- comparisons of vectors give masks (boolxN): select(m, a, b), m.any/all/count, & | ^ !;
f32x8(list, i, fill) and v.store_part(list, i) for the end of a list (masked load/store)
- v.shuffle(constant lanes), v.lo(), v.hi(), concat(a, b), table.lookup(indices):
pshufb / vpshufb on x86 (SSSE3 and up), tbl on ARM64 (checked in the assembly, LLVM 18 and 21)
- acc.dot_add(a, b): i32 lanes gain 4 byte products each, exactly; vpdpbusd with AVX-VNNI or AVX-512 VNNI,
usdot/sdot/udot on ARM64, the sum written out elsewhere (pmaddwd); run here with all three paths agreeing
- @cpu("avx2,fma") fn: LLVM target-features for one function (+evex512 for AVX-512, which LLVM 18+ needs);
cpu_has("avx2"): CPUID + XGETBV on x86, hwcaps on Linux ARM64, sysctl on macOS; a wide vector can't be passed
between functions with different @cpu (its registers differ), and the compiler says so
- the compiler asks clang which features --cpu means (its predefined macros, cached), so dot_add picks the instruction
- Nix's clang wrapper dropped -march=native in nix-shell (reproduced): finch tells it NIX_ENFORCE_NO_NATIVE=0, shell.nix too,
and warns if --cpu native still made no difference
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR
- a * b + c is two roundings on every processor and --cpu: the IR has no fast-math flags, and clang
compiles it with -ffp-contract=on (checked in the assembly: never fused, also with FMA hardware)
- fma(a, b, c) / v.fma(b, c): one rounding, always (llvm.fma)
- muladd(a, b, c) / v.muladd(b, c): the fastest way here (llvm.fmuladd), what v.fma was before
- f32_bits / f32_from_bits / f64_bits / f64_from_bits: bitcasts, also of vectors
- f16_to_f32 / f32_to_f16 (LLVM half), bf16_to_f32 / f32_to_bf16 (integer ops, ties to even, NaN kept)
- docs (en, pl): what is guaranteed about floats, the new built-ins
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR
- CI: macOS on Apple Silicon and Intel (macos-15, macos-15-intel); a tag v* attaches every package
(Linux x64/ARM64, macOS ARM64/x64, Windows x64, the VS Code extension) to its release
- vectors: fma is llvm.fmuladd (llvm.fma was a libm call per lane without FMA hardware: 3.5x slower)
- examples/simd.fch: a dot product eight at a time, timed against the plain loop
- README and the English guides: installing on Linux, macOS and Windows; vectors (SIMD); --cpu;
-g is DWARF on Windows too (lldb from llvm-mingw); no more mention of the C++ compiler
- comments left from the C++ compiler (src/target.cpp, src/codegen.cpp, "never frees") brought up to date
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR
- fNxM / iNxM / uNxM up to 512 bits, as LLVM vectors (<8 x float>): LLVM picks SSE, AVX2,
AVX-512 or NEON for them
- f32x8(x) in every lane, f32x8(a, ..., h), f32x8(list, i) loads, f32x8(i32x8 v) converts
- + - * / and the bit operators lane by lane, with a plain number on either side counting for every lane;
v[i], v[i] = x, v.len, == / != (every lane), print and str
- methods: sum min max dot abs sqrt floor ceil round fma store; min/max/abs/clamp lane by lane
- wide vectors (and structs holding them) are loaded and stored with align 16, what malloc and the stack give
- --cpu <name> (and cpu = "..." in finch.toml): native, x86-64-v3, apple-m1, ...
