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.
This may happen if no Status or Content-Type is set and if the written
content is longer than 100B, the compression is going to be applied
after coroutine.yield, even though it shouldn't be.
- 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
- Perform some housekeeping on scalar math function code
- Import ARM's Optimized Routines for SIMD string processing
- Upgrade to latest Chromium zlib and enable more SIMD optimizations
- Introduce epoll_pwait()
- Rewrite -ftrapv and ffs() libraries in C code
- Use more FreeBSD code in math function library
- Get significantly more tests passing on qemu-aarch64
- Fix many Musl long double functions that were broken on AARCH64
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
- Utilities like pledge.com now build
- kprintf() will no longer balk at 48-bit addresses
- There's a new aarch64-dbg build mode that should work
- gc() and defer() are mostly pacified; avoid using them on aarch64
- THIRD_PART_STB now has Arm Neon intrinsics for fast image handling
It's now possible to run commands like:
make -j8 m=aarch64 o/aarch64/test/libc/str
Which will cross-compile and run the test suites in a qemu-aarch64
binary that's vendored in the third_party/qemu/ folder within your
x86_64 build environment.
When redbean is functioning as a Lua interpreter, the `-e` flag should
behave the same way as other open source language interpreters. Namely
it should exit after evaluating the code rather than showing the REPL.
* Update redbean SQLite config to handle more options
This requires moving sqlite3_initialize call to open, as configuration
should be allowed before initialization is done. This call is effective
only for the first time and then no-op after that.
* Fix redbean SQLite for closing db with already finalized statements
There is a loop in cleanupdb that finalizes all vms that are associated
with that db when it's being closed. Under some circumstances (detailed
below) that loop may contain references pointing to already collected
objects, thus leading to SIGSEGV when those references are used.
This may happen with the following sequence of events ("VM" is the name
used in lsqlite and describes the same thing as "statement"):
1. A finalized statement is created (for example, by preparing an empty
string or a string with no statement that is still grammatically valid).
2. This statement goes out of scope before the DB object it's associated
with does and is garbage collected.
3. When it's garbage collected, dbvm_gc method is called, which checks
for svm->vm being not NULL.
4. Since the VM is already finalized, cleanupvm method is not called,
so the VM reference is not removed from the table of VMs tracked for
that DB.
5. When the DB is finally closed or garbage collected, all the VMs
associated with it are accessed to be finalized, including the ones that
have been garbage collected and have invalid references (thus leading
to a memory access error).
Here is an example of a stacktrace from the resulting SIGSEGV:
70000003de20 5df71a getgeneric+26
70000003fac0 5dfc7f luaH_get+111
70000003faf0 5e06c8 luaH_set+40
70000003fb20 5c5bd7 aux_rawset+55
70000003fb50 5c70cb lua_rawset+27
70000003fb60 4fa8e7 cleanupvm+71
70000003fb80 4fa988 cleanupdb+88
70000003fbc0 4fe899 db_gc+41
One way to fix this is to use userdata references (which anchor their
targets) instead of lightuserdata references (which do not), but this
would prevent the targets (VMs) from being garbage collected until the
DB itself is garbage collected, so this needs to be combined with
weakening the keys in the DB table. The code in cleanupdb to remove the
VM references is no longer needed, as this is handled by having weak keys.
The patch also switches to using close_v2, as it is intended for use
with garbage collected languages where the order in which destructors
are called is arbitrary, as is the case here.
* Remove GC collection from redbean SQLite session
The behavior of sqlite3session_delete is undefined after the DB
connection is closed, so we need to avoid calling it from gc handler.
It's been superseded by token bucket processing, does not take time
into considerations (only the number of fragments), and affects file
uploads that may require a large number of reads.
