+ return TRUE;
+}
+
+/* Called via elf_link_hash_traverse to allocate stubs for any _SPUEAR_
+ symbols. */
+
+static bfd_boolean
+allocate_spuear_stubs (struct elf_link_hash_entry *h, void *inf)
+{
+ /* Symbols starting with _SPUEAR_ need a stub because they may be
+ invoked by the PPU. */
+ struct bfd_link_info *info = inf;
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+ asection *sym_sec;
+
+ if ((h->root.type == bfd_link_hash_defined
+ || h->root.type == bfd_link_hash_defweak)
+ && h->def_regular
+ && strncmp (h->root.root.string, "_SPUEAR_", 8) == 0
+ && (sym_sec = h->root.u.def.section) != NULL
+ && sym_sec->output_section != bfd_abs_section_ptr
+ && spu_elf_section_data (sym_sec->output_section) != NULL
+ && (spu_elf_section_data (sym_sec->output_section)->u.o.ovl_index != 0
+ || htab->params->non_overlay_stubs))
+ {
+ return count_stub (htab, NULL, NULL, nonovl_stub, h, NULL);
+ }
+
+ return TRUE;
+}
+
+static bfd_boolean
+build_spuear_stubs (struct elf_link_hash_entry *h, void *inf)
+{
+ /* Symbols starting with _SPUEAR_ need a stub because they may be
+ invoked by the PPU. */
+ struct bfd_link_info *info = inf;
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+ asection *sym_sec;
+
+ if ((h->root.type == bfd_link_hash_defined
+ || h->root.type == bfd_link_hash_defweak)
+ && h->def_regular
+ && strncmp (h->root.root.string, "_SPUEAR_", 8) == 0
+ && (sym_sec = h->root.u.def.section) != NULL
+ && sym_sec->output_section != bfd_abs_section_ptr
+ && spu_elf_section_data (sym_sec->output_section) != NULL
+ && (spu_elf_section_data (sym_sec->output_section)->u.o.ovl_index != 0
+ || htab->params->non_overlay_stubs))
+ {
+ return build_stub (info, NULL, NULL, nonovl_stub, h, NULL,
+ h->root.u.def.value, sym_sec);
+ }
+
+ return TRUE;
+}
+
+/* Size or build stubs. */
+
+static bfd_boolean
+process_stubs (struct bfd_link_info *info, bfd_boolean build)
+{
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+ bfd *ibfd;
+
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ {
+ extern const bfd_target spu_elf32_vec;
+ Elf_Internal_Shdr *symtab_hdr;
+ asection *isec;
+ Elf_Internal_Sym *local_syms = NULL;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ /* We'll need the symbol table in a second. */
+ symtab_hdr = &elf_tdata (ibfd)->symtab_hdr;
+ if (symtab_hdr->sh_info == 0)
+ continue;
+
+ /* Walk over each section attached to the input bfd. */
+ for (isec = ibfd->sections; isec != NULL; isec = isec->next)
+ {
+ Elf_Internal_Rela *internal_relocs, *irelaend, *irela;
+
+ /* If there aren't any relocs, then there's nothing more to do. */
+ if ((isec->flags & SEC_RELOC) == 0
+ || isec->reloc_count == 0)
+ continue;
+
+ if (!maybe_needs_stubs (isec))
+ continue;
+
+ /* Get the relocs. */
+ internal_relocs = _bfd_elf_link_read_relocs (ibfd, isec, NULL, NULL,
+ info->keep_memory);
+ if (internal_relocs == NULL)
+ goto error_ret_free_local;
+
+ /* Now examine each relocation. */
+ irela = internal_relocs;
+ irelaend = irela + isec->reloc_count;
+ for (; irela < irelaend; irela++)
+ {
+ enum elf_spu_reloc_type r_type;
+ unsigned int r_indx;
+ asection *sym_sec;
+ Elf_Internal_Sym *sym;
+ struct elf_link_hash_entry *h;
+ enum _stub_type stub_type;
+
+ r_type = ELF32_R_TYPE (irela->r_info);
+ r_indx = ELF32_R_SYM (irela->r_info);
+
+ if (r_type >= R_SPU_max)
+ {
+ bfd_set_error (bfd_error_bad_value);
+ error_ret_free_internal:
+ if (elf_section_data (isec)->relocs != internal_relocs)
+ free (internal_relocs);
+ error_ret_free_local:
+ if (local_syms != NULL
+ && (symtab_hdr->contents
+ != (unsigned char *) local_syms))
+ free (local_syms);
+ return FALSE;
+ }
+
+ /* Determine the reloc target section. */
+ if (!get_sym_h (&h, &sym, &sym_sec, &local_syms, r_indx, ibfd))
+ goto error_ret_free_internal;
+
+ stub_type = needs_ovl_stub (h, sym, sym_sec, isec, irela,
+ NULL, info);
+ if (stub_type == no_stub)
+ continue;
+ else if (stub_type == stub_error)
+ goto error_ret_free_internal;
+
+ if (htab->stub_count == NULL)
+ {
+ bfd_size_type amt;
+ amt = (htab->num_overlays + 1) * sizeof (*htab->stub_count);
+ htab->stub_count = bfd_zmalloc (amt);
+ if (htab->stub_count == NULL)
+ goto error_ret_free_internal;
+ }
+
+ if (!build)
+ {
+ if (!count_stub (htab, ibfd, isec, stub_type, h, irela))
+ goto error_ret_free_internal;
+ }
+ else
+ {
+ bfd_vma dest;
+
+ if (h != NULL)
+ dest = h->root.u.def.value;
+ else
+ dest = sym->st_value;
+ dest += irela->r_addend;
+ if (!build_stub (info, ibfd, isec, stub_type, h, irela,
+ dest, sym_sec))
+ goto error_ret_free_internal;
+ }
+ }
+
+ /* We're done with the internal relocs, free them. */
+ if (elf_section_data (isec)->relocs != internal_relocs)
+ free (internal_relocs);
+ }
+
+ if (local_syms != NULL
+ && symtab_hdr->contents != (unsigned char *) local_syms)
+ {
+ if (!info->keep_memory)
+ free (local_syms);
+ else
+ symtab_hdr->contents = (unsigned char *) local_syms;
+ }
+ }
+
+ return TRUE;
+}
+
+/* Allocate space for overlay call and return stubs.
+ Return 0 on error, 1 if no overlays, 2 otherwise. */
+
+int
+spu_elf_size_stubs (struct bfd_link_info *info)
+{
+ struct spu_link_hash_table *htab;
+ bfd *ibfd;
+ bfd_size_type amt;
+ flagword flags;
+ unsigned int i;
+ asection *stub;
+
+ if (!process_stubs (info, FALSE))
+ return 0;
+
+ htab = spu_hash_table (info);
+ elf_link_hash_traverse (&htab->elf, allocate_spuear_stubs, info);
+ if (htab->stub_err)
+ return 0;
+
+ ibfd = info->input_bfds;
+ if (htab->stub_count != NULL)
+ {
+ amt = (htab->num_overlays + 1) * sizeof (*htab->stub_sec);
+ htab->stub_sec = bfd_zmalloc (amt);
+ if (htab->stub_sec == NULL)
+ return 0;
+
+ flags = (SEC_ALLOC | SEC_LOAD | SEC_CODE | SEC_READONLY
+ | SEC_HAS_CONTENTS | SEC_IN_MEMORY);
+ stub = bfd_make_section_anyway_with_flags (ibfd, ".stub", flags);
+ htab->stub_sec[0] = stub;
+ if (stub == NULL
+ || !bfd_set_section_alignment (ibfd, stub,
+ ovl_stub_size_log2 (htab->params)))
+ return 0;
+ stub->size = htab->stub_count[0] * ovl_stub_size (htab->params);
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ /* Extra space for linked list entries. */
+ stub->size += htab->stub_count[0] * 16;
+
+ for (i = 0; i < htab->num_overlays; ++i)
+ {
+ asection *osec = htab->ovl_sec[i];
+ unsigned int ovl = spu_elf_section_data (osec)->u.o.ovl_index;
+ stub = bfd_make_section_anyway_with_flags (ibfd, ".stub", flags);
+ htab->stub_sec[ovl] = stub;
+ if (stub == NULL
+ || !bfd_set_section_alignment (ibfd, stub,
+ ovl_stub_size_log2 (htab->params)))
+ return 0;
+ stub->size = htab->stub_count[ovl] * ovl_stub_size (htab->params);
+ }
+ }
+
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ {
+ /* Space for icache manager tables.
+ a) Tag array, one quadword per cache line.
+ b) Rewrite "to" list, one quadword per cache line.
+ c) Rewrite "from" list, one byte per outgoing branch (rounded up to
+ a power-of-two number of full quadwords) per cache line. */
+
+ flags = SEC_ALLOC;
+ htab->ovtab = bfd_make_section_anyway_with_flags (ibfd, ".ovtab", flags);
+ if (htab->ovtab == NULL
+ || !bfd_set_section_alignment (ibfd, htab->ovtab, 4))
+ return 0;
+
+ htab->ovtab->size = (16 + 16 + (16 << htab->fromelem_size_log2))
+ << htab->num_lines_log2;
+
+ flags = SEC_ALLOC | SEC_LOAD | SEC_HAS_CONTENTS | SEC_IN_MEMORY;
+ htab->init = bfd_make_section_anyway_with_flags (ibfd, ".ovini", flags);
+ if (htab->init == NULL
+ || !bfd_set_section_alignment (ibfd, htab->init, 4))
+ return 0;
+
+ htab->init->size = 16;
+ }
+ else if (htab->stub_count == NULL)
+ return 1;
+ else
+ {
+ /* htab->ovtab consists of two arrays.
+ . struct {
+ . u32 vma;
+ . u32 size;
+ . u32 file_off;
+ . u32 buf;
+ . } _ovly_table[];
+ .
