On Wed, Jun 21, 2023 at 2:06 PM Jan Beulich via Gcc-patches
<[email protected]> wrote:
>
> ... in vec_dupv4sf / *vec_dupv4si. The respective broadcast insns are
> never longer (yet sometimes shorter) than the corresponding VSHUFPS /
> VPSHUFD, due to the immediate operand of the shuffle insns balancing the
> possible need for VEX3 in the broadcast ones. When EVEX encoding is
> required the broadcast insns are always shorter.
>
> Add new alternatives to cover the AVX2 and AVX512 cases as appropriate.
>
> gcc/
>
> * config/i386/sse.md (vec_dupv4sf): Make first alternative use
> vbroadcastss for AVX2. New AVX512F alternative.
> (*vec_dupv4si): New AVX2 and AVX512F alternatives using
> vpbroadcastd.
> ---
> Especially with the added "enabled" attribute I didn't really see how to
> (further) fold alternatives 0 and 1. Instead *vec_dupv4si might benefit
> from using sse2_noavx2 instead of sse2 for alternative 2, except that
> there is no sse2_noavx2, only sse2_noavx.
>
> Is there a reason why vec_dupv4sf uses sseshuf1 for its shuffle
> alternatives, but *vec_dupv4si uses sselog1? I'd be happy to correct
> this in whichever is the appropriate direction, while touching this
> anyway.
It should be sseshuf1(or sseshuf depending on input operands number in
the pattern) for shufps, sselog means logical instructions.
I think it comes from
Intel SDM Vlolumes1:
5.6.1
Intel® SSE2 Packed and Scalar Double Precision Floating-Point Instructions
there're descriptions for different kinds of instructions.
>
> I'm working from the assumption that the isa attributes to the original
> 1st and 2nd alternatives don't need further restricting (to sse2_noavx2
> or avx_noavx2 as applicable), as the new earlier alternatives cover all
> operand forms already when at least AVX2 is enabled.
>
> Isn't prefix_extra use bogus here? What extra prefix does vbroadcastss
According to comments, yes, no extra prefix is needed.
;; There are also additional prefixes in 3DNOW, SSSE3.
;; ssemuladd,sse4arg default to 0f24/0f25 and DREX byte,
;; sseiadd1,ssecvt1 to 0f7a with no DREX byte.
;; 3DNOW has 0f0f prefix, SSSE3 and SSE4_{1,2} 0f38/0f3a.
> use? (Same further down in *vec_dupv4si and avx2_vbroadcasti128_<mode>
> and elsewhere.)
> ---
> v2: Correct operand constraints. Respect -mprefer-vector-width=. Fold
> two alternatives of vec_dupv4sf.
>
> --- a/gcc/config/i386/sse.md
> +++ b/gcc/config/i386/sse.md
> @@ -26141,41 +26141,64 @@
> (const_int 1)))])
>
> (define_insn "vec_dupv4sf"
> - [(set (match_operand:V4SF 0 "register_operand" "=v,v,x")
> + [(set (match_operand:V4SF 0 "register_operand" "=v,v,v,x")
> (vec_duplicate:V4SF
> - (match_operand:SF 1 "nonimmediate_operand" "Yv,m,0")))]
> + (match_operand:SF 1 "nonimmediate_operand" "Yv,v,m,0")))]
> "TARGET_SSE"
> "@
> - vshufps\t{$0, %1, %1, %0|%0, %1, %1, 0}
> + * return TARGET_AVX2 ? \"vbroadcastss\t{%1, %0|%0, %1}\" :
> \"vshufps\t{$0, %d1, %0|%0, %d1, 0}\";
> + vbroadcastss\t{%1, %g0|%g0, %1}
> vbroadcastss\t{%1, %0|%0, %1}
> shufps\t{$0, %0, %0|%0, %0, 0}"
> - [(set_attr "isa" "avx,avx,noavx")
> - (set_attr "type" "sseshuf1,ssemov,sseshuf1")
> - (set_attr "length_immediate" "1,0,1")
