The patch removes the the ELF_DURING_STARTUP optimization and assume
both .rel.dyn and .rel.plt might not be subsequent. This allows some
code simplification since relocation will be handled independently
where it is done on bootstrap.
At least on x86_64_64, I can not measure any performance implications.
Running 10000 time the command
LD_DEBUG=statistics ./elf/ld.so ./libc.so
And filtering the "total startup time in dynamic loader" result,
the geometric mean is:
patched master
Ryzen 7 5900x 24140 24952
i7-4510U 45957 45982
(The results do show some variation, I did not make any statistical
analysis).
It also allows build arm with lld, since it inserts ".ARM.exidx"
between ".rel.dyn" and ".rel.plt" for the loader.
Checked on x86_64-linux-gnu and arm-linux-gnueabihf.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Based on random input arguments. About 85% tuples have exponents
of the two arguments close together (+-1 range).
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Instead of inputs based on the algorithm implementation details.
About 85% tuples have exponents of the two arguments close
together (+-1 range).
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
The lld linker does not support TLSDESC for arm. The have-arm-tls-desc
is a leftover of 56583289b1 to support NaCL.
Reviewed-by: Fangrui Song <maskray@google.com>
The lld does not support R_ARM_GOTOFF32 to preemptible symbol (_dl_argv
has default visibility). Use the internal alias instead (one option
would to use HIDDEN_JUMPTARGET, bu the macro is not defined for
!__ASSEMBLER__ and I made this patch arm-specific to avoid require to
check extensivelly on other architecture it this might break something).
Checked on arm-linux-gnueabihf.
Reviewed-by: Fangrui Song <maskray@google.com>
Remove Prefer_AVX2_STRCMP to enable EVEX strcmp. When comparing 2 32-byte
strings, EVEX strcmp has been improved to require 1 load, 1 VPTESTM, 1
VPCMP, 1 KMOVD and 1 INCL instead of 2 loads, 3 VPCMPs, 2 KORDs, 1 KMOVD
and 1 TESTL while AVX2 strcmp requires 1 load, 2 VPCMPEQs, 1 VPMINU, 1
VPMOVMSKB and 1 TESTL. EVEX strcmp is now faster than AVX2 strcmp by up
to 40% on Tiger Lake and Ice Lake.
In strcmp-evex.S, to compare 2 32-byte strings, replace
VMOVU (%rdi, %rdx), %YMM0
VMOVU (%rsi, %rdx), %YMM1
/* Each bit in K0 represents a mismatch in YMM0 and YMM1. */
VPCMP $4, %YMM0, %YMM1, %k0
VPCMP $0, %YMMZERO, %YMM0, %k1
VPCMP $0, %YMMZERO, %YMM1, %k2
/* Each bit in K1 represents a NULL in YMM0 or YMM1. */
kord %k1, %k2, %k1
/* Each bit in K1 represents a NULL or a mismatch. */
kord %k0, %k1, %k1
kmovd %k1, %ecx
testl %ecx, %ecx
jne L(last_vector)
with
VMOVU (%rdi, %rdx), %YMM0
VPTESTM %YMM0, %YMM0, %k2
/* Each bit cleared in K1 represents a mismatch or a null CHAR
in YMM0 and 32 bytes at (%rsi, %rdx). */
VPCMP $0, (%rsi, %rdx), %YMM0, %k1{%k2}
kmovd %k1, %ecx
incl %ecx
jne L(last_vector)
It makes EVEX strcmp faster than AVX2 strcmp by up to 40% on Tiger Lake
and Ice Lake.
Co-Authored-By: Noah Goldstein <goldstein.w.n@gmail.com>
Add acosf function to bench-math and copy acosf-inputs to benchtests.
