[tree-optimization] : Avoid !=/== 0/1 comparisons for boolean-typed argument

2011/8/2 Richard Guenther <richard.guenther@gmail.com>:
> On Tue, Aug 2, 2011 at 3:14 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>> 2011/8/2 Richard Guenther <richard.guenther@gmail.com>:
>>> On Tue, Aug 2, 2011 at 12:17 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>>> Hello,
>>>>
>>>> this patch removes in forward-propagation useless comparisons X != 0
>>>> and X != ~0 for boolean-typed X.  For one-bit precision typed X we
>>>> simplifiy X == 0 (and X != ~0) to ~X, and for X != 0 (and X == ~0) to
>>>> X.
>>>> For none one-bit precisione typed X, we simplify here X == 0 -> X ^ 1,
>>>> and for X != 0 -> X.  We can do this as even for Ada - which has only
>>>> boolean-type with none-one-bit precision - the truth-value is one.
>>>
>>> This isn't a simplification but a canonicalization and thus should be
>>> done by fold_stmt instead (we are not propagating anything after all).
>>> In fact, fold_stmt should do parts of this already by means of its
>>> canonicalizations via fold.
>>
>> Well, it simplifies and canonicalizes.  But to put this into
>> gimple-fold looks better.
>>
>>>> Additionally this patch changes for function
>>>> forward_propagate_comparison the meaning of true-result.  As this
>>>> result wasn't used and it is benefitial to use this propagation also
>>>
>>> which is a bug - for a true return value we need to set cfg_changed to true.
>>
>> I addressed this in my updated patch (see below)
>>
>>>> in second loop in function ssa_forward_propagate_and_combine, it
>>>> returns true iff statement was altered.  Additionally this function
>>>> handles now the boolean-typed simplifications.
>>>
>>> why call it twice?  How should that be "beneficial"?  I think that
>>> forward_propagate_into_comparison should instead fold the changed
>>> statement.
>>
>> Well, due missing fold_stmt call, there were still none-converted
>> comparisons. I've added here the call to fold_stmt_inplace, and it
>> solved the issue.
>>
>>>> For the hunk in gimple.c for function canonicalize_cond_expr_cond:
>>>> This change seems to show no real effect, but IMHO it makes sense to
>>>> add here the check for cast from boolean-type to be consitant.
>>>
>>> Probably yes.
>>>
>>> Thanks,
>>> Richard.
>>
>>
>> 2011-08-02  Kai Tietz  <ktietz@redhat.com>
>>
>>       * gimple.c (canonicalize_cond_expr_cond): Handle cast from boolean-type.
>>       (ssa_forward_propagate_and_combine): Interprete result of
>>       forward_propagate_comparison.
>>       * gcc/gimple-fold.c (fold_gimple_assign): Add canonicalization for
>>       boolean-typed operands for comparisons.
>>
>> 2011-08-02  Kai Tietz  <ktietz@redhat.com>
>>
>>        * gcc.dg/tree-ssa/forwprop-15.c: New testcase.
>>
>> Regression tested and bootstrapped for all languages (including Ada
>> and Obj-C++).  Ok for apply?
>
> Comments below
>
>> Regards,
>> Kai
>>
>> Index: gcc/gcc/testsuite/gcc.dg/tree-ssa/forwprop-15.c
>> ===================================================================
>> --- /dev/null
>> +++ gcc/gcc/testsuite/gcc.dg/tree-ssa/forwprop-15.c
>> @@ -0,0 +1,14 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2 -fdump-tree-forwprop1" }  */
>> +
>> +_Bool
>> +foo (_Bool a, _Bool b, _Bool c
>> +{
>> +  _Bool r1 = a == 0 & b != 0;
>> +  _Bool r2 = b != 0 & c == 0;
>> +  return (r1 == 0 & r2 == 0);
>> +}
>> +
>> +/* { dg-final { scan-tree-dump-times " == " 0 "forwprop1" } } */
>> +/* { dg-final { scan-tree-dump-times " != " 0 "forwprop1" } } */
>> +/* { dg-final { cleanup-tree-dump "forwprop1" } } */
>> Index: gcc/gcc/gimple-fold.c
>> ===================================================================
>> --- gcc.orig/gcc/gimple-fold.c
>> +++ gcc/gcc/gimple-fold.c
>> @@ -814,6 +814,34 @@ fold_gimple_assign (gimple_stmt_iterator
>>                                             gimple_assign_rhs1 (stmt),
>>                                             gimple_assign_rhs2 (stmt));
>>        }
>> +      else if (gimple_assign_rhs_code (stmt) == EQ_EXPR
>> +               || gimple_assign_rhs_code (stmt) == NE_EXPR)
>> +        {
>> +         tree op1 = gimple_assign_rhs1 (stmt);
>> +         tree op2 = gimple_assign_rhs2 (stmt);
>> +         tree type = TREE_TYPE (op1);
>> +         if (useless_type_conversion_p (TREE_TYPE (gimple_assign_lhs (stmt)),
>> +                                        type)
>> +             && TREE_CODE (op2) == INTEGER_CST)
>
> first check op2, it's cheaper.  put the lhs into a local var to avoid the
> excessive long line.

