On 15.1.2016 10.08, Eli Zaretskii wrote:
> Right, this can explain everything.  It probably means that a build
> configured --with-wide-int should also be tested there, as it has
> different alignment needs.

I'll test that as well, then.

>> How should we fix this?  I can write and test the fix, but I'd like to
>> hear your opinion on the preferred mechanism.
>
> What fix did you have in mind?  Over-allocating and recording the
> offset in the initial part of the block that we don't pass to the
> application?

That was my first idea, yes.  Of course, this introduces some memory 
overhead, especially if there are many small allocations.

A proposal patch is attached.  Reallocs needed some additional care, as 
there are two special cases to handle (source being in dumped memory, 
and source and destination alignments mismatching).

-- 
Jussi Lahdenniemi