MIR-only rlibs

[saethlin]

Status update: The strategy I implemented in this PR is to make builds of non-rlib crates monomorphize everything they need. I was quite thorough about this and had to add a handful of hacks to make it happen.

But I now believe that was a mistake, because of how it breaks proc-macro crates. I think those crates can only work if they are able to cooperate with the compiler correctly to share types across the proc-macro bridge. Currently this is achieved by linking both against the libstd dylib, which contains monomorphizations of the allocator shims and probably a handful of other things. So in order to make this work, I at least need to figure out how to make the compiler expose all the monomorphizations of the standard library, and make the proc-macro crates link against those instead of monomorphizing their own.

---

Originally I was just trying to imitate MIR-only rlibs with -Zcross-crate-inline-threshold=yes, but I think I have enough skill to actually implement the behavior. We shall see.

At time of writing, set RUSTFLAGS=-Zmir-only-libs and if each session that Cargo kicks off only compiles an rlib or an executable, every rlib should contain only MIR and the build should work. Any other configuration probably explodes unceremoniously.

r? @ghost

[saethlin]

@bors try @rust-timer queue

[bors]

⌛ Trying commit d9e0f19 with merge f4568e9...

[bors]

☀️ Try build successful - checks-actions
Build commit: f4568e9 (f4568e9b4c437be978fa03a4b3ebc07f5800c83d)

[rust-timer]

Finished benchmarking commit (f4568e9): comparison URL.

Overall result: ❌✅ regressions and improvements - ACTION NEEDED

Benchmarking this pull request likely means that it is perf-sensitive, so we're automatically marking it as not fit for rolling up. While you can manually mark this PR as fit for rollup, we strongly recommend not doing so since this PR may lead to changes in compiler perf.

Next Steps: If you can justify the regressions found in this try perf run, please indicate this with @rustbot label: +perf-regression-triaged along with sufficient written justification. If you cannot justify the regressions please fix the regressions and do another perf run. If the next run shows neutral or positive results, the label will be automatically removed.

@bors rollup=never
@rustbot label: -S-waiting-on-perf +perf-regression

Instruction count

This is a highly reliable metric that was used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	156.8%	[0.3%, 3121.3%]	166
Regressions ❌ (secondary)	1167.6%	[0.1%, 21506.7%]	116
Improvements ✅ (primary)	-0.8%	[-6.6%, -0.2%]	47
Improvements ✅ (secondary)	-1.8%	[-11.6%, -0.1%]	79
All ❌✅ (primary)	122.0%	[-6.6%, 3121.3%]	213

Max RSS (memory usage)

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	32.5%	[0.4%, 448.7%]	179
Regressions ❌ (secondary)	27.0%	[0.7%, 183.5%]	151
Improvements ✅ (primary)	-2.8%	[-5.1%, -0.5%]	2
Improvements ✅ (secondary)	-8.9%	[-9.2%, -8.6%]	3
All ❌✅ (primary)	32.1%	[-5.1%, 448.7%]	181

Cycles

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	185.0%	[0.7%, 2991.0%]	151
Regressions ❌ (secondary)	1094.9%	[1.3%, 18459.6%]	114
Improvements ✅ (primary)	-1.6%	[-1.6%, -1.6%]	1
Improvements ✅ (secondary)	-3.5%	[-5.0%, -2.1%]	5
All ❌✅ (primary)	183.8%	[-1.6%, 2991.0%]	152

Binary size

Results

This is a less reliable metric that may be of interest but was not used to determine the overall result at the top of this comment.

	mean	range	count
Regressions ❌ (primary)	73.5%	[0.0%, 423.5%]	134
Regressions ❌ (secondary)	440.4%	[0.1%, 4194.7%]	52
Improvements ✅ (primary)	-18.0%	[-49.7%, -0.8%]	17
Improvements ✅ (secondary)	-32.3%	[-50.4%, -0.4%]	73
All ❌✅ (primary)	63.2%	[-49.7%, 423.5%]	151

Bootstrap: 672.461s -> 1381.575s (105.45%)
Artifact size: 312.46 MiB -> 339.60 MiB (8.69%)

[rustbot]

Failed to set assignee to ghost: invalid assignee

Note: Only org members with at least the repository "read" role, users with write permissions, or people who have commented on the PR may be assigned.

[bors]

☔ The latest upstream changes (presumably #119843) made this pull request unmergeable. Please resolve the merge conflicts.

[oli-obk]

The MCP has been accepted

[saethlin]

👍

This PR is still not really mergeable because it breaks proc macros. I think because I'm monomorphizing the allocator shims wrong.

[bors]

☔ The latest upstream changes (presumably #121644) made this pull request unmergeable. Please resolve the merge conflicts.

[bors]

☔ The latest upstream changes (presumably #122037) made this pull request unmergeable. Please resolve the merge conflicts.

[bors]

☔ The latest upstream changes (presumably #122568) made this pull request unmergeable. Please resolve the merge conflicts.

[bors]

☔ The latest upstream changes (presumably #124558) made this pull request unmergeable. Please resolve the merge conflicts.

[joshtriplett]

How feasible would it be to opt into a MIR-only rlib for a particular dependency? That might sidestep some of the difficulties associated with using it everywhere, while allowing large crates with very few items used to opt into it to reduce how much gets compiled.

[saethlin]

while allowing large crates with very few items used to opt into it to reduce how much gets compiled.

If you want to try using this to cut down build times, you can investigate whether it would help by setting -Zcross-crate-inline-threshold=always for just that crate. I suspect that doing so will regress build times.

But thanks for pinging this PR! I'll try to rebase it up soon just to keep it alive.

[joshtriplett]

@saethlin

I tested this on one of my projects:

Baseline:

Finished release profile [optimized] target(s) in 10m 31s

With the three biggest crates in the dependency tree (which I use a tiny fraction of the items of) using -Zcross-crate-inline-threshold=always:

Finished release profile [optimized] target(s) in 6m 31s

[saethlin]

Note that CI may pass, because in line with:

That might sidestep some of the difficulties associated with using it everywhere

I removed the line that was making the compiler bootstrap with -Zmir-only-rlibs set. So the option is now untested in CI, I think.

[joshtriplett]

Update (same project, flags applied to the same set of 3 large crates with most items unused):

Build	Time
Baseline	10m 30s
-Zcross-crate-inline-threshold=always	6m 12s
-Zcross-crate-inline-threshold=always -Zmir-opt-level=0	5m 34s
-Zcross-crate-inline-threshold=always -Zmir-opt-level=1	5m 48s
-Zcross-crate-inline-threshold=always -Zinline-mir=no	5m 51s

[bors]

☔ The latest upstream changes (presumably #132173) made this pull request unmergeable. Please resolve the merge conflicts.