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

[Richard Guenther <richard dot guenther at gmail dot com>]

On Mon, Oct 17, 2011 at 1:31 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
> 2011/10/17 Richard Guenther <richard.guenther@gmail.com>:
>> On Mon, Oct 17, 2011 at 12:59 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>> 2011/10/17 Richard Guenther <richard.guenther@gmail.com>:
>>>> On Fri, Oct 14, 2011 at 9:43 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>>>> Hello,
>>>>>
>>>>> So I committed the gimplify patch separate. ?And here is the remaining
>>>>> fold-const patch.
>>>>> The important tests here are in gcc.dg/tree-ssa/builtin-expect[1-4].c, which
>>>>> cover the one special-case for branching. Also tree-ssa/20040204-1.c covers
>>>>> tests for branching code (on targets having high-engough BRANCH_COST and no
>>>>> special-casing - like MIPS, S/390, and AVR.
>>>>>
>>>>> ChangeLog
>>>>>
>>>>> 2011-10-14 ?Kai Tietz ?<ktietz@redhat.com>
>>>>>
>>>>> ? ? ? ?* fold-const.c (simple_operand_p_2): New function.
>>>>> ? ? ? ?(fold_truthop): Rename to
>>>>> ? ? ? ?(fold_truth_andor_1): function name.
>>>>> ? ? ? ?Additionally remove branching creation for logical and/or.
>>>>> ? ? ? ?(fold_truth_andor): Handle branching creation for logical and/or here.
>>>>>
>>>>> Bootstrapped and regression-tested for all languages plus Ada and
>>>>> Obj-C++ on x86_64-pc-linux-gnu.
>>>>> Ok for apply?
>>>>
>>>> Ok with ...
>>>>
>>>>> Regards,
>>>>> Kai
>>>>>
>>>>> Index: gcc/gcc/fold-const.c
>>>>> ===================================================================
>>>>> --- gcc.orig/gcc/fold-const.c
>>>>> +++ gcc/gcc/fold-const.c
>>>>> @@ -112,13 +112,13 @@ static tree decode_field_reference (loca
>>>>> ?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 bool simple_operand_p_2 (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 +3500,7 @@ 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.
>>>>>
>>>>> @@ -3668,7 +3668,7 @@ 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
>>>>> @@ -3678,7 +3678,7 @@ simple_operand_p (const_tree exp)
>>>>> ? STRIP_NOPS (exp);
>>>>>
>>>>> ? return (CONSTANT_CLASS_P (exp)
>>>>> - ? ? ? ? || TREE_CODE (exp) == SSA_NAME
>>>>> + ? ? ? ? || TREE_CODE (exp) == SSA_NAME
>>>>> ? ? ? ? ?|| (DECL_P (exp)
>>>>> ? ? ? ? ? ? ?&& ! TREE_ADDRESSABLE (exp)
>>>>> ? ? ? ? ? ? ?&& ! TREE_THIS_VOLATILE (exp)
>>>>> @@ -3692,6 +3692,46 @@ simple_operand_p (const_tree exp)
>>>>> ? ? ? ? ? ? ? ? registers aren't expensive. ?*/
>>>>> ? ? ? ? ? ? ?&& (! TREE_STATIC (exp) || DECL_REGISTER (exp))));
>>>>> ?}
>>>>> +
>>>>> +/* Subroutine for fold_truth_andor: determine if an operand is simple enough
>>>>> + ? to be evaluated unconditionally.
>>>>> + ? I addition to simple_operand_p, we assume that comparisons and logic-not
>>>>> + ? operations are simple, if their operands are simple, too. ?*/
>>>>> +
>>>>> +static bool
>>>>> +simple_operand_p_2 (tree exp)
>>>>> +{
>>>>> + ?enum tree_code code;
>>>>> +
>>>>> + ?/* Strip any conversions that don't change the machine mode. ?*/
>>>>> + ?STRIP_NOPS (exp);
>>>>> +
>>>>> + ?code = TREE_CODE (exp);
>>>>> +
>>>>> + ?if (TREE_CODE_CLASS (code) == tcc_comparison)
>>>>> + ? ?return (!tree_could_trap_p (exp)
>>>>> + ? ? ? ? ? && simple_operand_p_2 (TREE_OPERAND (exp, 0))
>>>>> + ? ? ? ? ? && simple_operand_p_2 (TREE_OPERAND (exp, 1)));
>>>>
>>>> recurse with simple_operand_p.
>>>
>>> No, as this again would reject simple operations and additionally
>>> wouldn't check for trapping.
>>
>> ? ?Your code allows arbitrarily complex expressions. ?Also
>> tree_could_trap_p obviously extents to operands.
>
> Ah, ok. I wasn't aware that it walks into tree.
>
>>>
>>>>> +
>>>>> + ?if (TREE_SIDE_EFFECTS (exp)
>>>>> + ? ? ?|| tree_could_trap_p (exp))
>>>>
>>>> Move this check before the tcc_comparison check and remove the
>>>> then redundant tree_could_trap_p check there.
>>>
>>> Ok
>>>
>>>>> + ? ?return false;
>>>>> +
>>>>> + ?switch (code)
>>>>> + ? ?{
>>>>> + ? ?case SSA_NAME:
>>>>> + ? ? ?return true;
>>>>
>>>> Do not handle here, it's handled in simple_operand_p.
>>>
>>> Well, was more a short-cut here.
>>>
>>>>> + ? ?case TRUTH_NOT_EXPR:
>>>>> + ? ? ?return simple_operand_p_2 (TREE_OPERAND (exp, 0));
>>>>> + ? ?case BIT_NOT_EXPR:
>>>>> + ? ? ?if (TREE_CODE (TREE_TYPE (exp)) != BOOLEAN_TYPE)
>>>>> + ? ? ? return false;
>>>>
>>>> Remove the BIT_NOT_EXPR handling. ?Thus, simply change this switch
>>>> to
