This patch optimizes the performance of memcpy/memmove for A64FX [1]
which implements ARMv8-A SVE and has L1 64KB cache per core and L2 8MB
cache per NUMA node.
The performance optimization makes use of Scalable Vector Register
with several techniques such as loop unrolling, memory access
alignment, cache zero fill, and software pipelining.
SVE assembler code for memcpy/memmove is implemented as Vector Length
Agnostic code so theoretically it can be run on any SOC which supports
ARMv8-A SVE standard.
We confirmed that all testcases have been passed by running 'make
check' and 'make xcheck' not only on A64FX but also on ThunderX2.
And also we confirmed that the SVE 512 bit vector register performance
is roughly 4 times better than Advanced SIMD 128 bit register and 8
times better than scalar 64 bit register by running 'make bench'.
[1] https://github.com/fujitsu/A64FX
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Reviewed-by: Szabolcs Nagy <Szabolcs.Nagy@arm.com>
This patch defines BTI_C and BTI_J macros conditionally for
performance.
If HAVE_AARCH64_BTI is true, BTI_C and BTI_J are defined as HINT
instruction for ARMv8.5 BTI (Branch Target Identification).
If HAVE_AARCH64_BTI is false, both BTI_C and BTI_J are defined as
NOP.
Remove generic tlsdesc code related to lazy tlsdesc processing since
lazy tlsdesc relocation is no longer supported. This includes removing
GL(dl_load_lock) from _dl_make_tlsdesc_dynamic which is only called at
load time when that lock is already held.
Added a documentation comment too.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
DL_UNMAP_IS_SPECIAL and DL_UNMAP were not defined. The definitions are
now copied from arm, since the same is needed on aarch64. The cleanup
of tlsdesc data is handled by the custom _dl_unmap.
Fixes bug 27403.
For j0f/j1f/y0f/y1f, the largest error for all binary32
inputs is reduced to at most 9 ulps for all rounding modes.
The new code is enabled only when there is a cancellation at the very end of
the j0f/j1f/y0f/y1f computation, or for very large inputs, thus should not
give any visible slowdown on average. Two different algorithms are used:
* around the first 64 zeros of j0/j1/y0/y1, approximation polynomials of
degree 3 are used, computed using the Sollya tool (https://www.sollya.org/)
* for large inputs, an asymptotic formula from [1] is used
[1] Fast and Accurate Bessel Function Computation,
John Harrison, Proceedings of Arith 19, 2009.
Inputs yielding the new largest errors are added to auto-libm-test-in,
and ulps are regenerated for various targets (thanks Adhemerval Zanella).
Tested on x86_64 with --disable-multi-arch and on powerpc64le-linux-gnu.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
This is a target hook for memory tagging, the original was a naive
implementation. Uses the same algorithm as __libc_mtag_tag_region,
but with instructions that also zero the memory. This was not
benchmarked on real cpu, but expected to be faster than the naive
implementation.
This is a target hook for memory tagging, the original was a naive
implementation. The optimized version relies on "dc gva" to tag 64
bytes at a time for large allocations and optimizes small cases without
adding too many branches. This was not benchmarked on real cpu, but
expected to be faster than the naive implementation.
This is a common operation when heap tagging is enabled, so inline the
instruction instead of using an extern call.
The .inst directive is used instead of the name of the instruction (or
acle intrinsics) because malloc.c is not compiled for armv8.5-a+memtag
architecture, runtime cpu support detection is used.
Prototypes are removed from the comments as they were not always
correct.
The memset api is suboptimal and does not provide much benefit. Memory
tagging only needs a zeroing memset (and only for memory that's sized
and aligned to multiples of the tag granule), so change the internal
api and the target hooks accordingly. This is to simplify the
implementation of the target hook.
Reviewed-by: DJ Delorie <dj@redhat.com>
This patch series removes all remaining slow paths and related code.
First asin/acos, tan, atan, atan2 implementations are updated, and the final
patch removes the unused mpa files, headers and probes. Passes buildmanyglibc.
Remove slow paths from asin/acos. Add ULP annotations based on previous slow
path checks (which are approximate). Update AArch64 and x86_64 libm-test-ulps.
Reviewed-By: Paul Zimmermann <Paul.Zimmermann@inria.fr>
It turns out the startup code in csu/elf-init.c has a perfect pair of
ROP gadgets (see Marco-Gisbert and Ripoll-Ripoll, "return-to-csu: A
New Method to Bypass 64-bit Linux ASLR"). These functions are not
needed in dynamically-linked binaries because DT_INIT/DT_INIT_ARRAY
are already processed by the dynamic linker. However, the dynamic
linker skipped the main program for some reason. For maximum
backwards compatibility, this is not changed, and instead, the main
map is consulted from __libc_start_main if the init function argument
is a NULL pointer.
For statically linked binaries, the old approach based on linker
symbols is still used because there is nothing else available.
A new symbol version __libc_start_main@@GLIBC_2.34 is introduced because
new binaries running on an old libc would not run their ELF
constructors, leading to difficult-to-debug issues.
Some IFUNC variants are not compatible with BTI and MTE so don't
set them as usable for testing and benchmarking on a BTI or MTE
enabled system.
As far as IFUNC selectors are concerned a system is BTI enabled if
the cpu supports it and glibc was built with BTI branch protection.
Most IFUNC variants are BTI compatible, but thunderx2 memcpy and
memmove use a jump table with indirect jump, without a BTI j.
Fixes bug 26818.