Motivation for this patch is to prepare for upcoming libmvec new
functions. Float and double version of libmvec functions stays
together.
acosf-inputs file generated from acos-inputs file using following
scaling formula:
f = d * (FLT_MAX/DBL_MAX)
Where d is input(double) and f is output(float). If scaled float value
is duplicate in new input file, nextafterf() function used to find next
float value, ensuring no duplicates.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Improve the random memcpy benchmark. Double the number of tests and increase
the size of the memory region to test between 32KB and 1024KB. This improves
accuracy on modern cores. Clean up formatting of the frequency array.
Reviewed-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
My build-many-glibcs.py bot shows -Waggressive-loop-optimizations
errors building the glibc testsuite for 32-bit architectures with GCC
mainline, which seem to have appeared between GCC commits
4abc0c196b10251dc80d0743ba9e8ab3e56c61ed and
d8edfadfc7a9795b65177a50ce44fd348858e844:
In function 'dynarray_long_noscratch_resize',
inlined from 'test_long_overflow' at tst-dynarray.c:489:5,
inlined from 'do_test' at tst-dynarray.c:571:3:
../malloc/dynarray-skeleton.c:391:36: error: iteration 1073741823 invokes undefined behavior [-Werror=aggressive-loop-optimizations]
391 | DYNARRAY_ELEMENT_INIT (&list->u.dynarray_header.array[i]);
tst-dynarray.c:39:37: note: in definition of macro 'DYNARRAY_ELEMENT_INIT'
39 | #define DYNARRAY_ELEMENT_INIT(e) (*(e) = 23)
| ^
In file included from tst-dynarray.c:42:
../malloc/dynarray-skeleton.c:389:37: note: within this loop
389 | for (size_t i = old_size; i < size; ++i)
| ~~^~~~~~
In function 'dynarray_long_resize',
inlined from 'test_long_overflow' at tst-dynarray.c:479:5,
inlined from 'do_test' at tst-dynarray.c:571:3:
../malloc/dynarray-skeleton.c:391:36: error: iteration 1073741823 invokes undefined behavior [-Werror=aggressive-loop-optimizations]
391 | DYNARRAY_ELEMENT_INIT (&list->u.dynarray_header.array[i]);
tst-dynarray.c:27:37: note: in definition of macro 'DYNARRAY_ELEMENT_INIT'
27 | #define DYNARRAY_ELEMENT_INIT(e) (*(e) = 17)
| ^
In file included from tst-dynarray.c:28:
../malloc/dynarray-skeleton.c:389:37: note: within this loop
389 | for (size_t i = old_size; i < size; ++i)
| ~~^~~~~~
I don't know what GCC change made these errors appear, or why they
only appear for 32-bit architectures. However, the warnings appear to
be both true (that iteration would indeed involve undefined behavior
if executed) and useless in this particular case (that iteration is
never executed, because the allocation size overflows and so the
allocation fails - but the check for allocation size overflow is in a
separate source file and so can't be seen by the compiler when
compiling this test). So use the DIAG_* macros to disable
-Waggressive-loop-optimizations around the calls in question to
dynarray_long_resize and dynarray_long_noscratch_resize in this test.
Tested with build-many-glibcs.py (GCC mainline) for arm-linux-gnueabi,
where it restores a clean testsuite build.
Compiling mmap_internal fails to compile when we use -1 for MMAP2_PAGE_UNIT
on 32 bit architectures. The error is as follows:
../sysdeps/unix/sysv/linux/mmap_internal.h:30:8: error: unknown type
name 'uint64_t'
|
30 | static uint64_t page_unit;
|
| ^~~~~~~~
Fix by adding including stdint.h.
Since some linkers (for instance lld for i386) does not support it
for all architectures.
Checked on i686-linux-gnu.
Reviewed-by: Fangrui Song <maskray@google.com>
GCC does not print the correct linker when -fuse-ld=lld is used with
the -print-prog-name=ld:
$ gcc -v 2>&1 | tail -n 1
gcc version 11.2.0 (Ubuntu 11.2.0-7ubuntu2)
$ gcc
ld
This is different than for gold:
$ gcc -fuse-ld=gold -print-prog-name=ld
ld.gold
Using ld.lld as the static linker name prints the expected result.
