[Bug target/51980] ARM - Neon code polluted by useless stores to the stack with vuzpq / vzipq / vtrnq

["ktkachov at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51980

ktkachov at gcc dot gnu.org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ktkachov at gcc dot gnu.org

--- Comment #8 from ktkachov at gcc dot gnu.org ---
> arm-none-eabi-gcc -march=armv7-a -mfpu=neon -mfloat-abi=softfp -O2 -mthumb:
> sqrlen4D_16u8:
>         vmov    d18, r0, r1  @ v16qi
>         vmov    d19, r2, r3
>         vld1.64 {d16-d17}, [sp:64]
>         vabd.u8 q8, q9, q8
>         vmull.u8        q9, d16, d16
>         vmull.u8        q8, d17, d17
>         vuzp.32 q9, q8
>         vpaddl.u16      q9, q9
>         vmov    q10, q9  @ v4si
>         vpadal.u16      q10, q8
>         vmov    r0, r1, d20  @ v4si
>         vmov    r2, r3, d21
>         bx      lr

With current trunk I'm getting for the softfp case:

        push    {lr}    @ 40    *push_multi     [length = 2]
        vmov    d16, r0, r1  @ v16qi    @ 37    *neon_movv16qi/6        [length
= 8]
        vmov    d17, r2, r3
        add     lr, sp, #4      @ 36    *arm_addsi3/5   [length = 4]
        vldr    d18, [sp, #4]   @ 3     *neon_movv16qi/4        [length = 8]
        vldr    d19, [sp, #12]
        vabd.u8 q9, q8, q9      @ 7     neon_vabdv16qi  [length = 4]
        vmull.u8        q8, d18, d18    @ 14    neon_vmullv8qi  [length = 4]
        vmull.u8        q9, d19, d19    @ 16    neon_vmullv8qi  [length = 4]
        vuzp.32 q8, q9  @ 18    *neon_vuzpv4si_insn     [length = 4]
        vpaddl.u16      q8, q8  @ 22    neon_vpaddlv8hi [length = 4]
        vpadal.u16      q8, q9  @ 28    neon_vpadalv8hi [length = 4]
        vmov    r0, r1, d16  @ v4si     @ 39    *neon_movv4si/5 [length = 8]
        vmov    r2, r3, d17
        ldr     pc, [sp], #4    @ 45    *ldr_with_return        [length = 4]


The move between the vpad*s is gone, but there's a couple of redundant loads
and some register spillage.