Use DefPathHash instead of HirId to break inlining cycles. #85321

cjgillot · 2021-05-15T09:59:38Z

The DefPathHash is stable across incremental compilation sessions, so provides a total order on LocalDefId. Using it instead of HirId ensures the MIR inliner has the same behaviour for incremental and non-incremental compilation.

A downside is that the cycle tie break is not as predictable is with HirId.

rust-highfive · 2021-05-15T09:59:41Z

r? @jackh726

(rust-highfive has picked a reviewer for you, use r? to override)

jackh726

I don't feel really comfortable reviewing this, since I'm not familiar with the intricacies here. I also don't know who would be best to review.

jackh726 · 2021-05-15T16:19:29Z

compiler/rustc_mir/src/transform/inline.rs

-            // So don't do it if that is enabled.
-            if !self.tcx.dep_graph.is_fully_enabled() && self.hir_id < callee_hir_id {
+            // a lower `DefPathHash` than the callee. This ensures that the callee will
+            // not inline us. This trick only works even with incremental compilation,


- This trick only works even with + This trick works even with

bjorn3 · 2021-05-17T14:03:25Z

compiler/rustc_mir/src/transform/inline.rs

+            // since `DefPathHash` is stable.
+            if self.tcx.def_path_hash(caller_def_id)
+                < self.tcx.def_path_hash(callee_def_id.to_def_id())
+            {
                return Ok(());
            }


Why does this still exist now that we have mir_callgraph_reachable?

Two advantages:

it halves the amount of calls to mir_callgraph_reachable,

it allows for inlining even in the presence of a cycle.

But this completely omits half of all inlining candidates. Also with hir id comparisons this would give a consistent result. Now any rustc version increment would cause the result to change as the def path hash contains the -Cmetadata calculated by cargo based on the compiler version, which would break the tests and seemingly randomly regress/improve performance when enabling mir inlining.

But this completely omits half of all inlining candidates.

To the contrary, this branch allows half of all candidates without further examination.

Also with hir id comparisons this would give a consistent result. Now any rustc version increment would cause the result to change as the def path hash contains the -Cmetadata calculated by cargo based on the compiler version, which would break the tests and seemingly randomly regress/improve performance when enabling mir inlining.

I will change the test to only consider the local part of the DefPathHash, removing the influence of -Cmetadata.

I would prefer to avoid optimization decisions being nondeterministic or otherwise uniform - we already see a lot of pain due to CGU partitioning, for example. This seems likely to be even less explainable, right?

The mir_callgraph_reachable is relatively expensive. If we can avoid calling it half the time that seems like a nice win!

The unpredictable nature of the check is a disadvantage, but the condition is effectively: caller_hash < callee_hash || !reachable(callee, caller), which means that it becomes unpredictable only in a presence of a cycle, since otherwise MIR would be available for inlining anyway. I think this is a reasonable trade-off to make.

BTW. Does comparing local_hash help? It seems to incorporate complete DefPathHash of a parent:

rust/compiler/rustc_hir/src/definitions.rs

Lines 143 to 156 in fa72878

parent.hash(&mut hasher);

let DisambiguatedDefPathData { ref data, disambiguator } = self.disambiguated_data;

std::mem::discriminant(data).hash(&mut hasher);

if let Some(name) = data.get_opt_name() {

// Get a stable hash by considering the symbol chars rather than

// the symbol index.

name.as_str().hash(&mut hasher);

}

disambiguator.hash(&mut hasher);

let local_hash: u64 = hasher.finish();

which means that it becomes unpredictable only in a presence of a cycle, since otherwise MIR would be available for inlining anyway. I think this is a reasonable trade-off to make.

I see. In that case I think it is fine.

jackh726 · 2021-06-01T02:06:10Z

@cjgillot any idea on a better reviewer?

jackh726 · 2021-06-24T14:45:05Z

Let's r? @bjorn3 😉

bjorn3 · 2021-06-24T14:59:12Z

The inliner isn't enabled by default. Any idea how to do a perf run for this?

Dylan-DPC-zz · 2021-07-12T17:54:00Z

The inliner isn't enabled by default. Any idea how to do a perf run for this?

cc @Mark-Simulacrum

Mark-Simulacrum · 2021-07-12T19:07:27Z

Generally speaking there's no trivial way - my recommendation is to edit a PR to change the defaults and then run a try build with that, then revert anything on the PR and run a try build with that. You want to make sure the two try builds have the same parent commit (from master). After that, perf runs for each of the try builds and compare the try build hashes in the UI.

bors · 2021-09-09T00:00:11Z

☔ The latest upstream changes (presumably #80522) made this pull request unmergeable. Please resolve the merge conflicts.

bjorn3 · 2021-09-16T14:37:16Z

@cjgillot I would like to see a perf run first before merging this. #85321 (comment) suggests how to do this.