This is only required when -fuse-ld=lld is used, if lld is used as
the 'ld' programs (through a symlink) LIBC_PROG_BINUTILS works
as expected.
Checked on x86_64-linux-gnu.
Reviewed-by: Fangrui Song <maskray@google.com>
These tests takes the address of a protected symbol (foo_protected)
and lld does not support copy relocations on protected data symbols.
Checked on x86_64-linux-gnu.
Reviewed-by: Fangrui Song <maskray@google.com>
Hoist the NULL check for malloc_usable_size into its entry points in
malloc-debug and malloc and assume non-NULL in all callees. This fixes
BZ #28506
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Reviewed-by: Florian Weimer <fweimer@redhat.com>
Reviewed-by: Richard W.M. Jones <rjones@redhat.com>
The following commit:
commit cf4fd28ea4
Author: Noah Goldstein <goldstein.w.n@gmail.com>
Date: Tue Oct 26 19:43:18 2021 -0500
Broke --disable-multi-arch build for x86_64 because x86_64/memcmpeq.S
was not defined outside of multiarch and the alias for __memcmpeq in
x86_64/memcmp.S was removed.
This commit fixes that issue by adding x86_64/memcmpeq.S.
make xcheck passes on x86_64 with and without --disable-multi-arch
There are several packages like sysvinit and buildroot that expect
-lutil to work. Rather than impacting them with having to change
the linker flags provide an empty libutil.a.
A non-local STV_DEFAULT defined symbol is by default preemptible in a
shared object. j/jal cannot target a preemptible symbol. On other
architectures, such a jump instruction either causes PLT [BZ #18822], or
if short-ranged, sometimes rejected by the linker (but not by GNU ld's
riscv port [ld PR/28509]).
Use HIDDEN_JUMPTARGET to target a non-preemptible symbol instead.
With this patch, ld.so and libc.so can be linked with LLD if source
files are compiled/assembled with -mno-relax/-Wa,-mno-relax.
Acked-by: Palmer Dabbelt <palmer@dabbelt.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
No bug. This commit adds new optimized __memcmpeq implementation for
evex.
The primary optimizations are:
1) skipping the logic to find the difference of the first mismatched
byte.
2) not updating src/dst addresses as the non-equals logic does not
need to be reused by different areas.
No bug. This commit adds new optimized __memcmpeq implementation for
avx2.
The primary optimizations are:
1) skipping the logic to find the difference of the first mismatched
byte.
2) not updating src/dst addresses as the non-equals logic does not
need to be reused by different areas.
No bug. This commit does not modify any of the memcmp
implementation. It just adds __memcmpeq ifdefs to skip obvious cases
where computing the proper 1/-1 required by memcmp is not needed.
No bug. This commit adds support for __memcmpeq to be implemented
seperately from memcmp. Support is added for versions optimized with
sse2, avx2, and evex.
No bug. This commit adds __memcmpeq benchmarks. The benchmarks just
use the existing ones in memcmp. This will be useful for testing
implementations of __memcmpeq that do not just alias memcmp.
No bug. This commit just adds __memcmpeq as a build target so that
implementations for __memcmpeq that are not just aliases to memcmp can
be supported.
No bug.
This commit adds tests for the new function __memcmpeq. The new tests
use the existing tests in 'test-memcmp.c' but relax the result
requirement to only check for zero or non-zero returns.
All string tests include test-memcmpeq are passing.
No bug.
