https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66112
--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to ktkachov from comment #12)
> r223115 caused a code quality regression that I'm seeing on aarch64 with PR
> 68381.
>
> We have
> int
> foo (unsigned short x, unsigned short y)
> {
> int r;
> if (__builtin_mul_overflow (x, y, &r))
> __builtin_abort ();
> return r;
> }
>
> and before the commit we used to generate:
> foo:
> uxth x0, w0
> uxth x1, w1
> mul x0, x0, x1
> cmp x0, x0, sxtw
> bne .L9
> ret
> .L9:
> stp x29, x30, [sp, -16]!
> add x29, sp, 0
> bl abort
>
> but after r223115 at -O3 on an aarch64 toolchain we generate:
> foo:
> uxth w0, w0
> uxth w1, w1
> stp x29, x30, [sp, -16]!
> umull x0, w0, w1
> add x29, sp, 0
> tbnz w0, #31, .L6
> mov w2, 0
> cbnz w2, .L6
> ldp x29, x30, [sp], 16
> ret
> .L6:
> bl abort
You really should consider adding {u,}{add,sub,mul}v4<mode> and/or negv3<mode>
expanders if you have some better sequences you can generate over the generic
ones on some target.
