* Optimized adler32 implementations.
- Adapted adler32 implementation from Crypto++ public domain library.
- Added unit tests for adler32 hashing algorithm.
* tabified adler32 implementation to match stdlib.
* Formatting to be consistent. Make unrolling use macro.
---------
Co-authored-by: soerlemans <sebasoerlemans+git@gmail.com>
Co-authored-by: Christoffer Lerno <christoffer@aegik.com>
* split `test_hash_vector_macro` into per-type functions
it went from a single 18367 instructions function, to eleven separate
functions with (1000 to 1600) instructions each
thats 15649 instructions total, 2718 less instructions overall.
```
$ build/c3c compile-test --suppress-run -O1 --print-large-functions -D
SLOW_TESTS test/unit/ 2>&1 | sort -n | grep "test_hash_vector"
1041 instructions found in
std::core::builtins:test_hash_vector_int128
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1041 instructions found in
std::core::builtins:test_hash_vector_uint128
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1305 instructions found in std::core::builtins:test_hash_vector_long
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1305 instructions found in
std::core::builtins:test_hash_vector_ulong
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1453 instructions found in std::core::builtins:test_hash_vector_int
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1453 instructions found in std::core::builtins:test_hash_vector_uint
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1565 instructions found in
std::core::builtins:test_hash_vector_short
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1565 instructions found in
std::core::builtins:test_hash_vector_ushort
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1629 instructions found in std::core::builtins:test_hash_vector_char
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1629 instructions found in
std::core::builtins:test_hash_vector_ichar
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
1663 instructions found in std::core::builtins:test_hash_vector_bool
(/home/mb/scripts/c3c/test/unit/stdlib/core/builtintests.c3) - function
is very long.
```
* made the `test_hash_vector_internal` generic
- replaced manual unrolling with loop structures and constant arrays
- instruction count reduced from 12445 to 4016
- maybe about 1 to 2% performance loss on some benchs but take this
number with a grain of salt.
* optimize `test_ct_intlog2` while still covering all 128 bit positions
* refactor whirlpool to reduce code bloat
replaced the fully unrolled round loop with a runtime loop, reducing
instruction count by 80k in `process_block` and yielding aprox 30%
performance boost due to improved cache locality.
* use compile-time arrays for `test_ct_intlog2`
- Add retry loops to all package manager commands (apt, apk, pacman,
pkg).
- Implement daily package caching to minimize external server hits.
- Enable Docker layer caching in CI using buildx and GHA backend.
- Sync CI Docker build steps with the build-with-docker.sh logic.
- We don't, or more accurately, can't actually use
`build-with-docker.sh` anymore in CI. Maybe we should move it to
scripts/tools/ to keep the root directory clean.
- add ramdisk for BSD, shave 30/60 seconds.
- change BDSs to cross-platform-actions/action for a cleaner ci and skip
SLOW_TESTS for BSDs: this shaves a few seconds
- set global CACHE_INVALIDATION_SEED
* codepage: add single-byte code page support
Add std::encoding::codepage with a shared engine for converting between
single-byte code pages and UTF-8 using table-driven mappings.
Introduce generated tables and wrappers for several code pages[1] each
exposing encode/decode helpers built on a common CodePageTable
structure.
The mapping data is generated by cpgen[2] from the Unicode Consortium’s
published code page mapping files and follows the Unicode standard’s
interpretation of control characters (abstract characters) rather than
historical VGA glyph shapes.
[1] Code page overview/groups:
DOS/OEM code pages (legacy PC):
cp437 cp737 cp775 cp850 cp852 cp855 cp857 cp860 cp861 cp862 cp863
cp864 cp865 cp866 cp869 cp874
Windows code pages (ANSI/Windows):
cp1250 cp1251 cp1252 cp1253 cp1254 cp1255 cp1256 cp1257 cp1258
ISO/IEC 8859 series (Latin/Regional):
iso_8859_1 iso_8859_2 iso_8859_3 iso_8859_4 iso_8859_5 iso_8859_6
iso_8859_7 iso_8859_8 iso_8859_9 iso_8859_10 iso_8859_11 iso_8859_13
iso_8859_14 iso_8859_15 iso_8859_16
[2] github.com/konimarti/cpgen
Signed-off-by: Koni Marti <koni.marti@gmail.com>
* codepage: change encoding format, streamline api
* Use enum to collect the data.
---------
Signed-off-by: Koni Marti <koni.marti@gmail.com>
Co-authored-by: Christoffer Lerno <christoffer@aegik.com>
- Uses `dlopen` to load libcurl at runtime for better portability.
- Replaces the FETCH_AVAILABLE macro with download_available() check.
- Provides descriptive error messages when libcurl is missing.
When C3_FETCH_LLVM=ON, the fetched LLVM artifact is self-contained.
Two issues caused the macOS release binary to embed hardcoded
/opt/homebrew paths, making it fail on systems without Homebrew:
1. /opt/homebrew/lib was unconditionally appended to LLVM_LIBRARY_DIRS,
causing find_library to discover Homebrew's LLD dylibs (whose
LC_ID_DYLIB points to /opt/homebrew/opt/lld/lib/...). Guard this
behind NOT C3_FETCH_LLVM so it only applies when using a
system-installed LLVM.
2. In the static linking block (C3_LINK_DYNAMIC=OFF), find_library NAMES
listed .dylib before .a, so shared libraries were preferred even when
static linking was intended. Reorder to prefer .a, keeping .dylib as
a fallback for systems that only ship shared libraries.
Co-authored-by: Christoffer Lerno <christoffer@aegik.com>
