#[cold] on match arms

[x17jiri]

Edit

This should be in T-lang. I'm not sure how I can change it.

There is discussion: https://rust-lang.zulipchat.com/#narrow/stream/213817-t-lang/topic/Allow.20.23.5Bcold.5D.20on.20match.20and.20if.20arms

Summary

Adds the possibility to use #[cold] attribute on match arms to hint the optimizer that the arm is unlikely to be taken.

Motivation

These hints are sometimes thought to help branch prediction, but the effect is probably marginal. Modern CPUs don't support hints on conditional branch instructions. They either have the current branch in the BTB (branch prediction buffer), or not, in which case the branch is predicted not to be taken.

These hints are, however, helpful in letting the compiler know what is the fast path, so it can be optimized at the expense of the slow path.

grep-ing the LLVM code for BlockFrequencyInfo and BranchProbabilityInfo shows that these hints are used at many places in the optimizer. Such as:

• block placement - improve locality by making the fast path compact and move everything else out of the way
• inlining, loop unrolling - these optimizations can be less aggressive on the cold path therefore reducing code size
• register allocation - preferably keep in registers the data needed on the fast path

History

RFC 1131 ( #26179 ) added likely and unlikely intrinsics, which get converted to llvm.expect.i8. However this LLVM instruction is fragile and may get removed by some optimization passes. The problems with the intrinsics have been reported several times: #96276 , #96275 , #88767

Other languages

Clang and GCC C++ compilers provide __builtin_expect. Since C++20, it is also possible to use [[likely]] and [[unlikely]] attributes.

Use:

if (__builtin_expect(condition, false)) { ... this branch is UNlikely ... }

if (condition) [[likely]] { ... this branch is likely... }

Note that while clang provides __builtin_expect, it does not convert it to llvm.expect.i8. Instead, it looks at the surrounding code and if there is a condition, emits branch weight metadata for conditional branches.

Design

Implementing likely/unlikely type of functions properly to emit branch weights would add significant complexity to the compiler. Additionally, these functions are not easy to use with match arms.

Replacing the functions with attributes is easier to implement and will also work with match.

A question remains whether these attributes should be named likely/unlikely as in C++, or if we could reuse the already existing #[cold] attribute. #[cold] has the same meaning as unlikely, i.e., marking the slow path, but it can currently only be used on entire functions.

I personally prefer #[cold] because it already exists in Rust and is a short word that looks better in code. It has one disadvantage though.
This code:

if cond #[likely] { ... }

becomes:

if cond { ... } #[cold] { ... empty cold branch ... }

In this PR, I implemented the possibility to add #[cold] attribute on match arms. Use is as follows:

match x {
    #[cold] true => { ... } // the true arm is UNlikely
    _ => { ... } // the false arm is likely
}

Limitations

The implementation only works on bool, or integers with single value arm and an otherwise arm. Extending it to other types and to if statements should not be too difficult.

[rustbot]

Thanks for the pull request, and welcome! The Rust team is excited to review your changes, and you should hear from @davidtwco (or someone else) soon.

Please see the contribution instructions for more information. Namely, in order to ensure the minimum review times lag, PR authors and assigned reviewers should ensure that the review label (S-waiting-on-review and S-waiting-on-author) stays updated, invoking these commands when appropriate:

• @rustbot author: the review is finished, PR author should check the comments and take action accordingly
• @rustbot review: the author is ready for a review, this PR will be queued again in the reviewer's queue

[rustbot]

These commits modify the Cargo.lock file. Unintentional changes to Cargo.lock can be introduced when switching branches and rebasing PRs.

If this was unintentional then you should revert the changes before this PR is merged.
Otherwise, you can ignore this comment.

This PR changes MIR

cc @oli-obk, @RalfJung, @JakobDegen, @davidtwco, @celinval, @vakaras

[rustbot]

There are merge commits (commits with multiple parents) in your changes. We have a no merge policy so these commits will need to be removed for this pull request to be merged.

You can start a rebase with the following commands:

$ # rebase
$ git rebase -i master
$ # delete any merge commits in the editor that appears
$ git push --force-with-lease

The following commits are merge commits:

• 1024689

[x17jiri]

This should be in T-lang. I'm not sure how I can change it.

There is discussion: https://rust-lang.zulipchat.com/#narrow/stream/213817-t-lang/topic/Allow.20.23.5Bcold.5D.20on.20match.20and.20if.20arms

[oli-obk]

I have opinions on the impl side, so I wanna review the impl once the lang design is complete

[traviscross]

@rustbot labels +I-lang-nominated

Let's nominate so we can discuss this. @scottmcm has proposed we might want to charter this as an experiment in such a way that we first make this work again in the compiler and then work on syntax.

