[DO NOT MERGE] More experiments with UnsafeBufferPointer.init #35075

lorentey · 2020-12-12T02:35:25Z

This is a spinoff of #34961 that reintroduces a _precondition for count >= 0; that particular static invariant may have a beneficial effect on overall performance.

Every other Unsafe(Mutable)BufferPointer precondition in this file is debug-mode only. This information is already part of the documentation for this function, and it's really hard to get rid of these branches and traps otherwise.

…uilds

…ckedBounds:) Unchecked APIs must still perform checks in debug builds to ensure that invariants aren’t violated. (I.e., these aren’t a license to perform invalid operations — they just let us get rid of the checks when we know for sure they are unnecessary.)

`Slice` does not technically guarantee that its indices are valid in its base, so these initializers accidentally allowed the creation of obviously out-of-bounds buffers.

…nnegative

This eliminates a couple of _debugPreconditions which seem to change inlining enough to interfere with some benchmarks.

When the stdlib says not to check, it’s a good idea to actually not have any checking, so that we leave existing code paths unchanged. (And because we do trust that the stdlib is responsible about such things.)

…nversions The UnboundRange → Range conversion path is complicated and it involves at least one unnecessary _precondition check for startIndex ..< endIndex.

…condition

… test failures

lorentey · 2020-12-12T02:35:35Z

@swift-ci benchmark

swift-ci · 2020-12-12T03:08:17Z

Performance: -O

Regression	OLD	NEW	DELTA	RATIO
ObjectiveCBridgeStubDataAppend	2360	3020	+28.0%	0.78x
ObjectiveCBridgeStringHash	99	118	+19.2%	0.84x
StringHasPrefixAscii	1760	2070	+17.6%	0.85x
Data.append.Sequence.64kB.Count0.RE	249	288	+15.7%	0.86x
Data.append.Sequence.64kB.Count0.RE.I	249	285	+14.5%	0.87x
Data.init.Sequence.64kB.Count0.RE	249	285	+14.5%	0.87x
Data.append.Sequence.809B.Count0.RE	348	398	+14.4%	0.87x (?)
Data.append.Sequence.64kB.Count0.I	248	283	+14.1%	0.88x
Data.init.Sequence.64kB.Count0.RE.I	250	285	+14.0%	0.88x
Data.append.Sequence.809B.Count0.RE.I	350	397	+13.4%	0.88x
Data.init.Sequence.64kB.Count0	248	281	+13.3%	0.88x
Data.append.Sequence.64kB.Count0	249	282	+13.3%	0.88x
StringHasSuffixAscii	1970	2230	+13.2%	0.88x
Data.append.Sequence.809B.Count0	345	390	+13.0%	0.88x (?)
Data.init.Sequence.64kB.Count0.I	249	281	+12.9%	0.89x
SubstringEqualString	365	411	+12.6%	0.89x
Data.append.Sequence.809B.Count0.I	345	388	+12.5%	0.89x (?)
DataToStringLargeUnicode	6250	7000	+12.0%	0.89x (?)
Data.init.Sequence.2049B.Count0.I	453	507	+11.9%	0.89x
Data.init.Sequence.2047B.Count0.I	453	506	+11.7%	0.90x (?)
ObjectiveCBridgeStringGetASCIIContents	396	440	+11.1%	0.90x (?)
String.data.LargeUnicode	119	132	+10.9%	0.90x (?)
Data.init.Sequence.809B.Count0.RE	406	449	+10.6%	0.90x (?)
Data.init.Sequence.809B.Count0.RE.I	406	448	+10.3%	0.91x (?)
InsertCharacterEndIndex	147	162	+10.2%	0.91x (?)
Data.init.Sequence.809B.Count0.I	407	447	+9.8%	0.91x (?)
ObjectiveCBridgeStubFromNSDateRef	4030	4420	+9.7%	0.91x (?)
Breadcrumbs.IdxToUTF16Range.longASCII	31	34	+9.7%	0.91x (?)
ObjectiveCBridgeStubFromNSDate	4550	4980	+9.5%	0.91x (?)
Data.init.Sequence.511B.Count0.I	424	464	+9.4%	0.91x (?)
Data.init.Sequence.513B.Count0.I	424	464	+9.4%	0.91x (?)
StringToDataLargeUnicode	3200	3500	+9.4%	0.91x (?)
Data.init.Sequence.809B.Count0	407	445	+9.3%	0.91x (?)
DataSubscriptMedium	56	61	+8.9%	0.92x
ObjectiveCBridgeStringCompare2	896	968	+8.0%	0.93x (?)
SubstringEquatable	617	666	+7.9%	0.93x (?)
Breadcrumbs.UTF16ToIdxRange.longASCII	26	28	+7.7%	0.93x
InsertCharacterEndIndexNonASCII	53	57	+7.5%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
ArrayAppendLatin1Substring	24048	15588	-35.2%	1.54x
ArrayAppendAsciiSubstring	23544	15264	-35.2%	1.54x
ArrayAppendUTF16Substring	23544	15300	-35.0%	1.54x
Data.init.Sequence.64kB.Count.I	52	35	-32.7%	1.49x
Data.init.Sequence.64kB.Count	52	35	-32.7%	1.49x
Data.init.Sequence.2047B.Count.I	91	65	-28.6%	1.40x
Data.init.Sequence.2049B.Count.I	91	65	-28.6%	1.40x
Data.init.Sequence.809B.Count	86	65	-24.4%	1.32x
Data.init.Sequence.809B.Count.I	86	65	-24.4%	1.32x
Data.init.Sequence.511B.Count.I	99	80	-19.2%	1.24x (?)
FlattenListLoop	1771	1449	-18.2%	1.22x (?)
Data.init.Sequence.513B.Count.I	100	82	-18.0%	1.22x (?)
FlattenListFlatMap	5207	4351	-16.4%	1.20x (?)
DataCreateSmallArray	2900	2500	-13.8%	1.16x
DataAccessBytesMedium	74	64	-13.5%	1.16x (?)
Data.hash.Empty	69	61	-11.6%	1.13x
UTF8Decode_InitFromData_ascii_as_ascii	652	600	-8.0%	1.09x (?)
RandomShuffleLCG2	432	400	-7.4%	1.08x (?)
Array2D	6464	6000	-7.2%	1.08x

