[SIL] Make owned function arguments lexical. #41228

nate-chandler · 2022-02-05T06:30:10Z

The destroys of owned arguments must not be hoisted over deinit barriers to respect the semantics of lexical lifetimes. Here, owned SILValues which are SILFunctionArguments are made to be lexical. Code which hoists destroys will check this field and not hoist destroys.

atrick · 2022-02-07T23:42:53Z

As discussed, we should fix ownership RAUW before doing this so most of the CSE/SILCombine optimizations will still work on owned arguments... although adjusting the tests to handle local variables instead was very clever and we may want to keep some of those new tests too

nate-chandler · 2022-02-08T04:07:03Z

@swift-ci please benchmark

nate-chandler · 2022-02-08T04:07:10Z

@swift-ci please test

swift-ci · 2022-02-08T07:21:16Z

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
Set.subtracting.Box.Empty	31	50	+61.3%	0.62x
Set.subtracting.Int.Empty	37	55	+48.6%	0.67x
Set.subtracting.Seq.Box.Empty	273	359	+31.5%	0.76x
Set.subtracting.Seq.Int.Empty	239	282	+18.0%	0.85x
NSStringConversion.MutableCopy.Rebridge.UTF8	738	843	+14.2%	0.88x (?)
DictionaryBridgeToObjC_Bridge	15	17	+13.3%	0.88x (?)
NSStringConversion.MutableCopy.UTF8	974	1084	+11.3%	0.90x (?)

Improvement	OLD	NEW	DELTA	RATIO
AnyHashableWithAClass	127500	96500	-24.3%	1.32x (?)
StringFromLongWholeSubstring	5	4	-20.0%	1.25x
Set.subtracting.Seq.Empty.Int	221	194	-12.2%	1.14x (?)
Set.subtracting.Empty.Int	44	40	-9.1%	1.10x (?)
String.replaceSubrange.RepChar.Small	515	469	-8.9%	1.10x (?)
EqualSubstringSubstring	42	39	-7.1%	1.08x (?)
EqualStringSubstring	42	39	-7.1%	1.08x (?)
EqualSubstringString	42	39	-7.1%	1.08x (?)
EqualSubstringSubstringGenericEquatable	43	40	-7.0%	1.07x (?)

Code size: -O

Regression	OLD	NEW	DELTA	RATIO
MonteCarloE.o	2817	2881	+2.3%	0.98x
FlattenList.o	3841	3921	+2.1%	0.98x
RangeAssignment.o	3275	3331	+1.7%	0.98x
RomanNumbers.o	6812	6924	+1.6%	0.98x
BucketSort.o	8551	8679	+1.5%	0.99x

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
Set.subtracting.Box.Empty	30	50	+66.7%	0.60x
Set.subtracting.Int.Empty	39	56	+43.6%	0.70x
Combos	417	488	+17.0%	0.85x
DictionaryBridgeToObjC_Bridge	15	17	+13.3%	0.88x (?)
ArrayAppendLatin1Substring	28512	30924	+8.5%	0.92x (?)
ArrayAppendUTF16Substring	27972	30312	+8.4%	0.92x (?)
ArrayAppendAsciiSubstring	27972	30276	+8.2%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
AnyHashableWithAClass	124000	96500	-22.2%	1.28x (?)
Set.isSubset.Seq.Int.Empty	231	195	-15.6%	1.18x
Set.isStrictSuperset.Seq.Empty.Int	230	198	-13.9%	1.16x
Set.isStrictSubset.Seq.Int.Empty	222	198	-10.8%	1.12x
DictionaryKeysContainsCocoa	25	23	-8.0%	1.09x (?)
String.replaceSubrange.RepChar.Small	514	476	-7.4%	1.08x (?)
LessSubstringSubstring	43	40	-7.0%	1.07x (?)
EqualSubstringSubstring	43	40	-7.0%	1.07x (?)
EqualStringSubstring	43	40	-7.0%	1.07x (?)
EqualSubstringSubstringGenericEquatable	43	40	-7.0%	1.07x (?)

