implemented Setup.hs to build boehm cpp libs and install them;
[hs-boehmgc.git] / gc-7.2 / mallocx.c
diff --git a/gc-7.2/mallocx.c b/gc-7.2/mallocx.c
new file mode 100644 (file)
index 0000000..e4902a7
--- /dev/null
@@ -0,0 +1,609 @@
+/*
+ * Copyright 1988, 1989 Hans-J. Boehm, Alan J. Demers
+ * Copyright (c) 1991-1994 by Xerox Corporation.  All rights reserved.
+ * Copyright (c) 1996 by Silicon Graphics.  All rights reserved.
+ * Copyright (c) 2000 by Hewlett-Packard Company.  All rights reserved.
+ *
+ * THIS MATERIAL IS PROVIDED AS IS, WITH ABSOLUTELY NO WARRANTY EXPRESSED
+ * OR IMPLIED.  ANY USE IS AT YOUR OWN RISK.
+ *
+ * Permission is hereby granted to use or copy this program
+ * for any purpose,  provided the above notices are retained on all copies.
+ * Permission to modify the code and to distribute modified code is granted,
+ * provided the above notices are retained, and a notice that the code was
+ * modified is included with the above copyright notice.
+ */
+
+#include "private/gc_priv.h"
+
+/*
+ * These are extra allocation routines which are likely to be less
+ * frequently used than those in malloc.c.  They are separate in the
+ * hope that the .o file will be excluded from statically linked
+ * executables.  We should probably break this up further.
+ */
+
+#include <stdio.h>
+#include <string.h>
+
+#ifdef MSWINCE
+# ifndef WIN32_LEAN_AND_MEAN
+#   define WIN32_LEAN_AND_MEAN 1
+# endif
+# define NOSERVICE
+# include <windows.h>
+#else
+# include <errno.h>
+#endif
+
+/* Some externally visible but unadvertised variables to allow access to */
+/* free lists from inlined allocators without including gc_priv.h        */
+/* or introducing dependencies on internal data structure layouts.       */
+void ** const GC_objfreelist_ptr = GC_objfreelist;
+void ** const GC_aobjfreelist_ptr = GC_aobjfreelist;
+void ** const GC_uobjfreelist_ptr = GC_uobjfreelist;
+# ifdef ATOMIC_UNCOLLECTABLE
+    void ** const GC_auobjfreelist_ptr = GC_auobjfreelist;
+# endif
+
+
+STATIC void * GC_generic_or_special_malloc(size_t lb, int knd)
+{
+    switch(knd) {
+#     ifdef STUBBORN_ALLOC
+        case STUBBORN:
+            return(GC_malloc_stubborn((size_t)lb));
+#     endif
+        case PTRFREE:
+            return(GC_malloc_atomic((size_t)lb));
+        case NORMAL:
+            return(GC_malloc((size_t)lb));
+        case UNCOLLECTABLE:
+            return(GC_malloc_uncollectable((size_t)lb));
+#       ifdef ATOMIC_UNCOLLECTABLE
+          case AUNCOLLECTABLE:
+            return(GC_malloc_atomic_uncollectable((size_t)lb));
+#       endif /* ATOMIC_UNCOLLECTABLE */
+        default:
+            return(GC_generic_malloc(lb,knd));
+    }
+}
+
+/* Change the size of the block pointed to by p to contain at least   */
+/* lb bytes.  The object may be (and quite likely will be) moved.     */
+/* The kind (e.g. atomic) is the same as that of the old.             */
+/* Shrinking of large blocks is not implemented well.                 */
+GC_API void * GC_CALL GC_realloc(void * p, size_t lb)
+{
+    struct hblk * h;
+    hdr * hhdr;
+    size_t sz;   /* Current size in bytes       */
+    size_t orig_sz;      /* Original sz in bytes        */
+    int obj_kind;
+
+    if (p == 0) return(GC_malloc(lb));  /* Required by ANSI */
+    h = HBLKPTR(p);
+    hhdr = HDR(h);
+    sz = hhdr -> hb_sz;
+    obj_kind = hhdr -> hb_obj_kind;
+    orig_sz = sz;
+
+    if (sz > MAXOBJBYTES) {
+        /* Round it up to the next whole heap block */
+          word descr;
+
+          sz = (sz+HBLKSIZE-1) & (~HBLKMASK);
+          hhdr -> hb_sz = sz;
