[stdlib] Mark `contains` methods on `Range/ClosedRange` transparent #81458

lorentey · 2025-05-12T22:15:57Z

These are really popular shorthands for implementing bounds checks, but currently they aren’t even marked @inline(__always), so in larger functions they tend to remain outlined, defeating optimizations that would otherwise happen.

The corresponding variants on the partial range types are @_transparent; that helps unoptimized performance, so let’s apply the same attribute here.

rdar://151177326

lorentey · 2025-05-12T22:16:12Z

@swift-ci test

lorentey · 2025-05-12T22:16:54Z

I'm not expecting any benchmark impact, but.

@swift-ci benchmark

benrimmington · 2025-05-13T02:30:36Z

Should you also update:

contains(_:) in the PartialRangeFrom extension?
~=(_:_:) in the RangeExpression extension?

glessard · 2025-05-13T18:14:18Z

@swift-ci please benchmark

glessard · 2025-05-13T20:59:00Z

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
ArrayAppendGenericStructs	1390.0	1765.0	+27.0%	0.79x (?)

Code size: -O

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
StringDistance.utf16.ascii	115.15	127.056	+10.3%	0.91x (?)

Code size: -Osize

Performance (x86_64): -Onone

Improvement	OLD	NEW	DELTA	RATIO
ClosedRangeContainsRange	15858.0	10739.0	-32.3%	1.48x
RangeContainsClosedRange	12057.0	10318.0	-14.4%	1.17x (?)
ArrayAppendReserved	2786.0	2530.0	-9.2%	1.10x (?)
Data.init.Sequence.809B.Count.RE.I	19137.0	17850.0	-6.7%	1.07x (?)

Code size: -swiftlibs

lorentey · 2025-05-16T04:53:10Z

Should you also update:

contains(_:) in the PartialRangeFrom extension?

This is not actually relevant to bounds checking, but sure, why not have them all transparent!

~=(_:_:) in the RangeExpression extension?

Oh yes indeed!

lorentey · 2025-05-17T01:42:57Z

@swift-ci smoke test

lorentey · 2025-05-19T19:53:31Z

@swift-ci smoke test

lorentey · 2025-05-20T17:20:39Z

Touching ~= was a mistake. I will take a look, but this is not worth all this, so I may end up reverting that part.

13:50:27 
 ******************** TEST 'Swift(linux-x86_64) :: Constraints/rdar105781521.swift' FAILED ********************
[...]
13:50:27  /home/build-user/swift/test/Constraints/rdar105781521.swift:13:17: error: unexpected note produced: overloads for '~=' exist with these partially matching parameter lists: (Substring, String)
13:50:27      case .first(true):
13:50:27                  ^
13:50:27  FAIL: Swift(linux-x86_64) :: Constraints/rdar105782480.swift (1298 of 19342)
13:50:27  /home/build-user/swift/test/Constraints/rdar105782480.swift:14:10: error: expected error not produced
13:50:27        // expected-error@-1 {{value of type 'MyEnum' has no member 'invalid'}}
13:50:27  ~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13:50:27  /home/build-user/swift/test/Constraints/rdar105782480.swift:15:30: error: incorrect message found
13:50:27        // expected-error@-2 {{'let' binding pattern cannot appear in an expression}}
13:50:27                               ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13:50:27                               type 'Range<MyEnum>' has no member 'second'
13:50:27  
13:50:27  --
13:50:27  
13:50:27  FAIL: Swift(linux-x86_64) :: Constraints/patterns.swift (1332 of 19342)
13:50:27  ******************** TEST 'Swift(linux-x86_64) :: Constraints/patterns.swift' FAILED ********************
13:50:27  /home/build-user/swift/test/Constraints/patterns.swift:125:7: error: unexpected error produced: reference to member 'HairForceOne' cannot be resolved without a contextual type
13:50:27  case .HairForceOne: // expected-error{{type 'any HairType' has no member 'HairForceOne'}}
13:50:27        ^
13:50:27  /home/build-user/swift/test/Constraints/patterns.swift:125:40: error: incorrect message found
13:50:27  case .HairForceOne: // expected-error{{type 'any HairType' has no member 'HairForceOne'}}
13:50:27                                         ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
13:50:27                                         binary operator '~=' cannot be applied to two 'any HairType' operands
13:50:27  FAIL: Swift(linux-x86_64) :: Constraints/result_builder_invalid_stmts.swift (1381 of 19342)
13:50:27  ******************** TEST 'Swift(linux-x86_64) :: Constraints/result_builder_invalid_stmts.swift' FAILED ********************
13:50:27  /home/build-user/swift/test/Constraints/result_builder_invalid_stmts.swift:17:10: error: unexpected note produced: overloads for '~=' exist with these partially matching parameter lists: (Substring, String)
13:50:27      case 1: // expected-error {{expression pattern of type 'Int' cannot match values of type 'String'}}
13:50:27           ^
13:50:27  /home/build-user/swift/test/Constraints/result_builder_invalid_stmts.swift:30:10: error: unexpected note produced: overloads for '~=' exist with these partially matching parameter lists: (Substring, String)
13:50:27      case 1: // expected-error {{expression pattern of type 'Int' cannot match values of type 'String'}}
13:50:27           ^
13:50:27  /home/build-user/swift/test/Constraints/result_builder_invalid_stmts.swift:44:10: error: unexpected note produced: overloads for '~=' exist with these partially matching parameter lists: (Substring, String)
13:50:27      case 1: // expected-error {{expression pattern of type 'Int' cannot match values of type 'String'}}
13:50:27           ^
13:50:27  /home/build-user/swift/test/Constraints/result_builder_invalid_stmts.swift:56:10: error: unexpected note produced: overloads for '~=' exist with these partially matching parameter lists: (Substring, String)
13:50:27      case 1: // expected-error {{expression pattern of type 'Int' cannot match values of type 'String'}}
13:50:27           ^
[etc]