Code size: -Osize

Regression	OLD	NEW	DELTA	RATIO
MonteCarloE.o	2797	2865	+2.4%	0.98x
RangeAssignment.o	3173	3241	+2.1%	0.98x
FlattenList.o	3744	3821	+2.1%	0.98x
ArraySetElement.o	1267	1287	+1.6%	0.98x
RomanNumbers.o	5149	5230	+1.6%	0.98x
COWArrayGuaranteedParameterOverhead.o	1327	1347	+1.5%	0.99x
Sim2DArray.o	1332	1352	+1.5%	0.99x
BucketSort.o	8385	8497	+1.3%	0.99x
Join.o	1580	1600	+1.3%	0.99x
ReversedCollections.o	7436	7521	+1.1%	0.99x
StringRemoveDupes.o	3720	3760	+1.1%	0.99x
Memset.o	1901	1921	+1.1%	0.99x
RangeOverlaps.o	6362	6428	+1.0%	0.99x
Array2D.o	2726	2754	+1.0%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
StrToInt.o	4633	4269	-7.9%	1.09x

Performance (x86_64): -Onone

Regression	OLD	NEW	DELTA	RATIO
RecursiveOwnedParameter	7430	8613	+15.9%	0.86x (?)
Dictionary2	1695	1900	+12.1%	0.89x (?)
Set.subtracting.Box.Empty	155	170	+9.7%	0.91x (?)
Set.subtracting.Int.Empty	166	182	+9.6%	0.91x (?)
NSStringConversion.Rebridge	501	540	+7.8%	0.93x (?)
StringBuilderLong	1940	2090	+7.7%	0.93x (?)
Hanoi	12950	13950	+7.7%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
AnyHashableWithAClass	165500	136000	-17.8%	1.22x (?)
ProtocolDispatch2	9924	9083	-8.5%	1.09x (?)
DictionaryKeysContainsNative	48	44	-8.3%	1.09x (?)
CharIndexing_korean_unicodeScalars_Backwards	187120	173480	-7.3%	1.08x (?)
CharIndexing_japanese_unicodeScalars_Backwards	223560	207400	-7.2%	1.08x (?)
CharIndexing_punctuated_unicodeScalars_Backwards	41200	38280	-7.1%	1.08x (?)
Set.isDisjoint.Seq.Int100	2435	2264	-7.0%	1.08x (?)

Code size: -swiftlibs

How to read the data

The tables contain differences in performance which are larger than 8% and differences in code size which are larger than 1%.

If you see any unexpected regressions, you should consider fixing the
regressions before you merge the PR.

Noise: Sometimes the performance results (not code size!) contain false
alarms. Unexpected regressions which are marked with '(?)' are probably noise.
If you see regressions which you cannot explain you can try to run the
benchmarks again. If regressions still show up, please consult with the
performance team (@eeckstein).

Hardware Overview

  Model Name: Mac Pro
  Model Identifier: MacPro6,1
  Processor Name: 12-Core Intel Xeon E5
  Processor Speed: 2.7 GHz
  Number of Processors: 1
  Total Number of Cores: 12
  L2 Cache (per Core): 256 KB
  L3 Cache: 30 MB
  Memory: 64 GB

nate-chandler · 2022-02-08T20:31:37Z

@swift-ci please benchmark

swift-ci · 2022-02-08T22:41:44Z

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
FlattenListFlatMap	3423	5746	+67.9%	0.60x (?)
Set.subtracting.Box.Empty	27	45	+66.7%	0.60x
FlattenListLoop	1479	2197	+48.5%	0.67x (?)
Set.subtracting.Int.Empty	33	49	+48.5%	0.67x
Set.subtracting.Seq.Box.Empty	245	326	+33.1%	0.75x (?)
ObjectiveCBridgeStubFromNSDateRef	3770	4520	+19.9%	0.83x (?)
Set.subtracting.Seq.Int.Empty	213	253	+18.8%	0.84x (?)
Combos	366	409	+11.7%	0.89x (?)
MapReduceShortString	12	13	+8.3%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
AnyHashableWithAClass	114000	87000	-23.7%	1.31x (?)
Set.subtracting.Seq.Empty.Int	197	174	-11.7%	1.13x (?)
String.replaceSubrange.RepChar.Small	463	422	-8.9%	1.10x (?)
DictionaryBridgeToObjC_Access	859	786	-8.5%	1.09x (?)
LessSubstringSubstring	38	35	-7.9%	1.09x (?)
EqualSubstringSubstring	38	35	-7.9%	1.09x (?)
EqualStringSubstring	38	35	-7.9%	1.09x (?)
EqualSubstringString	38	35	-7.9%	1.09x (?)
Set.subtracting.Empty.Int	39	36	-7.7%	1.08x (?)

