Re: [patch tree-optimization]: [2 of 3]: Boolify compares & more

[Kai Tietz <ktietz70@googlemail.com>]

2011/7/8 Richard Guenther <richard.guenther@gmail.com>:
> On Fri, Jul 8, 2011 at 1:32 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>> 2011/7/8 Richard Guenther <richard.guenther@gmail.com>:
>>> On Thu, Jul 7, 2011 at 6:07 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>>> Hello,
>>>>
>>>> This patch - second of series - adds boolification of comparisions in
>>>> gimplifier.  For this
>>>> casts from/to boolean are marked as not-useless. And in fold_unary_loc
>>>> casts to non-boolean integral types are preserved.
>>>> The hunk in tree-ssa-forwprop.c in combine_cond-expr_cond is not strictly
>>>> necessary - as long as fold-const handles 1-bit precision bitwise-expression
>>>> with truth-logic - but it has shown to short-cut some expensier folding. So
>>>> I kept it within this patch.
>>>
>>> Please split it out.  Also ...
>>>
>>>>
>>>> The adjusted testcase gcc.dg/uninit-15.c indicates that due
>>>> optimization we loose
>>>> in this case variables declaration.  But this might be to be expected.
>>>>
>>>> In vectorization we have a regression in gcc.dg/vect/vect-cond-3.c
>>>> test-case.  It's caused
>>>> by always having boolean-type on conditions.  So vectorizer sees
>>>> different types, which
>>>> aren't handled by vectorizer right now.  Maybe this issue could be
>>>> special-cased for
>>>> boolean-types in tree-vect-loop, by making operand for used condition
>>>> equal to vector-type.
>>>> But this is a subject for a different patch and not addressed by this series.
>>>>
>>>> There is a regressions in tree-ssa/vrp47.c, and the fix is addressed
>>>> by the 3rd patch of this
>>>> series.
>>>>
>>>> Bootstrapped and regression tested for all standard-languages (plus
>>>> Ada and Obj-C++) on host x86_64-pc-linux-gnu.
>>>>
>>>> Ok for apply?
>>>>
>>>> Regards,
>>>> Kai
>>>>
>>>>
>>>> ChangeLog
>>>>
>>>> 2011-07-07  Kai Tietz  <ktietz@redhat.com>
>>>>
>>>>        * fold-const.c (fold_unary_loc): Preserve
>>>>        non-boolean-typed casts.
>>>>        * gimplify.c (gimple_boolify): Handle boolification
>>>>        of comparisons.
>>>>        (gimplify_expr): Boolifiy non aggregate-typed
>>>>        comparisons.
>>>>        * tree-cfg.c (verify_gimple_comparison): Check result
>>>>        type of comparison expression.
>>>>        * tree-ssa.c (useless_type_conversion_p): Preserve incompatible
>>>>        casts from/to boolean,
>>>>        * tree-ssa-forwprop.c (combine_cond_expr_cond): Add simplification
>>>>        support for one-bit-precision typed X for cases X != 0 and X == 0.
>>>>        (forward_propagate_comparison): Adjust test of condition
>>>>        result.
>>>>
>>>>
>>>>        * gcc.dg/tree-ssa/builtin-expect-5.c: Adjusted.
>>>>        * gcc.dg/tree-ssa/pr21031.c: Likewise.
>>>>        * gcc.dg/tree-ssa/pr30978.c: Likewise.
>>>>        * gcc.dg/tree-ssa/ssa-fre-6.c: Likewise.
>>>>        * gcc.dg/binop-xor1.c: Mark it as expected fail.
>>>>        * gcc.dg/binop-xor3.c: Likewise.
>>>>        * gcc.dg/uninit-15.c: Adjust reported message.
>>>>
>>>> Index: gcc-head/gcc/fold-const.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/fold-const.c
>>>> +++ gcc-head/gcc/fold-const.c
>>>> @@ -7665,11 +7665,11 @@ fold_unary_loc (location_t loc, enum tre
>>>>             non-integral type.