This commit adds support for __memcmpeq() as a new ABI for all
targets. In this commit __memcmpeq() is implemented only as an alias
to the corresponding targets memcmp() implementation. __memcmpeq() is
added as a new symbol starting with GLIBC_2.35 and defined in string.h
with comments explaining its behavior. Basic tests that it is callable
and works where added in string/tester.c
As discussed in the proposal "Add new ABI '__memcmpeq()' to libc"
__memcmpeq() is essentially a reserved namespace for bcmp(). The means
is shares the same specifications as memcmp() except the return value
for non-equal byte sequences is any non-zero value. This is less
strict than memcmp()'s return value specification and can be better
optimized when a boolean return is all that is needed.
__memcmpeq() is meant to only be called by compilers if they can prove
that the return value of a memcmp() call is only used for its boolean
value.
All tests in string/tester.c passed. As well build succeeds on
x86_64-linux-gnu target.
When LIBC_LINKER_FEATURE is used to check a linker option with the equal
sign, it will likely fail because the LD -v --help output may look like
`-z lam-report=[none|warning|error]` while the needle is something like
`-z lam-report=warning`.
The LD -v --help filter doesn't save much time, so just remove it.
The global test is linked with globalmod1.so which dlopens reldepmod4.so.
Make global.out depend on reldepmod4.so. This fixes BZ #28457.
Reviewed-by: Florian Weimer <fweimer@redhat.com>
This commit replaces two usages of SSE2 'movups' with AVX 'vmovdqu'.
it could potentially be dangerous to use SSE2 if this function is ever
called without using 'vzeroupper' beforehand. While compilers appear
to use 'vzeroupper' before function calls if AVX2 has been used, using
SSE2 here is more brittle. Since it is not absolutely necessary it
should be avoided.
It costs 2-extra bytes but the extra bytes should only eat into
alignment padding.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
This second patch contains the actual implementation of a new sorting algorithm
for shared objects in the dynamic loader, which solves the slow behavior that
the current "old" algorithm falls into when the DSO set contains circular
dependencies.
The new algorithm implemented here is simply depth-first search (DFS) to obtain
the Reverse-Post Order (RPO) sequence, a topological sort. A new l_visited:1
bitfield is added to struct link_map to more elegantly facilitate such a search.
The DFS algorithm is applied to the input maps[nmap-1] backwards towards
maps[0]. This has the effect of a more "shallow" recursion depth in general
since the input is in BFS. Also, when combined with the natural order of
processing l_initfini[] at each node, this creates a resulting output sorting
closer to the intuitive "left-to-right" order in most cases.
Another notable implementation adjustment related to this _dl_sort_maps change
is the removing of two char arrays 'used' and 'done' in _dl_close_worker to
represent two per-map attributes. This has been changed to simply use two new
bit-fields l_map_used:1, l_map_done:1 added to struct link_map. This also allows
discarding the clunky 'used' array sorting that _dl_sort_maps had to sometimes
do along the way.
Tunable support for switching between different sorting algorithms at runtime is
also added. A new tunable 'glibc.rtld.dynamic_sort' with current valid values 1
(old algorithm) and 2 (new DFS algorithm) has been added. At time of commit
of this patch, the default setting is 1 (old algorithm).
Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
This is the first of a 2-part patch set that fixes slow DSO sorting behavior in
the dynamic loader, as reported in BZ #17645. In order to facilitate such a
large modification to the dynamic loader, this first patch implements a testing
framework for validating shared object sorting behavior, to enable comparison
between old/new sorting algorithms, and any later enhancements.
This testing infrastructure consists of a Python script
scripts/dso-ordering-test.py' which takes in a description language, consisting
of strings that describe a set of link dependency relations between DSOs, and
generates testcase programs and Makefile fragments to automatically test the
described situation, for example:
a->b->c->d # four objects linked one after another
a->[bc]->d;b->c # a depends on b and c, which both depend on d,
# b depends on c (b,c linked to object a in fixed order)
a->b->c;{+a;%a;-a} # a, b, c serially dependent, main program uses
# dlopen/dlsym/dlclose on object a
a->b->c;{}!->[abc] # a, b, c serially dependent; multiple tests generated
# to test all permutations of a, b, c ordering linked
# to main program
(Above is just a short description of what the script can do, more
documentation is in the script comments.)
