On Wed, 26 Jul 2017, Jakub Jelinek wrote:
> On Wed, Jul 26, 2017 at 04:13:30PM +0200, Richard Biener wrote:
> > > > You don't seem to use 'size' anywhere.
> > >
> > > size I thought about but then decided not to do anything with it.
> > > There are two cases, one is where there is no ADDR_EXPR and it actually
> > > a memory reference.
> > > In that case in theory the size could be used, but it would need
> > > to be used only for the positive offsets, so like:
> > > if (off > 0) {
> > > if (ptr + off + size < ptr)
> > > runtime_fail;
> > > } else if (ptr + off > ptr)
> > > runtime_fail;
> > > but when it is actually a memory reference, I suppose it will fail
> > > at runtime anyway when performing such an access, so I think it is
> > > unnecessary. And for the ADDR_EXPR case, the size is irrelevant, we
> > > are just taking address of the start of the object.
> > >
> > > > You fail to allow other handled components -- for no good reason?
> > >
> > > I was trying to have a quick bail out. What other handled components
> > > might
> > > be relevant? I guess IMAGPART_EXPR. For say BIT_FIELD_REF I don't think
> > > I can
> > > tree ptr = build1 (ADDR_EXPR, build_pointer_type (TREE_TYPE (t)), t);
> >
> > REALPART/IMAGPART_EXPR, yes. You can't address BIT_FIELD_REF
> > apart those on byte boundary (&vector[4] is eventually folded to
> > a BIT_FIELD_REF). Similar for VIEW_CONVERT_EXPR, but you are
> > only building the address on the base?
> >
> > > > You fail to handle
> > > > &MEM[ptr + CST] a canonical gimple invariant way of ptr +p CST,
> > > > the early out bitpos == 0 will cause non-instrumentation here.
> > >
> > > Guess I could use:
> > > if ((offset == NULL_TREE
> > > && bitpos == 0
> > > && (TREE_CODE (inner) != MEM_REF
> > > || integer_zerop (TREE_OPERAND (inner, 1))))
> > > The rest of the code will handle it.
> >
> > Yeah.
> >
> > >
> > > > (I'd just round down in the case of bitpos % BITS_PER_UNIT != 0)
> > >
> > > But then the
> > > tree ptr = build1 (ADDR_EXPR, build_pointer_type (TREE_TYPE (t)), t);
> > > won't work again.
> >
> > Hmm. So instead of building the address on the original tree you
> > could build the difference based on what get_inner_reference returns
> > in bitpos/offset?
>
> I'm building both addresses and subtracting them to get the offset.
> I guess the other option is to compute just the address of the base
> (i.e. base_addr), and add offset (if non-NULL) plus bitpos / BITS_PER_UNIT
> plus offset from the MEM_REF (if any). In that case it would probably
> handle any handled_component_p and bitfields too.
Yes. Can you try sth along this route? Should be a matter of
adding offset and bitpos / BITS_PER_UNIT (thus rounded down) plus
any MEM_REF offset on the base.
Thanks,
Richard.
