* GCC 3.3 compile speed regression - THE GRAPH
@ 2003-02-07 20:51 Ziemowit Laski
2003-02-07 22:08 ` Phil Edwards
` (2 more replies)
0 siblings, 3 replies; 24+ messages in thread
From: Ziemowit Laski @ 2003-02-07 20:51 UTC (permalink / raw)
To: gcc
[-- Attachment #1: Type: text/plain, Size: 397 bytes --]
The following graph shows the deterioration in compile-time performance
as seen on the gcc-3_3-branch (and, prior to the creation of the branch,
on the trunk itself). The test case is the same Qt fragment I used for
my earlier GC analysis; the test machine (a RedHat Linux 8.0 box) is
also the same. You can find the preprocessed source at:
http://homepage.mac.com/zlaski/FileSharing1.html
[-- Attachment #2: gcc_regress_chart.pdf --]
[-- Type: application/pdf, Size: 23179 bytes --]
[-- Attachment #3: Type: text/plain, Size: 65853 bytes --]
The two "inflection points" occur on 2002-06-04 and 2002-11-07,
respectively. They correspond to the following
two check-ins, with the ChangeLog entries reproduced in their gore and
glory below.
I've already spoken with Geoff regarding his fateful PCH merge. He did
not think
that doing a binary search on the pch-branch would be particularly
productive,
but instead suggested I look closely at his spew.c mods, which I shall
do now.
However, a more complete investigation of the PCH changes that were
brought in
would be beneficial.
Similarly, if Mark and/or others could take a look at the slow-down in
template parameter processing, the world will be a better place. :-)
--Zem
======================================================================
2002-06-03 Geoffrey Keating <geoffk@redhat.com>
Merge from pch-branch:
* gengtype.h (UNION_OR_STRUCT_P): New macro.
* gengtype.c (write_gc_structure_fields): Use it.
(write_gc_root): Use it here too.
* gengtype.c (write_gc_structure_fields): Assume that lengths
of typenames fit into an 'int'; don't pass a size_t to "%d" in
printf.
(write_gc_marker_routine_for_structure): Likewise.
(write_gc_types): Likewise.
(write_gc_root): Likewise.
* varray.h (VARRAY_CLEAR): New.
(varray_clear): Prototype.
* varray.c (varray_clear): New.
* cselib.c (reg_values_old): New.
(used_regs_old): New.
(cselib_init): Use cached varrays if available to avoid
generating large amounts of garbage.
(cselib_finish): Don't throw away old varrays.
* final.c (insn_addresses_): Move out of ifdef.
* varray.c (uses_ggc): Make more varray kinds GCable.
* varray.h (union varray_data_tag): Let gengtype see
fields 'generic' and 'te'.
* reg-stack.c: Include gt-reg-stack.h, ggc.h.
(stack_regs_mentioned_data): Move out of ifdef; mark with
gengtype.
(reg_to_stack): Don't call VARRAY_FREE.
* insn-addr.h (INSN_ADDRESSES_FREE): Don't use VARRAY_FREE.
(insn_addresses_): Use gengtype to mark.
* gengtype.c (write_gc_structure_fields): Handle arrays of
generic
pointers; handle generic pointers in unused union fields.
(get_output_file_with_visibility): Include cselib.h,
insn-addr.h in gtype-desc.c.
* function.c (prologue): Use gengtype to mark.
(epilogue): Likewise.
(sibcall_epilogue): Likewise.
* dependence.c: Include gt-dependence.h, ggc.h.
(struct def_use): Use gengtype to mark.
(struct loop): Likewise.
(struct induction): Likewise.
(struct dependence): Likewise.
(def_use_chain): Likewise.
(dep_chain): Likewise.
(loop_chain): Likewise.
(induction_chain): Likewise.
(init_dependence_analysis): Don't free anything, just clear
pointers.
(build_def_use): Use GGC to allocate def_use.
(add_loop): Use GGC to allocate loop.
(find_induction_variable): Use GGC to allocate induction.
(check_node_dependence): Use GGC to allocate induction,
dependence.
(dump_node_dependence): Don't free varrays.
(end_dependence_analysis): Likewise.
* cselib.h (struct cselib_val_struct): Use gengtype to mark.
(struct elt_loc_list): Likewise.
(struct elt_list): Likewise.
* cselib.c: Don't include obstack.h.
(hash_table): Use gengtype to mark.
(reg_values): Use gengtype to mark.
(used_regs): Use gengtype to mark.
(cselib_obstack): Delete.
(cselib_startobj): Delete.
(empty_vals): Mark as deletable.
(empty_elt_lists): Mark as deletable.
(empty_elt_loc_lists): Mark as deletable.
(new_elt_list): Use GGC to allocate struct elt_list.
(new_elt_loc_list): Use GGC to allocate struct elt_loc_list.
(clear_table): Don't delete obstack; don't unnecessarily clear
deletable variables.
(new_cselib_val): Use GGC to allocate struct cselib_val.
(cselib_init): Don't set up obstacks. Use GGC to allocate
hash table.
(cselib_finish): Just clear variables, don't free anything.
* Makefile.in (cselib.o): Remove dependency on $(OBSTACK_H).
(reg-stack.o): Add dependency on gt-reg-stack.h, $(GGC_H).
(dependence.o): Add dependency on gt-dependence.h, $(GGC_H).
(GTFILES): Add insn-addr.h, cselib.h, dependence.c, reg-stack.c.
(gt-reg-stack.h): New rule.
(gt-dependence.h): New rule.
(gtype-desc.o): Add cselib.h, insn-addr.h.
* varray.c: Use only necessary headers.
(element_size): New.
(uses_ggc): New.
(varray_init): Take type, not size. Use GGC if appropriate.
(varray_grow): Update for change to struct varray_head_tag.
Use GGC if appropriate.
* varray.h (struct const_equiv_data): Use gengtype.
(enum varray_data_enum): New.
(union varray_data_tag): Use gengtype.
(struct varray_head_tag): Use gengtype. Replace size field with
enum varray_data_enum.
(varray_init): Update prototype.
(VARRAY_SCHED_INIT): Delete.
(VARRAY_*_INIT): Update for change to varray_init.
(VARRAY_SCHED): Delete.
(VARRAY_PUSH_SCHED): Delete.
(VARRAY_TOP_SCHED): Delete.
* tree.h: Update for change to length specifier.
* tree-inline.c (optimize_inline_calls): Don't use VARRAY_FREE.
(clone_body): Likewise.
* ssa.h (ssa_definition): Use gengtype to mark.
* ssa.c (convert_from_ssa): Don't use VARRAY_FREE.
* ssa-dce.c (ssa_eliminate_dead_code): Don't use VARRAY_FREE.
* rtl.h (struct rtvec_def): Update for change to length
specifier.
* integrate.c (expand_inline_function): Don't use VARRAY_FREE.
(struct initial_value_struct): Update for change to length
specifier.
* ggc.h (ggc_add_rtx_varray_root): Delete prototype.
(ggc_add_tree_varray_root): Delete prototype.
(ggc_mark_rtx_varray): Delete prototype.
(ggc_mark_tree_varray): Delete prototype.
* ggc-common.c (ggc_add_rtx_varray_root): Delete.
(ggc_add_tree_varray_root): Delete.
(ggc_mark_rtx_varray): Delete.
(ggc_mark_tree_varray): Delete.
(ggc_mark_rtx_varray_ptr): Delete.
(ggc_mark_tree_varray_ptr): Delete.
* gengtype.h (enum typekind): Remove TYPE_VARRAY.
(create_varray): Delete prototype.
* gengtype.c (varrays): Delete.
(create_varray): Delete.
(adjust_field_type): Detect array of string pointers.
(process_gc_options): Remove code to handle varray_type option.
(set_gc_used_type): Remove TYPE_VARRAY case.
(output_escaped_param): New.
(write_gc_structure_fields): Use output_escaped_param on all
parameters. Handle 'skip' with 'use_param' option. Handle
arrays of strings. Remove TYPE_VARRAY handling.
(write_gc_roots): Use boolean to detect 'length' option.
* gengtype-yacc.y (VARRAY_TYPE): Delete token.
(struct_fields): Call adjust_field_type on array fields.
(type): Remove VARRAY_TYPE case.
(type_option): Likewise.
* gengtype-lex.l: Don't consider varray_type a keyword.
* function.h: Update for change to length specifier.
(free_eh_status): Delete prototype.
* function.c (free_after_compilation): Don't call
free_eh_status.
(reorder_blocks): Don't use VARRAY_FREE.
* except.c (struct eh_status): Update for change to length
specifier.
remove varray_type specifier.
(free_eh_status): Delete.
* dwarf2out.c: Include gt-dwarf2out.h.
(used_rtx_varray): Use gengtype to mark, move
outside ifdefs.
(incomplete_types): Likewise.
(decl_scope_table): Likewise.
(dwarf2out_init): Don't call ggc_add_tree_varray_root.
* cfglayout.c (scope_to_insns_finalize): Don't use VARRAY_FREE.
* c-tree.h (struct lang_type): Update for change to length
specifier.
* c-parse.in (yylexstring): Don't use VARRAY_FREE.
* c-objc-common.c: Include gt-c-objc-common.h.
(deferred_fns): Mark for gengtype.
(c_objc_common_init): Don't call ggc_add_tree_varray_root.
(expand_deferred_fns): Just set deferred_fns to 0 to free it.
* Makefile.in (c-objc-common.o): Add gt-c-objc-common.h.
(gtype-desc.o): Update dependencies.
(dwarf2out.o): Add gt-dwarf2out.h.
(varray.o): Update dependencies.
(GTFILES): Add varray.h, ssa.h, dwarf2out.c, c-objc-common.c.
(gt-c-objc-common.h): New rule.
(gt-dwarf2out.h): New rule.
* objc/objc-act.c (build_objc_string_object): Don't use
VARRAY_FREE.
* doc/gty.texi (GTY Options): Correct spelling.
(GGC Roots): Likewise.
* Makefile.in (TEXI_CPP_FILES): New.
(TEXI_GCC_FILES): New.
(TEXI_GCCINT_FILES): New.
(TEXI_CPPINT_FILES): New.
($(docdir)/cpp.info): Use new macros.
($(docdir)/gcc.info): Likewise.
($(docdir)/gccint.info): Likewise.
($(docdir)/cppinternals.info): Likewise.
(cpp.dvi): Likewise.
(gcc.dvi): Likewise.
(gccint.dvi): Likewise.
(cppinternals.dvi): Likewise.
* Makefile.in ($(docdir)/gccint.info): Depend on gty.texi.
* doc/gccint.texi (Top): Include gty.texi.
* doc/gty.texi: New file.
* bitmap.c: Include ggc.h, gt-bitmap.h.
(bitmap_ggc_free): New.
(bitmap_elem_to_freelist): New.
(bitmap_element_free): Use bitmap_elem_to_freelist.
(bitmap_element_allocate): Allow use of GGC.
(bitmap_clear): Use bitmap_elem_to_freelist.
(bitmap_copy): Update for change to bitmap_element_allocate.
(bitmap_set_bit): Likewise.
(bitmap_operation): Update for changes elsewhere.
(bitmap_initialize): Allow to create bitmaps that will use GGC.
* bitmap.h (struct bitmap_element_def): Use gengtype.
(struct bitmap_head_def): Likewise. Also add 'using_obstack'
field.
(bitmap_initialize): Add extra parameter.
(BITMAP_OBSTACK_ALLOC): Update for change to bitmap_initialize.
(BITMAP_ALLOCA): Delete.
(BITMAP_XMALLOC): Update for change to bitmap_initialize.
(BITMAP_GGC_ALLOC): New.
* Makefile.in (gtype-desc.o): Add bitmap.h.
(bitmap.o): Add gt-bitmap.h, $(GGC_H).
(GTFILES): Add bitmap.c.
(gt-bitmap.h): New rule.
($(HOST_PREFIX_1)bitmap.o): Add gt-bitmap.h.
* basic-block.h: Update for changes to bitmap_initialize.
* except.c (exception_handler_label_map): Move into...
