rustc_codegen_gcc: Progress Report #33
What is rustc_codegen_gcc?
rustc_codegen_gcc is a GCC ahead-of-time codegen for rustc, meaning that it can be loaded by the existing rustc frontend, but benefits from GCC by having more architectures supported and having access to GCC’s optimizations. It is not to be confused with gccrs, which is a GCC frontend for Rust.
GCC patches status
We implemented a new patch that is yet to be sent to the mailing list:
-
[PATCH] Allow gcc_jit_type_get_size to work with pointers (thanks to GuillaumeGomez!)
We got reviews of some patches that were sent in the past and some of them were merged:
State of rustc_codegen_gcc
Here’s what has been done during the past month:
-
Add support for Float16, Float32, Float64 and Float128 (thanks to zedar!)
-
ui pattern failure tests (thanks to lordshashank!)
-
prevent libgccjit.so download on unsupported os/arch (thanks to Skgland!)
-
Improve Readme.md format (thanks to GuillaumeGomez!)
-
simd: implement pointer provenance intrinsics (thanks to sadlerap!)
-
adding more env vars (thanks to Gerson2102!)
This month, the main highlight is the support for the new types f16
and f128
(thanks to zedar!).
We also were able to finish the sync with the Rust compiler, which was much harder than usual, so we’ll be able to get back on working on rustup
support.
During this sync, we implemented a dummy ThinLTO implementation, so while it doesn’t do ThinLTO, it won’t panic anymore when attempting to compile a project with ThinLTO enabled.
We did many attempts to sync with the Rust compiler since March because we had a very hard time keeping up with all the changes in the Rust compiler that break things every now and then:
It does not help that we have tooling issues for those syncs (git subtree
being the main culprit here).
Moreover, one change that was particularly difficult to mimic in libgccjit
was that rustc
will now change the type of some global variables after they were created.
We finally merged the sync PR by marking as ignored some tests that used to pass (to be able to move on and fix those later), and we had to generate fat LTO objects in some cases where only the bitcode was request, so I’ll try to fix those after my vacation.
We plan to enable more tests in the CI of the Rust repo for rustc_codegen_gcc
in order to avoid having that many problems in the future for the sync.
Next month, we’ll fix those issues and we’ll continue towards our goal of distributing this project through rustup
, which should hopefully happens in a couple of months.
Here’s a rough summary of what has been implemented:
Feature | Last month completion | Completion | Delta |
---|---|---|---|
Unwinding. |
80% |
80% |
|
LTO. |
80% |
80% |
|
More function and variable attributes. |
20% |
20% |
|
Target features (to detect what is supported in an architecture, like SIMD). |
82% |
82% |
|
Debug info. |
20% |
||
Thin LTO. |
0% |
5% |
+5% |
Rustup distribution. |
0% |
||
SIMD for other architectures than x86-64. |
0% |
||
Support for new architectures in libraries (libc, object, …) and rustc. |
0% |
||
Refactor to rustc_codegen_ssa to make it easier for the GCC codegen |
0% |
||
Endianness support for non-native 128-bit integers. |
Done |
||
SIMD (x86-64). |
Done |
||
Basic and aggregate types. |
Done |
||
Operations, local and global variables, constants, functions, basic blocks. |
Done |
||
Atomics. |
Done |
||
Thread-local storage. |
Done |
||
Inline assembly. |
Done |
||
Many intrinsics. |
Done |
||
Metadata. |
Done |
||
Setting optimization level. |
Done |
||
Packed structures. |
Done |
||
Alignment, symbol visibility, attributes. |
Done |
||
128-bit integers. |
Done |
UI tests progress
Here are the results of running the UI tests in the CI:
-
https://github.com/rust-lang/rustc_codegen_gcc/actions/runs/9813140371/job/27098581101#step:15:4635
-
https://github.com/rust-lang/rustc_codegen_gcc/actions/runs/9813140371/job/27098581731#step:15:4667
-
https://github.com/rust-lang/rustc_codegen_gcc/actions/runs/9812866864#summary-27097769703 (failures)
Category | Last Month | This Month | Delta |
---|---|---|---|
Passed |
5664 |
5931 |
+267 |
Failed |
71 |
88 |
+17 |
(I removed the LTO tests from the failed row because those tests pass in the CI job where LTO is enabled.)