The hwcap value is now in linux 5.10 and in glibc bits/hwcap.h, so use
that definition.
Move the definition to init-arch.h so all ifunc selectors can use it
and expose an "mte" shorthand for mte enabled runtime.
For now we allow user code to enable tag checks and use PROT_MTE
mappings without libc involvment, this is not guaranteed ABI, but
can be useful for testing and debugging with MTE.
In commit 863d775c48, kunpeng920 is added to default memcpy version,
however, there is performance degradation when the copy size is some large bytes, eg: 100k.
This is the result, tested in glibc-2.28:
before backport after backport Performance improvement
memcpy_1k 0.005 0.005 0.00%
memcpy_10k 0.032 0.029 10.34%
memcpy_100k 0.356 0.429 -17.02%
memcpy_1m 7.470 11.153 -33.02%
This is the demo
#include "stdio.h"
#include "string.h"
#include "stdlib.h"
char a[1024*1024] = {12};
char b[1024*1024] = {13};
int main(int argc, char *argv[])
{
int i = atoi(argv[1]);
int j;
int size = atoi(argv[2]);
for (j = 0; j < i; j++)
memcpy(b, a, size*1024);
return 0;
}
# gcc -g -O0 memcpy.c -o memcpy
# time taskset -c 10 ./memcpy 100000 1024
Co-authored-by: liqingqing <liqingqing3@huawei.com>
Add SUPPORT_STATIC_PIE that targets can define if they support
static PIE. This requires PI_STATIC_AND_HIDDEN support and various
linker features as described in
commit 9d7a3741c9
Add --enable-static-pie configure option to build static PIE [BZ #19574]
Currently defined on x86_64, i386 and aarch64 where static PIE is
known to work.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
AArch64 always uses pc relative access to static and hidden object
symbols, but the config setting was previously missing.
This affects ld.so start up code.
Remove the wordsize-64 implementations by merging them into the main dbl-64
directory. The second patch just moves all wordsize-64 files and removes a
few wordsize-64 uses in comments and Implies files.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
After sp is updated, the CFA offset should be set before next instruction.
Tested in glibc-2.28:
Thread 2 "xxxxxxx" hit Breakpoint 1, _dl_tlsdesc_dynamic () at ../sysdeps/aarch64/dl-tlsdesc.S:149
149 stp x1, x2, [sp, #-32]!
Missing separate debuginfos, use: dnf debuginfo-install libgcc-7.3.0-20190804.h24.aarch64
(gdb) bt
#0 _dl_tlsdesc_dynamic () at ../sysdeps/aarch64/dl-tlsdesc.S:149
#1 0x0000ffffbe4fbb44 in OurFunction (threadId=3194870184)
at /home/test/test_function.c:30
#2 0x0000000000400c08 in initaaa () at thread.c:58
#3 0x0000000000400c50 in thread_proc (param=0x0) at thread.c:71
#4 0x0000ffffbf6918bc in start_thread (arg=0xfffffffff29f) at pthread_create.c:486
#5 0x0000ffffbf5669ec in thread_start () at ../sysdeps/unix/sysv/linux/aarch64/clone.S:78
(gdb) ni
_dl_tlsdesc_dynamic () at ../sysdeps/aarch64/dl-tlsdesc.S:150
150 stp x3, x4, [sp, #16]
(gdb) bt
#0 _dl_tlsdesc_dynamic () at ../sysdeps/aarch64/dl-tlsdesc.S:150
#1 0x0000ffffbe4fbb44 in OurFunction (threadId=3194870184)
at /home/test/test_function.c:30
#2 0x0000000000000000 in ?? ()
Backtrace stopped: previous frame identical to this frame (corrupt stack?)
(gdb) ni
_dl_tlsdesc_dynamic () at ../sysdeps/aarch64/dl-tlsdesc.S:157
157 mrs x4, tpidr_el0
(gdb) bt
#0 _dl_tlsdesc_dynamic () at ../sysdeps/aarch64/dl-tlsdesc.S:157
#1 0x0000ffffbe4fbb44 in OurFunction (threadId=3194870184)
at /home/test/test_function.c:30
#2 0x0000000000400c08 in initaaa () at thread.c:58
#3 0x0000000000400c50 in thread_proc (param=0x0) at thread.c:71
#4 0x0000ffffbf6918bc in start_thread (arg=0xfffffffff29f) at pthread_create.c:486
#5 0x0000ffffbf5669ec in thread_start () at ../sysdeps/unix/sysv/linux/aarch64/clone.S:78
Signed-off-by: liqingqing <liqingqing3@huawei.com>
Signed-off-by: Shuo Wang <wangshuo47@huawei.com>
I used these shell commands:
../glibc/scripts/update-copyrights $PWD/../gnulib/build-aux/update-copyright
(cd ../glibc && git commit -am"[this commit message]")
and then ignored the output, which consisted lines saying "FOO: warning:
copyright statement not found" for each of 6694 files FOO.
I then removed trailing white space from benchtests/bench-pthread-locks.c
and iconvdata/tst-iconv-big5-hkscs-to-2ucs4.c, to work around this
diagnostic from Savannah:
remote: *** pre-commit check failed ...
remote: *** error: lines with trailing whitespace found
remote: error: hook declined to update refs/heads/master
DELOUSE was added to asm code to make them compatible with non-LP64
ABIs, but it is an unfortunate name and the code was not compatible
with ABIs where pointer and size_t are different. Glibc currently
only supports the LP64 ABI so these macros are not really needed or
tested, but for now the name is changed to be more meaningful instead
of removing them completely.
