This patch significantly improves performance of memmem using a novel
modified Horspool algorithm. Needles up to size 256 use a bad-character
table indexed by hashed pairs of characters to quickly skip past mismatches.
Long needles use a self-adapting filtering step to avoid comparing the whole
needle repeatedly.
By limiting the needle length to 256, the shift table only requires 8 bits
per entry, lowering preprocessing overhead and minimizing cache effects.
This limit also implies worst-case performance is linear.
Small needles up to size 2 use a dedicated linear search. Very long needles
use the Two-Way algorithm (to avoid increasing stack size inlining is now disabled).
The performance gain is 6.6 times on English text on AArch64 using random
needles with average size 8 (this is even faster than the recently improved strstr
algorithm, so I'll update that in the near future).
The size-optimized memmem has also been rewritten from scratch to get a
2.7x performance gain.
Tested against GLIBC testsuite and randomized tests.
Message-Id: <DB5PR08MB1030649D051FA8532A4512C883B20@DB5PR08MB1030.eurprd08.prod.outlook.com>
FreeBSD uses a 64-bit ino_t since 2017-05-23. We need this for the
pipe() support in libbsd.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Various structures exported by sysctl_rtsock() contain padding fields
which were not being zeroed.
Reported by: Thomas Barabosch, Fraunhofer FKIE
Reviewed by: ae
MFC after: 3 days
Security: kernel memory disclosure
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D18333
The threshold value at which powf overflows depends on the rounding mode
and the current check did not take this into account. So when the result
was rounded away from zero it could become infinity without setting
errno to ERANGE.
Example: pow(0x1.7ac7cp+5, 23) is 0x1.fffffep+127 + 0.1633ulp
If the result goes above 0x1.fffffep+127 + 0.5ulp then errno is set,
which is fine in nearest rounding mode, but
powf(0x1.7ac7cp+5, 23) is inf in upward rounding mode
powf(-0x1.7ac7cp+5, 23) is -inf in downward rounding mode
and the previous implementation did not set errno in these cases.
The fix tries to avoid affecting the common code path or calling a
function that may introduce a stack frame, so float arithmetics is used
to check the rounding mode and the threshold is selected accordingly.
for {n,u,m}stosbt
Integer overflows and wrong constants limited the accuracy of these
functions and created situatiosn where sbttoXs(Xstosbt(Y)) != Y. This
was especailly true in the ns case where we had millions of values
that were wrong.
Instead, used fixed constants because there's no way to say ceil(X)
for integer math. Document what these crazy constants are.
Also, use a shift one fewer left to avoid integer overflow causing
incorrect results, and adjust the equasion accordingly. Document this.
Allow times >= 1s to be well defined for these conversion functions
(at least the Xstosbt). There's too many users in the tree that they
work for >= 1s.
This fixes a failure on boot to program firmware on the mlx4
NIC. There was a msleep(1000) in the code. Prior to my recent rounding
changes, msleep(1000) worked, but msleep(1001) did not because the old
code rounded to just below 2^64 and the new code rounds to just above
it (overflowing, causing the msleep(1000) to really sleep 1ms).
A test program to test all cases will be committed shortly. The test
exaustively tries every value (thanks to bde for the test).
Sponsored by: Netflix, Inc
Differential Revision: https://reviews.freebsd.org/D18051
of the result rather than the floor(). Returning the floor means that
sbttoX(Xtosbt(y)) != y for almost all values of y. In practice, this
results in a difference of at most 1 in the lsb of the sbintime_t. This
difference is meaningless for all current users of these functions, but
is important for the newly introduced sysctl conversion routines which
implicitly rely on the transformation being idempotent.
Sponsored by: Netflix, Inc
Mainly focus on files that use BSD 3-Clause license.
The Software Package Data Exchange (SPDX) group provides a specification
to make it easier for automated tools to detect and summarize well known
opensource licenses. We are gradually adopting the specification, noting
that the tags are considered only advisory and do not, in any way,
superceed or replace the license texts.
Special thanks to Wind River for providing access to "The Duke of
Highlander" tool: an older (2014) run over FreeBSD tree was useful as a
starting point.
and decimal time units. Use them in some existing code that is
vulnerable to roundoff errors.
The existing constant SBT_1NS is a honeypot, luring unsuspecting folks into
writing code such as long_timeout_ns*SBT_1NS to generate the argument for a
sleep call. The actual value of 1ns in sbt units is ~4.3, leading to a
large roundoff error giving a shorter sleep than expected when multiplying
by the trucated value of 4 in SBT_1NS. (The evil honeypot aspect becomes
clear after you waste a whole day figuring out why your sleeps return early.)
Renumber cluase 4 to 3, per what everybody else did when BSD granted
them permission to remove clause 3. My insistance on keeping the same
numbering for legal reasons is too pedantic, so give up on that point.
Submitted by: Jan Schaumann <jschauma@stevens.edu>
Pull Request: https://github.com/freebsd/freebsd/pull/96
- Add CLOCK_REALTIME_COARSE, CLOCK_MONOTONIC_RAW,
CLOCK_MONOTONIC_COARSE and CLOCK_BOOTTIME
- Guard new values with __GNU_VISIBLE
- Add CLOCK_REALTIME_COARSE as (clockid_t) 0 for simplicity
(It allows to have all values < 8 and so be used as array
index into an array of clocks)
- Fix macro bracketing
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
A previous commit introduced the ability to use the semi-hosting
SYS_EXIT_EXTENDED operation to libgloss, this commit adds the same
ability to the sys/arm/ backend so that building newlib only will
provide the same capabilities.
This patch fixes an issue in the previous memset loop change. If the
zva size is >= 256 and there are more than 64 bytes left in the
tail, we could enter the loop and thus need to rebias dst by 32 as
well.
Since no known CPUs use this size this can't be tested natively, so I've
tested it on a simulator initialized with a large zva size.
--
Do not define __ATTRIBUTE_IMPURE_PTR__ for RTMES on the v850 target.
The previous definition lead to the following linker error in
combination with -fdata-sections:
relocation truncated to fit: R_V810_GPWLO_1 against symbol
`_global_impure_ptr' defined in .rodata._global_impure_ptr section in
libc.a(lib_a-impure.o)
relocation truncated to fit: R_V810_GPWLO_1 against symbol `_impure_ptr'
defined in .data._impure_ptr section in libc.a(lib_a-impure.o)
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
The <machine/param.h> header file exposes some unrelated stuff not
covered by C or POSIX. Avoid its use in <sys/_cpuset.h> since it is
included in <rtems.h>.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
This fixes an ineffiency in the non-zero memset. Delaying the writeback
until the end of the loop is slightly faster on some cores - this shows
~5% performance gain on Cortex-A53 when doing large non-zero memsets.
Tested against the GLIBC testsuite.
The new GCC port for OpenRISC will use the init_fini_array only and not
provide the init() and fini() functions. Disable the function usage by
default as its no longer needed.
Signed-off-by: Stafford Horne <shorne@gmail.com>
The malloc, alloc_size and alloc_aligned attributes must be only used in
case the function returns the pointer to the allocated memory.
See also:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87683
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
The following FreeBSD kernel methods are not in any standard and
prototypes/definitions were leaking into application space:
+ round_page()
+ trunc_page()
+ atop()
+ ptoa()
+ pgtok()
v3: Add support for read ahead using strnlen, giving an additional 25% speedup
on large inputs (both short and long needles).
This patch significantly improves performance of strstr by using Sunday's
Quick-Search algorithm. Due to its simplicity it has the best average
performance of string matching algorithms on almost all inputs. It uses a
bad-character shift table to skip past mismatches.
The needle length is limited to 254 - this reduces the shift table memory
4 to 8 times, lowering preprocessing overhead and minimizing cache effects.
The limit also implies its worst-case performance is linear.
Larger needles are processed by the Two-Way algorithm. The macro AVAILABLE
has been improved to use strnlen to read the input in chunks. This results
in a 2.5 times speedup for large needles, reducing the performance drop when
the Quick-Search algorithm can't be used.
The code for 1-4 byte needles has been simplified and now uses unsigned
char. Since the optimized code relies on 8-bit chars, we defer to the
size-optimized implementation if CHAR_BIT > 8.
The performance gain of finding a set of randomly chosen words of size 8 in
256 bytes of English text is 14 times on AArch64. For longer haystacks the
gain is well over 20 times.
The size-optimized strstr has also been rewritten from scratch to improve
performance. On the same test the performance gain is 69%.
Tested against GLIBC testsuite, randomized tests and the GNULIB strstr test
(https://git.savannah.gnu.org/cgit/gnulib.git/tree/tests/test-strstr.c).
--
Use existing HAVE_OPENDIR define to determine if a generic
implementation should be provided. Cygwin for example has its own
implementation of opendir() and dirfd().
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
This is used by the file system support of libstdc++ for example. Use
content from latest FreeBSD <sys/dirent.h>
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Move common content of the various <sys/dirent.h> and the latest FreeBSD
<dirent.h> to <dirent.h>.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Use O_RDONLY since you are not supposed to write to a directory.
Use O_DIRECTORY as mandated by POSIX (The Open Group Base Specifications
Issue 7, 2018 edition IEEE Std 1003.1-2017):
"If the type DIR is implemented using a file descriptor, the descriptor
shall be obtained as if the O_DIRECTORY flag was passed to open()."
Use O_CLOEXEC as mandated by POSIX:
"When a file descriptor is used to implement the directory stream, it
behaves as if the FD_CLOEXEC had been set for the file descriptor."
