[Bug tree-optimization/64946] [AArch64] gcc.target/aarch64/vect-abs-compile.c - "abs" vectorization fails for char/short types

[rguenther at suse dot de]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64946

--- Comment #18 from rguenther at suse dot de <rguenther at suse dot de> ---
On Mon, 6 Jun 2016, shiva0217 at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64946
> 
> --- Comment #17 from Shiva Chen <shiva0217 at gmail dot com> ---
> Hi, Richard
> 
> Thanks for the explanation :)
> 
> So the transformation (short)abs((int)short_var) -> abs (short_var)
> 
> should guard by TYPE_OVERFLOW_WRAPS 
> 
> because when TYPE_OVERFLOW_WRAPS is true, signed operation could wrap
> around.(ABS_EXPR in gimple could wrap around and rtl abs already modulo)

it _does_, not _could_ wrap around.

> 
> Therefore, the transformation is valid when TYPE_OVERFLOW_WRAPS is true.

Yes.

> It seems the last update of Matthew's patch in
> https://gcc.gnu.org/ml/gcc-patches/2016-01/msg00433.html
> still make sense.
> 
> Why would it be dropped, or there're still something we should consider ?

I don't see any "update" of the patch and the patch directly linked
is wrong.

> If we implement ABSU_EXPR, when should transfer 
> ABS_EXPR (x) -> (type of x) ABSU_EXPR (x) ?

When it makes sense to us to avoid introducing undefinedness into our IL
without losing the advantage of the undefinedness of ABS on INT_MIN.

> Could we define like if (!TYPE_OVERFLOW_WRAPS) then transfer ABS_EXPR (x) ->
> (type of x) ABSU_EXPR (x) in match.pd ?

If we do that unconditionally then we can as well simply say we always
treat ABS_EXPR as having defined behavior on overflow.  I think it's
useful to value-range analysis that we can assert that abs(x) >= 0
and thus disregard the special-case of x == INT_MIN which means we
do not want to lose that information in exchange for nothing.

> How to expand (type of x) ABSU_EXPR (x) to rtl ?
>
> Do we have to add a new naming pattern absu ?

As on RTL 'abs' has well-defined behavior on wrapping we can just use 
that.