Re: [PATCH] Allow non-overflow ops in vect_is_simple_reduction_1

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

On Fri, Jul 24, 2015 at 4:39 PM, Tom de Vries <Tom_deVries@mentor.com> wrote:
> Hi,
>
> this patch allows parallelization and vectorization of reduction operators
> that are guaranteed to not overflow (such as min and max operators),
> independent of the overflow behaviour of the type.
>
> Bootstrapped and reg-tested on x86_64.
>
> OK for trunk?

Hmm, I don't like that no_overflow_tree_code function.  We have a much more
clear understanding which codes may overflow or trap.  Thus please add
a operation specific variant of TYPE_OVERFLOW_{TRAPS,WRAPS,UNDEFINED} like

bool
operation_overflow_traps (tree type, enum tree_code code)
{
  if (!ANY_INTEGRAL_TYPE_P (type)
     || !TYPE_OVERFLOW_TRAPS (type))
    return false;
  switch (code)
    {
    case PLUS_EXPR:
    case MINUS_EXPR:
    case MULT_EXPR:
    case LSHIFT_EXPR:
       /* Can overflow in various ways */
    case TRUNC_DIV_EXPR:
    case EXACT_DIV_EXPR:
    case FLOOR_DIV_EXPR:
    case CEIL_DIV_EXPR:
       /* For INT_MIN / -1 */
    case NEGATE_EXPR:
    case ABS_EXPR:
       /* For -INT_MIN */
       return true;
    default:
       return false;
   }
}

and similar variants for _wraps and _undefined.  I think we decided at
some point
the compiler should not take advantage of the fact that lshift or
*_div have undefined
behavior on signed integer overflow, similar we only take advantage of
integral-type
overflow behavior, not vector or complex.  So we could reduce the
number of cases
the functions return true if we document that it returns true only for
the cases where
the compiler needs to / may assume wrapping behavior does not take place.
As for _traps for example we only have optabs and libfuncs for
plus,minus,mult,negate
and abs.

Thanks,
Richard.

> Thanks,
> - Tom