Re: C++ delayed folding branch review

[Jason Merrill <jason@redhat.com>]

On 08/27/2015 09:38 AM, Kai Tietz wrote:
> 2015-08-27 15:27 GMT+02:00 Jason Merrill <jason@redhat.com>:
>> On 08/27/2015 06:39 AM, Kai Tietz wrote:
>>>
>>> 2015-08-27 4:56 GMT+02:00 Jason Merrill <jason@redhat.com>:
>>>>
>>>> On 08/24/2015 03:15 AM, Kai Tietz wrote:
>>>>>
>>>>> 2015-08-03 17:39 GMT+02:00 Jason Merrill <jason@redhat.com>:
>>>>>>
>>>>>> On 08/03/2015 05:42 AM, Kai Tietz wrote:
>>>>>>>
>>>>>>> 2015-08-03 5:49 GMT+02:00 Jason Merrill <jason@redhat.com>:
>>>>>>>>
>>>>>>>> On 07/31/2015 05:54 PM, Kai Tietz wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> The "STRIP_NOPS-requirement in 'reduced_constant_expression_p'" I
>>>>>>>>> could
>>>>>>>>> remove, but for one case in constexpr.  Without folding we don't do
>>>>>>>>> type-sinking/raising.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Right.
>>>>>>>>
>>>>>>>>> So binary/unary operations might be containing cast, which were in
>>>>>>>>> the
>>>>>>>>> past unexpected.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> Why aren't the casts folded away?
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On such cast constructs, as for this vector-sample, we can't fold away
>>>>>>
>>>>>>
>>>>>>
>>>>>> Which testcase is this?
>>>>>
>>>>>
>>>>>
>>>>> It is the g++.dg/ext/vector20.C testcase.  IIRC I mentioned this
>>>>> testcase already earlier as reference, but I might be wrong here.
>>>>
>>>>
>>>>
>>>> I don't see any casts in that testcase.  So the compiler is introducing
>>>> introducing conversions back and forth between const and non-const, then?
>>>> I
>>>> suppose it doesn't so much matter where they come from, they should be
>>>> folded away regardless.
>>>
>>>
>>> The cast gets introduced in convert.c about line 836 in function
>>> convert_to_integer_1 AFAIK.  There should be the alternative solution
>>> for this issue by disallowing for PLUS/MINUS/... expressions the
>>> sinking of the cast into the expression, if dofold is false, and type
>>> has same width as inner_type, and is of vector-kind.
>>
>>
>> Why would we be calling convert_to_integer for conversions between vector
>> types?
>>
>>>>>>> the cast chain.  The difference here to none-delayed-folding branch is
>>>>>>> that the cast isn't moved out of the plus-expr.  What we see now is
>>>>>>> (plus ((vec) (const vector ...) { .... }), ...).  Before we had (vec)
>>>>>>> (plus (const vector ...) { ... }).
>>>>>>
>>>>>>
>>>>>>
>>>>>> How could a PLUS_EXPR be considered a reduced constant, regardless of
>>>>>> where
>>>>>> the cast is?
>>>>>
>>>>>
>>>>>
>>>>> Of course it is just possible to sink out a cast from PLUS_EXPR, in
>>>>> pretty few circumstance (eg. on constants if both types just differ in
>>>>> const-attribute, if conversion is no view-convert).
>>>>
>>>>
>>>>
>>>> I don't understand how this is an answer to my question.
>>>
>>>
>>> (vec) (const vector) { ... } expression can't be folded.
>>
>>
>> It currently isn't folded, but why can't we change that?
>>
>>> This cast to
>>> none-const variant happens due the 'constexpr v = v +
>>> <constant-value>' pattern in testcase.  v is still of type vec, even
>>> if function itself is constexpr.
>>
>>
>> I don't see that pattern in the testcase:
>>
>> typedef long vec __attribute__((vector_size (2 * sizeof (long))));
>> constexpr vec v = { 3, 4 };
>> constexpr vec s = v + v;
>> constexpr vec w = __builtin_shuffle (v, v);
>>
>> If we have v + constant-value, that's because we pulled out the constant
>> value of one of the v's, which we ought to be doing for both of them.
>>
>>>>>>>>> On verify_constant we check by reduced_constant_expression_p, if
>>>>>>>>> value
>>>>>>>>> is
>>>>>>>>> a constant.  We don't handle here, that NOP_EXPRs are something we
>>>>>>>>> want to
>>>>>>>>> look through here, as it doesn't change anything if this is a
>>>>>>>>> constant, or
>>>>>>>>> not.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> NOPs around constants should have been folded away by the time we get
>>>>>>>> there.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> Not in this cases, as the we actually have here a switch from const to
>>>>>>> none-const.  So there is an attribute-change, which we can't ignore in
>>>>>>> general.
>>>>>>
>>>>>>
>>>>>>
>>>>>> I wasn't suggesting we ignore it, we should be able to change the type
>>>>>> of
>>>>>> the vector_cst.
>>>>>
>>>>>
>>>>>
>>>>> Well, the vector_cst we can change type, but this wouldn't help
>>>>> AFAICS.  As there is still one cast surviving within PLUS_EXPR for the
>>>>> other operand.
>>>>
>>>>
>>>>
>>>> Isn't the other operand also constant?  In constexpr evaluation, either
>>>> we're dealing with a bunch of constants, in which case we should be
>>>> folding
>>>> things fully, including conversions between const and non-const, or we
>>>> don't
>>>> care.
>>>
>>>
>>> No other operand isn't a constant-value.  See code-pattern in
>>> testcase.  It is of type 'vec', which isn't constant (well, 'v' is,
>>> but constexpr doesn't know about it).
>>
>>
>> What do you mean, "constexpr doesn't know about it"?
>>
>>>>> So the way to solve it would be to move such conversion out of the
>>>>> expression.  For integer-scalars we do this, and for some
>>>>> floating-points too.  So it might be something we don't handle for
>>>>> operations with vector-type.
>>>>
>>>>
>>>>
>>>> We don't need to worry about that in constexpr evaluation, since we only
>>>> care about constant operands.
>>>
>>>
>>> Sure, but the variable 'v' is the problem, not a constant-value itself.
>>
>>
>>>>>>> But I agree that for constexpr's we could special case cast
>>>>>>> from const to none-const (as required in expressions like const vec v
>>>>>>> = v + 1).
>>>>>>
>>>>>>
>>>>>>
>>>>>> Right.  But really this should happen in convert.c, it shouldn't be
>>>>>> specific
>>>>>> to C++.
>>>>>
>>>>>
>>>>>
>>>>> Hmm, maybe.  But isn't one of our different goals to move such
>>>>> implicit code-modification to match.pd instead?
>>>>
>>>>
>>>> Folding const into a constant is hardly code modification.  But perhaps
>>>> it
>>>> should go into fold_unary_loc:VIEW_CONVERT_EXPR rather than into
>>>> convert.c.
>>>
>>>
>>> Hmm, it isn't related to a view-convert.  So moving it into
>>> fold_unary_loc wouldn't solve here anything.  Issue is in constexpr
>>> code, not in folding itself.
>>
>>
>> What TREE_CODE does the conversion (vec) (const vector) { ... } use?
>
> The tree code is a NOP_EXPR.

That's probably a bug, seeing as fold_convert and convert_to_vector use 
VIEW_CONVERT_EXPR.

Jason