/* Portable data dumper for XEmacs. Copyright (C) 1999-2000,2004 Olivier Galibert Copyright (C) 2001 Martin Buchholz Copyright (C) 2001, 2002, 2003, 2004, 2005, 2010 Ben Wing. This file is part of XEmacs. XEmacs is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version. XEmacs is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with XEmacs. If not, see . */ /* Synched up with: Not in FSF. */ /* This file has been Mule-ized, Ben Wing, 10-10-04. */ /* #### Put in much more assertions. Whenever we store fixups in the process or writing out data, make sure the fixups (offsets) point to the beginning of an object, i.e. are registered. Same whenever we read in -- verify offsets as registered, and when compute a fixup, verify the pointer is pointing within the pdump area. registered and check within pdump area. For specific types of pointers (e.g. to Lisp_Objects), check if they're pointing to the right kinds of types. It should be possible to check that a putative Lisp_Object is really a Lisp_Object since it will follow a strict format in its header. */ #include #include "lisp.h" #include "specifier.h" #include "file-coding.h" #include "elhash.h" #include "lstream.h" #include "sysfile.h" #include "console-stream.h" #ifdef WIN32_NATIVE #include "syswindows.h" #else #ifdef HAVE_MMAP #include #endif #ifdef DUMP_IN_EXEC #include "dump-data.h" #endif #endif typedef struct { const void *blockaddr; Bytecount size; const struct memory_description *desc; } pdump_root_block; typedef struct { Dynarr_declare (pdump_root_block); } pdump_root_block_dynarr; typedef struct { void **ptraddress; const struct sized_memory_description *desc; } pdump_root_block_ptr; typedef struct { Dynarr_declare (pdump_root_block_ptr); } pdump_root_block_ptr_dynarr; typedef struct { const void *object; void *data; Bytecount size; EMACS_INT offset; EMACS_INT dest_offset; EMACS_INT save_offset; const struct opaque_convert_functions *fcts; } pdump_cv_data_info; typedef struct { Dynarr_declare (pdump_cv_data_info); } pdump_cv_data_info_dynarr; typedef struct { EMACS_INT dest_offset; EMACS_INT save_offset; Bytecount size; } pdump_cv_data_dump_info; typedef struct { const void *object; void *data; Bytecount size; EMACS_INT index; EMACS_INT save_offset; const struct opaque_convert_functions *fcts; } pdump_cv_ptr_info; typedef struct { Dynarr_declare (pdump_cv_ptr_info); } pdump_cv_ptr_info_dynarr; typedef struct { EMACS_INT save_offset; Bytecount size; } pdump_cv_ptr_dump_info; typedef struct { EMACS_INT save_offset; Bytecount size; void *adr; } pdump_cv_ptr_load_info; typedef struct { Lisp_Object *address; Lisp_Object value; } pdump_static_Lisp_Object; typedef struct { Rawbyte **address; /* Rawbyte * for ease of doing relocation */ Rawbyte * value; } pdump_static_pointer; static pdump_root_block_dynarr *pdump_root_blocks; static pdump_root_block_ptr_dynarr *pdump_root_block_ptrs; static Lisp_Object_ptr_dynarr *pdump_root_lisp_objects; static Lisp_Object_ptr_dynarr *pdump_weak_object_chains; static pdump_cv_data_info_dynarr *pdump_cv_data; static pdump_cv_ptr_info_dynarr *pdump_cv_ptr; /* Mark SIZE bytes at non-heap address BLOCKADDR for dumping, described by DESC. Called by outside callers during XEmacs initialization. */ void dump_add_root_block (const void *blockaddr, Bytecount size, const struct memory_description *desc) { pdump_root_block info; info.blockaddr = blockaddr; info.size = size; info.desc = desc; if (pdump_root_blocks == NULL) pdump_root_blocks = Dynarr_new (pdump_root_block); Dynarr_add (pdump_root_blocks, info); } /* Mark the block described by DESC and pointed to by the pointer at non-heap address PTRADDRESS for dumping. All the objects reachable from this pointer will also be dumped. Called by outside callers during XEmacs initialization. */ void dump_add_root_block_ptr (void *ptraddress, const struct sized_memory_description *desc) { pdump_root_block_ptr info; info.ptraddress = (void **) ptraddress; info.desc = desc; if (pdump_root_block_ptrs == NULL) pdump_root_block_ptrs = Dynarr_new (pdump_root_block_ptr); Dynarr_add (pdump_root_block_ptrs, info); } /* Mark the Lisp_Object at non-heap address VARADDRESS for dumping. All the objects reachable from this var will also be dumped. Called by outside callers during XEmacs initialization. */ void dump_add_root_lisp_object (Lisp_Object *varaddress) { if (pdump_root_lisp_objects == NULL) pdump_root_lisp_objects = Dynarr_new2 (Lisp_Object_ptr_dynarr, Lisp_Object *); Dynarr_add (pdump_root_lisp_objects, varaddress); } /* Mark the list pointed to by the Lisp_Object at VARADDRESS for dumping. Called by outside callers during XEmacs initialization. */ void dump_add_weak_object_chain (Lisp_Object *varaddress) { if (pdump_weak_object_chains == NULL) pdump_weak_object_chains = Dynarr_new2 (Lisp_Object_ptr_dynarr, Lisp_Object *); Dynarr_add (pdump_weak_object_chains, varaddress); } inline static void pdump_align_stream (FILE *stream, Bytecount alignment) { OFF_T offset = FTELL (stream); OFF_T adjustment = ALIGN_SIZE (offset, alignment) - offset; if (adjustment) FSEEK (stream, adjustment, SEEK_CUR); } #define PDUMP_ALIGN_OUTPUT(type) pdump_align_stream (pdump_out, ALIGNOF (type)) #define PDUMP_WRITE(type, object) \ retry_fwrite (&object, sizeof (object), 1, pdump_out); #define PDUMP_WRITE_ALIGNED(type, object) do { \ PDUMP_ALIGN_OUTPUT (type); \ PDUMP_WRITE (type, object); \ } while (0) #define PDUMP_READ(ptr, type) \ (((type *) (ptr = (Rawbyte *) (((type *) ptr) + 1)))[-1]) #define PDUMP_READ_ALIGNED(ptr, type) \ ((ptr = (Rawbyte *) ALIGN_PTR (ptr, type)), PDUMP_READ (ptr, type)) typedef struct { const struct memory_description *desc; int count; } pdump_reloc_table; static Rawbyte *pdump_rt_list = 0; #ifndef NEW_GC void pdump_objects_unmark (void) { int i; Rawbyte *p = pdump_rt_list; if (p) for (;;) { pdump_reloc_table *rt = (pdump_reloc_table *)p; p += sizeof (pdump_reloc_table); if (rt->desc) { for (i=0; icount; i++) { struct lrecord_header *lh = * (struct lrecord_header **) p; #ifdef ALLOC_TYPE_STATS if (C_READONLY_RECORD_HEADER_P (lh)) tick_lrecord_stats (lh, ALLOC_IN_USE); else { tick_lrecord_stats (lh, MARKED_RECORD_HEADER_P (lh) ? ALLOC_IN_USE : ALLOC_ON_FREE_LIST); UNMARK_RECORD_HEADER (lh); } #else /* not ALLOC_TYPE_STATS */ if (! C_READONLY_RECORD_HEADER_P (lh)) UNMARK_RECORD_HEADER (lh); #endif /* (not) ALLOC_TYPE_STATS */ p += sizeof (EMACS_INT); } } else break; } } #endif /* not NEW_GC */ #ifdef NEW_GC /* The structure of the dump file looks like this: 0 - header - dumped objects stab_offset - mc allocation table (count, size, address) for individual allocation and relocation at load time. - nb_cv_data*struct(dest, adr) for in-object externally represented data - nb_cv_ptr*(adr) for pointed-to externally represented data - relocation table - nb_root_struct_ptrs*struct(void *, adr) for global pointers to structures - nb_root_blocks*struct(void *, size, info) for global objects to restore - root lisp object address/value couples with the count preceding the list */ #else /* not NEW_GC */ /* The structure of the dump file looks like this: 0 - header - dumped objects stab_offset - nb_cv_data*struct(dest, adr) for in-object externally represented data - nb_cv_ptr*(adr) for pointed-to externally represented data - nb_root_block_ptrs*struct(void *, adr) for global pointers to heap blocks - nb_root_blocks*struct(void *, size, info) for global data-segment blocks to restore - relocation table - root lisp object address/value couples with the count preceding the list */ #endif /* not NEW_GC */ #define PDUMP_SIGNATURE "XEmacsDP" #define PDUMP_SIGNATURE_LEN (sizeof (PDUMP_SIGNATURE) - 1) typedef struct { char signature[PDUMP_SIGNATURE_LEN]; unsigned int id; EMACS_UINT stab_offset; EMACS_UINT reloc_address; int nb_root_block_ptrs; int nb_root_blocks; int nb_cv_data; int nb_cv_ptr; } pdump_header; Rawbyte *pdump_start; Rawbyte *pdump_end; static Bytecount pdump_length; static pdump_cv_data_dump_info *pdump_loaded_cv_data; static pdump_cv_ptr_load_info *pdump_loaded_cv_ptr; #ifdef WIN32_NATIVE /* Handle for the dump file */ static HANDLE pdump_hFile = INVALID_HANDLE_VALUE; /* Handle for the file mapping object for the dump file */ static HANDLE pdump_hMap = INVALID_HANDLE_VALUE; #endif static void (*pdump_free) (void); static unsigned char pdump_align_table[] = { 64, 1, 2, 1, 4, 1, 2, 1, 8, 1, 2, 1, 4, 1, 2, 1, 16, 1, 2, 1, 4, 1, 2, 1, 8, 1, 2, 1, 4, 1, 