I have a number of long-term projects that I plan for on long timelines, on the order of decades or more. One of these projects is cozy, a C toolchain. I haven’t talked about this project in public before, so I’ll start by introducing you to the project. The main C toolchains in the “actually usable” category are GNU and LLVM, but I’m satisfied with neither and I want to build my own toolchain. I see no reason why compilers should be deep magic. Here are my goals for cozy:
- Self hosting and written in C
- An easy to grok codebase and internal design
- Focused on C. No built-in support for other languages
- Adding new targets architectures and ports should be straightforward
- Modular build pipeline with lots of opportunities for external integrations
- Trivially cross-compiles without building another version of the toolchain
- Includes a decent optimizer
Some other plans include opinionated warnings about code and minimal support for language extensions. Ambitious goals, right? That’s why this project is on my long-term schedule. I’ve found that large projects are entirely feasible, so long as you (1) start them and (2) keep working on them for a long time. I don’t need to rush this - gcc and clang may not be ideal, but they work today. In support of these goals, I’ll be writing these dev notes to explain my design choices and gather feedback — please email me if you have some!
Since I want to place an emphasis on portability and retargetability, I’m starting by designing the machine spec and its support code, which is used to add support for new architectures. I don’t like gcc’s lisp specs, and I really don’t like LLVM’s “huge pile of C++” approach. I think a really good machine spec meets these goals:
- Easy to write and human friendly
- More about data than code, but
- Easily extended with C to support architecture-specific nuances
- Provides loads of useful metadata about the target architecture
- Exposes information about the speed and side-effects of each instruction
- Can also be used to generate an assembler and disassembler
- Easily reused to create derivative architectures
Adding a new architecture should be a weekend project, and when you’re done the entire toolchain should both support and run on your new architecture. I set out to come up with a new syntax that could potentially meet these goals. I started with the Z80 architecture in mind because it’s simple, I’m intimately familiar with it, and I want cozy to be able to target 8-bit machines just as easily as 32 or 64 bit.
For reference, here are the gcc and LLVM guides on adding new targets:
The cozy machine spec is a cross between ini files, yaml, and a custom syntax. The format is somewhat complex, but once understood is intuitive and flexible. At the top level, it looks like an ini file:
[metadata] # ... [registers] # ... [macros] # ... [instructions] # ...
The metadata section contains some high-level information about the architecture design, and is the simplest section to understand. It currently looks like this for z80:
[metadata] name: z80 bits: 8 endianness: little signedness: twos-complement cache: none pipeline: none
This isn’t comprehensive, and I’ll be adding more metadata as it becomes
necessary. On LLVM, this sort of information is encoded into a string that looks
something like this:
"e-p:16:8:8-i8:8:8-i16:8:8-n8:16". This string is passed
LLVMTargetMachine base constructor in C++. I think we can do a hell
of a lot better than that!
The registers section describes the registers on this architecture.
[registers] BC: 16 B: 8 C: 8; offset=8 DE: 16 D: 8 E: 8; offset=8 HL: 16 H: 8 L: 8; offset=8 SP: 16; stack PC: 16; program
Here we can start to see some interesting syntax and get an idea of the design of cozy machine specs. The contents of each section are keys, which have values, attributes, and children. The format looks like this:
key: value; attributes, ... children...
In this example, we’ve defined the BC, DE, HL, SP, and PC registers. HL, DE, and BC are general purpose 16-bit registers, and each can also be used as two separate 8-bit registers. The attributes for these sub-registers indicates their offsets in the parent register. We also define the stack and program registers, SP and PC, which use the stack and program attributes to indicate their special purposes.
We can also describe CPU flags in this section:
[registers] AF: 16; special A: 8; accumulator F: 8; flags, offset 8;; flag _C: 1 _N: 1; offset 1 _PV: 1; offset 2 _3: 1; offset 3, undocumented _H: 1; offset 4 _5: 1; offset 5, undocumented _Z: 1; offset 6 _S: 1; offset 7
Here we introduce another feature of cozy specs with
F: 8; flags, offset 8;;
;; adds those attributes to all children of this key, so each of
_C, _N, etc have the
Take note of the “undocumented” attribute here. Some of the metadata included in a spec can be applied to cozy tools. Some of it, however, is there for other tools to utilize. We have a good opportunity to make a machine-readable description of the architecture, so I’ve opted to include a lot of extra details in machine specs that third parties could utilize (though there might be a -fno-undocumented compiler flag some day, I guess).
