OSDN Git Service

* config/i386/i386.md (x86_shrd): Add athlon_decode and
[pf3gnuchains/gcc-fork.git] / gcc / config / i386 / i386.md
index 47c585c..1e06ca0 100644 (file)
 \f
 ;; Processor type.
 (define_attr "cpu" "none,pentium,pentiumpro,geode,k6,athlon,k8,core2,atom,
-                   generic64,amdfam10,bdver1"
+                   generic64,amdfam10"
   (const (symbol_ref "ix86_schedule")))
 
 ;; A basic instruction type.  Refinements due to arguments to be
    && !x86_64_immediate_operand (operands[1], DImode) && 1"
   [(set (match_dup 0) (match_dup 1))
    (set (match_dup 2) (match_dup 3))]
-{
-  split_di (&operands[1], 1, &operands[2], &operands[3]);
-
-  operands[1] = gen_lowpart (DImode, operands[2]);
-  operands[2] = gen_rtx_MEM (SImode, gen_rtx_PLUS (DImode, stack_pointer_rtx,
-                                                  GEN_INT (4)));
-})
+  "split_di (&operands[1], 1, &operands[2], &operands[3]);
+   operands[1] = gen_lowpart (DImode, operands[2]);
+   operands[2] = gen_rtx_MEM (SImode, gen_rtx_PLUS (DImode, stack_pointer_rtx,
+                                                   GEN_INT (4)));
+  ")
 
 (define_split
   [(set (match_operand:DI 0 "push_operand" "")
    && !x86_64_immediate_operand (operands[1], DImode)"
   [(set (match_dup 0) (match_dup 1))
    (set (match_dup 2) (match_dup 3))]
-{
-  split_di (&operands[1], 1, &operands[2], &operands[3]);
-
-  operands[1] = gen_lowpart (DImode, operands[2]);
-  operands[2] = gen_rtx_MEM (SImode, gen_rtx_PLUS (DImode, stack_pointer_rtx,
-                                                  GEN_INT (4)));
-})
+  "split_di (&operands[1], 1, &operands[2], &operands[3]);
+   operands[1] = gen_lowpart (DImode, operands[2]);
+   operands[2] = gen_rtx_MEM (SImode, gen_rtx_PLUS (DImode, stack_pointer_rtx,
+                                                   GEN_INT (4)));
+  ")
 
 (define_insn "*pushdi2_prologue_rex64"
   [(set (match_operand:DI 0 "push_operand" "=<")
        case MODE_V4SF:
          return "%vxorps\t%0, %d0";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vxorpd\t%0, %d0";
+         return "%vxorpd\t%0, %d0";
        case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vpxor\t%0, %d0";
+         return "%vpxor\t%0, %d0";
        default:
          gcc_unreachable ();
        }
        case MODE_V4SF:
          return "%vmovaps\t{%1, %0|%0, %1}";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovapd\t{%1, %0|%0, %1}";
+         return "%vmovapd\t{%1, %0|%0, %1}";
        case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovdqa\t{%1, %0|%0, %1}";
+         return "%vmovdqa\t{%1, %0|%0, %1}";
        case MODE_DI:
          return "%vmovq\t{%1, %0|%0, %1}";
        case MODE_DF:
        case MODE_V4SF:
          return "%vxorps\t%0, %d0";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vxorpd\t%0, %d0";
+         return "%vxorpd\t%0, %d0";
        case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vpxor\t%0, %d0";
+         return "%vpxor\t%0, %d0";
        default:
          gcc_unreachable ();
        }
        case MODE_V4SF:
          return "%vmovaps\t{%1, %0|%0, %1}";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovapd\t{%1, %0|%0, %1}";
+         return "%vmovapd\t{%1, %0|%0, %1}";
        case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovdqa\t{%1, %0|%0, %1}";
+         return "%vmovdqa\t{%1, %0|%0, %1}";
        case MODE_DI:
          return "%vmovq\t{%1, %0|%0, %1}";
        case MODE_DF:
        case MODE_V4SF:
          return "xorps\t%0, %0";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "xorps\t%0, %0";
-         else
-           return "xorpd\t%0, %0";
+         return "xorpd\t%0, %0";
        case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "xorps\t%0, %0";
-         else
-           return "pxor\t%0, %0";
+         return "pxor\t%0, %0";
        default:
          gcc_unreachable ();
        }
        case MODE_V4SF:
          return "movaps\t{%1, %0|%0, %1}";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "movaps\t{%1, %0|%0, %1}";
-         else
-           return "movapd\t{%1, %0|%0, %1}";
+         return "movapd\t{%1, %0|%0, %1}";
        case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "movaps\t{%1, %0|%0, %1}";
-         else
-           return "movdqa\t{%1, %0|%0, %1}";
+         return "movdqa\t{%1, %0|%0, %1}";
        case MODE_DI:
          return "movq\t{%1, %0|%0, %1}";
        case MODE_DF:
   ""
   "ix86_expand_binary_operator (ASHIFT, <MODE>mode, operands); DONE;")
 
