- Use good ELF technique in cosmo_dlopen()
- Make strerror() conform more to other libc impls
- Introduce __clear_cache() and use it in cosmo_dlopen()
- Remove libc/fmt/fmt.h header (trying to kill off LIBC_FMT)
- We now serialize the file descriptor table when spawning / executing
processes on Windows. This means you can now inherit more stuff than
just standard i/o. It's needed by bash, which duplicates the console
to file descriptor #255. We also now do a better job serializing the
environment variables, so you're less likely to encounter E2BIG when
using your bash shell. We also no longer coerce environ to uppercase
- execve() on Windows now remotely controls its parent process to make
them spawn a replacement for itself. Then it'll be able to terminate
immediately once the spawn succeeds, without having to linger around
for the lifetime as a shell process for proxying the exit code. When
process worker thread running in the parent sees the child die, it's
given a handle to the new child, to replace it in the process table.
- execve() and posix_spawn() on Windows will now provide CreateProcess
an explicit handle list. This allows us to remove handle locks which
enables better fork/spawn concurrency, with seriously correct thread
safety. Other codebases like Go use the same technique. On the other
hand fork() still favors the conventional WIN32 inheritence approach
which can be a little bit messy, but is *controlled* by guaranteeing
perfectly clean slates at both the spawning and execution boundaries
- sigset_t is now 64 bits. Having it be 128 bits was a mistake because
there's no reason to use that and it's only supported by FreeBSD. By
using the system word size, signal mask manipulation on Windows goes
very fast. Furthermore @asyncsignalsafe funcs have been rewritten on
Windows to take advantage of signal masking, now that it's much more
pleasant to use.
- All the overlapped i/o code on Windows has been rewritten for pretty
good signal and cancelation safety. We're now able to ensure overlap
data structures are cleaned up so long as you don't longjmp() out of
out of a signal handler that interrupted an i/o operation. Latencies
are also improved thanks to the removal of lots of "busy wait" code.
Waits should be optimal for everything except poll(), which shall be
the last and final demon we slay in the win32 i/o horror show.
- getrusage() on Windows is now able to report RUSAGE_CHILDREN as well
as RUSAGE_SELF, thanks to aggregation in the process manager thread.
- Every unit test now passes on Apple Silicon. The final piece of this
puzzle was porting our POSIX threads cancelation support, since that
works differently on ARM64 XNU vs. AMD64. Our semaphore support on
Apple Silicon is also superior now compared to AMD64, thanks to the
grand central dispatch library which lets *NSYNC locks go faster.
- The Cosmopolitan runtime is now more stable, particularly on Windows.
To do this, thread local storage is mandatory at all runtime levels,
and the innermost packages of the C library is no longer being built
using ASAN. TLS is being bootstrapped with a 128-byte TIB during the
process startup phase, and then later on the runtime re-allocates it
either statically or dynamically to support code using _Thread_local.
fork() and execve() now do a better job cooperating with threads. We
can now check how much stack memory is left in the process or thread
when functions like kprintf() / execve() etc. call alloca(), so that
ENOMEM can be raised, reduce a buffer size, or just print a warning.
- POSIX signal emulation is now implemented the same way kernels do it
with pthread_kill() and raise(). Any thread can interrupt any other
thread, regardless of what it's doing. If it's blocked on read/write
then the killer thread will cancel its i/o operation so that EINTR can
be returned in the mark thread immediately. If it's doing a tight CPU
bound operation, then that's also interrupted by the signal delivery.
Signal delivery works now by suspending a thread and pushing context
data structures onto its stack, and redirecting its execution to a
trampoline function, which calls SetThreadContext(GetCurrentThread())
when it's done.
- We're now doing a better job managing locks and handles. On NetBSD we
now close semaphore file descriptors in forked children. Semaphores on
Windows can now be canceled immediately, which means mutexes/condition
variables will now go faster. Apple Silicon semaphores can be canceled
too. We're now using Apple's pthread_yield() funciton. Apple _nocancel
syscalls are now used on XNU when appropriate to ensure pthread_cancel
requests aren't lost. The MbedTLS library has been updated to support
POSIX thread cancelations. See tool/build/runitd.c for an example of
how it can be used for production multi-threaded tls servers. Handles
on Windows now leak less often across processes. All i/o operations on
Windows are now overlapped, which means file pointers can no longer be
inherited across dup() and fork() for the time being.
- We now spawn a thread on Windows to deliver SIGCHLD and wakeup wait4()
which means, for example, that posix_spawn() now goes 3x faster. POSIX
spawn is also now more correct. Like Musl, it's now able to report the
failure code of execve() via a pipe although our approach favors using
shared memory to do that on systems that have a true vfork() function.
- We now spawn a thread to deliver SIGALRM to threads when setitimer()
is used. This enables the most precise wakeups the OS makes possible.
- The Cosmopolitan runtime now uses less memory. On NetBSD for example,
it turned out the kernel would actually commit the PT_GNU_STACK size
which caused RSS to be 6mb for every process. Now it's down to ~4kb.
On Apple Silicon, we reduce the mandatory upstream thread size to the
smallest possible size to reduce the memory overhead of Cosmo threads.
