[stdlib] Refactor Unicode normalization

[karwa]

Refactors the standard library's internal normalization routines to make them independent of using String as a storage type.

This doesn't expose any public APIs or ABI symbols (it's all still just internal, and not @usableFromInline), and should be performance-neutral.

[karwa]

@swift-ci please benchmark

[Azoy]

@swift-ci please benchmark

[karwa]

Hmm, does seem to find a regression with zalgo (so, lots of combining characters), but checking my local results, when I ran it, it was okay. Hmm....

main: {"number": 973, "name": "StringComparison_zalgo", "num_samples": 200, "min": 26000.0, "median": 26150.0, "max": 27525.0}
this: {"number": 973, "name": "StringComparison_zalgo", "num_samples": 200, "min": 25925.0, "median": 26075.0, "max": 28300.0}

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
StringComparison_zalgo	49850.0	87350.0	+75.2%	0.57x
StringHashing_zalgo	2678.846	4522.222	+68.8%	0.59x
NormalizedIterator_zalgo	2808.333	4593.75	+63.6%	0.61x
Improvement	OLD	NEW	DELTA	RATIO
FlattenListLoop	1362.0	976.5	-28.3%	1.39x (?)
Array.removeAll.keepingCapacity.Object	7.375	5.775	-21.7%	1.28x (?)
StringHashing_abnormal	610.959	562.933	-7.9%	1.09x (?)
ObserverForwarderStruct	621.667	573.462	-7.8%	1.08x (?)

Code size: -O

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
StringComparison_zalgo	49800.0	87025.0	+74.7%	0.57x
StringHashing_zalgo	2670.0	4516.667	+69.2%	0.59x
NormalizedIterator_zalgo	2751.25	4610.417	+67.6%	0.60x
StringBuilder	233.143	253.75	+8.8%	0.92x (?)
Improvement	OLD	NEW	DELTA	RATIO
ArrayAppendGenericStructs	1765.0	1402.5	-20.5%	1.26x (?)
StringDistance.utf16.ascii	151.0	138.882	-8.0%	1.09x (?)
StringHashing_abnormal	610.588	562.857	-7.8%	1.08x (?)

Code size: -Osize

Performance (x86_64): -Onone

Regression	OLD	NEW	DELTA	RATIO
StringComparison_zalgo	50625.0	88075.0	+74.0%	0.57x
StringHashing_zalgo	2725.0	4579.167	+68.0%	0.60x
NormalizedIterator_zalgo	2781.579	4637.5	+66.7%	0.60x
ArrayAppendGenericStructs	1145.0	1437.5	+25.5%	0.80x (?)
NSStringConversion.InlineBuffer.ASCII	5071.0	5926.0	+16.9%	0.86x (?)
NSStringConversion.InlineBuffer.UTF8	3139.0	3564.0	+13.5%	0.88x (?)
StringWithCString2	0.009	0.01	+10.0%	0.91x (?)
Improvement	OLD	NEW	DELTA	RATIO
Dictionary4	687.5	630.0	-8.4%	1.09x (?)
CharacterLiteralsLarge	352.0	324.667	-7.8%	1.08x (?)
StringHashing_abnormal	686.774	639.063	-6.9%	1.07x (?)

[karwa]

@swift-ci please benchmark Apple silicon

[Azoy]

@swift-ci please benchmark Apple silicon

[karwa]

@swift-ci please benchmark

[karwa]

heeeey it worked, I can do it myself now :)

[karwa]

Okay that seems to have cleared it up. I'll take a look at things like LessSubstringSubstring, but most of the changes are now indicated as "dubious", which the benchmark script defines thusly:

# Indication of dubious changes: when result's MIN falls inside the
# (MIN, MAX) interval of result they are being compared with.

(as I understand it: their error-bars overlap). So overall I think results are fairly neutral, but I'll take another look anyway.

