If this implementation had existed in the 1980s, the C standard would have a rule that different tokens hashing to the same 16-bit value invoke undefined behavior, and optimizing compilers in the 2000s would simply optimize such tokens away to a no-op. ;)
Oh, it looks like my X86-16 boot sector C compiler that I made recently [1]. Writing boot sector games has a nostalgic magic to it, when programming was actually fun and showed off your skills. It's a shame that the AI era has terribly devalued these projects.
Er, what? The article describes a compiler for a not-quite-C programming language which fits entirely in 512B. Your project, if I see this correctly, can optionally produce code meant to execute as boot sector.
Both interesting projects, but other than the words 'boot sector', 'C' and 'compiler', I don't see a similarity.
An interesting use case - for the compiler as-is or for the essentiall idea of barely-C - might be in bootstrapping chains, i.e. starting from tiny platform-specific binaries one could verify the disassembly of, and gradually building more complex tools, interpreters, and compiler, so that eventually you get to something like a version of GCC and can then build an entire OS distribution.
This is very nice. I'm currently writing a minimalist C compiler although my goal isn't fitting in a boot sector, it's more targeted at 8-bit systems with a lot more room than that.
This is a great demonstration of how simple the bare bones of C are, which I think is one reason I and many others find it so appealing despite how Spartan it is. C really evolved from B which was a demake of Fortran, if Ken Thompson is to be trusted.
Would and how much would it shrink when if, while, and for were replaced by the simple goto routine? (after all, in assembly there is only jmp and no other fancy jump instruction (I assume) ).
And PS, it's "chose your own adventure". :-)
I love minimalism.
What fancy jumps are present in assembly depends on the CPU architecture. But there are always conditional jumps, like JNZ that jumps if the Zero flag isn't set.
It only does if & while, not for. A goto in a single-pass thing would need separate handling for forwards vs backwards jumps, which involves keeping track of data per name (in a form where you can tell when it's not yet set; whereas if/while data is freely held in recursion stack). And you'd still need to handle at least `if ( expr ) goto foo;` to do any conditionals at all.
The “fancy jump” is the branch instruction. As far as I know all ISAs have them. Even rv32i which is famously minimal has several branch instructions in addition to two forms of unconditional jump. Branches are typically used to construct if / for / while as well as && and || (because of short circuiting) and ternary (although some architectures may have special instructions for that that may or may not be faster than branches depending on the exact model). Without it you would have to use computed goto with a destination address computed without conditional execution using constant time techniques.
This is the kind of project that reminds you how far removed modern development is from the actual machine. We pile abstractions on abstractions until "Hello World" needs 200MB of node_modules, and then someone fits a C compiler in 512 bytes.
Not saying we should all write boot sector code, but reading through projects like this is genuinely humbling. Great educational resource too.
It's a fun comparison, but with the notable difference that that one can compile the Linux kernel and generate code for multiple different architectures, while this one can only compile a small proportion of valid C. It's a great project, but it's not so much a C compiler, as a compiler for a subset of C that allows all programs this compiler can compile to also be compiled by an actual C compiler, but not vice versa.
And it doesn't for the compiler in question either. As long as the headers exist in the places it looks for them. No compiler magically knows where the headers are if you haven't placed them in the right location
stddef.h (et al) should be shipped by the compiler itself, and so it should know where it is. But they rely on gcc for it, hence it doesn't always know where to look. Seems totally fine for a prototype.
Shipping GPL headers that explicitly state that they are part of GCC with a creative commons licensed compiler would probably make a lot of people rather unhappy, possibly even lawyers.
I've certainly encountered clang & gcc not finding or just not having header files a good couple times. Mostly around cross-compilation, but there was a period of time for which clang++ just completely failed to find any C++ headers on my system.
So we're down to a missing or unclear description of a dependency in a README - note following the instructions worked for others -, from implications the compiler didn't work.
I mean we know it can be done in little space, given the many tiny C compilers. I think what is most interesting about this one is exactly the creative shortcuts. It's an interesting design space for e.g. bootstrapping to impose extra restrictions.
