Python3 Continue Processing After Handling Exeption
You don't understand exceptions, but you should
When writing a compiler, it's hard to get exceptions right.
Exceptions seem straightforward: just unwind the stack until you hit a handler--right?
No. That's "not even wrong."
Exceptions are not fundamentally built upon stacks.
(After all, stackless implementations still have exception-handling.)
Trying to understand exceptions in terms of the stack is like trying to explain the world in purely Newtonian terms: it's a reasonable model, but it fails to capture (rare but important) corner-case behavior.
A "quantum" understanding of exceptions requires escape continuations.
Exceptions interact nontrivially with other language features like while
and for
loops (or rather, with break
and continue
), and even with return
.
When you throw in finally
blocks, the complexity skyrockets.
Consider some Python:
def f(): try: return 10 finally: print("got here!") return 20
The function above returns 10, but prints "got here!"
def f(): try: raise Exception() except: return 10 finally: print("got here!") return 20
The function above returns 10, but still prints "got here!"
def f(): try: raise Exception() except: return 10 finally: print("got here!") return 20
The function above returns 20 (and still prints "got here!")
while True: try: continue finally: print "got here!"
This is an infinite loop that prints "got here!" forever.
In this article, I'll explain how to implement return
, while
, break
, continue
, try
, catch
, throw
and finally
in terms of efficient escape continuations.
To understand exceptions is to implement exceptions.
The provided code uses Racket macros to add all of these features to the language. It desugars them in terms of a single construct: call/ec
.
(The techniques are not Racket-specific; for example, if you're compiling to C, you can fake call/ec
with setjmp
and longjmp
.)
My basic model for exception-handling is lifted directly from Andrew Appel's masterwork, Compiling with Continuations.
Why Python and Racket?
My undergrad compilers class has to implement a compiler for Python.
The intermediate representation used by their compiler is a cut-down Racket extended with Pythonic control-flow constructs.
That's why you'll find motivating examples in Python that get desugared into a Racket-like code.
A note on call/ec
This code makes heavy use of call/ec
.
Not all languages support escape continuations, but many support the (more general) full continuations. Full continuations work in place of escape continuations with a mild performance penalty.
You can also simulate call/ec
in C with setjmp
and longjmp
.
Implementing return
Newcomers to functional programming are often struck by the lack of a return
statement (and, I suppose, by the lack of statements entirely).
There are natural coding patterns that exploit the ability to bail out of a function early, like the following:
def f(): if easy case: return this if also easy case: return that do something complicated down here
Fortunately, it's easy to add a return statement to functional languages that support continuations.
A return statement is a "second-class escape continuation."
Conversely, think of an escape continuation as a first-class return statement.
That is, imagine that return
was just a special procedure bound every time you entered a function, and that to return from the function, you passed the return value to this special procedure.
In Racket, we can grab the escape continuation for a function with call-with-escape-continuation
, or call/ec
for short.
If we want to add a define/return
form to the language so that the following works:
(define/return (max a b) (when (> a b) (return a)) (return b))
we can create a macro that desugars it into:
(define (max a b) (call/ec (lambda (return) (when (> a b) (return a)) (return b))))
In Racket:
(define-syntax (define/return stx) (syntax-case stx () [(_ f-params body ...) ; => (with-syntax ([return (datum->syntax #'f-params 'return)]) #'(define f-params (call/ec (λ (return) body ...))))]))
It's just as easy to add a lambda/return
form.
Implementing while
A while
loop seems easy to implement in terms of tail-recursion or gotos.
In fact, it is.
But, break
and continue
complicate matters--especially when it comes time to implement exceptions.
Fortunately, break
and continue
, like return, are also escape continuations.
It's also best to implement these features as escape continuations.
Take the general while
loop form in Python for example:
while cond: body else: otherwise
This would be desugared in Racket as:
(call/ec (λ (break) (letrec ([loop (λ () (when cond (call/ec (λ (continue) body)) (loop)))]) (loop) otherwise)))
A call to break
bails out of the loop, skipping over the else case.
A call to continue
jumps to the next iteration of the loop.
A macro to transform (while cond body else)
forms is:
(define-syntax (while stx) (syntax-case stx () [(_ cond body else) ; => (with-syntax ([break (datum->syntax #'body 'break)] [continue (datum->syntax #'body 'continue)]) #'(call/ec (λ (break) (letrec ([loop (λ () (when cond (call/ec (λ (continue) body)) (loop)))]) (loop) else))))]))
Implementing try
Exceptions are the canonical example of escape continuations.
A reasonable approach for implementing exceptions is to create a special "handler" cell in memory. That cell contains the escape continuation to invoke once an exception is raised.
Let's start by compiling a (try body handler)
form. This is a try
form without finally
code.
We're going to add two global procedures to help--current-handler
and set-current-handler!
