[Bug tree-optimization/32461] [4.3 Regression] Segmentation fault in build_classic_dist_vector_1() at tree-data-ref.c:2700

[Richard Guenther]

On 6/23/07, Sebastian Pop <spop@gcc.gnu.org> wrote:
> Hi,
>
> In data dependence analysis, when the dependence distance is
> lexicographically negative, we recompute the dependence by inverting
> the source and the sink memory accesses, and thus we compute the
> dependence again but in the opposite direction.  In one way we
> compute:
>
> MINUS_EXPR
> (unsigned int) i_434 + 1
> (unsigned int) i_434
> =
> 1
>
> but when going the other way, backwards, we have to compute:
>
> MINUS_EXPR
> (unsigned int) i_434
> (unsigned int) i_434 + 1
> =
> ~(unsigned int) i_434 + (unsigned int) i_434
>
> and instead of getting an integer_cst, we get that ugly expression
> from fold, the dependence test fails, and everything goes wrong from
> that point.
>
> With this patch we get the expected 0xffffffffffffffff for the above
> case.  The patch teaches fold how to compute the minus expression for
> unsigned int and pointer expressions.
>
> Bootstrapped and tested on i686-linux.  Okay for trunk?

First - can you check why the asociate: case in fold_binary does not do
the transformation and maybe fix that instead if possible?

Then there is a ChangeLog missing.  Further...

+         && integer_zerop (fold_build2 (MINUS_EXPR, type, arg0,
+                                        TREE_OPERAND (arg1, 0))))

to avoid creating trees and just throwing them away you should use
operand_equal_p (arg0, TREE_OPERAND (arg1, 0)) instead.  You
similarly may want to check operand 1 for A - (B + A)?

+         && TYPE_UNSIGNED (type)

you should check TYPE_OVERFLOW_WRAPS (type)  instead.

Thanks,
Richard.