+          descr = GC_obj_kinds[obj_kind].ok_descriptor;
+          if (GC_obj_kinds[obj_kind].ok_relocate_descr) descr += sz;
+          hhdr -> hb_descr = descr;
+#         ifdef MARK_BIT_PER_OBJ
+            GC_ASSERT(hhdr -> hb_inv_sz == LARGE_INV_SZ);
+#         else
+            GC_ASSERT(hhdr -> hb_large_block &&
+                      hhdr -> hb_map[ANY_INDEX] == 1);
+#         endif
+          if (IS_UNCOLLECTABLE(obj_kind)) GC_non_gc_bytes += (sz - orig_sz);
+          /* Extra area is already cleared by GC_alloc_large_and_clear. */
+    }
+    if (ADD_SLOP(lb) <= sz) {
+        if (lb >= (sz >> 1)) {
+#           ifdef STUBBORN_ALLOC
+                if (obj_kind == STUBBORN) GC_change_stubborn(p);
+#           endif
+            if (orig_sz > lb) {
+              /* Clear unneeded part of object to avoid bogus pointer */
+              /* tracing.                                             */
+              /* Safe for stubborn objects.                           */
+                BZERO(((ptr_t)p) + lb, orig_sz - lb);
+            }
+            return(p);
+        } else {
+            /* shrink */
+              void * result =
+                        GC_generic_or_special_malloc((word)lb, obj_kind);
+
+              if (result == 0) return(0);
+                  /* Could also return original object.  But this       */
+                  /* gives the client warning of imminent disaster.     */
+              BCOPY(p, result, lb);
+#             ifndef IGNORE_FREE
+                GC_free(p);
+#             endif
+              return(result);
+        }
+    } else {
+        /* grow */
+          void * result =
+                GC_generic_or_special_malloc((word)lb, obj_kind);
+
+          if (result == 0) return(0);
+          BCOPY(p, result, sz);
+#         ifndef IGNORE_FREE
+            GC_free(p);
+#         endif
+          return(result);
+    }
+}
+
+# if defined(REDIRECT_MALLOC) && !defined(REDIRECT_REALLOC)
+#   define REDIRECT_REALLOC GC_realloc
+# endif
+
+# ifdef REDIRECT_REALLOC
+
+/* As with malloc, avoid two levels of extra calls here.        */
+
+# define GC_debug_realloc_replacement(p, lb) \
+        GC_debug_realloc(p, lb, GC_DBG_RA "unknown", 0)
+
+void * realloc(void * p, size_t lb)
+  {
+    return(REDIRECT_REALLOC(p, lb));
+  }
+
+# undef GC_debug_realloc_replacement
+# endif /* REDIRECT_REALLOC */
+
+
+/* Allocate memory such that only pointers to near the          */
+/* beginning of the object are considered.                      */
+/* We avoid holding allocation lock while we clear memory.      */
+GC_INNER void * GC_generic_malloc_ignore_off_page(size_t lb, int k)
+{
+    void *result;
+    size_t lg;
+    size_t lb_rounded;
+    word n_blocks;
+    GC_bool init;
+    DCL_LOCK_STATE;
+
+    if (SMALL_OBJ(lb))
+        return(GC_generic_malloc((word)lb, k));
+    lg = ROUNDED_UP_GRANULES(lb);
+    lb_rounded = GRANULES_TO_BYTES(lg);
+    if (lb_rounded < lb)
+        return((*GC_get_oom_fn())(lb));
+    n_blocks = OBJ_SZ_TO_BLOCKS(lb_rounded);
+    init = GC_obj_kinds[k].ok_init;
+    if (GC_have_errors) GC_print_all_errors();
+    GC_INVOKE_FINALIZERS();
+    LOCK();
+    result = (ptr_t)GC_alloc_large(ADD_SLOP(lb), k, IGNORE_OFF_PAGE);
+    if (0 != result) {
+        if (GC_debugging_started) {
+            BZERO(result, n_blocks * HBLKSIZE);
+        } else {
+#           ifdef THREADS
+              /* Clear any memory that might be used for GC descriptors */
+              /* before we release the lock.                          */
+                ((word *)result)[0] = 0;