The examples directory has a program called greenbean which can spawn
a web server on Linux with 10,000 worker threads and have the memory
usage of the process be ~77mb. The 1024 byte overhead of POSIX-style
thread-local storage is now optional; it won't be allocated until the
pthread_setspecific/getspecific functions are called. On Windows, the
threads that get spawned which are internal to the libc implementation
use reserve rather than commit memory, which shaves a few hundred kb.
- sigaltstack() is now supported on Windows, however it's currently not
able to be used to handle stack overflows, since crash signals are
still generated by WIN32. However the crash handler will still switch
to the alt stack, which is helpful in environments with tiny threads.
- Test binaries are now smaller. Many of the mandatory dependencies of
the test runner have been removed. This ensures many programs can do a
better job only linking the the thing they're testing. This caused the
test binaries for LIBC_FMT for example, to decrease from 200kb to 50kb
- long double is no longer used in the implementation details of libc,
except in the APIs that define it. The old code that used long double
for time (instead of struct timespec) has now been thoroughly removed.
- ShowCrashReports() is now much tinier in MODE=tiny. Instead of doing
backtraces itself, it'll just print a command you can run on the shell
using our new `cosmoaddr2line` program to view the backtrace.
- Crash report signal handling now works in a much better way. Instead
of terminating the process, it now relies on SA_RESETHAND so that the
default SIG_IGN behavior can terminate the process if necessary.
- Our pledge() functionality has now been fully ported to AARCH64 Linux.
This change fixes Cosmopolitan so it has fewer opinions about compiler
warnings. The whole repository had to be cleaned up to be buildable in
-Werror -Wall mode. This lets us benefit from things like strict const
checking. Some actual bugs might have been caught too.
- Fix mkdeps.com out of memory error
- Remove static memory from __get_cpu_count()
- Add support for passing hyphen to cat in cocmd
- Change more ZipOS errors from ENOTSUP to EROFS
- Specify mem_unit in sysinfo() output on BSD OSes
This way complex runtime features (e.g. ftrace, symbol tables) can
always yoink zipos support. This is important now that apelink.com
automates embedding symbol tables for multiple cpus.
This change ports APE Loader to Linux AARCH64, so that Raspberry Pi
users can run programs like redbean, without the executable needing
to modify itself. Progress has also slipped into this change on the
issue of making progress better conforming to user expectations and
industry standards regarding which symbols we're allowed to declare
- Found some bugs in LLVM compiler-rt library
- The useless LIBC_STUBS package is now deleted
- Improve the overflow checking story even further
- Get chibicc tests working in MODE=dbg mode again
- The libc/isystem/ headers now have correctly named guards
- Work towards improving non-optimized build support
- Introduce MODE=zero which is -O0 without ASAN/UBSAN
- Use system GCC when ~/.cosmo.mk has USE_SYSTEM_TOOLCHAIN=1
- Have package.com check .privileged code doesn't call non-privileged
This change implements a new approach to function call logging, that's
based on the GCC flag: -fpatchable-function-entry. Read the commentary
in build/config.mk to learn how it works.
llama.com can now load weights that use the new file format which was
introduced a few weeks ago. Note that, unlike llama.cpp, we will keep
support for old file formats in our tool so you don't need to convert
your weights when the upstream project makes breaking changes. Please
note that using ggjt v3 does make avx2 inference go 5% faster for me.
There's a new program named ape/ape-m1.c which will be used to build an
embeddable binary that can load ape and elf executables. The support is
mostly working so far, but still chasing down ABI issues.
- Fix UX issues with llama.com
- Do housekeeping on libm code
- Add more vectorization to GGML
- Get GGJT quantizer programs working well
- Have the quantizer keep the output layer as f16c
- Prefetching improves performance 15% if you use fewer threads
make -j8 o//third_party/radpajama/radpajama.com
make -j8 o//third_party/radpajama/radpajama-chat.com
This change gets the radpajama.mk config working. This package depends
on THIRD_PARTY_GGML but it's configured to call ggjt_v1(), so that the
library will provide the old quantizers. The ggml_quantize_chunk() API
will now dispatch to older quantizers based on the configured version.
This change makes quantized models (e.g. q4_0) go 10% faster on Macs
however doesn't offer much improvement for Intel PC hardware.
This change syncs llama.cpp 699b1ad7fe6f7b9e41d3cb41e61a8cc3ea5fc6b5
which recently made a breaking change to nearly all its file formats
without any migration. Since that'll break hundreds upon hundreds of
models on websites like HuggingFace llama.com will support both file
formats because llama.com will never ever break the GGJT file format
Example use case for JSON completion:
$ m=opt
$ make -j16 m=$m o/$m/third_party/ggml/llama.com
$ o/$m/third_party/ggml/llama.com -m llama.bin -p '{"key": "life", "val": ' -r '}'
42}
This provides better control. More sophisticated facilities for
controlling text generation will be provided soon enough.
- Introduce -v and --verbose flags
- Don't print stats / diagnostics unless -v is passed
- Reduce --top_p default from 0.95 to 0.70
- Change --reverse-prompt to no longer imply --interactive
- Permit --reverse-prompt specifying custom EOS if non-interactive