Some DELOUSE macros were dropped: clone, strlen and strnlen used it
unnecessarily.
The out of tree ILP32 patches are currently not maintained and will
likely need a rework to rebase them on top of the time64 changes.
This symbol is not in the implementation reserved namespace for static
linking and it was never used: it seems it was mistakenly added in the
orignal strlen_asimd commit 436e4d5b96
Re-mmap executable segments if possible instead of using mprotect
to add PROT_BTI. This allows using BTI protection with security
policies that prevent mprotect with PROT_EXEC.
If the fd of the ELF module is not available because it was kernel
mapped then mprotect is used and failures are ignored. To protect
the main executable even when mprotect is filtered the linux kernel
will have to be changed to add PROT_BTI to it.
The delayed failure reporting is mainly needed because currently
_dl_process_gnu_properties does not propagate failures such that
the required cleanups happen. Using the link_map_machine struct for
error propagation is not ideal, but this seemed to be the least
intrusive solution.
Fixes bug 26831.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
To handle GNU property notes on aarch64 some segments need to
be mmaped again, so the fd of the loaded ELF module is needed.
When the fd is not available (kernel loaded modules), then -1
is passed.
The fd is passed to both _dl_process_pt_gnu_property and
_dl_process_pt_note for consistency. Target specific note
processing functions are updated accordingly.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Handle unaligned executable load segments (the bfd linker is not
expected to produce such binaries, but other linkers may).
Computing the mapping bounds follows _dl_map_object_from_fd more
closely now.
Fixes bug 26988.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
The _dl_open_check and _rtld_main_check hooks are not called on the
dependencies of a loaded module, so BTI protection was missed on
every module other than the main executable and directly dlopened
libraries.
The fix just iterates over dependencies to enable BTI.
Fixes bug 26926.
Now __thread_gscope_wait (the function behind THREAD_GSCOPE_WAIT,
formerly __wait_lookup_done) can be implemented directly in ld.so,
eliminating the unprotected GL (dl_wait_lookup_done) function
pointer.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
This test fails without bug 26798 fixed because some integer registers
likely get clobbered by lazy binding and variant PCS only allows x16
and x17 to be clobbered at call time.
The test requires binutils 2.32.1 or newer for handling variant PCS
symbols. SVE registers are not covered by this test, to avoid the
complexity of handling multiple compile- and runtime feature support
cases.
The variant PCS support was ineffective because in the common case
linkmap->l_mach.plt == 0 but then the symbol table flags were ignored
and normal lazy binding was used instead of resolving the relocs early.
(This was a misunderstanding about how GOT[1] is setup by the linker.)
In practice this mainly affects SVE calls when the vector length is
more than 128 bits, then the top bits of the argument registers get
clobbered during lazy binding.
Fixes bug 26798.
Add CPU detection of Neoverse N2 and Neoverse V1, and select __memcpy_simd as
the memcpy/memmove ifunc.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
There are several compiler implementations that allow large stack
allocations to jump over the guard page at the end of the stack and
corrupt memory beyond that. See CVE-2017-1000364.
Compilers can emit code to probe the stack such that the guard page
cannot be skipped, but on aarch64 the probe interval is 64K by default
instead of the minimum supported page size (4K).
This patch enforces at least 64K guard on aarch64 unless the guard
is disabled by setting its size to 0. For backward compatibility
reasons the increased guard is not reported, so it is only observable
by exhausting the address space or parsing /proc/self/maps on linux.
On other targets the patch has no effect. If the stack probe interval
is larger than a page size on a target then ARCH_MIN_GUARD_SIZE can
be defined to get large enough stack guard on libc allocated stacks.
The patch does not affect threads with user allocated stacks.
Fixes bug 26691.
On some microarchitectures performance of the backwards memmove improves if
the stores use STR with decreasing addresses. So change the memmove loop
in memcpy_advsimd.S to use 2x STR rather than STP.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Make glibc MTE-safe on systems where MTE is available. This allows
using heap tagging with an LD_PRELOADed malloc implementation that
enables MTE. We don't document this as guaranteed contract yet, so
glibc may not be MTE safe when HWCAP2_MTE is set (older glibcs
certainly aren't). This is mainly for testing and debugging.
The HWCAP flag is not exposed in public headers until Linux adds it
to its uapi. The HWCAP value reservation will be in Linux 5.9.
Use PROT_READ and PROT_WRITE according to the load segment p_flags
when adding PROT_BTI.
This is before processing relocations which may drop PROT_BTI in
case of textrels. Executable stacks are not protected via PROT_BTI
either. PROT_BTI is hardening in case memory corruption happened,
it's value is reduced if there is writable and executable memory
available so missing it on such memory is fine, but we should
respect the p_flags and should not drop PROT_WRITE.
Optimize strlen using a mix of scalar and SIMD code. On modern micro
architectures large strings are 2.6 times faster than existing
strlen_asimd and 35% faster than the new MTE version of strlen.
On a random strlen benchmark using small sizes the speedup is 7% vs
strlen_asimd and 40% vs the MTE strlen. This fixes the main strlen
regressions on Cortex-A53 and other cores with a simple Neon unit.
Rename __strlen_generic to __strlen_mte, and select strlen_asimd when
MTE is not enabled (this is waiting on support for a HWCAP_MTE bit).
