long->sal_Int32 in tools/gen.hxx

[Caolán McNamara]

On Thu, 2018-04-05 at 09:53 +0200, Stephan Bergmann wrote:
> <https://cgit.freedesktop.org/libreoffice/core/commit/?id=8bc951daf79
> decbd8a599a409c6d33c5456710e0> 
> "long->sal_Int32 in tools/gen.hxx" changed tools/gen.hxx classes
> like Point and Rect from being based on long members to being based
> on sal_Int32 members.  It was probably a good idea to make the type
> of those members platform-independent, so that behavior will not 
> unnecessarily differ among platforms.  Relevant choices for 
> platform-independent type were apparently sal_Int32 and sal_Int64. 
> Choosing sal_Int32 seems reasonable, given that this choice
> obviously worked well already for all platforms where long is 32 bit
> (which includes all variants of Windows, x86 and x86-64).  Or did it?

> Other fallout from that "long->sal_Int32 in tools/gen.hxx" commit
> are <https://cgit.freedesktop.org/libreoffice/core/commit/?id=9762c9a
> b98a2b9ea86186a6da7b77031f1416bed> 
> "ofz: lots of integer overflow noise" 

FWIW, there were 28 new Integer overflow findings in oss-fuzz in the
last ~24 hours. Looks like having them as 64bit (by pure historical
accident) was really helpful in sanitizing external data. Lots of the
overflows happen far away from where the data is read in, after some
scaling, rotating and other tweaking, so I don't see a general route to
limiting them at input.