I actually "shipped" a parser using the symbols' hash(as the only identifier) for a test tool once. Hopefully, the users never used enough symbols to collide 32-bits.
That would be true of one using a libc, but in a boot sector, you only have the bios, so the atoi being referenced is the one defined in c near the beginning of the article
For me is not interesting because it fits in 512 bytes, it's interesting because it's very simple. I think it would be a great introduction to learning about compilers.
I've had enough of headlines that overpromise and underdeliver. It's essentially false advertising. It's not like the word "subset" would put it over the length limit.
Nice, now you can dd it to your boot sector and ... Wait, it is 2026, there are 1000 ways of booting and memory mapping on so-called unified ARM architecture @,@
If this implementation had existed in the 1980s, the C standard would have a rule that different tokens hashing to the same 16-bit value invoke undefined behavior, and optimizing compilers in the 2000s would simply optimize such tokens away to a no-op. ;)
"you don't have -wTokenHashCollision enabled! it's your own foolish ignorance that triggered UB; the spec is perfectly clear!"
Hey stop it with the ad hominems!
Too real! LMAO
Oh, it looks like my X86-16 boot sector C compiler that I made recently [1]. Writing boot sector games has a nostalgic magic to it, when programming was actually fun and showed off your skills. It's a shame that the AI era has terribly devalued these projects.
[1] https://github.com/Mati365/ts-c-compiler
Er, what? The article describes a compiler for a not-quite-C programming language which fits entirely in 512B. Your project, if I see this correctly, can optionally produce code meant to execute as boot sector.
Both interesting projects, but other than the words 'boot sector', 'C' and 'compiler', I don't see a similarity.
> when programming was actually fun and showed off your skills
Oh no. Now more people are able to do what I do. I'm not special anymore.
Seems like this is facetious but to me, “I’m not special” is a pretty valid thing to be sad about.
The two dos in "do what I do" do absolutely not carry the same meaning.
I may be the author.. enjoy! It was an absolute blast making this!
An interesting use case - for the compiler as-is or for the essentiall idea of barely-C - might be in bootstrapping chains, i.e. starting from tiny platform-specific binaries one could verify the disassembly of, and gradually building more complex tools, interpreters, and compiler, so that eventually you get to something like a version of GCC and can then build an entire OS distribution.
Examples:
https://github.com/cosinusoidally/mishmashvm/
and https://github.com/cosinusoidally/tcc_bootstrap_alt/
Related: the stage0/stage1 series of hex-to-c compiler bootstrapping tools https://github.com/oriansj/stage0?tab=readme-ov-file and OTCC https://bellard.org/otcc/
You may enjoy https://github.com/ludocode/onramp
It would be interesting to understand what non-toy programs can be coded in this subset of C. For example, could tcc be rewritten in this dialect?
https://bootstrapping.miraheze.org/wiki/Main_Page
(Why does the referenced short story remind me of "There Is No Antimemetics Division"?)
This is very nice. I'm currently writing a minimalist C compiler although my goal isn't fitting in a boot sector, it's more targeted at 8-bit systems with a lot more room than that.
This is a great demonstration of how simple the bare bones of C are, which I think is one reason I and many others find it so appealing despite how Spartan it is. C really evolved from B which was a demake of Fortran, if Ken Thompson is to be trusted.
Would and how much would it shrink when if, while, and for were replaced by the simple goto routine? (after all, in assembly there is only jmp and no other fancy jump instruction (I assume) ).
And PS, it's "chose your own adventure". :-) I love minimalism.
What fancy jumps are present in assembly depends on the CPU architecture. But there are always conditional jumps, like JNZ that jumps if the Zero flag isn't set.
It only does if & while, not for. A goto in a single-pass thing would need separate handling for forwards vs backwards jumps, which involves keeping track of data per name (in a form where you can tell when it's not yet set; whereas if/while data is freely held in recursion stack). And you'd still need to handle at least `if ( expr ) goto foo;` to do any conditionals at all.
