Re: PR55124 - tentative patch for ICE in PRE

[Richard Biener <richard dot guenther at gmail dot com>]

On Thu, Nov 29, 2012 at 1:14 PM, Tom de Vries <Tom_deVries@mentor.com> wrote:
> On 29/11/12 11:26, Richard Biener wrote:
>> I'm continuing trying to move value-ids back to PRE land.  Your patch
>> would be nice in a form that verifies the order is indeed topological,
>> maybe you can re-work it in a way that does this in
>> sorted_array_from_bitmap_set in a ENABLE_CHECKING piece?
>
> Richard,
>
> These are my current patches, tested together with tree-ssa.exp.
>
>
> The first patch checks the topological order in sorted_array_from_bitmap_set.
> Testing only this one with tree-ssa.exp gives 400 failures.
>
> Btw, I'm not 100% sure if this patch checks the required order. It's clear what
> topological order means if there is one expression per value. I've ran
> tree-ssa.exp with an assert that the number of expressions and values in the
> bitmap_set is equal in sorted_array_from_bitmap_set, and that passed, so that
> seems to be the case generally, but I don't know if that's by design.
> If there are more expressions with the same value, this patch is a 'weak' check,
> meaning a value is considered available if one expression with that value is
> available. A 'strong' check would consider a value available if all expressions
> with that value are available. I can imagine doing clean on a strongly or weakly
> ordered array could give different results.
>
>
> The second patch calculates value_id during pre. If you're working on the
> value_id part, I'll stop here.

Thanks.

For the moment I am fixing value-id assigns in tree-ssa-sccvn.c but there
seems to be an issue with possibly assigning the same value-ids for
conflicting topology:

  if ()
    {
      a = b + c;
      d = a + 1;
    }
  else
    {
      f = g + h;
      i = f - 1;
    }

if for some reason we value-number 'a' and 'i' the same and 'd' and 'f' the
same then there is no value-id ordering that sorts correctly according
to expression dependencies (some reason being for example
g == b + c, h = 2).  So I'm indeed not sure that simply sorting after
value-id and assigning "proper" value-ids will work at all.

patch that adjusts value-id assignment in SCCVN according to PREs
expectations attached.

Richard.

> Thanks,
> - Tom
>
> 2012-11-29  Tom de Vries  <tom@codesourcery.com>
>
>         * tree-ssa-pre.c (sorted_array_from_bitmap_set): Use
>         EXECUTE_IF_AND_IN_BITMAP instead of EXECUTE_IF_SET_IN_BITMAP.  Check
>         sort result ifdef ENABLE_CHECKING.  Check if the sorted array has the
>         same number of expressions as the bitmap_set.
>
>

Attachment: p2
Description: Binary data