[experimental] Don't partition generic and non-generic code into different CGUs during incr. comp.

[michaelwoerister]

At the moment we are splitting generic and non-generic code into different CGUs when partitioning the crate during incremental compilation. The reasoning behind this is that generic code might change more often than non-generic code. That reasoning was never properly validated though. Since it's simple to just not split things, let's give it a try and see how it performs.

[rust-highfive]

r? @nikomatsakis

(rust_highfive has picked a reviewer for you, use r? to override)

[michaelwoerister]

@bors try

[bors]

⌛ Trying commit c481aae427901e3821db90ef91d4f025ac0c497f with merge 01f9f95a596c37b93bbffdc2ef821b761a3e642f...

[michaelwoerister]

cc #53929

[pnkfelix]

@rust-timer build 01f9f95a596c37b93bbffdc2ef821b761a3e642f

[rust-timer]

Insufficient permissions to issue commands to rust-timer.

[michaelwoerister]

@pnkfelix We should wait for the try build to finish anyway.

[michaelwoerister]

@rust-timer build 01f9f95a596c37b93bbffdc2ef821b761a3e642f

[rust-timer]

Success: Queued 01f9f95a596c37b93bbffdc2ef821b761a3e642f with parent b0297f3, comparison URL.

[bors]

☀️ Test successful - status-travis
State: approved= try=True

[michaelwoerister]

I think we'll have to redo benchmarking after #53962 has been merged. The effects on script-servo and style-servo cases would be rather interesting.

[michaelwoerister]

There are no results yet for script-servo and style-servo but the ones we have don't look too good: https://perf.rust-lang.org/compare.html?start=b0297f3043e4ed592c63c0bcc11df3655ec8cf46&end=01f9f95a596c37b93bbffdc2ef821b761a3e642f&stat=wall-time

[michaelwoerister]

@bors try

[bors]

⌛ Trying commit 4e75f25 with merge 08bf2b2...

[nikomatsakis]

@michaelwoerister

The reasoning behind this is that generic code might change more often than non-generic code. That reasoning was never properly validated though. Since it's simple to just not split things, let's give it a try and see how it performs.

It seems like what we are seeing in these benchmarks is basically just "bigger CGU means more time to rebuild", I suspect?

[bors]

☀️ Test successful - status-travis
State: approved= try=True

[michaelwoerister]

Yes, it's a double-edged sword: If the partition is bigger and it is re-used you'll have saved time when generating it (less redundant work done). But you also increase the amount of work that needs to be re-done if it's not re-used because there are more functions in it that need to be re-compiled although it might just be one or two that actually have changed.

[michaelwoerister]

@rust-timer build 08bf2b2

[rust-timer]

Success: Queued 08bf2b2 with parent 35a5541, comparison URL.

[alexcrichton]

@michaelwoerister I wonder, could we perhaps ignore #[inline] entirely in debug mode? That may get more concrete instantiations in upstream crates (for example String::deref) to avoid generating code in each CGU? (this wouldn't affect optimized mode of course)

[michaelwoerister]

@alexcrichton I thought that we do that already:

rust/src/librustc_mir/monomorphize/item.rs

Lines 87 to 90 in b75b047

	let inline_in_all_cgus =
	tcx.sess.opts.debugging_opts.inline_in_all_cgus.unwrap_or_else(|| {
	tcx.sess.opts.optimize != OptLevel::No
	}) && !tcx.sess.opts.cg.link_dead_code;

[alexcrichton]

Oh hm perhaps! I guess I was thinking that if you have a crate like:

// crate "A"
#[inline]
pub fn foo() {}

and then depended on that with:

// crate "B"
extern crate a;

pub fn bar() { a::foo(); }

we'll codegen a::foo into B's CGU, right? Instead, though, if we ignored #[inline] entirely then A's CGU would have the code for foo and B would reference that.

(this doesn't apply super well to libstd because it's optimized, so it may not be a good idea in general)

[michaelwoerister]

Yeah, something like that should be doable via the existing -Zshare-generics infrastructure that does the same thing for monomorphizations already.

[michaelwoerister]

I opened #54089, so we don't lose track of the idea regarding inline functions. I'm closing this PR since it didn't really work out.