From: Greg Colvin (gcolvin_at_[hidden])
Date: 2001-03-19 12:21:54
From: Peter Dimov <pdimov_at_[hidden]>
> From: "Greg Colvin" <gcolvin_at_[hidden]>
>
> > > I agree that an exception is not appropriate. I'm arguing that (1)
> > > undocumented "assert" uses are bad, and (2) the proper way to document
> an
> > > "assert" is by using the term "undefined behavior."
> >
> > So the documentation is this? "If there is an error in my code,
> > or the compiler that compiled it, or the system code that it depends
> > on, or the hardware that the system code depends on, or any other
> > unforseen circumstance, then undefined behavior might occur. Use at
> > your own risk."
>
> No.
>
> When the code does not work like the documentation says it will work,
> including any documented "undefined behavior" cases, then it has a bug.
>
> I say that we should separate error reporting from bug reporting. It may not
> be appropriate to use the same mechanism for both.
>
> In other words:
>
> Error class #1: the programmer that wrote the application (that calls the
> library) has a bug in his/her code.
>
> Library Action: (documented) undefined behavior.
>
> Error class #2: a runtime error occurred in the library that the application
> programmer should handle.
>
> Library Action: (documented) throw an exception.
>
> Error class #3: the library, or the compiler, or the OS has a bug that can
> be detected. Depending on the situation, the bug may or may not prevent the
> library from functioning properly.
>
> Library Action: (undocumented) report the error via the appropriate
> mechanism.
>
> My point is that a plain assert is fine for #1, but may or may not be the
> best choice for #3.
Aha. We are in violent agreement then, and discussion comes down
to what is the appropriate mechanism? I still think plain assert
may be fine, especially if testing and review are supposed to have
caught whatever problems exist. A more configurable mechanism would
also be fine, but I would like to see any such mechanism boostified
too, rather than provided ad hoc for just the thread library.