+ . struct {
+ . u32 mapped;
+ . } _ovly_buf_table[];
+ . */
+
+ flags = SEC_ALLOC | SEC_LOAD | SEC_HAS_CONTENTS | SEC_IN_MEMORY;
+ htab->ovtab = bfd_make_section_anyway_with_flags (ibfd, ".ovtab", flags);
+ if (htab->ovtab == NULL
+ || !bfd_set_section_alignment (ibfd, htab->ovtab, 4))
+ return 0;
+
+ htab->ovtab->size = htab->num_overlays * 16 + 16 + htab->num_buf * 4;
+ }
+
+ htab->toe = bfd_make_section_anyway_with_flags (ibfd, ".toe", SEC_ALLOC);
+ if (htab->toe == NULL
+ || !bfd_set_section_alignment (ibfd, htab->toe, 4))
+ return 0;
+ htab->toe->size = 16;
+
+ return 2;
+}
+
+/* Called from ld to place overlay manager data sections. This is done
+ after the overlay manager itself is loaded, mainly so that the
+ linker's htab->init section is placed after any other .ovl.init
+ sections. */
+
+void
+spu_elf_place_overlay_data (struct bfd_link_info *info)
+{
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+ unsigned int i;
+
+ if (htab->stub_sec != NULL)
+ {
+ (*htab->params->place_spu_section) (htab->stub_sec[0], NULL, ".text");
+
+ for (i = 0; i < htab->num_overlays; ++i)
+ {
+ asection *osec = htab->ovl_sec[i];
+ unsigned int ovl = spu_elf_section_data (osec)->u.o.ovl_index;
+ (*htab->params->place_spu_section) (htab->stub_sec[ovl], osec, NULL);
+ }
+ }
+
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ (*htab->params->place_spu_section) (htab->init, NULL, ".ovl.init");
+
+ if (htab->ovtab != NULL)
+ {
+ const char *ovout = ".data";
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ ovout = ".bss";
+ (*htab->params->place_spu_section) (htab->ovtab, NULL, ovout);
+ }
+
+ if (htab->toe != NULL)
+ (*htab->params->place_spu_section) (htab->toe, NULL, ".toe");
+}
+
+/* Functions to handle embedded spu_ovl.o object. */
+
+static void *
+ovl_mgr_open (struct bfd *nbfd ATTRIBUTE_UNUSED, void *stream)
+{
+ return stream;
+}
+
+static file_ptr
+ovl_mgr_pread (struct bfd *abfd ATTRIBUTE_UNUSED,
+ void *stream,
+ void *buf,
+ file_ptr nbytes,
+ file_ptr offset)
+{
+ struct _ovl_stream *os;
+ size_t count;
+ size_t max;
+
+ os = (struct _ovl_stream *) stream;
+ max = (const char *) os->end - (const char *) os->start;
+
+ if ((ufile_ptr) offset >= max)
+ return 0;
+
+ count = nbytes;
+ if (count > max - offset)
+ count = max - offset;
+
+ memcpy (buf, (const char *) os->start + offset, count);
+ return count;
+}
+
+static int
+ovl_mgr_stat (struct bfd *abfd ATTRIBUTE_UNUSED,
+ void *stream,
+ struct stat *sb)
+{
+ struct _ovl_stream *os = (struct _ovl_stream *) stream;
+
+ memset (sb, 0, sizeof (*sb));
+ sb->st_size = (const char *) os->end - (const char *) os->start;
+ return 0;
+}
+
+bfd_boolean
+spu_elf_open_builtin_lib (bfd **ovl_bfd, const struct _ovl_stream *stream)
+{
+ *ovl_bfd = bfd_openr_iovec ("builtin ovl_mgr",
+ "elf32-spu",
+ ovl_mgr_open,
+ (void *) stream,
+ ovl_mgr_pread,
+ NULL,
+ ovl_mgr_stat);
+ return *ovl_bfd != NULL;
+}
+
+static unsigned int
+overlay_index (asection *sec)
+{
+ if (sec == NULL
+ || sec->output_section == bfd_abs_section_ptr)
+ return 0;
+ return spu_elf_section_data (sec->output_section)->u.o.ovl_index;
+}
+
+/* Define an STT_OBJECT symbol. */
+
+static struct elf_link_hash_entry *
+define_ovtab_symbol (struct spu_link_hash_table *htab, const char *name)
+{
+ struct elf_link_hash_entry *h;
+
+ h = elf_link_hash_lookup (&htab->elf, name, TRUE, FALSE, FALSE);
+ if (h == NULL)
+ return NULL;
+
+ if (h->root.type != bfd_link_hash_defined
+ || !h->def_regular)
+ {
+ h->root.type = bfd_link_hash_defined;
+ h->root.u.def.section = htab->ovtab;
+ h->type = STT_OBJECT;
+ h->ref_regular = 1;
+ h->def_regular = 1;
+ h->ref_regular_nonweak = 1;
+ h->non_elf = 0;
+ }
+ else if (h->root.u.def.section->owner != NULL)
+ {
+ (*_bfd_error_handler) (_("%B is not allowed to define %s"),
+ h->root.u.def.section->owner,
+ h->root.root.string);
+ bfd_set_error (bfd_error_bad_value);
+ return NULL;
+ }
+ else
+ {
+ (*_bfd_error_handler) (_("you are not allowed to define %s in a script"),
+ h->root.root.string);
+ bfd_set_error (bfd_error_bad_value);
+ return NULL;
+ }
+
+ return h;
+}
+
+/* Fill in all stubs and the overlay tables. */
+
+static bfd_boolean
+spu_elf_build_stubs (struct bfd_link_info *info)
+{
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+ struct elf_link_hash_entry *h;
+ bfd_byte *p;
+ asection *s;
+ bfd *obfd;
+ unsigned int i;
+
+ if (htab->num_overlays != 0)
+ {
+ for (i = 0; i < 2; i++)
+ {
+ h = htab->ovly_entry[i];
+ if (h != NULL
+ && (h->root.type == bfd_link_hash_defined
+ || h->root.type == bfd_link_hash_defweak)
+ && h->def_regular)
+ {
+ s = h->root.u.def.section->output_section;
+ if (spu_elf_section_data (s)->u.o.ovl_index)
+ {
+ (*_bfd_error_handler) (_("%s in overlay section"),
+ h->root.root.string);
+ bfd_set_error (bfd_error_bad_value);
+ return FALSE;
+ }
+ }
+ }
+ }
+
+ if (htab->stub_sec != NULL)
+ {
+ for (i = 0; i <= htab->num_overlays; i++)
+ if (htab->stub_sec[i]->size != 0)
+ {
+ htab->stub_sec[i]->contents = bfd_zalloc (htab->stub_sec[i]->owner,
+ htab->stub_sec[i]->size);
+ if (htab->stub_sec[i]->contents == NULL)
+ return FALSE;
+ htab->stub_sec[i]->rawsize = htab->stub_sec[i]->size;
+ htab->stub_sec[i]->size = 0;
+ }
+
+ /* Fill in all the stubs. */
+ process_stubs (info, TRUE);
+ if (!htab->stub_err)
+ elf_link_hash_traverse (&htab->elf, build_spuear_stubs, info);
+
+ if (htab->stub_err)
+ {
+ (*_bfd_error_handler) (_("overlay stub relocation overflow"));
+ bfd_set_error (bfd_error_bad_value);
+ return FALSE;
+ }
+
+ for (i = 0; i <= htab->num_overlays; i++)
+ {
+ if (htab->stub_sec[i]->size != htab->stub_sec[i]->rawsize)
+ {
+ (*_bfd_error_handler) (_("stubs don't match calculated size"));
+ bfd_set_error (bfd_error_bad_value);
+ return FALSE;
+ }
+ htab->stub_sec[i]->rawsize = 0;
+ }
+ }
+
+ if (htab->ovtab == NULL || htab->ovtab->size == 0)
+ return TRUE;
+
+ htab->ovtab->contents = bfd_zalloc (htab->ovtab->owner, htab->ovtab->size);
+ if (htab->ovtab->contents == NULL)
+ return FALSE;
+
+ p = htab->ovtab->contents;
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ {
+ bfd_vma off;
+
+ h = define_ovtab_symbol (htab, "__icache_tag_array");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 0;
+ h->size = 16 << htab->num_lines_log2;
+ off = h->size;
+
+ h = define_ovtab_symbol (htab, "__icache_tag_array_size");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 16 << htab->num_lines_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_rewrite_to");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = off;
+ h->size = 16 << htab->num_lines_log2;
+ off += h->size;
+
+ h = define_ovtab_symbol (htab, "__icache_rewrite_to_size");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 16 << htab->num_lines_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_rewrite_from");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = off;
+ h->size = 16 << (htab->fromelem_size_log2 + htab->num_lines_log2);
+ off += h->size;
+
+ h = define_ovtab_symbol (htab, "__icache_rewrite_from_size");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 16 << (htab->fromelem_size_log2
+ + htab->num_lines_log2);
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_log2_fromelemsize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->fromelem_size_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_base");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->ovl_sec[0]->vma;
+ h->root.u.def.section = bfd_abs_section_ptr;
+ h->size = htab->num_buf << htab->line_size_log2;
+
+ h = define_ovtab_symbol (htab, "__icache_linesize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 1 << htab->line_size_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_log2_linesize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->line_size_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_neg_log2_linesize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = -htab->line_size_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_cachesize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 1 << (htab->num_lines_log2 + htab->line_size_log2);
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_log2_cachesize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->num_lines_log2 + htab->line_size_log2;
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ h = define_ovtab_symbol (htab, "__icache_neg_log2_cachesize");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = -(htab->num_lines_log2 + htab->line_size_log2);
+ h->root.u.def.section = bfd_abs_section_ptr;
+
+ if (htab->init != NULL && htab->init->size != 0)
+ {
+ htab->init->contents = bfd_zalloc (htab->init->owner,
+ htab->init->size);
+ if (htab->init->contents == NULL)
+ return FALSE;
+
+ h = define_ovtab_symbol (htab, "__icache_fileoff");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 0;
+ h->root.u.def.section = htab->init;
+ h->size = 8;
+ }
+ }
+ else
+ {
+ /* Write out _ovly_table. */
+ /* set low bit of .size to mark non-overlay area as present. */
+ p[7] = 1;
+ obfd = htab->ovtab->output_section->owner;
+ for (s = obfd->sections; s != NULL; s = s->next)
+ {
+ unsigned int ovl_index = spu_elf_section_data (s)->u.o.ovl_index;
+
+ if (ovl_index != 0)
+ {
+ unsigned long off = ovl_index * 16;
+ unsigned int ovl_buf = spu_elf_section_data (s)->u.o.ovl_buf;
+
+ bfd_put_32 (htab->ovtab->owner, s->vma, p + off);
+ bfd_put_32 (htab->ovtab->owner, (s->size + 15) & -16,
+ p + off + 4);
+ /* file_off written later in spu_elf_modify_program_headers. */
+ bfd_put_32 (htab->ovtab->owner, ovl_buf, p + off + 12);
+ }
+ }
+
+ h = define_ovtab_symbol (htab, "_ovly_table");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = 16;
+ h->size = htab->num_overlays * 16;
+
+ h = define_ovtab_symbol (htab, "_ovly_table_end");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->num_overlays * 16 + 16;
+ h->size = 0;
+
+ h = define_ovtab_symbol (htab, "_ovly_buf_table");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->num_overlays * 16 + 16;
+ h->size = htab->num_buf * 4;
+
+ h = define_ovtab_symbol (htab, "_ovly_buf_table_end");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.value = htab->num_overlays * 16 + 16 + htab->num_buf * 4;
+ h->size = 0;
+ }
+
+ h = define_ovtab_symbol (htab, "_EAR_");
+ if (h == NULL)
+ return FALSE;
+ h->root.u.def.section = htab->toe;
+ h->root.u.def.value = 0;
+ h->size = 16;
+
+ return TRUE;
+}
+
+/* Check that all loadable section VMAs lie in the range
+ LO .. HI inclusive, and stash some parameters for --auto-overlay. */
+
+asection *
+spu_elf_check_vma (struct bfd_link_info *info)
+{
+ struct elf_segment_map *m;
+ unsigned int i;
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+ bfd *abfd = info->output_bfd;
+ bfd_vma hi = htab->params->local_store_hi;
+ bfd_vma lo = htab->params->local_store_lo;
+
+ htab->local_store = hi + 1 - lo;
+
+ for (m = elf_seg_map (abfd); m != NULL; m = m->next)
+ if (m->p_type == PT_LOAD)
+ for (i = 0; i < m->count; i++)
+ if (m->sections[i]->size != 0
+ && (m->sections[i]->vma < lo
+ || m->sections[i]->vma > hi
+ || m->sections[i]->vma + m->sections[i]->size - 1 > hi))
+ return m->sections[i];
+
+ return NULL;
+}
+
+/* OFFSET in SEC (presumably) is the beginning of a function prologue.
+ Search for stack adjusting insns, and return the sp delta.
+ If a store of lr is found save the instruction offset to *LR_STORE.
