OSDN Git Service

* config/i386/i386.c (struct_builtin_description)
[pf3gnuchains/gcc-fork.git] / gcc / config / i386 / sse.md
index e1f316b..d31aceb 100644 (file)
 (define_mode_attr ssevecsize [(V16QI "b") (V8HI "w") (V4SI "d") (V2DI "q")])
 
 ;; Mapping of the sse5 suffix
-(define_mode_attr ssemodesuffixf4 [(SF "ss") (DF "sd") (V4SF "ps") (V2DF "pd")])
-(define_mode_attr ssemodesuffixf2s [(SF "ss") (DF "sd") (V4SF "ss") (V2DF "sd")])
+(define_mode_attr ssemodesuffixf4 [(SF "ss") (DF "sd")
+                                  (V4SF "ps") (V2DF "pd")])
+(define_mode_attr ssemodesuffixf2s [(SF "ss") (DF "sd")
+                                   (V4SF "ss") (V2DF "sd")])
 (define_mode_attr ssemodesuffixf2c [(V4SF "s") (V2DF "d")])
 
 ;; Mapping of the max integer size for sse5 rotate immediate constraint
@@ -68,8 +70,8 @@
 ;; This is essential for maintaining stable calling conventions.
 
 (define_expand "mov<mode>"
-  [(set (match_operand:SSEMODEI 0 "nonimmediate_operand" "")
-       (match_operand:SSEMODEI 1 "nonimmediate_operand" ""))]
+  [(set (match_operand:SSEMODE 0 "nonimmediate_operand" "")
+       (match_operand:SSEMODE 1 "nonimmediate_operand" ""))]
   "TARGET_SSE"
 {
   ix86_expand_vector_move (<MODE>mode, operands);
@@ -77,8 +79,8 @@
 })
 
 (define_insn "*mov<mode>_internal"
-  [(set (match_operand:SSEMODEI 0 "nonimmediate_operand" "=x,x ,m")
-       (match_operand:SSEMODEI 1 "nonimmediate_or_sse_const_operand"  "C ,xm,x"))]
+  [(set (match_operand:SSEMODE 0 "nonimmediate_operand" "=x,x ,m")
+       (match_operand:SSEMODE 1 "nonimmediate_or_sse_const_operand"  "C ,xm,x"))]
   "TARGET_SSE
    && (register_operand (operands[0], <MODE>mode)
        || register_operand (operands[1], <MODE>mode))"
       return standard_sse_constant_opcode (insn, operands[1]);
     case 1:
     case 2:
-      if (get_attr_mode (insn) == MODE_V4SF)
-       return "movaps\t{%1, %0|%0, %1}";
-      else
-       return "movdqa\t{%1, %0|%0, %1}";
+      switch (get_attr_mode (insn))
+       {
+       case MODE_V4SF:
+         return "movaps\t{%1, %0|%0, %1}";
+       case MODE_V2DF:
+         return "movapd\t{%1, %0|%0, %1}";
+       default:
+         return "movdqa\t{%1, %0|%0, %1}";
+       }
     default:
       gcc_unreachable ();
     }
 }
   [(set_attr "type" "sselog1,ssemov,ssemov")
    (set (attr "mode")
-       (if_then_else
-         (ior (ior (ne (symbol_ref "optimize_size") (const_int 0))
-                   (eq (symbol_ref "TARGET_SSE2") (const_int 0)))
-              (and (eq_attr "alternative" "2")
-                   (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
-                       (const_int 0))))
-         (const_string "V4SF")
+       (cond [(ior (ior (ne (symbol_ref "optimize_size") (const_int 0))
+                        (eq (symbol_ref "TARGET_SSE2") (const_int 0)))
+                   (and (eq_attr "alternative" "2")
+                        (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
+                            (const_int 0))))
+                (const_string "V4SF")
+              (eq (const_string "<MODE>mode") (const_string "V4SFmode"))
+                (const_string "V4SF")
+              (eq (const_string "<MODE>mode") (const_string "V2DFmode"))
+                (const_string "V2DF")
+             ]
          (const_string "TI")))])
 
 ;; Move a DI from a 32-bit register pair (e.g. %edx:%eax) to an xmm.
       gcc_unreachable ();
 })
 
-(define_expand "mov<mode>"
-  [(set (match_operand:SSEMODEF2P 0 "nonimmediate_operand" "")
-       (match_operand:SSEMODEF2P 1 "nonimmediate_operand" ""))]
-  "TARGET_SSE"
-{
-  ix86_expand_vector_move (<MODE>mode, operands);
-  DONE;
-})
-
-(define_insn "*movv4sf_internal"
-  [(set (match_operand:V4SF 0 "nonimmediate_operand" "=x,x,m")
-       (match_operand:V4SF 1 "nonimmediate_or_sse_const_operand" "C,xm,x"))]
-  "TARGET_SSE
-   && (register_operand (operands[0], V4SFmode)
-       || register_operand (operands[1], V4SFmode))"
-{
-  switch (which_alternative)
-    {
-    case 0:
-      return standard_sse_constant_opcode (insn, operands[1]);
-    case 1:
-    case 2:
-      return "movaps\t{%1, %0|%0, %1}";
-    default:
-      gcc_unreachable ();
-    }
-}
-  [(set_attr "type" "sselog1,ssemov,ssemov")
-   (set_attr "mode" "V4SF")])
-
 (define_split
   [(set (match_operand:V4SF 0 "register_operand" "")
        (match_operand:V4SF 1 "zero_extended_scalar_load_operand" ""))]
   operands[2] = CONST0_RTX (V4SFmode);
 })
 
