Improve strstr performance. Strstr tends to be slow because it uses
many calls to memchr and a slow byte loop to scan for the next match.
Performance is significantly improved by using strnlen on larger blocks
and using strchr to search for the next matching character. strcasestr
can also use strnlen to scan ahead, and memmem can use memchr to check
for the next match.
On the GLIBC bench tests the performance gains on Cortex-A72 are:
strstr: +25%
strcasestr: +4.3%
memmem: +18%
On a 256KB dataset strstr performance improves by 67%, strcasestr by 47%.
Reviewd-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
These are a grab bag of changes where the testsuite was using internal
symbols of some variety, but this was straightforward to fix, and the
fixed code should work with or without the change to compile the
testsuite under _ISOMAC.
Four of these are just more #include adjustments, but I want to highlight
sysdeps/powerpc/fpu/tst-setcontext-fpscr.c, which appears to have been
written before the advent of sys/auxv.h. I think a big chunk of this file
could be replaced by a simple call to getauxval, but I'll let someone who
actually has a powerpc machine to test on do that.
dlfcn/tst-dladdr.c was including ldsodefs.h just so it could use
DL_LOOKUP_ADDRESS to print an additional diagnostic; as requested by Carlos,
I have removed this.
math/test-misc.c was using #ifndef NO_LONG_DOUBLE, which is an internal
configuration macro, to decide whether to do certain tests involving
'long double'. I changed the test to #if LDBL_MANT_DIG > DBL_MANT_DIG
instead, which uses only public float.h macros and is equivalent on
all supported platforms. (Note that NO_LONG_DOUBLE doesn't mean 'the
compiler doesn't support long double', it means 'long double is the
same as double'.)
tst-writev.c has a configuration macro 'ARTIFICIAL_LIMIT' that the
Makefiles are expected to define, and sysdeps/unix/sysv/linux/Makefile
was using the internal __getpagesize in the definition; changed to
sysconf(_SC_PAGESIZE) which is the POSIX equivalent.
ia64-linux doesn't supply 'clone', only '__clone2', which is not
defined in the public headers(!) All the other clone tests have local
extern declarations of __clone2, but tst-clone.c doesn't; it was
getting away with this because include/sched.h does declare __clone2.
* nss/tst-cancel-getpwuid_r.c: Include nss.h.
* string/strcasestr.c: No need to include config.h.
* sysdeps/powerpc/fpu/tst-setcontext-fpscr.c: Include
sys/auxv.h. Don't include sysdep.h.
* sysdeps/powerpc/tst-set_ppr.c: Don't include dl-procinfo.h.
* dlfcn/tst-dladdr.c: Don't include ldsodefs.h. Don't use
DL_LOOKUP_ADDRESS.
* math/test-misc.c: Instead of testing NO_LONG_DOUBLE, test whether
LDBL_MANT_DIG is greater than DBL_MANT_DIG.
* sysdeps/unix/sysv/linux/Makefile (CFLAGS-tst-writev.c): Use
sysconf (_SC_PAGESIZE) instead of __getpagesize in definition
of ARTIFICIAL_LIMIT.
* sysdeps/unix/sysv/linux/tst-clone.c [__ia64__]: Add extern
declaration of __clone2.
2008-03-29 Eric Blake <ebb9@byu.net>
Rewrite string searches to O(n) rather than O(n^2).
* string/str-two-way.h: New file. For linear fixed-allocation
string searching.
* string/memmem.c: New implementation.
* string/strstr.c: New implementation.
* string/strcasestr.c: New implementation.
* sysdeps/posix/getaddrinfo.c (getaddrinfo): Call _res_hconf_init
Wilco Dijkstra
Joseph Myers
Zack Weinberg
Allan McRae
Maxim Kuvyrkov
Paul Eggert
Ulrich Drepper
H.J. Lu
Jakub Jelinek