>>>>             Do not fold the result as that would not simplify further, also
>>>>             folding again results in recursions.  */
>>>> -         if (INTEGRAL_TYPE_P (type))
>>>> +         if (TREE_CODE (type) == BOOLEAN_TYPE)
>>>>            return build2_loc (loc, TREE_CODE (op0), type,
>>>>                               TREE_OPERAND (op0, 0),
>>>>                               TREE_OPERAND (op0, 1));
>>>> -         else
>>>> +         else if (!INTEGRAL_TYPE_P (type))
>>>>            return build3_loc (loc, COND_EXPR, type, op0,
>>>>                               fold_convert (type, boolean_true_node),
>>>>                               fold_convert (type, boolean_false_node));
>>>> Index: gcc-head/gcc/gimplify.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/gimplify.c
>>>> +++ gcc-head/gcc/gimplify.c
>>>> @@ -2842,18 +2842,23 @@ gimple_boolify (tree expr)
>>>>
>>>>     case TRUTH_NOT_EXPR:
>>>>       TREE_OPERAND (expr, 0) = gimple_boolify (TREE_OPERAND (expr, 0));
>>>> -      /* FALLTHRU */
>>>>
>>>> -    case EQ_EXPR: case NE_EXPR:
>>>> -    case LE_EXPR: case GE_EXPR: case LT_EXPR: case GT_EXPR:
>>>>       /* These expressions always produce boolean results.  */
>>>> -      TREE_TYPE (expr) = boolean_type_node;
>>>> +      if (TREE_CODE (type) != BOOLEAN_TYPE)
>>>> +       TREE_TYPE (expr) = boolean_type_node;
>>>>       return expr;
>>>>
>>>>     default:
>>>> +      if (COMPARISON_CLASS_P (expr))
>>>> +       {
>>>> +         /* There expressions always prduce boolean results.  */
>>>> +         if (TREE_CODE (type) != BOOLEAN_TYPE)
>>>> +           TREE_TYPE (expr) = boolean_type_node;
>>>> +         return expr;
>>>> +       }
>>>>       /* Other expressions that get here must have boolean values, but
>>>>         might need to be converted to the appropriate mode.  */
>>>> -      if (type == boolean_type_node)
>>>> +      if (TREE_CODE (type) == BOOLEAN_TYPE)
>>>>        return expr;
>>>>       return fold_convert_loc (loc, boolean_type_node, expr);
>>>>     }
>>>> @@ -6763,7 +6768,7 @@ gimplify_expr (tree *expr_p, gimple_seq
>>>>            tree org_type = TREE_TYPE (*expr_p);
>>>>
>>>>            *expr_p = gimple_boolify (*expr_p);
>>>> -           if (org_type != boolean_type_node)
>>>> +           if (!useless_type_conversion_p (org_type, TREE_TYPE (*expr_p)))
>>>>              {
>>>>                *expr_p = fold_convert (org_type, *expr_p);
>>>
>>> Use fold_convert_loc with saved_location
>>
>> Oh, good catch. Yes, I will adjust that.
>>
>>>>                ret = GS_OK;
>>>> @@ -7208,7 +7213,7 @@ gimplify_expr (tree *expr_p, gimple_seq
>>>>               fold_truth_not_expr) happily uses operand type and doesn't
>>>>               automatically uses boolean_type as result, we need to keep
>>>>               orignal type.  */
>>>> -           if (org_type != boolean_type_node)
>>>> +           if (!useless_type_conversion_p (org_type, TREE_TYPE (*expr_p)))
>>>>              {
>>>>                *expr_p = fold_convert (org_type, *expr_p);
>>>
>>> Likewise.  Maybe this fixes the diagnostic regression.
