A BQN implementation in C

CBQN-specific documentation • source code overview

Running

1. make
   • Add FFI=0 if your system doesn't have libffi
   • Use gmake on BSD
   • Add REPLXX=0 if C++ is unavailable (will remove line editing/coloring/name completion in the REPL)
   • Run sudo make install afterwards to install into /usr/local/bin/bqn (a PREFIX=/some/path argument will install to /some/path/bin/bqn); sudo make uninstall to uninstall
   • make clean to get to a clean build state
2. ./BQN somefile.bqn to execute a file, or ./BQN for a REPL

Alternatively, third-party packages (and other ways to run BQN) are listed here; CBQN packaging status:

Configuration options

The default configuration enables REPLXX & Singeli, and, if not done before, implicitly runs make for-build to build a CBQN for running build/src/build.bqn and compiling Singeli.

Builds with more performance

(TL;DR: use make o3n for local builds on x86-64, but make is fine on other architectures)

The default target (make o3) will target optimizations for the current architecture, but not any further extensions the specific CPU may have.

Thus, performance can be significantly improved by targeting the specific CPU via make o3n (with the usual drawback of -march=native of it not producing a binary portable to other CPUs of the same architecture).

On x86-64, a native build will, if available, enable usage of AVX2 (i.e. ability to use 256-bit SIMD vectors instead of 128-bit ones, among other things), and BMI2. But, on aarch64, NEON is always available, so a native build won't give significant benefits.

To produce a binary utilizing AVX2 not specific to any processor, it's possible to do make has=avx2. (has=avx2,bmi2 for targeting both AVX2 & BMI2)

Additionally, on AMD Zen 1 & Zen 2, make o3n has=slow-pdep will further improve certain builtins (Zen 1/2 support BMI2, but their implementation of pdep/pext is so slow that not using it for certain operations is very beneficial).

CBQN currently does not utilize AVX-512 or SVE, or have any SIMD optimizations specific to any architectures other than x86-64 and aarch64.

For native builds, targeted extensions are determined by /proc/cpuinfo (or sysctl machdep.cpu on macOS) and C macros defined as a result of -march=native.

Build flags

• notui=1 - display build progress in a plain-text format
• version=... - specify the version to report in --version (default is commit hash)
• nogit=1 - error if something attempts to use git
• CC=... - choose a different C compiler (default is clang, or cc if unavailable; CBQN is more tuned for clang, but gcc also works)
• CXX=... - choose a different C++ compiler; needed only for REPLXX (default is c++)
• j=8 - override the default parallel job count (default is the output of nproc)
• OUTPUT=path/to/somewhere - change output location; for emcc-o3 it will be the destination folder for BQN.js and BQN.wasm, for everything else - the filename
• target_arch=(x86-64|aarch64|generic) - target architecture; if abscent, inferred from uname, or CC if target_from_cc=1; used for enabling architecture-specific optimizations
• target_os=(linux|bsd|macos|windows) - target OS; if abscent, inferred from uname, or CC if target_from_cc=1; used for determining default output names and slight configuration changes
• target_from_cc=1 - infer the target architecture and OS from C macros that CC defines via -dM -E; additionally infers available extensions, allowing e.g. make f=-march=x86-64-v3 - target_from_cc=1 to optimize assuming AVX2, which would otherwise need has=avx2
• has=... - assume specified architecture extensions/properties (x86-64-only); takes a comma-separated list which, beyond what is architecturally guaranteed, infer additional extensions as noted which hold on existing hardware (at least as of the time of writing):
  • pclmul (implies SSE4.2)
  • avx2 (implies pclmul, POPCNT, BMI1)
  • bmi2 (implies avx2)
  • slow-pdep (implies bmi2; specifies Zen 1 & Zen 2's slow pdep/pext)

• REPLXX=0 - disable REPLXX
• singeli=0 - disable usage of Singeli
• FFI=0 - disable •FFI, thus not depending on libffi
• usz=32 - use 32-bit integers for array lengths (default is 64-bit)