Two files containing several new tests, elf/dso-sort-tests-[12].def are added,
including test scenarios for BZ #15311 and Redhat issue #1162810 [1].
Due to the nature of dynamic loader tests, where the sorting behavior and test
output occurs before/after main(), generating testcases to use
support/test-driver.c does not suffice to control meaningful timeout for ld.so.
Therefore a new utility program 'support/test-run-command', based on
test-driver.c/support_test_main.c has been added. This does the same testcase
control, but for a program specified through a command-line rather than at the
source code level. This utility is used to run the dynamic loader testcases
generated by dso-ordering-test.py.
[1] https://bugzilla.redhat.com/show_bug.cgi?id=1162810
Signed-off-by: Chung-Lin Tang <cltang@codesourcery.com>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
Currently the timeout for each iconv test is hard coded to 3 seconds.
On my OpenRISC test platform this is too slow and the test fails with a
HANG error.
This change uses the available TIMEOUTFACTOR to compute the timeout.
The default value is still 3.
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
GCC 4.9.0 added the alloc_align attribute to say that a function
argument specifies the alignment of the returned pointer. Clang supports
the attribute too. Using the attribute can allow a compiler to generate
better code if it knows the returned pointer has a minimum alignment.
See https://gcc.gnu.org/PR60092 for more details.
GCC implicitly knows the semantics of aligned_alloc and posix_memalign,
but not the obsolete memalign. As a result, GCC generates worse code
when memalign is used, compared to aligned_alloc. Clang knows about
aligned_alloc and memalign, but not posix_memalign.
This change adds a new __attribute_alloc_align__ macro to <sys/cdefs.h>
and then uses it on memalign (where it helps GCC) and aligned_alloc
(where GCC and Clang already know the semantics, but it doesn't hurt)
and xposix_memalign. It can't be used on posix_memalign because that
doesn't return a pointer (the allocated pointer is returned via a void**
parameter instead).
Unlike the alloc_size attribute, alloc_align only allows a single
argument. That means the new __attribute_alloc_align__ macro doesn't
really need to be used with double parentheses to protect a comma
between its arguments. For consistency with __attribute_alloc_size__
this patch defines it the same way, so that double parentheses are
required.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
Reviewed-by: Carlos O'Donell <carlos@redhat.com>
Tested-by: Carlos O'Donell <carlos@redhat.com>
According to C11 6.6p6, `const int` as an operand may not make up a
constant expression. GCC -O0 errors:
../sysdeps/unix/sysv/linux/opendir.c:107:19: error: static_assert expression is not an integral constant expression
_Static_assert (allocation_size >= sizeof (struct dirent64),
-O2 -Wpedantic has a similar warning.
See https://gcc.gnu.org/PR102502 for GCC's inconsistency.
Use enum which is guaranteed to be a constant expression.
This also makes the file compilable with Clang.
Fixes: 4b962c9e85 ("linux: Simplify opendir buffer allocation")
1. Add sysdeps/x86_64/fpu/Makeconfig to auto-generate libmvec.mk, which
contains libmvec ABI test dependencies and CFLAGS, in the build directory.
2. Include libmvec.mk for libmvec ABI test dependencies and CFLAGS.
Tested on SSE4, AVX, AVX2 and AVX512 machines.
Reviewed-by: Noah Goldstein <goldstein.w.n@gmail.com>
The powerpc optimization to provide a fast stacktrace requires some
ad-hoc code to handle Linux signal frames and the change is fragile
once the kernel decides to slight change its execution sequence [1].
The generic implementation work as-is and it should be future proof
since the kernel provides the expected CFI directives in vDSO shared
page.
Checked on powerpc-linux-gnu, powerpc64le-linux-gnu, and
powerpc64-linux-gnu.
