Building it

To build, you’ll need a Unixish machine (I develop on Linux) with some dependencies.

  • the Ninja build tool

  • Lua 5.1 (needed for the build)

  • the Pasmo Z80 assembler (needed to build part of the CP/M emulator)

  • the 64tass 6502 assembler (needed to build the 6502 code)

  • the libz80ex Z80 emulation library (needed for the CP/M emulator)

  • flex and bison and libbsd and libreadline (these are standard)

  • a C compiler and these GNU binutils packages:

    • i686-linux-gnu
    • arm-linux-gnueabihf
    • m68k-linux-gnu
    • m68k-atari-mint Packages are available from
    • powerpc-linux-gnu
  • the nasm ix86 assembler

  • the qemu userspace emulator

  • the gpp preprocessor

If you’re on a Debianish platform, you should be able to install them (after adding any custom package sources as above) with:

apt install ninja-build lua5.1 pasmo libz80ex-dev flex libbsd-dev libreadline-dev bison binutils-arm-linux-gnueabihf binutils-i686-linux-gnu binutils-powerpc-linux-gnu binutils-m68k-linux-gnu binutils-m68k-atari-mint qemu-user gpp 64tass nasm

Once done you can build the compiler itself with:


You’ll be left with a lot of stuff in the bin directory. The tools are all labeled as (name).(toolchain).(extension); however, several extensions also contain a dot. So, is cowfe, the main front-end compiler, targeting the 65c02, built with the ncpmz toolchain, which produced a executable.


The build process tries to build all the combinations of source (the toolchain used to build the compiler) and target (the machine the compiler is compiling for).

Cowgol defines these toolchains:

  • bootstrap: this is the compiler shipped in C with the distribution. It’s only used to build the first stage compiler.

  • ncgen: targeting C, built with the bootstrap compiler.

  • nncgen: targeting C, built with ncgen.

  • lx386: targeting Linux 80386 binaries, built with nncgen.

  • cpm: targeting CP/M 8080 binaries, built with nncgen.

  • cpmz: targeting CP/M Z80 binaries, built with nncgen.

  • bbct: targeting BBC Tube 65c02 binaries, built with nncgen.

  • bbcti: targeting BBC Tube 6502 bytecode binaries, built with nncgen.

  • bbct6502: targeting BBC Tube 6502 binaries, built with nncgen.

ncgen and nncgen should behave identically. The compiler is built with itself to make sure that nncgen was built with a compiler built from the current compiler source, which is invaluable for testing. On a PC when you’re cross-compiling you’ll most likely want to be using the nncgen binaries (with the nncgen.exe extension).

To run the cross compiler to generate a Linux 80386 binary, do:

$ bin/cowfe-80386.nncgen.exe -Irt/ -Irt/lx386/ examples/helloworld.cow helloworld.cob
$ bin/cowbe-80386.nncgen.exe helloworld.cob helloworld.coo
$ bin/cowlink-lx386.nncgen.exe .obj/rt/lx386/cowgol.coo helloworld.coo -o helloworld.s
$ i686-linux-gnu-as helloworld.s -o helloworld.o
$ i686-linux-gnu-ld helloworld.o -o helloworld
$ ./helloworld
Hello, world!

If you’re on a system which can run Linux i686 binaries, this will work too:

$ bin/cowfe-80386.lx386.lx386.exe -Irt/ -Irt/lx386/ examples/helloworld.cow helloworld.cob
$ bin/cowbe-80386.lx386.lx386.exe helloworld.cob helloworld.coo
$ bin/cowlink-lx386.lx386.lx386.exe .obj/rt/lx386/cowgol.coo helloworld.coo -o helloworld.s
$ i686-linux-gnu-as helloworld.s -o helloworld.o
$ i686-linux-gnu-ld helloworld.o -o helloworld
$ ./helloworld
Hello, world!

cowfe and cowbe

The compiler is split into two binaries, for the front end (cowfe) and the back end (cowbe). cowfe parses the Cowgol language, does type checking, generates the AST, determines variable and record sizes, etc. It emits a temporary file, the .cob file. cowbe then takes this temporary file and generates actual machine code, emitting a .coo file.

The .cob file is an implementation detail of the compiler and can be completely ignored. Don’t try to use mismatched cowfe and cowbe binaries — weird things will happen. If you’re interested, you can see the contents using the cowbdmp tool, but that’s primarily intended for my debugging, so it’s not very user friendly.

The general build process is:

  • cowfe and cowbe compile .cow files into .coo files.

  • cowwrap compiles .cos ‘assembly’ files into .coo files.

  • cowlink links together multiple .coo files, performs the global optimisation, and emits an ‘assembly’ file.

  • the ‘assembly’ file is ‘assembled’ with your native toolchain into an executable.

‘Assembly’ is in quotation marks because it’s not necessarily actual assembly. The cgen target actually emits C. This is treated just like any other assembly language, with cowlink emitting a (hideous and very large) C file, which is then compiled with, typically, gcc.

Cowlink is the tool which scans the .coo files, determines which subroutines are used, places their variable storage in memory, and converts them into a plain text file.

.coo files themselves are a chunked format containing both ‘assembly’ source and binary markup. They’re not considered readable or writeable by humans, although at a pinch you can look at them in a text editor. Because they’re not writeable, cowwrap is a tool for converting annotated text files into .coo files. If you look in rt/*/cowgol.cos you can see the runtime library for the various platforms. The format’s not documented yet, sorry.

Previous page Next page