Altering Java Semantics via Bytecode Manipulation Eric Tanter, - PowerPoint PPT Presentation

Jinline Altering Java Semantics via Bytecode Manipulation ´ Eric Tanter, Marc S´ egura-Devillechaise, Jacques Noy´ e, Jos´ e Piquer { etanter, jpiquer } @dcc.uchile.cl { msegura, jnoye } @emn.fr. DCC–University of Chile (Chile), OBASCO–Ecole des Mines de Nantes (France) E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.1/17

✁ � � � ✁ � ✁ ✁ Contents An Example Introduction Jinline Approach Overview Process Guarantees Current state, perspectives & conclusion E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.2/17

� ✁ ✁ ✁ ✁ An Example Set up a Factory pattern [GoF]: server handling requests with concurrent workers of different types available implementation starts a new thread for each request want to have a factory managing an instance pool replace new WorkerXXX(arg1, arg2, ...); with a generic code: Factory.getInstance("WorkerA", [..args..]); E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.3/17

✁ � ✁ ✁ ✁ � ✁ Introduction (1) Altering semantics of Java programs Areas of application: Application customization OTS Component integration Separation of concerns (reflection, AOP) Need for performing alterations on binaries (Java bytecode) binary components open distributed systems with dynamic class loading E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.4/17

✁ � � ✁ ✁ ✁ ✁ Introduction (2) Granularity of alterations Coarse-grained: outside methods class hierarchies, interfaces, class members, ... e.g. structural reflection with Javassist [Chiba], BCA [Keller] Fine-grained: within method bodies sometimes necessary (e.g., previous example) as of now, either low level, unsafe • general-purpose bytecode manipulation (BCEL [Dahm], JOIE [Cohen]...) • bytecode = language semantics + stack code or limited (not generic) • Javassist’ code converter • new WorkerA(a1, a2); → Factory.getWorkerA(a1, a2); E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.5/17

J i n l i n e E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.6/17

✁ � ✁ ✁ � ✁ ✁ � � ✁ Approach fine-grained alterations on bytecode load-time transformation by a metaprogram plugged into the Javassist framework source-level abstractions ease of use ensure consistency (no direct, error-prone, stack manipulation) high expressiveness approach can be used to transform almost any source language mechanism genericity reify language mechanism occurrences at run-time if needed E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.7/17

� � ✁ ✁ ✁ Overview transformation of/around language mechanisms occurrences inserting code before/after and instead of conceptually, insert a call to a method concretely, inline that method E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.8/17

Transformation i−1 i−1 i prologue i+1 adjusted body epilogue original bytecode sequence i+1 bytecode sequence after inlining body to inline E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.9/17

� ✁ ✁ Overview (2) abstraction level: source code specify target of transformation at source level: description classes specify inserted code as source code: standard methods E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.10/17

� ✁ ✁ � ✁ ✁ Descriptions Language mechanisms are represented as description classes MessageSend.class Instantiation.class An occurrence of a language mechanism is an instance of a description class aMessageSend.getReceiverType() anInstantiation.getArgumentsTypes() E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.11/17

� ✁ � Inlined code The inlined code is a standard, compiled method Restriction on the arguments: • either object array (all dynamic info) or no-arg (more efficient) • Object newMethod(Object[] jinArgs){ String classname = (String) jinArgs[2]; Object[] args = (Object[]) jinArgs[3]; return Factory.getInstance(classname, args); } • newMethod(){ System.out.println("here"); } code can be inlined in constructors, try, catch, finally, ... solution to the Factory example is given in the paper (< 15 lines) E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.12/17

Process of the metaprogram specify mechanisms of interests - filter occurrences - specify method to inline notifyUpon(type, jinlerA) 1 Jinliner process(myMethod) 2 return inlining 5 notify(description, context) 3 order (or null) analyze static data 4 jinlerA E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.13/17

✁ � ✁ ✁ ✁ � � ✁ Guarantees User has no direct access to bytecode generated code is valid (passes bytecode verification) Signature of methods to inline are checked args: either Object[] or void Integration of inserted code stack state for instruction i + 1 is left intact return value • automatic downcast and unwrapping of primitive types • if mismatch, then runtime ClassCastException thrown exceptions • injection within original exception handling code • if not expected, wrapped into UndeclaredThrowableException E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.14/17

✁ ✁ ✁ ✁ ✁ � � ✁ Perspectives (1) Code explosion might become an issue Jinline overhead is not the problem size of the inlined code is an issue detection is straightforward possibility of resolving this issue by inserting standard method call instead of inlining Composition of transformation 3 possibilities: • ignore Jinline code and inlined code • ignore Jinline code and transform inlined code • transform both solution based on the use of method attributes E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.15/17

✁ � ✁ ✁ ✁ ✁ � ✁ ✁ � Perspectives (2) Support for more mechanisms as of now: message/constructor send, instantiation and cast soon: field/array accesses, RTTI later: method enter/exit, try/catch/finally enter, ... Support for before/after as of now only instead of before/after is easier Applications Reification components for Reflex, an open runtime MOP [Tanter] • provide caller-side metacontrol Implementation alternative for EAOP [Douence] E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.16/17

� � � � � ✁ Conclusion fine-grained alterations no need for source code (bytecode / load time) source-level asbtractions complements Javassist prototype ready, but packaging is not contact etanter@dcc.uchile.cl or msegura@emn.fr E. Tanter et al. – Altering Java Semantics via Bytecode Manipulation – ACM GPCE 2002, Pittsburgh, MA, 6-8 October – p.17/17