(struct eh_status): Here.
(struct eh_region): Make 'aka' GCable.
(free_eh_status): Don't need to specially handle
exception_handler_label_map.
(add_ehl_entry): Update for changes to
exception_handler_label_map.
(find_exception_handler_labels): Likewise.
(remove_exception_handler_label): Likewise.
(maybe_remove_eh_handler): Likewise.
(for_each_eh_label): Likewise.
(remove_eh_handler): Allocate 'aka' using GGC.
* gengtype.c (get_output_file_with_visibility): Add bitmap.h
to list of includes.
* gengtype.c (write_gc_marker_routine_for_structure): Name
the routines 'gt_ggc_mx_*' instead of 'gt_ggc_m_*'.
(write_gc_types): Arrange for the tests with NULL to be inlined.
(write_gc_roots): Update uses of procedure pointers.
* ggc-common.c (gt_ggc_mx_rtx_def): Rename from
gt_ggc_m_rtx_def.
* Makefile.in (explow.o): Add dependency on gt-explow.h.
(sdbout.o): Add dependency on gt-sdbout.h.
* emit-rtl.c (const_int_htab): Use gengtype to clear unused
entries.
(mem_attrs_htab): Likewise.
(init_emit_once): Don't call ggc_add_deletable_htab.
* fold-const.c (size_htab): Use gengtype to clear unused
entries.
(size_int_type_wide): Don't call ggc_add_deletable_htab.
* gengtype.c (finish_root_table): Add LASTNAME and TNAME
parameters, use them, change callers.
(write_gc_root): Add IF_MARKED parameter, use it, change
callers.
(write_gc_roots): Handle 'if_marked' option.
(main): Don't need to call set_gc_used_type any more.
* ggc.h (ggc_htab_marked_p): Delete.
(ggc_htab_mark): Delete.
(struct ggc_cache_tab): New.
(gt_ggc_cache_rtab): New declaration.
* ggc-common.c (struct d_htab_root): Delete.
(d_htab_roots): Delete.
(ggc_add_deletable_htab): Delete.
(ggc_htab_delete): Handle new htab-deleting mechanism.
(ggc_mark_roots): Use new htab-deleting mechanism.
* tree.c (type_hash_table): Use gengtype to clear unused
entries.
Make static.
(init_obstacks): Don't call ggc_add_deletable_htab.
* objc/objc-act.h (struct hashed_attribute): Use gengtype.
(struct hashed_entry): Likewise.
(nst_method_hash_list): Likewise.
(cls_method_hash_list): Likewise.
(HASH_ALLOC_LIST_SIZE): Delete.
(ATTR_ALLOC_LIST_SIZE): Delete.
* objc/objc-act.c (hash_init): Use ggc to allocate
nst_method_hash_list, cls_method_hash_list.
(hash_enter): Use ggc to allocate; allocate one entry at a time.
(hash_add_attr): Likewise.
(ggc_mark_hash_table): Delete.
(objc_act_parse_init): Delete.
(objc_init): Delete reference to objc_act_parse_init.
* tlink.c: Replace hash.h with hashtab.h. Explicitly include
obstack.h. Replace references to 'struct hash_table' with
htab_t.
(struct symbol_hash_entry): Replace hash header with key field.
(struct file_hash_entry): Replace hash header with key field.
(struct demangled_hash_entry): Replace hash header with key
field.
(hash_string_eq): New.
(hash_string_hash): New.
(symbol_hash_newfunc): Delete.
(symbol_hash_lookup): Modify to use htab_t.
(file_hash_newfunc): Delete.
(file_hash_lookup): Modify to use htab_t.
(demangled_hash_newfunc): Delete.
(demangled_hash_lookup): Modify to use htab_t.
(tlink_init): Modify to use htab_t.
* hash.h: Delete.
* hash.c: Delete.
* ggc.h: Delete forward structure declarations.
Delete prototypes for deleted functions.
* ggc-common.c: Don't include hash.h.
(ggc_add_tree_hash_table_root): Delete.
(ggc_mark_tree_hash_table_entry): Delete.
(ggc_mark_tree_hash_table): Delete.
(ggc_mark_tree_hash_table_ptr): Delete.
* gengtype.c (write_gc_structure_fields): Allow param_is option.
(write_gc_marker_routine_for_structure): Use visibility of
the parameter if there is one.
* function.c: Replace hash.h with hashtab.h. Replace references
to 'struct hash_table *' with htab_t.
(struct insns_for_mem_entry): Include a plain key.
(fixup_var_refs_insns_with_hash): Update to use htab_t.
(insns_for_mem_newfunc): Delete.
(insns_for_mem_hash): Update to use htab_t.
(insns_for_mem_comp): Likewise.
(insns_for_mem_walk): Likewise.
* c-lang.c: Include ggc.h.
* Makefile.in (OBJS): Remove hash.o.
(c-lang.o): Add GGC_H.
(COLLECT2_OBJS): Remove hash.o.
(tlink.o): Remove hash.h, add HASHTAB_H and OBSTACK_H.
(ggc-common.o): Remove hash.h.
(function.o): Remove hash.h, add HASHTAB_H.
(genautomata.o): Remove hash.h, add HASHTAB_H.
* varasm.c (mark_const_str_htab_1): Delete.
(mark_const_str_htab): Delete.
(const_str_htab_del): Delete.
(const_str_htab): Use gengtype to mark.
(init_varasm_once): Use gengtype to mark hashtables. Use GC to
allocate them.
* tree.c (mark_tree_hashtable_entry): Delete.
(mark_tree_hashtable): Delete.
* tree.h (mark_tree_hashtable): Delete prototype.
* ggc.h (ggc_test_and_set_mark): Treat (void *)1 like NULL.
(ggc_mark): Likewise.
(ggc_calloc): New.
(htab_create_ggc): New.
* ggc-common.c (ggc_calloc): New.
* gengtype.h (enum typekind): Add TYPE_PARAM_STRUCT.
(struct type): Add param_struct structure.
* gengtype.c (param_structs): New.
(adjust_field_type): Handle param_is option.
(set_gc_used_type): Handle TYPE_PARAM_STRUCT.
(get_output_file_with_visibility): Include hashtab.h in
gtype-desc.c.
(write_gc_structure_fields): Add new PARAM parameter. Update
callers. Handle use_param option. Handle TYPE_PARAM_STRUCT.
(write_gc_marker_routine_for_structure): Add new PARAM
parameter.
Use it to generate function name. Update callers.
(write_gc_types): Add new PARAM_STRUCTS parameter. Update
callers.
Process them.
(write_gc_roots): Handle TYPE_PARAM_STRUCT. Allow param_is
option.
(main): Define PTR as pointer-to-scalar. Don't specially
mark deferred_string or ehl_map_entry.
* gengtype-yacc.y (PARAM_IS): Add new token.
(externstatic): Use adjust_field_type.
(type_option): Add PARAM_IS.
* gengtype-lex.l: Add rule for typedef of function pointers.
Add rule for PARAM_IS.
(IWORD): Add size_t.
* except.c (exception_handler_label_map): Use gengtype to mark.
(type_to_runtime_map): Likewise.
(mark_ehl_map_entry): Delete.
(mark_ehl_map): Delete.
(init_eh): Use gengtype for roots; use GC to allocate hash
tables.
(t2r_mark_1): Delete.
(t2r_mark): Delete.
* Makefile.in (gtype-desc.o): Correct dependencies.
(GTFILES): Add hashtab.h.
(genautomata.o): Actually uses hashtab.h.
* Makefile.in (stringpool.o): Add $(GGC_H).
(dwarf2asm.o): Likewise.
(GTFILES): Add hashtable.h.
* c-common.h (struct c_common_identifier): Use gengtype.
* c-decl.h (c_mark_tree): Delete.
* c-lang.c (LANG_HOOKS_MARK_TREE): Delete.
* c-tree.h (struct lang_identifier): Use gengtype.
(union lang_tree_node): New.
(c_mark_tree): Delete prototype.
* dwarf2out.c [!DWARF2_DEBUGGING_INFO]: Define dummy
dwarf2_debug_hooks.
* gengtype-lex.l (IWORD): Allow 'bool'.
(ptr_alias): Match.
* gengtype-yacc.y (ALIAS): New token.
(type_option): New rule.
(option): Use type_option.
* gengtype.c (process_gc_options): New.
(set_gc_used_type): Use it.
(write_gc_structure_fields): Add 'bitmap' parameter, change
callers.
Add new variable 't' to hold the type of the field being
processed.
Add more error checking. Use UNION_P when looking at 'desc'
option.
Handle language-specific structures containing other
language-specific structures.
(write_gc_types): Handle 'ptr_alias' option.
(main): Don't need to specially output lang_type, lang_decl,
lang_id2.
* ggc-common.c (ggc_pending_trees): Delete.
(ggc_mark_roots): Don't manipulate ggc_pending_trees.
(ggc_mark_trees): Delete.
(gt_ggc_m_tree_node): Delete.
* ggc.h (ggc_pending_trees): Delete.
(ggc_mark_tree): Make alias of gt_ggc_m_tree_node.
* hashtable.h (ht_identifier): Use gengtype.
* langhooks-def.h (LANG_HOOKS_MARK_TREE): Delete.
* langhooks.h (struct lang_hooks): Delete mark_tree.
* sdbout.c [! SDB_DEBUGGING_INFO]: Define dummy sdb_debug_hooks
anyway.
* system.h: Poison LANG_HOOKS_MARK_TREE.
* tree.c (tree_node_structure): New.
* tree.h (struct tree_common): Use gengtype.
(struct tree_int_cst): Likewise.
(struct tree_real_cst): Likewise.
(struct tree_string): Likewise.
(struct tree_complex): Likewise.
(struct tree_vector): Likewise.
(struct tree_identifier): Likewise.
(struct tree_list): Likewise.
(struct tree_vec): Likewise.
(struct tree_exp): Likewise.
(struct tree_block): Likewise.
(struct tree_type): Likewise.
(struct tree_decl): Likewise.
(enum tree_structure_enum): New.
(union tree_node): Use gengtype, with an alias.
(tree_node_structure): Prototype.
* objc/objc-lang.c (LANG_HOOKS_MARK_TREE): Delete.
Merge to tag pch-merge-20020430. The LANG_HOOKS_FUNCTION_MARK
macro was deleted. The LANG_HOOKS_FUNCTION_FREE macro was
renamed
to LANG_HOOKS_FUNCTION_FINAL.
* Makefile.in (GTFILES): Add bitmap.h.
* except.c (struct eh_region): Mark field 'aka' to be skipped.
* config/alpha/alpha.c [TARGET_ABI_UNICOSMK]
(alpha_init_machine_status): Give proper type.
* Makefile.in (c-lang.o): Depend on gtype-c.h.
(optabs.o): Depend on gt-optabs.h.
(GTFILES): Add optabs.o.
(gt-optabs.h): Add rule.
* optabs.c: Include gt-optabs.h.
* gengtype.c (set_gc_used_type): Correct some errors in last
change.
(write_gc_structure_fields): If a field which should be NULL is
not, abort.
* c-pragma.c: Move struct align_stack and variable
alignment_stack
out from the ifdef.
* config/xtensa/t-xtensa: Add dependencies for gt-xtensa.h.
* config/xtensa/xtensa.c: Include gt-cris.h.
(struct machine_function): Use gengtype to mark.
* config/mmix/mmix.h (struct machine_function): Use gengtype
to mark.
* config/cris/t-cris: Add dependencies for gt-cris.h.
* config/cris/cris.c: Include gt-cris.h.
(struct machine_function): Use gengtype to mark.
* config/rs6000/rs6000.h (struct machine_function): Use gengtype
to mark.
* doc/tm.texi (Per-Function Data): Delete references to
mark_machine_status.
* config/ia64/ia64.c (ia64_override_options): Don't set
mark_machine_status.
* config/i386/i386.c (override_options): Likewise.
* config/d30v/d30v.c (d30v_init_expanders): Likewise.
* config/arm/arm.c (arm_init_expanders): Likewise.
