Re: [patch gimplify]: Make sure comparison using boolean-type after gimplification

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

On Thu, May 26, 2011 at 1:28 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
> 2011/5/26 Richard Guenther <richard.guenther@gmail.com>:
>> On Thu, May 26, 2011 at 1:20 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>> 2011/5/26 Richard Guenther <richard.guenther@gmail.com>:
>>>> On Thu, May 26, 2011 at 1:11 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>>>> 2011/5/26 Richard Guenther <richard.guenther@gmail.com>:
>>>>>> On Thu, May 26, 2011 at 12:46 PM, Richard Guenther
>>>>>> <richard.guenther@gmail.com> wrote:
>>>>>>> On Thu, May 26, 2011 at 12:20 PM, Kai Tietz <ktietz@redhat.com> wrote:
>>>>>>>> Hello,
>>>>>>>>
>>>>>>>> this patch ensures that after gimplification also comparison expressions using FE's boolean_type_node.  As we need to deal here with C/C++'s (obj-c/c++ and java's), Ada's, and Fortran's specific boolean types, this patch alters some checks in tree-cfg for Ada's sake, and we need to deal in fold-const about type-conversion of comparisons special.
>>>>>>>> Additionally it takes care that in forwprop pass we don't do type hoising for boolean types.
>>>>>>>>
>>>>>>>> ChangeLog
>>>>>>>>
>>>>>>>> 2011-05-26  Kai Tietz
>>>>>>>>
>>>>>>>>          * gimplify.c (gimple_boolify): Boolify all comparison
>>>>>>>>          expressions.
>>>>>>>>          (gimplify_expr): Use 'useless_type_conversion_p' for comparing
>>>>>>>>          org_type with boolean_type_node for TRUTH-expressions and comparisons.
>>>>>>>>          * fold-const.c (fold_unary_loc): Handle comparison conversions with
>>>>>>>>          boolean-type special.
>>>>>>>>          * tree-cfg.c (verify_gimple_comparison): Adjust check for boolean
>>>>>>>>          or compatible types.
>>>>>>>>          (verify_gimple_assign_unary): Likewise.
>>>>>>>>          * tree-ssa-forwprop.c (forward_propagate_comparison): Handle
>>>>>>>>          boolean case special.
>>>>>>>>
>>>>>>>> Tested on x86_64-pc-linux-gnu (multilib) with regression test for all standard languages (C, C++, Obj-C, Fortran, Java) plus Obj-C++ and Ada. Ok for apply?
>>>>>>>
>>>>>>> It obviously isn't ok to apply before a patch has gone in that will resolve
>>>>>>> all of the FAILs you specify.  Comments on the patch:
>>>>>>>
>>>>>>> @@ -7281,9 +7284,28 @@ gimplify_expr (tree *expr_p, gimple_seq
>>>>>>>                 plain wrong if bitfields are involved.  */
>>>>>>>                {
>>>>>>>                  tree type = TREE_TYPE (TREE_OPERAND (*expr_p, 1));
>>>>>>> +                 tree org_type = TREE_TYPE (*expr_p);
>>>>>>> +
>>>>>>> +                 if (!useless_type_conversion_p (org_type, boolean_type_node))
>>>>>>> +                   {
>>>>>>> +                     TREE_TYPE (*expr_p) = boolean_type_node;
>>>>>>> +                     *expr_p = fold_convert_loc (saved_location, org_type, *expr_p);
>>>>>>> +                     ret = GS_OK;
>>>>>>> +                     goto dont_recalculate;
>>>>>>> +                   }
>>>>>>>
>>>>>>> The above should be only done for !AGGREGATE_TYPE_P.  Probably then
>>>>>>> the strange dont_recalcuate goto can go away as well.
>>>>>>>
>>>>>>>                  if (!AGGREGATE_TYPE_P (type))
>>>>>>> -                   goto expr_2;
>>>>>>> +                   {
>>>>>>> +                     enum gimplify_status r0, r1;
>>>>>>> +
>>>>>>> +                     r0 = gimplify_expr (&TREE_OPERAND (*expr_p, 0), pre_p,
>>>>>>> +                                         post_p, is_gimple_val, fb_rvalue);
>>>>>>> +                     r1 = gimplify_expr (&TREE_OPERAND (*expr_p, 1), pre_p,
>>>>>>> +                                         post_p, is_gimple_val, fb_rvalue);
>>>>>>> +
>>>>>>> +                     ret = MIN (r0, r1);
>>>>>>> +                   }
>>>>>>> +
>>>>>>>
>>>>>>> why change this?
>>>>>>>
>>>>>>> @@ -7641,6 +7641,12 @@ fold_unary_loc (location_t loc, enum tre
>>>>>>>        }
>>>>>>>       else if (COMPARISON_CLASS_P (arg0))
>>>>>>>        {
>>>>>>> +         /* Don't optimize type change, if op0 is of kind boolean_type_node.
>>>>>>> +            Otherwise this will lead to race-condition on gimplification
>>>>>>> +            trying to boolify comparison expression.  */
>>>>>>> +         if (TREE_TYPE (op0) == boolean_type_node)
>>>>>>> +           return NULL_TREE;
>>>>>>> +
>>>>>>>          if (TREE_CODE (type) == BOOLEAN_TYPE)
>>>>>>>            {
>>>>>>>              arg0 = copy_node (arg0);
>>>>>>>
>>>>>>> The code leading here looks quite strange to me ...
>>>>>>>
>>>>>>> tree
>>>>>>> fold_unary_loc (location_t loc, enum tree_code code, tree type, tree op0)
>>>>>>> {
>>>>>>> ...
>>>>>>>  if (TREE_CODE_CLASS (code) == tcc_unary)