[oli-obk]

some thoughts:

• You also need to implement the logic for codgen_gcc. It is totally fine to just not support it there and emit normal branches (and leave a FIXME comment for the codegen_gcc authors).
• Is there some reasonable way we can test something here? How are we testing #[cold] functions beyond checking that they don't get inlined?
• The ThinVec could also just be a single index, stating that all arms after that one are cold
  • this will require shuffling arms around so that the cold ones are at the back
  • may need an additional boolean to decide what to do here
• we'll need to run perf to see if this affects performance when not used at all, because it touches a hot type in the MIR
• I'd prefer if this change weren't so invasive (bubbling it through a bunch of places, needing to add special logic to backends that duplicates some of the existing branching logic, yes I'm aware these two concerns are contradictory). Not sure how to do it much better, I'll give it some more thought

@bors try @rust-timer queue

[bors]

⌛ Trying commit 71b694c with merge 4d4892d...

[bors]

☀️ Try build successful - checks-actions
Build commit: 4d4892d (4d4892d72d55a8d6e12c6c7869c8737c5772ab9a)

[x17jiri]

@oli-obk Thank you for comments!

* Is there some reasonable way we can test something here? How are we testing `#[cold]` functions beyond checking that they don't get inlined?

I will look how the rust tests are done.

What I could do is create code like this:

match x {
    true => hot_function(),
    #[cold] false => cold_function(),
}

Then generate llvm file and it should contain something like

br i1 %x, label %bb1, label %bb2, !prof !3
...
!3 = !{!"branch_weights", i32 2000, i32 1}

looking at the weights, %bb1 should contain call to hot_function(), %bb2 should contain call to cold_function().

* The `ThinVec` could also just be a single index, stating that all arms after that one are cold
  
  * this will require shuffling arms around so that the cold ones are at the back
  * may need an additional boolean to decide what to do here

I like the idea of using just an index. But I need to think what would be the complexity of the shuffling and if it is worth it, because the vector only allocates memory if cold branches are used.

* we'll need to run perf to see if this affects performance when not used at all, because it touches a hot type in the MIR

If the performance tests find a regression, is there a way I can run them on my own? I would try to identify what is the bottleneck.

[rust-timer]

Finished benchmarking commit (4d4892d): comparison URL.

Overall result: no relevant changes - no 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.

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

Instruction count

This benchmark run did not return any relevant results for this metric.

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)	-	-	0
Regressions ❌ (secondary)	2.6%	[1.6%, 3.6%]	2
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	-	-	0

Cycles

This benchmark run did not return any relevant results for this metric.

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)	0.1%	[0.0%, 0.1%]	24
Regressions ❌ (secondary)	0.0%	[0.0%, 0.0%]	6
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	0.1%	[0.0%, 0.1%]	24

Bootstrap: 663.975s -> 662.782s (-0.18%)
Artifact size: 308.39 MiB -> 308.43 MiB (0.01%)

[oli-obk]

If the performance tests find a regression, is there a way I can run them on my own? I would try to identify what is the bottleneck

https://github.com/rust-lang/rustc-perf/tree/master/collector/benchlib has instructions for building the perf runner. The command to actually run perf is shown in regressions (there are none, but here's an example: https://perf.rust-lang.org/detailed-query.html?commit=4d4892d72d55a8d6e12c6c7869c8737c5772ab9a&benchmark=helloworld-check&scenario=incr-unchanged&base_commit=366d112fa69164d79239ceeaa49e06497df5497f)

[oli-obk]

Then generate llvm file and it should contain something like

Ah great! We have tests like that in the tests/codegen dir. You can run them with ./x.py test codegen

[rustbot]

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

[rustbot]

There are merge commits (commits with multiple parents) in your changes. We have a no merge policy so these commits will need to be removed for this pull request to be merged.

You can start a rebase with the following commands:

$ # rebase
$ git rebase -i master
$ # delete any merge commits in the editor that appears
$ git push --force-with-lease

The following commits are merge commits:

• 1024689

[x17jiri]

@oli-obk I added the comments. Note that I also opened a second PR to fix likely and unlikely ( #120370 ). If it is accepted, I would rebase and reuse the backend change.

[Dylan-DPC]

Marking this as blocked on #120370

[x17jiri]

This should be unblocked now that #120370 is merged.

With how #120370 evolved, the following code suggested by the current PR:

match x {
    #[cold] true => { ... }
    _ => { ... }
}

can now be written as either:

match unlikely(x) {
    true => { ... }
    _ => { ... }
}

or

match x {
    true => { cold_path(); ... }
    _ => { ... }
}

The question now is whether having attribute brings any additional value or not. If yes, I could reimplement this PR to put cold_path() as the first thing in arms marked with #[cold]. If not, we can probably close this.

[Dylan-DPC]

@x17jiri this is no longer blocked so you can resolve the conflicts and we can push this forward

[x17jiri]

@Dylan-DPC Sure, I can update this to emit cold_path() call as the first thing in #[cold] match arms.

There is no decision from the lang team yet regarding this feature, right?

[Dylan-DPC]

Marked this as waiting on team. Till then can you remove the merge commit and rebase instead? thanks

[rustbot]

Some changes occurred in match lowering

cc @Nadrieril

[rustbot]

Some changes occurred to the intrinsics. Make sure the CTFE / Miri interpreter
gets adapted for the changes, if necessary.

cc @rust-lang/miri, @rust-lang/wg-const-eval

[safinaskar]

It seems we need #133852 if we actually want #[cold] to work