The “fancy jump” is the branch instruction. As far as I know all ISAs have them. Even rv32i which is famously minimal has several branch instructions in addition to two forms of unconditional jump. Branches are typically used to construct if / for / while as well as && and || (because of short circuiting) and ternary (although some architectures may have special instructions for that that may or may not be faster than branches depending on the exact model). Without it you would have to use computed goto with a destination address computed without conditional execution using constant time techniques.
It's "choose your own adventure"
thats the most important thing i noticed about the article, apart from the forth tokenising ideas.
Beautiful, but make sure to quickly add 2023 to the title.
Discussed at the time: https://news.ycombinator.com/item?id=36064971
Thanks! Macroexpanded:
SectorC: A C Compiler in 512 bytes - https://news.ycombinator.com/item?id=36064971 - May 2023 (80 comments)
why? and why "quickly?
This is the kind of project that reminds you how far removed modern development is from the actual machine. We pile abstractions on abstractions until "Hello World" needs 200MB of node_modules, and then someone fits a C compiler in 512 bytes.
Not saying we should all write boot sector code, but reading through projects like this is genuinely humbling. Great educational resource too.
This kind of comment reminds me of how broad "software development" is.
On other HN posts, they're stating something like "software development is dead", "LLM as a compiler", "Do you read compiled assembly?", and so on.
While some other posts like this contain huge mechanical sympathy and literally r/w the assembly directly.
Compare that to the C compiler in 100,000 lines written by Claude in two weeks for $20,000 (I think was posted on HN just yesterday)
It's a fun comparison, but with the notable difference that that one can compile the Linux kernel and generate code for multiple different architectures, while this one can only compile a small proportion of valid C. It's a great project, but it's not so much a C compiler, as a compiler for a subset of C that allows all programs this compiler can compile to also be compiled by an actual C compiler, but not vice versa.
But can it compile "Hello, World" example from its own README.md?
https://github.com/anthropics/claudes-c-compiler/issues/1
It's fascinating how few people read past the issue title
And this is exactly why coding with AI is not-so-slowly taking over.
Most people think they are more capable than they actually are.
Noticed the part where all it requires is to actually have the headers in the right location?
"The location of Standard C headers do not need to be supplied to a conformant compiler."
From https://news.ycombinator.com/item?id=46920922 discussion.
And it doesn't for the compiler in question either. As long as the headers exist in the places it looks for them. No compiler magically knows where the headers are if you haven't placed them in the right location
stddef.h (et al) should be shipped by the compiler itself, and so it should know where it is. But they rely on gcc for it, hence it doesn't always know where to look. Seems totally fine for a prototype.
Especially given they're not shipping anything. The GCC binaries can't find misplaced or not installed headers either.
Shipping GPL headers that explicitly state that they are part of GCC with a creative commons licensed compiler would probably make a lot of people rather unhappy, possibly even lawyers.
Would you accept the same quality of implementation from a human team?
I've certainly encountered clang & gcc not finding or just not having header files a good couple times. Mostly around cross-compilation, but there was a period of time for which clang++ just completely failed to find any C++ headers on my system.
Yes, clang is famously in this category.
If you copy the clang binary to a random place in your filesystem, it will fail to compile programs that include standard headers.
A compiler that can't magically know how to find headers that don't exist in the expected directory?
Yes, that is the case for pretty much every compiler. I suppose you could build the headers into the binary, but nobody does that.
Consider: content-addressed headers.
Then you might as well embed the headers, since in that case you can't update the compiler and headers separately anyway.
I guess you've heard of https://www.unison-lang.org/
Noticed the part where the exact instructions from the Readme were followed and it didn't work?
So we're down to a missing or unclear description of a dependency in a README - note following the instructions worked for others -, from implications the compiler didn't work.
Well I'm pretty sure the author can make a compliant C compiler in a few more sectors.
I mean we know it can be done in little space, given the many tiny C compilers. I think what is most interesting about this one is exactly the creative shortcuts. It's an interesting design space for e.g. bootstrapping to impose extra restrictions.