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
Array.removeAll.keepingCapacity.Int	0.077	0.102	+32.1%	0.76x (?)
Data.hash.Medium	24.259	30.483	+25.7%	0.80x (?)
LessSubstringSubstring	24.279	29.613	+22.0%	0.82x (?)
EqualSubstringString	24.382	29.719	+21.9%	0.82x (?)
LessSubstringSubstringGenericComparable	24.486	29.829	+21.8%	0.82x
EqualSubstringSubstring	24.583	29.926	+21.7%	0.82x (?)
EqualSubstringSubstringGenericEquatable	24.719	29.733	+20.3%	0.83x (?)
EqualStringSubstring	24.808	29.708	+19.8%	0.84x (?)
Array.removeAll.keepingCapacity.Object	6.057	7.253	+19.7%	0.84x (?)
StringComparison_longSharedPrefix	206.909	240.7	+16.3%	0.86x (?)
FlattenListFlatMap	3299.0	3821.0	+15.8%	0.86x (?)
NormalizedIterator_fastPrenormal	468.654	533.913	+13.9%	0.88x (?)
NormalizedIterator_emoji	316.0	349.63	+10.6%	0.90x (?)
NaiveRRC.init.largeContiguous	5.451	6.028	+10.6%	0.90x (?)
DropLastSequenceLazy	297.0	328.143	+10.5%	0.91x (?)
DropLastSequence	298.333	328.333	+10.1%	0.91x (?)
DropLastAnySequence	299.286	327.833	+9.5%	0.91x (?)
StringComparison_abnormal	945.882	1026.383	+8.5%	0.92x (?)
NormalizedIterator_latin1	170.778	183.7	+7.6%	0.93x (?)
Improvement	OLD	NEW	DELTA	RATIO
StringComparison_zalgo	49525.0	40425.0	-18.4%	1.23x (?)
NormalizedIterator_zalgo	2757.143	2393.182	-13.2%	1.15x
StringHashing_zalgo	2658.333	2312.5	-13.0%	1.15x (?)
StringBuilderLong	970.625	848.235	-12.6%	1.14x (?)
SubstringEquatable	335.167	297.0	-11.4%	1.13x (?)
SubstringEqualString	174.8	156.7	-10.4%	1.12x
StringBuilderWithLongSubstring	1135.385	1021.818	-10.0%	1.11x (?)
DropWhileAnySequenceLazy	412.25	378.0	-8.3%	1.09x (?)
Set.isDisjoint.Int.Empty	52.077	47.917	-8.0%	1.09x (?)
Set.isSuperset.Seq.Empty.Int	50.862	47.388	-6.8%	1.07x (?)

Code size: -O

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
FlattenListLoop	1031.0	1637.0	+58.8%	0.63x (?)
Array.removeAll.keepingCapacity.Object	5.62	7.198	+28.1%	0.78x (?)
Data.hash.Medium	24.258	30.483	+25.7%	0.80x (?)
EqualSubstringSubstringGenericEquatable	24.279	29.8	+22.7%	0.81x (?)
EqualSubstringSubstring	24.281	29.72	+22.4%	0.82x (?)
EqualStringSubstring	24.269	29.6	+22.0%	0.82x (?)
EqualSubstringString	24.816	30.16	+21.5%	0.82x (?)
LessSubstringSubstringGenericComparable	24.93	30.281	+21.5%	0.82x
LessSubstringSubstring	25.029	30.258	+20.9%	0.83x (?)
StringComparison_longSharedPrefix	202.8	238.3	+17.5%	0.85x (?)
Dictionary4	215.571	251.0	+16.4%	0.86x (?)
ArrayInClass	175.0	200.484	+14.6%	0.87x (?)
DataReplaceMedium	2818.605	3095.313	+9.8%	0.91x (?)
PrefixWhileSequence	200.778	219.667	+9.4%	0.91x (?)
PrefixWhileAnySequence	200.667	219.25	+9.3%	0.92x (?)
RemoveWhereFilterInts	22.455	24.359	+8.5%	0.92x (?)
StringComparison_abnormal	946.8	1025.217	+8.3%	0.92x (?)
Improvement	OLD	NEW	DELTA	RATIO
StringComparison_zalgo	49775.0	40325.0	-19.0%	1.23x (?)
StringHashing_zalgo	2665.0	2298.438	-13.8%	1.16x
NormalizedIterator_zalgo	2748.81	2400.0	-12.7%	1.15x
SubstringEquatable	346.0	303.833	-12.2%	1.14x (?)
LineSink.characters.alpha	74.576	66.167	-11.3%	1.13x (?)
SubstringEqualString	179.1	159.545	-10.9%	1.12x
Set.isSubset.Int.Empty	55.0	50.1	-8.9%	1.10x (?)
DropFirstAnySequence	1080.5	987.0	-8.7%	1.09x (?)
String.initRepeating.longMixStr.Count100	725.667	676.0	-6.8%	1.07x (?)

