Macro parser performance improvements and refactoring #37701
Conversation
brson
added
the
relnotes
|
Mark-Simulacrum
force-pushed
the
macro-parser-impvement
branch
from
11bf624
to
8c10253
Compare
| match Rc::try_unwrap(nt) { | ||
| Ok(token::NtTT(sub_tt)) => tt = sub_tt, | ||
| Ok(nt) => { | ||
| tt = TokenTree::Token(sp, token::Interpolated(Rc::new(nt))); |
There was a problem hiding this comment.
I believe this recreates the Rc that we just unwrapped. Perhaps this could be avoided with something like:
while let TokenTree::Token(sp, token::Interpolated(nt)) = tt {
if let token::NtTT(..) = *nt {
match Rc::try_unwrap(nt) {
Ok(token::NtTT(sub_tt)) => tt = sub_tt,
Ok(_) => unreachable!(),
Err(nt_rc) => match *nt_rc {
token::NtTT(ref sub_tt) => tt = sub_tt.clone(),
_ => unreachable!(),
},
}
} else {
tt = TokenTree::Token(sp, token::Interpolated(nt.clone()));
break
}
}
Mark-Simulacrum
force-pushed
the
macro-parser-impvement
branch
from
a5c2641
to
5fd8f3d
Compare
Also makes nameize non-public since it's only locally used.
Change multiple functions to be non-public. Change nameize to accept an iterator so as to avoid an allocation.
Mark-Simulacrum
force-pushed
the
macro-parser-impvement
branch
from
5fd8f3d
to
2189f57
Compare
| let sub = ei.matches[idx].clone(); | ||
| new_pos.matches[idx] | ||
| .push(Rc::new(MatchedSeq(sub, mk_sp(ei.sp_lo, | ||
| span.hi)))); | ||
| } |
There was a problem hiding this comment.
If there's a way to avoid doing this work here, that would likely bring significant wins. Callgrind reports ~1 trillion instructions (if I'm reading it right) spent on the let sub = ei.matches[idx].clone(); line for the workload I'm profiling.
There was a problem hiding this comment.
We might be able to avoid this work by refactoring Rc out of Vec in NamedMatch::MatchedSeq, that is MatchedSeq(Vec<Rc<NamedMatch>>, syntax_pos::Span) -> MatchedSeq(Rc<Vec<NamedMatch>>, syntax_pos::Span), perhaps with a RefCell between the Rc and Vec.
Then, sub would be an Rc so cloning would be cheap, and Rc<NamedMatch> would never be needed since cloning NamedMatch would be almost as cheap as an cloning an Rc.
|
After realizing that I was benchmarking old code (from the make build system, I switched to rustbuild mid-PR); latest perf gains are even higher: from ~890 seconds with |
|
I'll try the optimization suggested by @jseyfried and clean up history a little. |
Mark-Simulacrum
force-pushed
the
macro-parser-impvement
branch
from
abc57c4
to
2189f57
Compare
|
Okay, so the suggested optimization doesn't work without more changes, since then we start sharing the vector instead of duplicating it--and we don't want to share it. It's something to look into for the future, but I'll leave it as out of scope for this PR since the changes are unlikely to be trivial. |
|
@bors r+ |
|
📌 Commit 2189f57 has been approved by |
…fried Macro parser performance improvements and refactoring This PR locally increased performance of #37074 by ~6.6 minutes. Follow up to #37569, but doesn't focus explicitly on expansion performance. History is relatively clean, but I can/will do some more polishing if this is deemed mergeable. Partially posting this now so I can get Travis to run tests for me. r? @jseyfried
This PR locally increased performance of #37074 by ~6.6 minutes.
Follow up to #37569, but doesn't focus explicitly on expansion performance.
History is relatively clean, but I can/will do some more polishing if this is deemed mergeable. Partially posting this now so I can get Travis to run tests for me.
r? @jseyfried