This variant of strlen uses vector loads and operations to reduce the
size of the code and also eliminate the non-ascii fallback. This
works very well for falkor because of its two vector units and
efficient vector ops. In the best case it reduces latency of cases in
bench-strlen by 48%, with gains throughout the benchmark.
strlen-walk also sees uniform gains in the 5%-15% range.
Overall the routine appears to work better than the stock one for falkor
regardless of the benchmark, length of string or cache state.
The same cannot be said of a53 and a72 though. a53 performance was
greatly reduced and for a72 it was a bit of a mixed bag, slightly on the
negative side but I reckon it might be fast in some situations.
* sysdeps/aarch64/strlen.S (__strlen): Rename to STRLEN.
[!STRLEN](STRLEN): Set to __strlen.
* sysdeps/aarch64/multiarch/strlen.c: New file.
* sysdeps/aarch64/multiarch/strlen_generic.S: Likewise.
* sysdeps/aarch64/multiarch/strlen_asimd.S: Likewise.
* sysdeps/aarch64/multiarch/ifunc-impl-list.c
(__libc_ifunc_impl_list): Add strlen.
* sysdeps/aarch64/multiarch/Makefile (sysdep_routines): Add
strlen_generic and strlen_asimd.
Reviewed-By: szabolcs.nagy@arm.com
CC: pinskia@gmail.com
The internal functions __kernel_sinf and __kernel_cosf are used only by
lgammaf_r. Removing the internal functions and using the generic sinf
and cosf is better overall. Benchmarking on Cortex-A72 shows the generic
sinf and cosf are 1.4x and 2.3x faster in the range |x| < PI/4, and 0.66x
and 1.1x for |x| < PI/2, so it should make lgammaf_r faster on average.
GLIBC regression tests pass on AArch64.
* sysdeps/ieee754/flt-32/e_lgammaf_r.c (sin_pif): Use __sinf/__cosf.
* sysdeps/ieee754/flt-32/k_cosf.c (__kernel_cosf): Remove all code.
* sysdeps/ieee754/flt-32/k_sinf.c (__kernel_sinf): Likewise.
Fix a few missing spaces, it's now identical to the regenerated version.
Passes GLIBC tests on x64.
* sysdeps/x86_64/fpu/libm-test-ulps: Regenerate to fix spaces.
The second patch improves performance of sinf and cosf using the same
algorithms and polynomials. The returned values are identical to sincosf
for the same input. ULP definitions for AArch64 and x64 are updated.
sinf/cosf througput gains on Cortex-A72:
* |x| < 0x1p-12 : 1.2x
* |x| < M_PI_4 : 1.8x
* |x| < 2 * M_PI: 1.7x
* |x| < 120.0 : 2.3x
* |x| < Inf : 3.0x
* NEWS: Mention sinf, cosf, sincosf.
* sysdeps/aarch64/libm-test-ulps: Update ULP for sinf, cosf, sincosf.
* sysdeps/x86_64/fpu/libm-test-ulps: Update ULP for sinf and cosf.
* sysdeps/x86_64/fpu/multiarch/s_sincosf-fma.c: Add definitions of
constants rather than including generic sincosf.h.
* sysdeps/x86_64/fpu/s_sincosf_data.c: Remove.
* sysdeps/ieee754/flt-32/s_cosf.c (cosf): Rewrite.
* sysdeps/ieee754/flt-32/s_sincosf.h (reduced_sin): Remove.
(reduced_cos): Remove.
(sinf_poly): New function.
* sysdeps/ieee754/flt-32/s_sinf.c (sinf): Rewrite.
This patch updates sysdeps/unix/sysv/linux/syscall-names.list for
Linux 4.18. The io_pgetevents and rseq syscalls are added to the
kernel on various architectures, so need to be mentioned in this file.
Tested with build-many-glibcs.py.
* sysdeps/unix/sysv/linux/syscall-names.list: Update kernel
version to 4.18.
(io_pgetevents): New syscall.
(rseq): Likewise.
The install.texi documentation of uses of Perl and Python is
substantially out of date.
The description of Perl is "to test the installation" (which I
interpret as referring to test-installation.pl), but it's used for
more tests than that, and to build the manual, and to regenerate one
file in the source tree.
