Re: [PATCH] middle-end/113396 - int128 array index and value-ranges

2024-03-20 Thread Richard Biener
On Tue, 19 Mar 2024, Jakub Jelinek wrote:

> On Tue, Mar 19, 2024 at 03:47:37PM +0100, Richard Biener wrote:
> > The following fixes bogus truncation of a value-range for an int128
> > array index when computing the maximum extent for a variable array
> > reference.  Instead of possibly slowing things down by using
> > widest_int the following makes sure the range bounds fit within
> > the constraints offset_int were designed for.
> 
> Perhaps you could use wide_int/poly_wide_int with precision
> of offset_int if it is at most 64-bit precision and twice that precision
> otherwise.
> I think large BITINT_TYPEs shouldn't be a problem since r14-7200,
> so another fix might to truncate at gimplification time
> ARRAY_REF indexes wider than sizetype to sizetype.  Maybe GCC 15-ish
> material though.

Yeah, I was thinking of instead comparing to the effective biggest
index supporting the (half) address-space limit but I guess that's what
the get_precision check effectively does as well.

But yeah, I guess truncating ARRAY_REF indices like we truncate
shift amounts is the way to go in the end.  For GCC 15 indeed.

I've installed the patch as-is.

Thanks,
Richard.

> Anyway, guess your patch is ok as is too.
> 
> > PR middle-end/113396
> > * tree-dfa.cc (get_ref_base_and_extent): Use index range
> > bounds only if they fit within the address-range constraints
> > of offset_int.
> > 
> > * gcc.dg/torture/pr113396.c: New testcase.
> > ---
> >  gcc/testsuite/gcc.dg/torture/pr113396.c | 19 +++
> >  gcc/tree-dfa.cc |  6 --
> >  2 files changed, 23 insertions(+), 2 deletions(-)
> >  create mode 100644 gcc/testsuite/gcc.dg/torture/pr113396.c
> > 
> > diff --git a/gcc/testsuite/gcc.dg/torture/pr113396.c 
> > b/gcc/testsuite/gcc.dg/torture/pr113396.c
> > new file mode 100644
> > index 000..585f717bdda
> > --- /dev/null
> > +++ b/gcc/testsuite/gcc.dg/torture/pr113396.c
> > @@ -0,0 +1,19 @@
> > +/* { dg-do run } */
> > +/* { dg-require-effective-target int128 } */
> > +
> > +unsigned char m[] = {5, 79, 79, 79, 79};
> > +__int128 p;
> > +int main()
> > +{
> > +  int g1 = 0;
> > +  p = 0;
> > +  for (int aj = 0; aj < 256; aj++)
> > +   {
> > +  m[0] = -4;
> > +  for (; p >= 0; p -= 1) {
> > +g1 = m[p];
> > +  }
> > +  }
> > +  if (g1 != 0xfc)
> > +__builtin_abort();
> > +}
> > diff --git a/gcc/tree-dfa.cc b/gcc/tree-dfa.cc
> > index cbd3774b21f..93e53b29a6d 100644
> > --- a/gcc/tree-dfa.cc
> > +++ b/gcc/tree-dfa.cc
> > @@ -549,7 +549,8 @@ get_ref_base_and_extent (tree exp, poly_int64 *poffset,
> > /* Try to constrain maxsize with range information.  */
> > offset_int omax
> >   = offset_int::from (max, TYPE_SIGN (TREE_TYPE (index)));
> > -   if (known_lt (lbound, omax))
> > +   if (wi::get_precision (max) <= ADDR_MAX_BITSIZE
> > +   && known_lt (lbound, omax))
> >   {
> > poly_offset_int rmaxsize;
> > rmaxsize = (omax - lbound + 1)
> > @@ -567,7 +568,8 @@ get_ref_base_and_extent (tree exp, poly_int64 *poffset,
> > /* Try to adjust bit_offset with range information.  */
> > offset_int omin
> >   = offset_int::from (min, TYPE_SIGN (TREE_TYPE (index)));
> > -   if (known_le (lbound, omin))
> > +   if (wi::get_precision (min) <= ADDR_MAX_BITSIZE
> > +   && known_le (lbound, omin))
> >   {
> > poly_offset_int woffset
> >   = wi::sext (omin - lbound,
> > -- 
> > 2.35.3
> 
>   Jakub
> 
> 