Code size: -O

Regression	OLD	NEW	DELTA	RATIO
ArrayAppend.o	23637	24101	+2.0%	0.98x
CSVParsing.o	54339	55043	+1.3%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
ObjectiveCBridgingStubs.o	16165	14181	-12.3%	1.14x

Performance: -Osize

Regression	OLD	NEW	DELTA	RATIO
ObjectiveCBridgeStubDataAppend	2380	3000	+26.1%	0.79x
DataCopyBytesSmall	84	105	+25.0%	0.80x
String.data.Medium	104	126	+21.2%	0.83x (?)
StringHasPrefixAscii	1660	2000	+20.5%	0.83x
ObjectiveCBridgeStringHash	99	118	+19.2%	0.84x
StringHasSuffixAscii	1860	2150	+15.6%	0.87x
Data.append.Sequence.64kB.Count0.I	268	306	+14.2%	0.88x
Data.append.Sequence.64kB.Count0	269	307	+14.1%	0.88x
Data.init.Sequence.64kB.Count0.RE	265	300	+13.2%	0.88x
Breadcrumbs.IdxToUTF16Range.longASCII	31	35	+12.9%	0.89x (?)
Data.append.Sequence.809B.Count0.I	373	421	+12.9%	0.89x
Data.init.Sequence.64kB.Count0.RE.I	265	299	+12.8%	0.89x
Data.append.Sequence.809B.Count0	373	420	+12.6%	0.89x
CharacterLiteralsLarge	89	100	+12.4%	0.89x
InsertCharacterEndIndex	147	165	+12.2%	0.89x
Data.append.Sequence.64kB.Count0.RE.I	263	295	+12.2%	0.89x
Data.init.Sequence.64kB.Count0	269	299	+11.2%	0.90x
ObjectiveCBridgeStringGetASCIIContents	397	441	+11.1%	0.90x (?)
Data.init.Sequence.64kB.Count0.I	270	299	+10.7%	0.90x
SubstringEqualString	378	418	+10.6%	0.90x (?)
Data.append.Sequence.809B.Count0.RE.I	369	408	+10.6%	0.90x
Data.init.Sequence.2049B.Count0.I	486	537	+10.5%	0.91x (?)
Data.init.Sequence.2047B.Count0.I	486	535	+10.1%	0.91x
Data.init.Sequence.809B.Count0.RE	430	473	+10.0%	0.91x (?)
Data.init.Sequence.809B.Count0.RE.I	430	473	+10.0%	0.91x (?)
ObjectiveCBridgeStringCStringUsingEncoding	674	740	+9.8%	0.91x (?)
Data.init.Sequence.809B.Count0	435	476	+9.4%	0.91x (?)
Data.init.Sequence.809B.Count0.I	435	476	+9.4%	0.91x (?)
InsertCharacterEndIndexNonASCII	54	59	+9.3%	0.92x (?)
String.replaceSubrange.String.Small	55	60	+9.1%	0.92x
Breadcrumbs.UTF16ToIdxRange.longASCII	26	28	+7.7%	0.93x (?)
Data.init.Sequence.511B.Count0.I	456	491	+7.7%	0.93x (?)
Data.init.Sequence.513B.Count0.I	457	492	+7.7%	0.93x (?)
RomanNumbers2	536	577	+7.6%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
CharIteration_tweet_unicodeScalars	6360	4480	-29.6%	1.42x
CharIteration_ascii_unicodeScalars	3040	2320	-23.7%	1.31x
DataAccessBytesMedium	92	72	-21.7%	1.28x
DataCreateSmallArray	3250	2700	-16.9%	1.20x
Array2D	7280	6224	-14.5%	1.17x
RemoveWhereFilterInts	44	39	-11.4%	1.13x
RandomShuffleLCG2	432	384	-11.1%	1.12x
ArraySetElement	462	411	-11.0%	1.12x
CharIteration_punctuated_unicodeScalars	760	680	-10.5%	1.12x (?)
ObjectiveCBridgeStubDateMutation	256	230	-10.2%	1.11x (?)
CharIteration_punctuatedJapanese_unicodeScalars	800	720	-10.0%	1.11x
PrefixWhileSequence	365	333	-8.8%	1.10x
CharIteration_chinese_unicodeScalars	3240	3000	-7.4%	1.08x (?)
CharIteration_russian_unicodeScalars	3840	3560	-7.3%	1.08x
Data.hash.Empty	69	64	-7.2%	1.08x