-(define_insn "*movv2df_internal"
-  [(set (match_operand:V2DF 0 "nonimmediate_operand" "=x,x,m")
-       (match_operand:V2DF 1 "nonimmediate_or_sse_const_operand" "C,xm,x"))]
-  "TARGET_SSE
-   && (register_operand (operands[0], V2DFmode)
-       || register_operand (operands[1], V2DFmode))"
-{
-  switch (which_alternative)
-    {
-    case 0:
-      return standard_sse_constant_opcode (insn, operands[1]);
-    case 1:
-    case 2:
-      if (get_attr_mode (insn) == MODE_V4SF)
-       return "movaps\t{%1, %0|%0, %1}";
-      else
-       return "movapd\t{%1, %0|%0, %1}";
-    default:
-      gcc_unreachable ();
-    }
-}
-  [(set_attr "type" "sselog1,ssemov,ssemov")
-   (set (attr "mode")
-       (if_then_else
-         (ior (ior (ne (symbol_ref "optimize_size") (const_int 0))
-                   (eq (symbol_ref "TARGET_SSE2") (const_int 0)))
-              (and (eq_attr "alternative" "2")
-                   (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
-                       (const_int 0))))
-         (const_string "V4SF")
-         (const_string "V2DF")))])
-
 (define_split
   [(set (match_operand:V2DF 0 "register_operand" "")
        (match_operand:V2DF 1 "zero_extended_scalar_load_operand" ""))]
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_expand "neg<mode>2"
-  [(set (match_operand:SSEMODEF2P 0 "register_operand" "")
-       (neg:SSEMODEF2P (match_operand:SSEMODEF2P 1 "register_operand" "")))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
-  "ix86_expand_fp_absneg_operator (NEG, <MODE>mode, operands); DONE;")
-
-(define_expand "abs<mode>2"
+(define_expand "<code><mode>2"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "")
-       (abs:SSEMODEF2P (match_operand:SSEMODEF2P 1 "register_operand" "")))]
+       (absneg:SSEMODEF2P
+         (match_operand:SSEMODEF2P 1 "register_operand" "")))]
   "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
-  "ix86_expand_fp_absneg_operator (ABS, <MODE>mode, operands); DONE;")
+  "ix86_expand_fp_absneg_operator (<CODE>, <MODE>mode, operands); DONE;")
 
-(define_expand "<addsub><mode>3"
+(define_expand "<plusminus_insn><mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "")
        (plusminus:SSEMODEF2P
          (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "")
   "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
   "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "*<addsub><mode>3"
+(define_insn "*<plusminus_insn><mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
        (plusminus:SSEMODEF2P
          (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "<comm>0")
          (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "xm")))]
   "SSE_VEC_FLOAT_MODE_P (<MODE>mode)
    && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
-  "<addsub>p<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
+  "<plusminus_mnemonic>p<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sseadd")
    (set_attr "mode" "<MODE>")])
 
-(define_insn "<sse>_vm<addsub><mode>3"
+(define_insn "<sse>_vm<plusminus_insn><mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
        (vec_merge:SSEMODEF2P
          (plusminus:SSEMODEF2P
            (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "xm"))
          (match_dup 1)
          (const_int 1)))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)
-   && ix86_binary_operator_ok (<CODE>, V4SFmode, operands)"
-  "<addsub>s<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
+  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
+  "<plusminus_mnemonic>s<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sseadd")
    (set_attr "mode" "<ssescalarmode>")])
 
            (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "xm"))
          (match_dup 1)
          (const_int 1)))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)
-   && ix86_binary_operator_ok (MULT, <MODE>mode, operands)"
+  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
   "muls<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
   [(set_attr "type" "ssemul")
    (set_attr "mode" "<ssescalarmode>")])
                  (match_operand:V4SF 2 "nonimmediate_operand" "")))]
   "TARGET_SSE"
 {
-  ix86_fixup_binary_operands_no_copy (DIV, V4SFmode, operands);
-
   if (TARGET_SSE_MATH && TARGET_RECIP && !optimize_size
       && flag_finite_math_only && !flag_trapping_math
       && flag_unsafe_math_optimizations)
        (div:V2DF (match_operand:V2DF 1 "register_operand" "")
                  (match_operand:V2DF 2 "nonimmediate_operand" "")))]
   "TARGET_SSE2"
-  "ix86_fixup_binary_operands_no_copy (DIV, V2DFmode, operands);")
+  "")
 
 (define_insn "<sse>_div<mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
   [(set_attr "type" "sseadd")
    (set_attr "mode" "V2DF")])
 
-(define_insn "sse3_h<addsub>v4sf3"
+(define_insn "sse3_h<plusminus_insn>v4sf3"
   [(set (match_operand:V4SF 0 "register_operand" "=x")
        (vec_concat:V4SF
          (vec_concat:V2SF
              (vec_select:SF (match_dup 2) (parallel [(const_int 2)]))
              (vec_select:SF (match_dup 2) (parallel [(const_int 3)]))))))]
   "TARGET_SSE3"
-  "h<addsub>ps\t{%2, %0|%0, %2}"
+  "h<plusminus_mnemonic>ps\t{%2, %0|%0, %2}"
   [(set_attr "type" "sseadd")
    (set_attr "prefix_rep" "1")
    (set_attr "mode" "V4SF")])
 
-(define_insn "sse3_h<addsub>v2df3"
+(define_insn "sse3_h<plusminus_insn>v2df3"
   [(set (match_operand:V2DF 0 "register_operand" "=x")
        (vec_concat:V2DF
          (plusminus:DF
              (parallel [(const_int 0)]))
            (vec_select:DF (match_dup 2) (parallel [(const_int 1)])))))]
   "TARGET_SSE3"
-  "h<addsub>pd\t{%2, %0|%0, %2}"
+  "h<plusminus_mnemonic>pd\t{%2, %0|%0, %2}"
   [(set_attr "type" "sseadd")
    (set_attr "mode" "V2DF")])
 
