[PATCH, i386]: Improve V2SI extractions

[Uros Bizjak]

Attached patch adds missing SSE4.1 xmm->reg/mem extractions for V2SI modes.

2019-08-01  Uroš Bizjak  <ubizjak@gmail.com>

    * config/i386/mmx.md (vec_extractv2si_0): Add (r,x) alternative.
    (*vec_extractv2si_0_zext_sse4): New insn pattern.
    (*vec_extractv2si_0_zext): Ditto.
    (*vec_extractv2si_1): Add (rm,x) alternative.
    (*vec_extractv2si_1_zext): New insn pattern.
    (*vec_extractv2si_zext_mem): Add "TARGET_MMX || TARGET_MMX_WITH_SSE"
    insn constraint.

Bootstrapped and regression tested on x86_64-linux-gnu {,-m32}.

Committed to mainline SVN.

Uros.
-------------- next part --------------
Index: config/i386/mmx.md
===================================================================
--- config/i386/mmx.md	(revision 273969)
+++ config/i386/mmx.md	(working copy)
@@ -1620,9 +1620,9 @@
 ;; Avoid combining registers from different units in a single alternative,
 ;; see comment above inline_secondary_memory_needed function in i386.c
 (define_insn_and_split "*vec_extractv2si_0"
-  [(set (match_operand:SI 0 "nonimmediate_operand"     "=x,m,y, m,r")
+  [(set (match_operand:SI 0 "nonimmediate_operand"     "=x,m,y, m,r,r")
 	(vec_select:SI
-	  (match_operand:V2SI 1 "nonimmediate_operand" "xm,x,ym,y,m")
+	  (match_operand:V2SI 1 "nonimmediate_operand" "xm,x,ym,y,m,x")
 	  (parallel [(const_int 0)])))]
   "(TARGET_MMX || TARGET_MMX_WITH_SSE)
    && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
@@ -1630,33 +1630,76 @@
   "&& reload_completed"
   [(set (match_dup 0) (match_dup 1))]
   "operands[1] = gen_lowpart (SImode, operands[1]);"
-  [(set_attr "mmx_isa" "*,*,native,native,*")])
+  [(set_attr "isa" "*,*,*,*,*,sse2")
+   (set_attr "mmx_isa" "*,*,native,native,*,*")
+   (set (attr "preferred_for_speed")
+     (cond [(eq_attr "alternative" "5")
+	      (symbol_ref "TARGET_INTER_UNIT_MOVES_FROM_VEC")
+	   ]
+	   (symbol_ref "true")))])
 
+(define_insn "*vec_extractv2si_0_zext_sse4"
+  [(set (match_operand:DI 0 "register_operand" "=r,x")
+	(zero_extend:DI
+	  (vec_select:SI
+	    (match_operand:V2SI 1 "register_operand" "x,x")
+	    (parallel [(const_int 0)]))))]
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE) && TARGET_SSE4_1"
+  "#"
+  [(set_attr "isa" "x64,*")
+   (set (attr "preferred_for_speed")
+     (cond [(eq_attr "alternative" "0")
+	      (symbol_ref "TARGET_INTER_UNIT_MOVES_FROM_VEC")
+	   ]
+	   (symbol_ref "true")))])
+
+(define_insn "*vec_extractv2si_0_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+	(zero_extend:DI
+	  (vec_select:SI
+	    (match_operand:V2SI 1 "register_operand" "x")
+	    (parallel [(const_int 0)]))))]
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE)
+   && TARGET_64BIT && TARGET_SSE2 && TARGET_INTER_UNIT_MOVES_FROM_VEC"
+  "#")
+
+(define_split
+  [(set (match_operand:DI 0 "register_operand")
+	(zero_extend:DI
+	  (vec_select:SI
+	    (match_operand:V2SI 1 "register_operand")
+	    (parallel [(const_int 0)]))))]
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE)
+   && TARGET_SSE2 && reload_completed"
+  [(set (match_dup 0) (zero_extend:DI (match_dup 1)))]
+  "operands[1] = gen_lowpart (SImode, operands[1]);")
+
 ;; Avoid combining registers from different units in a single alternative,
 ;; see comment above inline_secondary_memory_needed function in i386.c
 (define_insn "*vec_extractv2si_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand"     "=y,x,x,y,x,r")
+  [(set (match_operand:SI 0 "nonimmediate_operand"     "=y,rm,x,x,y,x,r")
 	(vec_select:SI
-	  (match_operand:V2SI 1 "nonimmediate_operand" " 0,x,x,o,o,o")
+	  (match_operand:V2SI 1 "nonimmediate_operand" " 0,x ,x,x,o,o,o")
 	  (parallel [(const_int 1)])))]
   "(TARGET_MMX || TARGET_MMX_WITH_SSE)
    && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
   "@
    punpckhdq\t%0, %0
+   %vpextrd\t{$1, %1, %0|%0, %1, 1}
    %vpshufd\t{$0xe5, %1, %0|%0, %1, 0xe5}
    shufps\t{$0xe5, %1, %0|%0, %1, 0xe5}
    #
    #
    #"
-  [(set_attr "isa" "*,sse2,noavx,*,*,*")
-   (set_attr "mmx_isa" "native,*,*,native,*,*")
-   (set_attr "type" "mmxcvt,sseshuf1,sseshuf1,mmxmov,ssemov,imov")
+  [(set_attr "isa" "*,sse4,sse2,noavx,*,*,*")
+   (set_attr "mmx_isa" "native,*,*,*,native,*,*")
+   (set_attr "type" "mmxcvt,ssemov,sseshuf1,sseshuf1,mmxmov,ssemov,imov")
    (set (attr "length_immediate")
-     (if_then_else (eq_attr "alternative" "1,2")
+     (if_then_else (eq_attr "alternative" "1,2,3")
 		   (const_string "1")
 		   (const_string "*")))
-   (set_attr "prefix" "orig,maybe_vex,orig,orig,orig,orig")
-   (set_attr "mode" "DI,TI,V4SF,SI,SI,SI")])
+   (set_attr "prefix" "orig,maybe_vex,maybe_vex,orig,orig,orig,orig")
+   (set_attr "mode" "DI,TI,TI,V4SF,SI,SI,SI")])
 
 (define_split
   [(set (match_operand:SI 0 "register_operand")
@@ -1667,6 +1710,21 @@
   [(set (match_dup 0) (match_dup 1))]
   "operands[1] = adjust_address (operands[1], SImode, 4);")
 
+(define_insn "*vec_extractv2si_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+	(zero_extend:DI
+	  (vec_select:SI
+	    (match_operand:V2SI 1 "register_operand" "x")
+	    (parallel [(const_int 1)]))))]
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE)
+   && TARGET_64BIT && TARGET_SSE4_1"
+  "%vpextrd\t{$1, %1, %k0|%k0, %1, 1}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "length_immediate" "1")
+   (set_attr "prefix" "maybe_vex")
+   (set_attr "mode" "TI")])
+
 (define_insn_and_split "*vec_extractv2si_zext_mem"
   [(set (match_operand:DI 0 "register_operand" "=y,x,r")
 	(zero_extend:DI
@@ -1673,7 +1731,7 @@
 	  (vec_select:SI
 	    (match_operand:V2SI 1 "memory_operand" "o,o,o")
 	    (parallel [(match_operand:SI 2 "const_0_to_1_operand")]))))]
-  "TARGET_64BIT"
+  "(TARGET_MMX || TARGET_MMX_WITH_SSE) && TARGET_64BIT"
   "#"
   "&& reload_completed"
   [(set (match_dup 0) (zero_extend:DI (match_dup 1)))]