+ If a stack adjusting instruction is found, save that offset to
+ *SP_ADJUST. */
+
+static int
+find_function_stack_adjust (asection *sec,
+ bfd_vma offset,
+ bfd_vma *lr_store,
+ bfd_vma *sp_adjust)
+{
+ int reg[128];
+
+ memset (reg, 0, sizeof (reg));
+ for ( ; offset + 4 <= sec->size; offset += 4)
+ {
+ unsigned char buf[4];
+ int rt, ra;
+ int imm;
+
+ /* Assume no relocs on stack adjusing insns. */
+ if (!bfd_get_section_contents (sec->owner, sec, buf, offset, 4))
+ break;
+
+ rt = buf[3] & 0x7f;
+ ra = ((buf[2] & 0x3f) << 1) | (buf[3] >> 7);
+
+ if (buf[0] == 0x24 /* stqd */)
+ {
+ if (rt == 0 /* lr */ && ra == 1 /* sp */)
+ *lr_store = offset;
+ continue;
+ }
+
+ /* Partly decoded immediate field. */
+ imm = (buf[1] << 9) | (buf[2] << 1) | (buf[3] >> 7);
+
+ if (buf[0] == 0x1c /* ai */)
+ {
+ imm >>= 7;
+ imm = (imm ^ 0x200) - 0x200;
+ reg[rt] = reg[ra] + imm;
+
+ if (rt == 1 /* sp */)
+ {
+ if (reg[rt] > 0)
+ break;
+ *sp_adjust = offset;
+ return reg[rt];
+ }
+ }
+ else if (buf[0] == 0x18 && (buf[1] & 0xe0) == 0 /* a */)
+ {
+ int rb = ((buf[1] & 0x1f) << 2) | ((buf[2] & 0xc0) >> 6);
+
+ reg[rt] = reg[ra] + reg[rb];
+ if (rt == 1)
+ {
+ if (reg[rt] > 0)
+ break;
+ *sp_adjust = offset;
+ return reg[rt];
+ }
+ }
+ else if (buf[0] == 0x08 && (buf[1] & 0xe0) == 0 /* sf */)
+ {
+ int rb = ((buf[1] & 0x1f) << 2) | ((buf[2] & 0xc0) >> 6);
+
+ reg[rt] = reg[rb] - reg[ra];
+ if (rt == 1)
+ {
+ if (reg[rt] > 0)
+ break;
+ *sp_adjust = offset;
+ return reg[rt];
+ }
+ }
+ else if ((buf[0] & 0xfc) == 0x40 /* il, ilh, ilhu, ila */)
+ {
+ if (buf[0] >= 0x42 /* ila */)
+ imm |= (buf[0] & 1) << 17;
+ else
+ {
+ imm &= 0xffff;
+
+ if (buf[0] == 0x40 /* il */)
+ {
+ if ((buf[1] & 0x80) == 0)
+ continue;
+ imm = (imm ^ 0x8000) - 0x8000;
+ }
+ else if ((buf[1] & 0x80) == 0 /* ilhu */)
+ imm <<= 16;
+ }
+ reg[rt] = imm;
+ continue;
+ }
+ else if (buf[0] == 0x60 && (buf[1] & 0x80) != 0 /* iohl */)
+ {
+ reg[rt] |= imm & 0xffff;
+ continue;
+ }
+ else if (buf[0] == 0x04 /* ori */)
+ {
+ imm >>= 7;
+ imm = (imm ^ 0x200) - 0x200;
+ reg[rt] = reg[ra] | imm;
+ continue;
+ }
+ else if (buf[0] == 0x32 && (buf[1] & 0x80) != 0 /* fsmbi */)
+ {
+ reg[rt] = ( ((imm & 0x8000) ? 0xff000000 : 0)
+ | ((imm & 0x4000) ? 0x00ff0000 : 0)
+ | ((imm & 0x2000) ? 0x0000ff00 : 0)
+ | ((imm & 0x1000) ? 0x000000ff : 0));
+ continue;
+ }
+ else if (buf[0] == 0x16 /* andbi */)
+ {
+ imm >>= 7;
+ imm &= 0xff;
+ imm |= imm << 8;
+ imm |= imm << 16;
+ reg[rt] = reg[ra] & imm;
+ continue;
+ }
+ else if (buf[0] == 0x33 && imm == 1 /* brsl .+4 */)
+ {
+ /* Used in pic reg load. Say rt is trashed. Won't be used
+ in stack adjust, but we need to continue past this branch. */
+ reg[rt] = 0;
+ continue;
+ }
+ else if (is_branch (buf) || is_indirect_branch (buf))
+ /* If we hit a branch then we must be out of the prologue. */
+ break;
+ }
+
+ return 0;
+}
+
+/* qsort predicate to sort symbols by section and value. */
+
+static Elf_Internal_Sym *sort_syms_syms;
+static asection **sort_syms_psecs;
+
+static int
+sort_syms (const void *a, const void *b)
+{
+ Elf_Internal_Sym *const *s1 = a;
+ Elf_Internal_Sym *const *s2 = b;
+ asection *sec1,*sec2;
+ bfd_signed_vma delta;
+
+ sec1 = sort_syms_psecs[*s1 - sort_syms_syms];
+ sec2 = sort_syms_psecs[*s2 - sort_syms_syms];
+
+ if (sec1 != sec2)
+ return sec1->index - sec2->index;
+
+ delta = (*s1)->st_value - (*s2)->st_value;
+ if (delta != 0)
+ return delta < 0 ? -1 : 1;
+
+ delta = (*s2)->st_size - (*s1)->st_size;
+ if (delta != 0)
+ return delta < 0 ? -1 : 1;
+
+ return *s1 < *s2 ? -1 : 1;
+}
+
+/* Allocate a struct spu_elf_stack_info with MAX_FUN struct function_info
+ entries for section SEC. */
+
+static struct spu_elf_stack_info *
+alloc_stack_info (asection *sec, int max_fun)
+{
+ struct _spu_elf_section_data *sec_data = spu_elf_section_data (sec);
+ bfd_size_type amt;
+
+ amt = sizeof (struct spu_elf_stack_info);
+ amt += (max_fun - 1) * sizeof (struct function_info);
+ sec_data->u.i.stack_info = bfd_zmalloc (amt);
+ if (sec_data->u.i.stack_info != NULL)
+ sec_data->u.i.stack_info->max_fun = max_fun;
+ return sec_data->u.i.stack_info;
+}
+
+/* Add a new struct function_info describing a (part of a) function
+ starting at SYM_H. Keep the array sorted by address. */
+
+static struct function_info *
+maybe_insert_function (asection *sec,
+ void *sym_h,
+ bfd_boolean global,
+ bfd_boolean is_func)
+{
+ struct _spu_elf_section_data *sec_data = spu_elf_section_data (sec);
+ struct spu_elf_stack_info *sinfo = sec_data->u.i.stack_info;
+ int i;
+ bfd_vma off, size;
+
+ if (sinfo == NULL)
+ {
+ sinfo = alloc_stack_info (sec, 20);
+ if (sinfo == NULL)
+ return NULL;
+ }
+
+ if (!global)
+ {
+ Elf_Internal_Sym *sym = sym_h;
+ off = sym->st_value;
+ size = sym->st_size;
+ }
+ else
+ {
+ struct elf_link_hash_entry *h = sym_h;
+ off = h->root.u.def.value;
+ size = h->size;
+ }
+
+ for (i = sinfo->num_fun; --i >= 0; )
+ if (sinfo->fun[i].lo <= off)
+ break;
+
+ if (i >= 0)
+ {
+ /* Don't add another entry for an alias, but do update some
+ info. */
+ if (sinfo->fun[i].lo == off)
+ {
+ /* Prefer globals over local syms. */
+ if (global && !sinfo->fun[i].global)
+ {
+ sinfo->fun[i].global = TRUE;
+ sinfo->fun[i].u.h = sym_h;
+ }
+ if (is_func)
+ sinfo->fun[i].is_func = TRUE;
+ return &sinfo->fun[i];
+ }
+ /* Ignore a zero-size symbol inside an existing function. */
+ else if (sinfo->fun[i].hi > off && size == 0)
+ return &sinfo->fun[i];
+ }
+
+ if (sinfo->num_fun >= sinfo->max_fun)
+ {
+ bfd_size_type amt = sizeof (struct spu_elf_stack_info);
+ bfd_size_type old = amt;
+
+ old += (sinfo->max_fun - 1) * sizeof (struct function_info);
+ sinfo->max_fun += 20 + (sinfo->max_fun >> 1);
+ amt += (sinfo->max_fun - 1) * sizeof (struct function_info);
+ sinfo = bfd_realloc (sinfo, amt);
+ if (sinfo == NULL)
+ return NULL;
+ memset ((char *) sinfo + old, 0, amt - old);
+ sec_data->u.i.stack_info = sinfo;
+ }
+
+ if (++i < sinfo->num_fun)
+ memmove (&sinfo->fun[i + 1], &sinfo->fun[i],
+ (sinfo->num_fun - i) * sizeof (sinfo->fun[i]));
+ sinfo->fun[i].is_func = is_func;
+ sinfo->fun[i].global = global;
+ sinfo->fun[i].sec = sec;
+ if (global)
+ sinfo->fun[i].u.h = sym_h;
+ else
+ sinfo->fun[i].u.sym = sym_h;
+ sinfo->fun[i].lo = off;
+ sinfo->fun[i].hi = off + size;
+ sinfo->fun[i].lr_store = -1;
+ sinfo->fun[i].sp_adjust = -1;
+ sinfo->fun[i].stack = -find_function_stack_adjust (sec, off,
+ &sinfo->fun[i].lr_store,
+ &sinfo->fun[i].sp_adjust);
+ sinfo->num_fun += 1;
+ return &sinfo->fun[i];
+}
+
+/* Return the name of FUN. */
+
+static const char *
+func_name (struct function_info *fun)
+{
+ asection *sec;
+ bfd *ibfd;
+ Elf_Internal_Shdr *symtab_hdr;
+
+ while (fun->start != NULL)
+ fun = fun->start;
+
+ if (fun->global)
+ return fun->u.h->root.root.string;
+
+ sec = fun->sec;
+ if (fun->u.sym->st_name == 0)
+ {
+ size_t len = strlen (sec->name);
+ char *name = bfd_malloc (len + 10);
+ if (name == NULL)
+ return "(null)";
+ sprintf (name, "%s+%lx", sec->name,
+ (unsigned long) fun->u.sym->st_value & 0xffffffff);
+ return name;
+ }
+ ibfd = sec->owner;
+ symtab_hdr = &elf_tdata (ibfd)->symtab_hdr;
+ return bfd_elf_sym_name (ibfd, symtab_hdr, fun->u.sym, sec);
+}
+
+/* Read the instruction at OFF in SEC. Return true iff the instruction
+ is a nop, lnop, or stop 0 (all zero insn). */
+
+static bfd_boolean
+is_nop (asection *sec, bfd_vma off)
+{
+ unsigned char insn[4];
+
+ if (off + 4 > sec->size
+ || !bfd_get_section_contents (sec->owner, sec, insn, off, 4))
+ return FALSE;
+ if ((insn[0] & 0xbf) == 0 && (insn[1] & 0xe0) == 0x20)
+ return TRUE;
+ if (insn[0] == 0 && insn[1] == 0 && insn[2] == 0 && insn[3] == 0)
+ return TRUE;
+ return FALSE;
+}
+
+/* Extend the range of FUN to cover nop padding up to LIMIT.
+ Return TRUE iff some instruction other than a NOP was found. */
+
+static bfd_boolean
+insns_at_end (struct function_info *fun, bfd_vma limit)
+{
+ bfd_vma off = (fun->hi + 3) & -4;
+
+ while (off < limit && is_nop (fun->sec, off))
+ off += 4;
+ if (off < limit)
+ {
+ fun->hi = off;
+ return TRUE;
+ }
+ fun->hi = limit;
+ return FALSE;
+}
+
+/* Check and fix overlapping function ranges. Return TRUE iff there
+ are gaps in the current info we have about functions in SEC. */
+
+static bfd_boolean
+check_function_ranges (asection *sec, struct bfd_link_info *info)
+{
+ struct _spu_elf_section_data *sec_data = spu_elf_section_data (sec);
+ struct spu_elf_stack_info *sinfo = sec_data->u.i.stack_info;
+ int i;
+ bfd_boolean gaps = FALSE;
+
+ if (sinfo == NULL)
+ return FALSE;
+
+ for (i = 1; i < sinfo->num_fun; i++)
+ if (sinfo->fun[i - 1].hi > sinfo->fun[i].lo)
+ {
+ /* Fix overlapping symbols. */
+ const char *f1 = func_name (&sinfo->fun[i - 1]);
+ const char *f2 = func_name (&sinfo->fun[i]);
+
+ info->callbacks->einfo (_("warning: %s overlaps %s\n"), f1, f2);
+ sinfo->fun[i - 1].hi = sinfo->fun[i].lo;
+ }
+ else if (insns_at_end (&sinfo->fun[i - 1], sinfo->fun[i].lo))
+ gaps = TRUE;
+
+ if (sinfo->num_fun == 0)
+ gaps = TRUE;
+ else
+ {
+ if (sinfo->fun[0].lo != 0)
+ gaps = TRUE;
+ if (sinfo->fun[sinfo->num_fun - 1].hi > sec->size)
+ {
+ const char *f1 = func_name (&sinfo->fun[sinfo->num_fun - 1]);
+
+ info->callbacks->einfo (_("warning: %s exceeds section size\n"), f1);
+ sinfo->fun[sinfo->num_fun - 1].hi = sec->size;
+ }
+ else if (insns_at_end (&sinfo->fun[sinfo->num_fun - 1], sec->size))
+ gaps = TRUE;
+ }
+ return gaps;
+}
+
+/* Search current function info for a function that contains address
+ OFFSET in section SEC. */
+
+static struct function_info *
+find_function (asection *sec, bfd_vma offset, struct bfd_link_info *info)
+{
+ struct _spu_elf_section_data *sec_data = spu_elf_section_data (sec);
+ struct spu_elf_stack_info *sinfo = sec_data->u.i.stack_info;
+ int lo, hi, mid;
+
+ lo = 0;
+ hi = sinfo->num_fun;
+ while (lo < hi)
+ {
+ mid = (lo + hi) / 2;
+ if (offset < sinfo->fun[mid].lo)
+ hi = mid;
+ else if (offset >= sinfo->fun[mid].hi)
+ lo = mid + 1;
+ else
+ return &sinfo->fun[mid];
+ }
+ info->callbacks->einfo (_("%A:0x%v not found in function table\n"),
+ sec, offset);
+ bfd_set_error (bfd_error_bad_value);
+ return NULL;
+}
+
+/* Add CALLEE to CALLER call list if not already present. Return TRUE
+ if CALLEE was new. If this function return FALSE, CALLEE should
+ be freed. */
+
+static bfd_boolean
+insert_callee (struct function_info *caller, struct call_info *callee)
+{
+ struct call_info **pp, *p;
+
+ for (pp = &caller->call_list; (p = *pp) != NULL; pp = &p->next)
+ if (p->fun == callee->fun)
+ {
+ /* Tail calls use less stack than normal calls. Retain entry
+ for normal call over one for tail call. */
+ p->is_tail &= callee->is_tail;
+ if (!p->is_tail)
+ {
+ p->fun->start = NULL;
+ p->fun->is_func = TRUE;
+ }
+ p->count += callee->count;
+ /* Reorder list so most recent call is first. */
+ *pp = p->next;
+ p->next = caller->call_list;
+ caller->call_list = p;
+ return FALSE;
+ }
+ callee->next = caller->call_list;
+ caller->call_list = callee;
+ return TRUE;
+}
+
+/* Copy CALL and insert the copy into CALLER. */
+
+static bfd_boolean
+copy_callee (struct function_info *caller, const struct call_info *call)
+{
+ struct call_info *callee;
+ callee = bfd_malloc (sizeof (*callee));
+ if (callee == NULL)
+ return FALSE;
+ *callee = *call;
+ if (!insert_callee (caller, callee))
+ free (callee);
+ return TRUE;
+}
+
+/* We're only interested in code sections. Testing SEC_IN_MEMORY excludes
+ overlay stub sections. */
+
+static bfd_boolean
+interesting_section (asection *s)
+{
+ return (s->output_section != bfd_abs_section_ptr
+ && ((s->flags & (SEC_ALLOC | SEC_LOAD | SEC_CODE | SEC_IN_MEMORY))
+ == (SEC_ALLOC | SEC_LOAD | SEC_CODE))
+ && s->size != 0);
+}
+
+/* Rummage through the relocs for SEC, looking for function calls.
