[Python-Dev] with-statement heads-up (original) (raw)

Mike Bland mbland at acm.org
Tue Feb 28 19:25:41 CET 2006

On 2/28/06, Guido van Rossum <guido at python.org> wrote:

On 2/28/06, Mike Bland <mbland at acm.org> wrote: > On 2/28/06, Guido van Rossum <guido at python.org> wrote: > > I just realized that there's a bug in the with-statement as currently > > checked in. exit is supposed to re-raise the exception if there > > was one; if it returns normally, the finally clause is NOT to re-raise > > it. The fix is relatively simple (I believe) but requires updating > > lots of unit tests. It'll be a while. > > Hmm. My understanding was that exit was not to reraise it, but > was simply given the opportunity to record the exception-in-progress.

Yes, that's what the PEP said. :-( Unfortunately the way the PEP is specified, the intended equivalence between writing a try/except in a @contextmanager-decorated generator and writing things out explicitly is lost. The plan was that this: @contextmanager def foo(): try: yield except Exception: pass with foo(): 1/0 would be equivalent to this: try: 1/0 except Exception: pass IOW with GENERATOR(): BLOCK becomes a macro call, and GENERATOR() becomes a macro definition; its body is the macro expansion with "yield" replaced by BLOCK. But in order to get those semantics, it must be possible for exit() to signal that the exception passed into it should not be re-raised. The current expansion uses roughly this: finally: ctx.exit(*exc) and here the finally clause will re-raise the exception (if there was one). I ran into this when writing unit tests for @contextmanager.

This may just be my inexperience talking, and I don't have the code in front of me right this moment, but in my mind these semantics would simplify the original version of my patch, as we wouldn't have to juggle the stack at all. (Other than rotating the three exception objects, that is). We could then just pass the exception objects into exit without having to leave a copy on the stack, and could forego the END_FINALLY. (I think.) Does that make sense?

Mike

More information about the Python-Dev mailing list