2, 1, 32, 1, 2, 1, 4, 1, 2, 1, 8, 1, 2, 1, 4, 1, 2, 1, 16, 1, 2, 1, 4, 1, 2, 1, 8, 1, 2, 1, 4, 1, 2, 1 }; static inline int pdump_size_to_align (Bytecount size) { return pdump_align_table[size % countof (pdump_align_table)]; } /************************************************************************/ /* Registering memory blocks */ /************************************************************************/ /* "Registering" or recording a heap memory block (which will need to be written out, reloaded and relocated, and to which there may be pointers from other heap blocks or from the data segment) happens both in a list and in a hash table. There is a single hash table covering all registered blocks, but different lists for different kinds of blocks. There is one list for "opaque data" (stuff identified as XD_OPAQUE_DATA_PTR, XD_ASCII_STRING, XD_DOC_STRING), one list for each type of Lisp object, and one list for each different memory descriptor. This lets similar-sized and aligned objects be grouped together when they are written out, to save space. pdump_block_list is a list keeping track of registered memory blocks. pdump_block_list_elt is a single entry through the list, and the list is threaded through the NEXT pointer. The information in this list associated with a particular block of memory is -- address of the beginning -- number of elements at that address -- size of each element -- offset to this block in the dumped data pdump_desc_list is a list keeping track of the various descriptions that we've seen. The primary purpose of this is so that memory blocks can be grouped depending on the particular memory description appropriate for them. The format of the list is different from pdump_block_list -- a single array is used. (#### Dynarr should have been used!!!). The information in this list associated with a description is -- pointer to the description -- a pdump_block_list of blocks using that description Functions for working with lists of memory blocks: -- Add a memory block to a list using pdump_add_block() -- Get a memory block from a pointer to its beginning using pdump_get_block(). This uses the hash table, which lists everything. -- Return the memory-block list (pdump_block_list) associated with a descriptor, using pdump_get_block_list(). If no entry found in the pdump_desc_list, add a new one. */ typedef struct pdump_block_list_elt { struct pdump_block_list_elt *next; const void *obj; Bytecount size; int count; EMACS_INT save_offset; } pdump_block_list_elt; typedef struct { pdump_block_list_elt *first; int align; int count; } pdump_block_list; typedef struct pdump_desc_list_elt { pdump_block_list list; const struct memory_description *desc; } pdump_desc_list_elt; typedef struct { pdump_desc_list_elt *list; int count; int size; } pdump_desc_list; static pdump_block_list *pdump_object_table; static pdump_block_list pdump_opaque_data_list; static pdump_desc_list pdump_desc_table; static int *pdump_alert_undump_object; static unsigned long cur_offset; static Bytecount max_size; static int pdump_fd; static void *pdump_buf; static FILE *pdump_out; #ifdef NEW_GC /* PDUMP_HASHSIZE is a large prime. */ #define PDUMP_HASHSIZE 1000003 /* Nothing special about PDUMP_HASH_MULTIPLIER: arbitrary odd integer smaller than PDUMP_HASHSIZE. */ #define PDUMP_HASH_MULTIPLIER 12347 /* Nothing special about PDUMP_HASH_STEP: arbitrary integer for linear probing. */ #define PDUMP_HASH_STEP 574853 #else /* not NEW_GC */ #define PDUMP_HASHSIZE 200001 #endif /* not NEW_GC */ static pdump_block_list_elt **pdump_hash; #ifndef NEW_GC /* Since most pointers are eight bytes aligned, the >>3 allows for a better hash */ #endif /* not NEW_GC */ static int pdump_make_hash (const void *obj) { #ifdef NEW_GC return ((unsigned long)(obj) * PDUMP_HASH_MULTIPLIER) % PDUMP_HASHSIZE; #else /* not NEW_GC */ return ((unsigned long)(obj)>>3) % PDUMP_HASHSIZE; #endif /* not NEW_GC */ } /* Return the entry for an already-registered memory block at OBJ, or NULL if none. */ static pdump_block_list_elt * pdump_get_block (const void *obj) { int pos = pdump_make_hash (obj); pdump_block_list_elt *e; assert (obj != 0); while ((e = pdump_hash[pos]) != 0) { if (e->obj == obj) return e; pos++; if (pos == PDUMP_HASHSIZE) pos = 0; } return 0; } /* Register a new memory block on Return the entry for an already-registered heap (?) memory block at OBJ, or NULL if none. */ static void pdump_add_block (pdump_block_list *list, const void *obj, Bytecount size, int count) { pdump_block_list_elt *e; int pos = pdump_make_hash (obj); while ((e = pdump_hash[pos]) != 0) { if (e->obj == obj) return; pos++; if (pos == PDUMP_HASHSIZE) pos = 0; } e = xnew (pdump_block_list_elt); e->next = list->first; e->obj = obj; e->size = size; e->count = count; list->first = e; list->count += count; pdump_hash[pos] = e; { int align = pdump_size_to_align (size); if (align < list->align) list->align = align; } } #ifdef NEW_GC typedef struct mc_addr_elt { const void *obj; EMACS_INT addr; } mc_addr_elt; static mc_addr_elt *pdump_mc_hash; /* Return the entry for an already-registered memory block at OBJ, or NULL if none. */ static EMACS_INT pdump_get_mc_addr (const void *obj) { int pos = pdump_make_hash (obj); mc_addr_elt *mc_addr; assert (obj != 0); while (((mc_addr = &pdump_mc_hash[pos]) != 0) && (mc_addr->obj != 0)) { if (mc_addr->obj == obj) return mc_addr->addr; pos += PDUMP_HASH_STEP; if (pos >= PDUMP_HASHSIZE) pos -= PDUMP_HASHSIZE; } /* If this code is reached, an heap address occurred which has not been written to the lookup table before. This is a bug! */ ABORT(); return 0; } /* For indirect address lookups, needed for convertibles: Ptr points to an address within an object. Indirect gives the offset by how many bytes the address of the object has to be adjusted to do a lookup in the mc_addr translation table and get the new location of the data. */ #define pdump_get_indirect_mc_addr(ptr, indirect) \ pdump_get_mc_addr ((void *)((ptr) - indirect)) + indirect static void pdump_put_mc_addr (const void *obj, EMACS_INT addr) { mc_addr_elt *mc_addr; int pos = pdump_make_hash (obj); while (((mc_addr = &pdump_mc_hash[pos]) != 0) && (mc_addr->obj != 0)) { if (mc_addr->obj == obj) return; pos += PDUMP_HASH_STEP; if (pos >= PDUMP_HASHSIZE) pos -= PDUMP_HASHSIZE; } pdump_mc_hash[pos].obj = obj; pdump_mc_hash[pos].addr = addr; } #endif /* NEW_GC */ static pdump_block_list * pdump_get_block_list (const struct memory_description *desc) { int i; for (i=0; iname, backtrace[i].position, backtrace[i].offset); } } } static void pdump_unsupported_dump_type (enum memory_description_type type, int do_backtrace) { stderr_out ("Unsupported dump type : %d\n", type); #ifdef WIN32_NATIVE stderr_out ("Are you compiling with SUPPORT_EDIT_AND_CONTINUE?\n"); stderr_out ("See the PROBLEMS file.\n"); #endif if (do_backtrace) pdump_backtrace (); ABORT (); } static void pdump_bump_depth (void) { int me = pdump_depth++; if (me >= BACKTRACE_MAX) { stderr_out ("Backtrace overflow, loop ?\n"); ABORT (); } backtrace[me].obj = 0; backtrace[me].position = 0; backtrace[me].offset = 0; } static void pdump_register_object (Lisp_Object obj); #ifdef NEW_GC static void pdump_register_object_array (Lisp_Object data, Bytecount size, const struct memory_description *desc, int count); #endif /* NEW_GC */ static void pdump_register_block_contents (const void *data, Bytecount size, const struct memory_description * desc, int count); static void pdump_register_block (const void *data, Bytecount size, const struct memory_description *desc, int count); static void pdump_register_sub (const void *data, const struct memory_description *desc) { int pos; int me = pdump_depth - 1; for (pos = 0; desc[pos].type != XD_END; pos++) { const struct memory_description *desc1 = &desc[pos]; EMACS_INT offset = lispdesc_indirect_count (desc1->offset, desc, data); const void *rdata = (const Rawbyte *) data + offset; backtrace[me].position = pos; backtrace[me].offset = offset; union_switcheroo: /* If the flag says don't dump, then don't dump. */ if ((desc1->flags) & XD_FLAG_NO_PDUMP) continue; switch (desc1->type) { case XD_BYTECOUNT: case XD_ELEMCOUNT: case XD_HASHCODE: case XD_INT: case XD_LONG: case XD_INT_RESET: case XD_LO_LINK: break; case XD_OPAQUE_DATA_PTR: { EMACS_INT count = lispdesc_indirect_count (desc1->data1, desc, data); pdump_add_block (&pdump_opaque_data_list, *(void **)rdata, count, 1); break; } case