[1] https://sourceware.org/pipermail/libc-alpha/2021-January/122027.html
As noted in bug 28475, the access attribute on memfrob in <string.h>
is incorrect: the function both reads and writes the memory pointed to
by its argument, so it needs to use __read_write__, not
__write_only__. This incorrect attribute results in a build failure
for accessing uninitialized memory for s390x-linux-gnu-O3 with
build-many-glibcs.py using GCC mainline.
Correct the attribute. Fixing this shows up that some calls to
memfrob in elf/ tests are reading uninitialized memory; I'm not
entirely sure of the purpose of those calls, but guess they are about
ensuring that the stack space is indeed allocated at that point in the
function, and so it matters that they are calling a function whose
semantics are unknown to the compiler. Thus, change the first memfrob
call in those tests to use explicit_bzero instead, as suggested by
Florian in
<https://sourceware.org/pipermail/libc-alpha/2021-October/132119.html>,
to avoid the use of uninitialized memory.
Tested for x86_64, and with build-many-glibcs.py (GCC mainline) for
s390x-linux-gnu-O3.
In _FORTIFY_SOURCE=3, the size expression may be non-constant,
resulting in branches in the inline functions remaining intact and
causing a tiny overhead. Clang (and in future, gcc) make sure that
the -1 case is always safe, i.e. any comparison of the generated
expression with (size_t)-1 is always false so that bit is taken care
of. The rest is avoidable since we want the _chk variant whenever we
have a size expression and it's not -1.
Rework the conditionals in a uniform way to clearly indicate two
conditions at compile time:
- Either the size is unknown (-1) or we know at compile time that the
operation length is less than the object size. We can call the
original function in this case. It could be that either the length,
object size or both are non-constant, but the compiler, through
range analysis, is able to fold the *comparison* to a constant.
- The size and length are known and the compiler can see at compile
time that operation length > object size. This is valid grounds for
a warning at compile time, followed by emitting the _chk variant.
For everything else, emit the _chk variant.
This simplifies most of the fortified function implementations and at
the same time, ensures that only one call from _chk or the regular
function is emitted.
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Reviewed-by: Adhemerval Zanella <adhemerval.zanella@linaro.org>
In the context of a function definition, the size hints imply that the
size of an object pointed to by one parameter is another parameter.
This doesn't make sense for the fortified versions of the functions
since that's the bit it's trying to validate.
This is harmless with __builtin_object_size since it has fairly simple
semantics when it comes to objects passed as function parameters.
With __builtin_dynamic_object_size we could (as my patchset for gcc[1]
already does) use the access attribute to determine the object size in
the general case but it misleads the fortified functions.
Basically the problem occurs when access attributes are present on
regular functions that have inline fortified definitions to generate
_chk variants; the attributes get inherited by these definitions,
causing problems when analyzing them. For example with poll(fds, nfds,
timeout), nfds is hinted using the __attr_access as being the size of
fds.
Now, when analyzing the inline function definition in bits/poll2.h, the
compiler sees that nfds is the size of fds and tries to use that
information in the function body. In _FORTIFY_SOURCE=3 case, where the
object size could be a non-constant expression, this information results
in the conclusion that nfds is the size of fds, which defeats the
purpose of the implementation because we're trying to check here if nfds
does indeed represent the size of fds. Hence for this case, it is best
to not have the access attribute.
With the attributes gone, the expression evaluation should get delayed
until the function is actually inlined into its destinations.
Disable the access attribute for fortified function inline functions
when building at _FORTIFY_SOURCE=3 to make this work better. The
access attributes remain for the _chk variants since they can be used
by the compiler to warn when the caller is passing invalid arguments.
[1] https://gcc.gnu.org/pipermail/gcc-patches/2021-October/581125.html
Signed-off-by: Siddhesh Poyarekar <siddhesh@sourceware.org>
Unlike GCC, Clang parses asm statements and verifies they are valid
instructions/directives. Place the magic @@@ into a comment to avoid
a parse error.