Code size: -O

Regression	OLD	NEW	DELTA	RATIO
MonteCarloE.o	2817	2881	+2.3%	0.98x
FlattenList.o	3841	3921	+2.1%	0.98x
RangeAssignment.o	3275	3331	+1.7%	0.98x
RomanNumbers.o	6812	6924	+1.6%	0.98x
BucketSort.o	8551	8679	+1.5%	0.99x

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
Set.subtracting.Box.Empty	27	45	+66.7%	0.60x
Set.subtracting.Int.Empty	35	50	+42.9%	0.70x
Combos	377	437	+15.9%	0.86x (?)
ArrayAppendLatin1Substring	25560	27756	+8.6%	0.92x (?)
ArrayAppendUTF16Substring	25092	27180	+8.3%	0.92x (?)
ArrayAppendAsciiSubstring	25092	27144	+8.2%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
AnyHashableWithAClass	111500	86500	-22.4%	1.29x (?)
Set.isSubset.Seq.Int.Empty	207	175	-15.5%	1.18x (?)
Set.isStrictSuperset.Seq.Empty.Int	205	178	-13.2%	1.15x
Set.isStrictSubset.Seq.Int.Empty	200	178	-11.0%	1.12x (?)
String.replaceSubrange.RepChar.Small	464	423	-8.8%	1.10x (?)
Breadcrumbs.MutatedUTF16ToIdx.Mixed	258	238	-7.8%	1.08x (?)
ObjectiveCBridgeToNSDictionary	17250	16050	-7.0%	1.07x (?)
InsertCharacterEndIndex	159	148	-6.9%	1.07x (?)

Code size: -Osize

Regression	OLD	NEW	DELTA	RATIO
MonteCarloE.o	2797	2865	+2.4%	0.98x
RangeAssignment.o	3173	3241	+2.1%	0.98x
FlattenList.o	3744	3821	+2.1%	0.98x
ArraySetElement.o	1267	1287	+1.6%	0.98x
RomanNumbers.o	5149	5230	+1.6%	0.98x
COWArrayGuaranteedParameterOverhead.o	1327	1347	+1.5%	0.99x
Sim2DArray.o	1332	1352	+1.5%	0.99x
BucketSort.o	8385	8497	+1.3%	0.99x
Join.o	1580	1600	+1.3%	0.99x
ReversedCollections.o	7436	7521	+1.1%	0.99x
StringRemoveDupes.o	3720	3760	+1.1%	0.99x
Memset.o	1901	1921	+1.1%	0.99x
RangeOverlaps.o	6362	6428	+1.0%	0.99x
Array2D.o	2726	2754	+1.0%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
StrToInt.o	4633	4269	-7.9%	1.09x

Performance (x86_64): -Onone

Regression	OLD	NEW	DELTA	RATIO
RecursiveOwnedParameter	6677	7729	+15.8%	0.86x (?)
Set.subtracting.Box.Empty	139	153	+10.1%	0.91x (?)
Breadcrumbs.MutatedIdxToUTF16.ASCII	10	11	+10.0%	0.91x (?)
Set.subtracting.Int.Empty	149	163	+9.4%	0.91x (?)
UTF8Decode_InitFromCustom_contiguous	184	199	+8.2%	0.92x (?)
Hanoi	11630	12510	+7.6%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
FloatingPointPrinting_Double_description_small	26200	21100	-19.5%	1.24x (?)
AnyHashableWithAClass	147000	122500	-16.7%	1.20x (?)
DictionaryBridgeToObjC_Access	1192	1019	-14.5%	1.17x (?)
StringBuilderWithLongSubstring	4590	4030	-12.2%	1.14x (?)
DataAppendDataMediumToMedium	5080	4560	-10.2%	1.11x (?)
ProtocolDispatch2	8902	8146	-8.5%	1.09x (?)
StackPromo	54200	49800	-8.1%	1.09x (?)
CharIndexing_russian_unicodeScalars_Backwards	141240	129800	-8.1%	1.09x (?)
CharIndexing_chinese_unicodeScalars_Backwards	126480	117600	-7.0%	1.08x (?)
CharIndexing_japanese_unicodeScalars_Backwards	200600	187040	-6.8%	1.07x (?)
CharIndexing_punctuatedJapanese_unicodeScalars_Backwards	30960	28880	-6.7%	1.07x (?)

