Approach I: Do Nothing

Transcription

1 Exceptions and Continuations ndling in programming languages is a very limited form on ntinues after a function call that is still active when sed pproach II: Non-Standard Return to modify calls so that they look like this: _f OK to handle exception s mechanism to return a value with the exception, but new complexity for normal return ception: of exception in some standard register or memory locainstruction after normal return r the ia2(above) Easier on machines that allow returnster+constant offset address [why?] andling code decides whether it can handle the excepes another exception return if not equires small distributed overhead for every function 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 4 8: More Special Effects Exceptions and OOP Approach I: Do Nothing ep it simple; don t bother with exceptions code convention: C libraryfunctions often returneither for OK or nonvarious degrees of badness g to check ter s: makes value-returning functions less useful st in always checking return codes 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28

2 Approach III: Discussion pic: Dynamic Method Selection and OOP, call to setjmp appears to return twice, with two dif- uire help from compiler, tation is architecture-specific posed on every setjmp call plement try and catch, therefore, would impose cost ms involving variables that are stored in registers: uf typically has to store such registers, but s the value of some local variables may revert unprepon a longjmp language feature introduced by Simula 67, Smalltalk, he virtual function (to use C++ terminology) classes in a hierarchy of types of subtype is an instance of its supertype(s) ular, inherits their methods, but can override them c effect: Cannot in general tell from program text what ode executed by a given call tion difficulty (as usual) depends on details of a lanantics still static: functions, numbers of arguments are (usually) known can handle overloading by inventing new names for func-,g++encodesafunctionf(intx)inclassqas ZN1Q1fEi, x, int y) as ZN1Q1fEii 46:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 8 pproach III: Stack manipulation ave an exception mechanism built into its syntax, but outines: _point; ) { (catch_point) == ) { ase, which eventually Caller s n to Callee FP, SP, catch point: addr of exception setjmp call & others ) { xception: atch_point, 42); Callee s frame other frames Caller s frame Approach IV: PC tables mplementation uses a different approach nerates a table mapping instruction addresses (program values) to exception handlers for each function ompiler also leaves behind information necessary to refunction ( unwind the stack ) when exception thrown ception E: rent PC doesn t map to handler for E) tack to last caller approach, a try-catch incurs no cost unless there is an ut d handling the exception more expensive than other apd space 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 7

3 racteristics of Dynamic Approach os and Cons of Dynamic Approach stance is independent Contents of class definition until a new value is assigned to an attribute of the ines can be added freely to instances or to class ants of this approach, there are no classes at all, only d we get new instances by cloning existing objects, and adding new attributes lexible simple tion easy ead: every instance has pointers to all methods ad: lookup on each call ecking 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 12 I Fully Dynamic Approach on is completely dynamic: elf): return 42 = A () af() # Prints 2 42 a r, z: rw * z w = 5, ), af, aw # Prints , ), bf, bw # Error (x not a function) # Prints 2 a (self): 19 mplementing the Dynamic Approach egy: just put a dictionary in every instance, and in class stance by making fresh copy of class s dictionary alue of attribute in object s dictionary, then in that of perclass, etc at runtime or pointers) carry around dynamic type ), cf, cv # Prints 19, 2, 1 ), bf, bv # Prints 19, 2, 1 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 11

4 menting the Smalltalk-like Approach ngle Inheritance with Static Types ed not carry around copies of function pointers h class has a data structure mapping method names to d instance-variable names to offsets from the start of ): body1 ): body2 ): body ): body4 a: b: class: class: at offset 8 from start of instance 2 super x@4: super f: body h: body4 y@8: 2 va without interfaces Type can inherit from at most te superclass ss, xw, insist that compiler knows a supertype of x s e that defines w all possible overridings of a method have compatible pas and return values quesimilar to previous one, but put entriesfor all methr or not overridden) in each class data structure ata structures are called virtual tables or vtables in 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 16 ght Single Inheritance, Dynamic Typing s fixed set of methods and instance variables e fixed definition in each class herit from single superclass ypes of parameters, variables, etc, still dynamic nique in Smalltalk, Objective C os and Cons of Smalltalk Approach store modifiable things instance variables in instances ure can be a bit faster at accessing than fully dynamic much static checking possible, and of method names required 46:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 15