-(define_insn_and_split "*ashl<mode>3_doubleword"
+(define_insn "*ashl<mode>3_doubleword"
   [(set (match_operand:DWI 0 "register_operand" "=&r,r")
        (ashift:DWI (match_operand:DWI 1 "reg_or_pm1_operand" "n,0")
                    (match_operand:QI 2 "nonmemory_operand" "<S>c,<S>c")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (ASHIFT, <MODE>mode, operands)"
+  ""
   "#"
+  [(set_attr "type" "multi")])
+
+(define_split
+  [(set (match_operand:DWI 0 "register_operand" "")
+       (ashift:DWI (match_operand:DWI 1 "nonmemory_operand" "")
+                   (match_operand:QI 2 "nonmemory_operand" "")))
+   (clobber (reg:CC FLAGS_REG))]
   "(optimize && flag_peephole2) ? epilogue_completed : reload_completed"
   [(const_int 0)]
-  "ix86_split_ashl (operands, NULL_RTX, <MODE>mode); DONE;"
-  [(set_attr "type" "multi")])
+  "ix86_split_ashl (operands, NULL_RTX, <MODE>mode); DONE;")
 
 ;; By default we don't ask for a scratch register, because when DWImode
 ;; values are manipulated, registers are already at a premium.  But if
       gcc_assert (rtx_equal_p (operands[0], operands[1]));
       return "add{<imodesuffix>}\t%0, %0";
 
-    case TYPE_LEA:
-      return "#";
-
     default:
-      if (REG_P (operands[2]))
-       return "sal{<imodesuffix>}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
        return "sal{<imodesuffix>}\t%0";
       else
        return "sal{<imodesuffix>}\t{%2, %0|%0, %2}";
       gcc_assert (operands[2] == const1_rtx);
       return "add{l}\t%k0, %k0";
 
-    case TYPE_LEA:
-      return "#";
-
     default:
-      if (REG_P (operands[2]))
-       return "sal{l}\t{%b2, %k0|%k0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
        return "sal{l}\t%k0";
       else
        return "sal{l}\t{%2, %k0|%k0, %2}";
       return "add{w}\t%0, %0";
 
     default:
-      if (REG_P (operands[2]))
-       return "sal{w}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
        return "sal{w}\t%0";
       else
        return "sal{w}\t{%2, %0|%0, %2}";
     {
     case TYPE_LEA:
       return "#";
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      return "add{w}\t%0, %0";
 
     case TYPE_ALU:
       gcc_assert (operands[2] == const1_rtx);
       return "add{w}\t%0, %0";
 
     default:
-      if (REG_P (operands[2]))
-       return "sal{w}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
        return "sal{w}\t%0";
       else
        return "sal{w}\t{%2, %0|%0, %2}";
         return "add{b}\t%0, %0";
 
     default:
-      if (REG_P (operands[2]))
-       {
-         if (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t{%b2, %k0|%k0, %b2}";
-         else
-           return "sal{b}\t{%b2, %0|%0, %b2}";
-       }
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
        {
          if (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t%0";
+           return "sal{l}\t%k0";
          else
            return "sal{b}\t%0";
        }
     {
     case TYPE_LEA:
       return "#";
+
     case TYPE_ALU:
       gcc_assert (operands[2] == const1_rtx);
       if (REG_P (operands[1]) && !ANY_QI_REG_P (operands[1]))
         return "add{b}\t%0, %0";
 
     default:
-      if (REG_P (operands[2]))
-       {
-         if (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t{%b2, %k0|%k0, %b2}";
-         else
-           return "sal{b}\t{%b2, %0|%0, %b2}";
-       }
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
        {
          if (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t%0";
+           return "sal{l}\t%k0";
          else
            return "sal{b}\t%0";
        }
        (const_string "*")))
    (set_attr "mode" "QI,SI,SI")])
 
