Re: [patch tree-ssa-forwprop]: Improve binary and/or/xor folding

[Richard Guenther <richard.guenther@gmail.com>]

On Mon, Jun 27, 2011 at 12:58 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
> 2011/6/27 Richard Guenther <richard.guenther@gmail.com>:
>> On Wed, Jun 22, 2011 at 3:09 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>> Hello,
>>>
>>> This patch improves via type-sinking folding of binary and, or, and
>>> xor operations.
>>> First we do sinking also for compatible types with same precision, as
>>> those don't need to be preserved for these operations.
>>> Additional try to fold patterns (TYPE) X bin-op (Y CMP Z) and (TYPE) X
>>> bin-op !Y, if type of X is
>>> compatible to Y.
>>>
>>> ChangeLog gcc
>>>
>>> 2011-06-22  Kai Tietz  <ktietz@redhat.com>
>>>
>>>        * tree-ssa-forwprop.c (simplify_bitwise_binary):
>>>        Improve binary folding regarding casts.
>>>
>>>
>>> ChangeLog gcc/testsuite
>>>
>>> 2011-06-22  Kai Tietz  <ktietz@redhat.com>
>>>
>>>        * gcc.dg/binop-notand1a.c: New test.
>>>        * gcc.dg/binop-notand2a.c: New test.
>>>        * gcc.dg/binop-notand3a.c: New test.
>>>        * gcc.dg/binop-notand4a.c: New test.
>>>        * gcc.dg/binop-notand5a.c: New test.
>>>        * gcc.dg/binop-notand6a.c: New test.
>>>
>>> Bootstrapped and regression tested for all standard languages, Ada,
>>> and Obj-C++. Ok for apply?
>>
>> The first hunk is ok, the 2nd not - please don't use fold here.  Also
>> your comment says what it tries to match, but not what it tries
>> to produce.  So - what is the transformation you are trying to do?
>> The code is also two duplicates of exactly the same stuff.
>>
>> Btw, I see TRUTH_NOT_EXPR is still around, why's that so?
>>
>> Thanks,
>> Richard.
>
> Ok, I will sent first hunk as separate patch.  The second hunk shall
> try to do simple simple folding like X & !X -> 0 (which is handled by
> fold-const, too). As special case we have here also (type) X & !X,
> and for case X & (type) !X. Later case can happen as soon as we
> preserve casts from boolean-type.
> I was thinking about implementing here the optimizations for all
> binary and/or/xor the foldings to constant directly in
> forward-propagate. This might be the better choice. Should I put this
> into a separate function in forward-propagation, or should I put this
> folding function into a different file?

The function is fine I think, but if you want X & !X -> 0 and similar
patterns then I don't see why you need to hand things to fold at all.
Just pattern-match the cases you are interested in.  Eventually
add a helper function that can pattern-match !*X like

tree
match_unop_chain (enum tree_code code, tree name, tree stop_at
                             int *times)
{
  *times = 0;
  while (TREE_CODE (name) == SSA_NAME)
  {
  gimple def_stmt = SSA_NAME_DEF_STMT (name);
  if (gimple_assign_rhs_code (def_stmt) != code)
    break;
  ++*times;
  name = gimple_assign_rhs1 (def_stmt);
  if (name == stop_at)
   break;
  }
  return name;
}

and use that, checking for even/odd *times.  The above assumes
that code cancels itself out, like ~ or ! or -.  Untested of course.

Richard.


>
> Regards,
> Kai
>