The way hashing is used for tokens and for making a pseudo symbol table is such an elegant idea.
I think the same. Really nice project and good trick with hashing tokens.
PS. There left 21 bytes (21 * 0x00 - from 0x01e0 to 0x01fd). Maybe something can be packed there ;)
I actually "shipped" a parser using the symbols' hash(as the only identifier) for a test tool once. Hopefully, the users never used enough symbols to collide 32-bits.
I've had the idea before. Was never quite brave enough to do it. It's elegant until it isn't....
Such a great read! Reminds me of the bootsector OS I made some time ago[^1]
Maybe it's time to equip it with a C compiler...
[1]: https://github.com/shikaan/osle
This is so cool!
Fun fact, Tiny C Compiler was derived from such a C compiler submitted to the the International Obfuscated C Code Contest.
https://www.ioccc.org/2001/bellard/index.html
Further Fun fact, that submission was called OTCC. I reverse engineered it and that provided inspiration for SectorC.
https://xorvoid.com/otcc_deobfuscated.html https://github.com/xorvoid/otcc_deobfuscated
Meh, I did an entire awk interpreter in two lines:
There seems to be a good amount of interest for a boot sector compiler!!
If you're running on Linux, adjust the qemu call to use alsa rather than coreaudio.
I generated a pull request for this on Github. If the author is happy enough with my verbose shell scripting style :-) it might get included.
Reminds me of Allegro SizeHack where we made games in 10KB - but we were using C and Allegro library!
https://www.oocities.org/trentgamblin/sizehack/entries.html#...
> Big Insight #2 is that atoi() behaves as a (bad) hash function on ordinary text. It consumes characters and updates a 16-bit integer.
I could have sworn I remembered atoi() being defined to return 0 for invalid input (i.e. text not representing an integer in base ten).
That would be true of one using a libc, but in a boot sector, you only have the bios, so the atoi being referenced is the one defined in c near the beginning of the article
Ah, I somehow skipped over that exact code block on first read.
Brilliant! I love the stealing of Forth ideas to power this. Forth’s minimalism is highly underrated.
This is really beautiful (I feel like this sort of project is outsider art), thank you for sharing.
Great read. It would be neat to see a mini operating system under 1 kb of code.
C-subset, to be precise; but microcomputer C compilers were in the tens of KB range, for one that can actually compile real C.
For me is not interesting because it fits in 512 bytes, it's interesting because it's very simple. I think it would be a great introduction to learning about compilers.
Why is it called a C Compiler if it's a subset of C?
[flagged]
Why is your visceral reaction is to frame it as a quest for truth versus a great suppression of truth? Everything alright up there?
Literal second sentence in the article, in case it wasn't incredibly obvious to people anyways:
> It supports a subset of C that is large enough to write real and interesting programs.
I'm all for more boring headlines, but this characterization is ridiculous.
I've had enough of headlines that overpromise and underdeliver. It's essentially false advertising. It's not like the word "subset" would put it over the length limit.
Nice, now you can dd it to your boot sector and ... Wait, it is 2026, there are 1000 ways of booting and memory mapping on so-called unified ARM architecture @,@
> I wrote a fairly straight-forward and minimalist lexer and it took >150 lines of C code
was it supposed to be "<150"?
They're saying the naive implementation was more than 150 lines of C code (300-450 bytes), i.e. too big.
Lacking support for structs, I think this is too minimalistic to be called "a C compiler".
you bootstrap it into a library you can include optionally, duh
Weren't structs a fairly late addition to C?
And anyway, isn't that kind of missing the point. 512 bytes isn't much. Your comment is nearly a 5th of that budget.
[flagged]
> but it seems there are others here who don't want to speak of the truth
Or you know, just didn't get hung up on the blatantly obvious thing not being explicitly disclaimed right in the title, only in the preamble?
Not telling the whole truth, little-by-little, this is how honesty crumbles.
Nice. Very K&R-ish. Not a bad thing.
[dead]