More questionable C compiler ideas

After my pdb detour, I used the “prototype IDE” for a while to hack up a little platformer.

Not-even-saving-the-file code updates are definitely the way to go! It’s very fun when you can just type things and see them pop up immediately after the keystroke that types the last letter.

Crashy and fuzzy

I quickly discovered a variety of ways to crash the compiler (oops!¹), so I spent a while fixing all of those that I encountered. This made the editor/compiler/runtime combo stable enough that I could type random C for an hour or so without crashing everything.

I also added a fuzzer target hoping to further improve stability. It’s just a simple integration for libFuzzer (aka -fsanitize=fuzzer in clang) and it passes the fuzz input directly to the compiler update function. This found some bugs in the tokenizer and preprocessor so it was kind of a success.

But unfortunately, even with a helper corpus I only managed to find a couple parser bugs, and no codegen bugs. It is highly improbable that the reason I didn’t find more parser and codegen bugs because they don’t exist. Rather, the problem is that the fuzzer is unable to generate valid-enough semi-C-shaped inputs to make it to those codepaths.

I did a little hunting around for fuzzers that specialize in fuzzing compilers. It looks like some people have tried, but nothing too great, and I didn’t have much success integrating them, so better fuzzing will have to wait for now. (Suggestions?)

More ways to crash

This is of course a C compiler, and C compilers compile C code, and C code is eminently capable of writing its own crashes too!

So once the compiler wasn’t crashing, my next source of flow-interruption is when I write game code that dereferences null or reads out of bounds or whatever.

I haven’t tried to work on this problem yet, but I think that will be the next thing to do in the game shell. It should be possible to add some amount of sandboxing to the game’s C code so that it runs normally in-process, but can trap (say) access violations at the shell level and report a user error, while maintaining a functioning compiler REPL, so that the error can be repaired without restarting.

Easier embedding

When embedding the compiler into another program, libdyibicc.c and .h were required of course, but in addition, you also had to have a copy of (and point the compiler at) the compiler’s built-in include directory. This didn’t really spark joy because it gets into file system path manipulation, non-obvious initial setup steps, etc.

So instead now, all compiler built-in headers (stddef.h, etc.) are slapped into libdyibicc.c during packaging. So if you’re embedding, all you need is the .c and .h. It still uses the system headers (for example, for libc) so those are still required in the normal places (%WindowsSdkDir%, /usr/include, etc.).

Miscellaneous updates

Worse ideas

As a particularly lazy fellow, especially when prototyping, I’m annoyed and tired every time I have to make some janky linked list or

MyObj
objects[MAX_MY_OBJS]

. So in a fit of hubris, I decided to build containers directly into dyibicc as a language feature.

Now what does it mean to be a language feature vs. just a built-in library? To me, I think it means “syntax”, and that’s where we veer into this being a questionable idea.

There are a huge number of Absolutely Fine container libraries for C with various tradeoffs (ctl, Klib, mlib, sgc, STC, and hundreds more). I semi-arbitrarily picked STC as being featured-enough, but not too heavy.

As with all C libraries, it’s necessarily a bit “stutter”-y (especially with more complex types) as you need to repeat the namespace, the type, and the object in each function call.

methodcall

To try to make this nicer, I added some extensions to dyibicc. The first is a pretty simple one: __attribute__((methodcall(PREFIX))). This is an attribute that goes on a struct declaration that makes it “sort of callable”. So instead of writing:

struct my_vector_type { ... };

my_vector_type my_vec = {0};
my_vector_type_push_back(&my_vec, 123);
my_vector_type_push_back(&my_vec, 456);
my_vector_type_push_back(&my_vec, 789);

struct __attribute__((methodcall(my_vector_type_))) my_vector_type { ... };

my_vector_type my_vec = {0};
my_vec..push_back(123);
my_vec..push_back(456);
my_vec..push_back(789);

The .. syntax and methodcall attribute work together, and if the static type of the left-hand side has a methodcall attribute, the right-hand side is rewritten using the PREFIX and a “self” argument.

That is, with methodcall on the vector type, v..push(1) is rewritten to my_vector_type_push(&v, 1).

Additionally, .. follows pointers, so this works as expected, without needing to use -> or (*x).

int myfunc(my_vector_type* x) {
  x..push_back(14);
}

I thought about using . instead of .. for this, and while ambiguous (vs. normal struct field access) I think it could probably work fine. I wasn’t sure if it’d be more or less confusing, and this was slightly easier to implement for now.

Templating

The second simplifying feature was to make the “templating” part automatic. In STC, you need to pre-define various i_* keys, then include a specific header which generates the associated structures and functions and then undefines those keys. So you need to figure out which types you want to use up-front and predefine them, and think about how they’re going to be forward declared between translation units, and so on.

dyibicc has $vec and $map built in now, which deal with this automatically, so you can write $map(int, char*) or $vec(int) and the correct types will be instantiated and declared (or not) as necessary: simple test (notice no #includes). Assuming this doesn’t feel terrible after using it for a while, I guess probably a string type and maybe a set type would be commonly useful too.

whyyyyy would you do this

There are many good, valid, correct arguments against doing this. Primarily, it just isn’t really a C compiler any more, and your code isn’t going to work in another C compiler. Or maybe you just think it’s ugly! But what good is compiler power if you’re not going to abuse it! So I’m going with Embrace, Extend, Extinguish for now. Surely domination of the C compiler market is just around the corner.

End

Paraphrasing the funniest email signoff I’ve received in a while: There can’t possibly be a better way to end a blog post!

In my defense, while many issues were my bad (or more often lazy) code, some were inherited from chibicc. ↩
Always fun to iterate on “yaml scripting” without being able to run the actions locally so you have to make piles of trivial commits to check for typos. ↩