Code size: -swiftlibs

How to read the data

The tables contain differences in performance which are larger than 8% and differences in code size which are larger than 1%.

If you see any unexpected regressions, you should consider fixing the
regressions before you merge the PR.

Noise: Sometimes the performance results (not code size!) contain false
alarms. Unexpected regressions which are marked with '(?)' are probably noise.
If you see regressions which you cannot explain you can try to run the
benchmarks again. If regressions still show up, please consult with the
performance team (@eeckstein).

Hardware Overview

  Model Name: Mac Pro
  Model Identifier: MacPro6,1
  Processor Name: 8-Core Intel Xeon E5
  Processor Speed: 3 GHz
  Number of Processors: 1
  Total Number of Cores: 8
  L2 Cache (per Core): 256 KB
  L3 Cache: 25 MB
  Memory: 64 GB

nate-chandler · 2022-02-09T03:08:15Z

@swift-ci please test

eeckstein · 2022-02-09T07:22:55Z

@nate-chandler Can you please document that in SIL.rst? E.g. under "Function Types" and/or "Ownership SSA".
It would be nice to have a paragraph explaining lexical lifetimes in general (e.g. under "Ownership SSA").

swift-ci · 2022-02-09T08:16:54Z

Build failed
Swift Test OS X Platform
Git Sha - 6147b899b2116b5a67c57d9fed775d57e53c8aa4

nate-chandler · 2022-02-09T15:52:50Z

@eeckstein Absolutely! Sorry for leaving that out.

nate-chandler · 2022-02-09T23:43:08Z

Added some documentation to SIL.rst.

nate-chandler · 2022-02-09T23:43:14Z

@swift-ci please test

eeckstein

The documentation in SIL.rst looks very good!

Do you know the cause of the benchmark regressions?

docs/SIL.rst

lib/SILOptimizer/Utils/OwnershipOptUtils.cpp

UpdateSSA was being used to rewrite destroys, not for its intended purpose. Besides being a misuse, it's also wasteful. Instead, just rewrite the destroys that it would have updated.

lib/SILOptimizer/Utils/OwnershipOptUtils.cpp

RAUWing a lexical value with a non-lexical value is illegal because it would result in the value's lifetime being shortened without regard to deinit barriers. RAUWing _with_ a lexical value is LEGAL so long as doing so doesn't extend its lifetime.

The destroys of owned arguments must not be hoisted over deinit barriers to respect the semantics of lexical lifetimes. Here, owned SILValues which are SILFunctionArguments are made to be lexical. Code which hoists destroys will check this field and not hoist destroys. Allowed more RAUWing to be done with lexical values; specifically, allowed the replacement of a non-lexical value with a lexical value if the replacing value's lifetime wouldn't be extended as a result. Updated tests to get owned values via applies rather than having them passed as arguments.

