[Runtime] Avoid function pointer indirection in refcounting functions. #26516

mikeash · 2019-08-06T18:30:24Z

Functions like swift_retain call through a function pointer so that Instruments can interpose. This slows down the common case where there is no interposition. Instead, initialize the function pointers to NULL and call through directly to the real implementation when it's NULL. The compiler is smart enough to inline this call and the result is a single conditional branch rather than a function pointer call.

rdar://problem/18307425

Functions like swift_retain call through a function pointer so that Instruments can interpose. This slows down the common case where there is no interposition. Instead, initialize the function pointers to NULL and call through directly to the real implementation when it's NULL. The compiler is smart enough to inline this call and the result is a single conditional branch rather than a function pointer call. rdar://problem/18307425

mikeash · 2019-08-06T18:31:33Z

@swift-ci please benchmark

swift-ci · 2019-08-06T19:14:58Z

Performance: -O

Improvement	OLD	NEW	DELTA	RATIO
LazilyFilteredArrayContains	23000	18400	-20.0%	1.25x
UTF8Decode_InitDecoding	124	103	-16.9%	1.20x
ArraySetElement	305	262	-14.1%	1.16x
RandomShuffleLCG2	384	336	-12.5%	1.14x
Set.isSuperset.Seq.Empty.Int	53	47	-11.3%	1.13x (?)
Set.isDisjoint.Box.Empty	98	87	-11.2%	1.13x
RemoveWhereMoveInts	18	16	-11.1%	1.12x
Set.subtracting.Empty.Box	9	8	-11.1%	1.12x
Array2D	4144	3696	-10.8%	1.12x (?)
MapReduceAnyCollection	215	193	-10.2%	1.11x
Set.isSubset.Int.Empty	51	46	-9.8%	1.11x (?)
DataSetCountMedium	410	370	-9.8%	1.11x (?)
ObjectiveCBridgeStringHash	42	38	-9.5%	1.11x (?)
MapReduce	237	215	-9.3%	1.10x (?)
DistinctClassFieldAccesses	194	176	-9.3%	1.10x
ArrayLiteral2	78	71	-9.0%	1.10x (?)
DictionarySubscriptDefaultMutation	191	174	-8.9%	1.10x
RemoveWhereSwapInts	34	31	-8.8%	1.10x (?)
Integrate	217	198	-8.8%	1.10x (?)
FlattenListLoop	2512	2303	-8.3%	1.09x (?)
DropWhileAnySequence	2130	1953	-8.3%	1.09x (?)
DropFirstAnySequenceLazy	2188	2015	-7.9%	1.09x (?)
SubstringFromLongStringGeneric	13	12	-7.7%	1.08x (?)
PrefixWhileAnySeqCRangeIter	186	172	-7.5%	1.08x (?)
DropLastAnySequence	334	309	-7.5%	1.08x (?)
Set.isDisjoint.Seq.Box.Empty	67	62	-7.5%	1.08x (?)
Set.subtracting.Empty.Int	28	26	-7.1%	1.08x (?)
PrefixWhileAnySeqCntRange	186	173	-7.0%	1.08x
PrefixAnySequence	1769	1649	-6.8%	1.07x (?)
PrefixWhileSequence	195	182	-6.7%	1.07x (?)

Code size: -O

Performance: -Osize

Improvement	OLD	NEW	DELTA	RATIO
FlattenListLoop	2968	2169	-26.9%	1.37x (?)
UTF8Decode_InitDecoding	122	102	-16.4%	1.20x
ArraySetElement	305	262	-14.1%	1.16x
RangeAssignment	223	195	-12.6%	1.14x (?)
RemoveWhereSwapInts	33	29	-12.1%	1.14x
Set.isSuperset.Seq.Empty.Int	52	46	-11.5%	1.13x (?)
Set.isStrictSubset.Int.Empty	54	48	-11.1%	1.12x (?)
Array2D	3904	3472	-11.1%	1.12x
Set.isDisjoint.Box.Empty	110	98	-10.9%	1.12x (?)
Set.isDisjoint.Seq.Int.Empty	56	50	-10.7%	1.12x (?)
Set.isStrictSubset.Seq.Int.Empty	128	115	-10.2%	1.11x (?)
ObjectiveCBridgeStringHash	42	38	-9.5%	1.11x
DataCreateEmptyArray	1650	1500	-9.1%	1.10x (?)
Integrate	220	200	-9.1%	1.10x
MapReduce	258	237	-8.1%	1.09x (?)
MapReduceAnyCollection	259	238	-8.1%	1.09x (?)
RandomShuffleLCG2	400	368	-8.0%	1.09x (?)
Set.isSubset.Seq.Int.Empty	125	115	-8.0%	1.09x (?)
RemoveWhereFilterInts	25	23	-8.0%	1.09x (?)
SubstringFromLongStringGeneric	13	12	-7.7%	1.08x (?)
Set.isDisjoint.Seq.Box.Empty	95	88	-7.4%	1.08x (?)
DataSetCountMedium	410	380	-7.3%	1.08x (?)
PrefixWhileAnySeqCRangeIter	211	196	-7.1%	1.08x (?)
PrefixWhileSequence	201	187	-7.0%	1.07x (?)
PrefixWhileAnySeqCntRange	210	196	-6.7%	1.07x (?)

