Hi Iain,
On 1/25/24 17:26, Iain Sandoe wrote:
Hi Arthur,
On 25 Jan 2024, at 15:03, Arthur Cohen <arthur.co...@embecosm.com> wrote:
On 1/23/24 08:23, Richard Biener wrote:
On Mon, Jan 22, 2024 at 7:51 PM Arthur Cohen <arthur.co...@embecosm.com> wrote:
Hi everyone,
In order to increase the development speed of Rust features, we are
seeking feedback on reusing some Rust components directly within our
front-end. As mentioned in other conferences, the most important
component we would like to integrate into the front-end is the polonius
borrow-checker. Another component we've become interested in is the
`rustc_format_parser` library, responsible for handling parsing and
handling of format arguments as described in the documentation for
`std::fmt` [1].
However, since these libraries are written in Rust, GCC in itself is not
yet able to compile them. They all depend on the Rust standard library,
which we cannot yet compile and link properly. This obviously raises a
question - how to actually compile, integrate and distribute these
components?
We do have the option to rewrite them from scratch, but we feel that
spending time on these existing, correct, battle-tested and easily
integrable components instead of focusing on other aspects of the
language would be a mistake. Spending this time instead on Rust features
that we are missing for compiling these components would, in our
opinion, yield more results, as it would also help in compiling other
Rust programs.
We could either distribute these components as compiled libraries, or
look at integrating the official Rust compiler to our build system as a
temporary measure. I am aware that this would mean restricting the Rust
GCC front-end to platforms where the official Rust compiler is also
available, which is less than ideal.
But that's only for the host part - you can still cross compile to another
target and possibly, once the GCC frontend can compile these libraries
itself, also use them to bootstrap a hosted version on that target -
speaking of ..
However, this would only be
temporary - as soon as the Rust front-end is able to compile these
components, we would simply reuse them and compile them with gccrs as
part of our bootstrapping process.
.. gccrs would then need to be able to build itself without those modules,
at least during stage1 so that the stage1 compiler can then be used to
build them. Or you go like m2 and build a "mini-rust" that's just capable
of building the modules.
Right, that makes a lot of sense. We should definitely be able to build the
format string parser without a format string parser, as it does not use format
strings for error handling or anything. And even if it did, it would be pretty
easy to remove that and do the formatting by hand.
Similarly, the borrow checker is not "needed" for compilation and we do plan on
building stage1 without it, while making it mandatory for stage2/3 builds.
As long as you leave a route available for porting it to either new (or old)
systems which do not have rustc (or any other rust compiler available).
E.g. with Ada it is possible to port to a new platform by first building a
cross-compiler and then to use that cross-compiler to build a “native cross”
(build != host == target) to provide an initial compiler on the target platform.
Thanks, this is really helpful. I'll have to figure out how to make that
happen in the build system but this is definitely something we'd be
interested in doing.
Thanks,
Arthur
Those cross compilers are not bootstrapped (it is a single stage) - but it is
reasonable to make it a condition that the $build compiler is the same (even
exact) version as the sources you are trying to port to the new platform.
thanks
Iain
I think re-using parts already available is very sensible at this point. Note
that while we might temporarily require a host rust compiler to boostrap
gccrs I'd rather not have the build system download something from the
internet - so at least the sources of those dependences need to be in the
GCC repository, possibly in a new toplevel directory.
Okay, that makes a lot of sense. I was thinking of adding a basic check for the
Rust compiler to be present when compiling these components - and error out if
that isn't the case. Are you suggesting we embed a full copy of rustc in GCC
and build it from source when compiling the Rust frontend? Or am I
misunderstanding?
The documentation for `std::fmt` [1] describes all of the features
available in Rust format strings. It also contains a grammar for the
format-string parser, which we would need to re-implement on top of
supporting all the formatting features. As a prototype, I wrote an FFI
interface to the `rustc_format_parser` library and integrated it to our
macro expansion system, which took me less than a couple hours. In less
than an afternoon, we had bindings for all of the exported types and
functions in the library and had access to a compliant and performant
Rust format string parser. But re-implementing a correct
`format_args!()` parser - with the same performance as the Rust one, and
the same amount of features - would probably take days, if not weeks.
And we are talking about one of the simplest components we aim to reuse.
Something like a fully-compliant trait solver for the Rust programming
language would take months if not years to implement properly from scratch.
I would like to stress once again that relying on distributing compiled
libraries or using `rustc` in our build system would be temporary, and
that these measures would be removed as soon as gccrs is able to compile
these components from source.
I am looking for comments on this decision as well as the best practices
we could adopt. Have there been any similar decisions for other
self-hosted front-ends? Any previous email threads/commits that you
could point me to would be greatly appreciated.
Not in this very same way but the D frontend is a shim around the official
DMD frontend and the Go frontend imports the golang standard library
(but the frontend itself is written in C++ and doesn't use any part of it).
The C++ frontend uses part of the C++ standard library (but it can of
course build that itself - but it requires a host C++ compiler with library).
Thanks for the pointers. I was wondering if this is something that the Ada
frontend had faced at the beginning, but I've been told it does not have a lot
of dependencies anyway so this might not be helpful.
Richard.
Thanks,
Arthur
[1]: https://doc.rust-lang.org/std/fmt/
Thanks a lot for taking the time, I really appreciate it.
Best,
Arthur