lorentey · 2025-05-21T01:31:27Z

@swift-ci smoke test

lorentey · 2025-05-21T05:53:02Z

@swift-ci smoke test

lorentey · 2025-05-21T16:05:07Z

This looks like a fresh regression on main (unrelated).

23:56:39 
 ******************** TEST 'Swift(macosx-x86_64) :: SILOptimizer/performance-annotations.swift' FAILED ********************
23:56:39  in function specialized _ArrayBuffer._native.getter
23:56:39  /Users/ec2-user/jenkins/workspace/swift-PR-macos-smoke-test/branch-main/swift/test/SILOptimizer/performance-annotations.swift:38:11: error: unexpected error produced: this code performs reference counting operations which can cause locking
23:56:39    return a[i].x
23:56:39            ^
23:56:39  /Users/ec2-user/jenkins/workspace/swift-PR-macos-smoke-test/branch-main/swift/test/SILOptimizer/performance-annotations.swift:43:10: error: unexpected note produced: called from here
23:56:39    return noRTCallsForArrayGet(a, 0)
23:56:39           ^
23:56:39  
23:56:39  --
23:56:39  
23:56:39  ********************

lorentey · 2025-05-21T16:12:08Z

macOS runs look back in green this morning; Windows runs less so.

@swift-ci smoke test

lorentey · 2025-05-21T18:01:31Z

Now Linux is failing, too 🤡

10:17:20  UNRESOLVED: lldb-api :: lang/swift/completion/TestSwiftREPLCompletion.py (2 of 411)
10:17:20 
 ******************** TEST 'lldb-api :: lang/swift/completion/TestSwiftREPLCompletion.py' FAILED ********************

lorentey · 2025-05-22T04:09:05Z

@swift-ci smoke test

lorentey · 2025-05-22T23:45:02Z

This again; it looks like it may actually be triggered by this PR:

[2025-05-22T06:04:25.386Z] /Users/ec2-user/jenkins/workspace/swift-PR-macos-smoke-test/branch-main/swift/test/SILOptimizer/performance-annotations.swift:38:11: error: unexpected error produced: this code performs reference counting operations which can cause locking
[2025-05-22T06:04:25.386Z]   return a[i].x
[2025-05-22T06:04:25.386Z]           ^
[2025-05-22T06:04:25.386Z] /Users/ec2-user/jenkins/workspace/swift-PR-macos-smoke-test/branch-main/swift/test/SILOptimizer/performance-annotations.swift:43:10: error: unexpected note produced: called from here
[2025-05-22T06:04:25.386Z]   return noRTCallsForArrayGet(a, 0)
[2025-05-22T06:04:25.386Z]          ^

In my local builds, this keeps passing (on arm64):

PASS: Swift(macosx-arm64) :: SILOptimizer/performance-annotations.swift (4631 of 11741)

Array subscripts do have bounds checks; but it isn't clear to me how force-inlining Range.contains would trigger more retain/release traffic. 🤔 I guess I'll see if this reproduces on x86_64?

lorentey · 2025-05-22T23:55:48Z

Right, Array's subscript does not rely on Range.contains or ~= to validate the index; as far as I can tell, this PR does not touch that code.

lorentey · 2025-05-23T00:19:56Z

No, this doesn't locally reproduce on x86_64, either. I'll try rebasing next.

