[
https://issues.apache.org/jira/browse/STDCXX-914?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12646021#action_12646021
]
sebor edited comment on STDCXX-914 at 11/8/08 4:21 PM:
--------------------------------------------------------------
Here's a superficially tested patch to optimize
{{\_\_rw_locale::_C_is_managed()}} and {{\_\_rw_locale::_C_manage()}} in
{{[src/locale_body.cpp|http://svn.eu.apache.org/viewvc/stdcxx/trunk/src/locale_body.cpp?revision=651334&view=markup]}}.
It improves the performance of the test case by about 45% (down from 18.905s
to 12.147s on an Intel Core 2 6600 running at 2.40GHz) by having
{{\_\_rw_locale::_C_is_managed()}} avoid expensive tests for named faces in the
"C" locale and by using a more efficient way to detect the classic locale in
{{\_\_rw_locale::_C_manage()}} when invoked from {{locale::~locale()}}.
\\
\\
{noformat}
Index: src/locale_body.cpp
===================================================================
--- src/locale_body.cpp (revision 712407)
+++ src/locale_body.cpp (working copy)
@@ -859,7 +859,22 @@
return tmp;
}
+ if (plocale && plocale == classic) {
+ // optimize the "destruction" of the classic C locale
+ // the object is never destroyed and its reference count
+ // never drops to 0
+ _RWSTD_ASSERT (__rw_is_C (locname));
+ _RWSTD_ASSERT (__rw_is_C (plocale->_C_name));
+ const size_t ref =
+ _RWSTD_ATOMIC_PREDECREMENT (plocale->_C_ref, false);
+
+ _RWSTD_ASSERT (ref + 1U != 0);
+ _RWSTD_UNUSED (ref);
+
+ return 0;
+ }
+
// re-entrant to protect static local data structures
// (not the locales themselves)
_RWSTD_MT_STATIC_GUARD (_RW::__rw_locale);
@@ -1066,6 +1081,15 @@
return false;
}
+ _RWSTD_ASSERT (0 == _C_usr_facets);
+
+ if (_C_all == _C_std_facet_bits && 0 == _C_byname_facet_bits) {
+ // optimized for the C locale
+ _RWSTD_ASSERT (__rw_is_C (_C_name));
+
+ return true;
+ }
+
// unless all facets in the same category come either from
// the C locale or from some named locale the locale object
// containing the facets is not managed (this test doesn't
{noformat}
With the patch applied, the top 12 list looks like so:
\\
\\
{noformat}
Flat profile:
Each sample counts as 0.01 seconds.
% cumulative self self total
time seconds seconds calls ms/call ms/call name
16.70 0.97 0.97 50000000 0.00 0.00
__rw::__rw_locale::_C_manage(__rw::__rw_locale*, char const*)
12.57 1.70 0.73 10000000 0.00 0.00 std::istream&
std::operator>>(std::istream&, std::string&)
8.43 2.19 0.49 10000000 0.00 0.00
std::num_put::_C_put(std::ostreambuf_iterator, std::ios_base&, char, int, void
const*) const
7.06 2.60 0.41 10000001 0.00 0.00
std::string::operator=(std::string const&)
6.45 2.98 0.38 10000000 0.00 0.00 std::string
lex_cast<std::string, long>(long const&)
5.34 3.29 0.31 10000000 0.00 0.00 __rw::__rw_dtoa(char*,
unsigned long, unsigned int)
4.65 3.56 0.27 main
4.30 3.81 0.25 10000000 0.00 0.00 std::ostream&
__rw::__rw_insert(std::ostream&, long)
3.27 4.00 0.19 10000000 0.00 0.00
std::locale::locale(std::locale const&)
3.01 4.17 0.18 10000000 0.00 0.00 std::stringbuf::str(char
const*, unsigned long)
2.75 4.33 0.16 30000000 0.00 0.00
__rw::__rw_locale::_C_is_managed(int) const
2.75 4.49 0.16 30000000 0.00 0.00 std::locale::~locale()
{noformat}
was (Author: sebor):
Here's a superficially tested patch to optimize
{{\_\_rw_locale::_C_is_managed()}} and {{\_\_rw_locale::_C_manage()}} in
{{[src/locale_body.cpp|http://svn.eu.apache.org/viewvc/stdcxx/trunk/src/locale_body.cpp?revision=651334&view=markup]}}.
It improves the performance of the test case by about 45% (down from 18.905s
to 12.147s on an Intel Core 2 6600 running at 2.40GHz) by having
{{\_\_rw_locale::_C_is_managed()}} avoid expensive tests for named faces in the
"C" locale and by using a more efficient way to detect the classic locale in
{{\_\_rw_locale::_C_manage()}} when invoked from {{locale::~locale()}}.