• f=... - add extra C compiler flags for CBQN file compilation
• lf=... - add extra linking flags (LDFLAGS is a synonym)
• CCFLAGS=... - add flags for all CC/CXX/linking invocations
• REPLXX_FLAGS=... - override replxx build flags (default is -std=c++11 -Os)
• CXXFLAGS=... - add additional CXX flags

Alternatively, build/build (aka build.bqn) can be invoked manually, though note that it has slightly different argument naming (see build/build --help) and doesn't have predefined build types (i.e. make o3ng is done as build/build replxx singeli native g)

More build types

• make o3 - the default build
• make o3g - effectively make o3 f=-g (custom f=... can still be added on)
• make c - make o3 but without -O3
• make shared-o3 - produce a shared library libcbqn.so/libcbqn.dylib/cbqn.dll
• make shared-c - like make c but for a shared library
• make emcc-o3 - build with Emscripten emcc
• make wasi-o3 - build targeting WASI
• make wasi-reactor-o3 - build producing a WASI Reactor
• make debug - unoptimized build with extra assertion checks (also includes -g)
• make static-bin - build a statically linked executable (for a fully standalone binary, try make static-bin CC=musl-gcc REPLXX=0)
• make static-lib - build a static library archive

All of the above (and plain make) will go through one bootstrap build for running build/src/build.bqn.
make o3-makeonly and make c-makeonly are available for a very stripped-down makefile-only build, though those are only intended for debugging, and are missing features and configurability.

Limitations

• While 64-bit values are used for array sizes (unless configured otherwise), some builtins don't have implementations for input/output arrays with over 2⋆31 elements, and will result in an error or possibly give wrong results. As these cases are difficult to test, crashes or memory corruption are also possible. When implemented they may also perform disproportionately worse.

• Throwing & catching errors will permanently leak memory; the garbage collector has no way to scan for objects on the C stack, and thus checks for objects with an incorrect reference count (comparing to the expected from other heap objects / GC roots) to infer objects to preserve. Error throwing is implemented via a longjmp, thus making the GC permanently think those are still on the C stack. A full GC can run between REPL lines (as then there's a guarantee of no BQN code being mid-execution), but not in any other context.

• Highly nested objects will result in crashes: object freeing is done recursively, and there is nothing preventing arbitrarily-nested objects, resulting in stack overflows.

Requirements

CBQN requires:

• either gcc (≥ 9) or clang (≥ 10) as the C compiler (by default it attempts clang as things are primarily optimized for clang, but, if unavailable, it'll fall back to cc; override with CC=your-cc);
• libffi (≥ 3.3) for •FFI;
• for replxx, a C++ compiler (≥C++11; defaults to c++, override with CXX=your-c++).

Though other configurations are expected to work (except 32-bit x86 without SSE2¹), here are some that CBQN is tested on by dzaima:

x86-64 (Linux):
  gcc 9.5; gcc 14.2.0; clang 10.0.0; clang 22; clang 23 prerelease (++20260601033445+ef7cbc5f9c40-1~exp1~20260601153641.3163)
  libffi 3.4.6
  CPU microarchitecture: Haswell
  replxx: g++ 14.0.1; clang++ 22
x86 with SSE2 (Linux; running on the above x86-64 system, compiled with `make target_from_cc=1 REPLXX=0 FFI=0 CCFLAGS='-m32 -msse2' usz=32`):
  clang 21.0.0; gcc 14.2.0
AArch64 ARMv8-A (within Termux, Android 16):
  clang 21.1.8
  libffi 3.4.7
  replxx: clang++ 21.1.8
  May need `lf=-landroid-spawn` & `pkg install libandroid-spawn` on old Android versions

Additionally, CBQN is known to compile as-is on macOS. Windows builds can be made by cross-compilation (see Docker setup).

The build will attempt to use pkg-config to find libffi, uname to determine target_arch & target_os if not using target_from_cc, and nproc for parallel job count, with defaults if unavailable (-lffi for linking libffi (+ -ldl on non-BSD), target_arch=generic, target_os=linux, j=4; these can of course also be specified manually).

