[Bug rtl-optimization/102840] [12 Regression] gcc.target/i386/pr22076.c by r12-4475

["roger at nextmovesoftware dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102840

Roger Sayle <roger at nextmovesoftware dot com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Last reconfirmed|                            |2021-10-19
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1

--- Comment #1 from Roger Sayle <roger at nextmovesoftware dot com> ---
I believe this test case is poorly written, and not correctly testing the
original issue in PR target/22076 which concerned suboptimal moving of
arguments via memory (fixed by prohibiting reload using mmx registers).

Prior to my patch, with -m32 -O2 -fomit-frame-pointer -mmmx -mno-sse2, GCC
generated:

test:   movq    .LC1, %mm0
        paddb   .LC0, %mm0
        movq    %mm0, x
        ret

.x:     .zero 8
.LC0:   .byte   1
        .byte   2
        .byte   3
        .byte   4
        .byte   5
        .byte   6
        .byte   7
        .byte   8
.LC1:   .byte   11
        .byte   22
        .byte   33
        .byte   44
        .byte   55
        .byte   66
        .byte   77
        .byte   88

which indeed doesn't use movl, and requires two movq.

After my patch, we now generate the much more efficient (dare I say optimal):
test:   movl    $807671820, %eax
        movl    $1616136252, %edx
        movl    %eax, x
        movl    %edx, x+4
        ret

which has evaluated the _mm_add_pi8 at compile-time, and effectively memsets x
to the correct value in the minimum possible number of cycles.  In fact,
failing to evaluate this at compile-time is a regression since v4.1 (according
to godbolt)

[p.s. I predict other platforms might also notice changes in their testsuites,
as the middle-end now generates more efficient instruction sequences].