lorentey · 2025-05-23T00:20:08Z

Meanwhile,

@swift-ci smoke test macOS platform

lorentey · 2025-05-23T01:22:30Z

Nah, this still passes locally, even on current HEAD, on arm64 and x86_64 both.

lorentey · 2025-05-23T03:39:20Z

[2025-05-23T02:17:56.403Z] FAIL: Swift(macosx-x86_64) :: SILOptimizer/performance-annotations.swift (8067 of 19086)

¯\(ツ)/¯

These are really popular shorthands for implementing bounds checks, but currently they aren’t even marked `@inline(__always)`, so in larger functions they tend to remain outlined, defeating optimizations that would otherwise happen. The corresponding variants on the partial range types are `@_transparent`; that helps unoptimized performance, so let’s apply the same attribute here. rdar://151177326

…sparent is a bad idea

@benrimmington

Kudos to @benrimmington for spotting this

Making the generic ~= transparent does not help unoptimized code; adding a specialization does.

…en in unoptimized builds

…s-inlined instead The new overload messes up diagnostics too much.

lorentey · 2025-06-05T00:03:30Z

@swift-ci test

lorentey · 2025-06-05T22:06:23Z

@swift-ci smoke test

lorentey · 2025-06-05T22:08:19Z

I'm resorting to taking things out until the array lookup regression goes away. First up, a1e51a1 replaces @_transparent with @inline(__always), sacrificing debug performance.

lorentey · 2025-06-06T00:13:07Z

[2025-06-05T23:52:18.674Z] PASS: Swift(macosx-x86_64) :: SILOptimizer/performance-annotations.swift (8110 of 19207)

Cool; so evidently somehow Array.subscript does invoke one of these APIs. I'm guessing it might be ~=?

Making it transparent evidently induces new retain/release traffic in Array’s subscript, even though I can find no indication that `ClosedRange.contains` is ever called by that code path. Oh well.

lorentey · 2025-06-06T02:49:41Z

Right, macOS got a checkmark.

Of course, Linux would like some attention now:

[2025-06-06T00:24:46.474Z] ◇ Test run started.
[2025-06-06T00:24:46.474Z] ↳ Testing Library Version: 6.3-dev (ca81dffea6bd63d)
[2025-06-06T00:24:46.474Z] ↳ Target Platform: x86_64-unknown-linux-gnu
[2025-06-06T00:24:46.474Z] ◇ Suite "URL.Template Template" started.
[2025-06-06T00:24:46.474Z] ◇ Suite "AttributedString Index Tracking" started.
[2025-06-06T00:24:46.474Z] ◇ Suite "URL.Template Value" started.
[2025-06-06T00:24:46.474Z] ◇ Suite "URL.Template PercentEncoding" started.
[2025-06-06T00:24:46.474Z] ◇ Suite "URL.Template Expression" started.
[2025-06-06T00:24:46.474Z] ◇ Test stringRoundTrip(template:) started.
[2025-06-06T00:24:46.474Z] ◇ Test convertToNFCAndPercentEncode() started.
[2025-06-06T00:24:46.474Z] ◇ Test creating() started.
[2025-06-06T00:24:46.474Z] ◇ Test collapseRight() started.
[2025-06-06T00:24:46.474Z] ◇ Test insertionAtEmptyRange() started.
[2025-06-06T00:24:46.474Z] ◇ Test expressibleByLiteral() started.
[2025-06-06T00:24:46.474Z] ◇ Test normalizedAddingPercentEncoding_unreserved() started.
[2025-06-06T00:24:46.474Z] ◇ Test parsingWithMultipleNames() started.
[2025-06-06T02:28:53.851Z] Cancelling nested steps due to timeout
[2025-06-06T02:28:53.854Z] ◇ Test invalidInputRanges() startedSending interrupt signal to process
[2025-06-06T02:29:11.487Z] script returned exit code 143

lorentey · 2025-06-06T03:00:06Z

Further investigation indicates that the performance-annotations issue is caused by the @_transparent attribute on ClosedRange.contains(_: Element), but only if stdlib assertions are enabled, and only when the ObjC runtime is present.

