Jonathan Wright via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> Hi,
>
> As subject, this patch rewrites the [su]paddl[q] Neon intrinsics to use
> RTL builtins rather than inline assembly code, allowing for better
> scheduling and optimization.
>
> Regression tested and bootstrapped on aarch64-none-linux-gnu - no
> issues.
>
> Ok for master?
OK, thanks. For the record…
> __extension__ extern __inline uint64x1_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vpaddl_u32 (uint32x2_t __a)
> {
> - uint64x1_t __result;
> - __asm__ ("uaddlp %0.1d,%1.2s"
> - : "=w"(__result)
> - : "w"(__a)
> - : /* No clobbers */);
> - return __result;
> + return (uint64x1_t) __builtin_aarch64_uaddlpv2si_uu (__a);
> }
…I wasn't sure for this whether it would be better to use (uint64x1_t) {…}
instead of a scalar-to-vector conversion, since that seems to be the more
common style in the rest of arm_neon.h. But there are already instances
of this kind of conversion too, and if anything it should be more
efficient than creating a distinct vector object.
Richard
