
    }

    /* Now work forward until the end of this contiguous area is found,
     * marking all pages as dont_move. */
    page_index_t i;
    for (i = first_page; ;i++) {
        gc_assert(page_table[i].allocated == region_allocation);

        /* Mark the page static. */
        page_table[i].dont_move = 1;

        /* It is essential that the pages are not write protected as
         * they may have pointers into the old-space which need
         * scavenging. They shouldn't be write protected at this
         * stage. */
        gc_assert(!page_table[i].write_protected);

        /* Check whether this is the last page in this contiguous block.. */
        if (page_ends_contiguous_block_p(i, from_space))
            break;
    }

#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
    /* Do not do this for multi-page objects.  Those pages do not need
     * object wipeout anyway.
     */
    if (do_wipe_p && i == first_page) // single-page object
        pin_words(first_page, object_start);
#endif

    /* Check that the page is now static. */
    gc_assert(page_table[addr_page_index].dont_move != 0);
}
\f
/* If the given page is not write-protected, then scan it for pointers
 * to younger generations or the top temp. generation, if no
 * suspicious pointers are found then the page is write-protected.
 *
 * Care is taken to check for pointers to the current gc_alloc()
 * region if it is a younger generation or the temp. generation. This
 * frees the caller from doing a gc_alloc_update_page_tables(). Actually
 * the gc_alloc_generation does not need to be checked as this is only
 * called from scavenge_generation() when the gc_alloc generation is
 * younger, so it just checks if there is a pointer to the current
 * region.
 *
 * We return 1 if the page was write-protected, else 0. */
static int
update_page_write_prot(page_index_t page)
{
    generation_index_t gen = page_table[page].gen;
    sword_t j;
    int wp_it = 1;
    void **page_addr = (void **)page_address(page);
    sword_t num_words = page_table[page].bytes_used / N_WORD_BYTES;

    /* Shouldn't be a free page. */
    gc_assert(page_allocated_p(page));
    gc_assert(page_table[page].bytes_used != 0);

    /* Skip if it's already write-protected, pinned, or unboxed */
    if (page_table[page].write_protected
        /* FIXME: What's the reason for not write-protecting pinned pages? */
        || page_table[page].dont_move
        || page_unboxed_p(page))
        return (0);

    /* Scan the page for pointers to younger generations or the
     * top temp. generation. */

    /* This is conservative: any word satisfying is_lisp_pointer() is
     * assumed to be a pointer. To do otherwise would require a family
     * of scavenge-like functions. */
    for (j = 0; j < num_words; j++) {
        void *ptr = *(page_addr+j);
        page_index_t index;
        lispobj __attribute__((unused)) header;

        if (!is_lisp_pointer((lispobj)ptr))
            continue;
        /* Check that it's in the dynamic space */
        if ((index = find_page_index(ptr)) != -1) {
            if (/* Does it point to a younger or the temp. generation? */
                (page_allocated_p(index)
                 && (page_table[index].bytes_used != 0)
                 && ((page_table[index].gen < gen)
                     || (page_table[index].gen == SCRATCH_GENERATION)))

                /* Or does it point within a current gc_alloc() region? */
                || ((boxed_region.start_addr <= ptr)
                    && (ptr <= boxed_region.free_pointer))
                || ((unboxed_region.start_addr <= ptr)
                    && (ptr <= unboxed_region.free_pointer))) {
                wp_it = 0;
                break;
            }
        }
#ifdef LISP_FEATURE_IMMOBILE_SPACE
        else if ((index = find_immobile_page_index(ptr)) >= 0 &&
                 other_immediate_lowtag_p(header = *native_pointer((lispobj)ptr))) {
            // This is *possibly* a pointer to an object in immobile space,
            // given that above two conditions were satisfied.
            // But unlike in the dynamic space case, we need to read a byte
            // from the object to determine its generation, which requires care.
            // Consider an unboxed word that looks like a pointer to a word that
            // looks like fun-header-widetag. We can't naively back up to the
            // underlying code object since the alleged header might not be one.
            int obj_gen = gen; // Make comparison fail if we fall through
            if (lowtag_of((lispobj)ptr) != FUN_POINTER_LOWTAG) {
                obj_gen = __immobile_obj_generation(native_pointer((lispobj)ptr));
            } else if (widetag_of(header) == SIMPLE_FUN_HEADER_WIDETAG) {
                struct code* code =
                  code_obj_from_simple_fun((struct simple_fun *)
                                           ((lispobj)ptr - FUN_POINTER_LOWTAG));
                // This is a heuristic, since we're not actually looking for
                // an object boundary. Precise scanning of 'page' would obviate
                // the guard conditions here.
                if ((lispobj)code >= IMMOBILE_VARYOBJ_SUBSPACE_START
                    && widetag_of(code->header) == CODE_HEADER_WIDETAG)
                    obj_gen = __immobile_obj_generation((lispobj*)code);
            }
            // A bogus generation number implies a not-really-pointer,
            // but it won't cause misbehavior.
            if (obj_gen < gen || obj_gen == SCRATCH_GENERATION) {
                wp_it = 0;
                break;
           }
        }
#endif
    }

    if (wp_it == 1) {
        /* Write-protect the page. */
        /*FSHOW((stderr, "/write-protecting page %d gen %d\n", page, gen));*/

        os_protect((void *)page_addr,
                   GENCGC_CARD_BYTES,
                   OS_VM_PROT_READ|OS_VM_PROT_EXECUTE);

        /* Note the page as protected in the page tables. */
        page_table[page].write_protected = 1;
    }

    return (wp_it);
}

/* Scavenge all generations from FROM to TO, inclusive, except for
 * new_space which needs special handling, as new objects may be
 * added which are not checked here - use scavenge_newspace generation.
 *
 * Write-protected pages should not have any pointers to the
 * from_space so do need scavenging; thus write-protected pages are
 * not always scavenged. There is some code to check that these pages
 * are not written; but to check fully the write-protected pages need
 * to be scavenged by disabling the code to skip them.
 *
 * Under the current scheme when a generation is GCed the younger
 * generations will be empty. So, when a generation is being GCed it
 * is only necessary to scavenge the older generations for pointers
 * not the younger. So a page that does not have pointers to younger
 * generations does not need to be scavenged.
 *
 * The write-protection can be used to note pages that don't have
 * pointers to younger pages. But pages can be written without having
 * pointers to younger generations. After the pages are scavenged here
 * they can be scanned for pointers to younger generations and if
 * there are none the page can be write-protected.
 *
 * One complication is when the newspace is the top temp. generation.
 *
 * Enabling SC_GEN_CK scavenges the write-protected pages and checks
 * that none were written, which they shouldn't be as they should have
 * no pointers to younger generations. This breaks down for weak
 * pointers as the objects contain a link to the next and are written
 * if a weak pointer is scavenged. Still it's a useful check. */
static void
scavenge_generations(generation_index_t from, generation_index_t to)
{
    page_index_t i;
    page_index_t num_wp = 0;

#define SC_GEN_CK 0
#if SC_GEN_CK
    /* Clear the write_protected_cleared flags on all pages. */
    for (i = 0; i < page_table_pages; i++)
        page_table[i].write_protected_cleared = 0;
#endif

    for (i = 0; i < last_free_page; i++) {
        generation_index_t generation = page_table[i].gen;
        if (page_boxed_p(i)
            && (page_table[i].bytes_used != 0)
            && (generation != new_space)
            && (generation >= from)
            && (generation <= to)) {
            page_index_t last_page,j;
            int write_protected=1;

            /* This should be the start of a region */
            gc_assert(page_starts_contiguous_block_p(i));

            /* Now work forward until the end of the region */
            for (last_page = i; ; last_page++) {
                write_protected =
                    write_protected && page_table[last_page].write_protected;
                if (page_ends_contiguous_block_p(last_page, generation))
                    break;
            }
            if (!write_protected) {
                scavenge(page_address(i),
                         ((uword_t)(page_table[last_page].bytes_used
                                          + npage_bytes(last_page-i)))
                         /N_WORD_BYTES);

                /* Now scan the pages and write protect those that
                 * don't have pointers to younger generations. */
                if (enable_page_protection) {
                    for (j = i; j <= last_page; j++) {
                        num_wp += update_page_write_prot(j);
                    }
                }
                if ((gencgc_verbose > 1) && (num_wp != 0)) {
                    FSHOW((stderr,
                           "/write protected %d pages within generation %d\n",
                           num_wp, generation));
                }
            }
            i = last_page;
        }
    }

#if SC_GEN_CK
    /* Check that none of the write_protected pages in this generation
     * have been written to. */
    for (i = 0; i < page_table_pages; i++) {
        if (page_allocated_p(i)
            && (page_table[i].bytes_used != 0)
            && (page_table[i].gen == generation)
            && (page_table[i].write_protected_cleared != 0)) {
            FSHOW((stderr, "/scavenge_generation() %d\n", generation));
            FSHOW((stderr,
                   "/page bytes_used=%d scan_start_offset=%lu dont_move=%d\n",
                    page_table[i].bytes_used,
                    page_table[i].scan_start_offset,
                    page_table[i].dont_move));
            lose("write to protected page %d in scavenge_generation()\n", i);
        }
    }
#endif
}

\f
/* Scavenge a newspace generation. As it is scavenged new objects may
 * be allocated to it; these will also need to be scavenged. This
 * repeats until there are no more objects unscavenged in the
 * newspace generation.
 *
 * To help improve the efficiency, areas written are recorded by
 * gc_alloc() and only these scavenged. Sometimes a little more will be
 * scavenged, but this causes no harm. An easy check is done that the
 * scavenged bytes equals the number allocated in the previous
 * scavenge.
 *
 * Write-protected pages are not scanned except if they are marked
 * dont_move in which case they may have been promoted and still have
 * pointers to the from space.
 *
 * Write-protected pages could potentially be written by alloc however
 * to avoid having to handle re-scavenging of write-protected pages
 * gc_alloc() does not write to write-protected pages.
 *
 * New areas of objects allocated are recorded alternatively in the two
 * new_areas arrays below. */
static struct new_area new_areas_1[NUM_NEW_AREAS];
static struct new_area new_areas_2[NUM_NEW_AREAS];

