[patch tree-optimization]: Improve handling of conditional-branches on targets with high branch costs

Richard Guenther richard.guenther@gmail.com
Mon Oct 10 11:27:00 GMT 2011


On Fri, Oct 7, 2011 at 11:36 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
> Hello,
>
> this is the updated version with the suggestion
>
> 2011/10/7 Richard Guenther <richard.guenther@gmail.com>:
>> On Thu, Oct 6, 2011 at 4:25 PM, Kai Tietz <ktietz@redhat.com> wrote:
>>> +      && ((TREE_CODE_CLASS (TREE_CODE (arg1)) != tcc_comparison
>>> +           && TREE_CODE (arg1) != TRUTH_NOT_EXPR
>>> +           && simple_operand_p (arg1))
>>
>> As I said previously simple_operand_p already rejects covers
>> comparisons and TRUTH_NOT_EXPR.  Also arg1 had better
>> TREE_SIDE_EFFECTS set if the comparison might trap, as
>> it might just be hidden in something more complicated - so
>> the simple check isn't enough anyway (and if simple_operand_p
>> would cover it, the check would be better placed there).
>
> I reworked simple_operand_p so that it does this special-casing and additionally
> also checks for trapping.
>
>>> +      if (TREE_CODE (arg0) == code
>>> +          && !TREE_SIDE_EFFECTS (TREE_OPERAND (arg0, 1))
>>> +          && simple_operand_p (TREE_OPERAND (arg0, 1)))
>>> +       {
>>> +         tem = build2_loc (loc,
>>> +                           (code == TRUTH_ANDIF_EXPR ? TRUTH_AND_EXPR
>>> +                                                     : TRUTH_OR_EXPR),
>>> +                           type, TREE_OPERAND (arg0, 1), arg1);
>>> +         return build2_loc (loc, code, type, TREE_OPERAND (arg0, 0),
>>> +                            tem);
>>
>> All trees should be folded, don't use plain build without a good reason.
>
> Ok, done
>
>>> +       }
>>> +      /* Convert X TRUTH-ANDORIF Y to X TRUTH-ANDOR Y, if X and Y
>>> +        are simple operands and have no side-effects.  */
>>> +      if (simple_operand_p (arg0)
>>> +          && !TREE_SIDE_EFFECTS (arg0))
>>
>> Again, the checks you do for arg0 do not match those for arg1.  OTOH
>> it doesn't matter whether arg0 is simple or not or has side-effects or
>> not for this transformation, so why check it at all?
>
> It is required.  For left-hand operand, if it isn't a logical
> and/or/xor, we need to check for side-effects (and for trapping).  I
> see that calling of simple_operand_p is wrong here, as it rejects too
> much.  Nevertheless the check for side-effects is necessary for having
> valid sequence-points.  Without that checking a simple test
>
> int getter (void);
>
> int foo (void)
> {
>  int c, r = 0;
>  while ((c = getter ()) != '"' && c >= 0)
>    r +=c;
>  return r;
> }
>
> would give a warning about sequence-points.  As left-hand operand has
> side-effects, but right-hand not.  If we would combine it as AND, the
> operands are exchange-able.  So right-hand operand needs to be another
> ANDIF expression instead.
> Same apply on trapping.
>
>> In fold_truthop we still have the same (albeit more restricted transform),
>> but guarded with
>>
>>  if (BRANCH_COST (optimize_function_for_speed_p (cfun),
>>                   false) >= 2
>>
>> too.  Why not here?  Please delete redundant code in fold_truthop.
> Well, in general this is the default definition of
> LOGICAL_OP_NON_SHORT_CIRCUIT, so I missed that.  As for some targets
> the macro LOGICAL_OP_NON_SHORT_CIRCUIT might be defined differently,
> it might make sense to check for BRANCH_COST again.
>
>> It's also odd that this is only called from fold_truth_andor but has
>> a more generic name, so maybe rename it to fold_truth_andor_1 on the way.
>
> I renamed it.
>
>> Richard.
>
> ChangeLog
>
> 2011-10-07  Kai Tietz  <ktietz@redhat.com>
>
>        * fold-const.c (simple_operand_p): Make argument non-const
>        and add floating-point trapping check, and special cases for
>        comparisons, and logical-not's.
>        (fold_truthop): Rename to
>        (fold_truth_andor_1): function name.
>        Additionally remove here TRUTH-AND|OR_EXPR generation.
>        (fold_truth_andor): Handle branching at one place.
>
> Bootstrapped and regression-tested for all languages plus Ada and
> Obj-C++ on x86_64-pc-linux-gnu.
> Ok for apply?
>
> Regards,
> Kai
>
> Index: gcc/gcc/fold-const.c
> ===================================================================
> --- gcc.orig/gcc/fold-const.c
> +++ gcc/gcc/fold-const.c
> @@ -111,14 +111,13 @@ static tree decode_field_reference (loca
>                                    tree *, tree *);
>  static int all_ones_mask_p (const_tree, int);
>  static tree sign_bit_p (tree, const_tree);
> -static int simple_operand_p (const_tree);
> +static int simple_operand_p (tree);
>  static tree range_binop (enum tree_code, tree, tree, int, tree, int);
>  static tree range_predecessor (tree);
>  static tree range_successor (tree);
>  static tree fold_range_test (location_t, enum tree_code, tree, tree, tree);
>  static tree fold_cond_expr_with_comparison (location_t, tree, tree,
> tree, tree);
>  static tree unextend (tree, int, int, tree);
> -static tree fold_truthop (location_t, enum tree_code, tree, tree, tree);
>  static tree optimize_minmax_comparison (location_t, enum tree_code,
>                                        tree, tree, tree);
>  static tree extract_muldiv (tree, tree, enum tree_code, tree, bool *);
> @@ -3500,7 +3499,7 @@ optimize_bit_field_compare (location_t l
>   return lhs;
>  }
>
> -/* Subroutine for fold_truthop: decode a field reference.
> +/* Subroutine for fold_truth_andor_1: decode a field reference.
>
>    If EXP is a comparison reference, we return the innermost reference.
>
> @@ -3668,17 +3667,43 @@ sign_bit_p (tree exp, const_tree val)
>   return NULL_TREE;
>  }
>
> -/* Subroutine for fold_truthop: determine if an operand is simple enough
> +/* Subroutine for fold_truth_andor_1: determine if an operand is simple enough
>    to be evaluated unconditionally.  */
>
>  static int
> -simple_operand_p (const_tree exp)
> +simple_operand_p (tree exp)
>  {
> +  enum tree_code code;
>   /* Strip any conversions that don't change the machine mode.  */
>   STRIP_NOPS (exp);
>
> +  code = TREE_CODE (exp);
> +
> +  /* Handle some trivials  $$$$ */
> +  if (TREE_CODE_CLASS (code) == tcc_comparison)
> +    return (tree_could_trap_p (exp)
> +           && simple_operand_p (TREE_OPERAND (exp, 0))
> +           && simple_operand_p (TREE_OPERAND (exp, 1)));

