<div dir="ltr">Hi Levente,<div class="gmail_extra"><br><div class="gmail_quote">On Tue, Feb 28, 2017 at 2:43 PM, Levente Uzonyi <span dir="ltr"><<a href="mailto:leves@caesar.elte.hu" target="_blank">leves@caesar.elte.hu</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">Hi Eliot,<br>
<br>
I tried to change the selector in the special objects array, but it didn't help. I get the same numbers as before.<br>
<br>
Here's what I did:<br>
First, I changed #blockCopy to: #~~ in SmalltalkImage >> #recreateSpecialObjectsArray.<br>
Then, I evaluated the following snippet in a workspace:<br>
<br>
senders := SystemNavigation default allCallsOn: #~~.<br>
Smalltalk recreateSpecialObjectsArray.<br>
senders<br>
do: [ :methodReference |<br>
| class |<br>
class := methodReference actualClass.<br>
class recompile: methodReference selector from: class ]<br>
displayingProgress: 'Recompiling...'.<br>
<br>
And finally I ran the benchmark.<br>
<br>
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).<br></blockquote><div><br></div><div>I forgot to tell you that you need to do </div><div> VariableNode initialize. Decompiler initialize</div><div>See variableNode class>>#initialize. This fills StdSelectors, a class var of ParseNode, with the special selectors. Sorry!</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">
<br>
However, what's really interesting is how the JIT behaves in this case:<br>
#== alone is about 1.43x slower than #== + #ifTrue:ifFalse: in this test.</blockquote><div><br></div><div>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.</div><div><br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex"><span class="gmail-HOEnZb"><font color="#888888"><br>
<br>
Levente</font></span><div class="gmail-HOEnZb"><div class="gmail-h5"><br>
<br>
On Tue, 28 Feb 2017, Eliot Miranda wrote:<br>
<br>
<blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left-width:1px;border-left-color:rgb(204,204,204);border-left-style:solid;padding-left:1ex">
Hi Levente,<br>
On Tue, Feb 28, 2017 at 10:44 AM, Levente Uzonyi <<a href="mailto:leves@caesar.elte.hu" target="_blank">leves@caesar.elte.hu</a>> wrote:<br>
In case someone would like to defend the performance of #~~ and #notNil, here's a little benchmark I just wrote:<br>
<br>
| base ifNilIfNotNil notIdentical notEqual notNil |<br>
Smalltalk garbageCollect.<br>
base := (1 to: 5) collect: [ :run |<br>
[ 1 to: 100000000 do: [ :i | ] ] timeToRun ].<br>
<br>
ifNilIfNotNil := (1 to: 5) collect: [ :run |<br>
[ 1 to: 100000000 do: [ :i | i ifNil: [ false ] ifNotNil: [ true ] ] ] timeToRun ].<br>
<br>
notIdentical := (1 to: 5) collect: [ :run |<br>
[ 1 to: 100000000 do: [ :i | i ~~ nil ] ] timeToRun ].<br>
<br>
notEqual := (1 to: 5) collect: [ :run |<br>
[ 1 to: 100000000 do: [ :i | i ~= nil ] ] timeToRun ].<br>
<br>
<br>
notNil := (1 to: 5) collect: [ :run |<br>
[ 1 to: 100000000 do: [ :i | i notNil ] ] timeToRun ].<br>
<br>
{<br>
#ifNilIfNotNil -> ifNilIfNotNil.<br>
#notIdentical -> notIdentical.<br>
#notEqual -> notEqual.<br>
#notNil -> notNil<br>
}<br>
replace: [ :each |<br>
each key -> (each value average - base average) asFloat ];<br>
sort: #value ascending<br>
<br>
"{#ifNilIfNotNil->73.6 . #notNil->336.8 . #notIdentical->511.6 . #notEqual->3354.0}"<br>
<br>
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.<br>
<br>
<br>
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<br>
swap since putting it in the specialSelectors array may confuse allReferencesTo: et al.<br>
<br>
Levente<br>
<br>
<br>
On Tue, 28 Feb 2017, <a href="mailto:commits@source.squeak.org" target="_blank">commits@source.squeak.org</a> wrote:<br>
<br>
Levente Uzonyi uploaded a new version of Collections to project The Trunk:<br>
<a href="http://source.squeak.org/trunk/Collections-ul.739.mcz" rel="noreferrer" target="_blank">http://source.squeak.org/trunk<wbr>/Collections-ul.739.mcz</a><br>
<br>
==================== Summary ====================<br>
<br>
Name: Collections-ul.739<br>
Author: ul<br>
Time: 28 February 2017, 7:34:55.711173 pm<br>
UUID: 66c6ecf2-71bf-4e7d-80e6-5ace72<wbr>fa19ba<br>
Ancestors: Collections-ul.738<br>
<br>
- use micro optimization for #~~ and #notNil in common collection methods<br>
- reuse one of the OrderedCollections in RunArray >> #scanFrom:<br>
- two other minor tweaks in KeyedSet<br>
- removed accidentally commited Set >> #includes2:<br>
<br>
=============== Diff against Collections-ul.738 ===============<br>
<br>
Item was changed:<br>