Code size: -Osize

Performance (x86_64): -Onone

Regression	OLD	NEW	DELTA	RATIO
CxxStringConversion.cxx.to.swift	115.5	154.0	+33.3%	0.75x (?)
ArrayAppendGenericStructs	1387.5	1785.0	+28.6%	0.78x (?)
LessSubstringSubstringGenericComparable	27.224	32.667	+20.0%	0.83x (?)
EqualSubstringSubstringGenericEquatable	27.654	33.083	+19.6%	0.84x (?)
Data.hash.Medium	28.667	34.211	+19.3%	0.84x (?)
LessSubstringSubstring	28.381	33.862	+19.3%	0.84x (?)
EqualSubstringString	28.625	34.071	+19.0%	0.84x (?)
EqualStringSubstring	28.957	34.182	+18.0%	0.85x
Set.isSubset.Seq.Int50	1369.0	1585.0	+15.8%	0.86x (?)
EqualSubstringSubstring	29.4	34.034	+15.8%	0.86x (?)
Set.isSubset.Seq.Int25	685.667	793.667	+15.8%	0.86x (?)
Set.isSubset.Seq.Int100	2778.0	3165.0	+13.9%	0.88x (?)
SIMDReduce.Int8	6525.0	7175.0	+10.0%	0.91x (?)
Improvement	OLD	NEW	DELTA	RATIO
StringComparison_zalgo	50675.0	41350.0	-18.4%	1.23x (?)
StringHashing_zalgo	2718.333	2371.875	-12.7%	1.15x (?)
NormalizedIterator_zalgo	2763.333	2426.563	-12.2%	1.14x (?)
Set.subtracting.Seq.Empty.Int	353.667	321.714	-9.0%	1.10x (?)
Dictionary4	692.0	630.667	-8.9%	1.10x (?)
String.initRepeating.longMixStr.Count100	725.5	677.0	-6.7%	1.07x (?)

[Azoy]

@swift-ci please benchmark

[karwa]

@swift-ci please benchmark apple silicon

@shahmishal is there an issue with that command? Neither of my attempts actually got enqueued, and neither was Alejandro’s.

[karwa]

@Azoy After investigating, it appears EqualSubstringString and friends are just flaky. Firstly, even though it sounds normalisation-related, they don't actually hit the normalising slow-path. Secondly, lots of other PRs show the same kind of sporadic results for these benchmarks.

Example: An Array PR https://ci.swift.org/view/Pull%20Requests/job/swift-PR-macos-perf/920/

Benchmarks showed the same regressions as here:

LessSubstringSubstring                    | 24.061    | 29.719    | +23.5% | **0.81x (?)**
EqualSubstringSubstringGenericEquatable   | 24.506    | 30.156    | +23.1% | **0.81x (?)**
LessSubstringSubstringGenericComparable   | 24.382    | 29.938    | +22.8% | **0.81x (?)**
EqualSubstringSubstring                   | 24.51     | 30.053    | +22.6% | **0.82x**
EqualSubstringString                      | 24.4      | 29.842    | +22.3% | **0.82x (?)**
UTF8Decode_InitFromCustom_contiguous      | 143.733   | 175.545   | +22.1% | **0.82x (?)**
EqualStringSubstring                      | 24.806    | 30.269    | +22.0% | **0.82x**

Example 2: Optimised isASCII check https://ci.swift.org/view/Pull%20Requests/job/swift-PR-macos-perf/899/

Same thing:

EqualSubstringSubstring                    | 24.5    | 30.25     | +23.5% | **0.81x (?)**
EqualSubstringString                       | 24.279  | 29.839    | +22.9% | **0.81x (?)**
EqualStringSubstring                       | 24.385  | 29.96     | +22.9% | **0.81x (?)**
LessSubstringSubstring                     | 24.292  | 29.833    | +22.8% | **0.81x**
LessSubstringSubstringGenericComparable    | 24.594  | 30.038    | +22.1% | **0.82x (?)**
EqualSubstringSubstringGenericEquatable    | 24.5    | 29.84     | +21.8% | **0.82x (?)**
UTF8Decode_InitFromData                    | 149.417 | 175.727   | +17.6% | **0.85x (?)**
UTF8Decode_InitFromBytes                   | 153.818 | 179.75    | +16.9% | **0.86x (?)**
StringComparison_longSharedPrefix          | 211.444 | 242.5     | +14.7% | **0.87x (?)**
NormalizedIterator_fastPrenormal           | 516.429 | 579.75    | +12.3% | **0.89x (?)**

The zalgo thing was a much clearer indication of a performance issue from suboptimal inlining. If all we have left are some dubious results from known-flaky benchmarks (with the same kind of numbers, too), I don't think that's any indication of a regression.

[karwa]

Also, can you share any info about why the string normalisation tests are disabled? They work for me locally if I remove the REQUIRES lines.

[karwa]

@swift-ci please smoke test

[Azoy]

Looks good to me. I don't see anything here that affects the actual normalization logic, so the stream work looks pretty trivial on top of it.

[milseman]

@swift-ci please benchmark

[milseman]

@swift-ci please test macOS platform

[karwa]

@milseman Are you okay with me merging this?