* config/alpha/alpha.c (override_options): Likewise.
* gengtype.h (enum gc_used_enum): Add GC_MAYBE_POINTED_TO.
* gengtype.c (set_gc_used_type): Handle 'maybe_null' option.
(write_gc_structure_fields): Don't handle 'really' option.
Handle 'maybe_null' option.
(write_gc_types): Handle 'maybe_null' option.
* function.h (struct function): Don't use "really".
(mark_machine_status): Delete declaration.
(mark_lang_status): Delete declaration.
(gt_ggc_mr_machine_function): Delete prototype.
(gt_ggc_mr_language_function): Delete prototype.
* function.c (mark_machine_status): Delete.
(mark_lang_status): Delete.
(gt_ggc_mr_machine_function): Delete.
(gt_ggc_mr_language_function): Delete.
* c-tree.h (mark_c_function_context): Delete prototype.
* c-objc-common.c (c_objc_common_init): Don't set
mark_lang_status.
* c-decl.c (struct language_function): Rename from struct
c_language_function. Update uses. Use gengtype to mark.
(mark_c_function_context): Delete.
* c-common.h (struct c_language_function): Rename from struct
language_function.
(mark_stmt_tree): Delete prototype.
(c_mark_lang_decl): Delete prototype.
(mark_c_language_function): Delete prototype.
* c-common.c (mark_stmt_tree): Delete.
(c_mark_lang_decl): Delete.
(mark_c_language_function): Delete.
* gengtype.h (enum typekind): Add TYPE_LANG_STRUCT.
(lang_bitmap): New typedef. Use where appropriate.
(struct type): Add gc_used field, lang_struct field.
(UNION_P): New macro.
(new_structure): New prototype.
(find_structure): Remove 'pos' parameter. Change all callers.
* gengtype-lex.l: Update for changes to find_structure.
* gengtype-yacc.y (typedef_struct): Use new_structure.
(yacc_ids): Suppress warning.
(type): Use new_structure.
* gengtype.c (string_type): Update for changes to struct type.
(find_structure): Just find a structure, don't worry about
creating one.
(new_structure): New.
(note_yacc_type): Use new_structure.
(set_gc_used_type): New.
(set_gc_used): New.
(write_gc_structure_fields): Allow for pointers to
TYPE_LANG_STRUCT.
(write_gc_types): Handle TYPE_LANG_STRUCT.
(write_gc_marker_routine_for_structure): New.
(main): Call set_gc_used. Add some calls to set_gc_used_type
for places where GCC doesn't use gengtype properly yet.
* ggc.h (gt_ggc_m_rtx_def): Don't prototype.
(gt_ggc_m_tree_node): Likewise.
* varasm.c (copy_constant): Call expand_constant if we hit
something we can't recognise.
* ggc-common.c (ggc_mark_rtvec_children): Delete.
(ggc_mark_rtx_children): Use generic name for ggc_mark_rtvec.
(lang_mark_false_label_stack): Delete.
* rtl.h (struct rtvec_def): Use gengtype to mark.
* ggc.h (ggc_mark_rtvec): Delete.
(gt_ggc_m_rtvec_def): Delete.
(ggc_mark_nonnull_tree): Delete.
(ggc_mark_rtvec_children): Delete prototype.
(lang_mark_false_label_stack): Delete declaration.
* gengtype.h (note_yacc_type): Add prototype.
* gengtype.c (note_yacc_type): New function.
* gengtype-lex.l: Add lexer support for yacc files.
* gengtype-yacc.y (start): Extract union from yacc files.
(yacc_union): New rule.
(yacc_typematch): New rule.
(yacc_ids): New rule.
(enum_items): Tweak for efficiency.
(optionseq): Likewise.
* c-common.h (struct language_function): Use gengtype.
(struct c_lang_decl): Likewise.
* c-tree.h (struct lang_decl): Likewise.
(struct lang_type): Likewise.
* c-decl.c (lang_mark_tree): Use generated marker routines to
mark
tree language substructures.
* stringpool.c (mark_ident): Replace ggc_mark_nonnull_tree with
ggc_mark_tree.
* dwarf2asm.c (mark_indirect_pool_entry): Likewise.
* varasm.c (struct rtx_const): Remove 'skip' tags for scalar
arrays.
* stmt.c (struct nesting): Add discriminator. Use gengtype to
mark. Remove 'data.block.cleanup_ptr' field.
(struct stmt_status): Use usual technique to mark struct
nesting.
(gt_ggc_mr_nesting_cond): Delete.
(gt_ggc_mr_nesting_loop): Delete.
(gt_ggc_mr_nesting_block): Delete.
(gt_ggc_mr_nesting_case_stmt): Delete.
(expand_start_cond): Set discriminator.
(expand_start_loop): Likewise.
(expand_start_null_loop): Likewise.
(expand_start_bindings_and_block): Set discriminator. Don't set
deleted fields.
(expand_decl_cleanup): Replace 'cleanup_ptr' with
&thisblock->data.block.cleanups.
(expand_start_case): Set discriminator.
(expand_start_case_dummy): Set discriminator.
* ggc-callbacks.c: Remove.
* gengtype.h (struct type): Add 'u.s.bitmap' field.
(find_structure): Add 'pos' parameter.
* gengtype-lex.l: Update callers to find_structure.
* gengtype-yacc.y: Likewise.
* gengtype.c (find_structure): Allow for structures to be
defined
in multiple language backends.
(get_output_file_with_visibility): Include debug.h in
gtype-desc.c.
(counter): Rename to gc_counter.
(write_gc_structure_fields): Fail when writing out fields for
an incomplete structure. Ignore arrays of scalars. Handle
'tree_vec' special.
(write_gc_types): Reset counter for each procedure written.
* stmt.c (add_case_node): Use GGC to allocate struct case_node.
(free_case_nodes): Delete.
(expand_end_case_type): Delete call to free_case_nodes.
* Makefile.in (cselib.o): Include gt-<filename>.h.
(gcse.o): Likewise.
(profile.o): Likewise.
(alias.o): Likewise.
(GTFILES): Add alias.c, cselib.c, gcse.c, profile.c, and
alphabetize backend files.
(gt-alias.h, gt-cselib.h, gt-gcse.h, gt-profile.h): New rules.
* alias.c: Use gengtype for roots.
* c-common.h (struct stmt_tree_s): Use gengtype.
* c-decl.c: Use gengtype for roots.
* cselib.c: Use gengtype for roots.
* expr.c: Use gengtype for roots.
* fold-const.c: Use gengtype for roots.
* gcse.c: Use gengtype for roots.
* gengtype-lex.l: Handle typedefs of function types.
Allow for empty array bounds.
Allow processing to stop on initialisers.
* gengtype-yacc.y (externstatic): Stop processing on
initialisers.
(semiequal): New rule.
* gengtype.c (create_file): Tidy output files.
(get_output_file_with_visibility): Fix paren warning. Fix bug
involving multiple input files mapping to one output file.
(write_gc_structure_fields): Skip arrays of scalars.
(write_gc_types): Tidy output files.
(write_gc_root): New function.
(write_gc_roots): Fix bugs, add support for roots that are
structures.
* ggc-common.c (ggc_mark_rtx_ptr): Delete.
(ggc_mark_tree_ptr): Delete.
(ggc_add_rtx_root): Delete.
(ggc_add_tree_root): Delete.
(ggc_del_root): Delete.
* integrate.c (get_func_hard_reg_initial_val): Use ggc_alloc to
allocate struct initial_value_struct.
* profile.c: Use gengtype for roots.
* sdbout.c: Use gengtype for roots.
* varasm.c (mark_weak_decls): Delete unused prototype.
(mark_const_hash_entry): Delete unused function.
* config/darwin-protos.h: Use gengtype for roots.
(machopic_add_gc_roots): Delete.
* config/arm/arm.c: Use gengtype for roots.
* config/arm/arm.h: Use gengtype for roots.
* config/c4x/c4x-protos.h: Use gengtype for roots.
* config/c4x/c4x.c (c4x_add_gc_roots): Delete.
* config/d30v/d30v-protos.h: Use gengtype for roots.
* config/d30v/d30v.c (d30v_add_gc_roots): Delete.
* config/dsp16xx/dsp16xx.c (override_options): Use gengtype for
roots.
* config/dsp16xx/dsp16xx.h: Use gengtype for roots.
* config/ia64/ia64-protos.h: Use gengtype for roots.
* config/ia64/ia64.c (ia64_add_gc_roots): Delete.
* config/m68hc11/m68hc11-protos.h: Use gengtype for roots.
* config/m68hc11/m68hc11.c (z_reg): Make global.
(z_reg_qi): Make global.
(m68hc11_add_gc_roots): Delete.
* config/mcore/mcore-protos.h: Use gengtype for roots.
* config/mcore/mcore.c (mcore_add_gc_roots): Delete.
* config/mips/mips.c (mips_add_gc_roots): Delete.
* config/mips/mips.h: Use gengtype for roots.
* config/mmix/mmix.c (override_options): Use gengtype for roots.
* config/mmix/mmix.h: Use gengtype for roots.
* config/mn10200/mn10200.c (asm_file_start): Use gengtype for
roots.
* config/mn10200/mn10200.h: Use gengtype for roots.
* config/pa/pa.c: Use gengtype for roots, marking.
(struct deferred_plabel): Use GGC, gengtype.
(pa_add_gc_roots): Delete.
(mark_deferred_plabels): Delete.
* config/pj/pj-protos.h: Use gengtype for roots.
* config/pj/pj.h (OVERRIDE_OPTIONS): Don't define.
* config/rs6000/rs6000.c: Use gengtype for roots. Don't call
machopic_add_gc_roots.
* config/rs6000/rs6000.h: Use gengtype for roots.
* config/rs6000/t-darwin (darwin.o): Add dependency on
gt-darwin.h.
(gt-darwin.h): Add rule.
* config/sh/sh.c: Use gengtype for roots.
* config/sh/t-sh ($(out_object_file)): Add dependency on
gt-sh.h.
(gt-sh.h): Add rule.
* config/sparc/sparc.c: Use gengtype for roots.
(sparc_add_gc_roots): Delete.
(struct ultrasparc_pipeline_state): Use GGC, gengtype.
(mark_ultrasparc_pipeline_state): Delete.
* config/sparc/sparc.h: Use gengtype for roots.
* Makefile.in (c-parse.o): Update dependencies.
(c-common.o): Likewise.
(GTFILES): Add c-common.h, c-tree.h, c-common.c, c-parse.in.
Add dependencies for the files they generate.
* c-common.c: Replace ggc_add_* uses with GTY annotations.
* c-common.h: Likewise.
* c-decl.c: Likewise.
(gt_ggc_mp_binding_level): Delete.
* c-lang.c: Include gtype-c.h.
* c-parse.in: Replace ggc_add_* uses with GTY annotations.
Include
gt-c-parse.h.
* c-pragma.h: Replace ggc_add_* uses with GTY annotations.
(gt_ggc_mp_align_stack): Delete.
* c-tree.h: Replace ggc_add_* uses with GTY annotations.
* function.c: Replace ggc_add_* uses with GTY annotations.
(gt_ggc_mp_function): Delete.
* function.h: Replace ggc_add_* uses with GTY annotations.
* gengtype.c (lang_names): New.
(NUM_BASE_FILES): New.
(open_base_files): Create language base files.
(startswith): New.
(get_file_basename): New.
(get_base_file_bitmap): New.
(get_output_file_with_visibility): Rename from get_output_file.
Add more mappings for various C/Objc filenames.
(finish_root_table): New.
(write_gc_roots): Handle dependencies and scoping properly.
* gengtype.h: Add prototypes for new functions.
* ggc-common.c (struct deletable_root): Delete.
(deletables): Delete.
(ggc_add_deletable_root): Delete.
(ggc_mark_roots): No need to deal with deleted functionality.
* ggc.h (ggc_add_deletable_root): Delete prototype.
* objc/Make-lang.in (objc-act.o): Add gtype-objc.h dependency.
(gtype-objc.h): Add rule to create.
