Jonathan Wright via Gcc-patches <gcc-patches@gcc.gnu.org> writes:
> Hi,
>
> As subject, this patch rewrites the v[q]tbx 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. I see arm also reuses tbl1 for the scheduling type, which
makes sense. We should only add a separate type if something actually
needs it.
Richard
>
> Thanks,
> Jonathan
>
> ---
>
> gcc/ChangeLog:
>
> 2021-02-12 Jonathan Wright <jonathan.wri...@arm.com>
>
> * config/aarch64/aarch64-simd-builtins.def: Add tbx1 builtin
> generator macros.
> * config/aarch64/aarch64-simd.md (aarch64_tbx1<mode>):
> Define.
> * config/aarch64/arm_neon.h (vqtbx1_s8): USE RTL builtin
> instead of inline asm.
> (vqtbx1_u8): Likewise.
> (vqtbx1_p8): Likewise.
> (vqtbx1q_s8): Likewise.
> (vqtbx1q_u8): Likewise.
> (vqtbx1q_p8): Likewise.
> (vtbx2_s8): Likewise.
> (vtbx2_u8): Likewise.
> (vtbx2_p8): Likewise.
>
> diff --git a/gcc/config/aarch64/aarch64-simd-builtins.def
> b/gcc/config/aarch64/aarch64-simd-builtins.def
> index
> a5cfb6754456a1e8f0fca57c68b009a53e09789e..0f44ed84ff9d08d808b1b2dfe528db5208b134f5
> 100644
> --- a/gcc/config/aarch64/aarch64-simd-builtins.def
> +++ b/gcc/config/aarch64/aarch64-simd-builtins.def
> @@ -710,6 +710,10 @@
> VAR1 (BINOP, qtbl4, 0, NONE, v8qi)
> VAR1 (BINOP, qtbl4, 0, NONE, v16qi)
>
> + /* Implemented by aarch64_tbx1<mode>. */
> + VAR2 (TERNOP, tbx1, 0, NONE, v8qi, v16qi)
> + VAR2 (TERNOPU, tbx1, 0, NONE, v8qi, v16qi)
> +
> /* Implemented by aarch64_tbx4<mode>. */
> VAR1 (TERNOP, tbx4, 0, NONE, v8qi)
> VAR1 (TERNOP, tbx4, 0, NONE, v16qi)
> diff --git a/gcc/config/aarch64/aarch64-simd.md
> b/gcc/config/aarch64/aarch64-simd.md
> index
> 939c281d533261bb84dc451586da707953018fb8..5f701dd2775290156634ef8c6feccecd359e9ec9
> 100644
> --- a/gcc/config/aarch64/aarch64-simd.md
> +++ b/gcc/config/aarch64/aarch64-simd.md
> @@ -6905,6 +6905,17 @@
> [(set_attr "type" "neon_tbl1<q>")]
> )
>
> +(define_insn "aarch64_tbx1<mode>"
> + [(set (match_operand:VB 0 "register_operand" "=w")
> + (unspec:VB [(match_operand:VB 1 "register_operand" "0")
> + (match_operand:V16QI 2 "register_operand" "w")
> + (match_operand:VB 3 "register_operand" "w")]
> + UNSPEC_TBX))]
> + "TARGET_SIMD"
> + "tbx\\t%0.<Vtype>, {%2.16b}, %3.<Vtype>"
> + [(set_attr "type" "neon_tbl1<q>")]
> +)
> +
> ;; Two source registers.