+ If CALL_TREE is true, fill in call graph. If CALL_TREE is false,
+ mark destination symbols on calls as being functions. Also
+ look at branches, which may be tail calls or go to hot/cold
+ section part of same function. */
+
+static bfd_boolean
+mark_functions_via_relocs (asection *sec,
+ struct bfd_link_info *info,
+ int call_tree)
+{
+ Elf_Internal_Rela *internal_relocs, *irelaend, *irela;
+ Elf_Internal_Shdr *symtab_hdr;
+ void *psyms;
+ unsigned int priority = 0;
+ static bfd_boolean warned;
+
+ if (!interesting_section (sec)
+ || sec->reloc_count == 0)
+ return TRUE;
+
+ internal_relocs = _bfd_elf_link_read_relocs (sec->owner, sec, NULL, NULL,
+ info->keep_memory);
+ if (internal_relocs == NULL)
+ return FALSE;
+
+ symtab_hdr = &elf_tdata (sec->owner)->symtab_hdr;
+ psyms = &symtab_hdr->contents;
+ irela = internal_relocs;
+ irelaend = irela + sec->reloc_count;
+ for (; irela < irelaend; irela++)
+ {
+ enum elf_spu_reloc_type r_type;
+ unsigned int r_indx;
+ asection *sym_sec;
+ Elf_Internal_Sym *sym;
+ struct elf_link_hash_entry *h;
+ bfd_vma val;
+ bfd_boolean nonbranch, is_call;
+ struct function_info *caller;
+ struct call_info *callee;
+
+ r_type = ELF32_R_TYPE (irela->r_info);
+ nonbranch = r_type != R_SPU_REL16 && r_type != R_SPU_ADDR16;
+
+ r_indx = ELF32_R_SYM (irela->r_info);
+ if (!get_sym_h (&h, &sym, &sym_sec, psyms, r_indx, sec->owner))
+ return FALSE;
+
+ if (sym_sec == NULL
+ || sym_sec->output_section == bfd_abs_section_ptr)
+ continue;
+
+ is_call = FALSE;
+ if (!nonbranch)
+ {
+ unsigned char insn[4];
+
+ if (!bfd_get_section_contents (sec->owner, sec, insn,
+ irela->r_offset, 4))
+ return FALSE;
+ if (is_branch (insn))
+ {
+ is_call = (insn[0] & 0xfd) == 0x31;
+ priority = insn[1] & 0x0f;
+ priority <<= 8;
+ priority |= insn[2];
+ priority <<= 8;
+ priority |= insn[3];
+ priority >>= 7;
+ if ((sym_sec->flags & (SEC_ALLOC | SEC_LOAD | SEC_CODE))
+ != (SEC_ALLOC | SEC_LOAD | SEC_CODE))
+ {
+ if (!warned)
+ info->callbacks->einfo
+ (_("%B(%A+0x%v): call to non-code section"
+ " %B(%A), analysis incomplete\n"),
+ sec->owner, sec, irela->r_offset,
+ sym_sec->owner, sym_sec);
+ warned = TRUE;
+ continue;
+ }
+ }
+ else
+ {
+ nonbranch = TRUE;
+ if (is_hint (insn))
+ continue;
+ }
+ }
+
+ if (nonbranch)
+ {
+ /* For --auto-overlay, count possible stubs we need for
+ function pointer references. */
+ unsigned int sym_type;
+ if (h)
+ sym_type = h->type;
+ else
+ sym_type = ELF_ST_TYPE (sym->st_info);
+ if (sym_type == STT_FUNC)
+ {
+ if (call_tree && spu_hash_table (info)->params->auto_overlay)
+ spu_hash_table (info)->non_ovly_stub += 1;
+ /* If the symbol type is STT_FUNC then this must be a
+ function pointer initialisation. */
+ continue;
+ }
+ /* Ignore data references. */
+ if ((sym_sec->flags & (SEC_ALLOC | SEC_LOAD | SEC_CODE))
+ != (SEC_ALLOC | SEC_LOAD | SEC_CODE))
+ continue;
+ /* Otherwise we probably have a jump table reloc for
+ a switch statement or some other reference to a
+ code label. */
+ }
+
+ if (h)
+ val = h->root.u.def.value;
+ else
+ val = sym->st_value;
+ val += irela->r_addend;
+
+ if (!call_tree)
+ {
+ struct function_info *fun;
+
+ if (irela->r_addend != 0)
+ {
+ Elf_Internal_Sym *fake = bfd_zmalloc (sizeof (*fake));
+ if (fake == NULL)
+ return FALSE;
+ fake->st_value = val;
+ fake->st_shndx
+ = _bfd_elf_section_from_bfd_section (sym_sec->owner, sym_sec);
+ sym = fake;
+ }
+ if (sym)
+ fun = maybe_insert_function (sym_sec, sym, FALSE, is_call);
+ else
+ fun = maybe_insert_function (sym_sec, h, TRUE, is_call);
+ if (fun == NULL)
+ return FALSE;
+ if (irela->r_addend != 0
+ && fun->u.sym != sym)
+ free (sym);
+ continue;
+ }
+
+ caller = find_function (sec, irela->r_offset, info);
+ if (caller == NULL)
+ return FALSE;
+ callee = bfd_malloc (sizeof *callee);
+ if (callee == NULL)
+ return FALSE;
+
+ callee->fun = find_function (sym_sec, val, info);
+ if (callee->fun == NULL)
+ return FALSE;
+ callee->is_tail = !is_call;
+ callee->is_pasted = FALSE;
+ callee->broken_cycle = FALSE;
+ callee->priority = priority;
+ callee->count = nonbranch? 0 : 1;
+ if (callee->fun->last_caller != sec)
+ {
+ callee->fun->last_caller = sec;
+ callee->fun->call_count += 1;
+ }
+ if (!insert_callee (caller, callee))
+ free (callee);
+ else if (!is_call
+ && !callee->fun->is_func
+ && callee->fun->stack == 0)
+ {
+ /* This is either a tail call or a branch from one part of
+ the function to another, ie. hot/cold section. If the
+ destination has been called by some other function then
+ it is a separate function. We also assume that functions
+ are not split across input files. */
+ if (sec->owner != sym_sec->owner)
+ {
+ callee->fun->start = NULL;
+ callee->fun->is_func = TRUE;
+ }
+ else if (callee->fun->start == NULL)
+ {
+ struct function_info *caller_start = caller;
+ while (caller_start->start)
+ caller_start = caller_start->start;
+
+ if (caller_start != callee->fun)
+ callee->fun->start = caller_start;
+ }
+ else
+ {
+ struct function_info *callee_start;
+ struct function_info *caller_start;
+ callee_start = callee->fun;
+ while (callee_start->start)
+ callee_start = callee_start->start;
+ caller_start = caller;
+ while (caller_start->start)
+ caller_start = caller_start->start;
+ if (caller_start != callee_start)
+ {
+ callee->fun->start = NULL;
+ callee->fun->is_func = TRUE;
+ }
+ }
+ }
+ }
+
+ return TRUE;
+}
+
+/* Handle something like .init or .fini, which has a piece of a function.
+ These sections are pasted together to form a single function. */
+
+static bfd_boolean
+pasted_function (asection *sec)
+{
+ struct bfd_link_order *l;
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+ Elf_Internal_Sym *fake;
+ struct function_info *fun, *fun_start;
+
+ fake = bfd_zmalloc (sizeof (*fake));
+ if (fake == NULL)
+ return FALSE;
+ fake->st_value = 0;
+ fake->st_size = sec->size;
+ fake->st_shndx
+ = _bfd_elf_section_from_bfd_section (sec->owner, sec);
+ fun = maybe_insert_function (sec, fake, FALSE, FALSE);
+ if (!fun)
+ return FALSE;
+
+ /* Find a function immediately preceding this section. */
+ fun_start = NULL;
+ for (l = sec->output_section->map_head.link_order; l != NULL; l = l->next)
+ {
+ if (l->u.indirect.section == sec)
+ {
+ if (fun_start != NULL)
+ {
+ struct call_info *callee = bfd_malloc (sizeof *callee);
+ if (callee == NULL)
+ return FALSE;
+
+ fun->start = fun_start;
+ callee->fun = fun;
+ callee->is_tail = TRUE;
+ callee->is_pasted = TRUE;
+ callee->broken_cycle = FALSE;
+ callee->priority = 0;
+ callee->count = 1;
+ if (!insert_callee (fun_start, callee))
+ free (callee);
+ return TRUE;
+ }
+ break;
+ }
+ if (l->type == bfd_indirect_link_order
+ && (sec_data = spu_elf_section_data (l->u.indirect.section)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL
+ && sinfo->num_fun != 0)
+ fun_start = &sinfo->fun[sinfo->num_fun - 1];
+ }
+
+ /* Don't return an error if we did not find a function preceding this
+ section. The section may have incorrect flags. */
+ return TRUE;
+}
+
+/* Map address ranges in code sections to functions. */
+
+static bfd_boolean
+discover_functions (struct bfd_link_info *info)
+{
+ bfd *ibfd;
+ int bfd_idx;
+ Elf_Internal_Sym ***psym_arr;
+ asection ***sec_arr;
+ bfd_boolean gaps = FALSE;
+
+ bfd_idx = 0;
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ bfd_idx++;
+
+ psym_arr = bfd_zmalloc (bfd_idx * sizeof (*psym_arr));
+ if (psym_arr == NULL)
+ return FALSE;
+ sec_arr = bfd_zmalloc (bfd_idx * sizeof (*sec_arr));
+ if (sec_arr == NULL)
+ return FALSE;
+
+ for (ibfd = info->input_bfds, bfd_idx = 0;
+ ibfd != NULL;
+ ibfd = ibfd->link.next, bfd_idx++)
+ {
+ extern const bfd_target spu_elf32_vec;
+ Elf_Internal_Shdr *symtab_hdr;
+ asection *sec;
+ size_t symcount;
+ Elf_Internal_Sym *syms, *sy, **psyms, **psy;
+ asection **psecs, **p;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ /* Read all the symbols. */
+ symtab_hdr = &elf_tdata (ibfd)->symtab_hdr;
+ symcount = symtab_hdr->sh_size / symtab_hdr->sh_entsize;
+ if (symcount == 0)
+ {
+ if (!gaps)
+ for (sec = ibfd->sections; sec != NULL && !gaps; sec = sec->next)
+ if (interesting_section (sec))
+ {
+ gaps = TRUE;
+ break;
+ }
+ continue;
+ }
+
+ if (symtab_hdr->contents != NULL)
+ {
+ /* Don't use cached symbols since the generic ELF linker
+ code only reads local symbols, and we need globals too. */
+ free (symtab_hdr->contents);
+ symtab_hdr->contents = NULL;
+ }
+ syms = bfd_elf_get_elf_syms (ibfd, symtab_hdr, symcount, 0,
+ NULL, NULL, NULL);
+ symtab_hdr->contents = (void *) syms;
+ if (syms == NULL)
+ return FALSE;
+
+ /* Select defined function symbols that are going to be output. */
+ psyms = bfd_malloc ((symcount + 1) * sizeof (*psyms));
+ if (psyms == NULL)
+ return FALSE;
+ psym_arr[bfd_idx] = psyms;
+ psecs = bfd_malloc (symcount * sizeof (*psecs));
+ if (psecs == NULL)
+ return FALSE;
+ sec_arr[bfd_idx] = psecs;
+ for (psy = psyms, p = psecs, sy = syms; sy < syms + symcount; ++p, ++sy)
+ if (ELF_ST_TYPE (sy->st_info) == STT_NOTYPE
+ || ELF_ST_TYPE (sy->st_info) == STT_FUNC)
+ {
+ asection *s;
+
+ *p = s = bfd_section_from_elf_index (ibfd, sy->st_shndx);
+ if (s != NULL && interesting_section (s))
+ *psy++ = sy;
+ }
+ symcount = psy - psyms;
+ *psy = NULL;
+
+ /* Sort them by section and offset within section. */
+ sort_syms_syms = syms;
+ sort_syms_psecs = psecs;
+ qsort (psyms, symcount, sizeof (*psyms), sort_syms);
+
+ /* Now inspect the function symbols. */
+ for (psy = psyms; psy < psyms + symcount; )
+ {
+ asection *s = psecs[*psy - syms];
+ Elf_Internal_Sym **psy2;
+
+ for (psy2 = psy; ++psy2 < psyms + symcount; )
+ if (psecs[*psy2 - syms] != s)
+ break;
+
+ if (!alloc_stack_info (s, psy2 - psy))
+ return FALSE;
+ psy = psy2;
+ }
+
+ /* First install info about properly typed and sized functions.
