public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH v2] x86: make better use of VBROADCASTSS / VPBROADCASTD
@ 2023-06-21  6:06 Jan Beulich
  2023-06-21  7:37 ` Hongtao Liu
  0 siblings, 1 reply; 6+ messages in thread
From: Jan Beulich @ 2023-06-21  6:06 UTC (permalink / raw)
  To: gcc-patches; +Cc: Kirill Yukhin, Hongtao Liu

... 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.

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
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")

^ permalink raw reply	[flat|nested] 6+ messages in thread

end of thread, other threads:[~2023-06-25  1:23 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-06-21  6:06 [PATCH v2] x86: make better use of VBROADCASTSS / VPBROADCASTD Jan Beulich
2023-06-21  7:37 ` Hongtao Liu
2023-06-21  7:44   ` Jan Beulich
2023-06-21 12:39     ` Jan Beulich
2023-06-25  1:17       ` Liu, Hongtao
2023-06-25  1:23         ` Hongtao Liu

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).