Clearly wrong.  And what's $$$$ supposed to be?

> +
> +  if (FLOAT_TYPE_P (TREE_TYPE (exp))
> +      && tree_could_trap_p (exp))
> +    return false;
> +
> +  switch (code)
> +    {
> +    case SSA_NAME:
> +      return true;
> +    case TRUTH_NOT_EXPR:
> +      return simple_operand_p (TREE_OPERAND (exp, 0));
> +    case BIT_NOT_EXPR:
> +      if (TREE_CODE (TREE_TYPE (exp)) != BOOLEAN_TYPE)
> +        return false;
> +      return simple_operand_p (TREE_OPERAND (exp, 0));
> +    default:
> +      break;
> +    }
> +

You add a lot of cases here without a good reason.  Why in this patch?
Simply removing the tcc_comparison checks would have been enough ...

>   return (CONSTANT_CLASS_P (exp)
> -         || TREE_CODE (exp) == SSA_NAME
>          || (DECL_P (exp)
>              && ! TREE_ADDRESSABLE (exp)
>              && ! TREE_THIS_VOLATILE (exp)
> @@ -4888,7 +4913,7 @@ fold_range_test (location_t loc, enum tr
>   return 0;
>  }
>
> -/* Subroutine for fold_truthop: C is an INTEGER_CST interpreted as a P
> +/* Subroutine for fold_truth_andor_1: C is an INTEGER_CST interpreted as a P

