type-directed completion of partial expressions · i paint.net image editor i windows installer xml...

Type-Directed Completion of Partial Expressions

Daniel Perelman†

Sumit Gulwani‡ Thomas Ball‡ Dan Grossman†

†University of Washington

‡Microsoft Research Redmond

June 12, 2012

I want to shrink an image...

Document image = ...; Size newSize = ...;

image.:::::::Shrink(newSize)

2 / 54



image.

3 / 54



image.

4 / 54



image.

5 / 54



image.

6 / 54



PaintDotNet.

7 / 54



PaintDotNet.

8 / 54



PaintDotNet.

9 / 54



PaintDotNet.

10 / 54



PaintDotNet.Document.

11 / 54



PaintDotNet.Document.

12 / 54


13 / 54



PaintDotNet.Data.

14 / 54



PaintDotNet.Actions.

15 / 54




16 / 54




17 / 54




18 / 54



PaintDotNet.Actions.ResizeAction.

19 / 54




20 / 54



PaintDotNet.Actions.CanvasSizeAction.

21 / 54



PaintDotNet.Actions.CanvasSizeAction.

.ResizeDocument(

/* PaintDotNet.Document image */,

/* System.Drawing.Size size */,

/* PaintDotNet.AnchorEdge edge */,

/* PointDotNet.ColorBgra bgColor */);

22 / 54

Programmer thought process

I I have a Document and a Size

I I want to shrink the Document

I There must be a method

I Current code completionI Left-to-rightI Complete, alphabetic list of just next tokenI Very limited filtering

23 / 54





I Current code completionI Left-to-rightI Complete, alphabetic list of just next tokenI Very limited filtering

24 / 54

Proposed workflow


var newImage =:?({image, newSize})

25 / 54

Proposed workflow


var newImage =:?({image, newSize})

26 / 54





I Query should contain what the programmer knowsI Some values and types the expression should involveI Loose syntactic structure

I Query shouldn’t require what the programmer doesn’t knowI NamesI Argument orderI Other arguments

I Show “best” results first

I Similar in spirit to Prospector [Mandelin et. al., PLDI’05]

27 / 54

Overview

I Expression of API queries as partial expressions

I Algorithm to generate results quickly in ranked order

I Experiment showing simple queries represent real code well

28 / 54

Unknown method queries

I Ex.::?({image, size})

I ⇒ PaintDotNet.Actions.CanvasSizeAction

.ResizeDocument(img, size, �, �)I ⇒ PaintDotNet.Functional.Func.Bind(�, size, img)I ⇒ PaintDotNet.Pair.Create(size, img)I ⇒ PaintDotNet.Quadruple.Create(size, img, �, �)I ⇒ PaintDotNet.Triple.Create(size, img, �)I ⇒ PaintDotNet.PropertySystem

.StaticListChoiceProperty

.CreateForEnum(img, size, �)I ⇒ System.Drawing.Size.Equals(size, img)I ⇒ System.Object.ReferenceEquals(size, img)

29 / 54

Unknown lookup queries

I Ex. float f = pointPair.::*

I ⇒ pointPair.P1.XI ⇒ pointPair.P1.YI ⇒ pointPair.P2.XI ⇒ pointPair.P2.YI ⇒ pointPair.Midpoint.XI ⇒ pointPair.Midpoint.YI ⇒ pointPair.FirstValidValue().XI ⇒ pointPair.Length

30 / 54

Unknown expression queries

I Ex. XmlReader xr =:?

I ⇒ System.Xml.XmlReader.Create(�)I ⇒ new System.Xml.XmlNodeReader(�)I ⇒ System.Data.SqlTypes.SqlXml.Null.CreateReader()I ⇒ new System.Xml.XmlNodeReader(�).ReadSubtree()I ⇒ new System.Xml.XmlValidatingReader(�).ReaderI ⇒ Microsoft.SqlServer.Server.SqlContext

.TriggerContext.EventData.CreateReader()I ⇒ new System.Xml.XmlValidatingReader(�)

.Reader.ReadSubtree()

31 / 54

Partial expression language

(a) e ::= call | varName | e.fieldName | e:=e | e<ecall ::= methodName(e1, . . . ,en)

(b) e ::= a |:? | �

a ::= e | a.:* | call | e:=e | e<e

call ::=:?({e1, . . . ,en}) | methodName(e1, . . . ,en)

I Ex.::?({strBuilder.

:*, e.

:*})

⇒::?({strBuilder, e.StackTrace})

⇒ strBuilder.Append(e.StackTrace)

32 / 54

Partial expression language

(a) e ::= call | varName | e.fieldName | e:=e | e<ecall ::= methodName(e1, . . . ,en)

(b) e ::= a |:? | �

a ::= e | a.:* | call | e:=e | e<e

call ::=:?({e1, . . . ,en}) | methodName(e1, . . . ,en)

I Ex.::?({strBuilder.

:*, e.

:*})

⇒::?({strBuilder, e.StackTrace})

⇒ strBuilder.Append(e.StackTrace)

33 / 54

Algorithm

I Problem: given query, generate completions

34 / 54

Method index by parameter type

Object

2210 methods

Equals

GetHashCode

Registry.SetValue

Array.IndexOf

IList.Add

Console.WriteLine

...

ICloneable

2211 methods

Clone

IList

2257 methods

Add

Remove

...

ArrayList

2299 methods

BinarySearch

Reverse

...

