This sequence makes no less than 2 calls memmoving 0 bytes:
vector<char> content;
content.resize(12);
>From *0x0 no less!
On a 64 bit system, the effect is far less pronounced (25%), but on 32 bits,
this loop takes 2.3 seconds on a vanilla g++;
for(int n=0; n < 10000000; ++n) {
vector<char> content;
content.resize(12);
}
But if we add:
if(__last != __first)
before the __builtin_memmove on line 377 of stl_algobase.h, it takes 0.9
seconds!
On a real 32 bit benchmark of PowerDNS, this single line gave a 9% performance
boost.
Seems worthwhile. Andrew Pinski suggested _M_fill_insert may need to do this
check instead.
#0 *__GI_memmove (dest=0x8b8dc18, src=0x0, len=0) at memmove.c:47
#1 0x080557f0 in std::__copy_move<false, true,
std::random_access_iterator_tag>::__copy_m<unsigned char> (__first=0x0,
__last=0x0,
__result=0x8b8dc18 "") at /usr/include/c++/4.3/bits/stl_algobase.h:378
#2 0x08055a3f in std::__copy_move_a<false, unsigned char*, unsigned char*>
(__first=0x0, __last=0x0, __result=0x8b8dc18 "")
at /usr/include/c++/4.3/bits/stl_algobase.h:397
#3 0x08055a7e in std::__copy_move_a2<false, unsigned char*, unsigned char*>
(__first=0x0, __last=0x0, __result=0x8b8dc18 "")
at /usr/include/c++/4.3/bits/stl_algobase.h:436
#4 0x08055ab9 in std::copy<unsigned char*, unsigned char*> (__first=0x0,
__last=0x0, __result=0x8b8dc18 "")
at /usr/include/c++/4.3/bits/stl_algobase.h:467
#5 0x08055ade in std::__uninitialized_copy<true>::uninitialized_copy<unsigned
char*, unsigned char*> (__first=0x0, __last=0x0,
__result=0x8b8dc18 "") at /usr/include/c++/4.3/bits/stl_uninitialized.h:98
#6 0x08055aff in std::uninitialized_copy<unsigned char*, unsigned char*>
(__first=0x0, __last=0x0, __result=0x8b8dc18 "")
at /usr/include/c++/4.3/bits/stl_uninitialized.h:122
#7 0x08055b20 in std::__uninitialized_copy_a<unsigned char*, unsigned char*,
unsigned char> (__first=0x0, __last=0x0, __result=0x8b8dc18 "")
at /usr/include/c++/4.3/bits/stl_uninitialized.h:262
#8 0x08055b48 in std::__uninitialized_move_a<unsigned char*, unsigned char*,
std::allocator<unsigned char> > (__first=0x0, __last=0x0,
__result=0x8b8dc18 "", __all...@0xff815c2c) at
/usr/include/c++/4.3/bits/stl_uninitialized.h:272
#9 0x0805af27 in std::vector<unsigned char, std::allocator<unsigned char>
>::_M_fill_insert (this=0xff815c2c, __position={_M_current = 0x0},
__n=12, _...@0xff815ab0) at /usr/include/c++/4.3/bits/vector.tcc:399
#10 0x0805b034 in std::vector<unsigned char, std::allocator<unsigned char>
>::insert (this=0xff815c2c, __position={_M_current = 0x0}, __n=12,
_...@0xff815ab0) at /usr/include/c++/4.3/bits/stl_vector.h:792
#11 0x0805b0b7 in std::vector<unsigned char, std::allocator<unsigned char>
>::resize (this=0xff815c2c, __new_size=12, __x=0 '\0')
at /usr/include/c++/4.3/bits/stl_vector.h:509
#12 0x08083057 in DNSPacketWriter (this=0xff815b9c, conte...@0xff815c2c,
qna...@0xff815c04, qtype=6, qclass=1, opcode=0 '\0') at dnswriter.cc:21
#13 0x0804d449 in makeRootReferral () at speedtest.cc:259
#14 0x080601e9 in RootRefTest::operator() (this=0xff815e87) at speedtest.cc:329
#15 0x080602cc in doRun<RootRefTest> (c...@0xff815e87, mseconds=100) at
speedtest.cc:37
#16 0x0804da3e in main () at speedtest.cc:443
--
Summary: vector<>::resize() from an empty vector calls memmove
for 0 bytes, wasting a lot of cpu time in a production
PowerDNS
Product: gcc
Version: 4.4.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: ahu at ds9a dot nl
GCC build triplet: x86_64-unknown-linux-gnu
GCC host triplet: x86_64-unknown-linux-gnu
GCC target triplet: x86_64-unknown-linux-gnu
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41267
