[RFC/RFA] New function attribute "leaf"

Hi,
here is updated version of the patch for current mainline.  I also did some analysis
of the effectivity of the flag:

The reasons for ipa-reference giving up on analyzing cc1 functions can be summarized
as follows
     55    raise/ -> read all, write all
     57    qsort/ -> read all, write all
     63    strchr/ -> read all
     70    __builtin_memcpy/ -> read all, write all
     78    _obstack_begin/ -> read all, write all
     79    strncmp/ -> read all
     90    abort/ -> read all
    102    memmove/ -> read all, write all
    115    fputs_unlocked/ -> read all, write all
    123    obstack_free/ -> read all, write all
    132    __gmpz_set_si/ -> read all, write all
    141    dcgettext/ -> read all, write all
    154    __gmpz_clear/ -> read all, write all
    156    __builtin_fputs/ -> read all, write all
    175    __gmpz_init/ -> read all, write all
    214    realloc/ -> read all, write all
    230    malloc/ -> read all, write all
    243    strcmp/ -> read all
    295    snprintf/ -> read all, write all
    305    __builtin_alloca/ -> read all, write all
    319    sprintf/ -> read all, write all
    388    _obstack_newchunk/ -> read all, write all
    404    __overflow/ -> read all, write all
    464    __builtin_fwrite_unlocked/ -> read all, write all
    621    __builtin_fputc/ -> read all, write all
    675    strlen/ -> read all
    741    memcpy/ -> read all, write all
   2151    memset/ -> read all, write all
   2292    __builtin_fwrite/ -> read all, write all
   2330    fprintf/ -> read all, write all
   2840    free/ -> read all, write all
   5118    indirect call -> write all
   5119    indirect call -> read all

I must say I am bit surprised by frequency of indirect calls (it is abut 50% of
cases), but that can be handled by improving "devirtualization" and
constructing may edges.

with the patch it is turned into:

     56    raise/ -> read all, write all
     57    qsort/ -> read all, write all
     78    _obstack_begin/ -> read all, write all
    115    fputs_unlocked/ -> read all, write all
    123    obstack_free/ -> read all, write all
    132    __gmpz_set_si/ -> read all, write all
    141    dcgettext/ -> read all, write all
    154    __gmpz_clear/ -> read all, write all
    156    __builtin_fputs/ -> read all, write all
    175    __gmpz_init/ -> read all, write all
    214    realloc/ -> read all, write all
    230    malloc/ -> read all, write all
    295    snprintf/ -> read all, write all
    322    sprintf/ -> read all, write all
    388    _obstack_newchunk/ -> read all, write all
    404    __overflow/ -> read all, write all
    464    __builtin_fwrite_unlocked/ -> read all, write all
    621    __builtin_fputc/ -> read all, write all
   2295    __builtin_fwrite/ -> read all, write all
   2330    fprintf/ -> read all, write all
   2840    free/ -> read all, write all
   5127    indirect call -> read all

Note that possibly malloc/free and I/O can also be flagged leaf.  I was concerned
about memory profiles or user defined streams. 

Many of the common cases are gone (mostly string functions, alloca and abort),
reducing the number of cases we give up on direct call by 60%.  This leads to
about 5% increase of functions where we analyze read/written to be something
else than ALL (this is number after computing transitive closure).

The savings on cc1 binary compile with the leaf attribute enabled comopared to
without are about 3kb.

It is not earthshaking improvement, but ipa-reference is somewhat toy quality
pass still (it does not analyze structures we memset into that stops it on most
cc1 global structures that matters, like x_rtl) and the attribute is supposed
to improve control flow information for about every pass that cares about side
effects of external calls (i.e. IPA-PTA once it understands unit escape,
profiling, datastructure reorganization passes etc.)

Honza

	* doc/extend.texi: (attribute leaf): Document.

	* tree.c (local_define_builtin): Handle ECF_LEAF.
	(build_common_builtin_nodes): Set it on all relevant builtins.
	* tree.h (ECF_LEAF): New.
	* builtin-attrs.def (ATTR_LEAF): New attribute.
	(ATTR_NOVOPS_LEAF_LIST, ATTR_LEAF_LIST, ATTR_NOTHROW_LEAF_LIST,
	ATTR_CONST_NOTHROW_LEAF_LIST, ATTR_PURE_NOTHROW_LEAF_LIST,
	ATTR_PURE_NOTHROW_NOVOPS_LEAF_LIST, ATTR_NORETURN_NOTHROW_LEAF_LIST,
	ATTR_MALLOC_NOTHROW_LEAF_LIST, ATTR_SENTINEL_NOTHROW_LEAF_LIST,
	ATTR_NOTHROW_NONNULL_LEAF, ATTR_CONST_NOTHROW_NONNULL_LEAF,
	ATTR_CONST_NOTHROW_TYPEGENERIC_LEAF, ATTR_PURE_NOTHROW_NONNULL_LEAF,
	ATTR_MALLOC_NOTHROW_NONNULL_LEAF): New attribute lists.
	* sync-builtins.def: Annotate all builtins by leaf.
	* omp-builtins.def: Annotate all builtins by leaf.
	* ipa-reference.c (propagate_bits): For leaf calls propagate ever overwrittable
	and unavailable functions.
	(ipa_init): Put all_module_statics into optimization_summary_obstack.
	(copy_global_bitmap): Do not copy all_module_statics.
	(read_write_all_from_decl): Use cgraph_node argument; handle ECF_LEAF.
	(propagate): Handle overwritable and unavailable leaf functions;
	initialize global info for overwritable and unavailable leaf functions;
	do not free all module statics.
	* builtins.def: Annotate relevant builtins with leaf attribute.
	(ATTR_MATHFN_ERRNO, ATTR_MATHFN_FPROUNDING,
	ATTR_MATHFN_FPROUNDING_ERRNO, ATTR_MATHFN_FPROUNDING_STORE): Make
	leaf.
	* calls.c (flags_from_decl_or_type): Handle leaf.

	* c-family/c-common.c (handle_leaf_attribute): New.
	(c_common_attribute_table): Add leaf.

	* gcc-interface/utils.c (handle_leaf_attribute): New function.
	(gnat_internal_attribute_tables): Add leaf.

	* lto-lang.c (handle_leaf_attribute): New function.
	(lto_attribute_tables): Add leaf.

	* f95-lang.c (gfc_define_builtin): Make leaf.
	(gfc_init_builtin_functions): Handle only ATTR_CONST_NOTHROW_LEAF_LIST
	and ATTR_NOTHROW_LEAF_LIST.
	(DEF_SYNC_BUILTIN): Check ATTR_CONST_NOTHROW_LEAF_LIST.
	(DEF_GOMP_BUILTIN): Likewise.

[RFC/RFA] New function attribute "leaf"

Commit Message

Patch