The description of Python is only for pretty-printer tests, but it's
used for other tests / benchmarks as well (and for other internal uses
such as updating Unicode data, for which we already require Python 3,
but I think install.texi only needs to describe uses from the main
glibc Makefiles).
This patch updates the descriptions of what those tools are used for.
The Python information (and information about other tools for testing
pretty printers) was awkwardly in the middle of the general
description of building and testing glibc, rather than with the rest
of information about tools used in glibc build and test; this patch
moves the information about those tools into the main list.
Tested with regeneration of INSTALL as well as "make info" and "make
pdf".
* manual/install.texi (Configuring and compiling): Do not list
tools used for testing pretty printers here.
(Tools for Compilation): List Python, PExpect and GDB here.
Update descriptions of uses of Perl and Python.
* INSTALL: Regenerate.
Add the workload test properties (max-throughput, latency, etc.) to
the schema to prevent benchmark output validation from failing.
* benchtests/scripts/benchout.schema.json (properties): Add
new properties.
Add the duration and iterations attributes to the workloads tests to
make the json schema parser happy
* benchtests/bench-skeleton.c (main): Add duration and
iterations attributes.
Adjust the non-glibc code to agree with what Gawk needs for
rational range interpretation (RRI) for regular expression ranges.
In unibyte locales, Gawk wants ranges to use the underlying byte
rather than the character code point. This change does not affect
glibc proper.
* posix/regcomp.c (parse_byte) [!LIBC && RE_ENABLE_I18N]:
In unibyte locales, use the byte value rather than
running it through btowc.
Continuing moving macros out of math-tests.h to smaller headers
following typo-proof conventions instead of using #ifndef, this patch
moves the SNAN_TESTS_* macros for individual types out to their own
sysdeps header (while the type-generic SNAN_TESTS wrapper for those
macros remains in math-tests.h).
Tested for x86_64 and x86, and with build-many-glibcs.py.
* sysdeps/generic/math-tests-snan.h: New file.
* sysdeps/generic/math-tests.h: Include <math-tests-snan.h>.
(SNAN_TESTS_float): Do not define here.
(SNAN_TESTS_double): Likewise.
(SNAN_TESTS_long_double): Likewise.
(SNAN_TESTS_float128): Likewise.
* sysdeps/i386/fpu/math-tests-snan.h: New file.
* sysdeps/i386/fpu/math-tests.h: Remove file.
* sysdeps/ia64/math-tests-snan.h: New file.
* sysdeps/ia64/math-tests.h: Remove file.
* sysdeps/x86/math-tests.h: Likewise.
* sysdeps/x86_64/fpu/math-tests-snan.h: New file.
This patch is a complete rewrite of sincosf. The new version is
significantly faster, as well as simple and accurate.
The worst-case ULP is 0.5607, maximum relative error is 0.5303 * 2^-23 over
all 4 billion inputs. In non-nearest rounding modes the error is 1ULP.
The algorithm uses 3 main cases: small inputs which don't need argument
reduction, small inputs which need a simple range reduction and large inputs
requiring complex range reduction. The code uses approximate integer
comparisons to quickly decide between these cases.
The small range reducer uses a single reduction step to handle values up to
120.0. It is fastest on targets which support inlined round instructions.
The large range reducer uses integer arithmetic for simplicity. It does a
32x96 bit multiply to compute a 64-bit modulo result. This is more than
accurate enough to handle the worst-case cancellation for values close to
an integer multiple of PI/4. It could be further optimized, however it is
already much faster than necessary.
sincosf throughput gains on Cortex-A72:
* |x| < 0x1p-12 : 1.6x
* |x| < M_PI_4 : 1.7x
* |x| < 2 * M_PI: 1.5x
* |x| < 120.0 : 1.8x
* |x| < Inf : 2.3x
* math/Makefile: Add s_sincosf_data.c.
* sysdeps/ia64/fpu/s_sincosf_data.c: New file.
* sysdeps/ieee754/flt-32/s_sincosf.h (abstop12): Add new function.
(sincosf_poly): Likewise.
(reduce_small): Likewise.
(reduce_large): Likewise.
* sysdeps/ieee754/flt-32/s_sincosf.c (sincosf): Rewrite.