+ In an ideal world this will cover all code sections, except
+ when partitioning functions into hot and cold sections,
+ and the horrible pasted together .init and .fini functions. */
+ for (psy = psyms; psy < psyms + symcount; ++psy)
+ {
+ sy = *psy;
+ if (ELF_ST_TYPE (sy->st_info) == STT_FUNC)
+ {
+ asection *s = psecs[sy - syms];
+ if (!maybe_insert_function (s, sy, FALSE, TRUE))
+ return FALSE;
+ }
+ }
+
+ for (sec = ibfd->sections; sec != NULL && !gaps; sec = sec->next)
+ if (interesting_section (sec))
+ gaps |= check_function_ranges (sec, info);
+ }
+
+ if (gaps)
+ {
+ /* See if we can discover more function symbols by looking at
+ relocations. */
+ for (ibfd = info->input_bfds, bfd_idx = 0;
+ ibfd != NULL;
+ ibfd = ibfd->link.next, bfd_idx++)
+ {
+ asection *sec;
+
+ if (psym_arr[bfd_idx] == NULL)
+ continue;
+
+ for (sec = ibfd->sections; sec != NULL; sec = sec->next)
+ if (!mark_functions_via_relocs (sec, info, FALSE))
+ return FALSE;
+ }
+
+ for (ibfd = info->input_bfds, bfd_idx = 0;
+ ibfd != NULL;
+ ibfd = ibfd->link.next, bfd_idx++)
+ {
+ Elf_Internal_Shdr *symtab_hdr;
+ asection *sec;
+ Elf_Internal_Sym *syms, *sy, **psyms, **psy;
+ asection **psecs;
+
+ if ((psyms = psym_arr[bfd_idx]) == NULL)
+ continue;
+
+ psecs = sec_arr[bfd_idx];
+
+ symtab_hdr = &elf_tdata (ibfd)->symtab_hdr;
+ syms = (Elf_Internal_Sym *) symtab_hdr->contents;
+
+ gaps = FALSE;
+ for (sec = ibfd->sections; sec != NULL && !gaps; sec = sec->next)
+ if (interesting_section (sec))
+ gaps |= check_function_ranges (sec, info);
+ if (!gaps)
+ continue;
+
+ /* Finally, install all globals. */
+ for (psy = psyms; (sy = *psy) != NULL; ++psy)
+ {
+ asection *s;
+
+ s = psecs[sy - syms];
+
+ /* Global syms might be improperly typed functions. */
+ if (ELF_ST_TYPE (sy->st_info) != STT_FUNC
+ && ELF_ST_BIND (sy->st_info) == STB_GLOBAL)
+ {
+ if (!maybe_insert_function (s, sy, FALSE, FALSE))
+ return FALSE;
+ }
+ }
+ }
+
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ {
+ extern const bfd_target spu_elf32_vec;
+ asection *sec;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ /* Some of the symbols we've installed as marking the
+ beginning of functions may have a size of zero. Extend
+ the range of such functions to the beginning of the
+ next symbol of interest. */
+ for (sec = ibfd->sections; sec != NULL; sec = sec->next)
+ if (interesting_section (sec))
+ {
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+
+ sec_data = spu_elf_section_data (sec);
+ sinfo = sec_data->u.i.stack_info;
+ if (sinfo != NULL && sinfo->num_fun != 0)
+ {
+ int fun_idx;
+ bfd_vma hi = sec->size;
+
+ for (fun_idx = sinfo->num_fun; --fun_idx >= 0; )
+ {
+ sinfo->fun[fun_idx].hi = hi;
+ hi = sinfo->fun[fun_idx].lo;
+ }
+
+ sinfo->fun[0].lo = 0;
+ }
+ /* No symbols in this section. Must be .init or .fini
+ or something similar. */
+ else if (!pasted_function (sec))
+ return FALSE;
+ }
+ }
+ }
+
+ for (ibfd = info->input_bfds, bfd_idx = 0;
+ ibfd != NULL;
+ ibfd = ibfd->link.next, bfd_idx++)
+ {
+ if (psym_arr[bfd_idx] == NULL)
+ continue;
+
+ free (psym_arr[bfd_idx]);
+ free (sec_arr[bfd_idx]);
+ }
+
+ free (psym_arr);
+ free (sec_arr);
+
+ return TRUE;
+}
+
+/* Iterate over all function_info we have collected, calling DOIT on
+ each node if ROOT_ONLY is false. Only call DOIT on root nodes
+ if ROOT_ONLY. */
+
+static bfd_boolean
+for_each_node (bfd_boolean (*doit) (struct function_info *,
+ struct bfd_link_info *,
+ void *),
+ struct bfd_link_info *info,
+ void *param,
+ int root_only)
+{
+ bfd *ibfd;
+
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ {
+ extern const bfd_target spu_elf32_vec;
+ asection *sec;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ for (sec = ibfd->sections; sec != NULL; sec = sec->next)
+ {
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+
+ if ((sec_data = spu_elf_section_data (sec)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL)
+ {
+ int i;
+ for (i = 0; i < sinfo->num_fun; ++i)
+ if (!root_only || !sinfo->fun[i].non_root)
+ if (!doit (&sinfo->fun[i], info, param))
+ return FALSE;
+ }
+ }
+ }
+ return TRUE;
+}
+
+/* Transfer call info attached to struct function_info entries for
+ all of a given function's sections to the first entry. */
+
+static bfd_boolean
+transfer_calls (struct function_info *fun,
+ struct bfd_link_info *info ATTRIBUTE_UNUSED,
+ void *param ATTRIBUTE_UNUSED)
+{
+ struct function_info *start = fun->start;
+
+ if (start != NULL)
+ {
+ struct call_info *call, *call_next;
+
+ while (start->start != NULL)
+ start = start->start;
+ for (call = fun->call_list; call != NULL; call = call_next)
+ {
+ call_next = call->next;
+ if (!insert_callee (start, call))
+ free (call);
+ }
+ fun->call_list = NULL;
+ }
+ return TRUE;
+}
+
+/* Mark nodes in the call graph that are called by some other node. */
+
+static bfd_boolean
+mark_non_root (struct function_info *fun,
+ struct bfd_link_info *info ATTRIBUTE_UNUSED,
+ void *param ATTRIBUTE_UNUSED)
+{
+ struct call_info *call;
+
+ if (fun->visit1)
+ return TRUE;
+ fun->visit1 = TRUE;
+ for (call = fun->call_list; call; call = call->next)
+ {
+ call->fun->non_root = TRUE;
+ mark_non_root (call->fun, 0, 0);
+ }
+ return TRUE;
+}
+
+/* Remove cycles from the call graph. Set depth of nodes. */
+
+static bfd_boolean
+remove_cycles (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ struct call_info **callp, *call;
+ unsigned int depth = *(unsigned int *) param;
+ unsigned int max_depth = depth;
+
+ fun->depth = depth;
+ fun->visit2 = TRUE;
+ fun->marking = TRUE;
+
+ callp = &fun->call_list;
+ while ((call = *callp) != NULL)
+ {
+ call->max_depth = depth + !call->is_pasted;
+ if (!call->fun->visit2)
+ {
+ if (!remove_cycles (call->fun, info, &call->max_depth))
+ return FALSE;
+ if (max_depth < call->max_depth)
+ max_depth = call->max_depth;
+ }
+ else if (call->fun->marking)
+ {
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+
+ if (!htab->params->auto_overlay
+ && htab->params->stack_analysis)
+ {
+ const char *f1 = func_name (fun);
+ const char *f2 = func_name (call->fun);
+
+ info->callbacks->info (_("Stack analysis will ignore the call "
+ "from %s to %s\n"),
+ f1, f2);
+ }
+
+ call->broken_cycle = TRUE;
+ }
+ callp = &call->next;
+ }
+ fun->marking = FALSE;
+ *(unsigned int *) param = max_depth;
+ return TRUE;
+}
+
+/* Check that we actually visited all nodes in remove_cycles. If we
+ didn't, then there is some cycle in the call graph not attached to
+ any root node. Arbitrarily choose a node in the cycle as a new
+ root and break the cycle. */
+
+static bfd_boolean
+mark_detached_root (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ if (fun->visit2)
+ return TRUE;
+ fun->non_root = FALSE;
+ *(unsigned int *) param = 0;
+ return remove_cycles (fun, info, param);
+}
+
+/* Populate call_list for each function. */
+
+static bfd_boolean
+build_call_tree (struct bfd_link_info *info)
+{
+ bfd *ibfd;
+ unsigned int depth;
+
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ {
+ extern const bfd_target spu_elf32_vec;
+ asection *sec;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ for (sec = ibfd->sections; sec != NULL; sec = sec->next)
+ if (!mark_functions_via_relocs (sec, info, TRUE))
+ return FALSE;
+ }
+
+ /* Transfer call info from hot/cold section part of function
+ to main entry. */
+ if (!spu_hash_table (info)->params->auto_overlay
+ && !for_each_node (transfer_calls, info, 0, FALSE))
+ return FALSE;
+
+ /* Find the call graph root(s). */
+ if (!for_each_node (mark_non_root, info, 0, FALSE))
+ return FALSE;
+
+ /* Remove cycles from the call graph. We start from the root node(s)
+ so that we break cycles in a reasonable place. */
+ depth = 0;
+ if (!for_each_node (remove_cycles, info, &depth, TRUE))
+ return FALSE;
+
+ return for_each_node (mark_detached_root, info, &depth, FALSE);
+}
+
+/* qsort predicate to sort calls by priority, max_depth then count. */
+
+static int
+sort_calls (const void *a, const void *b)
+{
+ struct call_info *const *c1 = a;
+ struct call_info *const *c2 = b;
+ int delta;
+
+ delta = (*c2)->priority - (*c1)->priority;
+ if (delta != 0)
+ return delta;
+
+ delta = (*c2)->max_depth - (*c1)->max_depth;
+ if (delta != 0)
+ return delta;
+
+ delta = (*c2)->count - (*c1)->count;
+ if (delta != 0)
+ return delta;
+
+ return (char *) c1 - (char *) c2;
+}
+
+struct _mos_param {
+ unsigned int max_overlay_size;
+};
+
+/* Set linker_mark and gc_mark on any sections that we will put in
+ overlays. These flags are used by the generic ELF linker, but we
+ won't be continuing on to bfd_elf_final_link so it is OK to use
+ them. linker_mark is clear before we get here. Set segment_mark
+ on sections that are part of a pasted function (excluding the last
+ section).