#ifdef LISP_FEATURE_IMMOBILE_SPACE
extern unsigned int immobile_scav_queue_count;
extern void
  gc_init_immobile(),
  update_immobile_nursery_bits(),
  scavenge_immobile_roots(generation_index_t,generation_index_t),
  scavenge_immobile_newspace(),
  sweep_immobile_space(int raise),
  write_protect_immobile_space();
#else
#define immobile_scav_queue_count 0
#endif

/* Do one full scan of the new space generation. This is not enough to
 * complete the job as new objects may be added to the generation in
 * the process which are not scavenged. */
static void
scavenge_newspace_generation_one_scan(generation_index_t generation)
{
    page_index_t i;

    FSHOW((stderr,
           "/starting one full scan of newspace generation %d\n",
           generation));
    for (i = 0; i < last_free_page; i++) {
        /* Note that this skips over open regions when it encounters them. */
        if (page_boxed_p(i)
            && (page_table[i].bytes_used != 0)
            && (page_table[i].gen == generation)
            && ((page_table[i].write_protected == 0)
                /* (This may be redundant as write_protected is now
                 * cleared before promotion.) */
                || (page_table[i].dont_move == 1))) {
            page_index_t last_page;
            int all_wp=1;

            /* The scavenge will start at the scan_start_offset of
             * page i.
             *
             * We need to find the full extent of this contiguous
             * block in case objects span pages.
             *
             * Now work forward until the end of this contiguous area
             * is found. A small area is preferred as there is a
             * better chance of its pages being write-protected. */
            for (last_page = i; ;last_page++) {
                /* If all pages are write-protected and movable,
                 * then no need to scavenge */
                all_wp=all_wp && page_table[last_page].write_protected &&
                    !page_table[last_page].dont_move;

                /* Check whether this is the last page in this
                 * contiguous block */
                if (page_ends_contiguous_block_p(last_page, generation))
                    break;
            }

            /* Do a limited check for write-protected pages.  */
            if (!all_wp) {
                sword_t nwords = (((uword_t)
                               (page_table[last_page].bytes_used
                                + npage_bytes(last_page-i)
                                + page_table[i].scan_start_offset))
                               / N_WORD_BYTES);
                new_areas_ignore_page = last_page;

                scavenge(page_scan_start(i), nwords);

            }
            i = last_page;
        }
    }
    FSHOW((stderr,
           "/done with one full scan of newspace generation %d\n",
           generation));
}

/* Do a complete scavenge of the newspace generation. */
static void
scavenge_newspace_generation(generation_index_t generation)
{
    size_t i;

    /* the new_areas array currently being written to by gc_alloc() */
    struct new_area (*current_new_areas)[] = &new_areas_1;
    size_t current_new_areas_index;

    /* the new_areas created by the previous scavenge cycle */
    struct new_area (*previous_new_areas)[] = NULL;
    size_t previous_new_areas_index;

    /* Flush the current regions updating the tables. */
    gc_alloc_update_all_page_tables(0);

    /* Turn on the recording of new areas by gc_alloc(). */
    new_areas = current_new_areas;
    new_areas_index = 0;

    /* Don't need to record new areas that get scavenged anyway during
     * scavenge_newspace_generation_one_scan. */
    record_new_objects = 1;

    /* Start with a full scavenge. */
    scavenge_newspace_generation_one_scan(generation);

    /* Record all new areas now. */
    record_new_objects = 2;

    /* Give a chance to weak hash tables to make other objects live.
     * FIXME: The algorithm implemented here for weak hash table gcing
     * is O(W^2+N) as Bruno Haible warns in
     * http://www.haible.de/bruno/papers/cs/weak/WeakDatastructures-writeup.html
     * see "Implementation 2". */
    scav_weak_hash_tables();

    /* Flush the current regions updating the tables. */
    gc_alloc_update_all_page_tables(0);

    /* Grab new_areas_index. */
    current_new_areas_index = new_areas_index;

    /*FSHOW((stderr,
             "The first scan is finished; current_new_areas_index=%d.\n",
             current_new_areas_index));*/

    while (current_new_areas_index > 0 || immobile_scav_queue_count) {
        /* Move the current to the previous new areas */
        previous_new_areas = current_new_areas;
        previous_new_areas_index = current_new_areas_index;

        /* Scavenge all the areas in previous new areas. Any new areas
         * allocated are saved in current_new_areas. */

        /* Allocate an array for current_new_areas; alternating between
         * new_areas_1 and 2 */
        if (previous_new_areas == &new_areas_1)
            current_new_areas = &new_areas_2;
        else
            current_new_areas = &new_areas_1;

        /* Set up for gc_alloc(). */
        new_areas = current_new_areas;
        new_areas_index = 0;

#ifdef LISP_FEATURE_IMMOBILE_SPACE
        scavenge_immobile_newspace();
#endif
        /* Check whether previous_new_areas had overflowed. */
        if (previous_new_areas_index >= NUM_NEW_AREAS) {

            /* New areas of objects allocated have been lost so need to do a
             * full scan to be sure! If this becomes a problem try
             * increasing NUM_NEW_AREAS. */
            if (gencgc_verbose) {
                SHOW("new_areas overflow, doing full scavenge");
            }

            /* Don't need to record new areas that get scavenged
             * anyway during scavenge_newspace_generation_one_scan. */
            record_new_objects = 1;

            scavenge_newspace_generation_one_scan(generation);

            /* Record all new areas now. */
            record_new_objects = 2;

            scav_weak_hash_tables();

            /* Flush the current regions updating the tables. */
            gc_alloc_update_all_page_tables(0);

        } else {

            /* Work through previous_new_areas. */
            for (i = 0; i < previous_new_areas_index; i++) {
                page_index_t page = (*previous_new_areas)[i].page;
                size_t offset = (*previous_new_areas)[i].offset;
                size_t size = (*previous_new_areas)[i].size / N_WORD_BYTES;
                gc_assert((*previous_new_areas)[i].size % N_WORD_BYTES == 0);
                scavenge(page_address(page)+offset, size);
            }

            scav_weak_hash_tables();

            /* Flush the current regions updating the tables. */
            gc_alloc_update_all_page_tables(0);
        }

        current_new_areas_index = new_areas_index;

        /*FSHOW((stderr,
                 "The re-scan has finished; current_new_areas_index=%d.\n",
                 current_new_areas_index));*/
    }

    /* Turn off recording of areas allocated by gc_alloc(). */
    record_new_objects = 0;

#if SC_NS_GEN_CK
    {
        page_index_t i;
        /* Check that none of the write_protected pages in this generation
         * have been written to. */
        for (i = 0; i < page_table_pages; i++) {
            if (page_allocated_p(i)
                && (page_table[i].bytes_used != 0)
                && (page_table[i].gen == generation)
                && (page_table[i].write_protected_cleared != 0)
                && (page_table[i].dont_move == 0)) {
                lose("write protected page %d written to in scavenge_newspace_generation\ngeneration=%d dont_move=%d\n",
                     i, generation, page_table[i].dont_move);
            }
        }
    }
#endif
}
\f
/* Un-write-protect all the pages in from_space. This is done at the
 * start of a GC else there may be many page faults while scavenging
 * the newspace (I've seen drive the system time to 99%). These pages
 * would need to be unprotected anyway before unmapping in
 * free_oldspace; not sure what effect this has on paging.. */
static void
unprotect_oldspace(void)
{
    page_index_t i;
    void *region_addr = 0;
    void *page_addr = 0;
    uword_t region_bytes = 0;

    for (i = 0; i < last_free_page; i++) {
        if (page_allocated_p(i)
            && (page_table[i].bytes_used != 0)
            && (page_table[i].gen == from_space)) {

            /* Remove any write-protection. We should be able to rely
             * on the write-protect flag to avoid redundant calls. */
            if (page_table[i].write_protected) {
                page_table[i].write_protected = 0;
                page_addr = page_address(i);
                if (!region_addr) {
                    /* First region. */
                    region_addr = page_addr;
                    region_bytes = GENCGC_CARD_BYTES;
                } else if (region_addr + region_bytes == page_addr) {
                    /* Region continue. */
                    region_bytes += GENCGC_CARD_BYTES;
                } else {
                    /* Unprotect previous region. */
                    os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
                    /* First page in new region. */
                    region_addr = page_addr;
                    region_bytes = GENCGC_CARD_BYTES;
                }
            }
        }
    }
    if (region_addr) {
        /* Unprotect last region. */
        os_protect(region_addr, region_bytes, OS_VM_PROT_ALL);
    }
}

/* Work through all the pages and free any in from_space. This
 * assumes that all objects have been copied or promoted to an older
 * generation. Bytes_allocated and the generation bytes_allocated
 * counter are updated. The number of bytes freed is returned. */
static uword_t
free_oldspace(void)
{
    uword_t bytes_freed = 0;
    page_index_t first_page, last_page;

    first_page = 0;

    do {
        /* Find a first page for the next region of pages. */
        while ((first_page < last_free_page)
               && (page_free_p(first_page)
                   || (page_table[first_page].bytes_used == 0)
                   || (page_table[first_page].gen != from_space)))
            first_page++;

        if (first_page >= last_free_page)
            break;

        /* Find the last page of this region. */
        last_page = first_page;

        do {
            /* Free the page. */
            bytes_freed += page_table[last_page].bytes_used;
            generations[page_table[last_page].gen].bytes_allocated -=
                page_table[last_page].bytes_used;
            page_table[last_page].allocated = FREE_PAGE_FLAG;
            page_table[last_page].bytes_used = 0;
            /* Should already be unprotected by unprotect_oldspace(). */
            gc_assert(!page_table[last_page].write_protected);
            last_page++;
        }
        while ((last_page < last_free_page)
               && page_allocated_p(last_page)
               && (page_table[last_page].bytes_used != 0)
               && (page_table[last_page].gen == from_space));

