Re: [Patch, omp] Patch to omp-low.c to fix failures on IA64 HP-UX

["Richard Guenther" <richard dot guenther at gmail dot com>]

On Thu, Oct 30, 2008 at 11:02 PM, Steve Ellcey <sje@cup.hp.com> wrote:
> I have two libgomp tests, libgomp.c/loop-5.c and libgomp.c++/loop-8.C
> that fail on IA64 HP-UX in 32 bit mode.  In this mode pointers are 32
> bits but need to be extended to 64 bits before being dereferenced.  IA64
> has a  special instruction, addp4 to do that.  In omp-low.c
> (expand_omp_for_generic) we have these two instructions:
>
>        t1 = fold_convert (fd->iter_type, fd->loop.n2);
>        t0 = fold_convert (fd->iter_type, fd->loop.n1);
>
> Where fd->loop.n1 (and n2) are 32 bit pointer types and fd->iter_type is
> a 64 bit 'unsigned long long' type.  There are two ways of getting from
> a 32 bit pointer to a 64 bit unsigned long, one could use addp4 or one
> could convert the pointer to a 32 bit unsigned long and then zero extend
> that.

It turns out that the middle-end type checker would error on a conversion
from a pointer to an integer type of different precision.  Just it is never
executed after gimplification ... (and we specifically allow conversion
to/from sizetype to make ports with mismatching pointer/sizetype precision
happy).

For 4.5 I plan to enable this checking code after all passes which will
catch this errorneous situation.

> It turns out that using either method works as long as you use the same
> method for both of these statements but in the two failing tests one
> convert gets done one way and the other the other way and then the test
> fails.  They get done differently because one of the expressions being
> converted is a simple address and the other is an expression and this
> results in different choices in how to extend the expression.
>
> This patch fixes the problem by explicitly converting the pointer to
> 'unsigned long' first and then extending that to type 'unsigned long
> long'.  This forces a consistent conversion path for both expressions
> and fixes the bug.
>
> Tested on IA64 HP-UX and Linux with no regressions.
>
> OK for checkin?

I'll leave that to OMP maintainers.  It at least looks ugly ;)

Richard.

>
> 2008-10-30  Steve Ellcey  <sje@cup.hp.com>
>
>        * omp-low.c (expand_omp_for_generic): Change conversion code.
>
>
> Index: omp-low.c
> ===================================================================
> --- omp-low.c   (revision 141457)
> +++ omp-low.c   (working copy)
> @@ -3681,8 +3681,20 @@ expand_omp_for_generic (struct omp_regio
>       t4 = build_fold_addr_expr (iend0);
>       t3 = build_fold_addr_expr (istart0);
>       t2 = fold_convert (fd->iter_type, fd->loop.step);
> -      t1 = fold_convert (fd->iter_type, fd->loop.n2);
> -      t0 = fold_convert (fd->iter_type, fd->loop.n1);
> +      if (POINTER_TYPE_P (TREE_TYPE (fd->loop.n2))
> +         && (TYPE_PRECISION (fd->iter_type) >
> +             TYPE_PRECISION (long_unsigned_type_node)))
> +       t1 = fold_convert (fd->iter_type,
> +                          fold_convert (long_unsigned_type_node, fd->loop.n2));
> +      else
> +       t1 = fold_convert (fd->iter_type, fd->loop.n2);
> +      if (POINTER_TYPE_P (TREE_TYPE (fd->loop.n1))
> +         && (TYPE_PRECISION (fd->iter_type) >
> +             TYPE_PRECISION (long_unsigned_type_node)))
> +       t0 = fold_convert (fd->iter_type,
> +                          fold_convert (long_unsigned_type_node, fd->loop.n1));
> +      else
> +        t0 = fold_convert (fd->iter_type, fd->loop.n1);
>       if (bias)
>        {
>          t1 = fold_build2 (PLUS_EXPR, fd->iter_type, t1, bias);
>