This fixes big-endian bug 25824. Passes GLIBC regression tests.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Add a new memcpy using 128-bit Q registers - this is faster on modern
cores and reduces codesize. Similar to the generic memcpy, small cases
include copies up to 32 bytes. 64-128 byte copies are split into two
cases to improve performance of 64-96 byte copies. Large copies align
the source rather than the destination.
bench-memcpy-random is ~9% faster than memcpy_falkor on Neoverse N1,
so make this memcpy the default on N1 (on Centriq it is 15% faster than
memcpy_falkor).
Passes GLIBC regression tests.
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Given almost all uses of ENTRY are for string/memory functions,
align ENTRY to a cacheline to simplify things.
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
RETURN_ADDRESS is used at several places in glibc to mean a valid
code address of the call site, but with pac-ret it may contain a
pointer authentication code (PAC), so its definition is adjusted.
This is gcc PR target/94891: __builtin_return_address should not
expose signed pointers to user code where it can cause ABI issues.
In glibc RETURN_ADDRESS is only changed if it is built with pac-ret.
There is no detection for the specific gcc issue because it is
hard to test and the additional xpac does not cause problems.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Currently gcc -pg -mbranch-protection=pac-ret passes signed return
address to _mcount, so _mcount now has to always strip pac from the
frompc since that's from user code that may be built with pac-ret.
This is gcc PR target/94791: signed pointers should not escape and get
passed across extern call boundaries, since that's an ABI break, but
because existing gcc has this issue we work it around in glibc until
that is resolved. This is compatible with a fixed gcc and it is a nop
on systems without PAuth support. The bug was introduced in gcc-7 with
-msign-return-address=non-leaf|all support which in gcc-9 got renamed
to -mbranch-protection=pac-ret|pac-ret+leaf|standard.
strip_pac uses inline asm instead of __builtin_aarch64_xpaclri since
that is not a documented api and not available in all supported gccs.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Use return address signing in assembly files for functions that save
LR when pac-ret is enabled in the compiler.
The GNU property note for PAC-RET is not meaningful to the dynamic
linker so it is not strictly required, but it may be used to track
the security property of binaries. (The PAC-RET property is only set
if BTI is set too because BTI implies working GNU property support.)
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Return address signing requires unwinder support, which is
present in libgcc since >=gcc-7, however due to bugs the
support may be broken in <gcc-10 (and similarly there may
be issues in custom unwinders), so pac-ret is not always
safe to use. So in assembly code glibc should only use
pac-ret if the compiler uses it too. Unfortunately there
is no predefined feature macro for it set by the compiler
so pac-ret is inferred from the code generation.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
When glibc is built with branch protection (i.e. with a gcc configured
with --enable-standard-branch-protection), all glibc binaries should
be BTI compatible and marked as such.
It is easy to link BTI incompatible objects by accident and this is
silent currently which is usually not the expectation, so this is
changed into a link error. (There is no linker flag for failing on
BTI incompatible inputs so all warnings are turned into fatal errors
outside the test system when building glibc with branch protection.)
Unfortunately, outlined atomic functions are not BTI compatible in
libgcc (PR libgcc/96001), so to build glibc with current gcc use
'CC=gcc -mno-outline-atomics', this should be fixed in libgcc soon
and then glibc can be built and tested without such workarounds.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Binaries can opt-in to using BTI via an ELF object file marking.
The dynamic linker has to then mprotect the executable segments
with PROT_BTI. In case of static linked executables or in case
of the dynamic linker itself, PROT_BTI protection is done by the
operating system.
On AArch64 glibc uses PT_GNU_PROPERTY instead of PT_NOTE to check
the properties of a binary because PT_NOTE can be unreliable with
old linkers (old linkers just append the notes of input objects
together and add them to the output without checking them for
consistency which means multiple incompatible GNU property notes
can be present in PT_NOTE).
BTI property is handled in the loader even if glibc is not built
with BTI support, so in theory user code can be BTI protected
independently of glibc. In practice though user binaries are not
marked with the BTI property if glibc has no support because the
static linked libc objects (crt files, libc_nonshared.a) are
unmarked.
This patch relies on Linux userspace API that is not yet in a
linux release but in v5.8-rc1 so scheduled to be in Linux 5.8.
Co-authored-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Tailcalls must use x16 or x17 for the indirect branch instruction
to be compatible with code that uses BTI c at function entries.
(Other forms of indirect branches can only land on BTI j.)
Also added a BTI c at the ELF entry point of rtld, this is not
strictly necessary since the kernel does not use indirect branch
to get there, but it seems safest once building glibc itself with
BTI is supported.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
To enable building glibc with branch protection, assembly code
needs BTI landing pads and ELF object file markings in the form
of a GNU property note.
The landing pads are unconditionally added to all functions that
may be indirectly called. When the code segment is not mapped
with PROT_BTI these instructions are nops. They are kept in the
code when BTI is not supported so that the layout of performance
critical code is unchanged across configurations.
The GNU property notes are only added when there is support for
BTI in the toolchain, because old binutils does not handle the
notes right. (Does not know how to merge them nor to put them in
PT_GNU_PROPERTY segment instead of PT_NOTE, and some versions
of binutils emit warnings about the unknown GNU property. In
such cases the produced libc binaries would not have valid
ELF marking so BTI would not be enabled.)
Note: functions using ENTRY or ENTRY_ALIGN now start with an
additional BTI c, so alignment of the following code changes,
but ENTRY_ALIGN_AND_PAD was fixed so there is no change to the
existing code layout. Some string functions may need to be
tuned for optimal performance after this commit.
