Improve state machine codegen #258
Comments
|
This issue is intended for status updates only. For general questions or comments, please contact the owner(s) directly. |
|
This is, I believe, mostly waiting on us on the lang team to have a look, probably in a design meeting, to feel out what's in the realm of possibility for us to accept. |
|
@traviscross how would we make progress on that? So far we've mostly been talking to @joshtriplett, under the assumption that a Our current implementation handles the following #![feature(loop_match)]
enum State {
A,
B,
}
fn main() {
let mut state = State::A;
#[loop_match]
'outer: loop {
state = 'blk: {
match state {
State::A =>
{
#[const_continue]
break 'blk State::B
}
State::B => break 'outer,
}
}
}
}Crucially, this does not add syntax, only the attributes and internal logic in MIR lowering for statically performing the pattern match to pick the right branch to jump to. The main challenge is then to implement this in the compiler itself, which we've been working on (I'll post our tl;dr update shortly) |
|
Some benchmarks (as of march 18th) A benchmark of https://github.com/bjorn3/comrak/blob/loop_match_attr/autolink_email.rs, basically a big state machine that is a perfect fit for loop match
A benchmark of zlib decompression with chunks of 16 bytes (this makes the impact of The important points: |
|
TL;DR: We've started work on implementing Our next steps can be found in the todo file, and focus mostly on improving the code quality and robustness. |
|
Thanks for that update. Have reached out separately. |
Summary
We want to improve rustc codegen, based on this initialive by the Trifecta Tech Foundation. The work focusses on improving state machine code generation, and finding (and hopefully fixing) cases where clang produces better code than rustc for roughly equivalent input.
Tasks and status
The text was updated successfully, but these errors were encountered: