ZNCC

Zak's New(-ish) C(-ish) Compiler.

(Alternatively, ZNCC is a recursive acronym for ZNCC is Not a C Compiler.)

Introduction

This is the latest version of my C-like compiler, which is distantly based on LICE: https://github.com/dorktype/LICE

This is NOT meant to be a full replacement for standards-compliant C/C++ compilers. It's rather intended for bootstrapping new architectures and for use in minimalist high-level computing systems (e.g. where code needs to be strictly audited for security reasons or needs to be kept simple for educational purposes, but still needs some complex/modern functionality).

The compiler mostly targets modern, 64-bit platforms and focuses on converting simple pre-processed C-like code into simple assembler code for the target architecture. It can be easily adapted for other use-cases or bundled with additional tools.

Major Features

• Supports multiple architectures and is easy to retarget
  • Mostly tested on x86-64/AMD64 Linux
  • Partial support for Windows ABI
  • RV64 is about half-supported (basic tests work but there are many broken bits), with minimal support for RV32 (enough for "Hello world")
  • Linux targets should also work for FreeBSD/OpenBSD/Solaris/etc. with minimal re-tooling (macOS may require a little more work, but not much)
  • Also includes some minimal/experimental support for other/new architectures
• Single compiler tool supports multiple targets (no need for per-architecture compiler builds)
• Compiler is mostly self-hosting (at least on fully-supported targets)
• Supports most (not all) essential C features with some extensions
  • Includes some Objective C-like OOP extensions
  • Basic floating-point support is included (assuming the target has such support)

NOTE: More detailed language & design features are covered in their own sections.

Building & Usage

Command-Line Build (Linux/Unix/bash/...)

Using your default host C compiler (Unix-like best practices):

cc -ozncc zncc.c

Or specifically using GCC or similar:

gcc -ozncc zncc.c

Visual Studio IDE Build (Windows)

1. Create a new C++ Command Line project
2. Copy/paste the zncc.c code into your main .c++ file
3. Rename your main .c++ file to zncc.c (or something else with a .c ending)
4. Create new header files, zncp.h and zncg.h, copying/pasting the associated code in
5. Build/run and enjoy!

NOTE: This trick generally works for getting simple C programs working in Visual Studio, which doesn't seem to be configured well for C programming by default.

Basic Usage

The compiler takes as input C-like program code (without any preprocessor directives) and generates as output assembler code.

For easy testing with files, the --input and --output arguments can be given:

./zncc --input mycode.c --output mycode.s

The assembler code then needs to be assembled/linked (see beneath).

Extended Usage/Testing/Self-Hosting (Linux/Unix Only)

First build the compiler as above. Then download the ZNLC headers, you can place these anywhere convenient (e.g.including in the compiler directory).

Then, create a preprocessed version using GCC's frontend or another preprocessor (NOTE: This differs a little between platforms):

gcc -E -Ipath/to/ZNLC/include -D_ZCC -D_ZCC_X64 zncc.c > zncc.X64.c

If this succeeds, zncc.X64.c should be the raw, preprocessed C code for the appropriate target. The next step is to run the compiler, producing assembly code:

./zncc --input zncc.X64.c --output zncc.X64.s

Then you can assemble & link, again GCC's frontend comes in handy on Linux:

gcc -static -ozncc.X64 zncc.X64.s

This will produce the self-hosted version zncc.X64, which you can test by compiling itself as above:

./zncc.X64 --input zncc.X64.c --output zncc.X64.again.s
gcc -static -ozncc.X64.again zncc.X64.again.s

Target Options

The default settings for now reflect the testing environment (future/integrated versions may detect settings a little better).

More-specific options can be relayed through environment variables:

• The value of CCB_FAMILY controls the target architecture:
  • x86 or X86 for commonplace Intel/AMD processors used in most PCs/laptops (currently only supported in 64-bit mode)
  • risc-v/RISC-V/riscv/RISCV for RISC-V (RV32/RV64-based) or compatible targets
  • arm or ARM for ARM-based targets (currently mostly unimplemented)
  • Potentially other/experimental settings
• The value of CCB_WORDSIZE specifies the basic word-size of the target processor:
  • Only the value of 64 fully works at the moment (to target 64-bit PCs and RV64)
  • The value of 32 can be used to test the RV32 target (which is less complete than the 64-bit modes)
  • The value of 16 is also recognised but there are no 16-bit targets at this stage
• The value of CCB_CALLCONV controls the default calling conventions:
  • standard or STANDARD generally implies the "System V" or similar conventions used by Linux/BSD/Solaris systems
  • windows or WINDOWS specifies Microsoft Windows (or ReactOS/WINE) conventions
  • Note that there is some support for specifying calling conventions on a function-by-function basis, but this isn't fully fleshed-out
• The value of CCB_ASMFMT controls the assembler format:
  • gas is often most useful on Linux/similar systems, and conforms to GNU/GCC's default assembler syntax
  • fasm generates code for Flat Assembler which works on x86 systems: https://flatassembler.net/ (this may also be useful for porting to NASM & other targets)
  • raw uses a simplified syntax, i.e. for testing new targets without good/standard assemblers (this is mostly useless for PC & RISC-V targets for now)
• The value of CCB_BINFMT controls the binary format or linker semantics assumed in the assembler code:
  • elf is generally the default on modern Linux/BSD/Solaris systems, and is ideal for linking with GCC/clang code on those platforms
  • flat can be used for producing small "flat binary" code snippets, particularly with Flat Assembler

NOTE: These names reflect internal naming (CCB being short for "C-like Compiler Backend") will be updated before the final release.

Language Features & Limitations

The compiler generally accepts C-like code with some (experimental/incomplete) Objective C-style extensions. It would roughly be on-par with a pre-standard C compiler, except for modern targets.

This means that features like integers, functions, structs, arrays, pointers, etc. generally work as per usual, but there are some C features which are not implemented in ZNCC:

• Unpacking "vararg" parameters will not work
  • Varargs can be declared/called but not unpacked
  • This basically means you need to use a different/standard compiler to build any "printf"-like functions (but you can still access them)
• Passing structures as arguments or return values of functions will not work
  • In other words, arguments are expected to be either integer/pointer-sized or floating-point values
  • This may need to be revised in order to target 32-bit platforms properly (which may need to pass around 64-bit integers)
• String constants with the same text are not guaranteed to be == at runtime
  • This kind of stuff can generally be done and may sometimes be automatic, but depends on linker features
  • It's generally considered bad practice to rely on this anyway, but may be worth noting
• Literals are limited to common forms (e.g. don't expect it to support wide strings or pointers to struct literals)
• There will probably never be any support for bitfields
• There is currently no support for C++ style classes/namespaces/templates/..
  • Some minimal support may be added in the future, but likely not the whole lot
  • The Objective C-like features will be the main focus for OOP-like extensions in the short term

Bonus Bugs & Temporary Limitations

• Floating-point support exists and should generally "work" on supported targets, but is minimal
  • This can probably be expanded quite easily in future versions, but will eventually require some platform-specific options
• Large numbers of function arguments will partly work, but not reliably
  • Large numbers of integer/pointer arguments will "work", but may misbehave when combined with floating-point arguments (and may throw off the stack alignment for floating point inside the function)
  • This is just an incomplete feature (i.e. missing float support), and should be easy to fix incrementally
• Number conversions and precise signed/unsigned/etc. semantics have not been thoroughly fuzz-tested
  • There are likely to be some issues with specific types or combinations, but these are usually easy to fix once identified
• Pointer arithmetic exists, but is limited
  • For now, this means: Use simple addition/subtraction for pointers, don't rely on increment operators or inner pointer syntax working correctly to obtain offsets
  • These issues can probably be solved incrementally (if not, warnings/errors can be added to catch broken cases)
• Error reporting is present, but is not ideal for catching bugs/suggesting solutions
• There is no/minimal optimisation (this is intentional, the focus is on making the thing work first)
• The object-oriented extensions have potential, but deciding on an exact ABI is difficult
  • Efficient implementations would easily become platform-specific, while unoptimised implementations may be impractical
  • This means that there are some OOP features, but they need to be tweaked and integrate them into some kind of platform to be useful

Other Limitations/Design Features

There is no built-in preprocessor or associated tools. I recently began integrating such tooling, but ran into a couple of issues.

Most notably, any generally-useful preproccessor ends up more complex than the compiler, so self-hosting can become troublesome.

Secondly, the compiler itself (and LICE which was used as a foundation) are licensed under the terms of Unlicense, while other/third-party components usually imply other (or more-vague) licensing conditions.

And thirdly, the compiler part is more-or-less essential for all use cases. The preprocessor, linker, assembler, and so on are more naturally specialised for particular platforms or combinations of technologies. For example a regular desktop target might require a full preprocessor and other tools, whereas an embedded target may only need to recompile straightforward programs and may use the compiler directly with a special frontend & assembler limited to that target.

See Also

• https://github.com/ZYSF/ZNLC "fake libc" headers for self-hosting builds (to ensure the compiler doesn't choke on any obscure system headers)