Co-authored-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
The compiler can add required elf markings based on CFLAGS
but the assembler cannot, so using C code for empty files
creates less of a maintenance problem.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Check BTI support in the compiler and linker. The check also
requires READELF that understands the BTI GNU property note.
It is expected to succeed with gcc >=gcc-9 configured with
--enable-standard-branch-protection and binutils >=binutils-2.33.
Note: passing -mbranch-protection=bti in CFLAGS when building glibc
may not be enough to get a glibc that supports BTI because crtbegin*
and crtend* provided by the compiler needs to be BTI compatible too.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Add support for MTE to strncmp. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.
The existing implementation assumes that any access to the pages in which the
string resides is safe. This assumption is not true when MTE is enabled. This
patch updates the algorithm to ensure that accesses remain within the bounds
of an MTE tag (16-byte chunks) and improves overall performance.
Co-authored-by: Branislav Rankov <branislav.rankov@arm.com>
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Add support for MTE to strcmp. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.
The existing implementation assumes that any access to the pages in which the
string resides is safe. This assumption is not true when MTE is enabled. This
patch updates the algorithm to ensure that accesses remain within the bounds
of an MTE tag (16-byte chunks) and improves overall performance.
Co-authored-by: Branislav Rankov <branislav.rankov@arm.com>
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Add support for MTE to strrchr. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.
The existing implementation assumes that any access to the pages in which the
string resides is safe. This assumption is not true when MTE is enabled. This
patch updates the algorithm to ensure that accesses remain within the bounds
of an MTE tag (16-byte chunks) and improves overall performance.
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Add support for MTE to memrchr. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.
The existing implementation assumes that any access to the pages in which the
string resides is safe. This assumption is not true when MTE is enabled. This
patch updates the algorithm to ensure that accesses remain within the bounds
of an MTE tag (16-byte chunks) and improves overall performance.
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Add support for MTE to memchr. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.
The existing implementation assumes that any access to the pages in which the
string resides is safe. This assumption is not true when MTE is enabled. This
patch updates the algorithm to ensure that accesses remain within the bounds
of an MTE tag (16-byte chunks) and improves overall performance.
Co-authored-by: Gabor Kertesz <gabor.kertesz@arm.com>
Add support for MTE to strcpy. Regression tested with xcheck and benchmarked
with glibc's benchtests on the Cortex-A53, Cortex-A72, and Neoverse N1.
The existing implementation assumes that any access to the pages in which the
string resides is safe. This assumption is not true when MTE is enabled. This
patch updates the algorithm to ensure that accesses remain within the bounds
of an MTE tag (16-byte chunks) and improves overall performance.
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
The -fno-math-errno is already added by default and the minimum
required GCC to build glibc (6.2) make the -ffinite-math-only
superflous.
Checked on aarch64-linux-gnu.
The define is already set on the math-use-builtins-ceil.h, the patch
just removes the implementations (it was missed on c9feb1be93).
Checked on aarch64-linux-gnu.
Each symbol definitions are moved on a separated file and it
cover all symbol type definitions (float, double, long double,
and float128).
It allows to set support for architectures without the boiler
place of copying default values.
Checked with a build on the affected ABIs.
Introduce an Arm MTE compatible strlen implementation.
The existing implementation assumes that any access to the pages in
which the string resides is safe. This assumption is not true when
MTE is enabled. This patch updates the algorithm to ensure that
accesses remain within the bounds of an MTE tag (16-byte chunks) and
improves overall performance on modern cores. On cores with less
efficient Advanced SIMD implementation such as Cortex-A53 it can
be slower.
Benchmarked on Cortex-A72, Cortex-A53, Neoverse N1.
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Introduce an Arm MTE compatible strchr implementation.
The existing implementation assumes that any access to the pages in
which the string resides is safe. This assumption is not true when
MTE is enabled. This patch updates the algorithm to ensure that
accesses remain within the bounds of an MTE tag (16-byte chunks) and
improves overall performance.
Benchmarked on Cortex-A72, Cortex-A53, Neoverse N1.
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Introduce an Arm MTE compatible strchrnul implementation.
The existing implementation assumes that any access to the pages in
which the string resides is safe. This assumption is not true when
MTE is enabled. This patch updates the algorithm to ensure that
accesses remain within the bounds of an MTE tag (16-byte chunks) and
improves overall performance.
Benchmarked on Cortex-A72, Cortex-A53, Neoverse N1.
Co-authored-by: Wilco Dijkstra <wilco.dijkstra@arm.com>
Falkor's memcpy and memmove share some implementation details,
therefore, the two routines are moved to a single source file
for code reuse.
The two routines now share code for small and medium copies
(up to and including 128 bytes). Large copies in memcpy do not
handle overlap correctly, consequently, the loops for
moving/copying more than 128 bytes stay separate for memcpy
and memmove.
To increase code reuse a number of small modifications were made:
1. The old implementation of memcpy copied the first 16-bytes as
soon as the size of data was determined to be greater than 32 bytes.
For memcpy code to also work when copying small/medium overlapping
data, the first load and store was moved to the large copy case.
2. Medium memcpy case no longer assumes that 16 bytes were already
copied and uses 8 registers to copy up to 128 bytes.
3. Small case for memmove was enlarged to that of memcpy, which is
less than or equal to 32 bytes.
4. Medium case for memmove was enlarged to that of memcpy, which is
less than or equal to 128 bytes.
