On Fri, Nov 19, 2021 at 8:17 PM Jakub Jelinek via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> On Sat, Nov 20, 2021 at 12:31:19AM +0530, Siddhesh Poyarekar wrote:
> > > Neither of these are equivalent to what it used to do before.
> > > If some target has e.g. pointers wider than size_t, then previously we
> > > could
> > > compute bytes that doesn't fit into size_t and would return NULL which
> > > eventually would result in object_sizes_execute or
> > > expand_builtin_object_size
> > > resolving it to the unknown value. But fold_convert will perform modulo
> > > arithmetics on it, so say if size_t is 24-bit and bytes is 0x1000001, it
> > > will fold to (size_t) 1. I think we should still punt if it doesn't fit.
> > > Or do we ensure that what it returns always has sizetype type?
> >
> > compute_builtin_object_size should always return sizetype. I probably
> > haven't strictly ensured that but I could do that. Would it be necessary to
> > check for fit in sizetype -> size_type_node conversion too? I've been
> > assuming that they'd have the same precision.
>
> sizetype and size_type_node should be indeed always the same precision.
I don't think so. vms.h has
/* Always a 32 bit type. */
#undef SIZE_TYPE
#define SIZE_TYPE "unsigned int"
...
#define SIZETYPE (flag_vms_pointer_size == VMS_POINTER_SIZE_NONE ? \
"unsigned int" : "long long unsigned int")
interestingly that's the _only_ SIZETYPE override we have, so it does look like
we might want to try removing the 'SIZETYPE' target macro. I'm not sure the
above does anything good - it looks more like a ptr_mode vs Pmode thing.
Richard.
> Jakub
>