#ifdef READ_PROTECT_FREE_PAGES
        os_protect(page_address(first_page),
                   npage_bytes(last_page-first_page),
                   OS_VM_PROT_NONE);
#endif
        first_page = last_page;
    } while (first_page < last_free_page);

    bytes_allocated -= bytes_freed;
    return bytes_freed;
}
\f
#if 0
/* Print some information about a pointer at the given address. */
static void
print_ptr(lispobj *addr)
{
    /* If addr is in the dynamic space then out the page information. */
    page_index_t pi1 = find_page_index((void*)addr);

    if (pi1 != -1)
        fprintf(stderr,"  %p: page %d  alloc %d  gen %d  bytes_used %d  offset %lu  dont_move %d\n",
                addr,
                pi1,
                page_table[pi1].allocated,
                page_table[pi1].gen,
                page_table[pi1].bytes_used,
                page_table[pi1].scan_start_offset,
                page_table[pi1].dont_move);
    fprintf(stderr,"  %x %x %x %x (%x) %x %x %x %x\n",
            *(addr-4),
            *(addr-3),
            *(addr-2),
            *(addr-1),
            *(addr-0),
            *(addr+1),
            *(addr+2),
            *(addr+3),
            *(addr+4));
}
#endif

static int
is_in_stack_space(lispobj ptr)
{
    /* For space verification: Pointers can be valid if they point
     * to a thread stack space.  This would be faster if the thread
     * structures had page-table entries as if they were part of
     * the heap space. */
    struct thread *th;
    for_each_thread(th) {
        if ((th->control_stack_start <= (lispobj *)ptr) &&
            (th->control_stack_end >= (lispobj *)ptr)) {
            return 1;
        }
    }
    return 0;
}

// NOTE: This function can produces false failure indications,
// usually related to dynamic space pointing to the stack of a
// dead thread, but there may be other reasons as well.
static void
verify_space(lispobj *start, size_t words)
{
    extern int valid_lisp_pointer_p(lispobj);
    int is_in_dynamic_space = (find_page_index((void*)start) != -1);
    int is_in_readonly_space =
        (READ_ONLY_SPACE_START <= (uword_t)start &&
         (uword_t)start < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
#ifdef LISP_FEATURE_IMMOBILE_SPACE
    int is_in_immobile_space =
        (IMMOBILE_SPACE_START <= (uword_t)start &&
         (uword_t)start < SymbolValue(IMMOBILE_SPACE_FREE_POINTER,0));
#endif

    while (words > 0) {
        size_t count = 1;
        lispobj thing = *start;
        lispobj __attribute__((unused)) pointee;

        if (is_lisp_pointer(thing)) {
            page_index_t page_index = find_page_index((void*)thing);
            sword_t to_readonly_space =
                (READ_ONLY_SPACE_START <= thing &&
                 thing < SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0));
            sword_t to_static_space =
                (STATIC_SPACE_START <= thing &&
                 thing < SymbolValue(STATIC_SPACE_FREE_POINTER,0));
#ifdef LISP_FEATURE_IMMOBILE_SPACE
            sword_t to_immobile_space =
                (IMMOBILE_SPACE_START <= thing &&
                 thing < SymbolValue(IMMOBILE_FIXEDOBJ_FREE_POINTER,0)) ||
                (IMMOBILE_VARYOBJ_SUBSPACE_START <= thing &&
                 thing < SymbolValue(IMMOBILE_SPACE_FREE_POINTER,0));
#endif

            /* Does it point to the dynamic space? */
            if (page_index != -1) {
                /* If it's within the dynamic space it should point to a used page. */
                if (!page_allocated_p(page_index))
                    lose ("Ptr %p @ %p sees free page.\n", thing, start);
                if ((char*)thing - (char*)page_address(page_index)
                    >= page_table[page_index].bytes_used)
                    lose ("Ptr %p @ %p sees unallocated space.\n", thing, start);
                /* Check that it doesn't point to a forwarding pointer! */
                if (*native_pointer(thing) == 0x01) {
                    lose("Ptr %p @ %p sees forwarding ptr.\n", thing, start);
                }
                /* Check that its not in the RO space as it would then be a
                 * pointer from the RO to the dynamic space. */
                if (is_in_readonly_space) {
                    lose("ptr to dynamic space %p from RO space %x\n",
                         thing, start);
                }
#ifdef LISP_FEATURE_IMMOBILE_SPACE
                // verify all immobile space -> dynamic space pointers
                if (is_in_immobile_space && !valid_lisp_pointer_p(thing)) {
                    lose("Ptr %p @ %p sees junk.\n", thing, start);
                }
#endif
                /* Does it point to a plausible object? This check slows
                 * it down a lot (so it's commented out).
                 *
                 * "a lot" is serious: it ate 50 minutes cpu time on
                 * my duron 950 before I came back from lunch and
                 * killed it.
                 *
                 *   FIXME: Add a variable to enable this
                 * dynamically. */
                /*
                if (!valid_lisp_pointer_p((lispobj *)thing) {
                    lose("ptr %p to invalid object %p\n", thing, start);
                }
                */
#ifdef LISP_FEATURE_IMMOBILE_SPACE
            } else if (to_immobile_space) {
                // the object pointed to must not have been discarded as garbage
                if (!other_immediate_lowtag_p(*native_pointer(thing))
                    || immobile_filler_p(native_pointer(thing)))
                    lose("Ptr %p @ %p sees trashed object.\n", (void*)thing, start);
                // verify all pointers to immobile space
                if (!valid_lisp_pointer_p(thing))
                    lose("Ptr %p @ %p sees junk.\n", thing, start);
#endif
            } else {
                extern char __attribute__((unused)) funcallable_instance_tramp;
                /* Verify that it points to another valid space. */
                if (!to_readonly_space && !to_static_space
                    && !is_in_stack_space(thing)) {
                    lose("Ptr %p @ %p sees junk.\n", thing, start);
                }
            }
        } else {
            if (!(fixnump(thing))) {
                /* skip fixnums */
                switch(widetag_of(*start)) {

                    /* boxed objects */
                case SIMPLE_VECTOR_WIDETAG:
                case RATIO_WIDETAG:
                case COMPLEX_WIDETAG:
                case SIMPLE_ARRAY_WIDETAG:
                case COMPLEX_BASE_STRING_WIDETAG:
#ifdef COMPLEX_CHARACTER_STRING_WIDETAG
                case COMPLEX_CHARACTER_STRING_WIDETAG:
#endif
                case COMPLEX_VECTOR_NIL_WIDETAG:
                case COMPLEX_BIT_VECTOR_WIDETAG:
                case COMPLEX_VECTOR_WIDETAG:
                case COMPLEX_ARRAY_WIDETAG:
                case CLOSURE_HEADER_WIDETAG:
                case FUNCALLABLE_INSTANCE_HEADER_WIDETAG:
                case VALUE_CELL_HEADER_WIDETAG:
                case SYMBOL_HEADER_WIDETAG:
                case CHARACTER_WIDETAG:
#if N_WORD_BITS == 64
                case SINGLE_FLOAT_WIDETAG:
#endif
                case UNBOUND_MARKER_WIDETAG:
                    break;
                case FDEFN_WIDETAG:
#ifdef LISP_FEATURE_IMMOBILE_CODE
                    verify_space(start + 1, 2);
                    pointee = fdefn_raw_referent((struct fdefn*)start);
                    verify_space(&pointee, 1);
                    count = 4;
#endif
                    break;

                case INSTANCE_HEADER_WIDETAG:
                    if (instance_layout(start)) {
                        lispobj bitmap =
                            ((struct layout*)
                             native_pointer(instance_layout(start)))->bitmap;
                        sword_t nslots = instance_length(thing) | 1;
                        instance_scan(verify_space, start+1, nslots, bitmap);
                        count = 1 + nslots;
                    }
                    break;
                case CODE_HEADER_WIDETAG:
                    {
                        /* Check that it's not in the dynamic space.
                         * FIXME: Isn't is supposed to be OK for code
                         * objects to be in the dynamic space these days? */
                        /* It is for byte compiled code, but there's
                         * no byte compilation in SBCL anymore. */
                        if (is_in_dynamic_space
                            /* Only when enabled */
                            && verify_dynamic_code_check) {
                            FSHOW((stderr,
                                   "/code object at %p in the dynamic space\n",
                                   start));
                        }

                        struct code *code = (struct code *) start;
                        sword_t nheader_words = code_header_words(code->header);
                        /* Scavenge the boxed section of the code data block */
                        verify_space(start + 1, nheader_words - 1);

                        /* Scavenge the boxed section of each function
                         * object in the code data block. */
                        for_each_simple_fun(i, fheaderp, code, 1, {
                            verify_space(SIMPLE_FUN_SCAV_START(fheaderp),
                                         SIMPLE_FUN_SCAV_NWORDS(fheaderp)); });
                        count = nheader_words + code_instruction_words(code->code_size);
                        break;
                    }

