On 13-08-23 11:41 AM, Bill Myers wrote:
>> > 2. Distribute compilations and tests across a cluster of machines (like
>> > distcc)
>> >
>>
>> Compilation is 99% serial (the only things that happen in parallel
>> are rustpkg and rustdoc etc at the end, and they are almost nothing),
>> though tests could be distributed (and Graydon is working on doing
>> that afaik).
>
> Are you sure?
Pretty sure:
- We essentially always do whole-program / link-time optimization
in C++ terminology. That is, we run a whole crate through LLVM at
once. Which _would_ be ok (I think!) if we weren't generating
quite so much code. It is an AOT/runtime trade but one we
consciously designed-in to the language.
- We are self-hosted so always bootstrap. Stage N comes before stage
N+1. Likewise when we cross compile, host comes before target.
We have 3 stages and when doing an all-cross it grows to 9.
You can't do those in any more parallel.
- Parsing, type checking and similar middle-end passes could
probably be parallelized eventually -- they were designed to
support this in principle -- but in practice it would require
a lot of rearchitecture. Also: these aren't the major speed
points (possibly with the exception of typechecking, it's a
_little_ bottleneck). Here's a normal compile phase-timing:
$ rustc -O -Z time-passes syntax.rs
time: 0.334 s parsing
time: 0.002 s std macros injection
time: 0.062 s configuration 1
time: 0.355 s expansion
time: 0.109 s configuration 2
time: 0.107 s maybe building test harness
time: 0.106 s std injection
time: 0.113 s ast indexing
time: 0.122 s external crate/lib resolution
time: 0.011 s language item collection
time: 0.779 s resolution
time: 0.000 s looking for entry point
time: 0.029 s freevar finding
time: 0.064 s region resolution
time: 0.017 s region parameterization inference
time: 0.059 s type collecting
time: 0.241 s coherence checking
time: 5.623 s type checking
time: 0.153 s const marking
time: 0.013 s const checking
time: 0.060 s privacy checking
time: 0.021 s effect checking
time: 0.010 s loop checking
time: 0.021 s stack checking
time: 0.170 s compute moves
time: 0.073 s match checking
time: 0.104 s liveness checking
time: 0.478 s borrow checking
time: 0.068 s kind checking
time: 0.028 s reachability checking
time: 0.103 s lint checking
time: 3.974 s translation
time: 33.939 s LLVM passes
time: 0.080 s linking
As you can see, LLVM dominates. We generate a lot of code (I pointed
to several clear culprits at the head of this email thread).
-Graydon
_______________________________________________
Rust-dev mailing list
[email protected]
https://mail.mozilla.org/listinfo/rust-dev