+                ((word *)result)[1] = 0;
+                ((word *)result)[GRANULES_TO_WORDS(lg)-1] = 0;
+                ((word *)result)[GRANULES_TO_WORDS(lg)-2] = 0;
+#           endif
+        }
+    }
+    GC_bytes_allocd += lb_rounded;
+    if (0 == result) {
+        GC_oom_func oom_fn = GC_oom_fn;
+        UNLOCK();
+        return((*oom_fn)(lb));
+    } else {
+        UNLOCK();
+        if (init && !GC_debugging_started) {
+            BZERO(result, n_blocks * HBLKSIZE);
+        }
+        return(result);
+    }
+}
+
+GC_API void * GC_CALL GC_malloc_ignore_off_page(size_t lb)
+{
+    return((void *)GC_generic_malloc_ignore_off_page(lb, NORMAL));
+}
+
+GC_API void * GC_CALL GC_malloc_atomic_ignore_off_page(size_t lb)
+{
+    return((void *)GC_generic_malloc_ignore_off_page(lb, PTRFREE));
+}
+
+/* Increment GC_bytes_allocd from code that doesn't have direct access  */
+/* to GC_arrays.                                                        */
+GC_API void GC_CALL GC_incr_bytes_allocd(size_t n)
+{
+    GC_bytes_allocd += n;
+}
+
+/* The same for GC_bytes_freed.                         */
+GC_API void GC_CALL GC_incr_bytes_freed(size_t n)
+{
+    GC_bytes_freed += n;
+}
+
+# ifdef PARALLEL_MARK
+    STATIC volatile signed_word GC_bytes_allocd_tmp = 0;
+                        /* Number of bytes of memory allocated since    */
+                        /* we released the GC lock.  Instead of         */
+                        /* reacquiring the GC lock just to add this in, */
+                        /* we add it in the next time we reacquire      */
+                        /* the lock.  (Atomically adding it doesn't     */
+                        /* work, since we would have to atomically      */
+                        /* update it in GC_malloc, which is too         */
+                        /* expensive.)                                  */
+# endif /* PARALLEL_MARK */
+
+/* Return a list of 1 or more objects of the indicated size, linked     */
+/* through the first word in the object.  This has the advantage that   */
+/* it acquires the allocation lock only once, and may greatly reduce    */
+/* time wasted contending for the allocation lock.  Typical usage would */
+/* be in a thread that requires many items of the same size.  It would  */
+/* keep its own free list in thread-local storage, and call             */
+/* GC_malloc_many or friends to replenish it.  (We do not round up      */
+/* object sizes, since a call indicates the intention to consume many   */
+/* objects of exactly this size.)                                       */
+/* We assume that the size is a multiple of GRANULE_BYTES.              */
+/* We return the free-list by assigning it to *result, since it is      */
+/* not safe to return, e.g. a linked list of pointer-free objects,      */
+/* since the collector would not retain the entire list if it were      */
+/* invoked just as we were returning.                                   */
+/* Note that the client should usually clear the link field.            */
+GC_API void GC_CALL GC_generic_malloc_many(size_t lb, int k, void **result)
+{
+    void *op;
+    void *p;
+    void **opp;
+    size_t lw;      /* Length in words.     */
+    size_t lg;      /* Length in granules.  */
+    signed_word my_bytes_allocd = 0;
+    struct obj_kind * ok = &(GC_obj_kinds[k]);
+    DCL_LOCK_STATE;
+
+    GC_ASSERT(lb != 0 && (lb & (GRANULE_BYTES-1)) == 0);
+    if (!SMALL_OBJ(lb)) {
+        op = GC_generic_malloc(lb, k);
+        if(0 != op) obj_link(op) = 0;
+        *result = op;
+        return;
+    }
+    lw = BYTES_TO_WORDS(lb);
+    lg = BYTES_TO_GRANULES(lb);
+    if (GC_have_errors) GC_print_all_errors();
+    GC_INVOKE_FINALIZERS();
+    LOCK();
+    if (!GC_is_initialized) GC_init();
+    /* Do our share of marking work */
+      if (GC_incremental && !GC_dont_gc) {
+        ENTER_GC();
+        GC_collect_a_little_inner(1);
+        EXIT_GC();
+      }
+    /* First see if we can reclaim a page of objects waiting to be */
+    /* reclaimed.                                                  */
+    {
+        struct hblk ** rlh = ok -> ok_reclaim_list;
+        struct hblk * hbp;
+        hdr * hhdr;
+
+        rlh += lg;
+        while ((hbp = *rlh) != 0) {
+            hhdr = HDR(hbp);
+            *rlh = hhdr -> hb_next;
+            GC_ASSERT(hhdr -> hb_sz == lb);
+            hhdr -> hb_last_reclaimed = (unsigned short) GC_gc_no;
+#           ifdef PARALLEL_MARK
+              if (GC_parallel) {
+                  signed_word my_bytes_allocd_tmp = GC_bytes_allocd_tmp;
+
+                  GC_ASSERT(my_bytes_allocd_tmp >= 0);
+                  /* We only decrement it while holding the GC lock.    */
+                  /* Thus we can't accidentally adjust it down in more  */
+                  /* than one thread simultaneously.                    */
+                  if (my_bytes_allocd_tmp != 0) {
+                    (void)AO_fetch_and_add(
+                                (volatile void *)(&GC_bytes_allocd_tmp),
+                                (AO_t)(-my_bytes_allocd_tmp));
+                    GC_bytes_allocd += my_bytes_allocd_tmp;
+                  }
+                  GC_acquire_mark_lock();
+                  ++ GC_fl_builder_count;
+                  UNLOCK();
+                  GC_release_mark_lock();
+              }
+#           endif
+            op = GC_reclaim_generic(hbp, hhdr, lb,
+                                    ok -> ok_init, 0, &my_bytes_allocd);
+            if (op != 0) {
+              /* We also reclaimed memory, so we need to adjust         */
+              /* that count.                                            */
+              /* This should be atomic, so the results may be           */
+              /* inaccurate.                                            */
+              GC_bytes_found += my_bytes_allocd;
+#             ifdef PARALLEL_MARK
+                if (GC_parallel) {
+                  *result = op;
+                  (void)AO_fetch_and_add(
+                                (volatile AO_t *)(&GC_bytes_allocd_tmp),
+                                (AO_t)(my_bytes_allocd));
+                  GC_acquire_mark_lock();
+                  -- GC_fl_builder_count;
+                  if (GC_fl_builder_count == 0) GC_notify_all_builder();
+                  GC_release_mark_lock();
+                  (void) GC_clear_stack(0);
+                  return;
+                }
+#             endif
+              GC_bytes_allocd += my_bytes_allocd;
+              goto out;
+            }
+#           ifdef PARALLEL_MARK
+              if (GC_parallel) {
+                GC_acquire_mark_lock();
+                -- GC_fl_builder_count;
+                if (GC_fl_builder_count == 0) GC_notify_all_builder();
+                GC_release_mark_lock();
+                LOCK();
+                /* GC lock is needed for reclaim list access.   We      */
+                /* must decrement fl_builder_count before reaquiring GC */
+                /* lock.  Hopefully this path is rare.                  */
+              }
+#           endif
+        }
+    }