Git submodules are used for Singeli, replxx, and precompiled bytecode. To avoid automatic usage of git here, link local copies to build/singeliLocal, build/replxxLocal, and build/bytecodeLocal.

Furthermore, git is used to determine the version that --version should display (override with version=...). Use nogit=1 to disallow automatic git usage.

Precompiled bytecode

CBQN uses the self-hosted BQN compiler & some parts of the runtime, and therefore needs to be bootstrapped. By default, the CBQN will use precompiled bytecode.

In order to build everything from source, you can:

option 1: use another BQN implementation

1. dzaima/BQN is a BQN implementation that is completely implemented in Java (clone it & run ./build)
2. clone mlochbaum/BQN
3. mkdir -p build/bytecodeLocal/gen
4. other-bqn-impl ./build/genRuntime path/to/mlochbaum/BQN build/bytecodeLocal
   In the case of the Java impl, java -jar path/to/dzaima/BQN/BQN.jar ./build/genRuntime path/to/mlochbaum/BQN build/bytecodeLocal

option 2: use the bootstrap compilers

1. clone mlochbaum/BQN
2. mkdir -p build/bytecodeLocal/gen && make for-bootstrap && ./BQN build/bootstrap.bqn path/to/mlochbaum/BQN

Note that, after either of those, the compiled bytecode may become desynchronized if you later update CBQN without also rebuilding the bytecode. Usage of the submodule can be restored by removing build/bytecodeLocal.

Cross-compilation

You must manually set up a cross-compilation environment. It's possible to pass flags to all CC/CXX/linking invocations via CCFLAGS=..., and LDFLAGS=... to pass ones to the linking step specifically (more configuration options above).

A target_arch=(x86-64|aarch64|generic) make argument should be added (generic will work always, but a more specific argument will enable significant optimizations), as otherwise it'll choose based on uname. Similarly, target_os=(linux|bsd|macos|windows) should be present if the target OS differs from the host.

Alternatively, the target_from_cc=1 make argument can be used, replacing the need of target_arch and target_os (although they can be still set, overriding the values inferred from CC).

Furthermore, all build targets (except -makeonly ones) will need a non-cross-compiled version of CBQN at build time to run build.bqn and Singeli. For those, a make for-build will need to be ran before the primary build, configured to not cross-compile. (this step only needs a C compiler (default is CC=cc here), and doesn't need libffi, nor a C++ compiler).

Licensing

First, some exceptions to the general licensing:

• timsort implementation - src/builtins/sortTemplate.h: MIT; original repo
• Ryu - src/utils/ryu.c & files in src/utils/ryu/: Apache 2.0 or Boost 1.0; original repo

Additionally, REPLXX (optional, included by default) has its own licensing at build/replxxSubmodule/LICENSE.md.

Everything else (i.e. all files except src/builtins/sortTemplate.h, src/utils/ryu.c, and everything in src/utils/ryu/) may be treated as under any of the following licenses, as if they had the respective notice:

• GNU LGPLv3 only

Copyright (C) 2021-2025 dzaima and contributors

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU Lesser General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU Lesser General Public License for more details.

You should have received a copy of the GNU Lesser General Public License
along with this program.  If not, see <https://www.gnu.org/licenses/>.

• GNU GPLv3 only

Copyright (C) 2021-2025 dzaima and contributors

This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.

This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with this program.  If not, see <https://www.gnu.org/licenses/>.

• MPL 2.0

This Source Code Form is subject to the terms of the Mozilla Public
License, v. 2.0. If a copy of the MPL was not distributed with this
file, You can obtain one at http://mozilla.org/MPL/2.0/.

Footnotes

1. On x86, without SSE2, only gcc≥15 is known to avoid miscompiling CBQN; older gcc (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58416), and all tested clang versions, result in immediate crashes in optimized builds. Even when not miscompiled, such builds may produce unexpected floating-point results due to its non-standard extended-precision format. Enabling SSE2 via CCFLAGS='-msse2' resolves these problems, though of course won't run on hardware without support for it (...which is over 25 years old) ↩