+(define_insn "*ashlqi3_1_slp"
+  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
+       (ashift:QI (match_dup 0)
+                  (match_operand:QI 1 "nonmemory_operand" "cI")))
+   (clobber (reg:CC FLAGS_REG))]
+  "(optimize_function_for_size_p (cfun)
+    || !TARGET_PARTIAL_FLAG_REG_STALL
+    || (operands[1] == const1_rtx
+       && (TARGET_SHIFT1
+           || (TARGET_DOUBLE_WITH_ADD && REG_P (operands[0])))))"
+{
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ALU:
+      gcc_assert (operands[1] == const1_rtx);
+      return "add{b}\t%0, %0";
+
+    default:
+      if (operands[1] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "sal{b}\t%0";
+      else
+       return "sal{b}\t{%1, %0|%0, %1}";
+    }
+}
+  [(set (attr "type")
+     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
+                         (const_int 0))
+                     (match_operand 0 "register_operand" ""))
+                (match_operand 1 "const1_operand" ""))
+             (const_string "alu")
+          ]
+          (const_string "ishift1")))
+   (set (attr "length_immediate")
+     (if_then_else
+       (ior (eq_attr "type" "alu")
+           (and (eq_attr "type" "ishift1")
+                (and (match_operand 1 "const1_operand" "")
+                     (ne (symbol_ref "TARGET_SHIFT1 || optimize_function_for_size_p (cfun)")
+                         (const_int 0)))))
+       (const_string "0")
+       (const_string "*")))
+   (set_attr "mode" "QI")])
+
 ;; Convert lea to the lea pattern to avoid flags dependency.
 (define_split
   [(set (match_operand:DI 0 "register_operand" "")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
 
-;; See comment above `ashldi3' about how this works.
-
-(define_expand "ashrti3"
-  [(set (match_operand:TI 0 "register_operand" "")
-       (ashiftrt:TI (match_operand:TI 1 "register_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (ASHIFTRT, TImode, operands); DONE;")
-
-(define_insn "*ashrti3_1"
-  [(set (match_operand:TI 0 "register_operand" "=r")
-       (ashiftrt:TI (match_operand:TI 1 "register_operand" "0")
-                    (match_operand:QI 2 "nonmemory_operand" "Oc")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "#"
-  [(set_attr "type" "multi")])
-
-(define_peephole2
-  [(match_scratch:DI 3 "r")
-   (parallel [(set (match_operand:TI 0 "register_operand" "")
-                  (ashiftrt:TI (match_operand:TI 1 "register_operand" "")
-                               (match_operand:QI 2 "nonmemory_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])
-   (match_dup 3)]
-  "TARGET_64BIT"
-  [(const_int 0)]
-  "ix86_split_ashr (operands, operands[3], TImode); DONE;")
-
-(define_split
-  [(set (match_operand:TI 0 "register_operand" "")
-       (ashiftrt:TI (match_operand:TI 1 "register_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ((optimize > 0 && flag_peephole2)
-                   ? epilogue_completed : reload_completed)"
-  [(const_int 0)]
-  "ix86_split_ashr (operands, NULL_RTX, TImode); DONE;")
-
-(define_insn "x86_64_shrd"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "+r*m")
-        (ior:DI (ashiftrt:DI (match_dup 0)
-                 (match_operand:QI 2 "nonmemory_operand" "Jc"))
-               (ashift:DI (match_operand:DI 1 "register_operand" "r")
-                 (minus:QI (const_int 64) (match_dup 2)))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "shrd{q}\t{%s2%1, %0|%0, %1, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "prefix_0f" "1")
-   (set_attr "mode" "DI")
-   (set_attr "athlon_decode" "vector")
-   (set_attr "amdfam10_decode" "vector")])
-
-(define_expand "ashrdi3"
-  [(set (match_operand:DI 0 "shiftdi_operand" "")
-       (ashiftrt:DI (match_operand:DI 1 "shiftdi_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (ASHIFTRT, DImode, operands); DONE;")
+;; See comment above `ashl<mode>3' about how this works.
 