* objc/config-lang.in (gtfiles): New.
* objc/objc-act.c: Allocate imp_list using GGC. Replace uses of
ggc_add_* with GTY markers. Include gtype-objc.h.
(ggc_mark_imp_list): Delete.
* objc/objc-act.h: Replace uses of ggc_add_* with GTY markers.
* objc/objc-lang.c: Random Whitespace Change.
* except.h (exception_handler_labels): Delete.
(get_exception_handler_labels): New.
* except.c (exception_handler_labels): Delete.
(struct eh_status): Add exception_handler_labels field.
(doing_eh): Don't add exception_handler_labels as root.
(free_eh_status): Don't need to free exception_handler_labels.
(get_exception_handler_labels): New.
(find_exception_handler_labels): Update for move of
exception_handler_labels.
(remove_exception_handler_label): Likewise.
* cfgrtl.c (can_delete_label_p): Use
get_exception_handler_labels.
* jump.c (rebuild_jump_labels): Likewise.
* loop.c (find_and_verify_loops): Likewise.
* sched-rgn.c (is_cfg_nonregular): Likewise.
* gengtype.c (write_gc_structure_fields): Handle variable-length
TYPE_ARRAYs.
* varasm.c (struct weak_syms): Use GGC, gengtype.
(mark_weak_decls): Delete.
(weak_decls): Likewise.
(add_weak): Likewise.
(remove_from_pending_weak_list): Likewise.
(init_varasm_once): Likewise.
* Makefile.in (gtype-desc.o): Add libfuncs.h dependency.
(GTFILES): Add tree.h, libfuncs.h, emit-rtl.c, explow.c,
stor-layout.c, regclass.c, and lists.c.
Add dependencies of gt-emit-rtl.h gt-explow.h gt-stor-layout.h
gt-regclass.h and gt-lists.h on s-gtype.
* emit-rtl.c: Use gengtype for roots. Include gt-emit-rtl.h.
* except.c: Use gengtype for roots.
* explow.c: Use gengtype for roots. Include gt-explow.h.
* expr.h (init_stor_layout_once): Delete prototype.
* function.c: Use gengtype for roots.
* gengtype-lex.l: Add ENT_EXTERNSTATIC lexing.
* gengtype-yacc.y (start): Can also be an externstatic.
(externstatic): New production.
(struct_fields): Correct array bounds inversion for 2-d arrays.
* gengtype.c (variables): New variable.
(note_variable): New function.
(get_output_file): Include libfuncs.h into gtype-desc.c.
(get_output_file_name): New function.
(write_gc_structure_fields): Suppress warnings.
(write_gc_types): Make static.
(put_mangled_filename): New function.
(write_gc_roots): New function.
(main): Call write_gc_roots.
* gengtype.h (note_variable): Prototype.
(get_output_file_name): Prototype.
(write_gc_types): Delete prototype.
* ggc.h: Clean up unnecessary structure predefinitions.
(struct ggc_root_tab): Define.
(gt_ggc_m_rtx_def): Make function, not macro.
(gt_ggc_m_tree_node): Likewise.
* libfuncs.h: Use gengtype for roots.
* lists.c: Use gengtype for roots. Include gt-lists.h.
(init_EXPR_INSN_LIST_cache): Delete.
* optabs.c: Use gengtype for roots.
(gt_ggc_mp_optab): Delete.
* optabs.h: Use gengtype for roots.
* regclass.c: Use gengtype for roots. Include gt-regclass.h.
* rtl.h: Use gengtype for roots.
(init_EXPR_INSN_LIST_cache): Delete prototype.
* stor-layout.c: Use gengtype for roots.
Include gt-stor-layout.h.
(init_stor_layout_once): Delete.
* toplev.c: Use gengtype for roots. Delete calls to deleted
routines.
* tree.c: Use gengtype for roots.
* tree.h: Use gengtype for roots.
* varasm.c: Use gengtype for roots.
* Makefile.in (GTFILES): Add @all_gtfiles@.
* configure: Regenerate.
* configure.in: Construct all_gtfiles from the gtfiles
definitions
in config-lang.in.
* gengtype-yacc.y (type): Warn about duplicate structure names.
* gengtype.c (get_output_file): Handle .c files in language
subdirectories.
* Makefile.in (GTFILES): Run gengtype on all the config files
and on the target .c file.
* except.c (mark_eh_region): Delete.
(init_eh_for_function): Use GGC on struct eh_status.
(mark_eh_status): Delete.
(free_eh_status): Use GGC.
(expand_eh_region_start): Use GGC to
(collect_eh_region_array): Allocate last_region_number using
GGC.
(duplicate_eh_region_1): Use GGC to allocate struct eh_region.
(remove_eh_handler): Let GGC free struct eh_region.
(add_call_site): Use GGC to reallocate call_site_record array.
* function.c (init_machine_status): Update calling sequence.
(mark_machine_status): Likewise.
(mark_lang_status): Likewise.
(prepare_function_start): Update init_machine_status call.
(mark_function_status): Delete.
(maybe_mark_struct_function): Delete.
(ggc_mark_struct_function): Delete.
(gt_ggc_mp_function): New.
(gt_ggc_mr_machine_function): New.
(gt_ggc_mr_language_function): New.
(init_function_once): Use canonical names.
* function.h (struct function): Use gengtype.
(init_machine_status): Return the structure.
(mark_machine_status): Take a 'void *'.
(mark_lang_status): Likewise.
* ggc-common.c (ggc_mark_trees): Use canonical name for
ggc_mark_struct_function.
* tree.h (ggc_mark_struct_function): Delete prototype.
* config/alpha/alpha.c (alpha_mark_machine_status): Delete.
(alpha_init_machine_status): Likewise.
(override_options): Use canonical name for
alpha_mark_machine_status.
* config/alpha/unicosmk.h (struct machine_function): Use
gengtype.
* config/arm/arm.h (struct machine_function): Use gengtype.
* config/arm/arm.c (arm_mark_machine_status): Delete.
(arm_init_machine_status): Update calling sequence.
(arm_init_expanders): Use canonical name for
arm_mark_machine_status.
* config/cris/cris.c (cris_init_machine_status): Update
calling sequence.
* config/d30v/d30v.h (struct machine_function): Use gengtype.
* config/d30v/d30v.c (d30v_init_machine_status): Update
calling sequence.
(d30v_mark_machine_status): Delete.
* config/i386/i386.c: Include gt-i386.h.
(struct machine_function): Use gengtype.
(ix86_init_machine_status): Update calling sequence.
(ix86_mark_machine_status): Delete.
(override_options): Use canonical namke for
ix86_mark_machine_status.
* config/ia64/ia64.h (struct machine_function): Use gengtype.
* config/ia64/ia64.c (ia64_init_machine_status): Update calling
sequence.
(ia64_mark_machine_status): Delete.
(ia64_override_options): Use canonical name for
ia64_mark_machine_status.
* config/mmix/mmix.c (mmix_init_machine_status): Update calling
sequence.
* config/rs6000/rs6000.c (rs6000_init_machine_status): Likewise.
* config/xtensa/xtensa.c (xtensa_init_machine_status): Likewise.
* gengtype.c (get_output_file): Fix warning.
(main): Add prototype to suppress warning.
* tree.c: Remove tree_hash_mark prototype.
* tree.h (init_stmt): Delete prototype.
* toplev.c (lang_independent_init): Don't call init_stmt.
* stmt.c (ALLOC_NESTING): Use GGC for 'struct nesting'.
(stmt_obstack): Delete.
(POPSTACK): No need to free 'struct nesting'.
(gt_ggc_mr_nesting_cond): Use canonical names.
(gt_ggc_mr_nesting_loop): Use canonical names.
(gt_ggc_mr_nesting_block): Use canonical names.
(gt_ggc_mr_nesting_case_stmt): Use canonical names.
(mark_stmt_status): Delete.
(init_stmt): Delete.
(clear_last_expr): Clear both last_expr_type and
last_expr_value.
Use it everywhere that last_expr_type was cleared.
* lists.c (init_EXPR_INSN_LIST_cache): Use
ggc_add_deletable_root.
(zap_lists): Delete.
* ggc.h (ggc_add_deletable_root): Prototype.
(mark_stmt_status): Remove prototype.
* ggc-common.c (ggc_add_deletable_root): New.
(ggc_mark_roots): Handle deletable roots.
* function.c (ggc_mark_struct_function): Use canonical name
for mark_stmt_status.
* emit-rtl.c (free_sequence_stack): New.
(start_sequence): Use a freelist for sequences.
(end_sequence): Likewise.
(init_emit_once): Add free_sequence_stack as a deleteable root.
* c-pragma.c Include gt-c-pragma.h.
(struct align_stack): Use gengtype.
(push_alignment): Use GGC for struct align_stack.
(mark_align_stack): Delete.
(gt_ggc_mp_align_stack): New.
(init_pragma): Use canonical name for mark_align_stack.
* c-decl.c: Include gt-c-decl.h.
(struct binding_level): Use gengtype.
(make_binding_level): Use GGC; handle the freelist here.
(pop_binding_level): New.
(pushlevel): Move code into make_binding_level.
(push_label_level): Likewise.
(poplevel): Move code into pop_binding_level.
(pop_label_level): Likewise.
(mark_binding_level): Delete.
(gt_ggc_mp_binding_level): New.
(c_init_decl_processing): Use canonical name for
mark_binding_level.
Add free_binding_level as deletable root.
(mark_c_function_context): Use canonical name for
mark_binding_level.
* Makefile.in (c-decl.o): Add gt-c-decl.h.
(c-pragma.o): Add gt-c-pragma.h.
(GTFILES): Add c-decl.c and c-pragma.c.
(gt-c-decl.h, gt-c-pragma.h): Create using gengtype.
* tree.c (struct type_hash): Use gengtype.
(init_obstacks): Use canonical name for type_hash_mark.
(type_hash_mark): Delete.
Include gt-tree.h.
* rtl.h (struct mem_attrs): Use gengtype.
* optabs.h (struct optab): Use gengtype.
* optabs.c (expand_binop): Squish signed/unsigned warning.
(mark_optab): Make local, use canonical name, use autogenerated
marker procedure.
(init_optabs): Use canonical name for mark_optab.
(new_optab): Use GGC to allocate optabs.
* ggc.h: Delete mark_optab prototype.
* ggc-common.c (ggc_mark_rtx_children): Use canonical name for
mem_attrs marker procedure.
* gengtype.c (get_output_file): Include headers in gtype-desc.c
explicitly rather than deducing them from file names.
(write_gc_structure_fields): Handle arrays of structures.
(main): Return non-zero exit code if errors occur during output.
* emit-rtl.c (mem_attrs_mark): Delete.
(init_emit_once): Use canonical name for mem_attrs marker
procedure.
* Makefile.in (gtype-desc.o): Explicitly name dependencies.
(tree.o): Depend on gt-tree.h.
(GTFILES): Add rtl.h, optabs.h, tree.c.
(gt-tree.h): Add it to s-gtype rule.
* .cvsignore: Ignore gengtype flex/bison generated files.
* Makefile.in (GGC_H): Add gtype-desc.h.
(OBJS): Add gtype-desc.o.
(GEN): Add gengtype.
(STAGESTUFF): Add gengtype.
(varasm.o): Add gt-varasm.h.
(stmt.o): Add gt-stmt.h.
(except.o): Add gt-except.h.
(integrate.o): Add gt-integrate.h.
(GTFILES): New.
Add new rules for new files.
* configure: Regenerate.
* configure.in: Correct defaults.h paths.
* emit-rtl.c (mark_sequence_stack): Delete.
(mark_emit_status): Delete.
(start_sequence): Allocate sequence structures using GGC.
(end_sequence): Allocate sequence structures using GGC.
* except.c: Use gengtype for various structures. Include
gt-except.h.
* expr.c (mark_expr_status): Delete.
* function.c: Use gengtype for various structures. Include
gt-function.h.
(mark_function_status): Use standard gt_ggc names for marker
functions.
(ggc_mark_struct_function): Likewise.