* sysdeps/ieee754/flt-32/s_sincosf_data.c: New file with sincosf data.
* sysdeps/m68k/m680x0/fpu/s_sincosf_data.c: New file.
* sysdeps/x86_64/fpu/s_sincosf_data.c: New file.
This patch currently only affects aarch64.
The roundtoint and converttoint internal functions are only called with small
values, so 32 bit result is enough for converttoint and it is a signed int
conversion so the return type is changed to int32_t.
The original idea was to help the compiler keeping the result in uint64_t,
then it's clear that no sign extension is needed and there is no accidental
undefined or implementation defined signed int arithmetics.
But it turns out gcc does a good job with inlining so changing the type has
no overhead and the semantics of the conversion is less surprising this way.
Since we want to allow the asuint64 (x + 0x1.8p52) style conversion, the top
bits were never usable and the existing code ensures that only the bottom
32 bits of the conversion result are used.
On aarch64 the neon intrinsics (which round ties to even) are changed to
round and lround (which round ties away from zero) this does not affect the
results in a significant way, but more portable (relies on round and lround
being inlined which works with -fno-math-errno).
The TOINT_SHIFT and TOINT_RINT macros were removed, only keep separate code
paths for TOINT_INTRINSICS and !TOINT_INTRINSICS.
* sysdeps/aarch64/fpu/math_private.h (roundtoint): Use round.
(converttoint): Use lround.
* sysdeps/ieee754/flt-32/math_config.h (roundtoint): Declare and
document the semantics when TOINT_INTRINSICS is set.
(converttoint): Likewise.
(TOINT_RINT): Remove.
(TOINT_SHIFT): Remove.
* sysdeps/ieee754/flt-32/e_expf.c (__expf): Remove the TOINT_RINT code
path.
Commit 298d0e3129 ("Consolidate Linux
getdents{64} implementation") broke the implementation because it does
not take into account struct offset differences.
The new implementation is close to the old one, before the
consolidation, but has been cleaned up slightly.
* Since __fentry__ is almost the same as _mcount, reuse the code by
#including it twice with different #defines around.
* Remove LA usages - they are needed in 31-bit mode to clear the top
bit, but in 64-bit they appear to do nothing.
* Add CFI rule for the nonstandard return register. This rule applies
to the current function (binutils generates a new CIE - see
gas/dw2gencfi.c:select_cie_for_fde()), so it is not necessary to put
__fentry__ into a new file.
* Fix CFI offset for %r14.
* Add CFI rule for %r0.
* Fix unwound value of %r15 being off by 244 bytes.
* Unwinding in __fentry__@plt does not work, no plan to fix it - it
would require asking linker to generate CFI for return address in
%r0. From functional perspective keeping it broken is fine, since
the callee did not have a chance to do anything yet. From
convenience perspective it would be possible to enhance GDB in the
future to treat __fentry__@plt in a special way.
* Fix whitespace.
* Fix offsets in comments, which were copied from 32-bit code.
* 32-bit version will not be implemented, since it's not compatible
with the corresponding PLT stubs: they assume %r12 points to GOT,
which is not the case for gcc-emitted __fentry__ stub, which runs
before the prolog.
This patch adds the runtime support in glibc for the -mfentry
gcc feature introduced in [1] and [2].
[1] https://gcc.gnu.org/ml/gcc-patches/2018-07/msg00784.html
[2] https://gcc.gnu.org/ml/gcc-patches/2018-07/msg00912.html
ChangeLog:
* sysdeps/s390/s390-64/Versions (__fentry__): Add.
* sysdeps/s390/s390-64/s390x-mcount.S: Move the common
code to s390x-mcount.h and #include it.
* sysdeps/s390/s390-64/s390x-mcount.h: New file.
* sysdeps/unix/sysv/linux/s390/s390-64/libc.abilist
(__fentry__): Add.
__fentry__ symbol is currently not defined for other architectures.
Attempts to introduce it cause abicheck to fail, because it will be
available since 2.29 earliest, and not 2.13, which is the case for
Intel. With the new code, abicheck passes for i686-linux-gnu,
x86_64-linux-gnu and x86_64-linux-gnu32 triples.
ChangeLog:
* stdlib/Versions: Remove __fentry__.