-(define_expand "x86_64_shift_adj_3"
-  [(use (match_operand:DI 0 "register_operand" ""))
-   (use (match_operand:DI 1 "register_operand" ""))
-   (use (match_operand:QI 2 "register_operand" ""))]
+(define_expand "<shiftrt_insn><mode>3"
+  [(set (match_operand:SDWIM 0 "<shift_operand>" "")
+       (any_shiftrt:SDWIM (match_operand:SDWIM 1 "<shift_operand>" "")
+                          (match_operand:QI 2 "nonmemory_operand" "")))]
   ""
-{
-  rtx label = gen_label_rtx ();
-  rtx tmp;
-
-  emit_insn (gen_testqi_ccz_1 (operands[2], GEN_INT (64)));
-
-  tmp = gen_rtx_REG (CCZmode, FLAGS_REG);
-  tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
-  tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
-                             gen_rtx_LABEL_REF (VOIDmode, label),
-                             pc_rtx);
-  tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
-  JUMP_LABEL (tmp) = label;
-
-  emit_move_insn (operands[0], operands[1]);
-  emit_insn (gen_ashrdi3_63_rex64 (operands[1], operands[1], GEN_INT (63)));
-
-  emit_label (label);
-  LABEL_NUSES (label) = 1;
-
-  DONE;
-})
-
-(define_insn "ashrdi3_63_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=*d,rm")
-       (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "*a,0")
-                    (match_operand:DI 2 "const_int_operand" "i,i")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && INTVAL (operands[2]) == 63
-   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "@
-   {cqto|cqo}
-   sar{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "imovx,ishift")
-   (set_attr "prefix_0f" "0,*")
-   (set_attr "length_immediate" "0,*")
-   (set_attr "modrm" "0,1")
-   (set_attr "mode" "DI")])
-
-(define_insn "*ashrdi3_1_one_bit_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "sar{q}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-(define_insn "*ashrdi3_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,rm")
-       (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "J,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "@
-   sar{q}\t{%2, %0|%0, %2}
-   sar{q}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "DI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrdi3_one_bit_cmp_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "sar{q}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-(define_insn "*ashrdi3_one_bit_cconly_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "sar{q}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrdi3_cmp_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_63_operand" "J"))
-         (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "sar{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "DI")])
-
-(define_insn "*ashrdi3_cconly_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_63_operand" "J"))
-         (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
-  "sar{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "DI")])
+  "ix86_expand_binary_operator (<CODE>, <MODE>mode, operands); DONE;")
 
-(define_insn "*ashrdi3_1"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (ashiftrt:DI (match_operand:DI 1 "register_operand" "0")
-                    (match_operand:QI 2 "nonmemory_operand" "Jc")))
+(define_insn_and_split "*<shiftrt_insn><mode>3_doubleword"
+  [(set (match_operand:DWI 0 "register_operand" "=r")
+       (any_shiftrt:DWI (match_operand:DWI 1 "register_operand" "0")
+                        (match_operand:QI 2 "nonmemory_operand" "<S>c")))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT"
+  ""
   "#"
+  "(optimize && flag_peephole2) ? epilogue_completed : reload_completed"
+  [(const_int 0)]
+  "ix86_split_<shiftrt_insn> (operands, NULL_RTX, <MODE>mode); DONE;"
   [(set_attr "type" "multi")])
 
-;; By default we don't ask for a scratch register, because when DImode
+;; By default we don't ask for a scratch register, because when DWImode
 ;; values are manipulated, registers are already at a premium.  But if
 ;; we have one handy, we won't turn it away.
-(define_peephole2
-  [(match_scratch:SI 3 "r")
-   (parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (ashiftrt:DI (match_operand:DI 1 "register_operand" "")
-                               (match_operand:QI 2 "nonmemory_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])
-   (match_dup 3)]
-  "!TARGET_64BIT && TARGET_CMOVE"
-  [(const_int 0)]
-  "ix86_split_ashr (operands, operands[3], DImode); DONE;")
 