Other changes include:
1. Improve alignment of existing loop bodies.
2. 'Delouse' memmove and memcpy input arguments. Make sure that
upper 32-bits of input registers are zeroed if unused.
3. Do one more iteration in memmove loops and reduce the number of
copies made from the start/end of the buffer, depending on
the direction of the memmove loop.
Benchmarking:
Looking at the results from bench-memcpy-random.out, we can see that
now memmove_falkor is about 5% faster than memcpy_falkor_old, while
memmove_falkor_old was more than 15% slower. The memcpy implementation
remained largely unmodified, so there is no significant performance
change.
The reason for such a significant memmove performance gain is the
increase of the upper bound on the small copy case to 32 bytes and
the increase of the upper bound on the medium copy case to 128 bytes.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
This patch fixes the optimized implementation of strcpy and strnlen
on a big-endian arm64 machine.
The optimized method uses neon, which can process 128bit with one
instruction. On a big-endian machine, the bit order should be reversed
for the whole 128-bits double word. But with instuction
rev64 datav.16b, datav.16b
it reverses 64bits in the two halves rather than reversing 128bits.
There is no such instruction as rev128 to reverse the 128bits, but we
can fix this by loading the data registers accordingly.
Fixes 0237b61526e7("aarch64: Optimized implementation of strcpy") and
2911cb68ed3d("aarch64: Optimized implementation of strnlen").
Signed-off-by: Lexi Shao <shaolexi@huawei.com>
Reviewed-by: Szabolcs Nagy <szabolcs.nagy@arm.com>
With mathinline removal there is no need to keep building and testing
inline math tests.
The gen-libm-tests.py support to generate ULP_I_* is removed and all
libm-test-ulps files are updated to longer have the
i{float,double,ldouble} entries. The support for no-test-inline is
also removed from both gen-auto-libm-tests and the
auto-libm-test-out-* were regenerated.
Checked on x86_64-linux-gnu and i686-linux-gnu.
Further optimize integer memcpy. Small cases now include copies up
to 32 bytes. 64-128 byte copies are split into two cases to improve
performance of 64-96 byte copies. Comments have been rewritten.
This supersedes the init_array sysdeps directory. It allows us to
check for ELF_INITFINI in both C and assembler code, and skip DT_INIT
and DT_FINI processing completely on newer architectures.
A new header file is needed because <dl-machine.h> is incompatible
with assembler code. <sysdep.h> is compatible with assembler code,
but it cannot be included in all assembler files because on some
architectures, it redefines register names, and some assembler files
conflict with that.
<elf-initfini.h> is replicated for legacy architectures which need
DT_INIT/DT_FINI support. New architectures follow the generic default
and disable it.
This patch adds a new macro, libm_alias_finite, to define all _finite
symbol. It sets all _finite symbol as compat symbol based on its first
version (obtained from the definition at built generated first-versions.h).
The <fn>f128_finite symbols were introduced in GLIBC 2.26 and so need
special treatment in code that is shared between long double and float128.
It is done by adding a list, similar to internal symbol redifinition,
on sysdeps/ieee754/float128/float128_private.h.
Alpha also needs some tricky changes to ensure we still emit 2 compat
symbols for sqrt(f).
Passes buildmanyglibc.
Co-authored-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Due to the branch prediction issue of Kunpeng processor, we found
memset_generic has poor performance on middle sizes setting, and so
we reconstructed the logic, expanded the loop by 4 times in set_long
to solve the problem, even when setting below 1K sizes have benefit.
Another change is that DZ_ZVA seems no work when setting zero, so we
discarded it and used set_long to set zero instead. Fewer branches and
predictions also make the zero case have slightly improvement.
Checked on aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Optimize the strlen implementation by using vector operations and
loop unrolling in main loop.Compared to __strlen_generic,it reduces
latency of cases in bench-strlen by 7%~18% when the length of src
is greater than 128 bytes, with gains throughout the benchmark.
Checked on aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Considering the excellent performance of memchr.S on glibc 2.30, the
same algorithm is used to find chrin. Compared to memrchr.c, this
method with memrchr.S achieves an average performance improvement
of 58% based on benchtest and its extension cases.
Checked on aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Optimize the strlen implementation by using vector operations and
loop unrooling in main loop. Compared to aarch64/strnlen.S, it
reduces latency of cases in bench-strnlen by 11%~24% when the length
of src is greater than 64 bytes, with gains throughout the benchmark.
Checked on aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
Optimize the strcpy implementation by using vector loads and operations
in main loop.Compared to aarch64/strcpy.S, it reduces latency of cases
in bench-strlen by 5%~18% when the length of src is greater than 64
bytes, with gains throughout the benchmark.
Checked on aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
The loop body is expanded from a 16-byte comparison to a 64-byte
comparison, and the usage of ldp is replaced by the Post-index
mode to the Base plus offset mode. Hence, compare can faster 18%
around > 128 bytes in all.
Checked on aarch64-linux-gnu.
Reviewed-by: Wilco Dijkstra <Wilco.Dijkstra@arm.com>
This commit adds missing skip_ifunc checks to aarch64, arm, i386,
sparc, and x86_64. A new test case ensures that IRELATIVE IFUNC
resolvers do not run in various diagnostic modes of the dynamic
loader.
Reviewed-By: Szabolcs Nagy <szabolcs.nagy@arm.com>
This patch adds a default pthread-offsets.h based on default
thread definitions from struct_mutex.h and struct_rwlock.h.
The idea is to simplify new ports inclusion.
Checked with a build on affected abis.