35 / 54

Infinite results

I Problem: too many resultsI inefficient to generate thousands of results to show only 20 to

the programmerI programmer does not want to look at every resultI result set is often infinite

I Ex. var res = foo.:*;

I ⇒ fooI ⇒ foo.GetType()I ⇒ foo.GetType().GetType()I ⇒ foo.GetType().GetType().GetType()I ⇒ foo.GetType().GetType().GetType().GetType()I ⇒ . . .

I Solution: generate in ranked order

36 / 54

Algorithm

I Simple structually recursive algorithm

I Group by type to minimize redundant workI Generate results in ranking order

I Allows determination of top n without computing all results

37 / 54

Heuristics: Type distance

Object

Shape

Rectangle

2

1

IDrawingElement

2

38 / 54

Heuristics: Type distance

Object

Shape

Rectangle

2

1

0

IDrawingElement

2

39 / 54

Heuristics: Length

I Number of field/property lookups or method calls added

I:?({strBuilder.

:*,e.

:*})

Good (1): ⇒ strBuilder.Append(e.StackTrace)

Bad (3): ⇒ strBuilder.Clear().Append(e.Data.Count)

40 / 54

Heuristics: Length

I Number of field/property lookups or method calls added

I:?({strBuilder.

:*,e.

:*})

Good (1): ⇒ strBuilder.Append(e.StackTrace)

Bad (3): ⇒ strBuilder.Clear().Append(e.Data.Count)

41 / 54

Heuristics: Inferred abstract types

Example usages elsewhere in codebase:string f = Path.GetTempFileName(); ...;

File.Delete(f);

File.Delete(Path.Combine(dir, filename));

if(File.Exists(Path.Combine(otherDir, file))) {...}

Query:string p = Path.GetTempFileName();

:?({p})⇒ GetCursor(p)

⇒ File.Delete(p)

⇒ File.Exists(p)

42 / 54

Ranking function

I Linear combination of these and other heuristics

I Sensitivity analysis showed these are most important andcoefficients do not matter much

43 / 54

Outline

Motivation

ApproachLanguageAlgorithmRanking

ExperimentResults

Related work

Conclusion

44 / 54

Experiment

I Automated test of expressiveness of partial expressions

I Generated queries for each call and looked at rank of actualcall in query results

I Advantage: able to do many queries

I Disadvantage: many of the method calls are not ones aprogrammer would need API discovery for

45 / 54

Experiment

I Used Microsoft CCI to disassemble mature C# projectsI Converted every call with at least 3 arguments (including

receiver) to a query with 1 or 2 arguments (including receiver)I For ResizeDocument(document, size, anchorEdge,

background) 16 queries would be generated:⇒

:?(document)

⇒:?(size)

⇒:?(anchorEdge)

⇒:?(background)

⇒:?(document, size)

⇒:?(document, background)

⇒ . . .

I Report rank for best-performing query for each call

46 / 54

Projects used

I Paint.NET image editor

I Windows Installer XML library

I Gnome Do program launcher

I Banshee music player

I .NET core libraries

I Family.Show (WPF example application)

I LiveGeometry geometry visualizer

I Scale: .NET contains 280,000 methods in 30,000 types

I Analyzed 21,176 method calls in these applications

47 / 54

CDF of rank for best method query

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d c

alls

Rank of correct answer is < x

?({foo, bar})

baz.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d c

alls


?({foo, bar})

baz.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d c

alls


?({foo, bar})

baz.

Partial expressions

Code completion

48 / 54

CDF of rank for best method query (correct is static)

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d s

tati

c ca

lls


?({foo, bar})

NS.Baz.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d s

tati

c ca

lls


?({foo, bar})

NS.Baz.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d s

tati

c ca

lls


?({foo, bar})

NS.Baz.

Partial expressions

Code completion

49 / 54

CDF of rank for best method query

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d c

alls


?({foo, bar})

?({foo})

baz.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d c

alls


?({foo, bar})

?({foo})

baz.

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 10 20 30 0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1P

roport

ion o

f an

alyze

d c

alls


?({foo, bar})

?({foo})

baz.Using two arguments

Using one argument

Code completion

50 / 54

Other experiments

I Time: unknown method queries take under 0.1 second

I Ran similar experiments on other partial expression templates

I Similar results: one argument or one lookup could bepredicted within the top 10 about 80% of the time

51 / 54

Related work

I Lots of other work on API discovery discussed in paper

I Prospector (for Java) [Mandelin et. al., PLDI’05]I Input is target type

I Similar to XmlReader xr =:? query

I Uses mined expressions which convert from one type to anotherI Output is chain of mined expressions starting with some local

I Advantage: able to synthesize larger expressionsI Disadvantage: queries only specify a single input type and a

single output type

52 / 54

Related work

I Lots of other work on API discovery discussed in paper

I Prospector (for Java) [Mandelin et. al., PLDI’05]I Input is target type

I Similar to XmlReader xr =:? query

I Uses mined expressions which convert from one type to anotherI Output is chain of mined expressions starting with some local

I Advantage: able to synthesize larger expressionsI Disadvantage: queries only specify a single input type and a

single output type

53 / 54

Contributions

I Expressed API searches in terms of partial expressions

I Leveraged rich type structure to reduce information neededfor queries

I Automated experiments across large codebases show smallpartial expressions often match real method calls

I Created Visual Studio pluginI https://pec.codeplex.com/

54 / 54

https://pec.codeplex.com/

type-directed completion of partial expressions · i paint.net image editor i windows installer xml...

Documents