                    /* unboxed objects */
                case BIGNUM_WIDETAG:
#if N_WORD_BITS != 64
                case SINGLE_FLOAT_WIDETAG:
#endif
                case DOUBLE_FLOAT_WIDETAG:
#ifdef COMPLEX_LONG_FLOAT_WIDETAG
                case LONG_FLOAT_WIDETAG:
#endif
#ifdef COMPLEX_SINGLE_FLOAT_WIDETAG
                case COMPLEX_SINGLE_FLOAT_WIDETAG:
#endif
#ifdef COMPLEX_DOUBLE_FLOAT_WIDETAG
                case COMPLEX_DOUBLE_FLOAT_WIDETAG:
#endif
#ifdef COMPLEX_LONG_FLOAT_WIDETAG
                case COMPLEX_LONG_FLOAT_WIDETAG:
#endif
#ifdef SIMD_PACK_WIDETAG
                case SIMD_PACK_WIDETAG:
#endif
#include "genesis/specialized-vectors.inc"
                case SAP_WIDETAG:
                case WEAK_POINTER_WIDETAG:
#ifdef NO_TLS_VALUE_MARKER_WIDETAG
                case NO_TLS_VALUE_MARKER_WIDETAG:
#endif
                    count = (sizetab[widetag_of(*start)])(start);
                    break;

                default:
                    lose("Unhandled widetag %p at %p\n",
                         widetag_of(*start), start);
                }
            }
        }
        start += count;
        words -= count;
    }
}

static void verify_dynamic_space();

static void
verify_gc(void)
{
    /* FIXME: It would be nice to make names consistent so that
     * foo_size meant size *in* *bytes* instead of size in some
     * arbitrary units. (Yes, this caused a bug, how did you guess?:-)
     * Some counts of lispobjs are called foo_count; it might be good
     * to grep for all foo_size and rename the appropriate ones to
     * foo_count. */
#ifdef LISP_FEATURE_IMMOBILE_SPACE
#  ifdef __linux__
    // Try this verification if marknsweep was compiled with extra debugging.
    // But weak symbols don't work on macOS.
    extern void __attribute__((weak)) check_varyobj_pages();
    if (&check_varyobj_pages) check_varyobj_pages();
#  endif
    verify_space((lispobj*)IMMOBILE_SPACE_START,
                 (lispobj*)SymbolValue(IMMOBILE_FIXEDOBJ_FREE_POINTER,0)
                 - (lispobj*)IMMOBILE_SPACE_START);
    verify_space((lispobj*)IMMOBILE_VARYOBJ_SUBSPACE_START,
                 (lispobj*)SymbolValue(IMMOBILE_SPACE_FREE_POINTER,0)
                 - (lispobj*)IMMOBILE_VARYOBJ_SUBSPACE_START);
#endif
    sword_t read_only_space_size =
        (lispobj*)SymbolValue(READ_ONLY_SPACE_FREE_POINTER,0)
        - (lispobj*)READ_ONLY_SPACE_START;
    sword_t static_space_size =
        (lispobj*)SymbolValue(STATIC_SPACE_FREE_POINTER,0)
        - (lispobj*)STATIC_SPACE_START;
    struct thread *th;
    for_each_thread(th) {
    sword_t binding_stack_size =
        (lispobj*)get_binding_stack_pointer(th)
            - (lispobj*)th->binding_stack_start;
        verify_space(th->binding_stack_start, binding_stack_size);
    }
    verify_space((lispobj*)READ_ONLY_SPACE_START, read_only_space_size);
    verify_space((lispobj*)STATIC_SPACE_START   , static_space_size);
    verify_dynamic_space();
}

void
walk_generation(void (*proc)(lispobj*,size_t),
                generation_index_t generation)
{
    page_index_t i;
    int genmask = generation >= 0 ? 1 << generation : ~0;

    for (i = 0; i < last_free_page; i++) {
        if (page_allocated_p(i)
            && (page_table[i].bytes_used != 0)
            && ((1 << page_table[i].gen) & genmask)) {
            page_index_t last_page;

            /* This should be the start of a contiguous block */
            gc_assert(page_starts_contiguous_block_p(i));

            /* Need to find the full extent of this contiguous block in case
               objects span pages. */

            /* Now work forward until the end of this contiguous area is
               found. */
            for (last_page = i; ;last_page++)
                /* Check whether this is the last page in this contiguous
                 * block. */
                if (page_ends_contiguous_block_p(last_page, page_table[i].gen))
                    break;

            proc(page_address(i),
                 ((uword_t)(page_table[last_page].bytes_used
                           + npage_bytes(last_page-i)))
                         / N_WORD_BYTES);
            i = last_page;
        }
    }
}
static void verify_generation(generation_index_t generation)
{
  walk_generation(verify_space, generation);
}

/* Check that all the free space is zero filled. */
static void
verify_zero_fill(void)
{
    page_index_t page;

    for (page = 0; page < last_free_page; page++) {
        if (page_free_p(page)) {
            /* The whole page should be zero filled. */
            sword_t *start_addr = (sword_t *)page_address(page);
            sword_t i;
            for (i = 0; i < (sword_t)GENCGC_CARD_BYTES/N_WORD_BYTES; i++) {
                if (start_addr[i] != 0) {
                    lose("free page not zero at %x\n", start_addr + i);
                }
            }
        } else {
            sword_t free_bytes = GENCGC_CARD_BYTES - page_table[page].bytes_used;
            if (free_bytes > 0) {
                sword_t *start_addr = (sword_t *)((uword_t)page_address(page)
                                          + page_table[page].bytes_used);
                sword_t size = free_bytes / N_WORD_BYTES;
                sword_t i;
                for (i = 0; i < size; i++) {
                    if (start_addr[i] != 0) {
                        lose("free region not zero at %x\n", start_addr + i);
                    }
                }
            }
        }
    }
}

/* External entry point for verify_zero_fill */
void
gencgc_verify_zero_fill(void)
{
    /* Flush the alloc regions updating the tables. */
    gc_alloc_update_all_page_tables(1);
    SHOW("verifying zero fill");
    verify_zero_fill();
}

static void
verify_dynamic_space(void)
{
    verify_generation(-1);
    if (gencgc_enable_verify_zero_fill)
        verify_zero_fill();
}
\f
/* Write-protect all the dynamic boxed pages in the given generation. */
static void
write_protect_generation_pages(generation_index_t generation)
{
    page_index_t start;

    gc_assert(generation < SCRATCH_GENERATION);

    for (start = 0; start < last_free_page; start++) {
        if (protect_page_p(start, generation)) {
            void *page_start;
            page_index_t last;

            /* Note the page as protected in the page tables. */
            page_table[start].write_protected = 1;

            for (last = start + 1; last < last_free_page; last++) {
                if (!protect_page_p(last, generation))
                  break;
                page_table[last].write_protected = 1;
            }

            page_start = (void *)page_address(start);

            os_protect(page_start,
                       npage_bytes(last - start),
                       OS_VM_PROT_READ | OS_VM_PROT_EXECUTE);

            start = last;
        }
    }

    if (gencgc_verbose > 1) {
        FSHOW((stderr,
               "/write protected %d of %d pages in generation %d\n",
               count_write_protect_generation_pages(generation),
               count_generation_pages(generation),
               generation));
    }
}

#if defined(LISP_FEATURE_SB_THREAD) && (defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64))
static void
preserve_context_registers (os_context_t *c)
{
    void **ptr;
    /* On Darwin the signal context isn't a contiguous block of memory,
     * so just preserve_pointering its contents won't be sufficient.
     */
#if defined(LISP_FEATURE_DARWIN)||defined(LISP_FEATURE_WIN32)
#if defined LISP_FEATURE_X86
    preserve_pointer((void*)*os_context_register_addr(c,reg_EAX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_ECX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_EDX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_EBX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_ESI));
    preserve_pointer((void*)*os_context_register_addr(c,reg_EDI));
    preserve_pointer((void*)*os_context_pc_addr(c));
#elif defined LISP_FEATURE_X86_64
    preserve_pointer((void*)*os_context_register_addr(c,reg_RAX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_RCX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_RDX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_RBX));
    preserve_pointer((void*)*os_context_register_addr(c,reg_RSI));
    preserve_pointer((void*)*os_context_register_addr(c,reg_RDI));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R8));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R9));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R10));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R11));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R12));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R13));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R14));
    preserve_pointer((void*)*os_context_register_addr(c,reg_R15));
    preserve_pointer((void*)*os_context_pc_addr(c));
#else
    #error "preserve_context_registers needs to be tweaked for non-x86 Darwin"
#endif
#endif
#if !defined(LISP_FEATURE_WIN32)
    for(ptr = ((void **)(c+1))-1; ptr>=(void **)c; ptr--) {
        preserve_pointer(*ptr);
    }
#endif
}
#endif

static void
move_pinned_pages_to_newspace()
{
    page_index_t i;

    /* scavenge() will evacuate all oldspace pages, but no newspace
     * pages.  Pinned pages are precisely those pages which must not
     * be evacuated, so move them to newspace directly. */

    for (i = 0; i < last_free_page; i++) {
        if (page_table[i].dont_move &&
            /* dont_move is cleared lazily, so validate the space as well. */
            page_table[i].gen == from_space) {
            if (pinned_dwords(i) && do_wipe_p) {
                // do not move to newspace after all, this will be word-wiped
                continue;
            }
            page_table[i].gen = new_space;
            /* And since we're moving the pages wholesale, also adjust
             * the generation allocation counters. */
            generations[new_space].bytes_allocated += page_table[i].bytes_used;
            generations[from_space].bytes_allocated -= page_table[i].bytes_used;
        }
    }
}

/* Garbage collect a generation. If raise is 0 then the remains of the
 * generation are not raised to the next generation. */
static void
garbage_collect_generation(generation_index_t generation, int raise)
{
    page_index_t i;
    uword_t static_space_size;
    struct thread *th;

    gc_assert(generation <= HIGHEST_NORMAL_GENERATION);

    /* The oldest generation can't be raised. */
    gc_assert((generation != HIGHEST_NORMAL_GENERATION) || (raise == 0));

    /* Check if weak hash tables were processed in the previous GC. */
    gc_assert(weak_hash_tables == NULL);

    /* Initialize the weak pointer list. */
    weak_pointers = NULL;

    /* When a generation is not being raised it is transported to a
     * temporary generation (NUM_GENERATIONS), and lowered when
     * done. Set up this new generation. There should be no pages
     * allocated to it yet. */
    if (!raise) {
         gc_assert(generations[SCRATCH_GENERATION].bytes_allocated == 0);
    }

    /* Set the global src and dest. generations */
    from_space = generation;
    if (raise)
        new_space = generation+1;
    else
        new_space = SCRATCH_GENERATION;

    /* Change to a new space for allocation, resetting the alloc_start_page */
    gc_alloc_generation = new_space;
    generations[new_space].alloc_start_page = 0;
    generations[new_space].alloc_unboxed_start_page = 0;
    generations[new_space].alloc_large_start_page = 0;
    generations[new_space].alloc_large_unboxed_start_page = 0;

