On Wed, May 29, 2019 at 3:21 PM Jan Hubicka <hubi...@ucw.cz> wrote:
>
> >
> > Please also see if there are testcases that do anything meaningful
> > and FAIL after instead of
> >
> > /* Do access-path based disambiguation. */
> > if (ref1 && ref2
> > && (handled_component_p (ref1) || handled_component_p (ref2)))
> >
> > doing
> >
> > /* Do access-path based disambiguation. */
> > if (ref1 && ref2
> > && (handled_component_p (ref1) && handled_component_p (ref2)))
> >
> On tramp3d we get quite few matches which are attached. If ref1 is
> MEM_REF and ref2 has non-trivial access path then it seems we need:
> 1) ref1 and ref2 to conflict (ref1 is a record or alias set 0)
> 2) basetype2 to contain ref1 (so it conflicts too)
> 3) if ref1 is a record than the access path may go into a type
> contained as field of ref1 but via path not containing ref1 itself.
>
> I tried to construct testcase:
>
> truct foo {int val;} *fooptr;
> struct bar {struct foo foo; int val2;} *barptr;
> int test()
> {
> struct foo foo={0};
> barptr->val2 = 1;
> *fooptr=foo;
> return barptr->val2;
> }
>
> but we do not optimize it. I.e. optimized dump has:
>
> test ()
> {
> struct bar * barptr.0_1;
> struct foo * fooptr.1_2;
> int _6;
>
> <bb 2> [local count: 1073741824]:
> barptr.0_1 = barptr;
> barptr.0_1->val2 = 1;
> fooptr.1_2 = fooptr;
> MEM[(struct foo *)fooptr.1_2] = 0;
> _6 = barptr.0_1->val2;
> return _6;
> }
>
> I see no reason why we should not constant propagate the return value.
Indeed a good example. Make it work and add it to the testsuite ;)
I would have said get_alias_set () on the ref type should already have
disambiguated 'int' (barptr->val2) from *fooptr (struct foo) but of course
they conflict because foo contains 'int'.
I guess it doesn't work because 'struct foo' isn't part of the other
path. Here nonoverlapping_component_refs_of_decl_p would be
the vehicle to use (but IIRC that would also require a common
type in one of both paths).
Richard.
>
> Honza