fold_truth_andor

>    bit value.  Arrange things so the extra bits will be set to zero if and
>    only if C is signed-extended to its full width.  If MASK is nonzero,
>    it is an INTEGER_CST that should be AND'ed with the extra bits.  */
> @@ -5025,8 +5050,8 @@ merge_truthop_with_opposite_arm (locatio
>    We return the simplified tree or 0 if no optimization is possible.  */
>
>  static tree
> -fold_truthop (location_t loc, enum tree_code code, tree truth_type,
> -             tree lhs, tree rhs)
> +fold_truth_andor_1 (location_t loc, enum tree_code code, tree truth_type,
> +                   tree lhs, tree rhs)
>  {
>   /* If this is the "or" of two comparisons, we can do something if
>      the comparisons are NE_EXPR.  If this is the "and", we can do something
> @@ -5054,8 +5079,6 @@ fold_truthop (location_t loc, enum tree_
>   tree lntype, rntype, result;
>   HOST_WIDE_INT first_bit, end_bit;
>   int volatilep;
> -  tree orig_lhs = lhs, orig_rhs = rhs;
> -  enum tree_code orig_code = code;
>
>   /* Start by getting the comparison codes.  Fail if anything is volatile.
>      If one operand is a BIT_AND_EXPR with the constant one, treat it as if
> @@ -5119,8 +5142,7 @@ fold_truthop (location_t loc, enum tree_
>   /* If the RHS can be evaluated unconditionally and its operands are
>      simple, it wins to evaluate the RHS unconditionally on machines
>      with expensive branches.  In this case, this isn't a comparison
> -     that can be merged.  Avoid doing this if the RHS is a floating-point
> -     comparison since those can trap.  */
> +     that can be merged.  */
>
>   if (BRANCH_COST (optimize_function_for_speed_p (cfun),
>                   false) >= 2
> @@ -5149,13 +5171,6 @@ fold_truthop (location_t loc, enum tree_
>                           build2 (BIT_IOR_EXPR, TREE_TYPE (ll_arg),
>                                   ll_arg, rl_arg),
>                           build_int_cst (TREE_TYPE (ll_arg), 0));
> -
> -      if (LOGICAL_OP_NON_SHORT_CIRCUIT)
> -       {
> -         if (code != orig_code || lhs != orig_lhs || rhs != orig_rhs)
> -           return build2_loc (loc, code, truth_type, lhs, rhs);
> -         return NULL_TREE;
> -       }
>     }
>
>   /* See if the comparisons can be merged.  Then get all the parameters for
> @@ -8380,13 +8395,52 @@ fold_truth_andor (location_t loc, enum t
>      lhs is another similar operation, try to merge its rhs with our
>      rhs.  Then try to merge our lhs and rhs.  */
>   if (TREE_CODE (arg0) == code
> -      && 0 != (tem = fold_truthop (loc, code, type,
> -                                  TREE_OPERAND (arg0, 1), arg1)))
> +      && 0 != (tem = fold_truth_andor_1 (loc, code, type,
> +                                        TREE_OPERAND (arg0, 1), arg1)))
>     return fold_build2_loc (loc, code, type, TREE_OPERAND (arg0, 0), tem);
>
> -  if ((tem = fold_truthop (loc, code, type, arg0, arg1)) != 0)
> +  if ((tem = fold_truth_andor_1 (loc, code, type, arg0, arg1)) != 0)
>     return tem;
>
> +  if ((code == TRUTH_ANDIF_EXPR || code == TRUTH_ORIF_EXPR)
> +      && (BRANCH_COST (optimize_function_for_speed_p (cfun),
> +                      false) >= 2)
> +      && !TREE_SIDE_EFFECTS (arg1)
> +      && LOGICAL_OP_NON_SHORT_CIRCUIT
> +      && simple_operand_p (arg1))
> +    {
> +      enum tree_code ncode = (code == TRUTH_ANDIF_EXPR ? TRUTH_AND_EXPR
> +                                                      : TRUTH_OR_EXPR);
> +
> +      /* We don't want to pack more then two non-IF branches
> +         together.  Therefore we need to check, if rhs isn't
> +         already an TRUTH_(XOR|OR|AND)[IF]_EXPR.  */

which means, just check if (simple_operand_p (arg0)) ...

> +      if (TREE_CODE (arg0) == code
> +         && TREE_CODE (TREE_OPERAND (arg0, 1)) != TRUTH_AND_EXPR
> +         && TREE_CODE (TREE_OPERAND (arg0, 1)) != TRUTH_OR_EXPR
> +         && TREE_CODE (TREE_OPERAND (arg0, 1)) != TRUTH_XOR_EXPR
> +         && TREE_CODE (TREE_OPERAND (arg0, 1)) != TRUTH_ANDIF_EXPR
> +         && TREE_CODE (TREE_OPERAND (arg0, 1)) != TRUTH_ORIF_EXPR
> +         /* Needed for sequence points and trappings, or side-effects.  */
> +         && !TREE_SIDE_EFFECTS (TREE_OPERAND (arg0, 1))
> +         && !tree_could_trap_p (TREE_OPERAND (arg0, 1)))
> +       {
> +         tem = fold_build2_loc (loc, ncode, type, TREE_OPERAND (arg0, 1),
> +                               arg1);
> +         return fold_build2_loc (loc, code, type, TREE_OPERAND (arg0, 0),
> +                                tem);

Don't do this association here.  Why should it not be applied to
(a TRUTH_ANDIF_EXPR b) TRUTH_AND_EXPR c?  Thus, if you
want to associate the above if b and c do not have side-effects then
do so generally, not only when you are converting a TRUTH_ANDIF_EXPR
to TRUTH_AND_EXPR.

Thus, please remove this association code.

> +       }
> +     else if (TREE_CODE (arg0) != TRUTH_AND_EXPR
> +             && TREE_CODE (arg0) != TRUTH_OR_EXPR
> +             && TREE_CODE (arg0) != TRUTH_ANDIF_EXPR
> +             && TREE_CODE (arg0) != TRUTH_ORIF_EXPR
> +             && TREE_CODE (arg0) != TRUTH_XOR_EXPR
> +             /* Needed for sequence points and trappings, or side-effects.  */
> +             && !TREE_SIDE_EFFECTS (arg0)
> +             && !tree_could_trap_p (arg0))

See above.

You made the patch more complex when I asked for a simpler one.

Richard.

> +       return fold_build2_loc (loc, ncode, type, arg0, arg1);
> +    }
> +
>   return NULL_TREE;
>  }
>



More information about the Gcc-patches mailing list