This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH] Fix a fallout of PR14179 fix (3.3/3.4/4.0 regression)

On Wed, 29 Dec 2004, Jakub Jelinek wrote:
> I have tried two different patches.
> One simply kills all special handling of sizetype in force_fit_type
> (thus effectively undoing part of the
> change).  There were no testcases for that though, so I can't guess
> what was the reason for the patch.
> With this patch I have successfully bootstrapped on 7 arches,
> and there were no regressions in the CVS testcases.
> But there are regressions with my object size checking patch,
> particularly builtin-object-size-{1,3}.c tests.
> The interesting part there is:
>   struct T { char buf[10]; char buf2[10]; } t;
>   char *p = &t.buf2[-4];
> With CVS GCC, that -4 doesn't have overflow flag set, while
> with the first patch applied it suddenly has the overflow
> flag set and this later on leads in the object size checking
> code to unknow object size rather than the expected 14.

Hi Jakub,

I have a pair of patches that might cure this problem.  We're currently
moving away from the representation of "overflow" on trees, leaving it
to the front-ends to issue their required diagnostics, whilst leaving
the tree-ssa optimizers to do their thing without concern for overflow.

Could you check whether your builtin_object_size failures are fixed by
the patch in

Ignore the additional failure of gcc.dg/bitfld-2.c;  I'm currently
addressing that, but it's fix is in turn waiting upon review of the
patch in

Let me know if this works for you.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]