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_expand "and<mode>3"
-  [(set (match_operand:SSEMODEF2P 0 "register_operand" "")
-       (and:SSEMODEF2P
-         (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "")
-         (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "")))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
-  "ix86_fixup_binary_operands_no_copy (AND, <MODE>mode, operands);")
-
-(define_insn "*and<mode>3"
-  [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
-       (and:SSEMODEF2P
-         (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "xm")))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)
-   && ix86_binary_operator_ok (AND, <MODE>mode, operands)"
-  "andp<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "mode" "<MODE>")])
-
 (define_insn "<sse>_nand<mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
        (and:SSEMODEF2P
   [(set_attr "type" "sselog")
    (set_attr "mode" "<MODE>")])
 
-(define_expand "ior<mode>3"
-  [(set (match_operand:SSEMODEF2P 0 "register_operand" "")
-       (ior:SSEMODEF2P
-         (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "")
-         (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "")))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
-  "ix86_fixup_binary_operands_no_copy (IOR, <MODE>mode, operands);")
-
-(define_insn "*ior<mode>3"
-  [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
-       (ior:SSEMODEF2P
-         (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "xm")))]
-  "SSE_VEC_FLOAT_MODE_P (<MODE>mode)
-   && ix86_binary_operator_ok (IOR, <MODE>mode, operands)"
-  "orp<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "mode" "<MODE>")])
-
-(define_expand "xor<mode>3"
+(define_expand "<code><mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "")
-       (xor:SSEMODEF2P
+       (plogic:SSEMODEF2P
          (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "")
          (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "")))]
   "SSE_VEC_FLOAT_MODE_P (<MODE>mode)"
-  "ix86_fixup_binary_operands_no_copy (XOR, <MODE>mode, operands);")
+  "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "*xor<mode>3"
+(define_insn "*<code><mode>3"
   [(set (match_operand:SSEMODEF2P 0 "register_operand" "=x")
-       (xor:SSEMODEF2P
+       (plogic:SSEMODEF2P
          (match_operand:SSEMODEF2P 1 "nonimmediate_operand" "%0")
          (match_operand:SSEMODEF2P 2 "nonimmediate_operand" "xm")))]
   "SSE_VEC_FLOAT_MODE_P (<MODE>mode)
-   && ix86_binary_operator_ok (XOR, <MODE>mode, operands)"
-  "xorp<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "<plogicprefix>p<ssemodesuffixf2c>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog")
    (set_attr "mode" "<MODE>")])
 
 ;; allocation lossage.  These patterns do not allow memory operands
 ;; because the native instructions read the full 128-bits.
 
-(define_insn "*and<mode>3"
-  [(set (match_operand:MODEF 0 "register_operand" "=x")
-       (and:MODEF
-         (match_operand:MODEF 1 "register_operand" "0")
-         (match_operand:MODEF 2 "register_operand" "x")))]
-  "SSE_FLOAT_MODE_P (<MODE>mode)"
-  "andp<ssemodefsuffix>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "mode" "<ssevecmode>")])
-
 (define_insn "*nand<mode>3"
   [(set (match_operand:MODEF 0 "register_operand" "=x")
        (and:MODEF
   [(set_attr "type" "sselog")
    (set_attr "mode" "<ssevecmode>")])
 
-(define_insn "*ior<mode>3"
-  [(set (match_operand:MODEF 0 "register_operand" "=x")
-       (ior:MODEF
-         (match_operand:MODEF 1 "register_operand" "0")
-         (match_operand:MODEF 2 "register_operand" "x")))]
-  "SSE_FLOAT_MODE_P (<MODE>mode)"
-  "orp<ssemodefsuffix>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "mode" "<ssevecmode>")])
-
-(define_insn "*xor<mode>3"
+(define_insn "*<code><mode>3"
   [(set (match_operand:MODEF 0 "register_operand" "=x")
-       (xor:MODEF
+       (plogic:MODEF
          (match_operand:MODEF 1 "register_operand" "0")
          (match_operand:MODEF 2 "register_operand" "x")))]
   "SSE_FLOAT_MODE_P (<MODE>mode)"
-  "xorp<ssemodefsuffix>\t{%2, %0|%0, %2}"
+  "<plogicprefix>p<ssemodefsuffix>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog")
    (set_attr "mode" "<ssevecmode>")])
 
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
+(define_expand "sse_movhlps_exp"
+  [(set (match_operand:V4SF 0 "nonimmediate_operand" "")
+       (vec_select:V4SF
+         (vec_concat:V8SF
+           (match_operand:V4SF 1 "nonimmediate_operand" "")
+           (match_operand:V4SF 2 "nonimmediate_operand" ""))
+         (parallel [(const_int 6)
+                    (const_int 7)
+                    (const_int 2)
+                    (const_int 3)])))]
+  "TARGET_SSE"
+  "ix86_fixup_binary_operands (UNKNOWN, V4SFmode, operands);")
+
 (define_insn "sse_movhlps"
   [(set (match_operand:V4SF 0 "nonimmediate_operand"     "=x,x,m")
        (vec_select:V4SF
   [(set_attr "type" "ssemov")
    (set_attr "mode" "V4SF,V2SF,V2SF")])
 