>>>
>>>>                ret = GS_OK;
>>>> @@ -7288,7 +7293,19 @@ gimplify_expr (tree *expr_p, gimple_seq
>>>>                  tree type = TREE_TYPE (TREE_OPERAND (*expr_p, 1));
>>>>
>>>>                  if (!AGGREGATE_TYPE_P (type))
>>>> -                   goto expr_2;
>>>> +                   {
>>>> +                     tree org_type = TREE_TYPE (*expr_p);
>>>> +                     *expr_p = gimple_boolify (*expr_p);
>>>> +                     if (!useless_type_conversion_p (org_type,
>>>> +                                                     TREE_TYPE (*expr_p)))
>>>> +                       {
>>>> +                         *expr_p = fold_convert_loc (saved_location,
>>>> +                                                     org_type, *expr_p);
>>>> +                         ret = GS_OK;
>>>> +                       }
>>>> +                     else
>>>> +                       goto expr_2;
>>>> +                   }
>>>>                  else if (TYPE_MODE (type) != BLKmode)
>>>>                    ret = gimplify_scalar_mode_aggregate_compare (expr_p);
>>>>                  else
>>>> Index: gcc-head/gcc/tree-cfg.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/tree-cfg.c
>>>> +++ gcc-head/gcc/tree-cfg.c
>>>> @@ -3203,7 +3203,9 @@ 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))
>>>> +      || !INTEGRAL_TYPE_P (type)
>>>> +      || (TREE_CODE (type) != BOOLEAN_TYPE
>>>> +         && TYPE_PRECISION (type) != 1))
>>>>     {
>>>>       error ("type mismatch in comparison expression");
>>>>       debug_generic_expr (type);
>>>> Index: gcc-head/gcc/tree-ssa.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/tree-ssa.c
>>>> +++ gcc-head/gcc/tree-ssa.c
>>>> @@ -1306,10 +1306,10 @@ useless_type_conversion_p (tree outer_ty
>>>>          || TYPE_PRECISION (inner_type) != TYPE_PRECISION (outer_type))
>>>>        return false;
>>>>
>>>> -      /* Preserve conversions to BOOLEAN_TYPE if it is not of precision
>>>> -         one.  */
>>>> -      if (TREE_CODE (inner_type) != BOOLEAN_TYPE
>>>> -         && TREE_CODE (outer_type) == BOOLEAN_TYPE
>>>> +      /* Preserve conversions to/from BOOLEAN_TYPE if types are not
>>>> +        of precision one.  */
>>>> +      if (((TREE_CODE (inner_type) == BOOLEAN_TYPE)
>>>> +          != (TREE_CODE (outer_type) == BOOLEAN_TYPE))
>>>>          && TYPE_PRECISION (outer_type) != 1)
>>>>        return false;
>>>>
>>>> Index: gcc-head/gcc/testsuite/gcc.dg/tree-ssa/builtin-expect-5.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/testsuite/gcc.dg/tree-ssa/builtin-expect-5.c
>>>> +++ gcc-head/gcc/testsuite/gcc.dg/tree-ssa/builtin-expect-5.c
>>>> @@ -11,5 +11,5 @@ f (int i, float j)
>>>>
>>>>  /* { dg-final { scan-tree-dump-times { if } 2 "forwprop1"} } */
>>>>  /* { dg-final { scan-tree-dump {builtin_expect[^\n]*, 0\);\n[^\n]*if}
>>>> "forwprop1"} } */
>>>> -/* { dg-final { scan-tree-dump {builtin_expect[^\n]*, 1\);\n[^\n]*if}
>>>> "forwprop1"} } */
>>>> +/* { dg-final { scan-tree-dump-not {builtin_expect[^\n]*,
>>>> 1\);\n[^\n]*if} "forwprop1"} } */
>>>
>>> Hm?  Why that?