    /* Before any pointers are preserved, the dont_move flags on the
     * pages need to be cleared. */
    for (i = 0; i < last_free_page; i++)
        if(page_table[i].gen==from_space) {
            page_table[i].dont_move = 0;
            gc_assert(pinned_dwords(i) == NULL);
        }

    /* Un-write-protect the old-space pages. This is essential for the
     * promoted pages as they may contain pointers into the old-space
     * which need to be scavenged. It also helps avoid unnecessary page
     * faults as forwarding pointers are written into them. They need to
     * be un-protected anyway before unmapping later. */
    unprotect_oldspace();

    /* Scavenge the stacks' conservative roots. */

    /* there are potentially two stacks for each thread: the main
     * stack, which may contain Lisp pointers, and the alternate stack.
     * We don't ever run Lisp code on the altstack, but it may
     * host a sigcontext with lisp objects in it */

    /* what we need to do: (1) find the stack pointer for the main
     * stack; scavenge it (2) find the interrupt context on the
     * alternate stack that might contain lisp values, and scavenge
     * that */

    /* we assume that none of the preceding applies to the thread that
     * initiates GC.  If you ever call GC from inside an altstack
     * handler, you will lose. */

#if defined(LISP_FEATURE_X86) || defined(LISP_FEATURE_X86_64)
    /* And if we're saving a core, there's no point in being conservative. */
    if (conservative_stack) {
        for_each_thread(th) {
            void **ptr;
            void **esp=(void **)-1;
            if (th->state == STATE_DEAD)
                continue;
# if defined(LISP_FEATURE_SB_SAFEPOINT)
            /* Conservative collect_garbage is always invoked with a
             * foreign C call or an interrupt handler on top of every
             * existing thread, so the stored SP in each thread
             * structure is valid, no matter which thread we are looking
             * at.  For threads that were running Lisp code, the pitstop
             * and edge functions maintain this value within the
             * interrupt or exception handler. */
            esp = os_get_csp(th);
            assert_on_stack(th, esp);

            /* In addition to pointers on the stack, also preserve the
             * return PC, the only value from the context that we need
             * in addition to the SP.  The return PC gets saved by the
             * foreign call wrapper, and removed from the control stack
             * into a register. */
            preserve_pointer(th->pc_around_foreign_call);

            /* And on platforms with interrupts: scavenge ctx registers. */

            /* Disabled on Windows, because it does not have an explicit
             * stack of `interrupt_contexts'.  The reported CSP has been
             * chosen so that the current context on the stack is
             * covered by the stack scan.  See also set_csp_from_context(). */
#  ifndef LISP_FEATURE_WIN32
            if (th != arch_os_get_current_thread()) {
                long k = fixnum_value(
                    SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
                while (k > 0)
                    preserve_context_registers(th->interrupt_contexts[--k]);
            }
#  endif
# elif defined(LISP_FEATURE_SB_THREAD)
            sword_t i,free;
            if(th==arch_os_get_current_thread()) {
                /* Somebody is going to burn in hell for this, but casting
                 * it in two steps shuts gcc up about strict aliasing. */
                esp = (void **)((void *)&raise);
            } else {
                void **esp1;
                free=fixnum_value(SymbolValue(FREE_INTERRUPT_CONTEXT_INDEX,th));
                for(i=free-1;i>=0;i--) {
                    os_context_t *c=th->interrupt_contexts[i];
                    esp1 = (void **) *os_context_register_addr(c,reg_SP);
                    if (esp1>=(void **)th->control_stack_start &&
                        esp1<(void **)th->control_stack_end) {
                        if(esp1<esp) esp=esp1;
                        preserve_context_registers(c);
                    }
                }
            }
# else
            esp = (void **)((void *)&raise);
# endif
            if (!esp || esp == (void*) -1)
                lose("garbage_collect: no SP known for thread %x (OS %x)",
                     th, th->os_thread);
            for (ptr = ((void **)th->control_stack_end)-1; ptr >= esp;  ptr--) {
                preserve_pointer(*ptr);
            }
        }
    }
#else
    /* Non-x86oid systems don't have "conservative roots" as such, but
     * the same mechanism is used for objects pinned for use by alien
     * code. */
    for_each_thread(th) {
        lispobj pin_list = SymbolTlValue(PINNED_OBJECTS,th);
        while (pin_list != NIL) {
            struct cons *list_entry =
                (struct cons *)native_pointer(pin_list);
            preserve_pointer((void*)list_entry->car);
            pin_list = list_entry->cdr;
        }
    }
#endif

#if QSHOW
    if (gencgc_verbose > 1) {
        sword_t num_dont_move_pages = count_dont_move_pages();
        fprintf(stderr,
                "/non-movable pages due to conservative pointers = %ld (%lu bytes)\n",
                num_dont_move_pages,
                npage_bytes(num_dont_move_pages));
    }
#endif

    /* Now that all of the pinned (dont_move) pages are known, and
     * before we start to scavenge (and thus relocate) objects,
     * relocate the pinned pages to newspace, so that the scavenger
     * will not attempt to relocate their contents. */
    move_pinned_pages_to_newspace();

    /* Scavenge all the rest of the roots. */

#if !defined(LISP_FEATURE_X86) && !defined(LISP_FEATURE_X86_64)
    /*
     * If not x86, we need to scavenge the interrupt context(s) and the
     * control stack.
     */
    {
        struct thread *th;
        for_each_thread(th) {
            scavenge_interrupt_contexts(th);
            scavenge_control_stack(th);
        }

# ifdef LISP_FEATURE_SB_SAFEPOINT
        /* In this case, scrub all stacks right here from the GCing thread
         * instead of doing what the comment below says.  Suboptimal, but
         * easier. */
        for_each_thread(th)
            scrub_thread_control_stack(th);
# else
        /* Scrub the unscavenged control stack space, so that we can't run
         * into any stale pointers in a later GC (this is done by the
         * stop-for-gc handler in the other threads). */
        scrub_control_stack();
# endif
    }
#endif

    /* Scavenge the Lisp functions of the interrupt handlers, taking
     * care to avoid SIG_DFL and SIG_IGN. */
    for (i = 0; i < NSIG; i++) {
        union interrupt_handler handler = interrupt_handlers[i];
        if (!ARE_SAME_HANDLER(handler.c, SIG_IGN) &&
            !ARE_SAME_HANDLER(handler.c, SIG_DFL)) {
            scavenge((lispobj *)(interrupt_handlers + i), 1);
        }
    }
    /* Scavenge the binding stacks. */
    {
        struct thread *th;
        for_each_thread(th) {
            sword_t len= (lispobj *)get_binding_stack_pointer(th) -
                th->binding_stack_start;
            scavenge((lispobj *) th->binding_stack_start,len);
#ifdef LISP_FEATURE_SB_THREAD
            /* do the tls as well */
            len=(SymbolValue(FREE_TLS_INDEX,0) >> WORD_SHIFT) -
                (sizeof (struct thread))/(sizeof (lispobj));
            scavenge((lispobj *) (th+1),len);
#endif
        }
    }

    /* Scavenge static space. */
    static_space_size =
        (lispobj *)SymbolValue(STATIC_SPACE_FREE_POINTER,0) -
        (lispobj *)STATIC_SPACE_START;
    if (gencgc_verbose > 1) {
        FSHOW((stderr,
               "/scavenge static space: %d bytes\n",
               static_space_size * sizeof(lispobj)));
    }
    scavenge( (lispobj *) STATIC_SPACE_START, static_space_size);

    /* All generations but the generation being GCed need to be
     * scavenged. The new_space generation needs special handling as
     * objects may be moved in - it is handled separately below. */
#ifdef LISP_FEATURE_IMMOBILE_SPACE
    scavenge_immobile_roots(generation+1, SCRATCH_GENERATION);
#endif
    scavenge_generations(generation+1, PSEUDO_STATIC_GENERATION);

    scavenge_pinned_ranges();

    /* Finally scavenge the new_space generation. Keep going until no
     * more objects are moved into the new generation */
    scavenge_newspace_generation(new_space);

    /* FIXME: I tried reenabling this check when debugging unrelated
     * GC weirdness ca. sbcl-0.6.12.45, and it failed immediately.
     * Since the current GC code seems to work well, I'm guessing that
     * this debugging code is just stale, but I haven't tried to
     * figure it out. It should be figured out and then either made to
     * work or just deleted. */

#define RESCAN_CHECK 0
#if RESCAN_CHECK
    /* As a check re-scavenge the newspace once; no new objects should
     * be found. */
    {
        os_vm_size_t old_bytes_allocated = bytes_allocated;
        os_vm_size_t bytes_allocated;

        /* Start with a full scavenge. */
        scavenge_newspace_generation_one_scan(new_space);

        /* Flush the current regions, updating the tables. */
        gc_alloc_update_all_page_tables(1);

        bytes_allocated = bytes_allocated - old_bytes_allocated;

        if (bytes_allocated != 0) {
            lose("Rescan of new_space allocated %d more bytes.\n",
                 bytes_allocated);
        }
    }
#endif

    scan_weak_hash_tables();
    scan_weak_pointers();
    wipe_nonpinned_words();
#ifdef LISP_FEATURE_IMMOBILE_SPACE
    // Do this last, because until wipe_nonpinned_words() happens,
    // not all page table entries have the 'gen' value updated,
    // which we need to correctly find all old->young pointers.
    sweep_immobile_space(raise);
#endif

    /* Flush the current regions, updating the tables. */
    gc_alloc_update_all_page_tables(0);

    /* Free the pages in oldspace, but not those marked dont_move. */
    free_oldspace();

    /* If the GC is not raising the age then lower the generation back
     * to its normal generation number */
    if (!raise) {
        for (i = 0; i < last_free_page; i++)
            if ((page_table[i].bytes_used != 0)
                && (page_table[i].gen == SCRATCH_GENERATION))
                page_table[i].gen = generation;
        gc_assert(generations[generation].bytes_allocated == 0);
        generations[generation].bytes_allocated =
            generations[SCRATCH_GENERATION].bytes_allocated;
        generations[SCRATCH_GENERATION].bytes_allocated = 0;
    }

