[Bug target/106069] [12/13 Regression] wrong code with -O -fno-tree-forwprop -maltivec on ppc64le

["rguenther at suse dot de" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106069

--- Comment #22 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 3 Aug 2022, linkw at gcc dot gnu.org wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106069
> 
> --- Comment #21 from Kewen Lin <linkw at gcc dot gnu.org> ---
> I didn't look into this in details, but something in the culprit commit caught
> my eyes, take altivec_vmrghh as example:
> 
> Before the patch, the pattern
> 
>    [(set (match_operand:V8HI 0 "register_operand" "=v")
>          (vec_select:V8HI
>            (vec_concat:V16HI
>              (match_operand:V8HI 1 "register_operand" "v")
>              (match_operand:V8HI 2 "register_operand" "v"))
>            (parallel [(const_int 0) (const_int 8)
>                       (const_int 1) (const_int 9)
>                       (const_int 2) (const_int 10)
>                       (const_int 3) (const_int 11)])))]
> 
> can match vmrghh on BE while vmrglh on LE. It indicates this pattern has
> different semantic from underlying instruction perspectives.
> 
> After the patch, this pattern only matches vmrghh.
> 
> IMHO, this part has semantic change before and after the patch. The code before
> the patch looks more reasonable to me, since the pattern can have different
> meanings on BE and LE (underlying behavior).

Ideally we would avoid semantic difference of RTL depending on the target.
If that's not avoidable there should be target macros/hooks that specify
the desired semantics.  I assume the semantic difference is in
vec_concat behavior but that's just documented as

@findex vec_concat
@item (vec_concat:@var{m} @var{x1} @var{x2})
Describes a vector concat operation.  The result is a concatenation of the
vectors or scalars @var{x1} and @var{x2}; its length is the sum of the
lengths of the two inputs.

which is a bit unspecific.  To me it implies that
vec_select of a single lane N of the concat result can be distributed
to the operands of the vec_concat in the obvious way (if N >=
GET_MODE_NUNITS (x1) subtract GET_MODE_NUNITS and use x2)