Code size: -Osize

Performance: -Onone

Regression	OLD	NEW	DELTA	RATIO
ArrayOfPOD	665	740	+11.3%	0.90x (?)

Improvement	OLD	NEW	DELTA	RATIO
UTF8Decode_InitDecoding	143	123	-14.0%	1.16x (?)
ObjectiveCBridgeStringHash	42	38	-9.5%	1.11x

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: 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

compnerd · 2019-08-09T01:33:57Z

stdlib/public/runtime/HeapObject.cpp

+                                     size_t requiredAlignmentMask) {
+  if (SWIFT_UNLIKELY(_swift_allocObject))
+    return _swift_allocObject(metadata, requiredSize, requiredAlignmentMask);
+  else


The else is unnecessary, and similar throughout.

Thanks, updated.

rjmccall · 2019-08-14T04:33:57Z

Very nice!

…k against those values to call directly. Instruments relies on the old values being there so it can call the original implementation. This has very slightly worse codegen but the impact should be minimal.

mikeash · 2019-08-16T18:43:21Z

Turns out that Instruments doesn't work when the function pointers start out as NULL. Trying a different approach now where we initialize them as before, but then we check for that value and make the call directly. The compiler still inlines the implementation in this case. One additional leaq instruction is generated to get the address to compare with, which shouldn't hurt too much.

mikeash · 2019-08-16T18:44:09Z

@swift-ci please benchmark

swift-ci · 2019-08-16T19:04:37Z

Performance: -O

Regression	OLD	NEW	DELTA	RATIO
DistinctClassFieldAccesses	173	202	+16.8%	0.86x
SubstringFromLongString	7	8	+14.3%	0.88x
UTF8Decode_InitDecoding	110	119	+8.2%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
Set.isStrictSubset.Seq.Int.Empty	129	111	-14.0%	1.16x
Set.isDisjoint.Seq.Int.Empty	56	49	-12.5%	1.14x
ArrayLiteral2	82	72	-12.2%	1.14x (?)
Chars2	3350	2950	-11.9%	1.14x (?)
DataCreateEmptyArray	1700	1500	-11.8%	1.13x (?)
Set.isDisjoint.Box.Empty	99	88	-11.1%	1.12x (?)
Set.subtracting.Empty.Box	9	8	-11.1%	1.12x (?)
Set.isSubset.Seq.Int.Empty	126	113	-10.3%	1.12x
Set.isSuperset.Seq.Empty.Int	51	46	-9.8%	1.11x (?)
DataCreateSmallArray	2250	2050	-8.9%	1.10x (?)
NormalizedIterator_fastPrenormal	630	580	-7.9%	1.09x (?)
LazilyFilteredArrayContains	20200	18600	-7.9%	1.09x (?)
Set.isStrictSubset.Int.Empty	54	50	-7.4%	1.08x (?)
DictionarySubscriptDefaultMutation	190	176	-7.4%	1.08x
PrefixWhileAnySequence	1153	1071	-7.1%	1.08x (?)
Calculator	147	137	-6.8%	1.07x (?)
Breadcrumbs.IdxToUTF16Range.longMixed	288	269	-6.6%	1.07x (?)

Code size: -O

Performance: -Osize

Regression	OLD	NEW	DELTA	RATIO
LazilyFilteredArrayContains	38800	45000	+16.0%	0.86x (?)
HashTest	750	820	+9.3%	0.91x (?)
Hanoi	2160	2350	+8.8%	0.92x (?)

Improvement	OLD	NEW	DELTA	RATIO
Set.isSuperset.Seq.Empty.Int	54	46	-14.8%	1.17x (?)
Set.isDisjoint.Box.Empty	104	89	-14.4%	1.17x
Set.isDisjoint.Seq.Box.Empty	100	87	-13.0%	1.15x (?)
DataCreateEmptyArray	1700	1500	-11.8%	1.13x (?)
Set.isDisjoint.Int.Empty	61	54	-11.5%	1.13x (?)
Set.subtracting.Empty.Box	9	8	-11.1%	1.12x
Set.isSubset.Seq.Int.Empty	129	116	-10.1%	1.11x
Chars2	3550	3200	-9.9%	1.11x (?)
Set.isStrictSubset.Int.Empty	52	47	-9.6%	1.11x
Calculator	150	136	-9.3%	1.10x (?)
NormalizedIterator_fastPrenormal	660	600	-9.1%	1.10x (?)
Set.isStrictSubset.Seq.Int.Empty	130	119	-8.5%	1.09x (?)
Set.isDisjoint.Seq.Int.Empty	51	47	-7.8%	1.09x (?)
DataSetCountMedium	410	380	-7.3%	1.08x (?)
DictionaryBridgeToObjC_Bridge	15	14	-6.7%	1.07x (?)
Set.isStrictSuperset.Seq.Empty.Int	181	169	-6.6%	1.07x (?)

Code size: -Osize

Performance: -Onone

Regression	OLD	NEW	DELTA	RATIO
ArrayInClass	2185	2495	+14.2%	0.88x
StringWalk	7120	8120	+14.0%	0.88x
DataAppendDataSmallToSmall	2960	3360	+13.5%	0.88x (?)
StringFromLongWholeSubstring	9	10	+11.1%	0.90x
NopDeinit	113100	124700	+10.3%	0.91x (?)

Improvement	OLD	NEW	DELTA	RATIO
SubstringFromLongString	10	9	-10.0%	1.11x (?)
ArrayOfPOD	723	661	-8.6%	1.09x (?)
DataSetCountMedium	430	400	-7.0%	1.07x (?)
StringToDataSmall	750	700	-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 mini
  Model Identifier: Macmini8,1
  Processor Name: 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

mikeash · 2019-08-27T16:31:30Z

@swift-ci please test

swift-ci · 2019-08-27T16:35:19Z

Build failed
Swift Test OS X Platform
Git Sha - 9fd03e2

mikeash · 2019-08-27T21:47:45Z

@swift-ci please test os x platform

compnerd reviewed Aug 9, 2019

View reviewed changes

rjmccall approved these changes Aug 14, 2019

View reviewed changes

[Runtime] Initialize function pointers with the original values, chec…

5198724

…k against those values to call directly. Instruments relies on the old values being there so it can call the original implementation. This has very slightly worse codegen but the impact should be minimal.

mikeash merged commit 9e61d53 into swiftlang:master Aug 28, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[Runtime] Avoid function pointer indirection in refcounting functions. #26516

[Runtime] Avoid function pointer indirection in refcounting functions. #26516

mikeash commented Aug 6, 2019

mikeash commented Aug 6, 2019

swift-ci commented Aug 6, 2019

compnerd Aug 9, 2019

mikeash Aug 16, 2019

rjmccall commented Aug 14, 2019

mikeash commented Aug 16, 2019

mikeash commented Aug 16, 2019

swift-ci commented Aug 16, 2019

mikeash commented Aug 27, 2019

swift-ci commented Aug 27, 2019

mikeash commented Aug 27, 2019

[Runtime] Avoid function pointer indirection in refcounting functions. #26516

[Runtime] Avoid function pointer indirection in refcounting functions. #26516

Conversation

mikeash commented Aug 6, 2019

mikeash commented Aug 6, 2019

swift-ci commented Aug 6, 2019

Performance: -O

Code size: -O

Performance: -Osize

Code size: -Osize

Performance: -Onone

Code size: -swiftlibs

compnerd Aug 9, 2019

Choose a reason for hiding this comment

mikeash Aug 16, 2019

Choose a reason for hiding this comment

rjmccall commented Aug 14, 2019

mikeash commented Aug 16, 2019

mikeash commented Aug 16, 2019

swift-ci commented Aug 16, 2019

Performance: -O

Code size: -O

Performance: -Osize

Code size: -Osize

Performance: -Onone

Code size: -swiftlibs

mikeash commented Aug 27, 2019

swift-ci commented Aug 27, 2019

mikeash commented Aug 27, 2019