* function.h: Use gengtype for various structures.
* gengtype-lex.l: New file.
* gengtype-yacc.y: New file.
* gengtype.c: New file.
* gengtype.h: New file.
* ggc.h: Include gtype-desc.h. Alias some marker procedures to
the standard names. Remove some now-unnecessary prototypes.
* integrate.c: Use gengtype for various structures. Include
gt-integrate.h.
(mark_hard_reg_initial_vals): Delete.
* integrate.h (mark_hard_reg_initial_vals): Delete.
* stmt.c: Use gengtype for various structures. Include
gt-stmt.h.
(mark_case_node): Delete.
(mark_goto_fixup): Delete.
(mark_stmt_status): Use standard gt_ggc names for marker
functions.
* system.h: Define GTY to empty. In flex/bison files,
don't poison malloc or realloc, instead just define them to
xmalloc and xrealloc.
* varasm.c: Use gengtype for various structures. Include
gt-varasm.h. Use standard gt_ggc names for marker functions.
(mark_pool_constant): Delete.
(mark_varasm_status): Delete.
(decode_rtx_const): #if 0 out non-typesafe hack.
* function.h (free_lang_status): Mark as obsolete.
* function.c (free_lang_status): Mark as obsolete.
* c-decl.c (push_c_function_context): Use GC to allocate and
free
struct language_function.
(pop_c_function_context): Likewise.
* c-common.c (mark_c_language_function): Mark struct
language_function.
* doc/tm.texi (Per-Function Data): Don't document
free_machine_status.
Document that the machine_function structures must be allocated
using GC. Update mark_machine_status documentation.
* function.h: Don't declare free_machine_status.
* function.c (free_machine_status): Don't define.
(free_after_compilation): Don't call free_machine_status.
(ggc_mark_struct_function): Mark f->machine. Call
mark_machine_status only on non-NULL pointers.
* system.h: Poison free_machine_status.
* config/xtensa/xtensa.c (xtensa_init_machine_status): Use GC on
struct machine_function.
(xtensa_free_machine_status): Delete.
(override_options): Don't set free_machine_status.
* config/rs6000/rs6000.c (rs6000_override_options): Don't set
free_machine_status.
(rs6000_init_machine_status): Use GC on struct machine_function.
(rs6000_free_machine_status): Delete.
* config/ia64/ia64.c (ia64_init_machine_status): Use GC on
struct
machine_function.
(ia64_mark_machine_status): Likewise.
(ia64_free_machine_status): Delete.
(ia64_override_options): Don't set free_machine_status.
* config/i386/i386.c (override_options): Don't set
free_machine_status.
(ix86_init_machine_status): Use GC on struct machine_function.
(ix86_mark_machine_status): Likewise.
(ix86_free_machine_status): Delete.
* config/d30v/d30v.c: (d30v_init_machine_status): Use GC on
struct
machine_function.
(d30v_mark_machine_status): Likewise.
(d30v_free_machine_status): Delete.
(d30v_init_expanders): Don't set free_machine_status.
* config/arm/arm.c (arm_mark_machine_status): Use GC on struct
machine_function.
(arm_init_machine_status): Likewise.
(arm_free_machine_status): Delete.
(arm_init_expanders): Don't set free_machine_status.
* config/alpha/alpha.c (override_options): Don't set
free_machine_status.
(alpha_init_machine_status): Use GC on struct machine_function.
(alpha_mark_machine_status): Likewise.
(alpha_free_machine_status): Delete.
* varasm.c (compare_constant): Fix typo.
* varasm.c: Don't include obstack.h.
(struct varasm_status): x_const_rtx_hash_table is a hash of
rtxes.
(struct rtx_const): Give substructures names, improve
formatting.
(struct constant_descriptor): Delete.
(struct constant_descriptor_tree): New, based on
constant_descriptor.
(const_hash_table): Is a hash table of trees.
(mark_const_hash_entry): Is used for hashes of trees. Mark
constant_descriptor_tree structure.
(mark_const_str_htab_1): Mark deferred_string structure.
(compare_constant): Rewrite to compare trees.
(compare_constant_1): Delete.
(record_constant): Delete.
(record_constant_1): Delete.
(output_constant_def): Use struct constant_descriptor_tree.
Don't duplicate trees twice.
(struct constant_descriptor_rtx): New.
(struct pool_constant): Used for rtx constants.
(init_varasm_status): Update for change to struct varasm_status.
(mark_varasm_status): Likewise.
(free_varasm_status): Delete.
(compare_constant_rtx): Rewrite to handle
constant_descriptor_rtx.
(record_constant_rtx): Likewise.
(mem_for_const_double): Update to use struct
constant_descriptor_rtx.
(force_const_mem): Likewise.
* Makefile.in (varasm.o): Doesn't depend on obstack.h.
* function.c (free_after_compilation): Don't use
free_varasm_status.
* function.h: Don't prototype free_varasm_status.
* ggc-common.c (ggc_realloc): Handle X being NULL.
* ggc-common.c (ggc_realloc): New function.
* ggc.h: Prototype it.
* emit-rtl.c (free_emit_status): Delete.
(init_emit): Allocate emit subarrays using GC.
(gen_reg_rtx): Reallocate subarrays using GC.
(init_emit): Use GC to allocate 'struct emit_status' and its
subarrays.
(mark_emit_status): Mark structure and its subarrays.
* stmt.c (free_stmt_status): Delete.
* expr.c (free_expr_status): Delete.
* function.h: Remove prototypes for deleted functions.
* function.c (free_after_compilation): Don't use deleted
functions.
Don't call free() on x_parm_reg_stack_loc.
(free_after_parsing): Don't use free_stmt_status.
(assign_parms): Use GC to allocate and resize
x_parm_reg_stack_loc.
(mark_function_status): Mark x_parm_reg_stack_loc.
* varasm.c (init_varasm_status): Use GC to allocate
'struct varasm_status' and its fields x_const_rtx_hash_table
and x_const_rtx_sym_hash_table.
(mark_varasm_status): Mark them.
(free_varasm_status): Use GC to free them.
* expr.c (init_expr): Use GC to allocate 'struct expr_status'.
(mark_expr_status): Mark the structure itself.
(free_expr_status): Use GC to free the structure.
* stmt.c (free_stmt_status): Use GC to free 'struct
stmt_status'.
(mark_stmt_status): Mark the 'struct stmt_status' itself.
(init_stmt_for_function): Allocate the structure for GC.
* dwarf2out.c (lookup_type_die): Use TYPE_SYMTAB_DIE.
(equate_type_number_to_die): Likewise.
* tree.h (TYPE_SYMTAB_DIE): New macro.
(struct die_struct): Predeclare.
(struct tree_type): Add field symtab.die. Add a tag
to the union type of field symtab.
* varray.h (VARRAY_RTVEC_INIT): A varray of rtvec contains
'struct rtvec_def *', not 'struct rtvec_def'.
* function.h (original_arg_vector): Make a real rtvec.
* function.c (ggc_mark_struct_function): Adjust.
* integrate.c (expand_inline_function): Adjust.
[cp/Changelog]
* pt.c (inline_parm_levels): Mark for GC.
* mangle.c (start_mangling): Allocate G.substitutions here...
(init_mangle): ... rather than here.
(finish_mangling): Clear the varray pointer when done with it.
* spew.c (yylexstring): Don't use VARRAY_FREE.
* search.c (bfs_walk): Don't use VARRAY_FREE.
* decl2.c (pending_statics): Use gengtype to mark.
(deferred_fns): Likewise.
(ssdf_decls): Likewise.
(init_decl2): Delete.
* decl.c (pop_from_top_level): Don't use VARRAY_FREE.
(cxx_init_decl_processing): Don't call init_decl2.
(cxx_pop_function_context): Don't use VARRAY_FREE.
* cp-tree.h (struct saved_scope): No need for special marking
of varrays.
(struct language_function): Likewise.
(local_classes): Use gengtype to mark.
(init_decl2): Delete prototype.
* class.c (init_class_processing): Don't use
ggc_add_tree_varray_root.
(build_vtbl_initializer): Don't use VARRAY_FREE.
* decl.c (typename_compare): Don't use same_type_p.
* decl.c: Include hashtab.h instead of hash.h.
(typename_hash): Update to use htab_h.
(typename_compare): Likewise.
(typename_htab): Use gengtype to mark.
(build_typename_type): Update to use htab_h.
* Make-lang.in (cp/decl.o): Use HASHTAB_H instead of hash.h.
* Make-lang.in (gt-cp-tree.h): New rule.
(cp/tree.o): Depend on gt-cp-tree.h.
* config-lang.in (gtfiles): Add cp/tree.c.
* tree.c: Include gt-cp-tree.h.
(list_hash_table): Use gengtype to mark.
(init_tree): Use gengtype to mark trees.
* Make-lang.in (cp/decl.o): Add debug.h dependency.
* call.c (struct z_candidate): Use gengtype.
(USER_CONV_CAND): Use WRAPPER_ZC.
(convert_class_to_reference): Use build_zc_wrapper.
(build_type_conversion_1): Likewise.
(build_over_call): Use WRAPPER_ZC.
(add_warning): Use build_zc_wrapper.
* cp-lang.c (LANG_HOOKS_MARK_TREE): Delete.
* cp-tree.h (struct lang_identifier): Use gengtype.
(struct template_parm_index_s): Likewise.
(struct ptrmem_cst): Likewise.
(struct tree_binding): Likewise.
(struct tree_overload): Likewise.
(struct tree_srcloc): Likewise.
(struct tree_wrapper): Likewise. Also modify to have a pointer
to struct z_candidate rather than void.
(enum cp_tree_node_structure_enum): New.
(union lang_tree_node): New.
(cxx_mark_tree): Delete prototype.
(cp_tree_node_structure): New prototype.
(build_ptr_wrapper): Delete prototype.
(build_int_wrapper): Delete prototype.
(build_zc_wrapper): New prototype.
* decl.c: Include debug.h
(cxx_mark_tree): Delete.
(cp_tree_node_structure): New.
* tree.c (build_ptr_wrapper): Delete.
(build_int_wrapper): Delete.
(build_zc_wrapper): New.
* cp-tree.h [! ENABLE_TREE_CHECKING] (LANG_TYPE_PTRMEM_CHECK):
Correct typo. Patch from k_fukui@highway.ne.jp.
* semantics.c (current_stmt_tree): Update for change to
struct language_function.
(finish_mem_initializers): Likewise.
* decl.c (cxx_init_decl_processing): Don't set mark_lang_status.
* cp-tree.h (struct language_function): Rename from
cp_language_function. Change all uses.
(cp_function_chain): Don't need to cast.
* class.c (duplicate_tag_error): Reset discriminator.
(check_bases_and_members): Update for data structure changes.
* cp-tree.h (struct lang_id2): Use gengtype.
(flagged_type_tree): Likewise.
(SET_LANG_ID): Use GGC on struct lang_id2.
(struct cp_language_function): Use gengtype. Remove field
'x_vcalls_possible_p'.
(current_vcalls_possible_p): Delete.
(struct lang_type_header): New.
(struct lang_type_class): Rename from struct lang_type. Include
struct lang_type_header.
(struct lang_type_ptrmem): New.
(struct lang_type): New.
(LANG_TYPE_CLASS_CHECK): New. Use it in all the appropriate
macros.
(LANG_TYPE_PTRMEM_CHECK): New. Use it in all the appropriate
macros.
(TYPE_SET_PTRMEMFUNC_TYPE): Set discriminator, update for
changes.
(struct lang_decl_flags): Use gengtype. Add discriminators.
(struct lang_decl): Use gengtype. Add and use discriminators.
Update the macros that reference moved fields.
(LANG_DECL_U2_CHECK): New function. Use it when appropriate.
(SET_DECL_THUNK_P): Set discriminator too.
(clear_inline_text_obstack): Delete prototype.
(finish_inline_definitions): Delete prototype.
(mark_pending_inlines): Delete prototype.
(lang_check_failed): New prototype.
* decl.c (struct named_label_use_list): Use gengtype.
(struct named_label_list): Likewise.