Change-Id: I7785a9581e651feb80d1413b9e03b5ac0452668a
This patch adds a new generic __pthread_rwlock_arch_t definition meant
to be used by new ports. Its layout mimics the current usage on some
64 bits ports and it allows some ports to use the generic definition.
The arch __pthread_rwlock_arch_t definition is moved from
pthreadtypes-arch.h to another arch-specific header (struct_rwlock.h).
Also the static intialization macro for pthread_rwlock_t is set to use
an arch defined on (__PTHREAD_RWLOCK_INITIALIZER) which simplifies its
implementation.
The default pthread_rwlock_t layout differs from current ports with:
1. Internal layout is the same for 32 bits and 64 bits.
2. Internal flag is an unsigned short so it should not required
additional padding to align for word boundary (if it is the case
for the ABI).
Checked with a build on affected abis.
Change-Id: I776a6a986c23199929d28a3dcd30272db21cd1d0
The current way of defining the common mutex definition for POSIX and
C11 on pthreadtypes-arch.h (added by commit 06be6368da) is
not really the best options for newer ports. It requires define some
misleading flags that should be always defined as 0
(__PTHREAD_COMPAT_PADDING_MID and __PTHREAD_COMPAT_PADDING_END), it
exposes options used solely for linuxthreads compat mode
(__PTHREAD_MUTEX_USE_UNION and __PTHREAD_MUTEX_NUSERS_AFTER_KIND), and
requires newer ports to explicit define them (adding more boilerplate
code).
This patch adds a new default __pthread_mutex_s definition meant to
be used by newer ports. Its layout mimics the current usage on both
32 and 64 bits ports and it allows most ports to use the generic
definition. Only ports that use some arch-specific definition (such
as hardware lock-elision or linuxthreads compat) requires specific
headers.
For 32 bit, the generic definitions mimic the other 32-bit ports
of using an union to define the fields uses on adaptive and robust
mutexes (thus not allowing both usage at same time) and by using a
single linked-list for robust mutexes. Both decisions seemed to
follow what recent ports have done and make the resulting
pthread_mutex_t/mtx_t object smaller.
Also the static intialization macro for pthread_mutex_t is set to use
a macro __PTHREAD_MUTEX_INITIALIZER where the architecture can redefine
in its struct_mutex.h if it requires additional fields to be
initialized.
Checked with a build on affected abis.
Change-Id: I30a22c3e3497805fd6e52994c5925897cffcfe13
The new rwlock implementation added by cc25c8b4c1 (2.25) removed
support for lock-elision. This patch removes remaining the
arch-specific unused definitions.
Checked with a build against all affected ABIs.
Change-Id: I5dec8af50e3cd56d7351c52ceff4aa3771b53cd6
This patch new build tests to check for internal fields offsets for
internal pthread_rwlock_t definition. Althoug the '__data.__flags'
field layout should be preserved due static initializators, the patch
also adds tests for the futexes that may be used in a shared memory
(although using different libc version in such scenario is not really
supported).
Checked with a build against all affected ABIs.
Change-Id: Iccc103d557de13d17e4a3f59a0cad2f4a640c148
The offsets of pthread_mutex_t __data.__nusers, __data.__spins,
__data.elision, __data.list are not required to be constant over
the releases. Only the __data.__kind is used for static
initializers.
This patch also adds an additional size check for __data.__kind.
Checked with a build against affected ABIs.
Change-Id: I7a4e48cc91b4c4ada57e9a5d1b151fb702bfaa9f
Increase the upper bound on medium cases from 96 to 128 bytes.
Now, up to 128 bytes are copied unrolled.
Increase the upper bound on small cases from 16 to 32 bytes so that
copies of 17-32 bytes are not impacted by the larger medium case.
Benchmarking:
The attached figures show relative timing difference with respect
to 'memcpy_generic', which is the existing implementation.
'memcpy_med_128' denotes the the version of memcpy_generic with
only the medium case enlarged. The 'memcpy_med_128_small_32' numbers
are for the version of memcpy_generic submitted in this patch, which
has both medium and small cases enlarged. The figures were generated
using the script from:
https://www.sourceware.org/ml/libc-alpha/2019-10/msg00563.html
Depending on the platform, the performance improvement in the
bench-memcpy-random.c benchmark ranges from 6% to 20% between
the original and final version of memcpy.S
Tested against GLIBC testsuite and randomized tests.
With only two exceptions (sys/types.h and sys/param.h, both of which
historically might have defined BYTE_ORDER) the public headers that
include <endian.h> only want to be able to test __BYTE_ORDER against
__*_ENDIAN.
This patch creates a new bits/endian.h that can be included by any
header that wants to be able to test __BYTE_ORDER and/or
__FLOAT_WORD_ORDER against the __*_ENDIAN constants, or needs
__LONG_LONG_PAIR. It only defines macros in the implementation
namespace.
The existing bits/endian.h (which could not be included independently
of endian.h, and only defines __BYTE_ORDER and maybe __FLOAT_WORD_ORDER)
is renamed to bits/endianness.h. I also took the opportunity to
canonicalize the form of this header, which we are stuck with having
one copy of per architecture. Since they are so short, this means git
doesn’t understand that they were renamed from existing headers, sigh.
endian.h itself is a nonstandard header and its only remaining use
from a standard header is guarded by __USE_MISC, so I dropped the
__USE_MISC conditionals from around all of the public-namespace things
it defines. (This means, an application that requests strict library
conformance but includes endian.h will still see the definition of
BYTE_ORDER.)