+
+ Set up function rodata section if --overlay-rodata. We don't
+ currently include merged string constant rodata sections since
+
+ Sort the call graph so that the deepest nodes will be visited
+ first. */
+
+static bfd_boolean
+mark_overlay_section (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ struct call_info *call;
+ unsigned int count;
+ struct _mos_param *mos_param = param;
+ struct spu_link_hash_table *htab = spu_hash_table (info);
+
+ if (fun->visit4)
+ return TRUE;
+
+ fun->visit4 = TRUE;
+ if (!fun->sec->linker_mark
+ && (htab->params->ovly_flavour != ovly_soft_icache
+ || htab->params->non_ia_text
+ || strncmp (fun->sec->name, ".text.ia.", 9) == 0
+ || strcmp (fun->sec->name, ".init") == 0
+ || strcmp (fun->sec->name, ".fini") == 0))
+ {
+ unsigned int size;
+
+ fun->sec->linker_mark = 1;
+ fun->sec->gc_mark = 1;
+ fun->sec->segment_mark = 0;
+ /* Ensure SEC_CODE is set on this text section (it ought to
+ be!), and SEC_CODE is clear on rodata sections. We use
+ this flag to differentiate the two overlay section types. */
+ fun->sec->flags |= SEC_CODE;
+
+ size = fun->sec->size;
+ if (htab->params->auto_overlay & OVERLAY_RODATA)
+ {
+ char *name = NULL;
+
+ /* Find the rodata section corresponding to this function's
+ text section. */
+ if (strcmp (fun->sec->name, ".text") == 0)
+ {
+ name = bfd_malloc (sizeof (".rodata"));
+ if (name == NULL)
+ return FALSE;
+ memcpy (name, ".rodata", sizeof (".rodata"));
+ }
+ else if (strncmp (fun->sec->name, ".text.", 6) == 0)
+ {
+ size_t len = strlen (fun->sec->name);
+ name = bfd_malloc (len + 3);
+ if (name == NULL)
+ return FALSE;
+ memcpy (name, ".rodata", sizeof (".rodata"));
+ memcpy (name + 7, fun->sec->name + 5, len - 4);
+ }
+ else if (strncmp (fun->sec->name, ".gnu.linkonce.t.", 16) == 0)
+ {
+ size_t len = strlen (fun->sec->name) + 1;
+ name = bfd_malloc (len);
+ if (name == NULL)
+ return FALSE;
+ memcpy (name, fun->sec->name, len);
+ name[14] = 'r';
+ }
+
+ if (name != NULL)
+ {
+ asection *rodata = NULL;
+ asection *group_sec = elf_section_data (fun->sec)->next_in_group;
+ if (group_sec == NULL)
+ rodata = bfd_get_section_by_name (fun->sec->owner, name);
+ else
+ while (group_sec != NULL && group_sec != fun->sec)
+ {
+ if (strcmp (group_sec->name, name) == 0)
+ {
+ rodata = group_sec;
+ break;
+ }
+ group_sec = elf_section_data (group_sec)->next_in_group;
+ }
+ fun->rodata = rodata;
+ if (fun->rodata)
+ {
+ size += fun->rodata->size;
+ if (htab->params->line_size != 0
+ && size > htab->params->line_size)
+ {
+ size -= fun->rodata->size;
+ fun->rodata = NULL;
+ }
+ else
+ {
+ fun->rodata->linker_mark = 1;
+ fun->rodata->gc_mark = 1;
+ fun->rodata->flags &= ~SEC_CODE;
+ }
+ }
+ free (name);
+ }
+ }
+ if (mos_param->max_overlay_size < size)
+ mos_param->max_overlay_size = size;
+ }
+
+ for (count = 0, call = fun->call_list; call != NULL; call = call->next)
+ count += 1;
+
+ if (count > 1)
+ {
+ struct call_info **calls = bfd_malloc (count * sizeof (*calls));
+ if (calls == NULL)
+ return FALSE;
+
+ for (count = 0, call = fun->call_list; call != NULL; call = call->next)
+ calls[count++] = call;
+
+ qsort (calls, count, sizeof (*calls), sort_calls);
+
+ fun->call_list = NULL;
+ while (count != 0)
+ {
+ --count;
+ calls[count]->next = fun->call_list;
+ fun->call_list = calls[count];
+ }
+ free (calls);
+ }
+
+ for (call = fun->call_list; call != NULL; call = call->next)
+ {
+ if (call->is_pasted)
+ {
+ /* There can only be one is_pasted call per function_info. */
+ BFD_ASSERT (!fun->sec->segment_mark);
+ fun->sec->segment_mark = 1;
+ }
+ if (!call->broken_cycle
+ && !mark_overlay_section (call->fun, info, param))
+ return FALSE;
+ }
+
+ /* Don't put entry code into an overlay. The overlay manager needs
+ a stack! Also, don't mark .ovl.init as an overlay. */
+ if (fun->lo + fun->sec->output_offset + fun->sec->output_section->vma
+ == info->output_bfd->start_address
+ || strncmp (fun->sec->output_section->name, ".ovl.init", 9) == 0)
+ {
+ fun->sec->linker_mark = 0;
+ if (fun->rodata != NULL)
+ fun->rodata->linker_mark = 0;
+ }
+ return TRUE;
+}
+
+/* If non-zero then unmark functions called from those within sections
+ that we need to unmark. Unfortunately this isn't reliable since the
+ call graph cannot know the destination of function pointer calls. */
+#define RECURSE_UNMARK 0
+
+struct _uos_param {
+ asection *exclude_input_section;
+ asection *exclude_output_section;
+ unsigned long clearing;
+};
+
+/* Undo some of mark_overlay_section's work. */
+
+static bfd_boolean
+unmark_overlay_section (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ struct call_info *call;
+ struct _uos_param *uos_param = param;
+ unsigned int excluded = 0;
+
+ if (fun->visit5)
+ return TRUE;
+
+ fun->visit5 = TRUE;
+
+ excluded = 0;
+ if (fun->sec == uos_param->exclude_input_section
+ || fun->sec->output_section == uos_param->exclude_output_section)
+ excluded = 1;
+
+ if (RECURSE_UNMARK)
+ uos_param->clearing += excluded;
+
+ if (RECURSE_UNMARK ? uos_param->clearing : excluded)
+ {
+ fun->sec->linker_mark = 0;
+ if (fun->rodata)
+ fun->rodata->linker_mark = 0;
+ }
+
+ for (call = fun->call_list; call != NULL; call = call->next)
+ if (!call->broken_cycle
+ && !unmark_overlay_section (call->fun, info, param))
+ return FALSE;
+
+ if (RECURSE_UNMARK)
+ uos_param->clearing -= excluded;
+ return TRUE;
+}
+
+struct _cl_param {
+ unsigned int lib_size;
+ asection **lib_sections;
+};
+
+/* Add sections we have marked as belonging to overlays to an array
+ for consideration as non-overlay sections. The array consist of
+ pairs of sections, (text,rodata), for functions in the call graph. */
+
+static bfd_boolean
+collect_lib_sections (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ struct _cl_param *lib_param = param;
+ struct call_info *call;
+ unsigned int size;
+
+ if (fun->visit6)
+ return TRUE;
+
+ fun->visit6 = TRUE;
+ if (!fun->sec->linker_mark || !fun->sec->gc_mark || fun->sec->segment_mark)
+ return TRUE;
+
+ size = fun->sec->size;
+ if (fun->rodata)
+ size += fun->rodata->size;
+
+ if (size <= lib_param->lib_size)
+ {
+ *lib_param->lib_sections++ = fun->sec;
+ fun->sec->gc_mark = 0;
+ if (fun->rodata && fun->rodata->linker_mark && fun->rodata->gc_mark)
+ {
+ *lib_param->lib_sections++ = fun->rodata;
+ fun->rodata->gc_mark = 0;
+ }
+ else
+ *lib_param->lib_sections++ = NULL;
+ }
+
+ for (call = fun->call_list; call != NULL; call = call->next)
+ if (!call->broken_cycle)
+ collect_lib_sections (call->fun, info, param);
+
+ return TRUE;
+}
+
+/* qsort predicate to sort sections by call count. */
+
+static int
+sort_lib (const void *a, const void *b)
+{
+ asection *const *s1 = a;
+ asection *const *s2 = b;
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+ int delta;
+
+ delta = 0;
+ if ((sec_data = spu_elf_section_data (*s1)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL)
+ {
+ int i;
+ for (i = 0; i < sinfo->num_fun; ++i)
+ delta -= sinfo->fun[i].call_count;
+ }
+
+ if ((sec_data = spu_elf_section_data (*s2)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL)
+ {
+ int i;
+ for (i = 0; i < sinfo->num_fun; ++i)
+ delta += sinfo->fun[i].call_count;
+ }
+
+ if (delta != 0)
+ return delta;
+
+ return s1 - s2;
+}
+
+/* Remove some sections from those marked to be in overlays. Choose
+ those that are called from many places, likely library functions. */
+
+static unsigned int
+auto_ovl_lib_functions (struct bfd_link_info *info, unsigned int lib_size)
+{
+ bfd *ibfd;
+ asection **lib_sections;
+ unsigned int i, lib_count;
+ struct _cl_param collect_lib_param;
+ struct function_info dummy_caller;
+ struct spu_link_hash_table *htab;
+
+ memset (&dummy_caller, 0, sizeof (dummy_caller));
+ lib_count = 0;
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ {
+ extern const bfd_target spu_elf32_vec;
+ asection *sec;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ for (sec = ibfd->sections; sec != NULL; sec = sec->next)
+ if (sec->linker_mark
+ && sec->size < lib_size
+ && (sec->flags & SEC_CODE) != 0)
+ lib_count += 1;
+ }
+ lib_sections = bfd_malloc (lib_count * 2 * sizeof (*lib_sections));
+ if (lib_sections == NULL)
+ return (unsigned int) -1;
+ collect_lib_param.lib_size = lib_size;
+ collect_lib_param.lib_sections = lib_sections;
+ if (!for_each_node (collect_lib_sections, info, &collect_lib_param,
+ TRUE))
+ return (unsigned int) -1;
+ lib_count = (collect_lib_param.lib_sections - lib_sections) / 2;
+
+ /* Sort sections so that those with the most calls are first. */
+ if (lib_count > 1)
+ qsort (lib_sections, lib_count, 2 * sizeof (*lib_sections), sort_lib);
+
+ htab = spu_hash_table (info);
+ for (i = 0; i < lib_count; i++)
+ {
+ unsigned int tmp, stub_size;
+ asection *sec;
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+
+ sec = lib_sections[2 * i];
+ /* If this section is OK, its size must be less than lib_size. */
+ tmp = sec->size;
+ /* If it has a rodata section, then add that too. */
+ if (lib_sections[2 * i + 1])
+ tmp += lib_sections[2 * i + 1]->size;
+ /* Add any new overlay call stubs needed by the section. */
+ stub_size = 0;
+ if (tmp < lib_size
+ && (sec_data = spu_elf_section_data (sec)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL)
+ {
+ int k;
+ struct call_info *call;
+
+ for (k = 0; k < sinfo->num_fun; ++k)
+ for (call = sinfo->fun[k].call_list; call; call = call->next)
+ if (call->fun->sec->linker_mark)
+ {
+ struct call_info *p;
+ for (p = dummy_caller.call_list; p; p = p->next)
+ if (p->fun == call->fun)
+ break;
+ if (!p)
+ stub_size += ovl_stub_size (htab->params);
+ }
+ }
+ if (tmp + stub_size < lib_size)
+ {
+ struct call_info **pp, *p;
+
+ /* This section fits. Mark it as non-overlay. */
+ lib_sections[2 * i]->linker_mark = 0;
+ if (lib_sections[2 * i + 1])
+ lib_sections[2 * i + 1]->linker_mark = 0;
+ lib_size -= tmp + stub_size;
+ /* Call stubs to the section we just added are no longer
+ needed. */
+ pp = &dummy_caller.call_list;
+ while ((p = *pp) != NULL)
+ if (!p->fun->sec->linker_mark)
+ {
+ lib_size += ovl_stub_size (htab->params);
+ *pp = p->next;
+ free (p);
+ }
+ else
+ pp = &p->next;
+ /* Add new call stubs to dummy_caller. */
+ if ((sec_data = spu_elf_section_data (sec)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL)
+ {
+ int k;
+ struct call_info *call;
+
+ for (k = 0; k < sinfo->num_fun; ++k)
+ for (call = sinfo->fun[k].call_list;
+ call;
+ call = call->next)
+ if (call->fun->sec->linker_mark)
+ {
+ struct call_info *callee;
+ callee = bfd_malloc (sizeof (*callee));
+ if (callee == NULL)
+ return (unsigned int) -1;
+ *callee = *call;
+ if (!insert_callee (&dummy_caller, callee))
+ free (callee);
+ }
+ }
+ }
+ }
+ while (dummy_caller.call_list != NULL)
+ {
+ struct call_info *call = dummy_caller.call_list;
+ dummy_caller.call_list = call->next;
+ free (call);
+ }
+ for (i = 0; i < 2 * lib_count; i++)
+ if (lib_sections[i])
+ lib_sections[i]->gc_mark = 1;
+ free (lib_sections);
+ return lib_size;
+}
+
+/* Build an array of overlay sections. The deepest node's section is
+ added first, then its parent node's section, then everything called
+ from the parent section. The idea being to group sections to
+ minimise calls between different overlays. */
+
+static bfd_boolean
+collect_overlays (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ struct call_info *call;
+ bfd_boolean added_fun;
+ asection ***ovly_sections = param;
+
+ if (fun->visit7)
+ return TRUE;
+
+ fun->visit7 = TRUE;
+ for (call = fun->call_list; call != NULL; call = call->next)
+ if (!call->is_pasted && !call->broken_cycle)
+ {
+ if (!collect_overlays (call->fun, info, ovly_sections))
+ return FALSE;
+ break;
+ }
+
+ added_fun = FALSE;
+ if (fun->sec->linker_mark && fun->sec->gc_mark)
+ {
+ fun->sec->gc_mark = 0;
+ *(*ovly_sections)++ = fun->sec;
+ if (fun->rodata && fun->rodata->linker_mark && fun->rodata->gc_mark)
+ {
+ fun->rodata->gc_mark = 0;
+ *(*ovly_sections)++ = fun->rodata;
+ }
+ else
+ *(*ovly_sections)++ = NULL;
+ added_fun = TRUE;
+
+ /* Pasted sections must stay with the first section. We don't
+ put pasted sections in the array, just the first section.