:
(define $current-handler (lambda (ex) (error "top-level exception!"))) (define (current-handler) $current-handler) (define (set-current-handler! handler) (set! $current-handler handler))
The variable $current-handler
holds a pointer to the current exception handler.
An isolated desugaring of the try
form is not hard:
(let ([$old (current-handler)]) (call/ec (lambda (ec) (set-current-handler! (lambda (ex) (set-current-handler! $old) (ec (handler ex)))) (let ([rv body]) (set-current-handler! $old) rv))))
This code:
- saves the old exception handler in
$old
; - captures the current escape continuation in
ec
; - sets a new handler that invokes
ec
; and - runs the body.
Whether the body returns naturally, or by invoking an exception, the old handler gets restored.
But, what happens if we have code like the following?
def f(): try: return 10 except: print("you should never see me")
This function leaves the wrong handler installed after it returns!
This is because return
bypasses the clean-up routines we established with the desugaring of try
, leaving the handler installed.
To fix this, we need to bind new versions of return
(and break
and continue
) that perform handler clean-up before exiting the loop:
(let ([$old (current-handler)]) (let* ([return (λ args (set-current-handler! $old) (apply return args))] [continue (λ () (set-current-handler! $old) (continue))] [break (λ () (set-current-handler! $old) (break))]) (call/ec (λ (ec) (set-current-handler! (λ (ex) (set-current-handler! $old) (ec (handler ex)))) (let ([rv body]) (set-current-handler! $old) rv)))))
But, if we want to allow a finally
expression, through an expanded try
form (try body handler finally)
, things get more complicated.
The finally
code must run when we leave the dynamic extent of the try
.
It must run whether we return out of it, whether we break out of it, whether we continue out of it, or where we raise out of it.
But, what it must do after the finally
code depends on how we got into it:
- If we fall through by exiting the
try
body normally, it should continue after thefinally
. - If we return out, then after the
finally
code executes, it should immediately return the provided return value. - If we break out, then it should run the
finally
code and jump out of the enclosing loop form. - If we continue out, then it should run the
finally
code and jump to the top of the enclosing loop form. - If we raise out, then it should run the
finally
code and re-raise the exception--unless the exception was handled, in which case it should fall through the bottom offinally
.
We can make all of this possible by creating a pointer to a thunk that should perform the post-finally action.
We need to capture and redirect all escaping continuations to pass through finally
by having them install their ordinary behavior in that thunk.
There's a further wrinkle in that if during the course of the exception handling, we throw another exception, we still have to run through the finally
code.
In other words, the exception handler has to install an exception handler.
Here's the desugaring to make it all happen:
(let* ([$val (void)] [$fin (λ () $val)] [$old (current-handler)]) (call/ec (λ (fin) (let* ([return (λ args (set! $fin (λ () (apply return args))) (fin))] [continue (λ () (set! $fin continue) (fin))] [break (λ () (set! $fin break) (fin))]) (call/ec (λ (ec) (set-current-handler! (λ (ex) ; if the handler ; throws an exception, ; re-throw it after ; the finally block: (set-current-handler! (λ (ex*) (set! $fin (λ () (throw ex*))) (ec #f))) (ec (let ([rv (handler ex)]) (set! $fin (λ () rv)))))) (let ([rv body]) (set! $fin (λ () rv)) (fin))))))) (set-current-handler! $old) (set! $val finally) ($fin)))
We also need to consider a try
form without an exception handler, (try body finally)
. Fortunately, all it has to do is re-raise the exception:
(let* ([$val (void)] [$fin (λ () $val)] [$old (current-handler)]) (call/ec (λ (fin) (let* ([return (λ args (set! $fin (λ () (apply return args))) (fin))] [continue (λ () (set! $fin continue) (fin))] [break (λ () (set! $fin break) (fin))]) (call/ec (λ (ec) (set-current-handler! (λ (ex) ; re-throw after finally: (set! $fin (λ () (throw ex))) (fin))) (let ([rv body]) (set! $fin (λ () rv)) (fin))))))) (set-current-handler! $old) (set! $val finally) ($fin)))
Implementing throw
Compared to all the excitement above, throw
desugars with a wimper:
(define (throw ex) ((current-handler) ex))
Code
Macro-expanders for all of the above are available in compiling-with-exceptions
.
More resources
- I learned everything I know about compiling exceptions from Appel's Compiling with Continuations.
- My post on programming with continuations.
- My post on continuation-passing style in JavaScript.
- Reader Travis Cardwell pointed me to the this excellent resource by Graham Hutton and Joel Wright: Calculating an Exception Machine [video].
Source: https://matt.might.net/articles/implementing-exceptions/
0 Response to "Python3 Continue Processing After Handling Exeption"
Publicar un comentario