[shell32 5/6] include: Add remotable stubs for non-default-marshallable shobjidl calls.

[Rob Shearman]

2009/1/1 Michael Karcher <wine at mkarcher.dialup.fu-berlin.de>:
> Am Donnerstag, den 01.01.2009, 20:43 +0000 schrieb Rob Shearman:
>> No, the [string] attribute in this case in redundant and it should
>> apply to the first pointer in the parameter.
> Now, that makes sense.
>
>> I'm guessing by the
>> change that you had to make that widl doesn't handle this too well...
> Seems so. I don't want to hurt anyone, but I somehow get the impression
> that the quality of widl is somewhere around works-for-me.

The trouble in this case was a corner case that was only tested by
some bad IDL. Corner cases are never easy to test for and the best way
of fixing bugs is to eliminate them as much as possible. To that end I
have some patches queued up that add more of an API to access the
parse tree to try to make the generator code more robust. In
particular, the C generator code could probably do with a function
like type_get_ndr_type that deals with the nuances of detecting
strings, user-types, context handles, interface pointers, etc and that
would likely have fixed several of the issues you reported.

In terms of testing, I want to find and fix these issues before users
(i.e. other developers, like you) find them. To that end, I think a
good test may be to run "widl -p" on as many of the files in include/
as possible and fix/log any issues that come up, and then to compile
each file. There should be no errors in either case, since errors
should be reported up front (i.e. even for just generating a header
file or a syntax check). I also plan to write a fuzzer for widl
someday - if anyone knows of a free fuzzer framework out there
somewhere then that would be useful.

-- 
Rob Shearman