+(define_expand "sse_movlhps_exp"
+  [(set (match_operand:V4SF 0 "nonimmediate_operand" "")
+       (vec_select:V4SF
+         (vec_concat:V8SF
+           (match_operand:V4SF 1 "nonimmediate_operand" "")
+           (match_operand:V4SF 2 "nonimmediate_operand" ""))
+         (parallel [(const_int 0)
+                    (const_int 1)
+                    (const_int 4)
+                    (const_int 5)])))]
+  "TARGET_SSE"
+  "ix86_fixup_binary_operands (UNKNOWN, V4SFmode, operands);")
+
 (define_insn "sse_movlhps"
   [(set (match_operand:V4SF 0 "nonimmediate_operand"     "=x,x,o")
        (vec_select:V4SF
   [(set_attr "type" "ssemov")
    (set_attr "mode" "V2SF,V4SF,V2SF")])
 
+(define_expand "sse_loadhps_exp"
+  [(set (match_operand:V4SF 0 "nonimmediate_operand" "")
+       (vec_concat:V4SF
+         (vec_select:V2SF
+           (match_operand:V4SF 1 "nonimmediate_operand" "")
+           (parallel [(const_int 0) (const_int 1)]))
+         (match_operand:V2SF 2 "nonimmediate_operand" "")))]
+  "TARGET_SSE"
+  "ix86_fixup_binary_operands (UNKNOWN, V4SFmode, operands);")
+
 (define_insn "sse_loadhps"
   [(set (match_operand:V4SF 0 "nonimmediate_operand" "=x,x,o")
        (vec_concat:V4SF
   [(set_attr "type" "ssemov")
    (set_attr "mode" "V2SF,V4SF,V2SF")])
 
+(define_expand "sse_loadlps_exp"
+  [(set (match_operand:V4SF 0 "nonimmediate_operand" "")
+       (vec_concat:V4SF
+         (match_operand:V2SF 2 "nonimmediate_operand" "")
+         (vec_select:V2SF
+           (match_operand:V4SF 1 "nonimmediate_operand" "")
+           (parallel [(const_int 2) (const_int 3)]))))]
+  "TARGET_SSE"
+  "ix86_fixup_binary_operands (UNKNOWN, V4SFmode, operands);")
+
 (define_insn "sse_loadlps"
   [(set (match_operand:V4SF 0 "nonimmediate_operand" "=x,x,m")
        (vec_concat:V4SF
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
+(define_expand "sse2_unpckhpd_exp"
+  [(set (match_operand:V2DF 0 "nonimmediate_operand" "")
+       (vec_select:V2DF
+         (vec_concat:V4DF
+           (match_operand:V2DF 1 "nonimmediate_operand" "")
+           (match_operand:V2DF 2 "nonimmediate_operand" ""))
+         (parallel [(const_int 1)
+                    (const_int 3)])))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands (UNKNOWN, V2DFmode, operands);")
+
 (define_insn "sse2_unpckhpd"
   [(set (match_operand:V2DF 0 "nonimmediate_operand"     "=x,x,m")
        (vec_select:V2DF
   DONE;
 })
 
+(define_expand "sse2_unpcklpd_exp"
+  [(set (match_operand:V2DF 0 "nonimmediate_operand" "")
+       (vec_select:V2DF
+         (vec_concat:V4DF
+           (match_operand:V2DF 1 "nonimmediate_operand" "")
+           (match_operand:V2DF 2 "nonimmediate_operand" ""))
+         (parallel [(const_int 0)
+                    (const_int 2)])))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands (UNKNOWN, V2DFmode, operands);")
+
 (define_insn "sse2_unpcklpd"
   [(set (match_operand:V2DF 0 "nonimmediate_operand"     "=x,x,o")
        (vec_select:V2DF
   DONE;
 })
 
+(define_expand "sse2_loadhpd_exp"
+  [(set (match_operand:V2DF 0 "nonimmediate_operand" "")
+       (vec_concat:V2DF
+         (vec_select:DF
+           (match_operand:V2DF 1 "nonimmediate_operand" "")
+           (parallel [(const_int 0)]))
+         (match_operand:DF 2 "nonimmediate_operand" "")))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands (UNKNOWN, V2DFmode, operands);")
+
 (define_insn "sse2_loadhpd"
   [(set (match_operand:V2DF 0 "nonimmediate_operand"     "=x,x,x,o")
        (vec_concat:V2DF
   operands[0] = adjust_address (operands[0], DFmode, 8);
 })
 
+(define_expand "sse2_loadlpd_exp"
+  [(set (match_operand:V2DF 0 "nonimmediate_operand" "")
+       (vec_concat:V2DF
+         (match_operand:DF 2 "nonimmediate_operand" "")
+         (vec_select:DF
+           (match_operand:V2DF 1 "nonimmediate_operand" "")
+           (parallel [(const_int 1)]))))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands (UNKNOWN, V2DFmode, operands);")
+
 (define_insn "sse2_loadlpd"
   [(set (match_operand:V2DF 0 "nonimmediate_operand"    "=x,x,x,x,x,m")
        (vec_concat:V2DF
   "TARGET_SSE2"
   "operands[2] = force_reg (<MODE>mode, CONST0_RTX (<MODE>mode));")
 
