1 / 47

Two techniques for programming by sketching (Stanford, November 2004)

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 Sarkar IBM. Synthesis. Program synthesis

hestia
Download Presentation

Two techniques for programming by sketching (Stanford, November 2004)

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  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 Sarkar IBM

  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. program = completed sketch sketch Programming by sketching • 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!) • StreamBit: • behavioral spec + sketch  full implementation • 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. SLOW O(w) functionality sketch FAST implementation +  ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? 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 FAST O(log w)

  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% speedupover 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. 3 x y z 2 1 0 0 0 1 0 x y x = consumes a 3-bit chunk of input; produces a 2-bit of output. The behavioral spec (StreamIt) • StreamIt • synchronous dataflow language • filters represented internally as matrices

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

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

  16. 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 01 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 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 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 F.F_1 specify FAST bit shifting algorithm w/out sketching: 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 F.F_1 F F.F_2 F.F_3 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 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 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 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 F.F_2 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_3

  17. 32 bits 32 bits 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 Sketching: another example A permutation from DES cipher (64 bits  64 bits) Problem: when implemented as a table lookup, the table is very large • Idea: decompose into a pipeline of two permutations: • provided by the programmer: an inexpensive permutation • automatically derived from the sketch: two identical permutations (to be implemented as one smaller table)

  18. Start with a sketch Define xi,jas the amount bit i will move on step j Semantic equivalence imposes linear constraints on the xi,j Many of the constraints in the sketch also impose linear constraints on xi,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 ); Sketching: How it works

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

  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. Example programming task: parse a Java file into an AST IFile file = … ICompilationUnit cu = JavaCore.createCompilationUnitFrom(file); ASTNode node = AST.parseCompilationUnit(cu, false); IFile file = … ASTNode node = ? Why so hard to find? (productivity: 2LOC/hour) • class member browsers? two unknown classes used • follow expected design? two levels of file handlers • 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’ • 1st observation • Reuse problems can usually be described with a have-one-want-one queryq=(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 • 2nd 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)

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

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

  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. getSelection() Viewer’ getViewer() downcast ISelection’ IStructuredSelection’ getFirstElement() downcast Object’ downcast downcast Mining jungloids with downcasts (example) 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<JavaInspectExpression> Viewer<IStructuredSelection<JavaInspectExpression>> IDebugView getViewer() Viewer getSelection() getInput() ISelection IStructuredSelection getFirstElement() Object JavaInspectExpression

  39. x.a.(T) y.a.(S) 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

  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 • StreamBit: • behavioral spec + sketch  full implementation • goal: total correctness and performance • 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)

More Related