>>>
>>>>  /* { dg-final { cleanup-tree-dump "forwprop?" } } */
>>>> Index: gcc-head/gcc/testsuite/gcc.dg/tree-ssa/pr21031.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/testsuite/gcc.dg/tree-ssa/pr21031.c
>>>> +++ gcc-head/gcc/testsuite/gcc.dg/tree-ssa/pr21031.c
>>>> @@ -16,5 +16,5 @@ foo (int a)
>>>>     return 0;
>>>>  }
>>>>
>>>> -/* { dg-final { scan-tree-dump-times "Replaced" 2 "forwprop1"} } */
>>>> +/* { dg-final { scan-tree-dump-times "Replaced" 1 "forwprop1"} } */
>>>>  /* { dg-final { cleanup-tree-dump "forwprop1" } } */
>>>> Index: gcc-head/gcc/testsuite/gcc.dg/tree-ssa/pr30978.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/testsuite/gcc.dg/tree-ssa/pr30978.c
>>>> +++ gcc-head/gcc/testsuite/gcc.dg/tree-ssa/pr30978.c
>>>> @@ -10,5 +10,5 @@ int foo(int a)
>>>>   return e;
>>>>  }
>>>>
>>>> -/* { dg-final { scan-tree-dump "e_. = a_..D. > 0;" "optimized" } } */
>>>> +/* { dg-final { scan-tree-dump " = a_..D. > 0;" "optimized" } } */
>>>>  /* { dg-final { cleanup-tree-dump "optimized" } } */
>>>> Index: gcc-head/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-6.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-6.c
>>>> +++ gcc-head/gcc/testsuite/gcc.dg/tree-ssa/ssa-fre-6.c
>>>> @@ -2,5 +2,5 @@
>>>>  /* { dg-options "-O -fdump-tree-fre1-details" } */
>>>>
>>>>  int i; int foo(void) { i = 2; int j = i * 2; int k = i + 2; return j == k; }
>>>> -/* { dg-final { scan-tree-dump-times "Replaced " 5 "fre1" } } */
>>>> +/* { dg-final { scan-tree-dump-times "Replaced " 6 "fre1" } } */
>>>>  /* { dg-final { cleanup-tree-dump "fre1" } } */
>>>> Index: gcc-head/gcc/tree-ssa-forwprop.c
>>>> ===================================================================
>>>> --- gcc-head.orig/gcc/tree-ssa-forwprop.c
>>>> +++ gcc-head/gcc/tree-ssa-forwprop.c
>>>> @@ -367,9 +367,61 @@ combine_cond_expr_cond (location_t loc,
>>>>   gcc_assert (TREE_CODE_CLASS (code) == tcc_comparison);
>>>>
>>>>   t = fold_binary_loc (loc, code, type, op0, op1);
>>>> +
>>>> +  if (!t && INTEGRAL_TYPE_P (TREE_TYPE (op1))
>>>> +      && TYPE_PRECISION (TREE_TYPE (op1)) == 1
>>>> +      && (code == EQ_EXPR || code == NE_EXPR))
>>>> +    {
>>>> +      if (TREE_CODE (op1) == INTEGER_CST)
>>>> +        {
>>>> +         if (integer_onep (op1))
>>>> +           {
>>>> +             op1 = fold_convert_loc (loc, TREE_TYPE (op1), integer_zero_node);
>>>> +             code = (code == NE_EXPR ? EQ_EXPR : NE_EXPR);
>>>
>>> So you change truthvalue !=/== 1 to truthvalue ==/!= 0 and then
>>> recurse ... that doesn't make sense to me and is super-ugly.
>>> What's the testcase that made you add all this code?
>>
>> Well, the convert from truthvalue !=/== 1 to !=/== 0 limits the amount
>> of cases to handle. As for truthvalued X the we have then just to
>> handle two cases. X != 0 -> X, and X == 0 -> (X ^ 1).
>> The recursion is someting I saw as existing pattern (for the same
>> thing) in truth-op folding in fold-const.
>>
>> Actual I can remove this optimization here, as it should be convered
>> by VRP already (when VRP handles 1-bit precision bitwise ops proper).
>
> We should have a canonical form for those compares and change
> them accordingly, best in fold_stmt.
>
> Richard.

Hmm, I tried to add this code-pattern to fold_stmt, but for this kind
of branch it seems not to be invoked at all. At least not now without
boolification of compares.  One nit I found for GIMPLE_BINARY, as here
just patterns getting replaced, which have fewer number of ops then
original statement. This check looks a bit bogus.
For getting this normalization right now in a consistant way,
fold-const might be right now the better place to handle this.

Regards,
Kai