Code size: -Osize

Regression	OLD	NEW	DELTA	RATIO
CSVParsing.o	51591	52147	+1.1%	0.99x

Improvement	OLD	NEW	DELTA	RATIO
ObjectiveCBridgingStubs.o	15254	13872	-9.1%	1.10x
RemoveWhere.o	14270	13425	-5.9%	1.06x
StringMatch.o	4140	4027	-2.7%	1.03x
Combos.o	6287	6128	-2.5%	1.03x

Performance: -Onone

Regression	OLD	NEW	DELTA	RATIO
DataCopyBytesSmall	1455	2690	+84.9%	0.54x
DataCopyBytesMedium	1568	2806	+79.0%	0.56x
ArrayValueProp3	5965	8223	+37.9%	0.73x (?)
ArrayValueProp4	5951	7986	+34.2%	0.75x (?)
StackPromo	51800	65600	+26.6%	0.79x (?)
PopFrontArrayGeneric	5000	6200	+24.0%	0.81x (?)
StringToDataSmall	3600	4400	+22.2%	0.82x (?)
StringToDataEmpty	3450	4200	+21.7%	0.82x (?)
String.data.Empty	83	101	+21.7%	0.82x (?)
RandomIntegersDef	57700	70000	+21.3%	0.82x (?)
String.data.Small	87	105	+20.7%	0.83x (?)
DevirtualizeProtocolComposition	6858	8271	+20.6%	0.83x
ObjectiveCBridgeStringHash	99	118	+19.2%	0.84x
Walsh	7282	8631	+18.5%	0.84x
ObjectiveCBridgeStubFromNSDateRef	4800	5680	+18.3%	0.85x (?)
ObjectiveCBridgeStubDataAppend	3900	4580	+17.4%	0.85x
StringBuilderLong	2290	2660	+16.2%	0.86x (?)
ArrayInitFromSlice	484	555	+14.7%	0.87x (?)
ObjectiveCBridgeStringGetASCIIContents	397	441	+11.1%	0.90x
Sim2DArray	12462	13786	+10.6%	0.90x
String.replaceSubrange.String.Small	60	66	+10.0%	0.91x (?)
ObjectiveCBridgeStringCStringUsingEncoding	671	737	+9.8%	0.91x (?)
Data.init.Sequence.809B.Count.RE	29191	32021	+9.7%	0.91x (?)
Radix2CooleyTukeyf	79824	87504	+9.6%	0.91x (?)
RomanNumbers2	7767	8513	+9.6%	0.91x (?)
Data.append.Sequence.64kB.Count0.RE.I	23668	25904	+9.4%	0.91x (?)
Data.init.Sequence.64kB.Count0.RE.I	23678	25913	+9.4%	0.91x (?)
Data.append.Sequence.64kB.Count.RE	23702	25939	+9.4%	0.91x (?)
Data.append.Sequence.64kB.Count.RE.I	23797	26025	+9.4%	0.91x (?)
Data.init.Sequence.64kB.Count.RE	23824	26030	+9.3%	0.92x (?)
DataCreateSmall	72610	79320	+9.2%	0.92x (?)
Data.append.Sequence.809B.Count0.RE	29581	32191	+8.8%	0.92x (?)
NSStringConversion.Mutable	3276	3565	+8.8%	0.92x (?)
Data.init.Sequence.809B.Count.RE.I	29557	32094	+8.6%	0.92x (?)
Data.append.Sequence.809B.Count0.RE.I	29560	32085	+8.5%	0.92x (?)
Data.init.Sequence.64kB.Count0.RE	23848	25877	+8.5%	0.92x (?)
Data.append.Sequence.64kB.Count0.RE	23810	25832	+8.5%	0.92x (?)
Data.append.Sequence.809B.Count.RE.I	29820	32335	+8.4%	0.92x (?)
Data.init.Sequence.64kB.Count.RE.I	23783	25742	+8.2%	0.92x (?)
Data.init.Sequence.809B.Count0.RE.I	29752	32177	+8.2%	0.92x (?)
Data.append.Sequence.809B.Count.RE	29895	32288	+8.0%	0.93x (?)
Data.init.Sequence.809B.Count0.RE	29850	32165	+7.8%	0.93x (?)
DataAppendSequence	2978300	3203100	+7.5%	0.93x (?)