-(define_expand "add<mode>3"
+(define_expand "<plusminus_insn><mode>3"
   [(set (match_operand:SSEMODEI 0 "register_operand" "")
-       (plus:SSEMODEI (match_operand:SSEMODEI 1 "nonimmediate_operand" "")
-                      (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
+       (plusminus:SSEMODEI
+         (match_operand:SSEMODEI 1 "nonimmediate_operand" "")
+         (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
   "TARGET_SSE2"
-  "ix86_fixup_binary_operands_no_copy (PLUS, <MODE>mode, operands);")
-
-(define_insn "*add<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-       (plus:SSEMODEI
-         (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
-  "padd<ssevecsize>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sseiadd")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
-(define_insn "sse2_ssadd<mode>3"
-  [(set (match_operand:SSEMODE12 0 "register_operand" "=x")
-       (ss_plus:SSEMODE12
-         (match_operand:SSEMODE12 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODE12 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (SS_PLUS, <MODE>mode, operands)"
-  "padds<ssevecsize>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sseiadd")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
-(define_insn "sse2_usadd<mode>3"
-  [(set (match_operand:SSEMODE12 0 "register_operand" "=x")
-       (us_plus:SSEMODE12
-         (match_operand:SSEMODE12 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODE12 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (US_PLUS, <MODE>mode, operands)"
-  "paddus<ssevecsize>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sseiadd")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
-(define_expand "sub<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "")
-       (minus:SSEMODEI (match_operand:SSEMODEI 1 "register_operand" "")
-                       (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
-  "TARGET_SSE2"
-  "ix86_fixup_binary_operands_no_copy (MINUS, <MODE>mode, operands);")
+  "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "*sub<mode>3"
+(define_insn "*<plusminus_insn><mode>3"
   [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-       (minus:SSEMODEI
-         (match_operand:SSEMODEI 1 "register_operand" "0")
+       (plusminus:SSEMODEI
+         (match_operand:SSEMODEI 1 "nonimmediate_operand" "<comm>0")
          (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2"
-  "psub<ssevecsize>\t{%2, %0|%0, %2}"
+  "TARGET_SSE2 && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "p<plusminus_mnemonic><ssevecsize>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sseiadd")
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "sse2_sssub<mode>3"
-  [(set (match_operand:SSEMODE12 0 "register_operand" "=x")
-       (ss_minus:SSEMODE12
-         (match_operand:SSEMODE12 1 "register_operand" "0")
-         (match_operand:SSEMODE12 2 "nonimmediate_operand" "xm")))]
+(define_expand "sse2_<plusminus_insn><mode>3"
+  [(set (match_operand:SSEMODE12 0 "register_operand" "")
+       (sat_plusminus:SSEMODE12
+         (match_operand:SSEMODE12 1 "nonimmediate_operand" "")
+         (match_operand:SSEMODE12 2 "nonimmediate_operand" "")))]
   "TARGET_SSE2"
-  "psubs<ssevecsize>\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sseiadd")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
+  "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "sse2_ussub<mode>3"
+(define_insn "*sse2_<plusminus_insn><mode>3"
   [(set (match_operand:SSEMODE12 0 "register_operand" "=x")
-       (us_minus:SSEMODE12
-         (match_operand:SSEMODE12 1 "register_operand" "0")
+       (sat_plusminus:SSEMODE12
+         (match_operand:SSEMODE12 1 "nonimmediate_operand" "<comm>0")
          (match_operand:SSEMODE12 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2"
-  "psubus<ssevecsize>\t{%2, %0|%0, %2}"
+  "TARGET_SSE2 && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "p<plusminus_mnemonic><ssevecsize>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sseiadd")
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "sse2_umulv2siv2di3"
+(define_expand "sse2_umulv2siv2di3"
+  [(set (match_operand:V2DI 0 "register_operand" "")
+       (mult:V2DI
+         (zero_extend:V2DI
+           (vec_select:V2SI
+             (match_operand:V4SI 1 "nonimmediate_operand" "")
+             (parallel [(const_int 0) (const_int 2)])))
+         (zero_extend:V2DI
+           (vec_select:V2SI
+             (match_operand:V4SI 2 "nonimmediate_operand" "")
+             (parallel [(const_int 0) (const_int 2)])))))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands_no_copy (MULT, V4SImode, operands);")
+
+(define_insn "*sse2_umulv2siv2di3"
   [(set (match_operand:V2DI 0 "register_operand" "=x")
        (mult:V2DI
          (zero_extend:V2DI
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "sse4_1_mulv2siv2di3"
+(define_expand "sse4_1_mulv2siv2di3"
+  [(set (match_operand:V2DI 0 "register_operand" "")
+       (mult:V2DI
+         (sign_extend:V2DI
+           (vec_select:V2SI
+             (match_operand:V4SI 1 "nonimmediate_operand" "")
+             (parallel [(const_int 0) (const_int 2)])))
+         (sign_extend:V2DI
+           (vec_select:V2SI
+             (match_operand:V4SI 2 "nonimmediate_operand" "")
+             (parallel [(const_int 0) (const_int 2)])))))]
+  "TARGET_SSE4_1"
+  "ix86_fixup_binary_operands_no_copy (MULT, V4SImode, operands);")
+(define_insn "*sse4_1_mulv2siv2di3"
   [(set (match_operand:V2DI 0 "register_operand" "=x")
        (mult:V2DI
          (sign_extend:V2DI
    (set_attr "prefix_extra" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "sse2_pmaddwd"
+(define_expand "sse2_pmaddwd"
+  [(set (match_operand:V4SI 0 "register_operand" "")
+       (plus:V4SI
+         (mult:V4SI
+           (sign_extend:V4SI
+             (vec_select:V4HI
+               (match_operand:V8HI 1 "nonimmediate_operand" "")
+               (parallel [(const_int 0)
+                          (const_int 2)
+                          (const_int 4)
+                          (const_int 6)])))
+           (sign_extend:V4SI
+             (vec_select:V4HI
+               (match_operand:V8HI 2 "nonimmediate_operand" "")
+               (parallel [(const_int 0)
+                          (const_int 2)
+                          (const_int 4)
+                          (const_int 6)]))))
+         (mult:V4SI
+           (sign_extend:V4SI
+             (vec_select:V4HI (match_dup 1)
+               (parallel [(const_int 1)
+                          (const_int 3)
+                          (const_int 5)
+                          (const_int 7)])))
+           (sign_extend:V4SI
+             (vec_select:V4HI (match_dup 2)
+               (parallel [(const_int 1)
+                          (const_int 3)
+                          (const_int 5)
+                          (const_int 7)]))))))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands_no_copy (MULT, V8HImode, operands);")
+
+(define_insn "*sse2_pmaddwd"
   [(set (match_operand:V4SI 0 "register_operand" "=x")
        (plus:V4SI
          (mult:V4SI
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn "sse2_eq<mode>3"
+(define_expand "sse2_eq<mode>3"
+  [(set (match_operand:SSEMODE124 0 "register_operand" "")
+       (eq:SSEMODE124
+         (match_operand:SSEMODE124 1 "nonimmediate_operand" "")
+         (match_operand:SSEMODE124 2 "nonimmediate_operand" "")))]
+  "TARGET_SSE2 && !TARGET_SSE5"
+  "ix86_fixup_binary_operands_no_copy (EQ, <MODE>mode, operands);")
+
+(define_insn "*sse2_eq<mode>3"
   [(set (match_operand:SSEMODE124 0 "register_operand" "=x")
        (eq:SSEMODE124
          (match_operand:SSEMODE124 1 "nonimmediate_operand" "%0")
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "sse4_1_eqv2di3"
+(define_expand "sse4_1_eqv2di3"
+  [(set (match_operand:V2DI 0 "register_operand" "")
+       (eq:V2DI
+         (match_operand:V2DI 1 "nonimmediate_operand" "")
+         (match_operand:V2DI 2 "nonimmediate_operand" "")))]
+  "TARGET_SSE4_1"
+  "ix86_fixup_binary_operands_no_copy (EQ, V2DImode, operands);")
+
+(define_insn "*sse4_1_eqv2di3"
   [(set (match_operand:V2DI 0 "register_operand" "=x")
        (eq:V2DI
          (match_operand:V2DI 1 "nonimmediate_operand" "%0")
 (define_insn "sse4_2_gtv2di3"
   [(set (match_operand:V2DI 0 "register_operand" "=x")
        (gt:V2DI
-         (match_operand:V2DI 1 "nonimmediate_operand" "0")
+         (match_operand:V2DI 1 "register_operand" "0")
          (match_operand:V2DI 2 "nonimmediate_operand" "xm")))]
   "TARGET_SSE4_2"
   "pcmpgtq\t{%2, %0|%0, %2}"
   operands[2] = force_reg (<MODE>mode, gen_rtx_CONST_VECTOR (<MODE>mode, v));
 })
 
-(define_expand "and<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "")
-       (and:SSEMODEI (match_operand:SSEMODEI 1 "nonimmediate_operand" "")
-                     (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
-  "TARGET_SSE"
-  "ix86_fixup_binary_operands_no_copy (AND, <MODE>mode, operands);")
-
-(define_insn "*sse_and<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-        (and:SSEMODEI
-          (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
-          (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "(TARGET_SSE && !TARGET_SSE2)
-   && ix86_binary_operator_ok (AND, <MODE>mode, operands)"
-  "andps\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "mode" "V4SF")])
-
-(define_insn "*sse2_and<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-       (and:SSEMODEI
-         (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (AND, <MODE>mode, operands)"
-  "pand\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
 (define_insn "*sse_nand<mode>3"
   [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
        (and:SSEMODEI
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_expand "andtf3"
-  [(set (match_operand:TF 0 "register_operand" "")
-       (and:TF (match_operand:TF 1 "nonimmediate_operand" "")
-               (match_operand:TF 2 "nonimmediate_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_fixup_binary_operands_no_copy (AND, TFmode, operands);")
-
-(define_insn "*andtf3"
-  [(set (match_operand:TF 0 "register_operand" "=x")
-       (and:TF
-         (match_operand:TF 1 "nonimmediate_operand" "%0")
-         (match_operand:TF 2 "nonimmediate_operand" "xm")))]
-  "TARGET_64BIT && ix86_binary_operator_ok (AND, TFmode, operands)"
-  "pand\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
 (define_insn "*nandtf3"
   [(set (match_operand:TF 0 "register_operand" "=x")
        (and:TF
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_expand "ior<mode>3"
+(define_expand "<code><mode>3"
   [(set (match_operand:SSEMODEI 0 "register_operand" "")
-       (ior:SSEMODEI (match_operand:SSEMODEI 1 "nonimmediate_operand" "")
-                     (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
+       (plogic:SSEMODEI
+         (match_operand:SSEMODEI 1 "nonimmediate_operand" "")
+         (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
   "TARGET_SSE"
-  "ix86_fixup_binary_operands_no_copy (IOR, <MODE>mode, operands);")
+  "ix86_fixup_binary_operands_no_copy (<CODE>, <MODE>mode, operands);")
 
-(define_insn "*sse_ior<mode>3"
+(define_insn "*sse_<code><mode>3"
   [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-        (ior:SSEMODEI
+        (plogic:SSEMODEI
           (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
           (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
   "(TARGET_SSE && !TARGET_SSE2)
-   && ix86_binary_operator_ok (IOR, <MODE>mode, operands)"
-  "orps\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "mode" "V4SF")])
-
-(define_insn "*sse2_ior<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-       (ior:SSEMODEI
-         (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (IOR, <MODE>mode, operands)"
-  "por\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
-(define_expand "iortf3"
-  [(set (match_operand:TF 0 "register_operand" "")
-       (ior:TF (match_operand:TF 1 "nonimmediate_operand" "")
-               (match_operand:TF 2 "nonimmediate_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_fixup_binary_operands_no_copy (IOR, TFmode, operands);")
-
-(define_insn "*iortf3"
-  [(set (match_operand:TF 0 "register_operand" "=x")
-       (ior:TF
-         (match_operand:TF 1 "nonimmediate_operand" "%0")
-         (match_operand:TF 2 "nonimmediate_operand" "xm")))]
-  "TARGET_64BIT && ix86_binary_operator_ok (IOR, TFmode, operands)"
-  "por\t{%2, %0|%0, %2}"
-  [(set_attr "type" "sselog")
-   (set_attr "prefix_data16" "1")
-   (set_attr "mode" "TI")])
-
-(define_expand "xor<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "")
-       (xor:SSEMODEI (match_operand:SSEMODEI 1 "nonimmediate_operand" "")
-                     (match_operand:SSEMODEI 2 "nonimmediate_operand" "")))]
-  "TARGET_SSE"
-  "ix86_fixup_binary_operands_no_copy (XOR, <MODE>mode, operands);")
-
-(define_insn "*sse_xor<mode>3"
-  [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-       (xor:SSEMODEI
-         (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
-         (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "(TARGET_SSE && !TARGET_SSE2)
-   && ix86_binary_operator_ok (XOR, <MODE>mode, operands)"
-  "xorps\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "<plogicprefix>ps\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog")
    (set_attr "mode" "V4SF")])
 
-(define_insn "*sse2_xor<mode>3"
+(define_insn "*sse2_<code><mode>3"
   [(set (match_operand:SSEMODEI 0 "register_operand" "=x")
-       (xor:SSEMODEI
+       (plogic:SSEMODEI
          (match_operand:SSEMODEI 1 "nonimmediate_operand" "%0")
          (match_operand:SSEMODEI 2 "nonimmediate_operand" "xm")))]
-  "TARGET_SSE2 && ix86_binary_operator_ok (XOR, <MODE>mode, operands)"
-  "pxor\t{%2, %0|%0, %2}"
+  "TARGET_SSE2 && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "p<plogicprefix>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog")
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_expand "xortf3"
+(define_expand "<code>tf3"
   [(set (match_operand:TF 0 "register_operand" "")
-       (xor:TF (match_operand:TF 1 "nonimmediate_operand" "")
-               (match_operand:TF 2 "nonimmediate_operand" "")))]
+       (plogic:TF
+         (match_operand:TF 1 "nonimmediate_operand" "")
+         (match_operand:TF 2 "nonimmediate_operand" "")))]
   "TARGET_64BIT"
-  "ix86_fixup_binary_operands_no_copy (XOR, TFmode, operands);")
+  "ix86_fixup_binary_operands_no_copy (<CODE>, TFmode, operands);")
 
-(define_insn "*xortf3"
+(define_insn "*<code>tf3"
   [(set (match_operand:TF 0 "register_operand" "=x")
-       (xor:TF
+       (plogic:TF
          (match_operand:TF 1 "nonimmediate_operand" "%0")
          (match_operand:TF 2 "nonimmediate_operand" "xm")))]
-  "TARGET_64BIT && ix86_binary_operator_ok (XOR, TFmode, operands)"
-  "pxor\t{%2, %0|%0, %2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, TFmode, operands)"
+  "p<plogicprefix>\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog")
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 ;;
 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(define_insn "sse2_uavgv16qi3"
+(define_expand "sse2_uavgv16qi3"
+  [(set (match_operand:V16QI 0 "register_operand" "")
+       (truncate:V16QI
+         (lshiftrt:V16HI
+           (plus:V16HI
+             (plus:V16HI
+               (zero_extend:V16HI
+                 (match_operand:V16QI 1 "nonimmediate_operand" ""))
+               (zero_extend:V16HI
+                 (match_operand:V16QI 2 "nonimmediate_operand" "")))
+             (const_vector:V16QI [(const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)
+                                  (const_int 1) (const_int 1)]))
+           (const_int 1))))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands_no_copy (PLUS, V16QImode, operands);")
+
+(define_insn "*sse2_uavgv16qi3"
   [(set (match_operand:V16QI 0 "register_operand" "=x")
        (truncate:V16QI
          (lshiftrt:V16HI
    (set_attr "prefix_data16" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "sse2_uavgv8hi3"
+(define_expand "sse2_uavgv8hi3"
+  [(set (match_operand:V8HI 0 "register_operand" "")
+       (truncate:V8HI
+         (lshiftrt:V8SI
+           (plus:V8SI
+             (plus:V8SI
+               (zero_extend:V8SI
+                 (match_operand:V8HI 1 "nonimmediate_operand" ""))
+               (zero_extend:V8SI
+                 (match_operand:V8HI 2 "nonimmediate_operand" "")))
+             (const_vector:V8HI [(const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)]))
+           (const_int 1))))]
+  "TARGET_SSE2"
+  "ix86_fixup_binary_operands_no_copy (PLUS, V8HImode, operands);")
+
+(define_insn "*sse2_uavgv8hi3"
   [(set (match_operand:V8HI 0 "register_operand" "=x")
        (truncate:V8HI
          (lshiftrt:V8SI
    (set_attr "prefix_extra" "1")
    (set_attr "mode" "DI")])
 