XD_ASCII_STRING: { const Ascbyte *str = * (const Ascbyte **) rdata; if (str) pdump_add_block (&pdump_opaque_data_list, str, strlen (str) + 1, 1); break; } case XD_DOC_STRING: { const Ascbyte *str = * (const Ascbyte **) rdata; if ((EMACS_INT) str > 0) pdump_add_block (&pdump_opaque_data_list, str, strlen (str) + 1, 1); break; } case XD_LISP_OBJECT: { const Lisp_Object *pobj = (const Lisp_Object *) rdata; assert (desc1->data1 == 0); backtrace[me].offset = (const Rawbyte *) pobj - (const Rawbyte *) data; pdump_register_object (*pobj); break; } case XD_LISP_OBJECT_ARRAY: { int i; EMACS_INT count = lispdesc_indirect_count (desc1->data1, desc, data); for (i = 0; i < count; i++) { const Lisp_Object *pobj = ((const Lisp_Object *) rdata) + i; Lisp_Object dobj = *pobj; backtrace[me].offset = (const Rawbyte *) pobj - (const Rawbyte *) data; pdump_register_object (dobj); } break; } #ifdef NEW_GC case XD_INLINE_LISP_OBJECT_BLOCK_PTR: { EMACS_INT count = lispdesc_indirect_count (desc1->data1, desc, data); const struct sized_memory_description *sdesc = lispdesc_indirect_description (data, desc1->data2.descr); const Lisp_Object *pobj = (const Lisp_Object *) rdata; if (pobj) pdump_register_object_array (*pobj, sdesc->size, sdesc->description, count); break; } #endif /* NEW_GC */ case XD_BLOCK_PTR: { EMACS_INT count = lispdesc_indirect_count (desc1->data1, desc, data); const struct sized_memory_description *sdesc = lispdesc_indirect_description (data, desc1->data2.descr); const Rawbyte *dobj = *(const Rawbyte **)rdata; if (dobj) pdump_register_block (dobj, sdesc->size, sdesc->description, count); break; } case XD_BLOCK_ARRAY: { EMACS_INT count = lispdesc_indirect_count (desc1->data1, desc, data); const struct sized_memory_description *sdesc = lispdesc_indirect_description (data, desc1->data2.descr); pdump_register_block_contents (rdata, sdesc->size, sdesc->description, count); break; } case XD_UNION: case XD_UNION_DYNAMIC_SIZE: desc1 = lispdesc_process_xd_union (desc1, desc, data); if (desc1) goto union_switcheroo; break; case XD_OPAQUE_PTR_CONVERTIBLE: { pdump_cv_ptr_info info; info.object = *(void **)rdata; info.fcts = desc1->data2.funcs; if (!pdump_find_in_cv_ptr_dynarr (info.object)) { info.fcts->convert(info.object, &info.data, &info.size); Dynarr_add (pdump_cv_ptr, info); } break; } case XD_OPAQUE_DATA_CONVERTIBLE: { pdump_cv_data_info info; info.object = data; info.offset = offset; info.fcts = desc1->data2.funcs; info.fcts->convert(rdata, &info.data, &info.size); Dynarr_add (pdump_cv_data, info); break; } default: pdump_unsupported_dump_type (desc1->type, 1); } } } static void pdump_register_object (Lisp_Object obj) { struct lrecord_header *objh; const struct lrecord_implementation *imp; if (!POINTER_TYPE_P (XTYPE (obj))) return; objh = XRECORD_LHEADER (obj); if (!objh) return; if (pdump_get_block (objh)) return; imp = LHEADER_IMPLEMENTATION (objh); if (imp->description #ifdef NEW_GC /* Objects with finalizers cannot be dumped with the new allocator's asynchronous finalization strategy. */ && !imp->finalizer #endif /* not NEW_GC */ && RECORD_DUMPABLE (objh)) { pdump_bump_depth (); backtrace[pdump_depth - 1].obj = objh; pdump_add_block (pdump_object_table + objh->type, objh, detagged_lisp_object_size (objh), 1); pdump_register_sub (objh, imp->description); --pdump_depth; } else { pdump_alert_undump_object[objh->type]++; stderr_out ("Undumpable object type : %s\n", imp->name); pdump_backtrace (); } } #ifdef NEW_GC static void pdump_register_object_array (Lisp_Object obj, Bytecount size, const struct memory_description *desc, int count) { struct lrecord_header *objh; const struct lrecord_implementation *imp; if (!POINTER_TYPE_P (XTYPE (obj))) return; objh = XRECORD_LHEADER (obj); if (!objh) return; if (pdump_get_block (objh)) return; imp = LHEADER_IMPLEMENTATION (objh); if (imp->description && RECORD_DUMPABLE (objh)) { pdump_bump_depth (); backtrace[pdump_depth - 1].obj = objh; pdump_add_block (pdump_object_table + objh->type, objh, lispdesc_block_size_1 (objh, size, desc), count); pdump_register_block_contents (objh, size, desc, count); --pdump_depth; } else { pdump_alert_undump_object[objh->type]++; stderr_out ("Undumpable object type : %s\n", imp->name); pdump_backtrace (); } } #endif /* NEW_GC */ /* Register the referenced objects in the array of COUNT blocks located at DATA; each block is described by SIZE and DESC. "Block" here simply means any block of memory. This does not register the block of memory itself; it may, for example, be an array of structures inlined in another memory block and thus should not be registered. See pdump_register_block(), which does register the memory block. */ static void pdump_register_block_contents (const void *data, Bytecount size, const struct memory_description *desc, int count) { int i; Bytecount elsize; pdump_bump_depth (); elsize = lispdesc_block_size_1 (data, size, desc); for (i = 0; i < count; i++) { pdump_register_sub (((Rawbyte *) data) + elsize * i, desc); } --pdump_depth; } /* Register the array of COUNT blocks located at DATA; each block is described by SDESC. "Block" here simply means any block of memory, which is more accurate and less confusing than terms like `struct' and `object'. A `block' need not actually be a C "struct". It could be a single integer or Lisp_Object, for example, as long as the description is accurate. This is like pdump_register_block_contents() but also registers the memory block itself. */ static void pdump_register_block (const void *data, Bytecount size, const struct memory_description *desc, int count) { if (data && !pdump_get_block (data)) { pdump_add_block (pdump_get_block_list (desc), data, lispdesc_block_size_1 (data, size, desc), count); pdump_register_block_contents (data, size, desc, count); } } /* Store the already-calculated new pointer offsets for all pointers in the COUNT contiguous blocks of memory, each described by DESC and of size SIZE, whose original is located at ORIG_DATA and the modifiable copy at DATA. We examine the description to figure out where the pointers are, and then look up the replacement values using pdump_get_block(). This is done just before writing the modified block of memory to the dump file. The new pointer offsets have been carefully calculated so that the data being pointed gets written at that offset in the dump file. That way, the dump file is a correct memory image except perhaps for a constant that needs to be added to all pointers. (#### In fact, we SHOULD be starting up a dumped XEmacs, seeing where the dumped file gets loaded into memory, and then rewriting the dumped file after relocating all the pointers relative to this memory location. That way, if the file gets loaded again at the same location, which will be common, we don't have to do any relocating, which is both faster at startup and allows the read-only part of the dumped data to be shared read-only between different invocations of XEmacs.) #### Do we distinguish between read-only and writable dumped data? Should we? It's tricky because the dumped data, once loaded again, cannot really be free()d or garbage collected since it's all stored in one contiguous block of data with no malloc() headers, and we don't keep track of the pointers used internally in malloc() and the Lisp allocator to track allocated blocks of memory. */ static void pdump_store_new_pointer_offsets (int count, void *data, const void *orig_data, const struct memory_description *desc, int size) { int pos, i; /* Process each block one by one */ for (i = 0; i < count; i++) { /* CUR points to the beginning of each block in the new data. */ Rawbyte *cur = ((Rawbyte *)data) + i * size; /* Scan each line of the description for relocatable pointers */ for (pos = 0; desc[pos].type != XD_END; pos++) { /* RDATA points to the beginning of each element in the new data. */ const struct memory_description *desc1 = &desc[pos]; /* #### Change ORIG_DATA to DATA. See below. */ void *rdata = cur + lispdesc_indirect_count (desc1->offset, desc, orig_data); union_switcheroo: /* If the flag says don't dump, then don't dump. */ if ((desc1->flags) & XD_FLAG_NO_PDUMP) continue; switch (desc1->type) { case XD_BYTECOUNT: case XD_ELEMCOUNT: case XD_HASHCODE: case XD_INT: case XD_LONG: break; case XD_INT_RESET: { EMACS_INT val = lispdesc_indirect_count (desc1->data1, desc, orig_data); * (int *) rdata = val; break; } #ifdef NEW_GC case