Drop the fcntl() call in favour of O_CLOEXEC.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Make the POSIX O_CLOEXEC, O_NOFOLLOW, O_DIRECTORY, O_EXEC, and O_SEARCH
open() flags available also to non-Cygwin systems.
Make the BSD/glibc O_DIRECT open() flag available also to non-Cygwin
systems.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
Commit fbace81684
("Import correctly working strtold from David M. Gay.")
introduced two new files, strtorx.c and strtodg.c. The functions
are only called from strtold.c. However, while strtold.c is only
built if HAVE_LONG_DOUBLE is defined, the patch erroneously added
the two new files to GENERAL_SOURCES unconditionally.
Fix this by building both files only if HAVE_LONG_DOUBLE has been
defined.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Commit 6c212a8b78
("Fix strtod ("nan") and strtold ("nan") returns wrong negative NaN")
introduced an unconditional dependency to nanl and, in turn, to libm.
Rather than including nanl in libc as well, just call __builtin_nanl
from here. Requires GCC 3.3 or later.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
Drop Cygwin-specific nanl in favor of a generic implementation
in newlib. Requires GCC 3.3 or later.
Signed-off-by: Corinna Vinschen <corinna@vinschen.de>
These attributes help static analysis tools to produce less false
positives, e.g. double free warnings.
Signed-off-by: Sebastian Huber <sebastian.huber@embedded-brains.de>
AngelSWI_Reason_ReportException does not return accoring to the ARM
documentation, so it is valid to mark _kill() as noreturn. This way,
the compiler does not warn about _exit() returning a value despite
being noreturn.
2018-10-01 Christophe Lyon <christophe.lyon@linaro.org>
* libgloss/arm/_exit.c (_exit): Declare _kill() as noreturn.
* libgloss/arm/_exit.c (_kill): Likewise. Remove the return
statements.
* newlib/libc/sys/arm/syscalls.c (_kill): Likewise..
While working on the strstr patch I noticed several copyright headers
of the new math functions are missing closing quotes after ``AS IS.
I've added these. Also update spellings of Arm Ltd in a few places
(but still use ARM LTD in upper case portion). Finally add SPDX
identifiers to make everything consistent.
It's been a while... I see the CRIS port broke with the
time_t-default-to-64-bit change, observable by a few test-cases in the gcc
fortran(!) tests failing, regressing when trying a recent newlib.
This is a two-part belt-and-suspenders change: adjust the CRIS port
gettimeofday syscall (the only one in newlib/CRIS passing a time_t or
struct timeval) to handle a userspace 64-bit time_t and secondly default
time_t to 32-bit long anyway. I considered making the local
"kernel_timeval" copy in _gettimeofday conditional on (userspace) time_t
being 64 bits, but thought it not worth bothering with the few move insns.
The effect of a 64-bit time_t is however observable as longer simulation
time when running the gcc testsuite and as bigger binaries without any
actual upside from the larger time_t size, so I thought better make the
default for this port go back to being a "long" again.
Tested by running the gcc testsuite over the three combinations of two
parts of the patch and observing the expected changes. Committed.
newlib:
* configure.host (cris, crisv32): Default to "long" time_t.
Signed-off-by: Hans-Peter Nilsson <hp@axis.com>
hash.h: Use 32-bit type for data stored on disk, so code
works for 16 and 64-bit targets. Reduce maximum bucket size on 16-bit
targets, so it fits in available memory.
hash.c: Check bucket size isn't too big for target.
hash_buf.c: Fix overflow warning on 16-bit targets.
When __HAVE_LOCALE_INFO__ is not selected, directly access the
existing _ctype_ variable from __locale_ctype_ptr() and
__locale_ctype_ptr_l(), eliminating the need for any locale or reent
structure
Signed-off-by: Keith Packard <keithp@keithp.com>
v2:
locale: fix conflict with __locale_ctype_ptr macro
If we are building without __HAVE_LOCALE_INFO__, there is a
macro providing __locale_ctype_ptr which in turn fouls up this
declaration.
Signed-off-by: Michael Lyle <mlyle@lyle.org>
The string/float conversion functions need to get the locale decimal
point. Instead of calling __localeconv_l (which copies locale data
into lconv form from __get_numeric_locale), use __get_numeric_locale
directly.
Signed-off-by: Keith Packard <keithp@keithp.com>
This makes sure any system-defined limits are specified before the
defaults are checked. Without this, ARG_MAX and PATH_MAX end up
getting the default definitions from limits.h rather than the defines
from syslimits.h. This could potentially cause problems when
different files used different values for the same name.
Signed-off-by: Keith Packard <keithp@keithp.com>
Wilco Dijkstra
Jeff Johnston
Sebastian Huber
markj
Jon Beniston
Szabolcs Nagy
imp
ian
pfg
Corinna Vinschen
Matthew Malcomson
Stafford Horne
Joel Sherrill
Christophe Lyon
Hans-Peter Nilsson
Keith Packard