This performs very well. x86-64-v3 and x86-64-v4 kernels were written,
too, but performed worse than the baseline kernel on short strings.
These may be added at a future point in time if the performance issues
can be fixed.
os: FreeBSD
arch: amd64
cpu: 11th Gen Intel(R) Core(TM) i7-1165G7 @ 2.80GHz
│ strlen_scalar.out │ strlen_baseline.out │
│ B/s │ B/s vs base │
Short 1.667Gi ± 1% 2.676Gi ± 1% +60.55% (p=0.000 n=20)
Mid 5.459Gi ± 1% 8.756Gi ± 1% +60.39% (p=0.000 n=20)
Long 15.34Gi ± 0% 52.27Gi ± 0% +240.64% (p=0.000 n=20)
geomean 5.188Gi 10.70Gi +106.24%
Sponsored by: The FreeBSD Foundation
Approved by: kib
Reviewed by: mjg jrtc27
Differential Revision: https://reviews.freebsd.org/D40693
Add a framework for selecting from one of multiple implementations
of a function based on amd64 architecture level (cf. amd64 SysV
ABI supplement).
Sponsored by: The FreeBSD Foundation
Approved by: kib
Reviewed by: jrtc27
Differential Revision: https://reviews.freebsd.org/D40693
The SPDX folks have obsoleted the BSD-2-Clause-FreeBSD identifier. Catch
up to that fact and revert to their recommended match of BSD-2-Clause.
Discussed with: pfg
MFC After: 3 days
Sponsored by: Netflix
Use it instead of defining IN_RTLD by base sources that want P_OSREL_
defines in userspace, but are not rtld.
This allows to remove abuse of IN_RTLD from userspace.
Reviewed by: dchagin, markj, imp
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D38585
Enhanced REP MOVSB feature of CPUs starting from Ivy Bridge makes
REP MOVSB the fastest way to copy memory in most of cases. However
Intel Optimization Reference Manual says: "setting the DF to force
REP MOVSB to copy bytes from high towards low addresses will expe-
rience significant performance degradation". Measurements on Intel
Cascade Lake and Alder Lake, same as on AMD Zen3 show that it can
drop throughput to as low as 2.5-3.5GB/s, comparing to ~10-30GB/s
of REP MOVSQ or hand-rolled loop, used for non-ERMS CPUs.
This patch keeps ERMS use for forward ordered memory copies, but
removes it for backward overlapped moves where it does not work.
This is just a cosmetic sync with kernel, since libc does not use
ERMS at this time.
Reviewed by: mjg
MFC after: 2 weeks
Turns out clang converts "memcmp(foo, bar, len) == 0" and similar to
bcmp calls.
Reviewed by: emaste (previous version), jhb (previous version)
Differential Revision: https://reviews.freebsd.org/D34673
Preferably bcmp would just alias memcmp but there is build magic which
makes this problematic.
Reviewed by: jhb
Differential Revision: https://reviews.freebsd.org/D28846
Their uses have been replaced by _tcb_get() and _tcb_set() from
<machine/tls.h>.
Reviewed by: kib, jrtc27
Sponsored by: The University of Cambridge, Google Inc.
Differential Revision: https://reviews.freebsd.org/D33354
This is a tradeoff which saves jumps for smaller sizes while making
the 8-16 range slower (roughly in line with the other cases).
Tested with glibc test suite.
For example size 3 (most common with vfs namecache) (ops/s):
before: 407086026
after: 461391995
The regressed range of 8-16 (with 8 as example):
before: 540850489
after: 461671032
In all practical situations, the resolver visibility is static.
Requested by: markj
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Approved by: so (emaste)
Differential revision: https://reviews.freebsd.org/D20281
If dso uses initial exec TLS mode, rtld tries to allocate TLS in
static space. If there is no space left, the dlopen(3) fails. If space
if allocated, initial content from PT_TLS segment is distributed to
all threads' pcbs, which was missed and caused un-initialized TLS
segment for such dso after dlopen(3).
The mode is auto-detected either due to the relocation used, or if the
DF_STATIC_TLS dynamic flag is set. In the later case, the TLS segment
is tried to allocate earlier, which increases chance of the dlopen(3)
to succeed. LLD was recently fixed to properly emit the flag, ld.bdf
did it always.
Initial test by: dumbbell
Tested by: emaste (amd64), ian (arm)
Tested by: Gerald Aryeetey <aryeeteygerald_rogers.com> (arm64)
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
Differential revision: https://reviews.freebsd.org/D19072
In particular, use ifuncs for __getcontextx_size(), also calculate the
size of the extended save area in resolver. Same for __fillcontextx2().