> - (set_attr "prefix_extra" "0,1,*")
> - (set_attr "prefix" "maybe_evex,maybe_evex,orig")
> - (set_attr "mode" "V4SF")])
> + [(set_attr "isa" "avx,*,avx,noavx")
> + (set (attr "type")
> + (cond [(and (eq_attr "alternative" "0")
> + (match_test "!TARGET_AVX2"))
> + (const_string "sseshuf1")
> + (eq_attr "alternative" "3")
> + (const_string "sseshuf1")
> + ]
> + (const_string "ssemov")))
> + (set (attr "length_immediate")
> + (if_then_else (eq_attr "type" "sseshuf1")
> + (const_string "1")
> + (const_string "0")))
> + (set_attr "prefix_extra" "0,0,1,*")
> + (set_attr "prefix" "maybe_evex,evex,maybe_evex,orig")
> + (set_attr "mode" "V4SF,V16SF,V4SF,V4SF")
> + (set (attr "enabled")
> + (if_then_else (eq_attr "alternative" "1")
> + (symbol_ref "TARGET_AVX512F && !TARGET_AVX512VL
> + && !TARGET_PREFER_AVX256")
> + (const_string "*")))])
>
> (define_insn "*vec_dupv4si"
> - [(set (match_operand:V4SI 0 "register_operand" "=v,v,x")
> + [(set (match_operand:V4SI 0 "register_operand" "=v,v,v,v,x")
> (vec_duplicate:V4SI
> - (match_operand:SI 1 "nonimmediate_operand" "Yv,m,0")))]
> + (match_operand:SI 1 "nonimmediate_operand" "Yvm,v,Yv,m,0")))]
> "TARGET_SSE"
> "@
> + vpbroadcastd\t{%1, %0|%0, %1}
> + vpbroadcastd\t{%1, %g0|%g0, %1}
> %vpshufd\t{$0, %1, %0|%0, %1, 0}
> vbroadcastss\t{%1, %0|%0, %1}
> shufps\t{$0, %0, %0|%0, %0, 0}"
> - [(set_attr "isa" "sse2,avx,noavx")
> - (set_attr "type" "sselog1,ssemov,sselog1")
> - (set_attr "length_immediate" "1,0,1")
> - (set_attr "prefix_extra" "0,1,*")
> - (set_attr "prefix" "maybe_vex,maybe_evex,orig")
> - (set_attr "mode" "TI,V4SF,V4SF")
> + [(set_attr "isa" "avx2,*,sse2,avx,noavx")
> + (set_attr "type" "ssemov,ssemov,sselog1,ssemov,sselog1")
> + (set_attr "length_immediate" "0,0,1,0,1")
> + (set_attr "prefix_extra" "0,0,0,1,*")
> + (set_attr "prefix" "maybe_evex,evex,maybe_vex,maybe_evex,orig")
> + (set_attr "mode" "TI,XI,TI,V4SF,V4SF")
> (set (attr "preferred_for_speed")
> - (cond [(eq_attr "alternative" "1")
> + (cond [(eq_attr "alternative" "3")
> (symbol_ref "!TARGET_INTER_UNIT_MOVES_TO_VEC")
> ]
> - (symbol_ref "true")))])
> + (symbol_ref "true")))
> + (set (attr "enabled")
> + (if_then_else (eq_attr "alternative" "1")
> + (symbol_ref "TARGET_AVX512F && !TARGET_AVX512VL
> + && !TARGET_PREFER_AVX256")
> + (const_string "*")))])
>
> (define_insn "*vec_dupv2di"
> [(set (match_operand:V2DI 0 "register_operand" "=x,v,v,v,x")
Could you write a testcase for the newly added constraint?
You can use an inline assembly to explicitly use an evex xmm register.
.i.e.
typedef float v4sf __attribute__((vector_size(16)));
typedef int v4si __attribute__((vector_size(16)));
v4sf
foo (float a)
{
register float c __asm("xmm16") = a;
asm volatile ("": "+v"(c));
return __extension__(v4sf) {c, c, c, c};
}
v4si
foo1 (int a)
{
register int c __asm("xmm16") = a;
asm volatile ("": "+v"(c));
return __extension__(v4si) {c, c, c, c};
}
with your patch, the above testcase should generate
vbroadcastss/vpbroadcastd %xmm16, %zmm0 with -mavx512f -mno-avx512vl
-mprefer-vector-width=512.
Refer to gcc/testsuite/gcc.target/i386/avx512bw-pack-2.c
Otherwise LGTM.
--
BR,
Hongtao
