Changes for if-convert to recognize simple conditional reduction.

[Richard Biener]

On Mon, Apr 28, 2014 at 10:05 PM, Richard Henderson <rth@redhat.com> wrote:
> On 04/17/2014 06:09 AM, Yuri Rumyantsev wrote:
>> +  /* Build cond expression using COND and constant operand
>> +     of reduction rhs.  */
>> +  c = fold_build_cond_expr (TREE_TYPE (rhs1),
>> +                         unshare_expr (cond),
>> +                         swap? zero: op1,
>> +                         swap? op1: zero);
>
> Do we recognize somewhere the canonical value for the comparison is -1, and
> simplify this further?
>
> E.g.
>
>   if (A[i] != 0) num += 1;
>
>   _vec_cmp = (_vec_A != _vec_ZERO);
>   _vec_num -= _vec_cmp;
>
>
>   if (A[i] != 0) num += x;
>
>   _vec_cmp = (_vec_A != _vec_ZERO);
>   _vec_cmp *= _vec_x;
>   _vec_num -= _vec_cmp;

While the middle-end knows that vector comparisons result in
{0,0...} or {-1,-1,...} I doubt that anyone implemented the above
transform.

Note that it depends on the target on whether the transform
would be profitable - a target may only implement
a vector conditional move for example (the XOP vcond one).
So the correct place for the optimization would be the
expander or combine (unless we apply some canonicalization
rule on GIMPLE - but the back-transform is certainly more
complicated than optimally expanding vec_A != vec_ZERO ? 0 : 1)

Richard.

>
>
> r~