This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch 1/8 tree-optimization]: Bitwise logic for fold_truth_not_expr
On Wed, Jul 13, 2011 at 2:15 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
> 2011/7/13 Richard Guenther <richard.guenther@gmail.com>:
>> On Wed, Jul 13, 2011 at 1:08 PM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>> 2011/7/13 Richard Guenther <richard.guenther@gmail.com>:
>>>> On Wed, Jul 13, 2011 at 11:04 AM, Kai Tietz <ktietz70@googlemail.com> wrote:
>>>>> Sorrty, the TRUTH_NOT_EXPR isn't here the point at all. The underlying
>>>>> issue is that fold-const re-inttroduces TRUTH_AND/OR and co.
>>>>
>>>> I'm very sure it doesn't re-constrct TRUTH_ variants out of thin air
>>>> when you present it with BIT_ variants as input.
>>>
>>> Well, look into fold-const's fold_binary_loc function and see
>>>
>>> ?/* ARG0 is the first operand of EXPR, and ARG1 is the second operand.
>>>
>>> ? ? First check for cases where an arithmetic operation is applied to a
>>> ? ? compound, conditional, or comparison operation. ?Push the arithmetic
>>> ? ? operation inside the compound or conditional to see if any folding
>>> ? ? can then be done. ?Convert comparison to conditional for this purpose.
>>> ? ? The also optimizes non-constant cases that used to be done in
>>> ? ? expand_expr.
>>>
>>> ? ? Before we do that, see if this is a BIT_AND_EXPR or a BIT_IOR_EXPR,
>>> ? ? one of the operands is a comparison and the other is a comparison, a
>>> ? ? BIT_AND_EXPR with the constant 1, or a truth value. ?In that case, the
>>> ? ? code below would make the expression more complex. ?Change it to a
>>> ? ? TRUTH_{AND,OR}_EXPR. ?Likewise, convert a similar NE_EXPR to
>>> ? ? TRUTH_XOR_EXPR and an EQ_EXPR to the inversion of a TRUTH_XOR_EXPR. ?*/
>>>
>>> ?if ((code == BIT_AND_EXPR || code == BIT_IOR_EXPR
>>> ? ? ? || code == EQ_EXPR || code == NE_EXPR)
>>> ? ? ?&& ((truth_value_p (TREE_CODE (arg0))
>>> ? ? ? ? ? && (truth_value_p (TREE_CODE (arg1))
>>> ? ? ? ? ? ? ? || (TREE_CODE (arg1) == BIT_AND_EXPR
>>> ? ? ? ? ? ? ? ? ? && integer_onep (TREE_OPERAND (arg1, 1)))))
>>> ? ? ? ? ?|| (truth_value_p (TREE_CODE (arg1))
>>> ? ? ? ? ? ? ?&& (truth_value_p (TREE_CODE (arg0))
>>> ? ? ? ? ? ? ? ? ?|| (TREE_CODE (arg0) == BIT_AND_EXPR
>>> ? ? ? ? ? ? ? ? ? ? ?&& integer_onep (TREE_OPERAND (arg0, 1)))))))
>>> ? ?{
>>> ? ? ?tem = fold_build2_loc (loc, code == BIT_AND_EXPR ? TRUTH_AND_EXPR
>>> ? ? ? ? ? ? ? ? ? ? ? ? : code == BIT_IOR_EXPR ? TRUTH_OR_EXPR
>>> ? ? ? ? ? ? ? ? ? ? ? ? : TRUTH_XOR_EXPR,
>>> ? ? ? ? ? ? ? ? ? ? ? ? boolean_type_node,
>>> ? ? ? ? ? ? ? ? ? ? ? ? fold_convert_loc (loc, boolean_type_node, arg0),
>>> ? ? ? ? ? ? ? ? ? ? ? ? fold_convert_loc (loc, boolean_type_node, arg1));
>>>
>>> ? ? ?if (code == EQ_EXPR)
>>> ? ? ? ?tem = invert_truthvalue_loc (loc, tem);
>>>
>>> ? ? ?return fold_convert_loc (loc, type, tem);
>>> ? ?}
>>>
>>> Here unconditionally TRUTH_AND/TRUTH_OR gets introduced, if operands
>>> are of kind truth. ?This is btw the point, why you see that those
>>> cases are handled. ?But as soon as this part is turned off for BIT_-
>>> IOR/AND, we need to do the folding for 1-bit precision case explicit.
>>
>> First of all this checks for a quite complex pattern - where do we pass
>> such complex pattern from the gimple level to fold? ?For the EQ/NE_EXPR
>> case forwprop probably might be able to feed it that, but then how does
>> it go wrong? ?The above could also simply be guarded by !in_gimple_form.
>>
>> Richard.
>
> See reassoc pass as example
I don't see where.
> and this hacky maybe_fold_and_comparisons
> / maybe_fold_or_comparisons functions.
Yes, that case is known - but unless the folding result is sane we
drop it anyway.
> ?As indeed we want still be
> able to do comparison foldings without getting back an TRUTH-op.
Well, if we get back one we have canonicalize_cond_expr_cond to
make it sane again.
> Additionally we have a lot of passes - like vectorizer - which are
> happily try to build new condition on tree-level. ?This is another
That's true and you promised to clean up remaining TRUTH_*
consumers / producers when doing the canonicalization to BIT_*.
At the moment they get back to BIT_* via force_gimple_operand*
which is probably even fine.
> place I saw issues and tree-cfg failures. And last but not least those
> truth-ops might be reintroduced in gimple_fold, as soon as we see
> bitwise-ops on one-bit precision integral type as truth_value.
Well, I already said that one-bit-precision as truth_value is wrong.
Relying on fold-const.c here with its very strict type rules always
was a stretch. If it turns out there are now cases that we want to
do differently then we finally have to bite the bullet and do it.
There's always two options - duplicate it properly in fold_stmt
(or forwprop in case it needs to lookup defs), or remove it from fold
and make sure we fold things using fold_stmt at gimplifcation time.
Richard.
> Kai
>