- debug info shows vectors; the language server completes their methods; Kate and VS Code highlight them
- f32 x = -1.5 works: a minus in front of a float written in the code keeps it such a number
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Psc42moPLwP43Cmb9oZqSR
- on stderr (finch ir stays the IR), as yellow warnings in editors (@W in the index), once per place
- str, arrays and maps inside a struct don't count: passing it copies only their handles
- none in the compiler, the examples or the tests other than the new one
- docs: the warning instead of "not planned"; what was and wasn't checked for Android
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016GzNP6r8dg6WDxUVZy8wjL
- a static inline function in a header is called through a pointer a small C file keeps
(void *finch_inl_f = &f), built with the program's include folders and cached
- --shared on macOS links with -undefined dynamic_lookup, as a .so does on Linux
- "unable to find library -lx" (lld, MinGW) gives the library-not-found advice
- the language server test shows the index records when alias completion fails
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016GzNP6r8dg6WDxUVZy8wjL
- finch.toml: name, entry, paths, include, links, libdirs, cflags, ldflags, and [target.windows] etc.;
finch run / build / test with no file use its entry
- import thread: spawn, join, mutex, cpus; atomic_add/sub/load/store/swap/cas built in
- finch fmt [--check]: indentation by { }, trailing spaces, empty lines; also as LSP formatting
- link "x.cpp" (or in finch.toml): built with the C++ compiler, which then links (FINCH_CXX)
- #define RAYWHITE (Color){ 245, 245, 245, 255 } imports as a Color
- push(body) where push takes ptr[Body]: the same as push(addr(body))
- --target android (NDK clang from ANDROID_NDK_HOME, -llog)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016GzNP6r8dg6WDxUVZy8wjL
- "x is \(x)", "\(speed, 2)" (two digits after the point), also in """ text
- test "name" { ... } blocks; assert(cond) and assert(cond, why); finch test runs each on its own
- the program's object and its C files (link "x.c") are kept in the cache until they or their headers change
- runtime errors show the calls that led there (finch run, finch test, -g, -O0, --trace)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016GzNP6r8dg6WDxUVZy8wjL
- const NAME = value (numbers, text, other constants; typed: const f32 G = 9.8), also module.NAME
- x := if c { a } else { b }
- """ ... """: the closing quotes' indentation goes from every line
- min, max, abs, clamp for any numbers, unless your own function has the name
- import ui/button (a module in a folder), import path as p
- fs: read_bytes, write_bytes, exists, is_dir, mkdir (with parents), list, remove_dir, data_dir(app)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016GzNP6r8dg6WDxUVZy8wjL
Also frees main's args and the message of a failing main (valgrind was not clean on two tests).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Windows: the compiler made from bootstrap/windows-x86_64.ll runs under Wine, compiles itself to the same IR, builds an
exe of itself, and every test program passes through it. It needed: Windows-sized unsigned long for libclang's unsaved
file, no 'cross' when Windows builds for Windows, --target passed to libclang (libclang.dll assumes MSVC), the MinGW
triple always in the IR, FINCH_LDFLAGS, .exe names and FINCH_LLVM_INCLUDE in bootstrap/build.sh, winseed.sh/winbuild.sh.
macOS: seeds for arm64 and x86-64 made with the 11.3 SDK; IR is accepted by clang, nothing has run on a Mac yet.
CI gets a windows and a macos job. UNTESTED on real Windows/macOS runners.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015f73FesGZGYPkmDvUcj6QX
enum Shape { Circle(float), Rect(float, float), Point }; Shape.Circle(2.0) makes one, match names the values
(Circle(r), _ to ignore), exhaustive like plain enums, printing shows Rect(2, 3.5). Underneath it is a struct (tag +
every variant's values), so copying, freeing, methods and valgrind-cleanliness come with it. Also: stray note files removed.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015f73FesGZGYPkmDvUcj6QX
gen.fch never dropped anything (it was the bootstrap compiler); now it does what the C++ one did: owned variables are
dropped when their scope ends (defers interleaved, newest first), return moves a local out, assignments drop the old
value, temporaries nobody keeps are released, values an expression holds are dropped when try/or leave early,
clear/resize/remove/map.clear drop what goes, free() drops the pointee. The compiler itself now frees as it goes.
MEMCHECK=1 tests/run.sh: 108 passed under valgrind with --leak-check=full.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015f73FesGZGYPkmDvUcj6QX
Hover, go to definition, completion (after a dot too: fields, methods, a module's names, an enum's variants,
array/text/map methods), signature help, the outline and errors while typing -- everything the C++ one did,
and the same tests/lsp_test.py passes against it.
It doesn't type-check by itself: for every change it runs finch index on the editor's text (--overlay says
which file the text stands in for) and reads what the compiler learned about every name, which is also what
the compiler's own error messages are made of. Half-typed code keeps using the last version that compiled.
src/docs.fch is the one description of the built-ins that hover and completion share.
tests/run.sh runs the LSP test too. 107 passed.
boot/ is src/ now. The C++ compiler and CMake are gone from this branch (they are on before-bootstrap).
bootstrap/build.sh builds the compiler from bootstrap/<system>.ll -- the compiler's own LLVM IR -- then builds
it with itself and checks the two come out identical. bootstrap/build.sh seed writes the seed anew.
The driver is complete: the runtime is compiled once and kept in ~/.cache/finch; a program's own C files,
.o and .a files and frameworks link along; a linker failure is read and answered with link "lib" or the
library that is missing. The old tests/fail diagnostics are all in again (a missing import, a str compared with
null, 'write just the library's name', a map key that is a float, non-ASCII names, ...).
tests/run.sh is the one test runner: programs, refusals, -g builds and the shared library. 106 passed.