----- Method: Dictionary>>includesKey: (in category 'testing') -----<br>
includesKey: key<br>
"Answer whether the receiver has a key equal to the argument, key."<br>
<br>
+ (array at: (self scanFor: key)) ifNil: [ ^false ] ifNotNil: [ ^true ]!<br>
- ^(array at: (self scanFor: key)) notNil!<br>
<br>
Item was changed:<br>
----- Method: KeyedSet>>includesKey: (in category 'testing') -----<br>
includesKey: key<br>
<br>
+ (array at: (self scanFor: key)) ifNil: [ ^false ] ifNotNil: [ ^true ]!<br>
- ^ (array at: (self scanFor: key)) notNil!<br>
<br>
Item was changed:<br>
----- Method: KeyedSet>>remove:ifAbsent: (in category 'removing') -----<br>
remove: oldObject ifAbsent: aBlock<br>
<br>
| index |<br>
index := self scanFor: (keyBlock value: oldObject).<br>
+ (array at: index) ifNil: [ ^ aBlock value ].<br>
- (array at: index) == nil ifTrue: [ ^ aBlock value ].<br>
array at: index put: nil.<br>
tally := tally - 1.<br>
self fixCollisionsFrom: index.<br>
^ oldObject!<br>
<br>
Item was changed:<br>
----- Method: KeyedSet>>removeKey:ifAbsent: (in category 'removing') -----<br>
removeKey: key ifAbsent: aBlock<br>
<br>
| index obj |<br>
index := self scanFor: key.<br>
+ obj := (array at: index) ifNil: [ ^ aBlock value ].<br>
- (obj := array at: index) == nil ifTrue: [ ^ aBlock value ].<br>
array at: index put: nil.<br>
tally := tally - 1.<br>
self fixCollisionsFrom: index.<br>
^ obj enclosedSetElement!<br>
<br>
Item was changed:<br>
----- Method: RunArray class>>scanFrom: (in category 'instance creation') -----<br>
scanFrom: strm<br>
"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<br>
the implementors of writeScanOn:"<br>
| runs values attrList char |<br>
(strm peekFor: $( ) ifFalse: [^ nil].<br>
runs := OrderedCollection new.<br>
[strm skipSeparators.<br>
strm peekFor: $)] whileFalse:<br>
[runs add: (Number readFrom: strm)].<br>
values := OrderedCollection new. "Value array"<br>
attrList := OrderedCollection new. "Attributes list"<br>
[(char := strm peek) == nil] whileFalse: [<br>
(char isSeparator or: [ char = $!! ])<br>
ifTrue: [ "n.b. Skip $!! to meet expectations of RunArrayTest>>testScanFromTrai<wbr>ler.<br>
The example string used in that test does not seem to match the implemention<br>
of the fileOut serialization, but the test may be right and the implementation<br>
wrong. In any case, relax the parsing here to meet the test expectations, and to<br>
be more consistent with the original version of this method that assumed any<br>
unexpected charater to be a separator. -dtl Jan 2014"<br>
strm next "space, cr do nothing"]<br>
ifFalse: [char == $,<br>
ifTrue: [strm next.<br>
values add: attrList asArray.<br>
+ attrList reset ]<br>
- attrList := OrderedCollection new]<br>
ifFalse: [attrList add: (TextAttribute newFrom: strm)]<br>
]<br>
].<br>
values add: attrList asArray.<br>
^ self runs: runs asArray values: (values copyFrom: 1 to: runs size) asArray<br>
"<br>
RunArray scanFrom: (ReadStream on: '(14 50 312)f1,f1b,f1LInteger +;i')<br>
"!<br>
<br>
Item was removed:<br>
- ----- Method: Set>>includes2: (in category 'testing') -----<br>
- includes2: anObject - - ^((array at: (self scanFor: anObject)) == nil) not!<br>
<br>
Item was changed:<br>
----- Method: WeakIdentityDictionary>>includ<wbr>esKey: (in category 'testing') -----<br>
includesKey: key<br>
"Answer whether the receiver has a key equal to the argument, key."<br>
<br>
+ (array at: (self scanFor: key))<br>
+ ifNil: [<br>
+ "it just has been reclaimed"<br>
+ ^false]<br>
+ ifNotNil: [ :element |<br>
+ element == vacuum<br>
+ ifTrue: [ ^false ]<br>
+ ifFalse: [ ^true ] ]!<br>
- ^(array at: (self scanFor: key))<br>
- ifNil:<br>
- ["it just has been reclaimed"<br>
- false]<br>
- ifNotNil: [:element | element ~~ vacuum]!<br>
<br>
Item was changed:<br>
----- Method: WeakSet>>includes: (in category 'testing') -----<br>
includes: anObject<br>
<br>
(array at: (self scanFor: anObject))<br>
ifNil: [ ^false ]<br>
+ ifNotNil: [ :object |<br>
+ object == flag<br>
+ ifTrue: [ ^false ]<br>
+ ifFalse: [ ^true ] ]!<br>
- ifNotNil: [ :object | ^object ~~ flag ]!<br>
<br>
<br>
<br>
<br>
<br>
--<br>
_,,,^..^,,,_<br>
best, Eliot<br>
<br>
</blockquote>
</div></div><br><br>
<br></blockquote></div><br><br clear="all"><div><br></div>-- <br><div class="gmail_signature"><div dir="ltr"><div><span style="font-size:small;border-collapse:separate"><div>_,,,^..^,,,_<br></div><div>best, Eliot</div></span></div></div></div>
</div></div>