(mark_binding_level): Delete.
(mark_named_label_lists): Delete.
(push_local_name): Set discriminator on DECL_LANG_SPECIFIC.
(cxx_init_decl_processing): Use generated marker routine.
(begin_destructor_body): Delete dead set to
current_vcalls_possible_p.
(mark_lang_function): Delete.
(mark_cp_function_context): Delete.
(lang_mark_tree): Use generated marker routines.
* decl2.c (start_objects): Set discriminator when setting
GLOBAL_INIT_PRIORITY.
* lex.c (retrofit_lang_decl): Set discriminators.
(copy_lang_type): Update for changes to lang_type structure.
(cp_make_lang_type): Set discriminator.
* parse.y: Use gengtype on YYLVAL. Don't use dots in
identifiers.
* search.c: Include ggc.h.
* semantics.c (anon_aggr_type_p): Use the macro, don't
hand-code it.
(finish_inline_definitions): Delete.
* spew.c (struct token): Use gengtype.
(struct token_chunk): New.
(struct unparsed_text): Use gengtype. Store tokens in chunks.
(struct feed): Use gengtype.
(feed_obstack): Delete.
(feed): Mark as GC root.
(pending_inlines): Mark as GC root.
(pending_inlines_tail): Likewise.
(processing_these_inlines): Likewise.
(token_obstack): Make static.
(first_token): Likewise.
(init_spew): Don't initialise deleted things; use gengtype for
roots.
(clear_inline_text_obstack): Delete.
(feed_input): Use GC for struct feed. Update for changes to
struct unparsed_text.
(mark_pending_inlines): Delete.
(next_token): Rename from add_token. Change all callers.
Update
for changes to struct unparsed_text.
(space_for_token): New.
(remove_last_token): New.
(alloc_unparsed_text): New.
(snarf_block): Take an unparsed_text. Update for changes to
struct
unparsed_text.
(snarf_method): Update for changes to struct unparsed_text.
(snarf_defarg): Update for changes to struct unparsed_text.
* tree.c (lang_check_failed): New.
* Make-lang.in (gt-cp-call.h gt-cp-decl2.h gt-cp-parse.h
gt-cp-pt.h gt-cp-repo.h gt-cp-spew.h): New rules.
(cp/spew.o): Add dependency on gt-<filename>.h.
(cp/decl2.o): Add dependency on gt-<filename>.h.
(cp/call.o): Add dependency on gt-<filename>.h.
(cp/pt.o): Add dependency on gt-<filename>.h.
(cp/repo.o): Add dependency on gt-<filename>.h.
(cp/parse.o): Add dependency on gt-<filename>.h.
* call.c: Use gengtype for roots.
* config-lang.in (gtfiles): Add cp-tree.h decl.h lex.h call.c
decl2.c parse.y pt.c repo.c spew.c.
* cp-tree.h: Use gengtype for roots.
(struct saved_scope): Use GGC, gengtype.
(cp_parse_init): Delete prototype.
(init_pt): Delete prototype.
* decl.c: Use gengtype for roots.
(mark_saved_scope): Delete.
(cxx_init_decl_processing): Don't call deleted initilisation
routines.
(signed_size_zero_node): Delete, unused.
* decl.h: Use gengtype for roots.
* decl2.c: Use gengtype for roots.
* lex.h: Use gengtype for roots.
* parse.y: Use gengtype for roots.
(cp_parse_init): Delete.
* pt.c: Use gengtype for roots.
(init_pt): Delete.
* repo.c: Use gengtype for roots.
* spew.c: Use gengtype for roots.
* Make-lang.in: Allow for filename changes. Add gtype-cp.h.
(cp/decl.o): Add dependency on gtype-cp.h.
* decl.c: Remove use of add_deletable_root, use GTY marker
instead.
Include gtype-cp.h. Allow for filename changes.
* Make-lang.in (cp/gt-decl.h): Generate using gengtype.
(cp/decl.o): Add cp/gt-decl.h dependency.
* config-lang.in (gtfiles): New.
* tree.h: Rename struct binding_level to struct
cp_binding_level.
* decl.c: Rename struct binding_level to struct
cp_binding_level.
Include cp/gt-decl.h.
(struct cp_binding_level): Use gengtype.
(make_binding_level): Use GGC on struct cp_binding_level.
(mark_binding_level): Use gt_ggc_m_cp_binding_level.
(cxx_init_decl_processing): Mark free_binding_level as
deletable.
* decl.c (mark_cp_function_context): Update calling sequence.
* decl.c (start_function): Don't free 'struct
cp_language_function'.
(pop_cp_function_context): Likewise.
(save_function_data): Allocate it using GC.
* semantics.c (genrtl_start_function): Don't free 'struct
cp_language_function'.
2002-11-07 Mark Mitchell <mark@codesourcery.com>
PR c++/8338
* pt.c (for_each_template_parm): Add htab parameter.
(process_partial_specialization): Adjust call.
(push_template_decl_real): Likewise.
(pair_fn_data): Add visited.
(for_each_template_parm_r): Avoid walking duplicates more than
once.
(uses_template_parms): Adjust call to for_each_template_parm.
--------------------------------------------------------------
Ziemowit Laski 1 Infinite Loop, MS 301-2K
Mac OS X Compiler Group Cupertino, CA USA 95014-2083
Apple Computer, Inc. +1.408.974.6229 Fax .5477
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 20:51 GCC 3.3 compile speed regression - THE GRAPH Ziemowit Laski
@ 2003-02-07 22:08 ` Phil Edwards
2003-02-07 22:14 ` Neil Booth
2003-02-09 19:25 ` Fergus Henderson
2 siblings, 0 replies; 24+ messages in thread
From: Phil Edwards @ 2003-02-07 22:08 UTC (permalink / raw)
To: Ziemowit Laski; +Cc: gcc
On Fri, Feb 07, 2003 at 12:51:59PM -0800, Ziemowit Laski wrote:
>
> The two "inflection points" occur on 2002-06-04 and 2002-11-07,
> respectively. They correspond to the following
> two check-ins, with the ChangeLog entries reproduced in their gore and
> glory below.
Nice detective work.
Phil
--
I would therefore like to posit that computing's central challenge, viz. "How
not to make a mess of it," has /not/ been met.
- Edsger Dijkstra, 1930-2002
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 20:51 GCC 3.3 compile speed regression - THE GRAPH Ziemowit Laski
2003-02-07 22:08 ` Phil Edwards
@ 2003-02-07 22:14 ` Neil Booth
2003-02-07 22:22 ` Ziemowit Laski
` (2 more replies)
2003-02-09 19:25 ` Fergus Henderson
2 siblings, 3 replies; 24+ messages in thread
From: Neil Booth @ 2003-02-07 22:14 UTC (permalink / raw)
To: Ziemowit Laski; +Cc: gcc
Ziemowit Laski wrote:-
> I've already spoken with Geoff regarding his fateful PCH merge. He did
> not think
> that doing a binary search on the pch-branch would be particularly
> productive,
> but instead suggested I look closely at his spew.c mods, which I shall
> do now.
spew.c is dead in 3.4, so this would only be useful for 3.3. I suggest
you spend your time on something else.
Neil.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 22:14 ` Neil Booth
@ 2003-02-07 22:22 ` Ziemowit Laski
2003-02-07 22:25 ` Neil Booth
2003-02-07 22:43 ` Stan Shebs
2003-02-07 22:22 ` Joe Buck
2003-02-09 17:46 ` Marc Espie
2 siblings, 2 replies; 24+ messages in thread
From: Ziemowit Laski @ 2003-02-07 22:22 UTC (permalink / raw)
To: Neil Booth; +Cc: gcc
On Friday, Feb 7, 2003, at 14:14 US/Pacific, Neil Booth wrote:
> Ziemowit Laski wrote:-
>
>> I've already spoken with Geoff regarding his fateful PCH merge. He
>> did
>> not think
>> that doing a binary search on the pch-branch would be particularly
>> productive,
>> but instead suggested I look closely at his spew.c mods, which I shall
>> do now.
>
> spew.c is dead in 3.4, so this would only be useful for 3.3. I suggest
> you spend your time on something else.
No sale. Apple will be using 3.3, not 3.4. :-(
--Zem
--------------------------------------------------------------
Ziemowit Laski 1 Infinite Loop, MS 301-2K
Mac OS X Compiler Group Cupertino, CA USA 95014-2083
Apple Computer, Inc. +1.408.974.6229 Fax .5477
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 22:14 ` Neil Booth
2003-02-07 22:22 ` Ziemowit Laski
@ 2003-02-07 22:22 ` Joe Buck
2003-02-09 17:46 ` Marc Espie
2 siblings, 0 replies; 24+ messages in thread
From: Joe Buck @ 2003-02-07 22:22 UTC (permalink / raw)
To: Neil Booth; +Cc: Ziemowit Laski, gcc
Ziemowit Laski wrote:-
> > I've already spoken with Geoff regarding his fateful PCH merge. He did
> > not think
> > that doing a binary search on the pch-branch would be particularly
> > productive,
> > but instead suggested I look closely at his spew.c mods, which I shall
> > do now.
On Fri, Feb 07, 2003 at 10:14:24PM +0000, Neil Booth wrote:
> spew.c is dead in 3.4, so this would only be useful for 3.3. I suggest
> you spend your time on something else.
Assuming that we're going to ship 3.3, we do care if there is a massive
performance regression, even if it is fixed in 3.4, since the world will
have to put up with 3.3 as the official release series for at least a
six-month period.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 22:22 ` Ziemowit Laski
@ 2003-02-07 22:25 ` Neil Booth
2003-02-07 22:43 ` Stan Shebs
1 sibling, 0 replies; 24+ messages in thread
From: Neil Booth @ 2003-02-07 22:25 UTC (permalink / raw)
To: Ziemowit Laski; +Cc: gcc
Ziemowit Laski wrote:-
> >spew.c is dead in 3.4, so this would only be useful for 3.3. I suggest
> >you spend your time on something else.
>
> No sale. Apple will be using 3.3, not 3.4. :-(
OK, that's understandable then.
Neil.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 22:22 ` Ziemowit Laski
2003-02-07 22:25 ` Neil Booth
@ 2003-02-07 22:43 ` Stan Shebs
1 sibling, 0 replies; 24+ messages in thread
From: Stan Shebs @ 2003-02-07 22:43 UTC (permalink / raw)
To: Ziemowit Laski; +Cc: Neil Booth, gcc
Ziemowit Laski wrote:
>
> On Friday, Feb 7, 2003, at 14:14 US/Pacific, Neil Booth wrote:
>
>> Ziemowit Laski wrote:-
>>
>>> I've already spoken with Geoff regarding his fateful PCH merge. He did
>>> not think
>>> that doing a binary search on the pch-branch would be particularly
>>> productive,
>>> but instead suggested I look closely at his spew.c mods, which I shall
>>> do now.
>>
>>
>> spew.c is dead in 3.4, so this would only be useful for 3.3. I suggest
>> you spend your time on something else.
>
>
> No sale. Apple will be using 3.3, not 3.4. :-(
Lest anybody think this is an announcement of Apple product plans,
the situation is perhaps better phrased as "Apple currently has a
requirement to speed up 3.3". Much can change as the situation
develops.
Stan
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 22:14 ` Neil Booth
2003-02-07 22:22 ` Ziemowit Laski
2003-02-07 22:22 ` Joe Buck
@ 2003-02-09 17:46 ` Marc Espie
2003-02-09 18:56 ` GCC 3.3 compile speed regression - AN ANSWER Michael S. Zick
2003-02-09 21:06 ` GCC 3.3 compile speed regression - THE GRAPH Neil Booth
2 siblings, 2 replies; 24+ messages in thread
From: Marc Espie @ 2003-02-09 17:46 UTC (permalink / raw)
To: gcc
In article <20030207221424.GC17153@daikokuya.co.uk> you write:
>Ziemowit Laski wrote:-
>> I've already spoken with Geoff regarding his fateful PCH merge. He did
>> not think
>> that doing a binary search on the pch-branch would be particularly
>> productive,
>> but instead suggested I look closely at his spew.c mods, which I shall
>> do now.