The macros section is heavily tied to the instructions section. Most instruction sets are quite large, and I don’t want to burden spec authors with writing out the entire thing. We can speed up their work by providing macros.
z80 instructions have a few sets of common patterns in their encodings. Register groups are often represented by the same set of bits, and we can make our instruction set specification more concise by taking advantage of this. For example, here’s a macro that we can use for instructions that can use either the BC, DE, HL, or SP registers:
[macros] reg_BCDEHLSP: BC: 00 DE: 01 HL: 10 SP: 11
We have the name of the macro as the top-level key, in this case
We can later refer to this macro with
@reg_BCDEHLSP. Then, we have each of the
cases it can match on, and the binary values these correspond to when encoded in
The instructions section brings everything together and defines the actual instructions available on this architecture. Instructions can be organized into groups at the spec author’s pleasure, which can be referenced by derivative architectures. Here we can take a look at the “load” group:
[instructions] .load: ld: @reg_BCDEHLSP, @imm: 00 $1 0001 $2
On z80, the
ld instruction is similar to the
mov instruction on Intel
architectures. It assigns the second argument to the first. This could be used
to assign registers to each other (e.g.
ld a, b to set A = B), to set
registers to constants, and so on. Our example here uses our macro from earlier
to match instructions like this:
ld hl, 0x1234
The value for this key may reference the arguments with variables. $1 here
10, from the macro. The
imm built-in is implemented in C to match
constants and provides $2. An assembler could use this information to assemble
our example instruction into this machine code:
00100001 00110100 00010010
Which will load HL with the value 0x1234 when executed.
Lots more metadata
Now that we have the basics down, let’s dive into some deeper details. Cozy specs are designed to provide most of the information the entire toolchain needs to support an architecture. The information we have so far could be used to generate assemblers and disassemblers, but I want this file to be able to generate things like optimizers as well. You can add the necessary metadata to each instruction by utilizing attributes.
Consider the z80 instruction LDIR, which stands for “load/decrement/increment/repeat”. This instruction is used for memcpy operations. To use it, you set the HL register to a source address, the DE register to a destination address, and BC to a length. This instruction looks like this in the spec:
ldir: 11101101 10110000; uses[HL, DE, BC], \ affects[HL[+BC], DE[+BC], BC], \ flags[_H:0,_N:0,_PV:0], cycles[16 + BC * 5]
That’s a lot of attributes! The purpose of these attributes are to give the toolchain insights into the registers this instruction uses, its side effects, and how fast it is. These attributes can help us compare the efficiency of different approaches and understand the how the state of registers evolves during a function, which leads to all sorts of useful optimizations.
affects attribute, for example, tells us how each register is affected by
this instruction. We can see that after this instruction, HL and DE will have
had BC added to them, and BC will have been set to 0. We can make all sorts of
optimizations based on this knowledge. Here are some examples:
char *dest, *src; int len = 10; memcpy(dest, src, len); src += len;
The compiler can assign
src to HL,
dest to DE, and
len to BC. We can then
optimize out the final statement entirely because we know that the LDIR
instruction will have already added BC to HL for us.
char *dest, *src; int len = 10; memcpy(dest, src, len); int foobar = 0;
In this case, the register allocator can just assign BC to
foobar and avoid
initializing it because we know it’s already going to be zero. Many other
optimizations are made possible when we are keeping track of the side effects of
I’ve iterated over this spec design for a while now, and I’m pretty happy with
it. I would love to hear your feedback. Assuming that this looks good, my next
step is writing more specs, and a tool that parses and compiles them to C. These
C files are going to be linked into
libcozyspec, which will provide an API to
access all of this metadata from C. It will also include an instruction matcher,
which will be utilized by the next step - writing the assembler.
The assembler is going to take a while, because I don’t want to go the gas route of making a half-baked assembler that’s more useful for compiling the C compiler’s output than for anything else. I want to make an assembler that assembly programmers would want to use.
I have not yet designed an intermediate bytecode for the compiler to use, but one will have to be made. The machine spec will likely change somewhat to accommodate this. Some of the conversion from internal bytecode to target assembly can likely be inferred from metadata, but some will have to be done manually for each architecture.
Here’s the entire z80 spec I’ve been working on, for your reading pleasure.