JohnCSimon · 2021-10-03T06:40:05Z

triage: merge conflicts

bors · 2021-12-23T02:13:28Z

☔ The latest upstream changes (presumably #90408) made this pull request unmergeable. Please resolve the merge conflicts.

bors · 2022-02-18T22:02:50Z

☔ The latest upstream changes (presumably #94121) made this pull request unmergeable. Please resolve the merge conflicts.

Dylan-DPC · 2022-02-19T02:06:51Z

Marking as blocked on #82280

@cjgillot the blocked experimental PR has been closed, so is this still blocked?

pierwill · 2022-03-16T19:29:00Z

Marking as blocked on #82280

@cjgillot the blocked experimental PR has been closed, so is this still blocked?

Friendly, non-urgent ping for @cjgillot, since some work on #90317 is blocked here. :)

cjgillot · 2022-04-02T17:12:32Z

I recall that this PR was blocked on a perf run with inliner enabled. Enabling the inliner by default is not possible yet, because of 2 unresolved type normalization bugs.

#92233 will remove this inlining tie breaker. I see two options:

merge Remove ordering traits from HirId #92233, and merge this PR when the inliner is on by default;
merge this PR and adapt Remove ordering traits from HirId #92233.

compiler/rustc_mir_transform/src/inline.rs

bjorn3 · 2022-04-02T18:33:29Z

I think merging this PR first is best so as to not break MIR inlining even more. r=me with or without the nit fixed.

bjorn3 · 2022-04-03T07:49:32Z

@bors r+

bors · 2022-04-03T07:49:33Z

📌 Commit 297dde9 has been approved by bjorn3

bors · 2022-04-03T07:53:13Z

⌛ Testing commit 297dde9 with merge ec7b753...

bors · 2022-04-03T10:33:46Z

☀️ Test successful - checks-actions
Approved by: bjorn3
Pushing ec7b753 to master...

rust-timer · 2022-04-03T11:47:14Z

Finished benchmarking commit (ec7b753): comparison url.

Summary:

Primary benchmarks: no relevant changes found. 1 results were found to be statistically significant but too small to be relevant.
Secondary benchmarks: no relevant changes found. 1 results were found to be statistically significant but too small to be relevant.

	Regressions 😿 (primary)	Regressions 😿 (secondary)	Improvements 🎉 (primary)	Improvements 🎉 (secondary)	All 😿 🎉 (primary)
count¹	0	0	1	1	1
mean²	N/A	N/A	-0.3%	-3.3%	-0.3%
max	N/A	N/A	-0.3%	-3.3%	0.0%

If you disagree with this performance assessment, please file an issue in rust-lang/rustc-perf.

@rustbot label: -perf-regression

number of relevant changes ↩
the arithmetic mean of the percent change ↩

rust-highfive assigned jackh726 May 15, 2021

rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label May 15, 2021

jackh726 reviewed May 15, 2021

View reviewed changes

bjorn3 reviewed May 17, 2021

View reviewed changes

rust-highfive assigned bjorn3 and unassigned jackh726 Jun 24, 2021

JohnCSimon added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Jul 12, 2021

JohnCSimon added S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Jul 31, 2021

camelid added A-mir-opt-inlining Area: MIR inlining S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Aug 15, 2021

inquisitivecrystal added the T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. label Aug 24, 2021

bjorn3 added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Sep 16, 2021

JohnCSimon added S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. and removed S-waiting-on-author Status: This is awaiting some action (such as code changes or more information) from the author. labels Oct 19, 2021

cjgillot force-pushed the mir-cycle branch from bf23f4d to 5a2c4fe Compare October 22, 2021 21:30

This comment has been minimized.

Sign in to view

cjgillot force-pushed the mir-cycle branch from 5a2c4fe to a7c0a95 Compare October 23, 2021 09:04

cjgillot mentioned this pull request Dec 26, 2021

Remove ordering traits from HirId #92233

Closed

cjgillot force-pushed the mir-cycle branch from 367247a to 19bc4db Compare December 26, 2021 20:26

This comment has been minimized.

Sign in to view

cjgillot force-pushed the mir-cycle branch from 19bc4db to cd1108d Compare January 22, 2022 23:50

This comment has been minimized.

Sign in to view

cjgillot force-pushed the mir-cycle branch from cd1108d to 6933f79 Compare January 23, 2022 18:46

cjgillot force-pushed the mir-cycle branch from 6933f79 to ef7225b Compare March 31, 2022 17:15

bjorn3 reviewed Apr 2, 2022

View reviewed changes

compiler/rustc_mir_transform/src/inline.rs Outdated Show resolved Hide resolved

cjgillot added 3 commits April 2, 2022 23:23

Use DefPathHash instead of HirId to break cycles.

e1b36f5

Use only local hash.