This smells like we have a stray _internalInvariant on Array.subscript's bridging path somewhere that somehow touches ClosedRange, but for the life of me I cannot find it, or figure out how the compiler would think a transparent contains would induce retain/release traffic. Giving up is the right move at this point; this entry point shouldn't be on anyone's bounds checking hot path. I'll still make ClosedRange.contains force-inlined, but I leave marking it @_transparent to someone else.

lorentey · 2025-06-06T03:00:24Z

@swift-ci test

lorentey requested review from glessard, airspeedswift and Azoy May 12, 2025 22:15

lorentey requested a review from a team as a code owner May 12, 2025 22:15

Azoy approved these changes May 12, 2025

View reviewed changes

lorentey mentioned this pull request May 12, 2025

[6.2][stdlib] Mark contains methods on Range/ClosedRange transparent #81460

Closed

lorentey mentioned this pull request May 13, 2025

[benchmark] Add trivial benchmarks for [Closed]Range.contains(Bound) #81489

Closed

lorentey force-pushed the range-contains-perf branch from fb505b6 to ac58408 Compare May 17, 2025 01:42

lorentey force-pushed the range-contains-perf branch from ac58408 to bf01d07 Compare May 19, 2025 18:42

lorentey force-pushed the range-contains-perf branch from 415a190 to e9aaac8 Compare May 21, 2025 05:50

lorentey added 7 commits June 4, 2025 17:02

[stdlib] On second thought, making ClosedRange.contains(Range) tran…

d231213

…sparent is a bad idea

[stdlib] Make PartialRangeFrom.contains transparent

7cd41df

Kudos to @benrimmington for spotting this

[stdlib] Add a transparent specialization of ~= for Range

a59a6ca

Making the generic ~= transparent does not help unoptimized code; adding a specialization does.

[test] Verify that [Closed]Range.contains and Range.~= get inlined ev…

0826ac6

…en in unoptimized builds

[test] Add rudimentary test for transparent bounds checking helpers

a6ae23d

[stdlib] Delete transparent ~= overload; mark the regular one alway…

487baca

…s-inlined instead The new overload messes up diagnostics too much.

lorentey force-pushed the range-contains-perf branch from e9aaac8 to 487baca Compare June 5, 2025 00:03

[stdlib] ClosedRange.contains: Leave this force-inlined, not transparent

ff3cf98

Making it transparent evidently induces new retain/release traffic in Array’s subscript, even though I can find no indication that `ClosedRange.contains` is ever called by that code path. Oh well.

lorentey force-pushed the range-contains-perf branch from a1e51a1 to ff3cf98 Compare June 6, 2025 02:50

lorentey enabled auto-merge June 6, 2025 03:00

lorentey merged commit e473977 into swiftlang:main Jun 6, 2025
4 of 5 checks passed

[stdlib] Mark contains methods on Range/ClosedRange transparent #81458

[stdlib] Mark contains methods on Range/ClosedRange transparent #81458

Uh oh!

Conversation

lorentey commented May 12, 2025

Uh oh!

lorentey commented May 12, 2025

Uh oh!

lorentey commented May 12, 2025

Uh oh!

benrimmington commented May 13, 2025

Uh oh!

glessard commented May 13, 2025

Uh oh!

glessard commented May 13, 2025

Performance (x86_64): -O

Code size: -O

Performance (x86_64): -Osize

Code size: -Osize

Performance (x86_64): -Onone

Code size: -swiftlibs

Uh oh!

lorentey commented May 16, 2025

Uh oh!

lorentey commented May 17, 2025

Uh oh!

lorentey commented May 19, 2025

Uh oh!

lorentey commented May 20, 2025

Uh oh!

lorentey commented May 21, 2025

Uh oh!

lorentey commented May 21, 2025

Uh oh!

lorentey commented May 21, 2025

Uh oh!

lorentey commented May 21, 2025

Uh oh!

lorentey commented May 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lorentey commented May 22, 2025

Uh oh!

lorentey commented May 22, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lorentey commented May 22, 2025

Uh oh!

lorentey commented May 23, 2025

Uh oh!

lorentey commented May 23, 2025

Uh oh!

lorentey commented May 23, 2025

Uh oh!

lorentey commented May 23, 2025

Uh oh!

lorentey commented Jun 5, 2025

Uh oh!

lorentey commented Jun 5, 2025

Uh oh!

lorentey commented Jun 5, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lorentey commented Jun 6, 2025

Uh oh!

lorentey commented Jun 6, 2025

Uh oh!

lorentey commented Jun 6, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

lorentey commented Jun 6, 2025

Uh oh!

Uh oh!

Uh oh!

[stdlib] Mark `contains` methods on `Range/ClosedRange` transparent #81458

[stdlib] Mark `contains` methods on `Range/ClosedRange` transparent #81458

lorentey commented May 21, 2025 •

edited

Loading

lorentey commented May 22, 2025 •

edited

Loading

lorentey commented Jun 5, 2025 •

edited

Loading

lorentey commented Jun 6, 2025 •

edited

Loading