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
No references to any of these exist in the tree. The list was also
erratic with different architectures exporting different things
(arm64 and riscv exported none).
Reviewed by: kib
Sponsored by: DARPA, AFRL
Differential Revision: https://reviews.freebsd.org/D18425
See the review for sample test results.
Reviewed by: kib (kernel part)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D18401
Handling sizes of > 32 backwards will be updated later.
Reviewed by: kib (kernel part)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D18387
For non-ERMS case the code used handle possible trailing bytes with
movsb first and then followed it up with movsq. This also happened
to alter how calculations were done for other cases.
Handle the tail with regular movs, just like when copying forward.
Use leaq to calculate the right offset from the get go, instead of
doing separate add and sub.
This adjusts the offset for non-rep cases so that they can be used
to handle the tail.
The routine is still a work in progress.
Sponsored by: The FreeBSD Foundation
Instead of jumping to locations which store the exact number of bytes,
use displacement to move the destination.
In particular the following clears an area between 8-16 (inclusive)
branch-free:
movq %r10,(%rdi)
movq %r10,-8(%rdi,%rcx)
For instance for rcx of 10 the second line is rdi + 10 - 8 = rdi + 2.
Writing 8 bytes starting at that offset overlaps with 6 bytes written
previously and writes 2 new, giving 10 in total.
Provides a nice win for smaller stores. Other ones are erratic depending
on the microarchitecture.
General idea taken from NetBSD (restricted use of the trick) and bionic
string functions (use for various ranges like in this patch).
Reviewed by: kib (previous version)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17660
- tidy up memset to have rax set earlier for small sizes
- finish the tail in memset with an overlapping store
- align memset buffers to 16 bytes before using rep stos
Sponsored by: The FreeBSD Foundation
The function is of limited use and is an almost a direct clone of
memmove/memcpy (with arguments swapped). Introduction of ERMS variants
of string routines would mean avoidable growth of libc.
bcopy will get redefined to a __builtin_memmove later on with this
symbol only left for compatibility.
Reviewed by: kib
Approved by: re (gjb)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17539
bcopy is left alone as it is expected to be converted to a C func.
Due to header mess ALIGN_TEXT is temporarily defined explicitly in memmove.S
Reviewed by: kib
Approved by: re (gjb)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17538
See r339205 for details.
An unused ERMS support is retained in the macro. It will be activated
after ifunc support lands.
Reviewed by: kib
Approved by: re (gjb)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17405
This is a depessimization, see r334537 for an explanation. Routines
remain significantly slower than they have to be.
bzero was removed from the kernel but remains in libc. Macroify to
accommodate differences to memset (no return value, always setting to 0).
The bzero.S file is left in place due to libc build magic which pulls in
a C variant if a matching .S file is missing.
Reviewed by: kib
Approved by: re (gjb)
Differential Revision: https://reviews.freebsd.org/D17355
Both are significantly slower than hand-coded loops. See r338963 for
kernel commit.
bcmp differs from memcmp by always returning 1 when a difference is
found, as opposed to going for a value bigger or lower than 0
depending on what it is. This means it can do less work. For now the
code is duplicated and modified. This will get deduplicated after
another round of optimization when memcmp will get a longer-term form.
Both tested with the glibc suite. While the suite does not have a test
for bcmp, I created a wrapper routine which verified that values match
(0 vs 0, 1 vs non-zero).
Reviewed by: kib
Approved by: re (gjb)
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17336
The change resembles what was done in r334537 for kernel routines.
While here take care of i386 variants. Note that primitives remain
suboptimal.
Reviewed by: kib (previous version)
Approved by: re (gjb)
Differential Revision: https://reviews.freebsd.org/D17167
Previously, libc.so would initialize its notion of the break address
using _end, a special symbol emitted by the static linker following
the bss section. Compatibility issues between lld and ld.bfd could
cause the wrong definition of _end (libc.so's definition rather than
that of the executable) to be used, breaking the brk()/sbrk()
interface.
Avoid this problem and future interoperability issues by simply not
relying on _end. Instead, modify the break() system call to return
the kernel's view of the current break address, and have libc
initialize its state using an extra syscall upon the first use of the
interface. As a side effect, this appears to fix brk()/sbrk() usage
in executables run with rtld direct exec, since the kernel and libc.so
no longer maintain separate views of the process' break address.
PR: 228574
Reviewed by: kib (previous version)
MFC after: 2 months
Differential Revision: https://reviews.freebsd.org/D15663