-- 
Richard Biener 
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)


Re: [PATCH] middle-end/113396 - int128 array index and value-ranges

2024-03-19 Thread Jakub Jelinek
On Tue, Mar 19, 2024 at 03:47:37PM +0100, Richard Biener wrote:
> The following fixes bogus truncation of a value-range for an int128
> array index when computing the maximum extent for a variable array
> reference.  Instead of possibly slowing things down by using
> widest_int the following makes sure the range bounds fit within
> the constraints offset_int were designed for.

Perhaps you could use wide_int/poly_wide_int with precision
of offset_int if it is at most 64-bit precision and twice that precision
otherwise.
I think large BITINT_TYPEs shouldn't be a problem since r14-7200,
so another fix might to truncate at gimplification time
ARRAY_REF indexes wider than sizetype to sizetype.  Maybe GCC 15-ish
material though.

Anyway, guess your patch is ok as is too.

>   PR middle-end/113396
>   * tree-dfa.cc (get_ref_base_and_extent): Use index range
>   bounds only if they fit within the address-range constraints
>   of offset_int.
> 
>   * gcc.dg/torture/pr113396.c: New testcase.
> ---
>  gcc/testsuite/gcc.dg/torture/pr113396.c | 19 +++
>  gcc/tree-dfa.cc |  6 --
>  2 files changed, 23 insertions(+), 2 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.dg/torture/pr113396.c
> 
> diff --git a/gcc/testsuite/gcc.dg/torture/pr113396.c 
> b/gcc/testsuite/gcc.dg/torture/pr113396.c
> new file mode 100644
> index 000..585f717bdda
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/torture/pr113396.c
> @@ -0,0 +1,19 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target int128 } */
> +
> +unsigned char m[] = {5, 79, 79, 79, 79};
> +__int128 p;
> +int main()
> +{
> +  int g1 = 0;
> +  p = 0;
> +  for (int aj = 0; aj < 256; aj++)
> +   {
> +  m[0] = -4;
> +  for (; p >= 0; p -= 1) {
> +g1 = m[p];
> +  }
> +  }
> +  if (g1 != 0xfc)
> +__builtin_abort();
> +}
> diff --git a/gcc/tree-dfa.cc b/gcc/tree-dfa.cc
> index cbd3774b21f..93e53b29a6d 100644
> --- a/gcc/tree-dfa.cc
> +++ b/gcc/tree-dfa.cc
> @@ -549,7 +549,8 @@ get_ref_base_and_extent (tree exp, poly_int64 *poffset,
>   /* Try to constrain maxsize with range information.  */
>   offset_int omax
> = offset_int::from (max, TYPE_SIGN (TREE_TYPE (index)));
> - if (known_lt (lbound, omax))
> + if (wi::get_precision (max) <= ADDR_MAX_BITSIZE
> + && known_lt (lbound, omax))
> {
>   poly_offset_int rmaxsize;
>   rmaxsize = (omax - lbound + 1)
> @@ -567,7 +568,8 @@ get_ref_base_and_extent (tree exp, poly_int64 *poffset,
>   /* Try to adjust bit_offset with range information.  */
>   offset_int omin
> = offset_int::from (min, TYPE_SIGN (TREE_TYPE (index)));
> - if (known_le (lbound, omin))
> + if (wi::get_precision (min) <= ADDR_MAX_BITSIZE
> + && known_le (lbound, omin))
> {
>   poly_offset_int woffset
> = wi::sext (omin - lbound,
> -- 
> 2.35.3

Jakub