nate-chandler · 2022-02-11T00:29:07Z

@swift-ci please test

nate-chandler · 2022-02-11T00:29:14Z

@swift-ci please benchmark

swift-ci · 2022-02-11T02:22:55Z

Performance (x86_64): -O

Regression	OLD	NEW	DELTA	RATIO
Set.subtracting.Box.Empty	23	37	+60.9%	0.62x
Set.subtracting.Int.Empty	28	43	+53.6%	0.65x
Set.subtracting.Seq.Box.Empty	210	274	+30.5%	0.77x (?)
Breadcrumbs.MutatedIdxToUTF16.Mixed	186	222	+19.4%	0.84x
Breadcrumbs.MutatedUTF16ToIdx.Mixed	184	219	+19.0%	0.84x
Set.subtracting.Seq.Int.Empty	184	212	+15.2%	0.87x
NormalizedIterator_nonBMPSlowestPrenormal	370	410	+10.8%	0.90x
DictionaryBridgeToObjC_Bridge	11	12	+9.1%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
LessSubstringSubstring	29	22	-24.1%	1.32x
EqualSubstringSubstring	29	22	-24.1%	1.32x
EqualStringSubstring	29	22	-24.1%	1.32x
EqualSubstringString	29	22	-24.1%	1.32x
EqualSubstringSubstringGenericEquatable	30	23	-23.3%	1.30x
LessSubstringSubstringGenericComparable	29	23	-20.7%	1.26x
ConvertFloatingPoint.MockFloat64Exactly2	16	14	-12.5%	1.14x
FlattenListLoop	1582	1385	-12.5%	1.14x (?)
StringComparison_longSharedPrefix	333	296	-11.1%	1.12x (?)
SortStringsUnicode	2215	2030	-8.4%	1.09x
BridgeString.find.native.longNonASCII	434	402	-7.4%	1.08x (?)
ObjectiveCBridgeStringHash	95	88	-7.4%	1.08x (?)
Set.isStrictSubset.Seq.Box0	286	266	-7.0%	1.08x
Diffing.Similar	376	350	-6.9%	1.07x (?)

Code size: -O

Regression	OLD	NEW	DELTA	RATIO
MonteCarloE.o	2817	2881	+2.3%	0.98x
FlattenList.o	3841	3921	+2.1%	0.98x
RangeAssignment.o	3275	3331	+1.7%	0.98x
RomanNumbers.o	6812	6924	+1.6%	0.98x
BucketSort.o	8551	8679	+1.5%	0.99x

Performance (x86_64): -Osize

Regression	OLD	NEW	DELTA	RATIO
Set.subtracting.Box.Empty	23	38	+65.2%	0.61x
Set.subtracting.Int.Empty	28	42	+50.0%	0.67x
DropFirstAnyCollection	84	124	+47.6%	0.68x (?)
PrefixAnySeqCRangeIterLazy	94	134	+42.6%	0.70x
PrefixAnySeqCntRangeLazy	94	134	+42.6%	0.70x
DropFirstAnySeqCRangeIterLazy	112	143	+27.7%	0.78x (?)
DropFirstAnySeqCntRangeLazy	112	143	+27.7%	0.78x (?)
PrefixWhileAnyCollection	142	178	+25.4%	0.80x
ArrayAppendRepeatCol	420	500	+19.0%	0.84x (?)
Breadcrumbs.MutatedIdxToUTF16.Mixed	187	222	+18.7%	0.84x
Breadcrumbs.MutatedUTF16ToIdx.Mixed	183	217	+18.6%	0.84x
DropWhileAnyCollectionLazy	148	174	+17.6%	0.85x (?)
DropLastAnyCollection	39	45	+15.4%	0.87x (?)
PrefixAnyCollection	111	127	+14.4%	0.87x
NormalizedIterator_fastPrenormal	430	480	+11.6%	0.90x (?)
FilterEvenUsingReduce	650	720	+10.8%	0.90x (?)
DropWhileAnySeqCntRangeLazy	142	157	+10.6%	0.90x (?)
DropWhileAnySeqCRangeIterLazy	142	156	+9.9%	0.91x (?)
DropWhileAnyCollection	122	134	+9.8%	0.91x (?)
NormalizedIterator_latin1	154	168	+9.1%	0.92x (?)
DictionaryBridgeToObjC_Bridge	11	12	+9.1%	0.92x (?)
OpenClose	58	63	+8.6%	0.92x (?)
DataReplaceMedium	2600	2800	+7.7%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
FlattenListLoop	1449	1039	-28.3%	1.39x (?)
LessSubstringSubstring	30	22	-26.7%	1.36x
EqualSubstringSubstringGenericEquatable	30	22	-26.7%	1.36x
EqualSubstringString	30	22	-26.7%	1.36x
LessSubstringSubstringGenericComparable	29	22	-24.1%	1.32x
EqualStringSubstring	30	23	-23.3%	1.30x
RecursiveOwnedParameter	284	221	-22.2%	1.29x (?)
EqualSubstringSubstring	29	23	-20.7%	1.26x
SequenceAlgosArray	2140	1710	-20.1%	1.25x
UTF8Decode_InitFromCustom_noncontiguous_ascii	688	569	-17.3%	1.21x (?)
ArrayAppendUTF16Substring	22032	18792	-14.7%	1.17x (?)
ArrayAppendLatin1Substring	22464	19188	-14.6%	1.17x (?)
ArrayAppendAsciiSubstring	22032	18828	-14.5%	1.17x (?)
UTF8Decode_InitFromCustom_noncontiguous_ascii_as_ascii	782	670	-14.3%	1.17x
StringComparison_longSharedPrefix	333	295	-11.4%	1.13x (?)
UTF8Decode_InitFromCustom_noncontiguous	283	253	-10.6%	1.12x (?)
ObjectiveCBridgeStringHash	95	87	-8.4%	1.09x
SortStringsUnicode	2210	2035	-7.9%	1.09x (?)
BridgeString.find.native.longNonASCII	433	402	-7.2%	1.08x (?)

