Closed
Description
Building Servo with CARGO_INCREMENTAL=1, the script crate takes an unusually long time to build and uses a lot of RAM. Here is an extract of the output of -Z time-passes:
time: 0.224; rss: 122MB parsing
[…]
time: 1.328; rss: 2689MB MIR optimisations
time: 0.296; rss: 2784MB write metadata
time: 2.442; rss: 2871MB translation item collection
time: 0.717; rss: 2897MB codegen unit partitioning
time: 0.608; rss: 10324MB internalize symbols
time: 78.431; rss: 10325MB translation
time: 0.000; rss: 10325MB assert dep graph
time: 1603.681; rss: 10048MB serialize dep graph
time: 0.000; rss: 8851MB llvm function passes [2]
[…]
time: 12.903; rss: 9397MB codegen passes [42]
time: 51.898; rss: 9394MB LLVM passes
time: 0.000; rss: 9394MB serialize work products
time: 9.727; rss: 1411MB linking
Most of the time (over 26 minutes) is spent in "serialize dep graph". During that time, RSS memory usage slowly grows from ~10 GB to ~18 GB. It goes back down to ~10 GB before the end of this pass.
For comparison, this is the same crate without CARGO_INCREMENTAL=1 (but with -C codegen-units=4):
time: 0.204; rss: 118MB parsing
[…]
time: 1.278; rss: 2093MB MIR optimisations
time: 0.279; rss: 2185MB write metadata
time: 2.365; rss: 2269MB translation item collection
time: 0.987; rss: 2295MB codegen unit partitioning
time: 0.582; rss: 4633MB internalize symbols
time: 33.135; rss: 4633MB translation
time: 0.000; rss: 4633MB assert dep graph
time: 0.000; rss: 4633MB serialize dep graph
time: 0.904; rss: 3123MB llvm function passes [3]
[…]
time: 30.495; rss: 3369MB codegen passes [0]
time: 34.953; rss: 3369MB LLVM passes
time: 0.000; rss: 3369MB serialize work products
time: 2.195; rss: 468MB linking
Finished dev [unoptimized + debuginfo] target(s) in 101.86 secs