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. 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 > > > > > > > > > > > -- Univ.-Prof. Dr. rer. nat. Martin Uecker Graz University of Technology Institute of Biomedical Imaging