>>>
>>> Why should we reject simple ~X operations from gimplified code here?
>>
>> Because this is FE triggered code. ?From gimple you won't ever see
>> such complex expressions (never even the TRUTH_AND*_EXPR variants).
>
> Hmm, I thought we might see such thing in fold and/or. ?But well, you
> might be right.
>
>>> I admit that from FE-code we won't see that, as always an integer-cast
>>> is done for foo (_Bool x) { ... if (~x) ... }, but from
>>> gimplified-code this is the general description of an boolean-typed !=
>>> 0?
>>>
>>>> if (code == TRUTH_NOT_EXPR)
>>>> ?return simple_operand_p_2 (TREE_OPERAND (exp, 0));
>>>>
>>>> return simple_operand_p (exp);
>>>>
>>>>> + ? ? ?return simple_operand_p_2 (TREE_OPERAND (exp, 0));
>>>>> + ? ?default:
>>>>> + ? ? ?return simple_operand_p (exp);
>>>>> + ? ?}
>>>>> +}
>>>>> +
>>>>>
>>>>> ?/* The following functions are subroutines to fold_range_test and allow it to
>>>>> ? ?try to change a logical combination of comparisons into a range test.
>>>>> @@ -4888,7 +4928,7 @@ 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
>>>>> ? ?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 +5065,8 @@ 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 +5094,6 @@ 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 +5157,7 @@ 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 +5186,6 @@ 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 +8410,49 @@ 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)
>>>>> + ? ? ?&& LOGICAL_OP_NON_SHORT_CIRCUIT
>>>>> + ? ? ?&& simple_operand_p_2 (arg1))
>>>>> + ? ?{
>>>>> + ? ? ?enum tree_code ncode = (code == TRUTH_ANDIF_EXPR ? TRUTH_AND_EXPR
>>>>> + ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?: TRUTH_OR_EXPR);
>>>>> +
>>>>> + ? ? ?/* Transform ((A AND-IF B) AND-IF C) into (A AND-IF (B AND C)),
>>>>> + ? ? ? ? or ((A OR-IF B) OR-IF C) into (A OR-IF (B OR C))
>>>>> + ? ? ? ? We don't want to pack more than two leafs to a non-IF AND/OR
>>>>> + ? ? ? ? expression.
>>>>> + ? ? ? ? If tree-code of left-hand operand isn't an AND/OR-IF code and not
>>>>> + ? ? ? ? equal to CODE, then we don't want to add right-hand operand.
>>>>> + ? ? ? ? If the inner right-hand side of left-hand operand has side-effects,
>>>>> + ? ? ? ? or isn't simple, then we can't add to it, as otherwise we might
>>>>> + ? ? ? ? destroy if-sequence. ?*/
>>>>> + ? ? ?if (TREE_CODE (arg0) == code
>>>>> + ? ? ? ? /* Needed for sequence points to handle trappings, and
>>>>> + ? ? ? ? ? ?side-effects. ?*/
>>>>> + ? ? ? ? && simple_operand_p_2 (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);
>>>>> + ? ? ? }
>>>>
>>>> I see you insist on this change. ?Let me explain again. ?You do this
>>>> for ((A AND-IF B) AND-IF C) but you don't do this for
>>>> ((A AND-IF B) AND C). ?Why? ?That is what doesn't make sense ot me.
>>>> Thus omit this hunk.
>>>
>>> Well, first ((A AND-IF B) AND C) would be an ill sequence, ?as AND is
>>> associative. So we would simply break sequence points for && and ||.
>>> If left-hand operand is an AND/OR-IF then outer operand has to always
>>> an ?-IF operation, too.
>>
>> Why? ?It's something like (ptr && *ptr) & x. ?Whether you evaluate
>> x or (ptr && *ptr) first does not matter. ?But you have to check
>> whether ptr is non-null before dereferencing it. ?So it's clearly not
>> ill-formed. ?You may argue the transform is pointless and we should
>> associate the & instead. ?Do you?
>
> well, if you are explict writing such thing as binary-and, it would be
> associative anyway and code doesn't change here anything. ?Binary and
> != logical and. The point about if we see something as (A TRUTH-IF B)
> TRUTH B), we don't want to change it at all. ?The outer if for this
> already checks that this operation is just to be used on TRUTH-IF. ?To
> modify a TRUTH to a TRUTH is pretty point-less, isn't it? ?If we would
> allow to sink the case (A TRUTH-IF B) TRUTH C to (A TRUTH-IF (B TRUTH
> C)), which might be of some intererest, but still would change
> association rule here from point of C specification. ?By C standard
> each ||,&& is treated as a separate sequence-point. ?Only in case that
> previous and next &&/|| operand have no side-effects, we can apply to
> them associative law. ?Or do I read C-spec here wrong?

Certainly if for (A TRUTH-IF B) TRUTH-IF C it is valid to associate
it as A TRUTH-IF (B IF C) then it is valid to do the same for
(A TRUTH-IF B) IF C.

> Regards,
> Kai
>