+    /* Next try to use prefix of global free list if there is one.      */
+    /* We don't refill it, but we need to use it up before allocating   */
+    /* a new block ourselves.                                           */
+      opp = &(GC_obj_kinds[k].ok_freelist[lg]);
+      if ( (op = *opp) != 0 ) {
+        *opp = 0;
+        my_bytes_allocd = 0;
+        for (p = op; p != 0; p = obj_link(p)) {
+          my_bytes_allocd += lb;
+          if ((word)my_bytes_allocd >= HBLKSIZE) {
+            *opp = obj_link(p);
+            obj_link(p) = 0;
+            break;
+          }
+        }
+        GC_bytes_allocd += my_bytes_allocd;
+        goto out;
+      }
+    /* Next try to allocate a new block worth of objects of this size.  */
+    {
+        struct hblk *h = GC_allochblk(lb, k, 0);
+        if (h != 0) {
+          if (IS_UNCOLLECTABLE(k)) GC_set_hdr_marks(HDR(h));
+          GC_bytes_allocd += HBLKSIZE - HBLKSIZE % lb;
+#         ifdef PARALLEL_MARK
+            if (GC_parallel) {
+              GC_acquire_mark_lock();
+              ++ GC_fl_builder_count;
+              UNLOCK();
+              GC_release_mark_lock();
+
+              op = GC_build_fl(h, lw,
+                        (ok -> ok_init || GC_debugging_started), 0);
+
+              *result = op;
+              GC_acquire_mark_lock();
+              -- GC_fl_builder_count;
+              if (GC_fl_builder_count == 0) GC_notify_all_builder();
+              GC_release_mark_lock();
+              (void) GC_clear_stack(0);
+              return;
+            }
+#         endif
+          op = GC_build_fl(h, lw, (ok -> ok_init || GC_debugging_started), 0);
+          goto out;
+        }
+    }
+
+    /* As a last attempt, try allocating a single object.  Note that    */
+    /* this may trigger a collection or expand the heap.                */
+      op = GC_generic_malloc_inner(lb, k);
+      if (0 != op) obj_link(op) = 0;
+
+  out:
+    *result = op;
+    UNLOCK();
+    (void) GC_clear_stack(0);
+}
+
+/* Note that the "atomic" version of this would be unsafe, since the    */
+/* links would not be seen by the collector.                            */
+GC_API void * GC_CALL GC_malloc_many(size_t lb)
+{
+    void *result;
+    GC_generic_malloc_many((lb + EXTRA_BYTES + GRANULE_BYTES-1)
+                           & ~(GRANULE_BYTES-1),
+                           NORMAL, &result);
+    return result;
+}
+
+/* Not well tested nor integrated.      */
+/* Debug version is tricky and currently missing.       */
+#include <limits.h>
+
+GC_API void * GC_CALL GC_memalign(size_t align, size_t lb)
+{
+    size_t new_lb;
+    size_t offset;
+    ptr_t result;
+
+    if (align <= GRANULE_BYTES) return GC_malloc(lb);
+    if (align >= HBLKSIZE/2 || lb >= HBLKSIZE/2) {
+        if (align > HBLKSIZE) {
+          return (*GC_get_oom_fn())(LONG_MAX-1024); /* Fail */
+        }
+        return GC_malloc(lb <= HBLKSIZE? HBLKSIZE : lb);
+            /* Will be HBLKSIZE aligned.        */
+    }
+    /* We could also try to make sure that the real rounded-up object size */
+    /* is a multiple of align.  That would be correct up to HBLKSIZE.      */
+    new_lb = lb + align - 1;
+    result = GC_malloc(new_lb);
+    offset = (word)result % align;
+    if (offset != 0) {
+        offset = align - offset;
+        if (!GC_all_interior_pointers) {
+            if (offset >= VALID_OFFSET_SZ) return GC_malloc(HBLKSIZE);
+            GC_register_displacement(offset);
+        }
+    }
+    result = (void *) ((ptr_t)result + offset);
+    GC_ASSERT((word)result % align == 0);
+    return result;