A few changes to specific bits/endian(ness).h variants deserve
mention:
- sysdeps/unix/sysv/linux/ia64/bits/endian.h is moved to
sysdeps/ia64/bits/endianness.h. If I remember correctly, ia64 did
have selectable endianness, but we have assembly code in
sysdeps/ia64 that assumes it’s little-endian, so there is no reason
to treat the ia64 endianness.h as linux-specific.
- The C-SKY port does not fully support big-endian mode, the compile
will error out if __CSKYBE__ is defined.
- The PowerPC port had extra logic in its bits/endian.h to detect a
broken compiler, which strikes me as unnecessary, so I removed it.
- The only files that defined __FLOAT_WORD_ORDER always defined it to
the same value as __BYTE_ORDER, so I removed those definitions.
The SH bits/endian(ness).h had comments inconsistent with the
actual setting of __FLOAT_WORD_ORDER, which I also removed.
- I *removed* copyright boilerplate from the few bits/endian(ness).h
headers that had it; these files record a single fact in a fashion
dictated by an external spec, so I do not think they are copyrightable.
As long as I was changing every copy of ieee754.h in the tree, I
noticed that only the MIPS variant includes float.h, because it uses
LDBL_MANT_DIG to decide among three different versions of
ieee854_long_double. This patch makes it not include float.h when
GCC’s intrinsic __LDBL_MANT_DIG__ is available.
* string/endian.h: Unconditionally define LITTLE_ENDIAN,
BIG_ENDIAN, PDP_ENDIAN, and BYTE_ORDER. Condition byteswapping
macros only on !__ASSEMBLER__. Move the definitions of
__BIG_ENDIAN, __LITTLE_ENDIAN, __PDP_ENDIAN, __FLOAT_WORD_ORDER,
and __LONG_LONG_PAIR to...
* string/bits/endian.h: ...this new file, which includes
the renamed header bits/endianness.h for the definition of
__BYTE_ORDER and possibly __FLOAT_WORD_ORDER.
* string/Makefile: Install bits/endianness.h.
* include/bits/endian.h: New wrapper.
* bits/endian.h: Rename to bits/endianness.h.
Add multiple-include guard. Rewrite the comment explaining what
the machine-specific variants of this file should do.
* sysdeps/unix/sysv/linux/ia64/bits/endian.h:
Move to sysdeps/ia64.
* sysdeps/aarch64/bits/endian.h
* sysdeps/alpha/bits/endian.h
* sysdeps/arm/bits/endian.h
* sysdeps/csky/bits/endian.h
* sysdeps/hppa/bits/endian.h
* sysdeps/ia64/bits/endian.h
* sysdeps/m68k/bits/endian.h
* sysdeps/microblaze/bits/endian.h
* sysdeps/mips/bits/endian.h
* sysdeps/nios2/bits/endian.h
* sysdeps/powerpc/bits/endian.h
* sysdeps/riscv/bits/endian.h
* sysdeps/s390/bits/endian.h
* sysdeps/sh/bits/endian.h
* sysdeps/sparc/bits/endian.h
* sysdeps/x86/bits/endian.h:
Rename to endianness.h; canonicalize form of file; remove
redundant definitions of __FLOAT_WORD_ORDER.
* sysdeps/powerpc/bits/endianness.h: Remove logic to check for
broken compilers.
* ctype/ctype.h
* sysdeps/aarch64/nptl/bits/pthreadtypes-arch.h
* sysdeps/arm/nptl/bits/pthreadtypes-arch.h
* sysdeps/csky/nptl/bits/pthreadtypes-arch.h
* sysdeps/ia64/ieee754.h
* sysdeps/ieee754/ieee754.h
* sysdeps/ieee754/ldbl-128/ieee754.h
* sysdeps/ieee754/ldbl-128ibm/ieee754.h
* sysdeps/m68k/nptl/bits/pthreadtypes-arch.h
* sysdeps/microblaze/nptl/bits/pthreadtypes-arch.h
* sysdeps/mips/ieee754/ieee754.h
* sysdeps/mips/nptl/bits/pthreadtypes-arch.h
* sysdeps/nios2/nptl/bits/pthreadtypes-arch.h
* sysdeps/nptl/pthread.h
* sysdeps/riscv/nptl/bits/pthreadtypes-arch.h
* sysdeps/sh/nptl/bits/pthreadtypes-arch.h
* sysdeps/sparc/sparc32/ieee754.h
* sysdeps/unix/sysv/linux/generic/bits/stat.h
* sysdeps/unix/sysv/linux/generic/bits/statfs.h
* sysdeps/unix/sysv/linux/sys/acct.h
* wctype/bits/wctype-wchar.h:
Include bits/endian.h, not endian.h.
* sysdeps/unix/sysv/linux/hppa/pthread.h: Don’t include endian.h.
* sysdeps/mips/ieee754/ieee754.h: Use __LDBL_MANT_DIG__
in ifdefs, instead of LDBL_MANT_DIG. Only include float.h
when __LDBL_MANT_DIG__ is not predefined, in which case
define __LDBL_MANT_DIG__ to equal LDBL_MANT_DIG.
* sysdeps/aarch64/multiarch/memset_base64.S (DC_ZVA_THRESHOLD):
Disable DC ZVA code if this macro is defined as zero.
* sysdeps/aarch64/multiarch/memset_emag.S (DC_ZVA_THRESHOLD):
Change to zero to disable using DC ZVA.