>spew.c is dead in 3.4, so this would only be useful for 3.3. I suggest
>you spend your time on something else.
>Neil.
You're going to say I'm always beating the same dead horse, but I believe
this is THE stem of all of gcc's performance problems:
the next version is always going to be better.
I've heard this at least ten times since 2.95 came out (between the gc for
stability -> flush speed out the window.. now it's pch, which is going to
be only in 3.4), and in reality, the current stable release always is pretty
bad, speed-wise, but who cares ?
the next release is going to be better...
After a few releases, it gets really hard to believe how the next release
is going to be better, since reality seems to prove otherwise.
And there you have the conundrum for distribution builders: each new release
of gcc has marked improvemtns in some areas, and marked losses in some others,
so it is fairly difficult to choose. Especially when you're not building
for mainstream environment...
The latest `improvement' in the release process seems to me of the
`sweep issues under the carpet' nature: mark as regression stuff that is
a regression over the last immediate release, and more or less ignore less
recent problems that did not exist in, say, gcc 2.95.3.
In the end, because there is no definite win in upgrading, various people
stay with various versions of gcc (for instance, OpenBSD will stay stuck
with 2.95.x until the compiler's speed markedly improves), thus dispersing
forces, handling lots more bug-reports than could happen otherwise...
And yes, the speed issue can be solved by cross-compiling. Unfortunately,
this is also the last nail in the coffin for old, slow architectures. Think
about it: if you start cross-compiling everything, you are getting rid of a
very good stress test (compiling the whole system), and thus help some bugs
not getting noticed...
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - AN ANSWER
2003-02-09 17:46 ` Marc Espie
@ 2003-02-09 18:56 ` Michael S. Zick
2003-02-09 21:09 ` Neil Booth
2003-02-09 21:06 ` GCC 3.3 compile speed regression - THE GRAPH Neil Booth
1 sibling, 1 reply; 24+ messages in thread
From: Michael S. Zick @ 2003-02-09 18:56 UTC (permalink / raw)
To: gcc
On Sunday 09 February 2003 11:45 am, Marc Espie wrote:
>
> the next version is always going to be better.
>
> After a few releases, it gets really hard to believe how the next release
> is going to be better, since reality seems to prove otherwise.
>
Taking several steps back from the current problem(s) for a broader
view...
Perhaps it is the Open Source Development, Staffed by Volunteers
model that is the central point to be dealt with.
Compiler execution speed;
Adherence to standards;
Effectiveness of the code generated
(your choice of definition of "Effectiveness");
Those are all measures of consideration to be given when selecting
a production compiler.
People volunteer for what interests them (or interests their employer).
Adding the hottest new fad in compiler science to a widely used
compiler can gather widespread interest.
Even adding established, proven (recent) compiler techniques to
that compiler can gather widespread interest.
There is rarely a lack of volunteers for such intellectual challenges.
Although challenging in its own way, bug fixing just doesn't have the
same intellectual appeal.
Sorting out things like a ten year accumulation of data structures into
a fast, elegant, production compiler core is just plain hard work.
Fortunately, there are a few people who like that sort of thing but there
are most certainly fewer of them in the world than of the compiler
science explorers.
GCC as a widely used, Open Source, compiler is certainly a good
vehicle for the compiler science explorers. But where does that
leave the persons hoping it will become a "killer" of a production
compiler?
So, lets deal with that question.
Set up, right at the top of the tree, two branches. Perhaps even
name one of them something other than GCC (pGCC?).
This new, production compiler, branch will not accept "outside"
merges. It will evolve ONLY by "pulls" from the current GCC tree,
still maintained by the compiler science explorers in the current
fashion.
This will require some agreement on how proven a compiler change
in the GCC tree must be to qualify as "ready to pull into production";
but presuming that can be established...
How do you get people interested in volunteering for "grunt work"
such as this?
Present it as an intellectual challenge, possibly also a technical
challenge, to bring that new, well established, well proven, GCC
compiler change into pGCC without degrading any of those first
three measures of a production compiler.
That should be one source of volunteers.
Another source of volunteers would be those "volunteered" by their
employers. In this case, the employers would be those whose interest
in a solid production compiler is stronger than their interest in a
"bleeding edge" compiler.
My suggestion:
Create two "mainline" compiler projects; the current GCC project
and a production compiler version project.
Mike
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-07 20:51 GCC 3.3 compile speed regression - THE GRAPH Ziemowit Laski
2003-02-07 22:08 ` Phil Edwards
2003-02-07 22:14 ` Neil Booth
@ 2003-02-09 19:25 ` Fergus Henderson
2003-02-09 20:18 ` Diego Novillo
` (2 more replies)
2 siblings, 3 replies; 24+ messages in thread
From: Fergus Henderson @ 2003-02-09 19:25 UTC (permalink / raw)
To: Ziemowit Laski; +Cc: gcc
Those who are interested in compilation speed may be interested in
the Tiny C Compiler, tcc <http://fabrice.bellard.free.fr/tcc/>.
I tried several C compilers on some C code generated by the Mercury
compiler, with the following results:
compiler version options compilation time
-------- ------- ------- ----------------
gcc 2.95.4 -c -O2 ... 5.53
gcc 2.95.4 -c -O1 ... 3.19
gcc 2.95.4 -c -O0 ... 2.65
lcc 4.1 -c 1.04
tcc 1.9.16 -c 0.34
Obviously gcc has room for improvement!
P.S.
Some details:
- tests carried out on a 1100MHz AMD Athlon
64kb L1 icache, 64kb L1 dcache, 256k L2 cache, 256Mb RAM
Debian GNU/Linux 3.0
binutils version 2.12.90.0.1 20020307
- "..." above is "-fomit-frame-pointer -ansi -Wall -Wwrite-strings -Wshadow
-Wstrict-prototypes -Wmissing-prototypes -Wno-unused -Wno-uninitialized".
There were also a few "-I" options passed to all C compilers.
- test code produced by `mmc -C -H tree234.m' using tree234.m from
the library directory in the Mercury 0.11.0 distribution.
--
Fergus Henderson <fjh@cs.mu.oz.au> | "I have always known that the pursuit
The University of Melbourne | of excellence is a lethal habit"
WWW: <http://www.cs.mu.oz.au/~fjh> | -- the last words of T. S. Garp.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 19:25 ` Fergus Henderson
@ 2003-02-09 20:18 ` Diego Novillo
2003-02-09 20:28 ` Aaron Lehmann
2003-02-09 20:46 ` Fergus Henderson
2003-02-09 21:15 ` Neil Booth
2003-02-09 23:24 ` Joseph S. Myers
2 siblings, 2 replies; 24+ messages in thread
From: Diego Novillo @ 2003-02-09 20:18 UTC (permalink / raw)
To: Fergus Henderson; +Cc: Ziemowit Laski, gcc
On Mon, 10 Feb 2003, Fergus Henderson wrote:
> compiler version options compilation time
> -------- ------- ------- ----------------
> gcc 2.95.4 -c -O2 ... 5.53
> gcc 2.95.4 -c -O1 ... 3.19
> gcc 2.95.4 -c -O0 ... 2.65
> lcc 4.1 -c 1.04
> tcc 1.9.16 -c 0.34
>
What optimizations were used by 'lcc' and 'tcc'? i.e., is GCC 7
times or 16 times slower than 'tcc'? What about the generated
code? I want to make sure that we are comparing apples to
apples.
Diego.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 20:18 ` Diego Novillo
@ 2003-02-09 20:28 ` Aaron Lehmann
2003-02-09 20:46 ` Fergus Henderson
1 sibling, 0 replies; 24+ messages in thread
From: Aaron Lehmann @ 2003-02-09 20:28 UTC (permalink / raw)
To: Diego Novillo; +Cc: Fergus Henderson, Ziemowit Laski, gcc
On Sun, Feb 09, 2003 at 03:17:15PM -0500, Diego Novillo wrote:
> What optimizations were used by 'lcc' and 'tcc'? i.e., is GCC 7
> times or 16 times slower than 'tcc'? What about the generated
> code? I want to make sure that we are comparing apples to
> apples.
tcc is like "gcc -O-1". Its code generation is freakily primitive.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 20:18 ` Diego Novillo
2003-02-09 20:28 ` Aaron Lehmann
@ 2003-02-09 20:46 ` Fergus Henderson
1 sibling, 0 replies; 24+ messages in thread
From: Fergus Henderson @ 2003-02-09 20:46 UTC (permalink / raw)
To: Diego Novillo; +Cc: Ziemowit Laski, gcc
On 09-Feb-2003, Diego Novillo <dnovillo@redhat.com> wrote:
> On Mon, 10 Feb 2003, Fergus Henderson wrote:
>
> > compiler version options compilation time
> > -------- ------- ------- ----------------
> > gcc 2.95.4 -c -O2 ... 5.53
> > gcc 2.95.4 -c -O1 ... 3.19
> > gcc 2.95.4 -c -O0 ... 2.65
> > lcc 4.1 -c 1.04
> > tcc 1.9.16 -c 0.34
> >
> What optimizations were used by 'lcc' and 'tcc'? i.e., is GCC 7
> times or 16 times slower than 'tcc'? What about the generated
> code? I want to make sure that we are comparing apples to
> apples.
Neither lcc nor tcc are optimizing compilers.
So the right one to compare compilation speed with is `-O0'.
--
Fergus Henderson <fjh@cs.mu.oz.au> | "I have always known that the pursuit
The University of Melbourne | of excellence is a lethal habit"
WWW: <http://www.cs.mu.oz.au/~fjh> | -- the last words of T. S. Garp.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 17:46 ` Marc Espie
2003-02-09 18:56 ` GCC 3.3 compile speed regression - AN ANSWER Michael S. Zick
@ 2003-02-09 21:06 ` Neil Booth
1 sibling, 0 replies; 24+ messages in thread
From: Neil Booth @ 2003-02-09 21:06 UTC (permalink / raw)
To: Marc Espie; +Cc: gcc
Marc Espie wrote:-
> You're going to say I'm always beating the same dead horse, but I believe
> this is THE stem of all of gcc's performance problems:
>
> the next version is always going to be better.
I didn't say that. I said improving this code doesn't help future
GCC's. I think that was quite clear.
Neil.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - AN ANSWER
2003-02-09 18:56 ` GCC 3.3 compile speed regression - AN ANSWER Michael S. Zick
@ 2003-02-09 21:09 ` Neil Booth
2003-02-10 14:33 ` Michael S. Zick
0 siblings, 1 reply; 24+ messages in thread
From: Neil Booth @ 2003-02-09 21:09 UTC (permalink / raw)
To: Michael S.Zick; +Cc: gcc
Michael S.Zick wrote:-
> Taking several steps back from the current problem(s) for a broader
> view...
>
> Perhaps it is the Open Source Development, Staffed by Volunteers
> model that is the central point to be dealt with.
Please, let's stop this volunteers excuse. It really grates with
me, being a volunteer. I'll merely note that a very large proportion
of slow-downs are in code committed by people paid to work on GCC.
Neil.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 19:25 ` Fergus Henderson
2003-02-09 20:18 ` Diego Novillo
@ 2003-02-09 21:15 ` Neil Booth
2003-02-09 23:24 ` Joseph S. Myers
2 siblings, 0 replies; 24+ messages in thread
From: Neil Booth @ 2003-02-09 21:15 UTC (permalink / raw)
To: Fergus Henderson; +Cc: Ziemowit Laski, gcc
Fergus Henderson wrote:-
> Those who are interested in compilation speed may be interested in
> the Tiny C Compiler, tcc <http://fabrice.bellard.free.fr/tcc/>.
However, please let's be fair in comparisons. As I noted on the
kernel list recently, tcc is not a fair comparison:
o It doesn't have half the capabilities that GCC's C front end
has. But one simple example is -E.
o tcc doesn't have to worry about other OS's or other target
arch's.
o tcc has almost *no* semantic or syntactic checking. It accepts
all kinds of garbage. It's easy to be fast when you don't do half
the work. While its author claims C89 compliance, in reality
it's nowhere near.
o tcc doesn't even expand macros correctly. I could tell that
just glancing at the source. It's hard to demonstrate owing
to lack of -E.
