1. Two techniques for programming by sketching (Stanford, November 2004) Rastislav Bodik, David Mandelin, Armando Solar-Lezama, Lin Xu UC Berkeley Rodric Rabbah MIT Kemal Ebcioglu, Doug Kimelman, Vivek SarkarIBM

2. Synthesis Program synthesis –given a specification, synthesize a program meeting this spec –synthesis inverse to verification –most work in reactive systems (Pnueli, Kupferman, …) Synthesis vs. compilation –synthesis involves a search for the desired program Benefits –“less coding, more correctness”

3. Programming by sketching sketch program = completed sketch  Our approach –apply synthesis to software –“sketching”: specification is partial (underspecified)

4. Two sketching techniques Sketch: –partial implementation, provided by programmer Sketch resolution: –completing the sketch into a full implementation –which one? (sketch completes into many implementations!) 1.StreamBit: –behavioral spec + sketch  full implementation 2.Prospector: –sketch  several full implementations –user selects implementation with desired behavior

5. StreamBit: Sketching high-performance implementations of bitstream programs Project lead: Armando Solar-Lezama

6. Bitstream Programs Bitstream programs: a growing domain –crypto: DES, Serpent, Rijndael, … –coding in general, NSA/BitTwiddle Bitstream programs operate under strict constraints –performance is very important up to 95% of server cycles spent in security-related processing –correctness is crucial subtle bug in Blowfish implementation allowed over half the keys to be cracked in less than 10 minutes

7. Example “Drop every third bit in the bit stream.” exhibits many features of complicated permutations –exponentially many choices –greedy choice is suboptimal fast implementation can be sketched SLOW O(w) FAST O(log w) functionalityFAST implementation + 

8. Full sketch (13 lines of code) WSIZE=16; subsequence = Unroll[WSIZE](subsequence); subsequence = PermutFactor[ [shift(1:2 by 0), shift(17:18 by 0), shift(33:34 by 0)], [shift(1:16 by ?), shift(17:32 by ?), shift(33:48 by ?)] ] ( subsequence ); subsequence.subsequence_1=DiagSplit[WSIZE](subsequence); for(i=0; i<3; ++i) { subsequence.subsequence_1.filter(i) = PermutFactor[ [shift(1:16 by 0 || 1)], [shift(1:16 by 0 || 2)], [shift(1:16 by 0 || 4)] ]( subsequence.subsequence_1.filter(i) ); } Size: 13 lines Compare with 100+ lines of such FORTRAN code (from BitTwiddle)... DATA MASKB2 /Z'FFC003FF000FFC00', Z'3FF000FFC003FF00', Z'0FFC003FF000FFC0', Z'03FF000FFC003FC0',... c Compress 5-bit groups together TB = IAND(TB + ISHFT(TB, SKIPBC), MASKB2(J)) TC = IAND(TC + ISHFT(TC, SKIPBC), MASKC2(J))...

9. What you gain DropThird benchmark: –Speedups over naïve code with a 14 line sketch: 32 bit on a Pentium IV: 83.8% 64 bit on an Itanium II: 233% DES benchmark: –32 bit on a Pentium IV with 30 line sketch: 634% speedup over naïve within 11% of hand optimized libDES –64 bit IA64 and IBM SP2 we beat libDES by 8%

10. What is sketching Key idea: separation of concerns –specify behavior without concern for performance –create implementation without concern for bugs domain expert: –writes a behavioral specification of her crypto algorithm –as clean as possible, no optimizations performance expert: –describes an efficient implementation of the clean algorithm –neither reimplements nor describes in full –he only sketches an outline of the implementation; compiler fills in details –if sketch is wrong, compiler complains  no bugs can be introduced

11. Compilation strategy A sketch overrides a naïve compiler: –naïve compiler translates the clean algorithm into target code, with a simple sequence of semantics-preserving transformations: (1) make all filters word-size (unroll and split) (2) decompose word-size filters into machine instructions –sketch “inserts” a step into the naïve sequence Ex.: sketch decomposes a filter into a pipeline of filters after sketch is applied, naïve compiler continues

12. The behavioral spec (StreamIt) StreamIt –synchronous dataflow language –filters represented internally as matrices 1 0 0 0 1 0 3 2 consumes a 3-bit chunk of input; produces a 2-bit of output. xyzxyz xyxy x =

13. Naïve compilation Example: Drop Third Bit (word size W = 4 bits) –Unroll filter –decompose into filters operating on W=4 bits of input. –decompose into filters producing W=4 bits of output

14. Naïve compilation (cont.) Make each filter correspond to one basic operation available in the hardware 1 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 duplicate or 0 0 0 0 1 0 0 0 t1 = in AND 1100 t2 = in SHIFTL 1 t3 = t2 AND 0010 out = t1 OR t3 in

15. The Full Picture Level of abstraction (high low) Implementations Task Description

16. F.F_1 Decomposition without sketching User provides high level decomposition of F into F.F_i System Takes care of compiling F.F_i Correctness is guaranteed as long as [F.F_3]  [F.F_2]  [F.F_1] = F Avoid spelling out the decomposition: Sketch It! [some properties]  [some properties]  [some properties] = F 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 F F.F_2 F.F_3 F.F_1 F.F_2 F.F_3 specify FAST bit shifting algorithm w/out sketching:

17. Sketching: another example A permutation from DES cipher (64 bits  64 bits) 32 bits Problem: when implemented as a table lookup, the table is very large Idea: decompose into a pipeline of two permutations: 1. provided by the programmer: an inexpensive permutation 2. automatically derived from the sketch: two identical permutations (to be implemented as one smaller table) shift(1:64 by 0 || 33 || -33), shift(1:2:31 by -33), shift(34:2:64 by 33), [] // unspecifed; filled in by compiler Sketch

18. Sketching: How it works Start with a sketch Define x i,j as the amount bit i will move on step j Semantic equivalence imposes linear constraints on the x i,j Many of the constraints in the sketch also impose linear constraints on x i,j Solving the linear constraints produces a space of possible solutions Map the nonlinear constraints to this solution space Search SketchDecomp[ [shift(1:32 by 0 || 1)], [shift(1:32 by 0 || 2)], [shift(1:32 by 0 || 4)], [shift(1:32 by 0 || 8)] ]( Filter );

19. User Study (time to first solution) C StreamBit

20. User Study (developing a good implementation)

21. Implementing the fastest DES How fast can we match the fastest DES implementation? –6 different implementations in 4 hours –includes all but one trick used in libDES –so fast partly because sketching avoids bugs

22. Concluding Remarks StreamBit allows for –Task specification oblivious to performance –Implementation specification without bugs Same idea may apply in other domains –If people currently resort to very low level coding –If some algebraic structure can be imposed on the task –It may be amenable to implementation sketching.

23. Mining Jungloids: Helping to Navigate the API Jungle Project lead: David Mandelin

24. A software reuse problem big components reusable [Lampson’99] –OS, DBMS, browser small components challenging –flexibility: functionality cut finely, for fine control –size: in J2SE, 21,000 methods in 1000s of classes cost to understand and use –one of three obstacles to reuse [Lampson’99] searching for information –nearly ¼ of developer time [metallect.com]  often give up reuse and reimplement

25. IFile file = … ICompilationUnit cu = JavaCore.createCompilationUnitFrom(file); ASTNode node = AST.parseCompilationUnit(cu, false); Example programming task: parse a Java file into an AST IFile file = … ASTNode node = ? Why so hard to find? (productivity: 2LOC/hour) 1.class member browsers? two unknown classes used 2.follow expected design? two levels of file handlers 3.grep? method returns a subclass

26. The morale? type signatures –not very useful in finding desired code –but once found, can be used to verify so why not search existing code base? –somebody must have written these two lines before! –yes, but not in same method for software engineering reasons –or even same program e.g.: parse an editor buffer, not a file still, sample code useful, as we will see …

27. Our goal We want a programmer’s “search engine” that –doesn’t merely find an example code –instead, it synthesizes the desired code –from two favorite sources: type signatures existing code examples

28. More precisely mining input: –the API (type signatures from class definitions) –corpus of API client code search input: –a query specifying programmer’s intent output: –synthesized code –ready for insertion into user program –give several candidates (user selects one)

29. Formulating the code search problem We must decide on the structure of: –input query (coding intent) easy to express for the user yet specific enough for the search engine –output code (synthesized code) easy to understand and validate (by reading docs) code should complete the program under construction

30. The query: from ‘have’ to ‘want’ 1 st observation –Reuse problems can usually be described with a have-one-want-one query q=(h,w): “What code will transform a (single) object of (static) type h into a (single) object of (static) type w?” Our parsing example: q = ( IFile, ASTNode ) IFile file = … ICompilationUnit cu = JavaCore.createCompilationUnitFrom(file); ASTNode node = AST.parseCompilationUnit(cu, false);

31. Output code: jungloid 2 nd observation: –most queries can be answered with a jungloid jungloid: –a unary expression composed of unary expressions: field access call to an instance method with 0 arguments call to a static method or constructor with 1 argument conversion to supertype (multi-argument methods decomposed into unary ones) IFile file = … ICompilationUnit cu = JavaCore.createCompilationUnitFrom(file); ASTNode node = AST.parseCompilationUnit(cu, false);

32. Coverage An informal experiment: –using 16 coding headaches, collected by us Can the query express interesting problems? –yes, for 12 out of 16 coding problems Can queries be answered with a jungloid? –yes, all 12 queries answered with jungloids 9 of them are simple jungloids 3 of them use some multi-argument methods

33. Prospector: our prototype Eclipse plugin –integrated with “code completion assist” var.[CTRL+SPACE] –the “want” type w WantType x = [CTRL+SPACE] –a set H of “has” types obtained from context local variables, arguments, class fields, globals –issue queries (h,w) for each h  H field foo() bar(int len, Object key) field foo() bar(int len, Object key)