    /* Reset the alloc_start_page for generation. */
    generations[generation].alloc_start_page = 0;
    generations[generation].alloc_unboxed_start_page = 0;
    generations[generation].alloc_large_start_page = 0;
    generations[generation].alloc_large_unboxed_start_page = 0;

    if (generation >= verify_gens) {
        if (gencgc_verbose) {
            SHOW("verifying");
        }
        verify_gc();
    }

    /* Set the new gc trigger for the GCed generation. */
    generations[generation].gc_trigger =
        generations[generation].bytes_allocated
        + generations[generation].bytes_consed_between_gc;

    if (raise)
        generations[generation].num_gc = 0;
    else
        ++generations[generation].num_gc;

}

/* Update last_free_page, then SymbolValue(ALLOCATION_POINTER). */
sword_t
update_dynamic_space_free_pointer(void)
{
    page_index_t last_page = -1, i;

    for (i = 0; i < last_free_page; i++)
        if (page_allocated_p(i) && (page_table[i].bytes_used != 0))
            last_page = i;

    last_free_page = last_page+1;

    set_alloc_pointer((lispobj)(page_address(last_free_page)));
    return 0; /* dummy value: return something ... */
}

static void
remap_page_range (page_index_t from, page_index_t to)
{
    /* There's a mysterious Solaris/x86 problem with using mmap
     * tricks for memory zeroing. See sbcl-devel thread
     * "Re: patch: standalone executable redux".
     */
#if defined(LISP_FEATURE_SUNOS)
    zero_and_mark_pages(from, to);
#else
    const page_index_t
            release_granularity = gencgc_release_granularity/GENCGC_CARD_BYTES,
                   release_mask = release_granularity-1,
                            end = to+1,
                   aligned_from = (from+release_mask)&~release_mask,
                    aligned_end = (end&~release_mask);

    if (aligned_from < aligned_end) {
        zero_pages_with_mmap(aligned_from, aligned_end-1);
        if (aligned_from != from)
            zero_and_mark_pages(from, aligned_from-1);
        if (aligned_end != end)
            zero_and_mark_pages(aligned_end, end-1);
    } else {
        zero_and_mark_pages(from, to);
    }
#endif
}

static void
remap_free_pages (page_index_t from, page_index_t to, int forcibly)
{
    page_index_t first_page, last_page;

    if (forcibly)
        return remap_page_range(from, to);

    for (first_page = from; first_page <= to; first_page++) {
        if (page_allocated_p(first_page) ||
            (page_table[first_page].need_to_zero == 0))
            continue;

        last_page = first_page + 1;
        while (page_free_p(last_page) &&
               (last_page <= to) &&
               (page_table[last_page].need_to_zero == 1))
            last_page++;

        remap_page_range(first_page, last_page-1);

        first_page = last_page;
    }
}

generation_index_t small_generation_limit = 1;

/* GC all generations newer than last_gen, raising the objects in each
 * to the next older generation - we finish when all generations below
 * last_gen are empty.  Then if last_gen is due for a GC, or if
 * last_gen==NUM_GENERATIONS (the scratch generation?  eh?) we GC that
 * too.  The valid range for last_gen is: 0,1,...,NUM_GENERATIONS.
 *
 * We stop collecting at gencgc_oldest_gen_to_gc, even if this is less than
 * last_gen (oh, and note that by default it is NUM_GENERATIONS-1) */
void
collect_garbage(generation_index_t last_gen)
{
    generation_index_t gen = 0, i;
    int raise, more = 0;
    int gen_to_wp;
    /* The largest value of last_free_page seen since the time
     * remap_free_pages was called. */
    static page_index_t high_water_mark = 0;

    FSHOW((stderr, "/entering collect_garbage(%d)\n", last_gen));
    log_generation_stats(gc_logfile, "=== GC Start ===");

    gc_active_p = 1;

    if (last_gen > HIGHEST_NORMAL_GENERATION+1) {
        FSHOW((stderr,
               "/collect_garbage: last_gen = %d, doing a level 0 GC\n",
               last_gen));
        last_gen = 0;
    }

    /* Flush the alloc regions updating the tables. */
    gc_alloc_update_all_page_tables(1);

    /* Verify the new objects created by Lisp code. */
    if (pre_verify_gen_0) {
        FSHOW((stderr, "pre-checking generation 0\n"));
        verify_generation(0);
    }

    if (gencgc_verbose > 1)
        print_generation_stats();

#ifdef LISP_FEATURE_IMMOBILE_SPACE
    /* Immobile space generation bits are lazily updated for gen0
       (not touched on every object allocation) so do it now */
    update_immobile_nursery_bits();
#endif

    do {
        /* Collect the generation. */

        if (more || (gen >= gencgc_oldest_gen_to_gc)) {
            /* Never raise the oldest generation. Never raise the extra generation
             * collected due to more-flag. */
            raise = 0;
            more = 0;
        } else {
            raise =
                (gen < last_gen)
                || (generations[gen].num_gc >= generations[gen].number_of_gcs_before_promotion);
            /* If we would not normally raise this one, but we're
             * running low on space in comparison to the object-sizes
             * we've been seeing, raise it and collect the next one
             * too. */
            if (!raise && gen == last_gen) {
                more = (2*large_allocation) >= (dynamic_space_size - bytes_allocated);
                raise = more;
            }
        }

        if (gencgc_verbose > 1) {
            FSHOW((stderr,
                   "starting GC of generation %d with raise=%d alloc=%d trig=%d GCs=%d\n",
                   gen,
                   raise,
                   generations[gen].bytes_allocated,
                   generations[gen].gc_trigger,
                   generations[gen].num_gc));
        }

        /* If an older generation is being filled, then update its
         * memory age. */
        if (raise == 1) {
            generations[gen+1].cum_sum_bytes_allocated +=
                generations[gen+1].bytes_allocated;
        }

        garbage_collect_generation(gen, raise);

        /* Reset the memory age cum_sum. */
        generations[gen].cum_sum_bytes_allocated = 0;

        if (gencgc_verbose > 1) {
            FSHOW((stderr, "GC of generation %d finished:\n", gen));
            print_generation_stats();
        }

        gen++;
    } while ((gen <= gencgc_oldest_gen_to_gc)
             && ((gen < last_gen)
                 || more
                 || (raise
                     && (generations[gen].bytes_allocated
                         > generations[gen].gc_trigger)
                     && (generation_average_age(gen)
                         > generations[gen].minimum_age_before_gc))));

    /* Now if gen-1 was raised all generations before gen are empty.
     * If it wasn't raised then all generations before gen-1 are empty.
     *
     * Now objects within this gen's pages cannot point to younger
     * generations unless they are written to. This can be exploited
     * by write-protecting the pages of gen; then when younger
     * generations are GCed only the pages which have been written
     * need scanning. */
    if (raise)
        gen_to_wp = gen;
    else
        gen_to_wp = gen - 1;

    /* There's not much point in WPing pages in generation 0 as it is
     * never scavenged (except promoted pages). */
    if ((gen_to_wp > 0) && enable_page_protection) {
        /* Check that they are all empty. */
        for (i = 0; i < gen_to_wp; i++) {
            if (generations[i].bytes_allocated)
                lose("trying to write-protect gen. %d when gen. %d nonempty\n",
                     gen_to_wp, i);
        }
        write_protect_generation_pages(gen_to_wp);
    }
#ifdef LISP_FEATURE_IMMOBILE_SPACE
    write_protect_immobile_space();
#endif

    /* Set gc_alloc() back to generation 0. The current regions should
     * be flushed after the above GCs. */
    gc_assert((boxed_region.free_pointer - boxed_region.start_addr) == 0);
    gc_alloc_generation = 0;

    /* Save the high-water mark before updating last_free_page */
    if (last_free_page > high_water_mark)
        high_water_mark = last_free_page;

    update_dynamic_space_free_pointer();

    /* Update auto_gc_trigger. Make sure we trigger the next GC before
     * running out of heap! */
    if (bytes_consed_between_gcs <= (dynamic_space_size - bytes_allocated))
        auto_gc_trigger = bytes_allocated + bytes_consed_between_gcs;
    else
        auto_gc_trigger = bytes_allocated + (dynamic_space_size - bytes_allocated)/2;

    if(gencgc_verbose)
        fprintf(stderr,"Next gc when %"OS_VM_SIZE_FMT" bytes have been consed\n",
                auto_gc_trigger);

    /* If we did a big GC (arbitrarily defined as gen > 1), release memory
     * back to the OS.
     */
    if (gen > small_generation_limit) {
        if (last_free_page > high_water_mark)
            high_water_mark = last_free_page;
        remap_free_pages(0, high_water_mark, 0);
        high_water_mark = 0;
    }

    gc_active_p = 0;
    large_allocation = 0;

    log_generation_stats(gc_logfile, "=== GC End ===");
    SHOW("returning from collect_garbage");
}
\f
void
gc_init(void)
{
    page_index_t i;

#if defined(LISP_FEATURE_SB_SAFEPOINT)
    alloc_gc_page();
#endif

    /* Compute the number of pages needed for the dynamic space.
     * Dynamic space size should be aligned on page size. */
    page_table_pages = dynamic_space_size/GENCGC_CARD_BYTES;
    gc_assert(dynamic_space_size == npage_bytes(page_table_pages));

    /* Default nursery size to 5% of the total dynamic space size,
     * min 1Mb. */
    bytes_consed_between_gcs = dynamic_space_size/(os_vm_size_t)20;
    if (bytes_consed_between_gcs < (1024*1024))
        bytes_consed_between_gcs = 1024*1024;

    /* The page_table must be allocated using "calloc" to initialize
     * the page structures correctly. There used to be a separate
     * initialization loop (now commented out; see below) but that was
     * unnecessary and did hurt startup time. */
    page_table = calloc(page_table_pages, sizeof(struct page));
    gc_assert(page_table);
#ifdef LISP_FEATURE_IMMOBILE_SPACE
    gc_init_immobile();
#endif

    size_t pins_map_size_in_bytes =
      (n_dwords_in_card / N_WORD_BITS) * sizeof (uword_t) * page_table_pages;
    /* We use mmap directly here so that we can use a minimum of
       system calls per page during GC.
       All we need here now is a madvise(DONTNEED) at the end of GC. */
    page_table_pinned_dwords
      = (in_use_marker_t*)os_validate(NULL, pins_map_size_in_bytes);
    /* We do not need to zero */
    gc_assert(page_table_pinned_dwords);

    scavtab[WEAK_POINTER_WIDETAG] = scav_weak_pointer;
    transother[SIMPLE_ARRAY_WIDETAG] = trans_boxed_large;

    heap_base = (void*)DYNAMIC_SPACE_START;

    /* The page structures are initialized implicitly when page_table
     * is allocated with "calloc" above. Formerly we had the following
     * explicit initialization here (comments converted to C99 style
     * for readability as C's block comments don't nest):
     *
     * // Initialize each page structure.
     * for (i = 0; i < page_table_pages; i++) {
     *     // Initialize all pages as free.
     *     page_table[i].allocated = FREE_PAGE_FLAG;
     *     page_table[i].bytes_used = 0;
     *
     *     // Pages are not write-protected at startup.
     *     page_table[i].write_protected = 0;
     * }
     *
     * Without this loop the image starts up much faster when dynamic
     * space is large -- which it is on 64-bit platforms already by
     * default -- and when "calloc" for large arrays is implemented
     * using copy-on-write of a page of zeroes -- which it is at least
     * on Linux. In this case the pages that page_table_pages is stored
     * in are mapped and cleared not before the corresponding part of
     * dynamic space is used. For example, this saves clearing 16 MB of
     * memory at startup if the page size is 4 KB and the size of
     * dynamic space is 4 GB.
     * FREE_PAGE_FLAG must be 0 for this to work correctly which is
     * asserted below: */
    {
      /* Compile time assertion: If triggered, declares an array
       * of dimension -1 forcing a syntax error. The intent of the
       * assignment is to avoid an "unused variable" warning. */
      char assert_free_page_flag_0[(FREE_PAGE_FLAG) ? -1 : 1];
      assert_free_page_flag_0[0] = assert_free_page_flag_0[0];
    }

    bytes_allocated = 0;

    /* Initialize the generations. */
    for (i = 0; i < NUM_GENERATIONS; i++) {
        generations[i].alloc_start_page = 0;
        generations[i].alloc_unboxed_start_page = 0;
        generations[i].alloc_large_start_page = 0;
        generations[i].alloc_large_unboxed_start_page = 0;
        generations[i].bytes_allocated = 0;
        generations[i].gc_trigger = 2000000;
        generations[i].num_gc = 0;
        generations[i].cum_sum_bytes_allocated = 0;
        /* the tune-able parameters */
        generations[i].bytes_consed_between_gc
            = bytes_consed_between_gcs/(os_vm_size_t)HIGHEST_NORMAL_GENERATION;
        generations[i].number_of_gcs_before_promotion = 1;
        generations[i].minimum_age_before_gc = 0.75;
    }

    /* Initialize gc_alloc. */
    gc_alloc_generation = 0;
    gc_set_region_empty(&boxed_region);
    gc_set_region_empty(&unboxed_region);

    last_free_page = 0;
}

/*  Pick up the dynamic space from after a core load.
 *
 *  The ALLOCATION_POINTER points to the end of the dynamic space.
 */

static void
gencgc_pickup_dynamic(void)
{
    page_index_t page = 0;
    void *alloc_ptr = (void *)get_alloc_pointer();
    lispobj *prev=(lispobj *)page_address(page);
    generation_index_t gen = PSEUDO_STATIC_GENERATION;

    bytes_allocated = 0;

    do {
        lispobj *first,*ptr= (lispobj *)page_address(page);

        if (!gencgc_partial_pickup || page_allocated_p(page)) {
          /* It is possible, though rare, for the saved page table
           * to contain free pages below alloc_ptr. */
          page_table[page].gen = gen;
          page_table[page].bytes_used = GENCGC_CARD_BYTES;
          page_table[page].large_object = 0;
          page_table[page].write_protected = 0;
          page_table[page].write_protected_cleared = 0;
          page_table[page].dont_move = 0;
          page_table[page].need_to_zero = 1;

          bytes_allocated += GENCGC_CARD_BYTES;
        }

        if (!gencgc_partial_pickup) {
            page_table[page].allocated = BOXED_PAGE_FLAG;
            first = gc_search_space3(ptr, prev, (ptr+2));
            if(ptr == first)
                prev=ptr;
            page_table[page].scan_start_offset =
                page_address(page) - (void *)prev;
        }
        page++;
    } while (page_address(page) < alloc_ptr);

    last_free_page = page;

    generations[gen].bytes_allocated = bytes_allocated;

    gc_alloc_update_all_page_tables(1);
    write_protect_generation_pages(gen);
}

void
gc_initialize_pointers(void)
{
    gencgc_pickup_dynamic();
}
\f

/* alloc(..) is the external interface for memory allocation. It
 * allocates to generation 0. It is not called from within the garbage
 * collector as it is only external uses that need the check for heap
 * size (GC trigger) and to disable the interrupts (interrupts are
 * always disabled during a GC).
 *
 * The vops that call alloc(..) assume that the returned space is zero-filled.
 * (E.g. the most significant word of a 2-word bignum in MOVE-FROM-UNSIGNED.)
 *
 * The check for a GC trigger is only performed when the current
 * region is full, so in most cases it's not needed. */

static inline lispobj *
general_alloc_internal(sword_t nbytes, int page_type_flag, struct alloc_region *region,
                       struct thread *thread)
{
#ifndef LISP_FEATURE_WIN32
    lispobj alloc_signal;
#endif
    void *new_obj;
    void *new_free_pointer;
    os_vm_size_t trigger_bytes = 0;

    gc_assert(nbytes > 0);

    /* Check for alignment allocation problems. */
    gc_assert((((uword_t)region->free_pointer & LOWTAG_MASK) == 0)
              && ((nbytes & LOWTAG_MASK) == 0));

#if !(defined(LISP_FEATURE_WIN32) && defined(LISP_FEATURE_SB_THREAD))
    /* Must be inside a PA section. */
    gc_assert(get_pseudo_atomic_atomic(thread));
#endif

    if ((os_vm_size_t) nbytes > large_allocation)
        large_allocation = nbytes;

    /* maybe we can do this quickly ... */
    new_free_pointer = region->free_pointer + nbytes;
    if (new_free_pointer <= region->end_addr) {
        new_obj = (void*)(region->free_pointer);
        region->free_pointer = new_free_pointer;
        return(new_obj);        /* yup */
    }

    /* We don't want to count nbytes against auto_gc_trigger unless we
     * have to: it speeds up the tenuring of objects and slows down
     * allocation. However, unless we do so when allocating _very_
     * large objects we are in danger of exhausting the heap without
     * running sufficient GCs.
     */
    if ((os_vm_size_t) nbytes >= bytes_consed_between_gcs)
        trigger_bytes = nbytes;

    /* we have to go the long way around, it seems. Check whether we
     * should GC in the near future
     */
    if (auto_gc_trigger && (bytes_allocated+trigger_bytes > auto_gc_trigger)) {
        /* Don't flood the system with interrupts if the need to gc is
         * already noted. This can happen for example when SUB-GC
         * allocates or after a gc triggered in a WITHOUT-GCING. */
        if (SymbolValue(GC_PENDING,thread) == NIL) {
            /* set things up so that GC happens when we finish the PA
             * section */
            SetSymbolValue(GC_PENDING,T,thread);
            if (SymbolValue(GC_INHIBIT,thread) == NIL) {
#ifdef LISP_FEATURE_SB_SAFEPOINT
                thread_register_gc_trigger();
#else
                set_pseudo_atomic_interrupted(thread);
#ifdef GENCGC_IS_PRECISE
                /* PPC calls alloc() from a trap
                 * look up the most context if it's from a trap. */
                {
                    os_context_t *context =
                        thread->interrupt_data->allocation_trap_context;
                    maybe_save_gc_mask_and_block_deferrables
                        (context ? os_context_sigmask_addr(context) : NULL);
                }
#else
                maybe_save_gc_mask_and_block_deferrables(NULL);
#endif
#endif
            }
        }
    }
    new_obj = gc_alloc_with_region(nbytes, page_type_flag, region, 0);

#ifndef LISP_FEATURE_WIN32
    /* for sb-prof, and not supported on Windows yet */
    alloc_signal = SymbolValue(ALLOC_SIGNAL,thread);
    if ((alloc_signal & FIXNUM_TAG_MASK) == 0) {
        if ((sword_t) alloc_signal <= 0) {
            SetSymbolValue(ALLOC_SIGNAL, T, thread);
            raise(SIGPROF);
        } else {
            SetSymbolValue(ALLOC_SIGNAL,
                           alloc_signal - (1 << N_FIXNUM_TAG_BITS),
                           thread);
        }
    }
#endif

    return (new_obj);
}

lispobj *
general_alloc(sword_t nbytes, int page_type_flag)
{
    struct thread *thread = arch_os_get_current_thread();
    /* Select correct region, and call general_alloc_internal with it.
     * For other then boxed allocation we must lock first, since the
     * region is shared. */
    if (BOXED_PAGE_FLAG & page_type_flag) {
#ifdef LISP_FEATURE_SB_THREAD
        struct alloc_region *region = (thread ? &(thread->alloc_region) : &boxed_region);
#else
        struct alloc_region *region = &boxed_region;
#endif
        return general_alloc_internal(nbytes, page_type_flag, region, thread);
    } else if (UNBOXED_PAGE_FLAG == page_type_flag) {
        lispobj * obj;
        int result;
        result = thread_mutex_lock(&allocation_lock);
        gc_assert(!result);
        obj = general_alloc_internal(nbytes, page_type_flag, &unboxed_region, thread);
        result = thread_mutex_unlock(&allocation_lock);
        gc_assert(!result);
        return obj;
    } else {
        lose("bad page type flag: %d", page_type_flag);
    }
}

lispobj AMD64_SYSV_ABI *
alloc(sword_t nbytes)
{
#ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
    struct thread *self = arch_os_get_current_thread();
    int was_pseudo_atomic = get_pseudo_atomic_atomic(self);
    if (!was_pseudo_atomic)
        set_pseudo_atomic_atomic(self);
#else
    gc_assert(get_pseudo_atomic_atomic(arch_os_get_current_thread()));
#endif

    lispobj *result = general_alloc(nbytes, BOXED_PAGE_FLAG);