-(define_insn "ssse3_pmaddubswv8hi3"
+(define_insn "ssse3_pmaddubsw128"
   [(set (match_operand:V8HI 0 "register_operand" "=x")
        (ss_plus:V8HI
          (mult:V8HI
            (zero_extend:V8HI
              (vec_select:V4QI
-               (match_operand:V16QI 1 "nonimmediate_operand" "0")
+               (match_operand:V16QI 1 "register_operand" "0")
                (parallel [(const_int 0)
                           (const_int 2)
                           (const_int 4)
    (set_attr "prefix_extra" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "ssse3_pmaddubswv4hi3"
+(define_insn "ssse3_pmaddubsw"
   [(set (match_operand:V4HI 0 "register_operand" "=y")
        (ss_plus:V4HI
          (mult:V4HI
            (zero_extend:V4HI
              (vec_select:V4QI
-               (match_operand:V8QI 1 "nonimmediate_operand" "0")
+               (match_operand:V8QI 1 "register_operand" "0")
                (parallel [(const_int 0)
                           (const_int 2)
                           (const_int 4)
    (set_attr "prefix_extra" "1")
    (set_attr "mode" "DI")])
 
-(define_insn "ssse3_pmulhrswv8hi3"
+(define_expand "ssse3_pmulhrswv8hi3"
+  [(set (match_operand:V8HI 0 "register_operand" "")
+       (truncate:V8HI
+         (lshiftrt:V8SI
+           (plus:V8SI
+             (lshiftrt:V8SI
+               (mult:V8SI
+                 (sign_extend:V8SI
+                   (match_operand:V8HI 1 "nonimmediate_operand" ""))
+                 (sign_extend:V8SI
+                   (match_operand:V8HI 2 "nonimmediate_operand" "")))
+               (const_int 14))
+             (const_vector:V8HI [(const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)]))
+           (const_int 1))))]
+  "TARGET_SSSE3"
+  "ix86_fixup_binary_operands_no_copy (MULT, V8HImode, operands);")
+
+(define_insn "*ssse3_pmulhrswv8hi3"
   [(set (match_operand:V8HI 0 "register_operand" "=x")
        (truncate:V8HI
          (lshiftrt:V8SI
    (set_attr "prefix_extra" "1")
    (set_attr "mode" "TI")])
 
-(define_insn "ssse3_pmulhrswv4hi3"
+(define_expand "ssse3_pmulhrswv4hi3"
+  [(set (match_operand:V4HI 0 "register_operand" "")
+       (truncate:V4HI
+         (lshiftrt:V4SI
+           (plus:V4SI
+             (lshiftrt:V4SI
+               (mult:V4SI
+                 (sign_extend:V4SI
+                   (match_operand:V4HI 1 "nonimmediate_operand" ""))
+                 (sign_extend:V4SI
+                   (match_operand:V4HI 2 "nonimmediate_operand" "")))
+               (const_int 14))
+             (const_vector:V4HI [(const_int 1) (const_int 1)
+                                 (const_int 1) (const_int 1)]))
+           (const_int 1))))]
+  "TARGET_SSSE3"
+  "ix86_fixup_binary_operands_no_copy (MULT, V4HImode, operands);")
+
+(define_insn "*ssse3_pmulhrswv4hi3"
   [(set (match_operand:V4HI 0 "register_operand" "=y")
        (truncate:V4HI
          (lshiftrt:V4SI
 }
   [(set_attr "type" "ssecmp")
    (set_attr "mode" "TI")])
+
+(define_insn "aesenc"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "register_operand" "0")
+                      (match_operand:V2DI 2 "nonimmediate_operand" "xm")]
+                     UNSPEC_AESENC))]
+  "TARGET_AES"
+  "aesenc\t{%2, %0|%0, %2}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])
+
+(define_insn "aesenclast"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "register_operand" "0")
+                      (match_operand:V2DI 2 "nonimmediate_operand" "xm")]
+                     UNSPEC_AESENCLAST))]
+  "TARGET_AES"
+  "aesenclast\t{%2, %0|%0, %2}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])
+
+(define_insn "aesdec"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "register_operand" "0")
+                      (match_operand:V2DI 2 "nonimmediate_operand" "xm")]
+                     UNSPEC_AESDEC))]
+  "TARGET_AES"
+  "aesdec\t{%2, %0|%0, %2}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])
+
+(define_insn "aesdeclast"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "register_operand" "0")
+                      (match_operand:V2DI 2 "nonimmediate_operand" "xm")]
+                     UNSPEC_AESDECLAST))]
+  "TARGET_AES"
+  "aesdeclast\t{%2, %0|%0, %2}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])
+
+(define_insn "aesimc"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "nonimmediate_operand" "xm")]
+                     UNSPEC_AESIMC))]
+  "TARGET_AES"
+  "aesimc\t{%1, %0|%0, %1}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])
+
+(define_insn "aeskeygenassist"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "nonimmediate_operand" "xm")
+                     (match_operand:SI 2 "const_0_to_255_operand" "n")]
+                    UNSPEC_AESKEYGENASSIST))]
+  "TARGET_AES"
+  "aeskeygenassist\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])
+
+(define_insn "pclmulqdq"
+  [(set (match_operand:V2DI 0 "register_operand" "=x")
+       (unspec:V2DI [(match_operand:V2DI 1 "register_operand" "0")
+                     (match_operand:V2DI 2 "nonimmediate_operand" "xm")
+                     (match_operand:SI 3 "const_0_to_255_operand" "n")]
+                    UNSPEC_PCLMUL))]
+  "TARGET_PCLMUL"
+  "pclmulqdq\t{%3, %2, %0|%0, %2, %3}"
+  [(set_attr "type" "sselog1")
+   (set_attr "prefix_extra" "1")
+   (set_attr "mode" "TI")])