>>>>>>>    {
>>>>>>> ...
>>>>>>>      else if (COMPARISON_CLASS_P (arg0))
>>>>>>>        {
>>>>>>>          if (TREE_CODE (type) == BOOLEAN_TYPE)
>>>>>>>            {
>>>>>>>              arg0 = copy_node (arg0);
>>>>>>>              TREE_TYPE (arg0) = type;
>>>>>>>              return arg0;
>>>>>>>            }
>>>>>>>          else if (TREE_CODE (type) != INTEGER_TYPE)
>>>>>>>            return fold_build3_loc (loc, COND_EXPR, type, arg0,
>>>>>>>                                fold_build1_loc (loc, code, type,
>>>>>>>                                             integer_one_node),
>>>>>>>                                fold_build1_loc (loc, code, type,
>>>>>>>                                             integer_zero_node));
>>>>>>>        }
>>>>>>>
>>>>>>> so, for any tcc_unary, like NEGATE_EXPR, with BOOLEAN_TYPE,
>>>>>>> return arg0 ... sure.  Same for the 2nd case.  ~ (a == b) isn't
>>>>>>> the same as a == b ? ~1 : ~0.  I _suppose_ those cases were
>>>>>>> ment for CONVERT_EXPR_CODE_P (code) instead of all of tcc_unary,
>>>>>>> in which case they should be dropped or moved below where we
>>>>>>> handle conversions explicitly.
>>>>>>>
>>>>>>> That said - does anyone remember anything about that above code?
>>>>>>> Trying to do some svn blame history tracking now ...
>>>>>>
>>>>>> Oh, the patch continues...
>>>>>>
>>>>>> @@ -3208,7 +3208,10 @@ verify_gimple_comparison (tree type, tre
>>>>>>        && (!POINTER_TYPE_P (op0_type)
>>>>>>           || !POINTER_TYPE_P (op1_type)
>>>>>>           || TYPE_MODE (op0_type) != TYPE_MODE (op1_type)))
>>>>>> -      || !INTEGRAL_TYPE_P (type))
>>>>>> +      || !(TREE_CODE (type) == BOOLEAN_TYPE
>>>>>> +          || (TREE_TYPE (type) && TREE_CODE (type) == INTEGER_TYPE
>>>>>> +              && TREE_CODE (TREE_TYPE (type)) == BOOLEAN_TYPE)
>>>>>> +          || (INTEGRAL_TYPE_P (type) && TYPE_PRECISION (type) == 1)))
>>>>>>     {
>>>>>>
>>>>>> why that strange TREE_TYPE (TREE_TYPE ())) thing again?  Drop
>>>>>> that.
>>>>>>
>>>>>> @@ -3352,6 +3355,8 @@ verify_gimple_assign_unary (gimple stmt)
>>>>>>     case TRUTH_NOT_EXPR:
>>>>>>       /* We require two-valued operand types.  */
>>>>>>       if (!(TREE_CODE (rhs1_type) == BOOLEAN_TYPE
>>>>>> +           || (TREE_TYPE (rhs1_type) && TREE_CODE (rhs1_type) == INTEGER_TYPE
>>>>>> +               && TREE_CODE (TREE_TYPE (rhs1_type)) == BOOLEAN_TYPE)
>>>>>>            || (INTEGRAL_TYPE_P (rhs1_type)
>>>>>>                && TYPE_PRECISION (rhs1_type) == 1)))
>>>>>>         {
>>>>>>
>>>>>> likewise.
>>>>>
>>>>> Well, those checks are necessary for Ada and its crippled
>>>>> boolean_type_node and computed boolean-based integer construct.  Ada
>>>>> derives here the boolean-type to an integer with range 0..1 and the
>>>>> only way to find out that it is in fact such a beast is by looking
>>>>> into TREE_TYPE of type.  See here Ada's code for getting base-type
>>>>> information.
>>>>> As such things are treated as compatible they can appear for TRUTH_NOT
>>>>> expressions and comparisons.
>>>>
>>>> No they can't as we boolify them.
>>>>
>>>> Richard.
>>>
>>> Well, they do.  Test can prove this by running fortran's dg tests.
>>>
>>> The logic we do in gimplification is:
>>>
>>> 1) save original type of expression.
>>> 2) set expression's type to boolean_type_node.
>>> 3) if conversion from boolean_type_node to saved type isn't useless
>>> (and here we have the issue about Fortran's different booleans with
>>> different modes, which make those kinds not useless) then cast
>>> expression's result to original type and retry gimplification.
>>> 4) Otherwise gimplify operands and continue.
>>
>> You don't make sense.  Btw,
>>
>>>>>> +                 if (!useless_type_conversion_p (org_type, boolean_type_node))
>>>>>> +                   {
>>>>>> +                     TREE_TYPE (*expr_p) = boolean_type_node;
>>>>>> +                     *expr_p = fold_convert_loc (saved_location, org_type,
>>
>> is bogus it should be
>>
>> TREE_TYPE (*expr_p) = boolean_type_node;
>> if (!useless_type_conversion_p (org_type, boolean_type_node))
>>  {
>>    *expr_p = fold_convert_loc (saved_location, org_type,
>
> Well, I spared here the type conversion to boolean_type_node for
> useless conversion,

useless_type_conversion_p isn't symmetric, so you can't do that.

> but it doesn't change anything about the endless
> recursion here. As long as fold_convert_loc modifies *expr_p's type
> back to original type it will loop for ever for those cases I
> described above.

Indeed which is why we also touch fold-const.c!

> Please don't forget that boolean_type_node is FE type here and not the
> later 1-bit BOOL_SIZE type set in free_lang_data.

I perfectly know that.  Still no valid reason.

Richard.

> Kai
>