On Mon, 6 May 2024, Martin Uecker wrote:

> Am Montag, dem 06.05.2024 um 11:07 +0200 schrieb Richard Biener:
> > On Mon, 6 May 2024, Martin Uecker wrote:
> > 
> > > Am Montag, dem 06.05.2024 um 09:00 +0200 schrieb Richard Biener:
> > > > On Sat, 4 May 2024, Martin Uecker wrote:
> > > > 
> > > > > Am Freitag, dem 03.05.2024 um 21:16 +0200 schrieb Jakub Jelinek:
> > > > > > > On Fri, May 03, 2024 at 09:11:20PM +0200, Martin Uecker wrote:
> > > > > > > > > > > TYPE_CANONICAL as used by the middle-end cannot express this but
> > > > > > > > > 
> > > > > > > > > Hm. so how does it work now for arrays?
> > > > > > > 
> > > > > > > Do you have a testcase which doesn't work correctly with the arrays?
> > > > > 
> > > > > I am mostly trying to understand better how this works. But
> > > > > if I am not mistaken, the following example would indeed
> > > > > indicate that we do incorrect aliasing decisions for types
> > > > > derived from arrays:
> > > > > 
> > > > > https://godbolt.org/z/rTsE3PhKc
> > > > 
> > > > This example is about pointer-to-array types, int (*)[2] and
> > > > int (*)[1] are supposed to be compatible as in receive the same alias
> > > > set. 
> > > 
> > > In C, char (*)[2] and char (*)[1] are not compatible. But with
> > > COMPAT set, the example operates^1 with char (*)[] and char (*)[1]
> > > which are compatible.  If we form equivalence classes, then
> > > all three types would need to be treated as equivalent. 
> > > 
> > > ^1 Actually, pointer to functions returning pointers
> > > to arrays. Probably this example can still be simplified...
> > > 
> > > >  This is ensured by get_alias_set POINTER_TYPE_P handling,
> > > > the alias set is supposed to be the same as that of int *.  It seems
> > > > we do restrict the handling a bit, the code does
> > > > 
> > > >       /* Unnest all pointers and references.
> > > >          We also want to make pointer to array/vector equivalent to 
> > > > pointer to
> > > >          its element (see the reasoning above). Skip all those types, too.  
> > > > */
> > > >       for (p = t; POINTER_TYPE_P (p)
> > > >            || (TREE_CODE (p) == ARRAY_TYPE
> > > >                && (!TYPE_NONALIASED_COMPONENT (p)
> > > >                    || !COMPLETE_TYPE_P (p)
> > > >                    || TYPE_STRUCTURAL_EQUALITY_P (p)))
> > > >            || TREE_CODE (p) == VECTOR_TYPE;
> > > >            p = TREE_TYPE (p))
> > > > 
> > > > where the comment doesn't exactly match the code - but C should
> > > > never have TYPE_NONALIASED_COMPONENT (p).
> > > > 
> > > > But maybe I misread the example or it goes wrong elsewhere.
> > > 
> > > If I am not confusing myself too much, the example shows that
> > > aliasing analysis treats the the types as incompatible in
> > > both cases, because it does not reload *a with -O2. 
> > > 
> > > For char (*)[1] and char (*)[2] this would be correct (but an
> > > implementation exploiting this would need to do structural
> > > comparisons and not equivalence classes) but for 
> > > char (*)[2] and char (*)[] it is not.
> > 
> > Oh, these are function pointers, so it's about the alias set of
> > a pointer to FUNCTION_TYPE.  I don't see any particular code
> > trying to make char[] * (*)() and char[1] *(*)() inter-operate
> > for TBAA iff the FUNCTION_TYPEs themselves are not having the
> > same TYPE_CANONICAL.
> > 
> > Can you open a bugreport and please point to the relevant parts
> > of the C standard that tells how pointer-to FUNCTION_TYPE TBAA
> > is supposed to work?
> 
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114959

I have now pushed the patch after consulting with Jakub.

Richard.

> Martin
> > 
> 
> > Thanks,
> > Richard.
> > 
> > > Martin
> > > 
> > > 
> > > > 
> > > > Richard.
> > > > 
> > > > > Martin
> > > > > 
> > > > > > > 
> > > > > > > E.g. same_type_for_tbaa has
> > > > > > >   type1 = TYPE_MAIN_VARIANT (type1);
> > > > > > >   type2 = TYPE_MAIN_VARIANT (type2);
> > > > > > > 
> > > > > > >   /* Handle the most common case first.  */
> > > > > > >   if (type1 == type2)
> > > > > > >     return 1;
> > > > > > > 
> > > > > > >   /* If we would have to do structural comparison bail out.  */
> > > > > > >   if (TYPE_STRUCTURAL_EQUALITY_P (type1)
> > > > > > >       || TYPE_STRUCTURAL_EQUALITY_P (type2))
> > > > > > >     return -1;
> > > > > > > 
> > > > > > >   /* Compare the canonical types.  */
> > > > > > >   if (TYPE_CANONICAL (type1) == TYPE_CANONICAL (type2))
> > > > > > >     return 1;
> > > > > > > 
> > > > > > >   /* ??? Array types are not properly unified in all cases as we have
> > > > > > >      spurious changes in the index types for example.  Removing this
> > > > > > >      causes all sorts of problems with the Fortran frontend.  */
> > > > > > >   if (TREE_CODE (type1) == ARRAY_TYPE
> > > > > > >       && TREE_CODE (type2) == ARRAY_TYPE)
> > > > > > >     return -1;
> > > > > > > ...
> > > > > > > and later compares alias sets and the like.
> > > > > > > So, even if int[] and int[0] have different TYPE_CANONICAL, they
> > > > > > > will be considered maybe the same.  Also, guess get_alias_set
> > > > > > > has some ARRAY_TYPE handling...
> > > > > > > 
> > > > > > > Anyway, I think we should just go with Richi's patch.
> > > > > > > 
> > > > > > > 	Jakub
> > > > > > > 
> > > > > 
> > > > > 
> > > > > 
> > > > 
> > > 
> > > 
> > 
> 
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)