XD_INLINE_LISP_OBJECT_BLOCK_PTR: #endif /* NEW_GC */ case XD_OPAQUE_DATA_PTR: case XD_ASCII_STRING: case XD_BLOCK_PTR: { void *ptr = * (void **) rdata; if (ptr) * (EMACS_INT *) rdata = pdump_get_block (ptr)->save_offset; break; } case XD_LO_LINK: { /* As described in lrecord.h, this is a weak link. Thus, we need to link this object not (necessarily) to the object directly pointed to, but to the next referenced object in the chain. None of the intermediate objects will be written out, so we traverse down the chain of objects until we find a referenced one. (The Qnil or Qunbound that ends the chain will always be a referenced object.) */ Lisp_Object obj = * (Lisp_Object *) rdata; pdump_block_list_elt *elt1; /* #### Figure out how to handle indirect offsets here. #### In general, when computing indirect counts, do we really need to use the orig_data pointer? Why not just use the new stuff? No, we don't usually need orig_data. We only need it when fetching pointers out of the data, not integers. This currently occurs only with description maps. We should change the other places to DATA to emphasize this. */ assert (!XD_IS_INDIRECT (desc1->offset)); for (;;) { elt1 = pdump_get_block (XRECORD_LHEADER (obj)); if (elt1) break; obj = * (Lisp_Object *) (desc1->offset + (Rawbyte *) (XRECORD_LHEADER (obj))); } * (EMACS_INT *) rdata = elt1->save_offset; break; } case XD_LISP_OBJECT: { Lisp_Object *pobj = (Lisp_Object *) rdata; assert (desc1->data1 == 0); if (POINTER_TYPE_P (XTYPE (*pobj)) && XRECORD_LHEADER (*pobj)) * (EMACS_INT *) pobj = pdump_get_block (XRECORD_LHEADER (*pobj))->save_offset; break; } case XD_LISP_OBJECT_ARRAY: { EMACS_INT num = lispdesc_indirect_count (desc1->data1, desc, orig_data); int j; for (j = 0; j < num; j++) { Lisp_Object *pobj = ((Lisp_Object *) rdata) + j; if (POINTER_TYPE_P (XTYPE (*pobj)) && XRECORD_LHEADER (*pobj)) * (EMACS_INT *) pobj = pdump_get_block (XRECORD_LHEADER (*pobj))->save_offset; } break; } case XD_DOC_STRING: { EMACS_INT str = *(EMACS_INT *)rdata; if (str > 0) * (EMACS_INT *) rdata = pdump_get_block ((void *)str)->save_offset; break; } case XD_BLOCK_ARRAY: { EMACS_INT num = lispdesc_indirect_count (desc1->data1, desc, orig_data); const struct sized_memory_description *sdesc = lispdesc_indirect_description (orig_data, desc1->data2.descr); pdump_store_new_pointer_offsets (num, rdata, ((Rawbyte *) rdata - (Rawbyte *) data) + (Rawbyte *) orig_data, sdesc->description, lispdesc_block_size (((Rawbyte *) rdata - (Rawbyte *) data) + (Rawbyte *) orig_data, sdesc)); break; } case XD_UNION: case XD_UNION_DYNAMIC_SIZE: desc1 = lispdesc_process_xd_union (desc1, desc, orig_data); if (desc1) goto union_switcheroo; break; case XD_OPAQUE_PTR_CONVERTIBLE: *(EMACS_INT *)rdata = pdump_find_in_cv_ptr_dynarr (*(void **)rdata)->index; break; case XD_OPAQUE_DATA_CONVERTIBLE: /* in-object, nothing to do */ break; default: pdump_unsupported_dump_type (desc1->type, 0); } } } } /* Write out to global file descriptor PDUMP_OUT the element (one or more contiguous blocks of identical size/description) recorded in ELT and described by DESC. The element is first copied to a buffer and then all pointers (this includes Lisp_Objects other than integer/character) are relocated to the (pre-computed) offset in the dump file. */ static void pdump_dump_data (pdump_block_list_elt *elt, const struct memory_description *desc) { Bytecount size = elt->size; int count = elt->count; if (desc) { /* Copy to temporary buffer */ memcpy (pdump_buf, elt->obj, size*count); /* Store new offsets into all pointers in block */ pdump_store_new_pointer_offsets (count, pdump_buf, elt->obj, desc, size); } retry_fwrite (desc ? pdump_buf : elt->obj, size, count, pdump_out); } #ifdef NEW_GC /* To be able to relocate during load time, more information about the dumped objects are needed: The count (for array-like data structures), the size of the object, and the location in the dumped data. */ static void pdump_dump_mc_data (pdump_block_list_elt *elt, const struct memory_description *UNUSED(desc)) { EMACS_INT rdata = pdump_get_block (elt->obj)->save_offset; int j; PDUMP_WRITE_ALIGNED (int, elt->count); PDUMP_WRITE_ALIGNED (Bytecount, elt->size); for (j = 0; j < elt->count; j++) { PDUMP_WRITE_ALIGNED (EMACS_INT, rdata); rdata += elt->size; } } static void pdump_scan_lisp_objects_by_alignment (void (*f) (pdump_block_list_elt *, const struct memory_description *)) { int align; for (align = ALIGNOF (max_align_t); align; align>>=1) { int i; pdump_block_list_elt *elt; for (i=0; inext) { f (elt, lrecord_implementations_table[i]->description); } } } static void pdump_scan_non_lisp_objects_by_alignment (void (*f) (pdump_block_list_elt *, const struct memory_description *)) { int align; for (align = ALIGNOF (max_align_t); align; align>>=1) { int i; pdump_block_list_elt *elt; for (i=0; inext) f (elt, list.desc); } for (elt = pdump_opaque_data_list.first; elt; elt = elt->next) if (pdump_size_to_align (elt->size) == align) f (elt, 0); } } static void pdump_reloc_one_mc (void *data, const struct memory_description *desc) { int pos; for (pos = 0; desc[pos].type != XD_END; pos++) { const struct memory_description *desc1 = &desc[pos]; void *rdata = (Rawbyte *) data + lispdesc_indirect_count (desc1->offset, desc, data); union_switcheroo: /* If the flag says don't dump, then don't dump. */ if ((desc1->flags) & XD_FLAG_NO_PDUMP) continue; switch (desc1->type) { case XD_BYTECOUNT: case XD_ELEMCOUNT: case XD_HASHCODE: case XD_INT: case XD_LONG: case XD_INT_RESET: break; case XD_INLINE_LISP_OBJECT_BLOCK_PTR: case XD_OPAQUE_DATA_PTR: case XD_ASCII_STRING: case XD_BLOCK_PTR: case XD_LO_LINK: { EMACS_INT ptr = *(EMACS_INT *) rdata; if (ptr) *(EMACS_INT *) rdata = pdump_get_mc_addr ((void *) ptr); break; } case XD_LISP_OBJECT: { Lisp_Object *pobj = (Lisp_Object *) rdata; assert (desc1->data1 == 0); if (POINTER_TYPE_P (XTYPE (*pobj)) && ! EQ (*pobj, Qnull_pointer)) *pobj = wrap_pointer_1 ((Rawbyte *) pdump_get_mc_addr (XPNTR (*pobj))); break; } case XD_LISP_OBJECT_ARRAY: { EMACS_INT num = lispdesc_indirect_count (desc1->data1, desc, data); int j; for (j=0; j 0) *(EMACS_INT *) rdata = pdump_get_mc_addr ((void *) str); break; } case XD_BLOCK_ARRAY: { EMACS_INT num = lispdesc_indirect_count (desc1->data1, desc, data); int j; const struct sized_memory_description *sdesc = lispdesc_indirect_description (data, desc1->data2.descr); Bytecount size = lispdesc_block_size (rdata, sdesc); /* Note: We are recursing over data in the block itself */ for (j = 0; j < num; j++) pdump_reloc_one_mc ((Rawbyte *) rdata + j * size, sdesc->description); break; } case XD_UNION: case XD_UNION_DYNAMIC_SIZE: desc1 = lispdesc_process_xd_union (desc1, desc, data); if (desc1) goto union_switcheroo; break; case XD_OPAQUE_PTR_CONVERTIBLE: { pdump_cv_ptr_load_info *p = pdump_loaded_cv_ptr + *(EMACS_INT *)rdata; if (!p->adr) p->adr = desc1->data2.funcs->deconvert(0, pdump_start + p->save_offset, p->size); *(void **)rdata = p->adr; break; } case XD_OPAQUE_DATA_CONVERTIBLE: { EMACS_INT dest_offset = (EMACS_INT) rdata; EMACS_INT indirect = lispdesc_indirect_count (desc1->offset, desc, data); pdump_cv_data_dump_info *p; for(p = pdump_loaded_cv_data; pdump_get_indirect_mc_addr (p->dest_offset, indirect) != dest_offset; p++); desc1->data2.funcs->deconvert(rdata, pdump_start + p->save_offset, p->size); break; } default: pdump_unsupported_dump_type (desc1->type, 0); } } } #else /* not NEW_GC */ /* Relocate a single memory block at DATA, described by DESC, from its assumed load location to its actual one by adding DELTA to all pointers in the block. Does not recursively relocate any other memory blocks pointed to. (We already have a list of all memory blocks in the dump file.) This is used once the dump data has been loaded back in, both for blocks sitting in the dumped data (former heap blocks) and in global data-sgment blocks whose contents have been restored from the dumped data. */ static void pdump_reloc_one (void *data, EMACS_INT delta, const struct memory_description *desc) { int pos; for (pos = 0; desc[pos].type != XD_END; pos++) { const struct memory_description *desc1 = &desc[pos]; void *rdata = (Rawbyte *) data + lispdesc_indirect_count (desc1->offset, desc, data); union_switcheroo: /* If the flag says don't dump, then don't dump. */ if ((desc1->flags) & XD_FLAG_NO_PDUMP) continue; switch (desc1->type) { case XD_BYTECOUNT: case XD_ELEMCOUNT: case XD_HASHCODE: case XD_INT: case XD_LONG: case XD_INT_RESET: break; case XD_OPAQUE_DATA_PTR: case XD_ASCII_STRING: case XD_BLOCK_PTR: case XD_LO_LINK: { EMACS_INT ptr = *(EMACS_INT *)rdata; if (ptr) *(EMACS_INT *)rdata = ptr+delta; break; } case XD_LISP_OBJECT: { Lisp_Object *pobj = (Lisp_Object *) rdata; assert (desc1->data1 == 0); if (POINTER_TYPE_P (XTYPE (*pobj)) && ! EQ (*pobj, Qnull_pointer)) *pobj = wrap_pointer_1 ((Rawbyte *) XPNTR (*pobj) + delta); break; } case XD_LISP_OBJECT_ARRAY: { EMACS_INT num = lispdesc_indirect_count (desc1->data1, desc, data); int j; for (j=0; j 0) *(EMACS_INT *)rdata = str + delta; break; } case XD_BLOCK_ARRAY: { EMACS_INT num = lispdesc_indirect_count (desc1->data1, desc, data); int j; const struct sized_memory_description *sdesc = lispdesc_indirect_description (data, desc1->data2.descr); Bytecount size = lispdesc_block_size (rdata, sdesc); /* Note: We are recursing over data in the block itself */ for (j = 0; j < num; j++) pdump_reloc_one ((Rawbyte *) rdata + j * size, delta, sdesc->description); break; } case XD_UNION: case XD_UNION_DYNAMIC_SIZE: desc1 = lispdesc_process_xd_union (desc1, desc, data); if (desc1) goto union_switcheroo; break; case XD_OPAQUE_PTR_CONVERTIBLE: { pdump_cv_ptr_load_info *p = pdump_loaded_cv_ptr + *(EMACS_INT *)rdata; if (!p->adr) p->adr = desc1->data2.funcs->deconvert(0, pdump_start + p->save_offset, p->size); *(void **)rdata = p->adr; break; } case XD_OPAQUE_DATA_CONVERTIBLE: { EMACS_INT dest_offset = (Rawbyte *)rdata - pdump_start; pdump_cv_data_dump_info *p; for(p = pdump_loaded_cv_data; p->dest_offset != dest_offset; p++); desc1->data2.funcs->deconvert(rdata, pdump_start + p->save_offset, p->size); break; } default: pdump_unsupported_dump_type (desc1->type, 0); } } } #endif /* not NEW_GC */ static void pdump_allocate_offset (pdump_block_list_elt *elt, const struct memory_description *UNUSED (desc)) { Bytecount size = elt->count * elt->size; elt->save_offset = cur_offset; if (size > max_size) max_size = size; cur_offset += size; } /* Write out to global file descriptor PDUMP_OUT the result of an external element. It's just opaque data. */ static void pdump_dump_cv_data (pdump_cv_data_info *elt) { retry_fwrite (elt->data, elt->size, 1, pdump_out); } static void pdump_dump_cv_ptr (pdump_cv_ptr_info *elt) { retry_fwrite (elt->data, elt->size, 1, pdump_out); } static void pdump_allocate_offset_cv_data (pdump_cv_data_info *elt) { elt->save_offset = cur_offset; if (elt->size>max_size) max_size = elt->size; cur_offset += elt->size; } static void pdump_allocate_offset_cv_ptr (pdump_cv_ptr_info *elt) { elt->save_offset = cur_offset; if (elt->size>max_size) max_size = elt->size; cur_offset += elt->size; } /* Traverse through all the heap blocks, once the "register" stage of dumping has finished. To compress space as much as possible, we logically sort all blocks by alignment, hitting all blocks with alignment == the maximum (which may be 8 bytes, for doubles), then all blocks with the next lower alignment (4 bytes), etc. Within each alignment we hit -- first the Lisp objects, type-by-type -- then the heap memory blocks that are not Lisp objects, description-by- description -- i.e. all blocks with the same description will be placed together -- then the "opaque" data objects declared as XD_OPAQUE_DATA_PTR, XD_ASCII_STRING and XD_DOC_STRING. The idea is to have as little blank space as possible in the laid-out data. For each item that we have hit, we process it by calling F, the function passed it. In dumper.c, pdump_scan_by_alignment() is called twice with two different functions -- pdump_allocate_offset() in stage 2 to compute the offset to each block, and pdump_dump_data() in stage 3 to successively write each block to disk. It's extremely important that the SAME traversal order gets invoked in both stage 2 and 3. */ static void pdump_scan_by_alignment (void (*f)(pdump_block_list_elt *, const struct memory_description *), void (*g)(pdump_cv_data_info *), void (*h)(pdump_cv_ptr_info *)) { int align; for (align = ALIGNOF (max_align_t); align; align>>=1) { int i; pdump_block_list_elt *elt; for (i=0; inext) f (elt, lrecord_implementations_table[i]->description); for (i=0; inext) f (elt, list.desc); } for (elt = pdump_opaque_data_list.first; elt; elt = elt->next) if (pdump_size_to_align (elt->size) == align) f (elt, 0); for (i=0; i < Dynarr_length (pdump_cv_data); i++) if (pdump_size_to_align (Dynarr_atp (pdump_cv_data, i)->size) == align) g (Dynarr_atp (pdump_cv_data, i)); for (i=0; i < Dynarr_length (pdump_cv_ptr); i++) if (pdump_size_to_align (Dynarr_atp (pdump_cv_ptr, i)->size) == align) h (Dynarr_atp (pdump_cv_ptr, i)); } } static void pdump_dump_cv_data_info (void) { int i; Elemcount count = Dynarr_length (pdump_cv_data); pdump_cv_data_dump_info *data = alloca_array (pdump_cv_data_dump_info, count); for (i = 0; i < count; i++) { data[i].dest_offset = Dynarr_at (pdump_cv_data, i).dest_offset; data[i].save_offset = Dynarr_at (pdump_cv_data, i).save_offset; data[i].size = Dynarr_at (pdump_cv_data, i).size; } PDUMP_ALIGN_OUTPUT (pdump_cv_data_dump_info); retry_fwrite (data, sizeof (pdump_cv_data_dump_info), count, pdump_out); } static void pdump_dump_cv_ptr_info (void) { int i; Elemcount count = Dynarr_length (pdump_cv_ptr); pdump_cv_ptr_dump_info *data = alloca_array (pdump_cv_ptr_dump_info, count); for (i = 0; i < count; i++) { data[i].save_offset = Dynarr_at (pdump_cv_ptr, i).save_offset; data[i].size = Dynarr_at (pdump_cv_ptr, i).size; } PDUMP_ALIGN_OUTPUT (pdump_cv_ptr_dump_info); retry_fwrite (data, sizeof (pdump_cv_ptr_dump_info), count, pdump_out); } /* Dump out the root block pointers, part of stage 3 (the "WRITE" stage) of dumping. For each pointer we dump out a structure containing the location of the pointer and its value, replaced by the appropriate offset into the dumped data. */ static void pdump_dump_root_block_ptrs (void) { int i; Elemcount count = Dynarr_length (pdump_root_block_ptrs); pdump_static_pointer *data = alloca_array (pdump_static_pointer, count); for (i = 0; i < count; i++) { data[i].address = (Rawbyte **) Dynarr_atp (pdump_root_block_ptrs, i)->ptraddress; data[i].value = (Rawbyte *) pdump_get_block (* data[i].address)->save_offset; } PDUMP_ALIGN_OUTPUT (pdump_static_pointer); retry_fwrite (data, sizeof (pdump_static_pointer), count, pdump_out); } /* Dump out the root blocks, part of stage 3 (the "WRITE" stage) of dumping. For each block we dump a structure containing info about the block (its location, size and description) and then the block itself, with its pointers replaced with offsets into the dump data. */ static void pdump_dump_root_blocks (void) { int i; for (i = 0; i < Dynarr_length (pdump_root_blocks); i++) { pdump_root_block info = Dynarr_at (pdump_root_blocks, i); PDUMP_WRITE_ALIGNED (pdump_root_block, info); if (info.desc) { /* Copy to temporary buffer */ memcpy (pdump_buf, info.blockaddr, info.size); /* Store new offsets into all pointers in block */ pdump_store_new_pointer_offsets (1, pdump_buf, info.blockaddr, info.desc, info.size); } retry_fwrite (info.desc ? pdump_buf : info.blockaddr, info.size, 1, pdump_out); } } static void pdump_dump_rtables (void) { int i; pdump_block_list_elt *elt; pdump_reloc_table rt; for (i=0; idescription; rt.count = pdump_object_table[i].count; PDUMP_WRITE_ALIGNED (pdump_reloc_table, rt); while (elt) { EMACS_INT rdata = pdump_get_block (elt->obj)->save_offset; #ifdef NEW_GC int j; for (j=0; jcount; j++) { PDUMP_WRITE_ALIGNED (EMACS_INT, rdata); rdata += elt->size; } #else /* not NEW_GC */ PDUMP_WRITE_ALIGNED (EMACS_INT, rdata); #endif /* not NEW_GC */ elt = elt->next; } } rt.desc = 0; rt.count = 0; PDUMP_WRITE_ALIGNED (pdump_reloc_table, rt); for (i=0; iobj)->save_offset; int j; for (j=0; jcount; j++) { PDUMP_WRITE_ALIGNED (EMACS_INT, rdata); rdata += elt->size; } elt = elt->next; } } rt.desc = 0; rt.count = 0; PDUMP_WRITE_ALIGNED (pdump_reloc_table, rt); } static void pdump_dump_root_lisp_objects (void) { Elemcount count = (Dynarr_length (pdump_root_lisp_objects) + Dynarr_length (pdump_weak_object_chains)); Elemcount i; PDUMP_WRITE_ALIGNED (Elemcount, count); PDUMP_ALIGN_OUTPUT (pdump_static_Lisp_Object); for (i = 0; i < Dynarr_length (pdump_root_lisp_objects); i++) { pdump_static_Lisp_Object obj; obj.address = Dynarr_at (pdump_root_lisp_objects, i); obj.value = * obj.address; if (POINTER_TYPE_P (XTYPE (obj.value))) obj.value = wrap_pointer_1 ((void *) pdump_get_block (XRECORD_LHEADER (obj.value))->save_offset); PDUMP_WRITE (pdump_static_Lisp_Object, obj); } for (i = 0; i < Dynarr_length (pdump_weak_object_chains); i++) { pdump_block_list_elt *elt; pdump_static_Lisp_Object obj; obj.address = Dynarr_at (pdump_weak_object_chains, i); obj.value = * obj.address; for (;;) { const struct memory_description *desc; int pos; elt = pdump_get_block (XRECORD_LHEADER (obj.value)); if (elt) break; desc = XRECORD_LHEADER_IMPLEMENTATION (obj.value)->description; for (pos = 0; desc[pos].type != XD_LO_LINK; pos++) assert (desc[pos].type != XD_END); /* #### Figure out how to handle indirect offsets here. */ assert (!XD_IS_INDIRECT (desc[pos].offset)); obj.value = * (Lisp_Object *) (desc[pos].offset + (Rawbyte *) (XRECORD_LHEADER (obj.value))); } obj.value = wrap_pointer_1 ((void *) elt->save_offset); PDUMP_WRITE (pdump_static_Lisp_Object, obj); } } /*######################################################################## # Pdump # ######################################################################## [ben] DISCUSSION OF DUMPING: The idea of dumping is to record the state of XEmacs in a file, so that it can be reloaded later. This avoids having to reload all of the basic Lisp code each time XEmacs is run, which is a rather time-consuming process. (Less so on new machines, but still noticeable. As an example of a program with similar issues but which does not have a dumping process and as a result has a slow startup time, consider Adobe Photoshop 5.0 or Adobe Photoshop Elements 2.0.) We don't actually record ALL the state of XEmacs (some of it, for example, is dependent on the run-time environment and needs to be initialized whenever XEmacs is run), but whatever state we don't record needs to be reinitialized every time XEmacs is run. The old way of dumping was to make a new executable file with the data segment expanded to contain the heap and written out from memory. This is what the unex* files do. Unfortunately this process is extremely system-specific and breaks easily with OS changes. Another simple, more portable trick, the "static heap" method, involves replacing the allocator with our own allocator which allocates all space out of a very large array declared in our data segment until we run out, then uses the underlying malloc() to start allocating on the heap. If we ensure that the large array is big enough to hold all data allocated during the dump stage, then all of the data we need to save is in the data segment, and it's easy to calculate the location and size of the data segment we want to save (we don't want to record and reinitialize the data segment of library functions) by using appropriately declared variables in the first and last file linked. This method is known as the "static heap" method, and is used by the non-pdump version of the dumper under Cygwin, and was also used under VMS and in Win-Emacs. The "static heap" method works well in practice. Nonetheless, a more complex method of dumping was written by Olivier Galibert, which requires that structural descriptions of all data allocated in the heap be provided and the roots of all pointers into the heap be noted through function calls to the pdump API. This way, all the heap data can be traversed and written out to a file, and then reloaded at run-time and the pointers relocated to point at the new location of the loaded data. This is the "pdump" method used in this file. There are two potential advantages of "pdump" over the "static heap": (1) It doesn't require any tricks to calculate the beginning and end of the data segment, or even that the XEmacs section of the data segment be contiguous. (It's not clear whether this is an issue in practice.) (2) Potentially, it could handle an OS that does not always load the static data segment at a predictable location. The "static heap" method by its nature needs the data segment to stay in the same place from invocation to invocation, since it simply dumps out memory and reloads it, without any pointer relocation. I say "potentially" because as it is currently written pdump does assume that the data segment is never relocated. However, changing pdump to remove this assumption is probably not difficult, as all the mechanism to handle pointer relocation is already present. DISCUSSION OF PDUMP WORKINGS: See man/internals/internals.texi for more information. NOTE that we have two kinds of memory to handle: memory on the heap (i.e. allocated through malloc()) or the like, and static memory in the data segment of the program, i.e. stuff declared as global or static. All heap memory needs to be written out to the dump file and reproduced (i.e. reloaded and any necessary relocations performed). Data-segment memory that is not statically initialized (i.e. through declarations in the C code) needs either to be written out and reloaded, or reinitialized. In addition, any pointers in data-segment memory to heap memory must be written out, reloaded and relocated. NOTE that we currently don't handle relocation of pointers into data- segment memory. (See overview discussion above.) These are treated in the descriptions as opaque data not needing relocation. If this becomes a problem, it can be fixed through new kinds of types in enum memory_description_type. Three basic steps to dumping out: (1) "REGISTER": Starting with all sources of relocatable memory (currently this means all data-segment pointers to heap memory -- see above about pointers to data-segment memory), recursively traverse the tree of pointers and "register" (make a note of) every memory block seen. (2) "LAYOUT": Go through all of the registered blocks and compute the location of each one in the dump data (i.e. the "offset" that will be added to the address corresponding to start of the loaded-in data to get the new pointer referring to this block). The blocks will be laid out sequentially according to the order we traverse them. Also note the maximum-sized block for use in step 3. (3) "WRITE": After writing some header stuff, go through all of the registered blocks and write out each one to the dump file. Note that we are simply writing out the blocks sequentially as we see them, and our traversal path is identical to that in step 2, so blocks will end up at the locations computed for them. In order to write out a block, first copy it to a temporary location (hence the maximum-block-size computation in the previous step), then for each relocatable pointer in the block, write in its place the offset to the heap block in the dump data. When the dump data is loaded, the address of the beginning of the dump data will be added to the offset in each pointer, and thence become accurate. --ben */ void pdump (void) { int i; Lisp_Object t_console, t_device, t_frame; int none; pdump_header header; in_pdump = 1; pdump_object_table = xnew_array (pdump_block_list, lrecord_type_count); pdump_alert_undump_object = xnew_array (int, lrecord_type_count); assert (ALIGNOF (max_align_t) <= pdump_align_table[0]); for (i = 0; i < countof (pdump_align_table); i++) if (pdump_align_table[i] > ALIGNOF (max_align_t)) pdump_align_table[i] = ALIGNOF (max_align_t); flush_all_buffer_local_cache (); /* These appear in a DEFVAR_LISP, which does a staticpro() */ t_console = Vterminal_console; Vterminal_console = Qnil; t_frame = Vterminal_frame; Vterminal_frame = Qnil; t_device = Vterminal_device; Vterminal_device = Qnil; dump_add_opaque (&lrecord_implementations_table, lrecord_type_count * sizeof (lrecord_implementations_table[0])); #ifdef USE_KKCC dump_add_opaque (&lrecord_memory_descriptions, lrecord_type_count * sizeof (lrecord_memory_descriptions[0])); #else /* not USE_KKCC */ dump_add_opaque (&lrecord_markers, lrecord_type_count * sizeof (lrecord_markers[0])); #endif /* not USE_KKCC */ pdump_hash = xnew_array_and_zero (pdump_block_list_elt *, PDUMP_HASHSIZE); for (i = 0; iname, pdump_alert_undump_object[i]); } if (!none) { in_pdump = 0; return; } /* (2) Register out the data-segment pointer variables to heap blocks */ for (i = 0; i < Dynarr_length (pdump_root_block_ptrs); i++) { pdump_root_block_ptr info = Dynarr_at (pdump_root_block_ptrs, i); pdump_register_block (*(info.ptraddress), info.desc->size, info.desc->description, 1); } /* (3) Register out the data-segment blocks, maybe with pointers to heap blocks */ for (i = 0; i < Dynarr_length (pdump_root_blocks); i++) { pdump_root_block *info = Dynarr_atp (pdump_root_blocks, i); if (info->desc) { /* Size may have been given as 0 meaning "compute later". Compute now and update. If no DESC, size must always be correct as there is no other way of computing it. */ info->size = lispdesc_block_size_1 (info->blockaddr, info->size, info->desc); pdump_register_block_contents (info->blockaddr, info->size, info->desc, 1); } } /* (II) The "layout" stage: Compute the offsets and max-size */ /* (1) Determine header size */ memcpy (header.signature, PDUMP_SIGNATURE, PDUMP_SIGNATURE_LEN); header.id = dump_id; header.reloc_address = 0; header.nb_root_block_ptrs = Dynarr_length (pdump_root_block_ptrs); header.nb_root_blocks = Dynarr_length (pdump_root_blocks); header.nb_cv_data = Dynarr_length (pdump_cv_data); header.nb_cv_ptr = Dynarr_length (pdump_cv_ptr); cur_offset = MAX_ALIGN_SIZE (sizeof (header)); max_size = 0; /* (2) Traverse all heap blocks and compute their offsets; keep track of maximum block size seen */ pdump_scan_by_alignment (pdump_allocate_offset, pdump_allocate_offset_cv_data, pdump_allocate_offset_cv_ptr); cur_offset = MAX_ALIGN_SIZE (cur_offset); header.stab_offset = cur_offset; /* (3) Update maximum size based on root (data-segment) blocks */ for (i = 0; i < Dynarr_length (pdump_root_blocks); i++) { pdump_root_block info = Dynarr_at (pdump_root_blocks, i); /* If no DESC, no relocation needed and we copy directly instead of into a temp buffer. */ if (info.desc) { if (info.size > max_size) max_size = info.size; } } /* (III) The "write "stage: Dump out the data, storing the offsets in place of pointers whenever we write out memory blocks */ pdump_buf = xmalloc (max_size); /* EMACS_PROGNAME is entirely ASCII so this should be Mule-safe */ pdump_fd = open (EMACS_PROGNAME ".dmp", O_WRONLY | O_CREAT | O_TRUNC | OPEN_BINARY, 0666); if (pdump_fd < 0) report_file_error ("Unable to open dump file", build_ascstring (EMACS_PROGNAME ".dmp")); pdump_out = fdopen (pdump_fd, "w"); if (pdump_out < 0) report_file_error ("Unable to open dump file for writing", build_ascstring (EMACS_PROGNAME ".dmp")); retry_fwrite (&header, sizeof (header), 1, pdump_out); PDUMP_ALIGN_OUTPUT (max_align_t); for (i = 0; i < Dynarr_length (pdump_cv_data); i++) { pdump_cv_data_info *elt = Dynarr_atp (pdump_cv_data, i); elt->dest_offset = pdump_get_block (elt->object)->save_offset + elt->offset; } for (i = 0; i < Dynarr_length (pdump_cv_ptr); i++) Dynarr_at (pdump_cv_ptr, i).index = i; pdump_scan_by_alignment (pdump_dump_data, pdump_dump_cv_data, pdump_dump_cv_ptr); for (i = 0; i < Dynarr_length (pdump_cv_data); i++) { pdump_cv_data_info *elt = Dynarr_atp (pdump_cv_data, i); if(elt->fcts->convert_free) elt->fcts->convert_free(elt->object, elt->data, elt->size); } for (i = 0; i < Dynarr_length (pdump_cv_ptr); i++) { pdump_cv_ptr_info *elt = Dynarr_atp (pdump_cv_ptr, i); if(elt->fcts->convert_free) elt->fcts->convert_free(elt->object, elt->data, elt->size); } FSEEK (pdump_out, header.stab_offset, SEEK_SET); #ifdef NEW_GC { EMACS_INT zero = 0; pdump_scan_lisp_objects_by_alignment (pdump_dump_mc_data); PDUMP_WRITE_ALIGNED (EMACS_INT, zero); pdump_scan_non_lisp_objects_by_alignment (pdump_dump_mc_data); PDUMP_WRITE_ALIGNED (EMACS_INT, zero); } #endif /* NEW_GC */ pdump_dump_cv_data_info (); pdump_dump_cv_ptr_info (); #ifdef NEW_GC pdump_dump_rtables (); #endif /* NEW_GC */ pdump_dump_root_block_ptrs (); pdump_dump_root_blocks (); #ifndef NEW_GC pdump_dump_rtables (); #endif /* not NEW_GC */ pdump_dump_root_lisp_objects (); retry_fclose (pdump_out); /* pdump_fd is already closed by the preceding call to fclose. retry_close (pdump_fd); */ free (pdump_buf); free (pdump_hash); Vterminal_console = t_console; Vterminal_frame = t_frame; Vterminal_device = t_device; in_pdump = 0; } static int pdump_load_check (void) { return (!memcmp (((pdump_header *) pdump_start)->signature, PDUMP_SIGNATURE, PDUMP_SIGNATURE_LEN) && ((pdump_header *)pdump_start)->id == dump_id); } /*----------------------------------------------------------------------*/ /* Reading the dump file */ /*----------------------------------------------------------------------*/ static int pdump_load_finish (void) { int i; Rawbyte *p; EMACS_INT delta; EMACS_INT count; pdump_header *header = (pdump_header *) pdump_start; #ifdef NEW_GC /* This is a DEFVAR_BOOL and gets dumped, but the actual value was already determined by vdb_install_signal_handler () in vdb-mprotect.c, which could be different from the value in the dump file. So store it here and restore it after loading the dump file. */ int allow_inc_gc = allow_incremental_gc; #endif /* NEW_GC */ pdump_end = pdump_start + pdump_length; delta = ((EMACS_INT) pdump_start) - header->reloc_address; p = pdump_start + header->stab_offset; #ifdef NEW_GC pdump_mc_hash = xnew_array_and_zero (mc_addr_elt, PDUMP_HASHSIZE); /* Allocate space for each object individually. First the Lisp_Objects, then the blocks. */ count = 2; for (;;) { EMACS_INT elt_count = PDUMP_READ_ALIGNED (p, EMACS_INT); if (elt_count) { Rawbyte *mc_addr = 0; Bytecount size = PDUMP_READ_ALIGNED (p, Bytecount); for (i = 0; i < elt_count; i++) { EMACS_INT rdata = PDUMP_READ_ALIGNED (p, EMACS_INT); if (i == 0) { Bytecount real_size = size * elt_count; if (count == 2) { if (elt_count <= 1) mc_addr = (Rawbyte *) mc_alloc (real_size); else mc_addr = (Rawbyte *) mc_alloc_array (size, elt_count); #ifdef ALLOC_TYPE_STATS inc_lrecord_stats (real_size, (const struct lrecord_header *) ((Rawbyte *) rdata + delta)); #endif /* ALLOC_TYPE_STATS */ } else mc_addr = (Rawbyte *) xmalloc_and_zero (real_size); } else mc_addr += size; pdump_put_mc_addr ((void *) rdata, (EMACS_INT) mc_addr); memcpy (mc_addr, (Rawbyte *) rdata + delta, size); } } else if (!(--count)) break; } #endif /* NEW_GC */ /* Get the cv_data array */ p = (Rawbyte *) ALIGN_PTR (p, pdump_cv_data_dump_info); pdump_loaded_cv_data = (pdump_cv_data_dump_info *)p; p += header->nb_cv_data*sizeof(pdump_cv_data_dump_info); /* Build the cv_ptr array */ p = (Rawbyte *) ALIGN_PTR (p, pdump_cv_ptr_dump_info); pdump_loaded_cv_ptr = alloca_array (pdump_cv_ptr_load_info, header->nb_cv_ptr); for (i = 0; i < header->nb_cv_ptr; i++) { pdump_cv_ptr_dump_info info = PDUMP_READ (p, pdump_cv_ptr_dump_info); pdump_loaded_cv_ptr[i].save_offset = info.save_offset; pdump_loaded_cv_ptr[i].size = info.size; pdump_loaded_cv_ptr[i].adr = 0; } #ifdef NEW_GC /* Relocate the heap objects */ pdump_rt_list = p; count = 2; for (;;) { pdump_reloc_table rt = PDUMP_READ_ALIGNED (p, pdump_reloc_table); p = (Rawbyte *) ALIGN_PTR (p, Rawbyte *); if (rt.desc) { Rawbyte **reloc = (Rawbyte **) p; for (i = 0; i < rt.count; i++) { reloc[i] = (Rawbyte *) pdump_get_mc_addr (reloc[i]); pdump_reloc_one_mc (reloc[i], rt.desc); } p += rt.count * sizeof (Rawbyte *); } else if (!(--count)) break; } #endif /* NEW_GC */ /* Put back the pdump_root_block_ptrs */ p = (Rawbyte *) ALIGN_PTR (p, pdump_static_pointer); for (i = 0; i < header->nb_root_block_ptrs; i++) { pdump_static_pointer ptr = PDUMP_READ (p, pdump_static_pointer); #ifdef NEW_GC (* ptr.address) = (Rawbyte *) pdump_get_mc_addr (ptr.value); #else /* not NEW_GC */ (* ptr.address) = ptr.value + delta; #endif /* not NEW_GC */ } /* Put back the pdump_root_blocks and relocate */ for (i = 0; i < header->nb_root_blocks; i++) { pdump_root_block info = PDUMP_READ_ALIGNED (p, pdump_root_block); memcpy ((void *) info.blockaddr, p, info.size); if (info.desc) #ifdef NEW_GC pdump_reloc_one_mc ((void *) info.blockaddr, info.desc); #else /* not NEW_GC */ pdump_reloc_one ((void *) info.blockaddr, delta, info.desc); #endif /* not NEW_GC */ p += info.size; } #ifndef NEW_GC /* Relocate the heap objects */ pdump_rt_list = p; count = 2; for (;;) { pdump_reloc_table rt = PDUMP_READ_ALIGNED (p, pdump_reloc_table); p = (Rawbyte *) ALIGN_PTR (p, Rawbyte *); if (rt.desc) { Rawbyte **reloc = (Rawbyte **) p; for (i = 0; i < rt.count; i++) { reloc[i] += delta; pdump_reloc_one (reloc[i], delta, rt.desc); } p += rt.count * sizeof (Rawbyte *); } else if (!