>
> (define_insn "aarch64_tbl2v16qi"
> diff --git a/gcc/config/aarch64/arm_neon.h b/gcc/config/aarch64/arm_neon.h
> index
> 46f919fb254b98f887db4748d3b410b7d18e8a4e..1c48c166b5b9aaf052761f95121c26845221dae9
> 100644
> --- a/gcc/config/aarch64/arm_neon.h
> +++ b/gcc/config/aarch64/arm_neon.h
> @@ -9617,72 +9617,46 @@ __extension__ extern __inline int8x8_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vqtbx1_s8 (int8x8_t __r, int8x16_t __tab, uint8x8_t __idx)
> {
> - int8x8_t __result = __r;
> - __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
> - : "+w"(__result)
> - : "w"(__tab), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return __builtin_aarch64_tbx1v8qi (__r, __tab, (int8x8_t) __idx);
> }
>
> __extension__ extern __inline uint8x8_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vqtbx1_u8 (uint8x8_t __r, uint8x16_t __tab, uint8x8_t __idx)
> {
> - uint8x8_t __result = __r;
> - __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
> - : "+w"(__result)
> - : "w"(__tab), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return __builtin_aarch64_tbx1v8qi_uuuu (__r, __tab, __idx);
> }
>
> __extension__ extern __inline poly8x8_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vqtbx1_p8 (poly8x8_t __r, poly8x16_t __tab, uint8x8_t __idx)
> {
> - poly8x8_t __result = __r;
> - __asm__ ("tbx %0.8b,{%1.16b},%2.8b"
> - : "+w"(__result)
> - : "w"(__tab), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return (poly8x8_t) __builtin_aarch64_tbx1v8qi ((int8x8_t) __r,
> + (int8x16_t) __tab,
> + (int8x8_t) __idx);
> }
>
> __extension__ extern __inline int8x16_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vqtbx1q_s8 (int8x16_t __r, int8x16_t __tab, uint8x16_t __idx)
> {
> - int8x16_t __result = __r;
> - __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
> - : "+w"(__result)
> - : "w"(__tab), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return __builtin_aarch64_tbx1v16qi (__r, __tab, (int8x16_t) __idx);
> }
>
> __extension__ extern __inline uint8x16_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vqtbx1q_u8 (uint8x16_t __r, uint8x16_t __tab, uint8x16_t __idx)
> {
> - uint8x16_t __result = __r;
> - __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
> - : "+w"(__result)
> - : "w"(__tab), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return __builtin_aarch64_tbx1v16qi_uuuu (__r, __tab, __idx);
> }
>
> __extension__ extern __inline poly8x16_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vqtbx1q_p8 (poly8x16_t __r, poly8x16_t __tab, uint8x16_t __idx)
> {
> - poly8x16_t __result = __r;
> - __asm__ ("tbx %0.16b,{%1.16b},%2.16b"
> - : "+w"(__result)
> - : "w"(__tab), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return (poly8x16_t) __builtin_aarch64_tbx1v16qi ((int8x16_t) __r,
> + (int8x16_t) __tab,
> + (int8x16_t) __idx);
> }
>
> /* V7 legacy table intrinsics. */
> @@ -9846,39 +9820,26 @@ __extension__ extern __inline int8x8_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vtbx2_s8 (int8x8_t __r, int8x8x2_t __tab, int8x8_t __idx)
> {
> - int8x8_t __result = __r;
> int8x16_t __temp = vcombine_s8 (__tab.val[0], __tab.val[1]);
> - __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
> - : "+w"(__result)
> - : "w"(__temp), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return __builtin_aarch64_tbx1v8qi (__r, __temp, __idx);
> }
>
> __extension__ extern __inline uint8x8_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vtbx2_u8 (uint8x8_t __r, uint8x8x2_t __tab, uint8x8_t __idx)
> {
> - uint8x8_t __result = __r;
> uint8x16_t __temp = vcombine_u8 (__tab.val[0], __tab.val[1]);
> - __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
> - : "+w"(__result)
> - : "w"(__temp), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return __builtin_aarch64_tbx1v8qi_uuuu (__r, __temp, __idx);
> }
>
> __extension__ extern __inline poly8x8_t
> __attribute__ ((__always_inline__, __gnu_inline__, __artificial__))
> vtbx2_p8 (poly8x8_t __r, poly8x8x2_t __tab, uint8x8_t __idx)
> {
> - poly8x8_t __result = __r;
> poly8x16_t __temp = vcombine_p8 (__tab.val[0], __tab.val[1]);
> - __asm__ ("tbx %0.8b, {%1.16b}, %2.8b"
> - : "+w"(__result)
> - : "w"(__temp), "w"(__idx)
> - : /* No clobbers */);
> - return __result;
> + return (poly8x8_t) __builtin_aarch64_tbx1v8qi ((int8x8_t) __r,
> + (int8x16_t) __temp,
> + (int8x8_t) __idx);
> }
>
> /* End of temporary inline asm. */