Code size: -Osize

Regression	OLD	NEW	DELTA	RATIO
MonteCarloE.o	2797	2865	+2.4%	0.98x
RangeAssignment.o	3173	3241	+2.1%	0.98x
FlattenList.o	3744	3821	+2.1%	0.98x
ArraySetElement.o	1267	1287	+1.6%	0.98x
RomanNumbers.o	5149	5230	+1.6%	0.98x
COWArrayGuaranteedParameterOverhead.o	1327	1347	+1.5%	0.99x
Sim2DArray.o	1332	1352	+1.5%	0.99x
BucketSort.o	8385	8497	+1.3%	0.99x
Join.o	1580	1600	+1.3%	0.99x
ReversedCollections.o	7436	7521	+1.1%	0.99x
StringRemoveDupes.o	3720	3760	+1.1%	0.99x
Memset.o	1901	1921	+1.1%	0.99x
RangeOverlaps.o	6362	6428	+1.0%	0.99x
Array2D.o	2726	2754	+1.0%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
StrToInt.o	4633	4269	-7.9%	1.09x

Performance (x86_64): -Onone

Regression	OLD	NEW	DELTA	RATIO
Breadcrumbs.MutatedIdxToUTF16.Mixed	195	231	+18.5%	0.84x
Breadcrumbs.MutatedUTF16ToIdx.Mixed	186	220	+18.3%	0.85x
Set.subtracting.Box.Empty	99	112	+13.1%	0.88x (?)
COWArrayGuaranteedParameterOverhead	5350	6050	+13.1%	0.88x (?)
Set.subtracting.Int.Empty	110	122	+10.9%	0.90x (?)
Set.subtracting.Seq.Int.Empty	571	629	+10.2%	0.91x (?)
ArrayValueProp	3027	3283	+8.5%	0.92x (?)
CharIteration_tweet_unicodeScalars	94480	102360	+8.3%	0.92x (?)
Set.subtracting.Seq.Box.Empty	592	641	+8.3%	0.92x (?)
ArrayAppendGenericStructs	610	660	+8.2%	0.92x (?)
ArrayLiteral2	1746	1889	+8.2%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
EqualSubstringSubstring	62	53	-14.5%	1.17x (?)
EqualSubstringString	62	53	-14.5%	1.17x (?)
LessSubstringSubstringGenericComparable	60	52	-13.3%	1.15x
LessSubstringSubstring	62	54	-12.9%	1.15x
EqualStringSubstring	62	54	-12.9%	1.15x (?)
EqualSubstringSubstringGenericEquatable	60	53	-11.7%	1.13x
ArrayOfPOD	735	669	-9.0%	1.10x (?)
DataReplaceLargeBuffer	12	11	-8.3%	1.09x (?)
String.replaceSubrange.ArrChar.Small	51	47	-7.8%	1.09x (?)
ObjectiveCBridgeStringHash	95	88	-7.4%	1.08x (?)
BridgeString.find.native.longNonASCII	433	402	-7.2%	1.08x (?)

Code size: -swiftlibs

How to read the data

The tables contain differences in performance which are larger than 8% and differences in code size which are larger than 1%.

If you see any unexpected regressions, you should consider fixing the
regressions before you merge the PR.

Noise: Sometimes the performance results (not code size!) contain false
alarms. Unexpected regressions which are marked with '(?)' are probably noise.
If you see regressions which you cannot explain you can try to run the
benchmarks again. If regressions still show up, please consult with the
performance team (@eeckstein).