\\
\\
{noformat}
Index: src/locale_body.cpp
===================================================================
--- src/locale_body.cpp (revision 712407)
+++ src/locale_body.cpp (working copy)
@@ -859,7 +859,22 @@
return tmp;
}
+ if (plocale && plocale == classic) {
+ // optimize the "destruction" of the classic C locale
+ // the object is never destroyed and its reference count
+ // never drops to 0
+ _RWSTD_ASSERT (__rw_is_C (locname));
+ _RWSTD_ASSERT (__rw_is_C (plocale->_C_name));
+ const size_t ref =
+ _RWSTD_ATOMIC_PREDECREMENT (plocale->_C_ref, false);
+
+ _RWSTD_ASSERT (ref + 1U != 0);
+ _RWSTD_UNUSED (ref);
+
+ return 0;
+ }
+
// re-entrant to protect static local data structures
// (not the locales themselves)
_RWSTD_MT_STATIC_GUARD (_RW::__rw_locale);
@@ -1066,6 +1081,15 @@
return false;
}
+ _RWSTD_ASSERT (0 == _C_usr_facets);
+
+ if (_C_all == _C_std_facet_bits && 0 == _C_byname_facet_bits) {
+ // optimized for the C locale
+ _RWSTD_ASSERT (__rw_is_C (_C_name));
+
+ return true;
+ }
+
// unless all facets in the same category come either from
// the C locale or from some named locale the locale object
// containing the facets is not managed (this test doesn't
{noformat}
With the patch applied, the top 12 list looks like so:
\\
\\
{noformat}
Flat profile:
Each sample counts as 0.01 seconds.
% cumulative self self total
time seconds seconds calls ms/call ms/call name
16.70 0.97 0.97 50000000 0.00 0.00
__rw::__rw_locale::_C_manage(__rw::__rw_locale*, char const*)
12.57 1.70 0.73 10000000 0.00 0.00 std::basic_istream<char,
std::char_traits<char> >& std::operator>><char, std::char_traits<char>,
std::allocator<char> >(std::basic_istream<char, std::char_traits<char> >&,
std::basic_string<char, std::char_traits<char>, std::allocator<char> >&)
8.43 2.19 0.49 10000000 0.00 0.00 std::num_put<char,
std::ostreambuf_iterator<char, std::char_traits<char> >
>::_C_put(std::ostreambuf_iterator<char, std::char_traits<char> >,
std::ios_base&, char, int, void const*) const
7.06 2.60 0.41 10000001 0.00 0.00
std::string::operator=(std::string const&)
6.45 2.98 0.38 10000000 0.00 0.00 std::string
lex_cast<std::string, long>(long const&)
5.34 3.29 0.31 10000000 0.00 0.00 __rw::__rw_dtoa(char*,
unsigned long, unsigned int)
4.65 3.56 0.27 main
4.30 3.81 0.25 10000000 0.00 0.00 std::basic_ostream<char,
std::char_traits<char> >& __rw::__rw_insert<char, std::char_traits<char>,
long>(std::basic_ostream<char, std::char_traits<char> >&, long)
3.27 4.00 0.19 10000000 0.00 0.00
std::locale::locale(std::locale const&)
3.01 4.17 0.18 10000000 0.00 0.00
std::basic_stringbuf<char, std::char_traits<char>, std::allocator<char>
>::str(char const*, unsigned long)
2.75 4.33 0.16 30000000 0.00 0.00
__rw::__rw_locale::_C_is_managed(int) const
2.75 4.49 0.16 30000000 0.00 0.00 std::locale::~locale()
{noformat}
> sstream ctors inefficient in reentrant modes
> --------------------------------------------
>
> Key: STDCXX-914
> URL: https://issues.apache.org/jira/browse/STDCXX-914
> Project: C++ Standard Library
> Issue Type: Improvement
> Components: 27. Input/Output
> Affects Versions: 4.1.2, 4.1.3, 4.1.4, 4.2.0, 4.2.1
> Reporter: Martin Sebor
> Priority: Critical
> Fix For: 4.2.2
>
> Attachments: stdcxx-914-gprof-gcc-4.1.2-12D.txt,
> stdcxx-914-gprof-gcc-4.3.0-12S.txt
>
> Original Estimate: 12h
> Time Spent: 2.5h
> Remaining Estimate: 9.5h
>
> As discussed in this [thread|http://markmail.org/message/hqlsw5dq23gx7d4o]
> stream ctors in thread-safe builds are inefficient due to the initialization
> of the mutex data member in every stream, even in those that never use it. As
> soon as binary compatibility rules permit it we should remove the mutex
> and/or defer its initialization until it's needed. It might be possible to
> implement the deferred initialization as early as 4.2.2, or maybe 4.3.
> Complete removal will need to wait until 5.0.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