+}
+
+/* This one exists largerly to redirect posix_memalign for leaks finding. */
+GC_API int GC_CALL GC_posix_memalign(void **memptr, size_t align, size_t lb)
+{
+  /* Check alignment properly.  */
+  if (((align - 1) & align) != 0 || align < sizeof(void *)) {
+#   ifdef MSWINCE
+      return ERROR_INVALID_PARAMETER;
+#   else
+      return EINVAL;
+#   endif
+  }
+
+  if ((*memptr = GC_memalign(align, lb)) == NULL) {
+#   ifdef MSWINCE
+      return ERROR_NOT_ENOUGH_MEMORY;
+#   else
+      return ENOMEM;
+#   endif
+  }
+  return 0;
+}
+
+#ifdef ATOMIC_UNCOLLECTABLE
+  /* Allocate lb bytes of pointerfree, untraced, uncollectable data     */
+  /* This is normally roughly equivalent to the system malloc.          */
+  /* But it may be useful if malloc is redefined.                       */
+  GC_API void * GC_CALL GC_malloc_atomic_uncollectable(size_t lb)
+  {
+    void *op;
+    void **opp;
+    size_t lg;
+    DCL_LOCK_STATE;
+
+    if( SMALL_OBJ(lb) ) {
+        if (EXTRA_BYTES != 0 && lb != 0) lb--;
+                  /* We don't need the extra byte, since this won't be  */
+                  /* collected anyway.                                  */
+        lg = GC_size_map[lb];
+        opp = &(GC_auobjfreelist[lg]);
+        LOCK();
+        if( (op = *opp) != 0 ) {
+            *opp = obj_link(op);
+            obj_link(op) = 0;
+            GC_bytes_allocd += GRANULES_TO_BYTES(lg);
+            /* Mark bit was already set while object was on free list. */
+            GC_non_gc_bytes += GRANULES_TO_BYTES(lg);
+            UNLOCK();
+        } else {
+            UNLOCK();
+            op = (ptr_t)GC_generic_malloc(lb, AUNCOLLECTABLE);
+        }
+        GC_ASSERT(0 == op || GC_is_marked(op));
+        return((void *) op);
+    } else {
+        hdr * hhdr;
+
+        op = (ptr_t)GC_generic_malloc(lb, AUNCOLLECTABLE);
+        if (0 == op) return(0);
+
+        GC_ASSERT(((word)op & (HBLKSIZE - 1)) == 0);
+        hhdr = HDR(op);
+
+        LOCK();
+        set_mark_bit_from_hdr(hhdr, 0); /* Only object. */
+#       ifndef THREADS
+          GC_ASSERT(hhdr -> hb_n_marks == 0);
+#       endif
+        hhdr -> hb_n_marks = 1;
+        UNLOCK();
+        return((void *) op);
+    }
+  }
+#endif /* ATOMIC_UNCOLLECTABLE */
+
+/* provide a version of strdup() that uses the collector to allocate the
+   copy of the string */
+GC_API char * GC_CALL GC_strdup(const char *s)
+{
+  char *copy;
+  size_t lb;
+  if (s == NULL) return NULL;
+  lb = strlen(s) + 1;
+  if ((copy = GC_malloc_atomic(lb)) == NULL) {
+#   ifndef MSWINCE
+      errno = ENOMEM;
+#   endif
+    return NULL;
+  }
+# ifndef MSWINCE
+    strcpy(copy, s);
+# else
+    /* strcpy() is deprecated in WinCE */
+    memcpy(copy, s, lb);
+# endif
+  return copy;
+}
+
+GC_API char * GC_CALL GC_strndup(const char *str, size_t size)
+{
+  char *copy;
+  size_t len = strlen(str); /* str is expected to be non-NULL  */
+  if (len > size)
+    len = size;
+  copy = GC_malloc_atomic(len + 1);
+  if (copy == NULL) {
+#   ifndef MSWINCE
+      errno = ENOMEM;
+#   endif
+    return NULL;
+  }
+  BCOPY(str, copy, len);
+  copy[len] = '\0';
+  return copy;
+}
+
+#ifdef GC_REQUIRE_WCSDUP
+# include <wchar.h> /* for wcslen() */
+
+  GC_API wchar_t * GC_CALL GC_wcsdup(const wchar_t *str)
+  {
+    size_t lb = (wcslen(str) + 1) * sizeof(wchar_t);
+    wchar_t *copy = GC_malloc_atomic(lb);
+    if (copy == NULL) {
+#     ifndef MSWINCE
+        errno = ENOMEM;
+#     endif
+      return NULL;
+    }
+    BCOPY(str, copy, lb);
+    return copy;
+  }
+#endif /* GC_REQUIRE_WCSDUP */