How to contribute
rustc_codegen_gcc
If you want to help on the project itself, please do the following:
-
Run the tests locally.
-
Choose a test that fails.
-
Investigate why it fails.
-
Fix the problem.
Even if you can’t fix the problem, your investigation could help, so if you enjoy staring at assembly code, have fun!
Crates and rustc
If you would like to contribute on adding support for Rust on
currently unsupported platforms, you can help by adding the support
for those platforms in some crates like libc
and object
and also
in the rust compiler itself.
Test this project
Otherwise, you can test this project on new platforms and also compare the assembly with LLVM to see if some optimization is missing.
To do so, follow these instructions to build the project and run a program via the cargo
command of our script.
If you find a bug, please open an issue.
Good first issue
Finally, another good way to help is to look at good first issues. Those are issues that should be easier to start with.
Thanks for your support!
I wanted to personally thank all the people that sponsor this project: your support is very much appreciated.
A special thanks to the following sponsors:
-
Futurewei
-
Shnatsel
-
Rust Foundation
A big thank you to bjorn3 for his help, contributions and reviews. And a big thank you to lqd and GuillaumeGomez for answering my questions about rustc’s internals. Another big thank you to Commeownist for his contributions.
Also, a big thank you to the rest of my sponsors:
-
kpp
-
0x7CFE
-
oleid
-
acshi
-
joshtriplett
-
djc
-
sdroege
-
pcn
-
alanfalloon
-
davidlattimore
-
colelawrence
-
zmanian
-
alexkirsz
-
berkus
-
belzael
-
yvt
-
Shoeboxam
-
yerke
-
bes
-
srijs
-
kkysen
-
riking
-
Lemmih
-
memoryruins
-
senden9
-
robjtede
-
Jonas Platte
-
Sam Harrington
-
Jonas
-
Eugene Bulkin
-
Joseph Garvin
-
MarcoFalke
-
athre0z
-
Sebastian Zivota
-
Oskar Nehlin
-
Nicolas Barbier
-
Daniel
-
Justin Ossevoort
-
kiyoshigawa
-
Daniel Sheehan
-
Marvin Löbel
-
nacaclanga
-
0x0177b11f
-
L.apz
-
JockeTF
-
davidcornu
-
stuhood
-
Mauve
-
icewind1991
-
nicholasbishop
-
David Vasak
-
Eric Driggers
-
Olaf Leidinger
-
UtherII
-
simonlindholm
-
lemmih
-
Eddddddd
-
rrbutani
-
Mateusz K
-
thk1
-
0xdeafbeef
-
teh
-
KirilMihaylov
-
Vladislav Sukhmel
-
CohenArthur
-
ximou
-
Kate Kiesel
and a few others who preferred to stay anonymous.
Former sponsors/patreons:
-
igrr
-
embark-studios
-
saethlin
-
Traverse-Research
-
finfet
-
Alovchin91
-
wezm
-
mexus
-
raymanfx
-
ghost
-
gilescope
-
olanod
-
Denis Zaletaev
-
Chai T. Rex
-
Paul Ellenbogen
-
Dakota Brink
-
Botlabs
-
Cass
-
Oliver Marshall
-
pthariensflame
-
tedbyron
-
sstadick
-
Absolucy
-
rafaelcaricio
-
dandxy89
-
luizirber
-
regiontog
-
vincentdephily
-
zebp
-
Hofer-Julian
-
messense
-
fanquake
-
jam1garner
-
sbstp
-
evanrichter
-
Nehliin
-
nevi-me
-
TimNN
-
steven-joruk
-
seanpianka
-
spike grobstein
-
Jeff Muizelaar
-
robinmoussu
-
Chris Butler
-
sierrafiveseven
-
icewind
-
Thomas Colliers
-
Tommy Thorn
-
Bálint Horváth
-
Matthew Conolly
-
Lapz
-
Myrik Lord
-
T
-
Emily A. Bellows
-
Chris
-
repi
-
opensrcsec