Investigate lto & codegen units

[Keats]

Right now the Cargo.toml has:

[profile.release]
lto = true
codegen-units = 1

in it for better performance at the cost of compilation speed.
Investigate how much additional time it takes & the perf difference.
This should be ideally done after #782

[southerntofu]

I just tried commenting out these two lines from Cargo.toml and the results are interesting. A very naive benchmark on my side (simply renaming a variable in src/main.rs and building) shows:

• the current config takes 2-3 minutes to rebuild and yields a 17Mb binary
• commented-out lto/codegen-units takes 8-20 seconds to rebuild and yields a 22Mb binary

[samford]

For what it's worth, looking back at the original post on the forum, this addition was proposed as a way of reducing binary size (rather than increasing performance), which still appears to hold true.

[Keats]

Yep, I don't think many people are actually building the release version locally except for me. Maybe a better way would be to inject it only on CI? So users still get the small binary but don't suffer any penalties locally.

[samford]

I had a similar thought, since it seems like the best of both worlds, but the concern with only having the release optimizations on CI is that it might lead to unoptimized binaries in other places.

For example, I remember that @ignatenkobrain maintains Zola on Fedora and it may be compiled from source in that context (rather than just downloading a binary from GitHub), so they would end up with an unoptimized release binary with the aforementioned setup. Anyone who maintains a separate build of Zola like this would have to know to add these optimizations to reduce the binary size (even if it was just uncommenting these lines in Cargo.toml).

I feel like it's better to have the release optimizations by default and anyone who needs faster release compilation locally can comment out the [profile.release] section in Cargo.toml.

[ignatenkobrain]

Well, yes and no. We do pass codegen-units=1 in our RUSTFLAGS. But we don't have -C lto now. I am not sure why we haven't enabled it yet.

[Keats]

Compelling argument @samford
Closing that.