Improvement	OLD	NEW	DELTA	RATIO
SevenBoom	1928	1658	-14.0%	1.16x (?)
DataCreateSmallArray	36750	33250	-9.5%	1.11x (?)
ArrayAppend	6860	6290	-8.3%	1.09x (?)
Integrate	3651	3385	-7.3%	1.08x (?)
ArrayAppendReserved	6570	6120	-6.8%	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

lorentey · 2020-12-12T03:13:37Z

Nah, this is slightly worse than the original.

karwa and others added 15 commits December 4, 2020 02:06

Make UBP nil/count checks debug-mode only

74a8254

Every other Unsafe(Mutable)BufferPointer precondition in this file is debug-mode only. This information is already part of the documentation for this function, and it's really hard to get rid of these branches and traps otherwise.

Make raw buffer nil/count checks debug-mode only

83c14bc

[stdlib] Encourage more inlining on String’s UTF-8 decoding fast path

90d58fe

[test] ArrayTraps: Skip trap test on negative UBP counts in release b…

7f63b42

…uilds

[stdlib] Add a missing debug mode check to U*BP.init(rebasing:)

d52cd82

`Slice` does not technically guarantee that its indices are valid in its base, so these initializers accidentally allowed the creation of obviously out-of-bounds buffers.

[stdlib] Eliminate an overflow check in U[M]BP.distance

27978d1

[stdlib] U[M]RBP.count: Let the compiler assume that the result is no…

2607428

…nnegative

[stdlib] Substring: Restore original Range paths

513a7d9

This eliminates a couple of _debugPreconditions which seem to change inlining enough to interfere with some benchmarks.

[stdlib] Use [Closed]Range.init(_uncheckedBounds:) in more places

d1ef4c4

When the stdlib says not to check, it’s a good idea to actually not have any checking, so that we leave existing code paths unchanged. (And because we do trust that the stdlib is responsible about such things.)

[stdlib] Don’t use unbound ranges (foo[…]) in String → Substring co…

9336821

…nversions The UnboundRange → Range conversion path is complicated and it involves at least one unnecessary _precondition check for startIndex ..< endIndex.

[stdlib] U[R]BP: Restore compile-time condition lost with removed pre…

282a140

…condition

[stdlib] Add an unsafe U[M]BP initializer to work around some inliner…

1f92df0

… test failures

[stdlib] Document sources of code bloat

94a7eee

[stdlib] Reintroduce a precondition for count >= 0

8cc91b2

lorentey closed this Dec 12, 2020

lorentey deleted the buffers-need-to-be-fast-but-not-too-fast2 branch December 12, 2020 03:13

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[DO NOT MERGE] More experiments with UnsafeBufferPointer.init #35075

[DO NOT MERGE] More experiments with UnsafeBufferPointer.init #35075

Uh oh!

lorentey commented Dec 12, 2020

Uh oh!

lorentey commented Dec 12, 2020

Uh oh!

swift-ci commented Dec 12, 2020

Uh oh!

lorentey commented Dec 12, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[DO NOT MERGE] More experiments with UnsafeBufferPointer.init #35075

[DO NOT MERGE] More experiments with UnsafeBufferPointer.init #35075

Uh oh!

Conversation

lorentey commented Dec 12, 2020

Uh oh!

lorentey commented Dec 12, 2020

Uh oh!

swift-ci commented Dec 12, 2020

Performance: -O

Code size: -O

Performance: -Osize

Code size: -Osize

Performance: -Onone

Code size: -swiftlibs

Uh oh!

lorentey commented Dec 12, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants