[squeak-dev] The Trunk: Collections-ul.739.mczClément Bera bera.clement at gmail.comWed Mar 1 16:04:38 UTC 2017
On Wed, Mar 1, 2017 at 12:45 AM, Eliot Miranda <eliot.miranda at gmail.com> wrote: > Hi Levente, > > On Tue, Feb 28, 2017 at 2:43 PM, Levente Uzonyi <leves at caesar.elte.hu> > wrote: > >> Hi Eliot, >> >> I tried to change the selector in the special objects array, but it >> didn't help. I get the same numbers as before. >> >> Here's what I did: >> First, I changed #blockCopy to: #~~ in SmalltalkImage >> >> #recreateSpecialObjectsArray. >> Then, I evaluated the following snippet in a workspace: >> >> senders := SystemNavigation default allCallsOn: #~~. >> Smalltalk recreateSpecialObjectsArray. >> senders >> do: [ :methodReference | >> | class | >> class := methodReference actualClass. >> class recompile: methodReference selector from: class ] >> displayingProgress: 'Recompiling...'. >> >> And finally I ran the benchmark. >> >> However, now #blockCopy: will just behave as #~~ should. And, it gives >> the same number as #==. So, I suppose there are few other places to change >> to get the right behavior (Compiler, Encoder, Whatever). >> > > I forgot to tell you that you need to do > VariableNode initialize. Decompiler initialize > See variableNode class>>#initialize. This fills StdSelectors, a class var > of ParseNode, with the special selectors. Sorry! > > >> However, what's really interesting is how the JIT behaves in this case: >> #== alone is about 1.43x slower than #== + #ifTrue:ifFalse: in this test. > > > Yes, I chose to only inline when a comparison operator is followed by a > conditional branch. So any of #= #~= #== #~~ #> #< #>= #<= if used to > compute a value will be compiled into a real send That could be a > mistake. Let me know if you think it will be critical to performance. The > notion is that inlining the branch saves a lot (reifying the result, > testing the result for being true, false or something else), but inlining > just to obtain the result saves only the inline cached send overhead. That > could be the wrong call, but that's what I decided to do back in 2009. > > Hi both, I believe #== and #~~ are inlined even if they're not followed by a branch in #genIdenticalNoBranchArgIsConstant:rcvrIsConstant:argReg:rcvrReg:orNotIf:. It's inlined to allow registers to stay live. The other special selectors Eliot mentioned (#= #~= #> #< #>= #<=) are however inlined only if followed by a branch. > >> >> Levente >> >> >> On Tue, 28 Feb 2017, Eliot Miranda wrote: >> >> Hi Levente, >>> On Tue, Feb 28, 2017 at 10:44 AM, Levente Uzonyi <leves at caesar.elte.hu> >>> wrote: >>> In case someone would like to defend the performance of #~~ and >>> #notNil, here's a little benchmark I just wrote: >>> >>> | base ifNilIfNotNil notIdentical notEqual notNil | >>> Smalltalk garbageCollect. >>> base := (1 to: 5) collect: [ :run | >>> [ 1 to: 100000000 do: [ :i | ] ] timeToRun ]. >>> >>> ifNilIfNotNil := (1 to: 5) collect: [ :run | >>> [ 1 to: 100000000 do: [ :i | i ifNil: [ false ] ifNotNil: >>> [ true ] ] ] timeToRun ]. >>> >>> notIdentical := (1 to: 5) collect: [ :run | >>> [ 1 to: 100000000 do: [ :i | i ~~ nil ] ] timeToRun ]. >>> >>> notEqual := (1 to: 5) collect: [ :run | >>> [ 1 to: 100000000 do: [ :i | i ~= nil ] ] timeToRun ]. >>> >>> >>> notNil := (1 to: 5) collect: [ :run | >>> [ 1 to: 100000000 do: [ :i | i notNil ] ] timeToRun ]. >>> >>> { >>> #ifNilIfNotNil -> ifNilIfNotNil. >>> #notIdentical -> notIdentical. >>> #notEqual -> notEqual. >>> #notNil -> notNil >>> } >>> replace: [ :each | >>> each key -> (each value average - base average) >>> asFloat ]; >>> sort: #value ascending >>> >>> "{#ifNilIfNotNil->73.6 . #notNil->336.8 . #notIdentical->511.6 . >>> #notEqual->3354.0}" >>> >>> So it's 4.5x faster than #notNil and 7x faster than #~~ on 64-bit >>> Spur on Linux. Once we have jitted #~~, I'll change the code again. >>> >>> >>> IIRC, all one needs to get jitted #~~ is to store #~~ in place of >>> blockCopy: in the specialSelectors array and recompile all senders of #~~. >>> You may need to collect all senders of #~~ before you make the >>> swap since putting it in the specialSelectors array may confuse >>> allReferencesTo: et al. >>> >>> Levente >>> >>> >>> On Tue, 28 Feb 2017, commits at source.squeak.org wrote: >>> >>> Levente Uzonyi uploaded a new version of Collections to >>> project The Trunk: >>> http://source.squeak.org/trunk/Collections-ul.739.mcz >>> >>> ==================== Summary ==================== >>> >>> Name: Collections-ul.739 >>> Author: ul >>> Time: 28 February 2017, 7:34:55.711173 pm >>> UUID: 66c6ecf2-71bf-4e7d-80e6-5ace72fa19ba >>> Ancestors: Collections-ul.738 >>> >>> - use micro optimization for #~~ and #notNil in common >>> collection methods >>> - reuse one of the OrderedCollections in RunArray >> >>> #scanFrom: >>> - two other minor tweaks in KeyedSet >>> - removed accidentally commited Set >> #includes2: >>> >>> =============== Diff against Collections-ul.738 >>> =============== >>> >>> Item was changed: >>> ----- Method: Dictionary>>includesKey: (in category >>> 'testing') ----- >>> includesKey: key >>> "Answer whether the receiver has a key equal to the >>> argument, key." >>> >>> + (array at: (self scanFor: key)) ifNil: [ ^false ] >>> ifNotNil: [ ^true ]! >>> - ^(array at: (self scanFor: key)) notNil! >>> >>> Item was changed: >>> ----- Method: KeyedSet>>includesKey: (in category >>> 'testing') ----- >>> includesKey: key >>> >>> + (array at: (self scanFor: key)) ifNil: [ ^false ] >>> ifNotNil: [ ^true ]! >>> - ^ (array at: (self scanFor: key)) notNil! >>> >>> Item was changed: >>> ----- Method: KeyedSet>>remove:ifAbsent: (in category >>> 'removing') ----- >>> remove: oldObject ifAbsent: aBlock >>> >>> | index | >>> index := self scanFor: (keyBlock value: oldObject). >>> + (array at: index) ifNil: [ ^ aBlock value ]. >>> - (array at: index) == nil ifTrue: [ ^ aBlock value ]. >>> array at: index put: nil. >>> tally := tally - 1. >>> self fixCollisionsFrom: index. >>> ^ oldObject! >>> >>> Item was changed: >>> ----- Method: KeyedSet>>removeKey:ifAbsent: (in category >>> 'removing') ----- >>> removeKey: key ifAbsent: aBlock >>> >>> | index obj | >>> index := self scanFor: key. >>> + obj := (array at: index) ifNil: [ ^ aBlock value ]. >>> - (obj := array at: index) == nil ifTrue: [ ^ aBlock >>> value ]. >>> array at: index put: nil. >>> tally := tally - 1. >>> self fixCollisionsFrom: index. >>> ^ obj enclosedSetElement! >>> >>> Item was changed: >>> ----- Method: RunArray class>>scanFrom: (in category >>> 'instance creation') ----- >>> scanFrom: strm >>> "Read the style section of a fileOut or sources >>> file. nextChunk has already been done. We need to return a RunArray of >>> TextAttributes of various kinds. These are written by >>> the implementors of writeScanOn:" >>> | runs values attrList char | >>> (strm peekFor: $( ) ifFalse: [^ nil]. >>> runs := OrderedCollection new. >>> [strm skipSeparators. >>> strm peekFor: $)] whileFalse: >>> [runs add: (Number readFrom: strm)]. >>> values := OrderedCollection new. "Value array" >>> attrList := OrderedCollection new. "Attributes >>> list" >>> [(char := strm peek) == nil] whileFalse: [ >>> (char isSeparator or: [ char = $!! ]) >>> ifTrue: [ "n.b. Skip $!! to meet >>> expectations of RunArrayTest>>testScanFromTrailer. >>> The example string >>> used in that test does not seem to match the implemention >>> of the fileOut >>> serialization, but the test may be right and the implementation >>> wrong. In any case, >>> relax the parsing here to meet the test expectations, and to >>> be more consistent >>> with the original version of this method that assumed any >>> unexpected charater >>> to be a separator. -dtl Jan 2014" >>> strm next "space, cr do >>> nothing"] >>> ifFalse: [char == $, >>> ifTrue: [strm next. >>> values add: >>> attrList asArray. >>> + attrList >>> reset ] >>> - attrList := >>> OrderedCollection new] >>> ifFalse: [attrList >>> add: (TextAttribute newFrom: strm)] >>> ] >>> ]. >>> values add: attrList asArray. >>> ^ self runs: runs asArray values: (values copyFrom: >>> 1 to: runs size) asArray >>> " >>> RunArray scanFrom: (ReadStream on: '(14 50 >>> 312)f1,f1b,f1LInteger +;i') >>> "! >>> >>> Item was removed: >>> - ----- Method: Set>>includes2: (in category 'testing') ----- >>> - includes2: anObject - - ^((array at: (self scanFor: >>> anObject)) == nil) not! >>> >>> Item was changed: >>> ----- Method: WeakIdentityDictionary>>includesKey: (in >>> category 'testing') ----- >>> includesKey: key >>> "Answer whether the receiver has a key equal to the >>> argument, key." >>> >>> + (array at: (self scanFor: key)) >>> + ifNil: [ >>> + "it just has been reclaimed" >>> + ^false] >>> + ifNotNil: [ :element | >>> + element == vacuum >>> + ifTrue: [ ^false ] >>> + ifFalse: [ ^true ] ]! >>> - ^(array at: (self scanFor: key)) >>> - ifNil: >>> - ["it just has been reclaimed" >>> - false] >>> - ifNotNil: [:element | element ~~ vacuum]! >>> >>> Item was changed: >>> ----- Method: WeakSet>>includes: (in category 'testing') >>> ----- >>> includes: anObject >>> >>> (array at: (self scanFor: anObject)) >>> ifNil: [ ^false ] >>> + ifNotNil: [ :object | >>> + object == flag >>> + ifTrue: [ ^false ] >>> + ifFalse: [ ^true ] ]! >>> - ifNotNil: [ :object | ^object ~~ flag ]! >>> >>> >>> >>> >>> >>> -- >>> _,,,^..^,,,_ >>> best, Eliot >>> >>> >> >> >> > > > -- > _,,,^..^,,,_ > best, Eliot > > > > -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.squeakfoundation.org/pipermail/squeak-dev/attachments/20170301/be5ba2ac/attachment.html>
More information about the Squeak-dev mailing list |