Firefox compile time regression

[glandium]

Preliminary note: this may or may not be related to #43141, but just in case, I file this separately instead of piling up there.

I'm sorry I don't have much more fine-grained data at the moment, but -Ztime-passes, when multiple things are built in parallel (while building Firefox), is not entirely helpful, because it doesn't tell to which crate the interleaving numbers belong to.

Rust build times on Firefox are already bad with 1.18. Some of the most time consuming crates have build times that look like the following:

• 74.992975322 webrender
• 99.960699075 geckoservo
• 259.393332042 gkrust
• 259.639402886 gkrust_gtest
• 363.456949809 style

With the latest nightly as of writing (1.20.0-nightly (f85579d 2017-07-12)), those times become:

• 85.71714646 webrender
• 124.097890719 geckoservo
• 384.619852505 gkrust_gtest
• 396.106775997 gkrust
• 455.281318295 style

I'm trying to narrow it down to older nightlies, but it started badly, because I got an ICE building the style crate on the last 1.18.0 nightly (2017-04-29)

Those are times for --release with lto enabled (I have a separate bug to file about lto, but I don't think that's relevant to this regression).

CC: @froydnj, @rillian

[glandium]

Mmmmm with 1.19.0-nightly (75b0568 2017-05-15):

• 60.430645418 webrender
• 116.540179527 0 geckoservo
• 370.565763687 0 gkrust_gtest
• 372.559032193 0 gkrust
• 433.088609909 0 style

With 1.19.0-nightly (d3abc80 2017-05-09):

• 63.26174462 webrender
• 117.30302075 geckoservo
• 319.482715603 gkrust
• 320.997348082 gkrust_gtest
• 387.128239057 style

There would seem to be different regression points for different things, and early in the 1.19 development, at that (I tried 75b0568 because of range given in #43141 )

Also note that the order in which the mentioned crates are built is webrender, style, geckoservo, gkrust, gkrust_gtest (the last two are interchangeable depending on timing), and are all happening sequentially because, to simplify things, they are a dependency chain (except webrender, which may build in parallel of style, but it's not like it would matter much).

Also note that this long tail of builds takes so long that by the time geckoservo starts to build after style, there's only that happening on the build machine (everything C++ that doesn't depend on rust is done), and in fact, style has been the only thing happening on the build machine for a while too...

[glandium]

Note, I'm not excluding that as a counterpart, the resulting builds might have better code, but I haven't verified that (and I don't know which specific tests in the Firefox test suites would tell me that about the rust code)

[alexcrichton]

cc @rust-lang/compiler

[arielb1]

@glandium

I remember that LLVM 4.0 introduced a 10% regression in LLVM passes time, but that was #40123 (merged Apr 24).

[brson]

P-?

[michaelwoerister]

It would be interesting to know if turning off LTO makes a difference here.

[nikomatsakis]

More data points would also be useful -- how hard are they to gather?

[glandium]

What kind of data points? Different versions? Different configurations?

[michaelwoerister]

@glandium Different Rust versions, yes, and if possible, timings for specific passes.

[michaelwoerister]

Some more questions:

• Can you give some instructions on how to build Firefox with Stylo enabled? Is it just ./mach build?
• Does FF use the system provided Rust? I.e. can I just switch the Rust version via rustup and expect the FF build system to pick that up?
• Where does the Rust code in FF live? (I tried changing files in the servo subdirectory but that did not seem to do anything)