+ Mark subsequent sections as already considered. */
+ if (fun->sec->segment_mark)
+ {
+ struct function_info *call_fun = fun;
+ do
+ {
+ for (call = call_fun->call_list; call != NULL; call = call->next)
+ if (call->is_pasted)
+ {
+ call_fun = call->fun;
+ call_fun->sec->gc_mark = 0;
+ if (call_fun->rodata)
+ call_fun->rodata->gc_mark = 0;
+ break;
+ }
+ if (call == NULL)
+ abort ();
+ }
+ while (call_fun->sec->segment_mark);
+ }
+ }
+
+ for (call = fun->call_list; call != NULL; call = call->next)
+ if (!call->broken_cycle
+ && !collect_overlays (call->fun, info, ovly_sections))
+ return FALSE;
+
+ if (added_fun)
+ {
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+
+ if ((sec_data = spu_elf_section_data (fun->sec)) != NULL
+ && (sinfo = sec_data->u.i.stack_info) != NULL)
+ {
+ int i;
+ for (i = 0; i < sinfo->num_fun; ++i)
+ if (!collect_overlays (&sinfo->fun[i], info, ovly_sections))
+ return FALSE;
+ }
+ }
+
+ return TRUE;
+}
+
+struct _sum_stack_param {
+ size_t cum_stack;
+ size_t overall_stack;
+ bfd_boolean emit_stack_syms;
+};
+
+/* Descend the call graph for FUN, accumulating total stack required. */
+
+static bfd_boolean
+sum_stack (struct function_info *fun,
+ struct bfd_link_info *info,
+ void *param)
+{
+ struct call_info *call;
+ struct function_info *max;
+ size_t stack, cum_stack;
+ const char *f1;
+ bfd_boolean has_call;
+ struct _sum_stack_param *sum_stack_param = param;
+ struct spu_link_hash_table *htab;
+
+ cum_stack = fun->stack;
+ sum_stack_param->cum_stack = cum_stack;
+ if (fun->visit3)
+ return TRUE;
+
+ has_call = FALSE;
+ max = NULL;
+ for (call = fun->call_list; call; call = call->next)
+ {
+ if (call->broken_cycle)
+ continue;
+ if (!call->is_pasted)
+ has_call = TRUE;
+ if (!sum_stack (call->fun, info, sum_stack_param))
+ return FALSE;
+ stack = sum_stack_param->cum_stack;
+ /* Include caller stack for normal calls, don't do so for
+ tail calls. fun->stack here is local stack usage for
+ this function. */
+ if (!call->is_tail || call->is_pasted || call->fun->start != NULL)
+ stack += fun->stack;
+ if (cum_stack < stack)
+ {
+ cum_stack = stack;
+ max = call->fun;
+ }
+ }
+
+ sum_stack_param->cum_stack = cum_stack;
+ stack = fun->stack;
+ /* Now fun->stack holds cumulative stack. */
+ fun->stack = cum_stack;
+ fun->visit3 = TRUE;
+
+ if (!fun->non_root
+ && sum_stack_param->overall_stack < cum_stack)
+ sum_stack_param->overall_stack = cum_stack;
+
+ htab = spu_hash_table (info);
+ if (htab->params->auto_overlay)
+ return TRUE;
+
+ f1 = func_name (fun);
+ if (htab->params->stack_analysis)
+ {
+ if (!fun->non_root)
+ info->callbacks->info (_(" %s: 0x%v\n"), f1, (bfd_vma) cum_stack);
+ info->callbacks->minfo (_("%s: 0x%v 0x%v\n"),
+ f1, (bfd_vma) stack, (bfd_vma) cum_stack);
+
+ if (has_call)
+ {
+ info->callbacks->minfo (_(" calls:\n"));
+ for (call = fun->call_list; call; call = call->next)
+ if (!call->is_pasted && !call->broken_cycle)
+ {
+ const char *f2 = func_name (call->fun);
+ const char *ann1 = call->fun == max ? "*" : " ";
+ const char *ann2 = call->is_tail ? "t" : " ";
+
+ info->callbacks->minfo (_(" %s%s %s\n"), ann1, ann2, f2);
+ }
+ }
+ }
+
+ if (sum_stack_param->emit_stack_syms)
+ {
+ char *name = bfd_malloc (18 + strlen (f1));
+ struct elf_link_hash_entry *h;
+
+ if (name == NULL)
+ return FALSE;
+
+ if (fun->global || ELF_ST_BIND (fun->u.sym->st_info) == STB_GLOBAL)
+ sprintf (name, "__stack_%s", f1);
+ else
+ sprintf (name, "__stack_%x_%s", fun->sec->id & 0xffffffff, f1);
+
+ h = elf_link_hash_lookup (&htab->elf, name, TRUE, TRUE, FALSE);
+ free (name);
+ if (h != NULL
+ && (h->root.type == bfd_link_hash_new
+ || h->root.type == bfd_link_hash_undefined
+ || h->root.type == bfd_link_hash_undefweak))
+ {
+ h->root.type = bfd_link_hash_defined;
+ h->root.u.def.section = bfd_abs_section_ptr;
+ h->root.u.def.value = cum_stack;
+ h->size = 0;
+ h->type = 0;
+ h->ref_regular = 1;
+ h->def_regular = 1;
+ h->ref_regular_nonweak = 1;
+ h->forced_local = 1;
+ h->non_elf = 0;
+ }
+ }
+
+ return TRUE;
+}
+
+/* SEC is part of a pasted function. Return the call_info for the
+ next section of this function. */
+
+static struct call_info *
+find_pasted_call (asection *sec)
+{
+ struct _spu_elf_section_data *sec_data = spu_elf_section_data (sec);
+ struct spu_elf_stack_info *sinfo = sec_data->u.i.stack_info;
+ struct call_info *call;
+ int k;
+
+ for (k = 0; k < sinfo->num_fun; ++k)
+ for (call = sinfo->fun[k].call_list; call != NULL; call = call->next)
+ if (call->is_pasted)
+ return call;
+ abort ();
+ return 0;
+}
+
+/* qsort predicate to sort bfds by file name. */
+
+static int
+sort_bfds (const void *a, const void *b)
+{
+ bfd *const *abfd1 = a;
+ bfd *const *abfd2 = b;
+
+ return filename_cmp ((*abfd1)->filename, (*abfd2)->filename);
+}
+
+static unsigned int
+print_one_overlay_section (FILE *script,
+ unsigned int base,
+ unsigned int count,
+ unsigned int ovlynum,
+ unsigned int *ovly_map,
+ asection **ovly_sections,
+ struct bfd_link_info *info)
+{
+ unsigned int j;
+
+ for (j = base; j < count && ovly_map[j] == ovlynum; j++)
+ {
+ asection *sec = ovly_sections[2 * j];
+
+ if (fprintf (script, " %s%c%s (%s)\n",
+ (sec->owner->my_archive != NULL
+ ? sec->owner->my_archive->filename : ""),
+ info->path_separator,
+ sec->owner->filename,
+ sec->name) <= 0)
+ return -1;
+ if (sec->segment_mark)
+ {
+ struct call_info *call = find_pasted_call (sec);
+ while (call != NULL)
+ {
+ struct function_info *call_fun = call->fun;
+ sec = call_fun->sec;
+ if (fprintf (script, " %s%c%s (%s)\n",
+ (sec->owner->my_archive != NULL
+ ? sec->owner->my_archive->filename : ""),
+ info->path_separator,
+ sec->owner->filename,
+ sec->name) <= 0)
+ return -1;
+ for (call = call_fun->call_list; call; call = call->next)
+ if (call->is_pasted)
+ break;
+ }
+ }
+ }
+
+ for (j = base; j < count && ovly_map[j] == ovlynum; j++)
+ {
+ asection *sec = ovly_sections[2 * j + 1];
+ if (sec != NULL
+ && fprintf (script, " %s%c%s (%s)\n",
+ (sec->owner->my_archive != NULL
+ ? sec->owner->my_archive->filename : ""),
+ info->path_separator,
+ sec->owner->filename,
+ sec->name) <= 0)
+ return -1;
+
+ sec = ovly_sections[2 * j];
+ if (sec->segment_mark)
+ {
+ struct call_info *call = find_pasted_call (sec);
+ while (call != NULL)
+ {
+ struct function_info *call_fun = call->fun;
+ sec = call_fun->rodata;
+ if (sec != NULL
+ && fprintf (script, " %s%c%s (%s)\n",
+ (sec->owner->my_archive != NULL
+ ? sec->owner->my_archive->filename : ""),
+ info->path_separator,
+ sec->owner->filename,
+ sec->name) <= 0)
+ return -1;
+ for (call = call_fun->call_list; call; call = call->next)
+ if (call->is_pasted)
+ break;
+ }
+ }
+ }
+
+ return j;
+}
+
+/* Handle --auto-overlay. */
+
+static void
+spu_elf_auto_overlay (struct bfd_link_info *info)
+{
+ bfd *ibfd;
+ bfd **bfd_arr;
+ struct elf_segment_map *m;
+ unsigned int fixed_size, lo, hi;
+ unsigned int reserved;
+ struct spu_link_hash_table *htab;
+ unsigned int base, i, count, bfd_count;
+ unsigned int region, ovlynum;
+ asection **ovly_sections, **ovly_p;
+ unsigned int *ovly_map;
+ FILE *script;
+ unsigned int total_overlay_size, overlay_size;
+ const char *ovly_mgr_entry;
+ struct elf_link_hash_entry *h;
+ struct _mos_param mos_param;
+ struct _uos_param uos_param;
+ struct function_info dummy_caller;
+
+ /* Find the extents of our loadable image. */
+ lo = (unsigned int) -1;
+ hi = 0;
+ for (m = elf_seg_map (info->output_bfd); m != NULL; m = m->next)
+ if (m->p_type == PT_LOAD)
+ for (i = 0; i < m->count; i++)
+ if (m->sections[i]->size != 0)
+ {
+ if (m->sections[i]->vma < lo)
+ lo = m->sections[i]->vma;
+ if (m->sections[i]->vma + m->sections[i]->size - 1 > hi)
+ hi = m->sections[i]->vma + m->sections[i]->size - 1;
+ }
+ fixed_size = hi + 1 - lo;
+
+ if (!discover_functions (info))
+ goto err_exit;
+
+ if (!build_call_tree (info))
+ goto err_exit;
+
+ htab = spu_hash_table (info);
+ reserved = htab->params->auto_overlay_reserved;
+ if (reserved == 0)
+ {
+ struct _sum_stack_param sum_stack_param;
+
+ sum_stack_param.emit_stack_syms = 0;
+ sum_stack_param.overall_stack = 0;
+ if (!for_each_node (sum_stack, info, &sum_stack_param, TRUE))
+ goto err_exit;
+ reserved = (sum_stack_param.overall_stack
+ + htab->params->extra_stack_space);
+ }
+
+ /* No need for overlays if everything already fits. */
+ if (fixed_size + reserved <= htab->local_store
+ && htab->params->ovly_flavour != ovly_soft_icache)
+ {
+ htab->params->auto_overlay = 0;
+ return;
+ }
+
+ uos_param.exclude_input_section = 0;
+ uos_param.exclude_output_section
+ = bfd_get_section_by_name (info->output_bfd, ".interrupt");
+
+ ovly_mgr_entry = "__ovly_load";
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ ovly_mgr_entry = "__icache_br_handler";
+ h = elf_link_hash_lookup (&htab->elf, ovly_mgr_entry,
+ FALSE, FALSE, FALSE);
+ if (h != NULL
+ && (h->root.type == bfd_link_hash_defined
+ || h->root.type == bfd_link_hash_defweak)
+ && h->def_regular)
+ {
+ /* We have a user supplied overlay manager. */
+ uos_param.exclude_input_section = h->root.u.def.section;
+ }
+ else
+ {
+ /* If no user overlay manager, spu_elf_load_ovl_mgr will add our
+ builtin version to .text, and will adjust .text size. */
+ fixed_size += (*htab->params->spu_elf_load_ovl_mgr) ();
+ }
+
+ /* Mark overlay sections, and find max overlay section size. */
+ mos_param.max_overlay_size = 0;
+ if (!for_each_node (mark_overlay_section, info, &mos_param, TRUE))
+ goto err_exit;
+
+ /* We can't put the overlay manager or interrupt routines in
+ overlays. */
+ uos_param.clearing = 0;
+ if ((uos_param.exclude_input_section
+ || uos_param.exclude_output_section)
+ && !for_each_node (unmark_overlay_section, info, &uos_param, TRUE))
+ goto err_exit;
+
+ bfd_count = 0;
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ ++bfd_count;
+ bfd_arr = bfd_malloc (bfd_count * sizeof (*bfd_arr));
+ if (bfd_arr == NULL)
+ goto err_exit;
+
+ /* Count overlay sections, and subtract their sizes from "fixed_size". */
+ count = 0;
+ bfd_count = 0;
+ total_overlay_size = 0;
+ for (ibfd = info->input_bfds; ibfd != NULL; ibfd = ibfd->link.next)
+ {
+ extern const bfd_target spu_elf32_vec;
+ asection *sec;
+ unsigned int old_count;
+
+ if (ibfd->xvec != &spu_elf32_vec)
+ continue;
+
+ old_count = count;
+ for (sec = ibfd->sections; sec != NULL; sec = sec->next)
+ if (sec->linker_mark)
+ {
+ if ((sec->flags & SEC_CODE) != 0)
+ count += 1;
+ fixed_size -= sec->size;
+ total_overlay_size += sec->size;
+ }
+ else if ((sec->flags & (SEC_ALLOC | SEC_LOAD)) == (SEC_ALLOC | SEC_LOAD)
+ && sec->output_section->owner == info->output_bfd
+ && strncmp (sec->output_section->name, ".ovl.init", 9) == 0)
+ fixed_size -= sec->size;
+ if (count != old_count)
+ bfd_arr[bfd_count++] = ibfd;
+ }
+
+ /* Since the overlay link script selects sections by file name and
+ section name, ensure that file names are unique. */
+ if (bfd_count > 1)
+ {
+ bfd_boolean ok = TRUE;
+
+ qsort (bfd_arr, bfd_count, sizeof (*bfd_arr), sort_bfds);
+ for (i = 1; i < bfd_count; ++i)
+ if (filename_cmp (bfd_arr[i - 1]->filename, bfd_arr[i]->filename) == 0)
+ {
+ if (bfd_arr[i - 1]->my_archive == bfd_arr[i]->my_archive)
+ {
+ if (bfd_arr[i - 1]->my_archive && bfd_arr[i]->my_archive)
+ info->callbacks->einfo (_("%s duplicated in %s\n"),
+ bfd_arr[i]->filename,
+ bfd_arr[i]->my_archive->filename);
+ else
+ info->callbacks->einfo (_("%s duplicated\n"),
+ bfd_arr[i]->filename);
+ ok = FALSE;
+ }
+ }
+ if (!ok)
+ {
+ info->callbacks->einfo (_("sorry, no support for duplicate "
+ "object files in auto-overlay script\n"));
+ bfd_set_error (bfd_error_bad_value);
+ goto err_exit;
+ }
+ }
+ free (bfd_arr);
+
+ fixed_size += reserved;
+ fixed_size += htab->non_ovly_stub * ovl_stub_size (htab->params);
+ if (fixed_size + mos_param.max_overlay_size <= htab->local_store)
+ {
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ {
+ /* Stubs in the non-icache area are bigger. */
+ fixed_size += htab->non_ovly_stub * 16;
+ /* Space for icache manager tables.
+ a) Tag array, one quadword per cache line.
+ - word 0: ia address of present line, init to zero. */
+ fixed_size += 16 << htab->num_lines_log2;
+ /* b) Rewrite "to" list, one quadword per cache line. */
+ fixed_size += 16 << htab->num_lines_log2;
+ /* c) Rewrite "from" list, one byte per outgoing branch (rounded up
+ to a power-of-two number of full quadwords) per cache line. */
+ fixed_size += 16 << (htab->fromelem_size_log2
+ + htab->num_lines_log2);
+ /* d) Pointer to __ea backing store (toe), 1 quadword. */
+ fixed_size += 16;
+ }
+ else
+ {
+ /* Guess number of overlays. Assuming overlay buffer is on
+ average only half full should be conservative. */
+ ovlynum = (total_overlay_size * 2 * htab->params->num_lines
+ / (htab->local_store - fixed_size));
+ /* Space for _ovly_table[], _ovly_buf_table[] and toe. */
+ fixed_size += ovlynum * 16 + 16 + 4 + 16;
+ }
+ }
+
+ if (fixed_size + mos_param.max_overlay_size > htab->local_store)
+ info->callbacks->einfo (_("non-overlay size of 0x%v plus maximum overlay "
+ "size of 0x%v exceeds local store\n"),
+ (bfd_vma) fixed_size,
+ (bfd_vma) mos_param.max_overlay_size);
+
+ /* Now see if we should put some functions in the non-overlay area. */
+ else if (fixed_size < htab->params->auto_overlay_fixed)
+ {
+ unsigned int max_fixed, lib_size;
+
+ max_fixed = htab->local_store - mos_param.max_overlay_size;
+ if (max_fixed > htab->params->auto_overlay_fixed)
+ max_fixed = htab->params->auto_overlay_fixed;
+ lib_size = max_fixed - fixed_size;
+ lib_size = auto_ovl_lib_functions (info, lib_size);
+ if (lib_size == (unsigned int) -1)
+ goto err_exit;
+ fixed_size = max_fixed - lib_size;
+ }
+
+ /* Build an array of sections, suitably sorted to place into
+ overlays. */
+ ovly_sections = bfd_malloc (2 * count * sizeof (*ovly_sections));
+ if (ovly_sections == NULL)
+ goto err_exit;
+ ovly_p = ovly_sections;
+ if (!for_each_node (collect_overlays, info, &ovly_p, TRUE))
+ goto err_exit;
+ count = (size_t) (ovly_p - ovly_sections) / 2;
+ ovly_map = bfd_malloc (count * sizeof (*ovly_map));
+ if (ovly_map == NULL)
+ goto err_exit;
+
+ memset (&dummy_caller, 0, sizeof (dummy_caller));
+ overlay_size = (htab->local_store - fixed_size) / htab->params->num_lines;
+ if (htab->params->line_size != 0)
+ overlay_size = htab->params->line_size;
+ base = 0;
+ ovlynum = 0;
+ while (base < count)
+ {
+ unsigned int size = 0, rosize = 0, roalign = 0;
+
+ for (i = base; i < count; i++)
+ {
+ asection *sec, *rosec;
+ unsigned int tmp, rotmp;
+ unsigned int num_stubs;
+ struct call_info *call, *pasty;
+ struct _spu_elf_section_data *sec_data;
+ struct spu_elf_stack_info *sinfo;
+ unsigned int k;
+
+ /* See whether we can add this section to the current
+ overlay without overflowing our overlay buffer. */
+ sec = ovly_sections[2 * i];
+ tmp = align_power (size, sec->alignment_power) + sec->size;
+ rotmp = rosize;
+ rosec = ovly_sections[2 * i + 1];
+ if (rosec != NULL)
+ {
+ rotmp = align_power (rotmp, rosec->alignment_power) + rosec->size;
+ if (roalign < rosec->alignment_power)
+ roalign = rosec->alignment_power;
+ }
+ if (align_power (tmp, roalign) + rotmp > overlay_size)
+ break;
+ if (sec->segment_mark)
+ {
+ /* Pasted sections must stay together, so add their
+ sizes too. */
+ pasty = find_pasted_call (sec);
+ while (pasty != NULL)
+ {
+ struct function_info *call_fun = pasty->fun;
+ tmp = (align_power (tmp, call_fun->sec->alignment_power)
+ + call_fun->sec->size);
+ if (call_fun->rodata)
+ {
+ rotmp = (align_power (rotmp,
+ call_fun->rodata->alignment_power)
+ + call_fun->rodata->size);
+ if (roalign < rosec->alignment_power)
+ roalign = rosec->alignment_power;
+ }
+ for (pasty = call_fun->call_list; pasty; pasty = pasty->next)
+ if (pasty->is_pasted)
+ break;
+ }
+ }
+ if (align_power (tmp, roalign) + rotmp > overlay_size)
+ break;
+
+ /* If we add this section, we might need new overlay call
+ stubs. Add any overlay section calls to dummy_call. */
+ pasty = NULL;
+ sec_data = spu_elf_section_data (sec);
+ sinfo = sec_data->u.i.stack_info;
+ for (k = 0; k < (unsigned) sinfo->num_fun; ++k)
+ for (call = sinfo->fun[k].call_list; call; call = call->next)
+ if (call->is_pasted)
+ {
+ BFD_ASSERT (pasty == NULL);
+ pasty = call;
+ }
+ else if (call->fun->sec->linker_mark)
+ {
+ if (!copy_callee (&dummy_caller, call))
+ goto err_exit;
+ }
+ while (pasty != NULL)
+ {
+ struct function_info *call_fun = pasty->fun;
+ pasty = NULL;
+ for (call = call_fun->call_list; call; call = call->next)
+ if (call->is_pasted)
+ {
+ BFD_ASSERT (pasty == NULL);
+ pasty = call;
+ }
+ else if (!copy_callee (&dummy_caller, call))
+ goto err_exit;
+ }
+
+ /* Calculate call stub size. */
+ num_stubs = 0;
+ for (call = dummy_caller.call_list; call; call = call->next)
+ {
+ unsigned int stub_delta = 1;
+
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ stub_delta = call->count;
+ num_stubs += stub_delta;
+
+ /* If the call is within this overlay, we won't need a
+ stub. */
+ for (k = base; k < i + 1; k++)
+ if (call->fun->sec == ovly_sections[2 * k])
+ {
+ num_stubs -= stub_delta;
+ break;
+ }
+ }
+ if (htab->params->ovly_flavour == ovly_soft_icache
+ && num_stubs > htab->params->max_branch)
+ break;
+ if (align_power (tmp, roalign) + rotmp
+ + num_stubs * ovl_stub_size (htab->params) > overlay_size)
+ break;
+ size = tmp;
+ rosize = rotmp;
+ }
+
+ if (i == base)
+ {
+ info->callbacks->einfo (_("%B:%A%s exceeds overlay size\n"),
+ ovly_sections[2 * i]->owner,
+ ovly_sections[2 * i],
+ ovly_sections[2 * i + 1] ? " + rodata" : "");
+ bfd_set_error (bfd_error_bad_value);
+ goto err_exit;
+ }
+
+ while (dummy_caller.call_list != NULL)
+ {
+ struct call_info *call = dummy_caller.call_list;
+ dummy_caller.call_list = call->next;
+ free (call);
+ }
+
+ ++ovlynum;
+ while (base < i)
+ ovly_map[base++] = ovlynum;
+ }
+
+ script = htab->params->spu_elf_open_overlay_script ();
+
+ if (htab->params->ovly_flavour == ovly_soft_icache)
+ {
+ if (fprintf (script, "SECTIONS\n{\n") <= 0)
+ goto file_err;
+
+ if (fprintf (script,
+ " . = ALIGN (%u);\n"
+ " .ovl.init : { *(.ovl.init) }\n"
+ " . = ABSOLUTE (ADDR (.ovl.init));\n",
+ htab->params->line_size) <= 0)
+ goto file_err;
+
+ base = 0;
+ ovlynum = 1;
+ while (base < count)
+ {
+ unsigned int indx = ovlynum - 1;
+ unsigned int vma, lma;
+
+ vma = (indx & (htab->params->num_lines - 1)) << htab->line_size_log2;
+ lma = vma + (((indx >> htab->num_lines_log2) + 1) << 18);
+
+ if (fprintf (script, " .ovly%u ABSOLUTE (ADDR (.ovl.init)) + %u "
+ ": AT (LOADADDR (.ovl.init) + %u) {\n",
+ ovlynum, vma, lma) <= 0)
+ goto file_err;
+
+ base = print_one_overlay_section (script, base, count, ovlynum,
+ ovly_map, ovly_sections, info);
+ if (base == (unsigned) -1)
+ goto file_err;
+
+ if (fprintf (script, " }\n") <= 0)
+ goto file_err;
+
+ ovlynum++;
+ }
+
+ if (fprintf (script, " . = ABSOLUTE (ADDR (.ovl.init)) + %u;\n",
+ 1 << (htab->num_lines_log2 + htab->line_size_log2)) <= 0)
+ goto file_err;
+
+ if (fprintf (script, "}\nINSERT AFTER .toe;\n") <= 0)
+ goto file_err;
+ }
+ else
+ {
+ if (fprintf (script, "SECTIONS\n{\n") <= 0)
+ goto file_err;
+
+ if (fprintf (script,
+ " . = ALIGN (16);\n"
+ " .ovl.init : { *(.ovl.init) }\n"
+ " . = ABSOLUTE (ADDR (.ovl.init));\n") <= 0)
+ goto file_err;
+
+ for (region = 1; region <= htab->params->num_lines; region++)
+ {
+ ovlynum = region;
+ base = 0;
+ while (base < count && ovly_map[base] < ovlynum)
+ base++;
+
+ if (base == count)
+ break;
+
+ if (region == 1)
+ {
+ /* We need to set lma since we are overlaying .ovl.init. */
+ if (fprintf (script,
+ " OVERLAY : AT (ALIGN (LOADADDR (.ovl.init) + SIZEOF (.ovl.init), 16))\n {\n") <= 0)
+ goto file_err;
+ }
+ else
+ {
+ if (fprintf (script, " OVERLAY :\n {\n") <= 0)
+ goto file_err;
+ }
+
+ while (base < count)
+ {
+ if (fprintf (script, " .ovly%u {\n", ovlynum) <= 0)
+ goto file_err;
+
+ base = print_one_overlay_section (script, base, count, ovlynum,
+ ovly_map, ovly_sections, info);
+ if (base == (unsigned) -1)
+ goto file_err;
+
+ if (fprintf (script, " }\n") <= 0)
+ goto file_err;
+
+ ovlynum += htab->params->num_lines;
+ while (base < count && ovly_map[base] < ovlynum)
+ base++;
+ }
+
+ if (fprintf (script, " }\n") <= 0)
+ goto file_err;
+ }
+
+ if (fprintf (script, "}\nINSERT BEFORE .text;\n") <= 0)
+ goto file_err;
+ }
+
+ free (ovly_map);
+ free (ovly_sections);
+
+ if (fclose (script) != 0)
+ goto file_err;
+
+ if (htab->params->auto_overlay & AUTO_RELINK)
+ (*htab->params->spu_elf_relink) ();
+
+ xexit (0);
+
+ file_err:
+ bfd_set_error (bfd_error_system_call);
+ err_exit:
+ info->callbacks->einfo ("%F%P: auto overlay error: %E\n");
+ xexit (1);
projects / deliverable / binutils-gdb.git / blobdiff