So please, let's not compare with tcc until it's a real compiler.
Comparisons with LCC are fair. You'll note that I'm pushing hard
for a better C front end, and not making excuses for GCC. We
certainly are way too slow.
Neil.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 19:25 ` Fergus Henderson
2003-02-09 20:18 ` Diego Novillo
2003-02-09 21:15 ` Neil Booth
@ 2003-02-09 23:24 ` Joseph S. Myers
2003-02-10 7:43 ` Fergus Henderson
2 siblings, 1 reply; 24+ messages in thread
From: Joseph S. Myers @ 2003-02-09 23:24 UTC (permalink / raw)
To: Fergus Henderson; +Cc: gcc
On Mon, 10 Feb 2003, Fergus Henderson wrote:
> gcc 2.95.4 -c -O2 ... 5.53
> gcc 2.95.4 -c -O1 ... 3.19
> gcc 2.95.4 -c -O0 ... 2.65
What about -O0 -fsyntax-only? (To see how much time is spent in the front
end and how much generating code.) And how much worse is 3.3 (with each
of the options)?
--
Joseph S. Myers
jsm28@cam.ac.uk
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-09 23:24 ` Joseph S. Myers
@ 2003-02-10 7:43 ` Fergus Henderson
2003-02-11 20:25 ` Neil Booth
2003-02-12 19:57 ` Fergus Henderson
0 siblings, 2 replies; 24+ messages in thread
From: Fergus Henderson @ 2003-02-10 7:43 UTC (permalink / raw)
To: Joseph S. Myers; +Cc: gcc
On 09-Feb-2003, Joseph S. Myers <jsm28@cam.ac.uk> wrote:
> What about -O0 -fsyntax-only?
> And how much worse is 3.3 (with each of the options)?
I don't happen to have 3.3 handy, but I built the current CVS head
(3.4 20030209) yesterday. Here are the results.
gcc 2.95.4 -c -O2 ... 5.53
gcc 2.95.4 -c -O1 ... 3.19
gcc 2.95.4 -c -O0 ... 2.65
gcc 2.95.4 -c -O0 -fsyntax-only ... 1.16
gcc 3.4 20030209 -c -O2 ... 10.16
gcc 3.4 20030209 -c -O1 ... 5.10
gcc 3.4 20030209 -c -O0 ... 4.03
gcc 3.4 20030209 -c -O0 -fsyntax-only ... 0.87
lcc 4.1 -c 1.04
tcc 0.9.16 -c 0.34
There is at least *some* good news: for this benchmark,
`-fsyntax-only' is now 33% faster than it was in 2.95.4!
On the other hand, for `-O0' we've gone from 7.8 times slower than
tcc to 11.8 times slower than tcc -- in other words, just the amount by
which we slowed down since 2.95.4 was itself four times the total that
tcc takes.
--
Fergus Henderson <fjh@cs.mu.oz.au> | "I have always known that the pursuit
The University of Melbourne | of excellence is a lethal habit"
WWW: <http://www.cs.mu.oz.au/~fjh> | -- the last words of T. S. Garp.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - AN ANSWER
2003-02-09 21:09 ` Neil Booth
@ 2003-02-10 14:33 ` Michael S. Zick
0 siblings, 0 replies; 24+ messages in thread
From: Michael S. Zick @ 2003-02-10 14:33 UTC (permalink / raw)
To: Neil Booth; +Cc: gcc
On Sunday 09 February 2003 03:09 pm, Neil Booth wrote:
> Michael S.Zick wrote:-
>
> > Taking several steps back from the current problem(s) for a broader
> > view...
> >
> > Perhaps it is the Open Source Development, Staffed by Volunteers
> > model that is the central point to be dealt with.
>
> Please, let's stop this volunteers excuse. It really grates with
> me, being a volunteer. I'll merely note that a very large proportion
> of slow-downs are in code committed by people paid to work on GCC.
>
> Neil.
Please read further into the post Neil.
I said good things about the volunteers and did not posit it as a
reason or an excuse.
Mike
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-10 7:43 ` Fergus Henderson
@ 2003-02-11 20:25 ` Neil Booth
2003-02-11 23:35 ` Joseph S. Myers
2003-02-12 19:57 ` Fergus Henderson
1 sibling, 1 reply; 24+ messages in thread
From: Neil Booth @ 2003-02-11 20:25 UTC (permalink / raw)
To: Fergus Henderson; +Cc: Joseph S. Myers, gcc
Fergus Henderson wrote:-
> There is at least *some* good news: for this benchmark,
> `-fsyntax-only' is now 33% faster than it was in 2.95.4!
Most, if not all of that is probably integrated CPP.
Neil.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-11 20:25 ` Neil Booth
@ 2003-02-11 23:35 ` Joseph S. Myers
0 siblings, 0 replies; 24+ messages in thread
From: Joseph S. Myers @ 2003-02-11 23:35 UTC (permalink / raw)
To: Neil Booth; +Cc: gcc
On Tue, 11 Feb 2003, Neil Booth wrote:
> Fergus Henderson wrote:-
>
> > There is at least *some* good news: for this benchmark,
> > `-fsyntax-only' is now 33% faster than it was in 2.95.4!
>
> Most, if not all of that is probably integrated CPP.
Which absorbed less than one release's worth of slack in the rest of the
compiler; and speeding up the C front end might absorb another less than
one release's worth of slack - useful, but we need to get rid of the slack
in the parts of the compiler that accumulate it (which isn't mainly the
front ends), after which improvements in the speed of front ends might be
more noticeable.
--
Joseph S. Myers
jsm28@cam.ac.uk
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
2003-02-10 7:43 ` Fergus Henderson
2003-02-11 20:25 ` Neil Booth
@ 2003-02-12 19:57 ` Fergus Henderson
1 sibling, 0 replies; 24+ messages in thread
From: Fergus Henderson @ 2003-02-12 19:57 UTC (permalink / raw)
To: Joseph S. Myers; +Cc: gcc
On 10-Feb-2003, Fergus Henderson <fjh@cs.mu.oz.au> wrote:
> gcc 2.95.4 -c -O2 ... 5.53
> gcc 2.95.4 -c -O1 ... 3.19
> gcc 2.95.4 -c -O0 ... 2.65
> gcc 2.95.4 -c -O0 -fsyntax-only ... 1.16
>
> gcc 3.4 20030209 -c -O2 ... 10.16
> gcc 3.4 20030209 -c -O1 ... 5.10
> gcc 3.4 20030209 -c -O0 ... 4.03
> gcc 3.4 20030209 -c -O0 -fsyntax-only ... 0.87
Sorry, I screwed up. Those results that I posted earlier were bogus,
because I forgot to configure with --disable-checking.
My apologies for the misinformation!
Here are the correct figures. For `-O0', we've only slowed down to 9.7 times
slower than tcc, not 11.8 times slower as I wrongly said in my earlier post,
and the net slowdown since 0.95 is only 1.4 times the time it takes tcc,
not four times.
I also have figures for gcc 3.2 now.
tcc 0.9.16 -c 0.34
lcc 4.1 -c 1.04
gcc 2.95.4 -c -O0 ... 2.65 base
gcc 3.2 -c -O0 ... 2.92 +10%
gcc 3.4 -c -O0 ... 3.30 +25%
gcc 2.95.4 -c -O1 ... 3.00 base
gcc 3.2 -c -O1 ... 3.62 +21%
gcc 3.4 -c -O1 ... 4.10 +37%
gcc 2.95.4 -c -Os ... 5.57 base
gcc 3.2 -c -Os ... 6.66 +20%
gcc 3.4 -c -Os ... 7.35 +32%
gcc 2.95.4 -c -O2 -fomit-frame-pointer ... 5.53 base
gcc 3.2 -c -O2 -fomit-frame-pointer ... 7.08 +28%
gcc 3.4 -c -O2 -fomit-frame-pointer ... 8.45 +53%
gcc 2.95.4 -c -O0 -fsyntax-only ... 1.16 base
gcc 3.2 -c -O0 -fsyntax-only ... 0.67 -42%
gcc 3.4 20030209 -c -O0 -fsyntax-only ... 0.67 -42%
The results are not very consistent from run to run (it was an unloaded
machine, but I didn't switch to single-user mode). I took the best of
five runs. But some of these figures might be off by 5% or thereabouts.
--
Fergus Henderson <fjh@cs.mu.oz.au> | "I have always known that the pursuit
The University of Melbourne | of excellence is a lethal habit"
WWW: <http://www.cs.mu.oz.au/~fjh> | -- the last words of T. S. Garp.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
@ 2003-02-10 4:11 Robert Dewar
0 siblings, 0 replies; 24+ messages in thread
From: Robert Dewar @ 2003-02-10 4:11 UTC (permalink / raw)
To: dnovillo, fjh; +Cc: gcc, zlaski
> Neither lcc nor tcc are optimizing compilers.
> So the right one to compare compilation speed with is `-O0'.
Even that is not an apples/apples comparison. Clearly if one does not
care about code quality, then it should be possible to write a C compiler
running at millions of lines a minute on a modern PC (compare for example
the speed of the Realia COBOL compiler). Even at -O0 you are paying
a significant price for the general structure of gcc which makes advanced
optimizations feasible (not to mention the very general table driven
sturcture of the code generator).
Certainly I agree that it should be possible to speed up gcc by a
significant factor while retaining these general characteristics, but
I am not sure it is instructive or useful to compare gcc performance
with tcc.
^ permalink raw reply [flat|nested] 24+ messages in thread
* Re: GCC 3.3 compile speed regression - THE GRAPH
@ 2003-02-08 7:46 Tim Josling
0 siblings, 0 replies; 24+ messages in thread
From: Tim Josling @ 2003-02-08 7:46 UTC (permalink / raw)
To: gcc
> The following graph shows the deterioration in compile-time performance
> as seen on the gcc-3_3-branch (and, prior to the creation of the branch,
> on the trunk itself). The test case is the same Qt fragment I used for
> my earlier GC analysis; the test machine (a RedHat Linux 8.0 box) is
> also the same. You can find the preprocessed source at:
> http://homepage.mac.com/zlaski/FileSharing1.html
Good work.
With this program the GC cost is around 35%, including 2.6% cost for
allocating using ggc_alloc (versus sequential allocation).
Tim Josling
^ permalink raw reply [flat|nested] 24+ messages in thread
end of thread, other threads:[~2003-02-12 19:52 UTC | newest]
Thread overview: 24+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2003-02-07 20:51 GCC 3.3 compile speed regression - THE GRAPH Ziemowit Laski
2003-02-07 22:08 ` Phil Edwards
2003-02-07 22:14 ` Neil Booth
2003-02-07 22:22 ` Ziemowit Laski
2003-02-07 22:25 ` Neil Booth
2003-02-07 22:43 ` Stan Shebs
2003-02-07 22:22 ` Joe Buck
2003-02-09 17:46 ` Marc Espie
2003-02-09 18:56 ` GCC 3.3 compile speed regression - AN ANSWER Michael S. Zick
2003-02-09 21:09 ` Neil Booth
2003-02-10 14:33 ` Michael S. Zick
2003-02-09 21:06 ` GCC 3.3 compile speed regression - THE GRAPH Neil Booth
2003-02-09 19:25 ` Fergus Henderson
2003-02-09 20:18 ` Diego Novillo
2003-02-09 20:28 ` Aaron Lehmann
2003-02-09 20:46 ` Fergus Henderson
2003-02-09 21:15 ` Neil Booth
2003-02-09 23:24 ` Joseph S. Myers
2003-02-10 7:43 ` Fergus Henderson
2003-02-11 20:25 ` Neil Booth
2003-02-11 23:35 ` Joseph S. Myers
2003-02-12 19:57 ` Fergus Henderson
2003-02-08 7:46 Tim Josling
2003-02-10 4:11 Robert Dewar
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).