Hardware Overview

  Model Name: Mac mini
  Model Identifier: Macmini8,1
  Processor Name: 6-Core Intel Core i7
  Processor Speed: 3.2 GHz
  Number of Processors: 1
  Total Number of Cores: 6
  L2 Cache (per Core): 256 KB
  L3 Cache: 12 MB
  Memory: 64 GB

nate-chandler · 2022-02-11T03:32:45Z

@swift-ci please test windows platform

atrick

Great!

nate-chandler · 2022-02-12T02:11:48Z

The root cause of the Set.subtracting.Box.Empty benchmark regression is that we don't yet have the lexical destroy hoisting utility to run during copy propagation on owned lexical values (move_value [lexical] and @owned arguments). That results in extra ARC traffic in $sSh11subtractingyShyxGABF.

In detail, the problem is that CanonicalizeOSSALifetime, which doesn't respect lexical lifetimes, no longer hoists the destroys of %1 of

sil [serialized] [ossa] @$sSh11subtractingyShyxGABF : $@convention(method) <Element where Element : Hashable> (@guaranteed Set<Element>, @owned Set<Element>) -> @owned Set<Element> {
bb0(%0 : @guaranteed $Set<Element>, %1 : @owned $Set<Element>):

to just after its last uses. Specifically, the destroy is not hoisted up from the end of the function to just after the apply of $sSh12_subtractingyShyxGqd__7ElementQyd__RszSTRd__lF.

Before this change, SILCombine saw

  %102 = copy_value %1 : $Set<Element>
  %105 = function_ref @$sSh12_subtractingyShyxGqd__7ElementQyd__RszSTRd__lF : $@convention(method) <τ_0_0 where τ_0_0 : Hashable><τ_1_0 where τ_0_0 == τ_1_0.Element, τ_1_0 : Sequence> (@in_guaranteed τ_1_0, @owned Set<τ_0_0>) -> @owned Set<τ_0_0>
  %106 = apply %105<Element, Set<Element>>(%103, %102) : $@convention(method) <τ_0_0 where τ_0_0 : Hashable><τ_1_0 where τ_0_0 == τ_1_0.Element, τ_1_0 : Sequence> (@in_guaranteed τ_1_0, @owned Set<τ_0_0>) -> @owned Set<τ_0_0>
  destroy_value %1 : $Set<Element>

and was able to rewrite it as

  %85 = function_ref @$sSh12_subtractingyShyxGqd__7ElementQyd__RszSTRd__lF : $@convention(method) <τ_0_0 where τ_0_0 : Hashable><τ_1_0 where τ_0_0 == τ_1_0.Element, τ_1_0 : Sequence> (@in_guaranteed τ_1_0, @owned Set<τ_0_0>) -> @owned Set<τ_0_0>
  %86 = apply %85<Element, Set<Element>>(%83, %1) : $@convention(method) <τ_0_0 where τ_0_0 : Hashable><τ_1_0 where τ_0_0 == τ_1_0.Element, τ_1_0 : Sequence> (@in_guaranteed τ_1_0, @owned Set<τ_0_0>) -> @owned Set<τ_0_0>

Now, though, the destroy_value isn't in the same block, so the extra copy can't be eliminated.

eeckstein · 2022-02-14T08:16:10Z

Is this something which is easy to fix?

nate-chandler · 2022-02-14T15:57:30Z

It should be reasonably straightforward. ShrinkBorrowScope has been restructured ( #41358 ) to make it easy to create the new LexicalDestroyHoisting from it. Filed rdar://88903962 .

nate-chandler marked this pull request as draft February 5, 2022 18:45

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch 5 times, most recently from d0dac47 to 97df9a3 Compare February 7, 2022 20:20

nate-chandler marked this pull request as ready for review February 7, 2022 20:20

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch from 97df9a3 to 959aeb6 Compare February 7, 2022 20:28

nate-chandler requested a review from atrick February 7, 2022 20:32

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch from 959aeb6 to f007b8b Compare February 8, 2022 03:35

swiftlang deleted a comment from swift-ci Feb 8, 2022

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch 2 times, most recently from 3827ab7 to e8ff4ae Compare February 8, 2022 03:59