34. Type signature graph Any path from h to w is a (h,w)-jungloid 3 rd observation: –desired jungloid typically among k shortest paths (k=5) IFileCompilationUnit ICompilationUnit ASTNode IClassFile JavaCore.createCompilationUnitFrom() AST.parseCompilationUnit() supertype AST.parseCompilationUnit() JavaCore.createClassFileFrom() IJavaElementIResource supertype getResource() IContainer getParent()

35. Jungloids with downcasts IDebugView debugger =... Viewer viewer = debugger.getViewer(); IStructuredSelection sel = (IStructuredSelection) viewer.getSelection(); JavaInspectExpression expr = (JavaInspectExpression) sel.getFirstElement(); IDebugView Viewer ISelection IStructuredSelection JavaInspectExpression Object getViewer() getSelection() getFirstElement() getInput() downcast

36. Our solution Besides downcasts, this problem appears in –method arguments of type Object (only accept a JavaBean) –String objects (strings are highly polymorphic) Potential solutions –parametric type inference, alias analysis Our solution –mine a corpus of API uses for legal downcasts

37. Mining jungloids with downcasts Ideally, only correct jungloids are synthesized –correct = it must be possible to write a client code in which the jungloid’s downcast succeeds, for at least one input This ideal can be approximated (overview): –use a corpus of API client code –extract jungloids with downcasts –use them to extend the signature graph In the limit, we meet the ideal –limit = infinitely large, bug-free corpus bug-free corpus –weak requirement: jungloids in corpus to succeed for one input

38. Mining jungloids with downcasts (example) IDebugView Viewer ISelection IStructuredSelection JavaInspectExpression Object getViewer() getSelection() getFirstElement() getInput() Viewer’ ISelection’ IStructuredSelection’ Object’ getViewer() getSelection() getFirstElement() downcast protected IJavaObject getObjectContext() IWorkbenchPage page = … IWorkbenchPart part = page.getActivePart(); IDebugView view = (IDebugView) part.getAdapter(); ISelection s = view.getViewer().getSelection(); IStructuredSelection sel = (IStructuredSelection)s; Object selection = sel.getFirstElement(); JavaInspectExpression exp = (JavaInspectExpression) selection;... } protected IJavaObject getObjectContext() IWorkbenchPage page = … IWorkbenchPart part = page.getActivePart(); IDebugView view = (IDebugView) part.getAdapter(); ISelection s = view.getViewer().getSelection(); IStructuredSelection sel = (IStructuredSelection)s; Object selection = sel.getFirstElement(); JavaInspectExpression exp = (JavaInspectExpression) selection;... } IStructuredSelection Viewer >

39. The jungloid mining algorithm (key idea) When extracting jungloids, how to determine the necessary downcast context (i.e., jungloid suffix)? w.x.a.(T) s.y.a.(S) What if the context is too short? –unsound: a query may synthesize a jungloid that will throw exception in any client code What if the context is too long? –incomplete: a query may fail to synthesize the jungloid even though the corpus contains the necessary example x.a.(T) y.a.(S)

40. Experiment 1 (ranking test) hypothesis: –to find the desired code, the user needs to examine only top 5 candidate jungloids. result: –desired code in “top 5” 17 out 20 times (10 out of 20, in “top 1”) –remaining three fixable methodology: –used 20 real-world coding tasks –collected from FAQs, newsgroups, our practice, emails to us

41. Experiment 2 (user study) hypothesis: –Prospector-equipped programmers are better at solving API programming problems than other programmers methodology: –6 problems, each user did 3 with Prospector and 3 without –problems formulated not to reveal the query –sample problem: “The new Java channel IO system represents files as channels. How do I get a channel that represents a String filename?” –somewhat sparse data (10 users), surveys still trickling in

42. Experiment 2 (user study). Results. Prospector shortens development time –some problems solved only by Prospector users –when both groups succeeded, Prospector users 30% faster Prospector may help enable reuse –non-Prospector users sometimes reimplemented Prospector may help avoid making mistakes –mistakes applying code found on internet into own code We expect even stronger results on a more robust infrastructure.

43. Future work Coding task we currently can’t handle: –print an AST as Java source The limitation: –task is expressible as a (have,want) query –but result is not a jungloid (as defined in this talk) ASTNode ast =... ASTFlattener visitor = new ASTFlattener(); ast.accept(visitor); String result = visitor.getResult(); ASTNode ast =... ASTFlattener visitor = new ASTFlattener(); ASTFlattener visitor2 = ast.accept(visitor); String result = visitor2.getResult();

44. Try it! Web demo –snobol.cs.berkeley.edu Eclipse plugin –coming soon –want to alpha test it?

45. Conclusion Sketch: –partial implementation, provided by programmer 1.StreamBit: –behavioral spec + sketch  full implementation –goal: total correctness and performance 2.Prospector: –sketch  several full implementations –user selects implementation with desired behavior –goal: software reuse

46. Backup slides

47. Programming with jungloids NodeItem node = (NodeItem) getModel(); GraphNodeFigure f = (GraphNodeFigure) getFigure(); f.getLabel().setName(node.getNodeName()); Rectangle r = new Rectangle(node.x, node.y, -1, -1); GraphicalEditPart parent = (GraphicalEditPart) getParent(); parent.setLayoutConstraint(this, f, r)