(--count)) break; } #endif /* not NEW_GC */ /* Put the pdump_root_lisp_objects variables in place */ i = PDUMP_READ_ALIGNED (p, Elemcount); p = (Rawbyte *) ALIGN_PTR (p, pdump_static_Lisp_Object); while (i--) { pdump_static_Lisp_Object obj = PDUMP_READ (p, pdump_static_Lisp_Object); if (POINTER_TYPE_P (XTYPE (obj.value))) #ifdef NEW_GC obj.value = wrap_pointer_1 ((Rawbyte *) pdump_get_mc_addr (XPNTR (obj.value))); #else /* not NEW_GC */ obj.value = wrap_pointer_1 ((Rawbyte *) XPNTR (obj.value) + delta); #endif /* not NEW_GC */ (* obj.address) = obj.value; } /* Final cleanups */ /* reorganize hash tables */ p = pdump_rt_list; for (;;) { pdump_reloc_table rt = PDUMP_READ_ALIGNED (p, pdump_reloc_table); p = (Rawbyte *) ALIGN_PTR (p, Lisp_Object); if (!rt.desc) break; if (rt.desc == hash_table_description) { for (i = 0; i < rt.count; i++) pdump_reorganize_hash_table (PDUMP_READ (p, Lisp_Object)); break; } else p += sizeof (Lisp_Object) * rt.count; } #ifdef NEW_GC xfree (pdump_mc_hash); #endif /* NEW_GC */ #ifdef NEW_GC allow_incremental_gc = allow_inc_gc; #endif /* NEW_GC */ return 1; } #ifdef WIN32_NATIVE /* Free the mapped file if we decide we don't want it after all */ static void pdump_file_unmap (void) { UnmapViewOfFile (pdump_start); CloseHandle (pdump_hFile); CloseHandle (pdump_hMap); } static int pdump_file_get (const Wexttext *wpath) { Extbyte *path; if (XEUNICODE_P) path = (Extbyte *) wpath; else path = WEXTTEXT_TO_MULTIBYTE (wpath); pdump_hFile = qxeCreateFile (path, GENERIC_READ + GENERIC_WRITE, /* Required for copy on write */ 0, /* Not shared */ NULL, /* Not inheritable */ OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, NULL); /* No template file */ if (pdump_hFile == INVALID_HANDLE_VALUE) return 0; pdump_length = GetFileSize (pdump_hFile, NULL); pdump_hMap = qxeCreateFileMapping (pdump_hFile, NULL, /* No security attributes */ PAGE_WRITECOPY, /* Copy on write */ 0, /* Max size, high half */ 0, /* Max size, low half */ NULL); /* Unnamed */ if (pdump_hMap == INVALID_HANDLE_VALUE) return 0; pdump_start = (Rawbyte *) MapViewOfFile (pdump_hMap, FILE_MAP_COPY, /* Copy on write */ 0, /* Start at zero */ 0, 0); /* Map all of it */ pdump_free = pdump_file_unmap; return 1; } /* pdump_resource_free is called (via the pdump_free pointer) to release any resources allocated by pdump_resource_get. Since the Windows API specs specifically state that you don't need to (and shouldn't) free the resources allocated by FindResource, LoadResource, and LockResource this routine does nothing. */ static void pdump_resource_free (void) { } static int pdump_resource_get (void) { HRSRC hRes; /* Handle to dump resource */ HRSRC hResLoad; /* Handle to loaded dump resource */ /* See Q126630 which describes how Windows NT and 95 trap writes to resource sections and duplicate the page to allow the write to proceed. It also describes how to make the resource section read/write (and hence private to each process). Doing this avoids the exceptions and related overhead, but causes the resource section to be private to each process that is running XEmacs. Since the resource section contains little other than the dumped data, which should be private to each process, we make the whole resource section read/write so we don't have to copy it. */ hRes = FindResourceA (NULL, MAKEINTRESOURCE (101), "DUMP"); if (hRes == NULL) return 0; /* Found it, use the data in the resource */ hResLoad = (HRSRC) LoadResource (NULL, hRes); if (hResLoad == NULL) return 0; pdump_start = (Rawbyte *) LockResource (hResLoad); if (pdump_start == NULL) return 0; pdump_free = pdump_resource_free; pdump_length = SizeofResource (NULL, hRes); if (pdump_length <= (Bytecount) sizeof (pdump_header)) { pdump_start = 0; return 0; } return 1; } #else /* !WIN32_NATIVE */ static void pdump_file_free (void) { xfree (pdump_start); } #ifdef HAVE_MMAP static void pdump_file_unmap (void) { munmap (pdump_start, pdump_length); } #endif static int pdump_file_get (const Wexttext *path) { int fd = wext_retry_open (path, O_RDONLY | OPEN_BINARY); if (fd < 0) return 0; pdump_length = lseek (fd, 0, SEEK_END); if (pdump_length < (Bytecount) sizeof (pdump_header)) { retry_close (fd); return 0; } lseek (fd, 0, SEEK_SET); #ifdef HAVE_MMAP /* Unix 98 requires that sys/mman.h define MAP_FAILED, but many earlier implementations don't. */ # ifndef MAP_FAILED # define MAP_FAILED ((void *) -1L) # endif pdump_start = (Rawbyte *) mmap (0, pdump_length, PROT_READ|PROT_WRITE, MAP_PRIVATE, fd, 0); if (pdump_start != (Rawbyte *) MAP_FAILED) { pdump_free = pdump_file_unmap; retry_close (fd); return 1; } #endif /* HAVE_MMAP */ pdump_start = xnew_array (Rawbyte, pdump_length); pdump_free = pdump_file_free; retry_read (fd, pdump_start, pdump_length); retry_close (fd); return 1; } #ifdef DUMP_IN_EXEC static int pdump_ram_try (void) { pdump_start = dumped_data_get (); pdump_length = dumped_data_size (); return pdump_load_check (); } #endif #endif /* !WIN32_NATIVE */ /* This used to be static, but there seems to be a bug in the GCC 4.1.2 optimizer that clobbers exe_path. */ int pdump_file_try (Wexttext*); int pdump_file_try (Wexttext *exe_path) { Wexttext *w = exe_path + wext_strlen (exe_path); /* We look for various names, including those with the version and dump ID, those with just the dump ID, and those without either. We first try adding directly to the executable name, then lopping off any extension (e.g. .exe) or version name in the executable (xemacs-21.5.18). */ do { wext_sprintf (w, WEXTSTRING ("-%s-%08x.dmp"), WEXTSTRING (EMACS_VERSION), dump_id); if (pdump_file_get (exe_path)) { if (pdump_load_check ()) return 1; pdump_free (); } wext_sprintf (w, WEXTSTRING ("-%08x.dmp"), dump_id); if (pdump_file_get (exe_path)) { if (pdump_load_check ()) return 1; pdump_free (); } wext_sprintf (w, WEXTSTRING (".dmp")); if (pdump_file_get (exe_path)) { if (pdump_load_check ()) return 1; pdump_free (); } do w--; /* !!#### See comment below about how this is unsafe. */ while (w > exe_path && !IS_DIRECTORY_SEP (*w) && (*w != '-') && (*w != '.')); } while (w > exe_path && !IS_DIRECTORY_SEP (*w)); return 0; } #define DUMP_SLACK 100 /* Enough to include dump ID, version name, .DMP */ int pdump_load (const Wexttext *argv0) { #ifdef WIN32_NATIVE Wexttext *exe_path = NULL; int bufsize = 4096; int cchpathsize; /* Copied from mswindows_get_module_file_name (). Not clear if it's kosher to malloc() yet. */ while (1) { exe_path = alloca_array (Wexttext, bufsize); cchpathsize = qxeGetModuleFileName (NULL, (Extbyte *) exe_path, bufsize); if (!cchpathsize) goto fail; if (cchpathsize + DUMP_SLACK <= bufsize) break; bufsize *= 2; } if (!XEUNICODE_P) { Wexttext *wexe = MULTIBYTE_TO_WEXTTEXT ((Extbyte *) exe_path); wext_strcpy (exe_path, wexe); } #else /* !WIN32_NATIVE */ Wexttext *exe_path; Wexttext *w; const Wexttext *dir, *p; #ifdef DUMP_IN_EXEC if (pdump_ram_try ()) { pdump_load_finish (); in_pdump = 0; return 1; } #endif in_pdump = 1; dir = argv0; if (dir[0] == '-') { /* XEmacs as a login shell, oh goody! */ dir = wext_getenv ("SHELL"); /* not egetenv -- not yet initialized and we want external-format data */ } p = dir + wext_strlen (dir); /* !!#### This is bad as it may fail with certain non-ASCII-compatible external formats such as JIS. Maybe we should be using the mb*() routines in libc? But can we reliably trust them on all Unix platforms? (We can't convert to internal since those conversion routines aren't yet initialized) */ while (p != dir && !IS_ANY_SEP (p[-1])) p--; if (p != dir) { /* invocation-name includes a directory component -- presumably it is relative to cwd, not $PATH. */ exe_path = alloca_array (Wexttext, 1 + wext_strlen (dir) + DUMP_SLACK); wext_strcpy (exe_path, dir); } else { const Wexttext *path = wext_getenv ("PATH"); /* not egetenv -- not yet init. */ const Wexttext *name = p; exe_path = alloca_array (Wexttext, 1 + DUMP_SLACK + max (wext_strlen (name), wext_strlen (path))); for (;;) { p = path; while (*p && *p != SEPCHAR) p++; if (p == path) { exe_path[0] = '.'; w = exe_path + 1; } else { memcpy (exe_path, path, (p - path) * sizeof (Wexttext)); w = exe_path + (p - path); } if (!IS_DIRECTORY_SEP (w[-1])) *w++ = '/'; wext_strcpy (w, name); { struct stat statbuf; if (wext_access (exe_path, X_OK) == 0 && wext_stat (exe_path, &statbuf) == 0 && ! S_ISDIR (statbuf.st_mode)) break; } if (!*p) { /* Oh well, let's have some kind of default */ wext_sprintf (exe_path, "./%s", name); break; } path = p + 1; } } #endif /* WIN32_NATIVE */ if (pdump_file_try (exe_path)) { pdump_load_finish (); in_pdump = 0; #ifdef NEW_GC pdump_free (); #endif /* NEW_GC */ return 1; } #ifdef WIN32_NATIVE if (pdump_resource_get ()) { if (pdump_load_check ()) { pdump_load_finish (); in_pdump = 0; #ifdef NEW_GC pdump_free (); #endif /* NEW_GC */ return 1; } pdump_free (); } fail: #endif in_pdump = 0; return 0; }