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch from e8ff4ae to 99f9f57 Compare February 8, 2022 20:31

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch 2 times, most recently from 11b4548 to 6147b89 Compare February 9, 2022 02:27

eeckstein reviewed Feb 10, 2022

View reviewed changes

docs/SIL.rst Show resolved Hide resolved

atrick reviewed Feb 10, 2022

View reviewed changes

docs/SIL.rst Outdated Show resolved Hide resolved

atrick reviewed Feb 10, 2022

View reviewed changes

docs/SIL.rst Outdated Show resolved Hide resolved

atrick requested changes Feb 10, 2022

View reviewed changes

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch from 2fbfbe6 to cade25f Compare February 10, 2022 20:04

nate-chandler requested a review from atrick February 10, 2022 20:19

nate-chandler added 3 commits February 10, 2022 15:21

[LexicalDestroyHoisting] Eliminate updateSSA.

9af40b3

UpdateSSA was being used to rewrite destroys, not for its intended purpose. Besides being a misuse, it's also wasteful. Instead, just rewrite the destroys that it would have updated.

[LexicalDestroyHoisting] Mark context const.

93b3ffb

[LexicalDestroyFolding] Used ArrayRef.

49044d0

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch 2 times, most recently from b95785f to 6a99b7e Compare February 10, 2022 23:27

atrick reviewed Feb 11, 2022

View reviewed changes

lib/SILOptimizer/Utils/OwnershipOptUtils.cpp Outdated Show resolved Hide resolved

nate-chandler added 5 commits February 10, 2022 16:28

[OwnershipUtils] Added find for simple value.

d23139f

[Gardening] Tweaked comment.

2aa70af

[SIL] Doc'd deinit barriers.

2da6ac1

nate-chandler force-pushed the lexical_lifetimes/owned_function_args_are_lexical branch from 6a99b7e to 2da6ac1 Compare February 11, 2022 00:28

atrick approved these changes Feb 11, 2022

View reviewed changes

nate-chandler merged commit 63161e6 into swiftlang:main Feb 12, 2022

nate-chandler deleted the lexical_lifetimes/owned_function_args_are_lexical branch February 12, 2022 02:12

[SIL] Make owned function arguments lexical. #41228

[SIL] Make owned function arguments lexical. #41228

Uh oh!

Conversation

nate-chandler commented Feb 5, 2022

Uh oh!

atrick commented Feb 7, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

nate-chandler commented Feb 8, 2022

Uh oh!

nate-chandler commented Feb 8, 2022

Uh oh!

swift-ci commented Feb 8, 2022

Performance (x86_64): -O

Code size: -O

Performance (x86_64): -Osize

Code size: -Osize

Performance (x86_64): -Onone

Code size: -swiftlibs

Uh oh!

nate-chandler commented Feb 8, 2022

Uh oh!

swift-ci commented Feb 8, 2022

Performance (x86_64): -O

Code size: -O

Performance (x86_64): -Osize

Code size: -Osize

Performance (x86_64): -Onone

Code size: -swiftlibs

Uh oh!

nate-chandler commented Feb 9, 2022

Uh oh!

eeckstein commented Feb 9, 2022

Uh oh!

swift-ci commented Feb 9, 2022

Uh oh!

nate-chandler commented Feb 9, 2022

Uh oh!

nate-chandler commented Feb 9, 2022

Uh oh!

nate-chandler commented Feb 9, 2022

Uh oh!

eeckstein left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

nate-chandler commented Feb 11, 2022

Uh oh!

nate-chandler commented Feb 11, 2022

Uh oh!

swift-ci commented Feb 11, 2022

Performance (x86_64): -O

Code size: -O

Performance (x86_64): -Osize

Code size: -Osize

Performance (x86_64): -Onone

Code size: -swiftlibs

Uh oh!

nate-chandler commented Feb 11, 2022

Uh oh!

atrick left a comment

Choose a reason for hiding this comment

Uh oh!

nate-chandler commented Feb 12, 2022

Uh oh!

eeckstein commented Feb 14, 2022

Uh oh!

nate-chandler commented Feb 14, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

atrick commented Feb 7, 2022 •

edited

Loading

eeckstein left a comment •

edited

Loading

nate-chandler commented Feb 14, 2022 •

edited

Loading