Ok

> And add a comment what you check here - cost me some 2nd thoguht.
> Like
>
>  /* Check whether the comparison operands are of the same boolean
>     type as the result type is.  */
>
>> +           {
>> +             gimple s;
>> +             bool inverted = (gimple_assign_rhs_code (stmt) == EQ_EXPR);
>> +             if (!integer_zerop (op2))
>> +               inverted = !inverted;
>
> For non-1-precision bools I believe you can have non-1 and non-0 op2.
> So you better explicitly check.  The code also isn't too easy to follow,
> just enumerating the four cases wouldn't cause too much bloat, no?

Well, in my tests I haven't saw this for Ada.  Anyway I added explicit
checks for one and zero, so range of constant is defined.
To make code here a bit more easier to readable, I renamed local var
and added some comment here.

>> +             if (inverted == false)
>> +               result = op1;
>> +             else if (TREE_CODE (op1) == SSA_NAME
>> +                      && (s = SSA_NAME_DEF_STMT (op1)) != NULL
>> +                      && is_gimple_assign (s)
>> +                      && gimple_assign_rhs_code (s) == BIT_NOT_EXPR)
>
> You shouldn't do stmt lookups here.

Hmm, well, I removed it.  This folding is done in forwprop IIRC.

>> +               result = gimple_assign_rhs1 (s);
>> +            else
>> +               result = build1_loc (gimple_location (stmt), BIT_NOT_EXPR, type, op1);
>
> What about the non-1-precision booleans?  You need to generate
> op1 ^ 1 here instead, no?

Yes, missed that (and had here a bug).  Corrected in patch.

>> +
>> +           }
>> +
>> +       }
>>
>>       if (!result)
>>         result = fold_binary_loc (loc, subcode,
>> Index: gcc/gcc/tree-ssa-forwprop.c
>> ===================================================================
>> --- gcc.orig/gcc/tree-ssa-forwprop.c
>> +++ gcc/gcc/tree-ssa-forwprop.c
>> @@ -469,6 +469,9 @@ forward_propagate_into_comparison (gimpl
>>     {
>>       gimple_assign_set_rhs_from_tree (gsi, tmp);
>>       update_stmt (stmt);
>> +      if (fold_stmt_inplace (stmt))
>> +        update_stmt (stmt);
>> +
>
> No, we need to update_stmt always as we already did change the stmt.

Why,  if fold_stmt_inplace returns false, comment says nothing was
changed.  So why should be a second update_stmt be necessary?  I
modified that in new patch, but this looks more costy.

>>       if (TREE_CODE (rhs1) == SSA_NAME)
>>        cfg_changed |= remove_prop_source_from_use (rhs1);
>>       if (TREE_CODE (rhs2) == SSA_NAME)
>> @@ -2407,7 +2410,8 @@ ssa_forward_propagate_and_combine (void)
>>            }
>>          else if (TREE_CODE_CLASS (code) == tcc_comparison)
>>            {
>> -             forward_propagate_comparison (stmt);
>> +             if (forward_propagate_comparison (stmt))
>> +               cfg_changed = true;
>>              gsi_next (&gsi);
>
> This hunk is ok.
>
> Thanks,
> Richard.
>
>>            }
>>          else

So revised patch with same changelog as before ...

Regards,
Kai

[tree-optimization] : Avoid !=/== 0/1 comparisons for boolean-typed argument

Commit Message

Comments

Patch