#ifdef LISP_FEATURE_SB_SAFEPOINT_STRICTLY
    if (!was_pseudo_atomic)
        clear_pseudo_atomic_atomic(self);
#endif

    return result;
}
\f
/*
 * shared support for the OS-dependent signal handlers which
 * catch GENCGC-related write-protect violations
 */
void unhandled_sigmemoryfault(void* addr);

/* Depending on which OS we're running under, different signals might
 * be raised for a violation of write protection in the heap. This
 * function factors out the common generational GC magic which needs
 * to invoked in this case, and should be called from whatever signal
 * handler is appropriate for the OS we're running under.
 *
 * Return true if this signal is a normal generational GC thing that
 * we were able to handle, or false if it was abnormal and control
 * should fall through to the general SIGSEGV/SIGBUS/whatever logic.
 *
 * We have two control flags for this: one causes us to ignore faults
 * on unprotected pages completely, and the second complains to stderr
 * but allows us to continue without losing.
 */
extern boolean ignore_memoryfaults_on_unprotected_pages;
boolean ignore_memoryfaults_on_unprotected_pages = 0;

extern boolean continue_after_memoryfault_on_unprotected_pages;
boolean continue_after_memoryfault_on_unprotected_pages = 0;

int
gencgc_handle_wp_violation(void* fault_addr)
{
    page_index_t page_index = find_page_index(fault_addr);

#if QSHOW_SIGNALS
    FSHOW((stderr,
           "heap WP violation? fault_addr=%p, page_index=%"PAGE_INDEX_FMT"\n",
           fault_addr, page_index));
#endif

    /* Check whether the fault is within the dynamic space. */
    if (page_index == (-1)) {
#ifdef LISP_FEATURE_IMMOBILE_SPACE
        extern int immobile_space_handle_wp_violation(void*);
        if (immobile_space_handle_wp_violation(fault_addr))
            return 1;
#endif

        /* It can be helpful to be able to put a breakpoint on this
         * case to help diagnose low-level problems. */
        unhandled_sigmemoryfault(fault_addr);

        /* not within the dynamic space -- not our responsibility */
        return 0;

    } else {
        int ret;
        ret = thread_mutex_lock(&free_pages_lock);
        gc_assert(ret == 0);
        if (page_table[page_index].write_protected) {
            /* Unprotect the page. */
            os_protect(page_address(page_index), GENCGC_CARD_BYTES, OS_VM_PROT_ALL);
            page_table[page_index].write_protected_cleared = 1;
            page_table[page_index].write_protected = 0;
        } else if (!ignore_memoryfaults_on_unprotected_pages) {
            /* The only acceptable reason for this signal on a heap
             * access is that GENCGC write-protected the page.
             * However, if two CPUs hit a wp page near-simultaneously,
             * we had better not have the second one lose here if it
             * does this test after the first one has already set wp=0
             */
            if(page_table[page_index].write_protected_cleared != 1) {
                void lisp_backtrace(int frames);
                lisp_backtrace(10);
                fprintf(stderr,
                        "Fault @ %p, page %"PAGE_INDEX_FMT" not marked as write-protected:\n"
                        "  boxed_region.first_page: %"PAGE_INDEX_FMT","
                        "  boxed_region.last_page %"PAGE_INDEX_FMT"\n"
                        "  page.scan_start_offset: %"OS_VM_SIZE_FMT"\n"
                        "  page.bytes_used: %"PAGE_BYTES_FMT"\n"
                        "  page.allocated: %d\n"
                        "  page.write_protected: %d\n"
                        "  page.write_protected_cleared: %d\n"
                        "  page.generation: %d\n",
                        fault_addr,
                        page_index,
                        boxed_region.first_page,
                        boxed_region.last_page,
                        page_table[page_index].scan_start_offset,
                        page_table[page_index].bytes_used,
                        page_table[page_index].allocated,
                        page_table[page_index].write_protected,
                        page_table[page_index].write_protected_cleared,
                        page_table[page_index].gen);
                if (!continue_after_memoryfault_on_unprotected_pages)
                    lose("Feh.\n");
            }
        }
        ret = thread_mutex_unlock(&free_pages_lock);
        gc_assert(ret == 0);
        /* Don't worry, we can handle it. */
        return 1;
    }
}
/* This is to be called when we catch a SIGSEGV/SIGBUS, determine that
 * it's not just a case of the program hitting the write barrier, and
 * are about to let Lisp deal with it. It's basically just a
 * convenient place to set a gdb breakpoint. */
void
unhandled_sigmemoryfault(void *addr)
{}

static void
update_thread_page_tables(struct thread *th)
{
    gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->alloc_region);
#if defined(LISP_FEATURE_SB_SAFEPOINT_STRICTLY) && !defined(LISP_FEATURE_WIN32)
    gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &th->sprof_alloc_region);
#endif
}

/* GC is single-threaded and all memory allocations during a
   collection happen in the GC thread, so it is sufficient to update
   all the the page tables once at the beginning of a collection and
   update only page tables of the GC thread during the collection. */
void gc_alloc_update_all_page_tables(int for_all_threads)
{
    /* Flush the alloc regions updating the tables. */
    struct thread *th;
    if (for_all_threads) {
        for_each_thread(th) {
            update_thread_page_tables(th);
        }
    }
    else {
        th = arch_os_get_current_thread();
        if (th) {
            update_thread_page_tables(th);
        }
    }
    gc_alloc_update_page_tables(UNBOXED_PAGE_FLAG, &unboxed_region);
    gc_alloc_update_page_tables(BOXED_PAGE_FLAG, &boxed_region);
}

void
gc_set_region_empty(struct alloc_region *region)
{
    region->first_page = 0;
    region->last_page = -1;
    region->start_addr = page_address(0);
    region->free_pointer = page_address(0);
    region->end_addr = page_address(0);
}

static void
zero_all_free_pages()
{
    page_index_t i;

    for (i = 0; i < last_free_page; i++) {
        if (page_free_p(i)) {
#ifdef READ_PROTECT_FREE_PAGES
            os_protect(page_address(i),
                       GENCGC_CARD_BYTES,
                       OS_VM_PROT_ALL);
#endif
            zero_pages(i, i);
        }
    }
}

/* Things to do before doing a final GC before saving a core (without
 * purify).
 *
 * + Pages in large_object pages aren't moved by the GC, so we need to
 *   unset that flag from all pages.
 * + The pseudo-static generation isn't normally collected, but it seems
 *   reasonable to collect it at least when saving a core. So move the
 *   pages to a normal generation.
 */
static void
prepare_for_final_gc ()
{
    page_index_t i;

#ifdef LISP_FEATURE_IMMOBILE_SPACE
    extern void prepare_immobile_space_for_final_gc();
    prepare_immobile_space_for_final_gc ();
#endif
    do_wipe_p = 0;
    for (i = 0; i < last_free_page; i++) {
        page_table[i].large_object = 0;
        if (page_table[i].gen == PSEUDO_STATIC_GENERATION) {
            int used = page_table[i].bytes_used;
            page_table[i].gen = HIGHEST_NORMAL_GENERATION;
            generations[PSEUDO_STATIC_GENERATION].bytes_allocated -= used;
            generations[HIGHEST_NORMAL_GENERATION].bytes_allocated += used;
        }
    }
}


/* Do a non-conservative GC, and then save a core with the initial
 * function being set to the value of the static symbol
 * SB!VM:RESTART-LISP-FUNCTION */
void
gc_and_save(char *filename, boolean prepend_runtime,
            boolean save_runtime_options, boolean compressed,
            int compression_level, int application_type)
{
    FILE *file;
    void *runtime_bytes = NULL;
    size_t runtime_size;

    file = prepare_to_save(filename, prepend_runtime, &runtime_bytes,
                           &runtime_size);
    if (file == NULL)
       return;

    conservative_stack = 0;

    /* The filename might come from Lisp, and be moved by the now
     * non-conservative GC. */
    filename = strdup(filename);

    /* Collect twice: once into relatively high memory, and then back
     * into low memory. This compacts the retained data into the lower
     * pages, minimizing the size of the core file.
     */
    prepare_for_final_gc();
    gencgc_alloc_start_page = last_free_page;
    collect_garbage(HIGHEST_NORMAL_GENERATION+1);

    prepare_for_final_gc();
    gencgc_alloc_start_page = -1;
    collect_garbage(HIGHEST_NORMAL_GENERATION+1);

    if (prepend_runtime)
        save_runtime_to_filehandle(file, runtime_bytes, runtime_size,
                                   application_type);

    /* The dumper doesn't know that pages need to be zeroed before use. */
    zero_all_free_pages();
    save_to_filehandle(file, filename, SymbolValue(RESTART_LISP_FUNCTION,0),
                       prepend_runtime, save_runtime_options,
                       compressed ? compression_level : COMPRESSION_LEVEL_NONE);
    /* Oops. Save still managed to fail. Since we've mangled the stack
     * beyond hope, there's not much we can do.
     * (beyond FUNCALLing RESTART_LISP_FUNCTION, but I suspect that's
     * going to be rather unsatisfactory too... */
    lose("Attempt to save core after non-conservative GC failed.\n");
}