* sysdeps/i386/Versions: Add __fentry__.
* sysdeps/x86_64/Versions: Add __fentry__.
The following combinations need to be tested:
* 32- (g5, esa and zarch) and 64-bit
* linux32 glibc/configure CC='gcc -m31 -march=g5'
* linux32 glibc/configure CC='gcc -m31'
* linux32 glibc/configure CC='gcc -m31 -mzarch'
* With and without VX:
* glibc/configure libc_cv_asm_s390_vx=no
* With and without profiling (using LD_PROFILE)
* With and without pltexit (using LD_AUDIT)
ChangeLog:
* sysdeps/s390/Makefile: Register the new tests.
* sysdeps/s390/tst-dl-runtime-mod.S: New file.
* sysdeps/s390/tst-dl-runtime-profile-audit.c: New file.
* sysdeps/s390/tst-dl-runtime-profile-noaudit.c: New file.
* sysdeps/s390/tst-dl-runtime-resolve-audit.c: New file.
* sysdeps/s390/tst-dl-runtime-resolve-noaudit.c: New file.
* sysdeps/s390/tst-dl-runtime.c: New file.
Following the recent discussion of using Python instead of Perl and
Awk for glibc build / test, this patch replaces gen-libm-test.pl with
a new gen-libm-test.py script. This script should work with all
Python versions supported by glibc (tested by hand with Python 2.7,
tested in the build system with Python 3.5; configure prefers Python 3
if available).
This script is designed to give identical output to gen-libm-test.pl
for ease of verification of the change, except for generated comments
referring to .py instead of .pl. (That is, identical for actual
inputs passed to the script, not necessarily for all possible input;
for example, this version more precisely follows the C standard syntax
for floating-point constants when deciding when to add LIT macro
calls.) In one place a comment notes that the generation of
NON_FINITE flags is replicating a bug in the Perl script to assist in
such comparisons (with the expectation that this bug can then be
separately fixed in the Python script later).
Tested for x86_64, including comparison of generated files (and hand
testing of the case of generating a sorted libm-test-ulps file, which
isn't covered by normal "make check").
I'd expect to follow this up by extending the new script to produce
the ulps tables for the manual as well (replacing
manual/libm-err-tab.pl, so that then we just have one ulps file
parser) - at which point the manual build would depend on both Perl
and Python (eliminating the Perl dependency would require someone to
rewrite summary.pl in Python, and that would only eliminate the
*direct* Perl dependency; current makeinfo is written in Perl so there
would still be an indirect dependency).
I think install.texi is more or less equally out-of-date regarding
Perl and Python uses before and after this patch, so I don't think
this patch depends on my patch
<https://sourceware.org/ml/libc-alpha/2018-08/msg00133.html> to update
install.texi regarding such uses (pending review).
* math/gen-libm-test.py: New file.
* math/gen-libm-test.pl: Remove.
* math/Makefile [$(PERL) != no]: Change condition to [PYTHON].
($(objpfx)libm-test-ulps.h): Use gen-libm-test.py instead of
gen-libm-test.pl.
($(libm-test-c-noauto-obj)): Likewise.
($(libm-test-c-auto-obj)): Likewise.
($(libm-test-c-narrow-obj)): Likewise.
(regen-ulps): Likewise.
* math/README.libm-test: Update references to gen-libm-test.pl.
* math/libm-test-driver.c (struct test_fj_f_data): Update comment
referencing gen-libm-test.pl.
* math/libm-test-nexttoward.inc (nexttoward_test_data): Likewise.
* math/libm-test-support.c: Likewise.
* math/libm-test-support.h: Likewise.
* sysdeps/generic/libm-test-ulps: Likewise.
MIN_PAGE_SIZE is normally set to 4096 but for testing it can be set to
16 so that it exercises the page crossing code for every misaligned
access. The value was set to 15, which is obviously wrong, so fixed
as obvious and tested.
* sysdeps/aarch64/strlen.S [TEST_PAGE_CROSS](MIN_PAGE_SIZE):
Fix value.
When libm tests were split into separate per-function .inc files, a
comment relating to the nexttoward tests ended up at the end of
libm-test-nextdown.inc (because the split was based on starting each
function's tests with the <function>_test_data definition, which
failed to allow for comments before such definitions). This patch
moves that comment to the correct location.