2d3d9b2

Less manipulation of the callee_def_id.

297dde9

cjgillot force-pushed the mir-cycle branch from ef7225b to 297dde9 Compare April 2, 2022 21:29

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-blocked Status: Blocked on something else such as an RFC or other implementation work. labels Apr 3, 2022

bors added the merged-by-bors This PR was explicitly merged by bors. label Apr 3, 2022

bors merged commit ec7b753 into rust-lang:master Apr 3, 2022

rustbot added this to the 1.61.0 milestone Apr 3, 2022

bors mentioned this pull request Apr 3, 2022

Add query to check for impossible predicates, use it to skip NoopMethodCall and Inline #95398

Closed

cjgillot deleted the mir-cycle branch April 3, 2022 11:26

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Use DefPathHash instead of HirId to break inlining cycles. #85321

Use DefPathHash instead of HirId to break inlining cycles. #85321

cjgillot commented May 15, 2021

rust-highfive commented May 15, 2021

jackh726 left a comment

jackh726 May 15, 2021 •

edited

Loading

bjorn3 May 17, 2021

cjgillot May 17, 2021

bjorn3 May 17, 2021

cjgillot May 17, 2021

Mark-Simulacrum May 17, 2021

tmiasko May 17, 2021

bjorn3 Jun 24, 2021

jackh726 commented Jun 1, 2021

jackh726 commented Jun 24, 2021

bjorn3 commented Jun 24, 2021

Dylan-DPC-zz commented Jul 12, 2021

Mark-Simulacrum commented Jul 12, 2021

bors commented Sep 9, 2021

bjorn3 commented Sep 16, 2021

JohnCSimon commented Oct 3, 2021

This comment has been minimized.

bors commented Dec 23, 2021

This comment has been minimized.

This comment has been minimized.

bors commented Feb 18, 2022

Dylan-DPC commented Feb 19, 2022

pierwill commented Mar 16, 2022 •

edited

Loading

cjgillot commented Apr 2, 2022

bjorn3 commented Apr 2, 2022

bjorn3 commented Apr 3, 2022

bors commented Apr 3, 2022

bors commented Apr 3, 2022

bors commented Apr 3, 2022

rust-timer commented Apr 3, 2022

	parent.hash(&mut hasher);

	let DisambiguatedDefPathData { ref data, disambiguator } = self.disambiguated_data;

	std::mem::discriminant(data).hash(&mut hasher);
	if let Some(name) = data.get_opt_name() {
	// Get a stable hash by considering the symbol chars rather than
	// the symbol index.
	name.as_str().hash(&mut hasher);
	}

	disambiguator.hash(&mut hasher);

	let local_hash: u64 = hasher.finish();

Use DefPathHash instead of HirId to break inlining cycles. #85321

Use DefPathHash instead of HirId to break inlining cycles. #85321

Conversation

cjgillot commented May 15, 2021

rust-highfive commented May 15, 2021

jackh726 left a comment

Choose a reason for hiding this comment

jackh726 May 15, 2021 • edited Loading

Choose a reason for hiding this comment

bjorn3 May 17, 2021

Choose a reason for hiding this comment

cjgillot May 17, 2021

Choose a reason for hiding this comment

bjorn3 May 17, 2021

Choose a reason for hiding this comment

cjgillot May 17, 2021

Choose a reason for hiding this comment

Mark-Simulacrum May 17, 2021

Choose a reason for hiding this comment

tmiasko May 17, 2021

Choose a reason for hiding this comment

bjorn3 Jun 24, 2021

Choose a reason for hiding this comment

jackh726 commented Jun 1, 2021

jackh726 commented Jun 24, 2021

bjorn3 commented Jun 24, 2021

Dylan-DPC-zz commented Jul 12, 2021

Mark-Simulacrum commented Jul 12, 2021

bors commented Sep 9, 2021

bjorn3 commented Sep 16, 2021

JohnCSimon commented Oct 3, 2021

This comment has been minimized.

bors commented Dec 23, 2021

This comment has been minimized.

This comment has been minimized.

bors commented Feb 18, 2022

Dylan-DPC commented Feb 19, 2022

pierwill commented Mar 16, 2022 • edited Loading

cjgillot commented Apr 2, 2022

bjorn3 commented Apr 2, 2022

bjorn3 commented Apr 3, 2022

bors commented Apr 3, 2022

bors commented Apr 3, 2022

bors commented Apr 3, 2022

rust-timer commented Apr 3, 2022

Footnotes

jackh726 May 15, 2021 •

edited

Loading

pierwill commented Mar 16, 2022 •

edited

Loading