5 Interfaces terface inheritance of any number of interface types o new bodies) tes life: consider class B { interface C { int y; f (); () { } g () { } } h () { } public f () { } } */ tends A class B2 extends B plements C implements C { } */ void f (C y) { yf () } // How can this work? Implementation II: Make Interface Values Different roach is to represent values of static type C (an interifferently value x2 of type B2 to C then causes C to point to a antity: o x2 o a cut-down virtual table containing just the f entry at offset ) converting to interface requires work and allocates stor- ile A and B without knowledge of C, A2, B2 make the virtual table of A2 and B2 compatible with that f is at same known offset regardless of whether of C is A2 or B2? (Above isn t hardest example!) 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 2 ation of Simple Static Single Inheritance { body1 } { body2 } ends A { { body } { body4 } ) ) a: b: 2 f: body h: body4 ore offsets of x and y; compiler knows where they are r knows where to find f, g, h in virtual tables ffsets of variables in instances and of method pointers les are known constants, the same for all subtypes nows how to call methods of b even if static type is A! 6:4 218 CS164: Lecture #28 17 rface Implementation I: Brute Force h is to have the system assign a different offset globdifferent function signature s f(int x) and f() have different function signatures) us example, the virtual tables can be: C: : pntr to Bg : unused 4: pntr to Bh 4: unused f 8: pntr to Bf 8: pntr to Cf f f method calls B2: : pntr to Bg 4: pntr to Bh 8: pntr to Bf ze of tables gets big (some optimization possible) ake into account all classes before laying out tables es dynamic linking 46:4 218 CS164: Lecture #28 19

6 oving Interface Implementation II void doing allocation to create value of interface type ultiple Inheritance: What Must Work ust solve the problem of insuring that extend the virtual table of all types to include an inor n this vector identifies an interface the type implehe table (eg C table for B in last slide) t C c = b from last slide, just copy pointer b, as for es when assigning to a variable whose type is a superalue assigned cg() from last slide, find the C table in the interfor object pointed to be c and fetch the entry for g sual the reader: How best to design the interface vector? hing of cg to be fast, avoid having to actually perform a search at execution? llowing work (Java syntax, but not quite Java!), and calls and instance-variable accesses involve small, fast l, non-looping code sequences 9; ) { x += 1; h(x + px); return x; } y) { print(x+y); } 2; ) { y += 1; h(y + py); return y; } z) { print(x*y); } ds A, B { t a) { thisf(this); thisg(this); print(x + y + a); } ; af(a); bg(b); ah(); bh(); df(a); dg(b); dh() 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 24 face Implementation II, Illustrated Full Multiple Inheritance { body1 } { body2 } { body } ; c: h: body multiple inheritance only via interfaces oint: interfaces don t have instance variables riables basically mess everything up for multiple inheriming we want to keep constant offsets to instance vari- { void g (); } ends A s C { } (); b: "interface object" from cvtbl, use cobj as this obj: C table for h: body class B { int y = 42; x h() } void g () { y h() } } void h () { } } class D extends A, B { // Where do x and y go? void h () { } } thenadfexpectsthat this pointstoana,adgexpects s to a B, but adh expects it to point to a D se all be true?? 6:4 218 CS164: Lecture # :4 218 CS164: Lecture #28 2

7 ting Full Multiple Inheritance I (contd) ction address of g from D table first add 8 to pointer value ofad so as to get a pointer art of ad ventually calls h (actually thish), rs to the B part of ad l table is D (B part) in the preceding slide from that table gives us Dh, e call, after first adding the -8 offset from the table nd up calling Dh with a this value that points to ad, cts 6:4 218 CS164: Lecture #28 26 menting Full Multiple Inheritance I tend the contents of the virtual table with an offset hod how to adjust the this pointer before calling es from the last slide: g: body of Bg h: body of Bh D (B part): g: body of Bg h: body of Dh -8 f: body of Af h: body of Ah D: f: body of Af h: body of Dh g: body of Bg 6:4 218 CS164: Lecture # ementing Full Multiple Inheritance II entation slows things down in all cases to accommodate tter if only the methods inherited from B (for example) a work design: use stubs to adjust the this pointer to add 8 to the this pointer and then call Bg; and Dh 1 8 and then call Dh: g: body of Bg h: body of Bh D (B part): g: body of Bg h: body of Dh 1 f: body of Af h: body of Ah D: f: body of Af h: body of Dh g: body of Bg 1 46:4 218 CS164: Lecture #28 27