Tested for x86_64.
* math/libm-test-nextdown.inc (do_test): Move comment to ....
* math/libm-test-nexttoward.inc (nexttoward_test_data): ... here.
Drop realloc_bufs in favour of making alloc_bufs transparently
reallocate the buffers if it had allocated before. Also consolidate
computation of buffer lengths so that they don't get repeated on every
reallocation.
* benchtests/bench-string.h (buf1_size, buf2_size): New
variables.
(init_sizes): New function.
(test_init): Use it.
(alloc_buf, exit_error): New functions.
(alloc_bufs): Use ALLOC_BUF.
(realloc_bufs): Remove.
* benchtests/bench-memcmp.c (do_test): Adjust.
* benchtests/bench-memset-large.c (do_test): Likewise.
* benchtests/bench-memset-walk.c (do_test): Likewise.
* benchtests/bench-memset.c (do_test): Likewise.
* benchtests/bench-strncmp.c (do_test): Likewise.
Since RISC-V stores the thread pointer in a general register libthread_db
can just ask the debugger for the register contents instead of trying to
call ps_get_thread_area. This enables thread debugging in gdb.
* sysdeps/riscv/nptl/tls.h (DB_THREAD_SELF): Use REGISTER instead
of CONST_THREAD_AREA.
Move STATE_SAVE_OFFSET and STATE_SAVE_MASK to sysdep.h to make
sysdeps/x86/cpu-features.h a C header file.
* sysdeps/x86/cpu-features.h (STATE_SAVE_OFFSET): Removed.
(STATE_SAVE_MASK): Likewise.
Don't check __ASSEMBLER__ to include <cpu-features-offsets.h>.
* sysdeps/x86/sysdep.h (STATE_SAVE_OFFSET): New.
(STATE_SAVE_MASK): Likewise.
* sysdeps/x86_64/dl-trampoline.S: Include <cpu-features-offsets.h>
instead of <cpu-features.h>.
* sysdeps/riscv/rv64/rvd/libm-test-ulps: Update.
Note: I regen'd these from scratch, but I'm only committing the
increases, as I only tested on hardware. There were a few 2->1
decreases that I omitted, possibly "for now".
* sysdeps/riscv/rvf/math_private.h (libc_feholdexcept_setround_riscv):
Fix rounding save-restore bug.
Fixes about a hundred off-by-ULP failures in the math testsuite.
Some TEST_* lines in libm-test-*.inc end with semicolons not commas.
This works at present because gen-libm-test.pl ignores whatever comes
after the TEST_* call, but is logically wrong (since the TEST_* calls
generate array elements, not statements) and the Python replacement
for gen-libm-test.pl that I'm working on is stricter about the syntax
here. This patch fixes the lines in question to use commas like most
such lines already do.
Tested for x86_64.
* math/libm-test-ilogb.inc (ilogb_test_data): Use ',' not ';'
after TEST_* calls.
* math/libm-test-llogb.inc (llogb_test_data): Likewise.
* math/libm-test-logb.inc (logb_test_data): Likewise.
Looking at the benchtests, both strstr and strcasestr spend a lot of time
in a slow initialization loop handling one character per iteration.
This can be simplified and use the much faster strlen/strnlen/strchr/memcmp.
Read ahead a few cachelines to reduce the number of strnlen calls, which
improves performance by ~3-4%. This patch improves the time taken for the
full strstr benchtest by >40%.
* string/strcasestr.c (STRCASESTR): Simplify and speedup first match.
* string/strstr.c (AVAILABLE): Likewise.
There is no need to include <init-arch.h> in assembly codes since all
x86 IFUNC selector functions are written in C. Tested on i686 and
x86-64. There is no code change in libc.so, ld.so and libmvec.so.
* sysdeps/i386/i686/multiarch/bzero-ia32.S: Don't include
<init-arch.h>.
* sysdeps/x86_64/fpu/multiarch/svml_d_sin8_core-avx2.S: Likewise.
* sysdeps/x86_64/fpu/multiarch/svml_s_expf16_core-avx2.S: Likewise.
* sysdeps/x86_64/multiarch/memset-sse2-unaligned-erms.S: Likewise.