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (ashiftrt:DI (match_operand:DI 1 "register_operand" "")
+(define_peephole2
+  [(match_scratch:DWIH 3 "r")
+   (parallel [(set (match_operand:<DWI> 0 "register_operand" "")
+                  (any_shiftrt:<DWI>
+                    (match_operand:<DWI> 1 "register_operand" "")
                     (match_operand:QI 2 "nonmemory_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && ((optimize > 0 && flag_peephole2)
-                    ? epilogue_completed : reload_completed)"
-  [(const_int 0)]
-  "ix86_split_ashr (operands, NULL_RTX, DImode); DONE;")
-
-(define_insn "x86_shrd"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "+r*m")
-        (ior:SI (ashiftrt:SI (match_dup 0)
-                 (match_operand:QI 2 "nonmemory_operand" "Ic"))
-               (ashift:SI (match_operand:SI 1 "register_operand" "r")
-                 (minus:QI (const_int 32) (match_dup 2)))))
-   (clobber (reg:CC FLAGS_REG))]
-  ""
-  "shrd{l}\t{%s2%1, %0|%0, %1, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "prefix_0f" "1")
-   (set_attr "pent_pair" "np")
-   (set_attr "mode" "SI")])
-
-(define_expand "x86_shift_adj_3"
-  [(use (match_operand:SI 0 "register_operand" ""))
-   (use (match_operand:SI 1 "register_operand" ""))
-   (use (match_operand:QI 2 "register_operand" ""))]
-  ""
-{
-  rtx label = gen_label_rtx ();
-  rtx tmp;
-
-  emit_insn (gen_testqi_ccz_1 (operands[2], GEN_INT (32)));
-
-  tmp = gen_rtx_REG (CCZmode, FLAGS_REG);
-  tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
-  tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
-                             gen_rtx_LABEL_REF (VOIDmode, label),
-                             pc_rtx);
-  tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
-  JUMP_LABEL (tmp) = label;
-
-  emit_move_insn (operands[0], operands[1]);
-  emit_insn (gen_ashrsi3_31 (operands[1], operands[1], GEN_INT (31)));
-
-  emit_label (label);
-  LABEL_NUSES (label) = 1;
-
-  DONE;
-})
-
-(define_expand "ashrsi3_31"
-  [(parallel [(set (match_operand:SI 0 "nonimmediate_operand" "=*d,rm")
-                  (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "*a,0")
-                               (match_operand:SI 2 "const_int_operand" "i,i")))
-              (clobber (reg:CC FLAGS_REG))])]
-  "")
-
-(define_insn "*ashrsi3_31"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=*d,rm")
-       (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "*a,0")
-                    (match_operand:SI 2 "const_int_operand" "i,i")))
-   (clobber (reg:CC FLAGS_REG))]
-  "INTVAL (operands[2]) == 31
-   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "@
-   {cltd|cdq}
-   sar{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "imovx,ishift")
-   (set_attr "prefix_0f" "0,*")
-   (set_attr "length_immediate" "0,*")
-   (set_attr "modrm" "0,1")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_31_zext"
-  [(set (match_operand:DI 0 "register_operand" "=*d,r")
-       (zero_extend:DI (ashiftrt:SI (match_operand:SI 1 "register_operand" "*a,0")
-                                    (match_operand:SI 2 "const_int_operand" "i,i"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
-   && INTVAL (operands[2]) == 31
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "@
-   {cltd|cdq}
-   sar{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "imovx,ishift")
-   (set_attr "prefix_0f" "0,*")
-   (set_attr "length_immediate" "0,*")
-   (set_attr "modrm" "0,1")
-   (set_attr "mode" "SI")])
-
-(define_expand "ashrsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (ASHIFTRT, SImode, operands); DONE;")
-
-(define_insn "*ashrsi3_1_one_bit"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_1_one_bit_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (ashiftrt:SI (match_operand:SI 1 "register_operand" "0")
-                                    (match_operand:QI 2 "const1_operand" ""))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t%k0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,rm")
-       (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "@
-   sar{l}\t{%2, %0|%0, %2}
-   sar{l}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (zero_extend:DI (ashiftrt:SI (match_operand:SI 1 "register_operand" "0,0")
-                                    (match_operand:QI 2 "nonmemory_operand" "I,c"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "@
-   sar{l}\t{%2, %k0|%k0, %2}
-   sar{l}\t{%b2, %k0|%k0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrsi3_one_bit_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:SI (match_dup 1) (match_dup 2)))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_one_bit_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_one_bit_cmp_zext"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:SI (match_operand:SI 1 "register_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (ashiftrt:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t%k0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrsi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:SI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashrsi3_cmp_zext"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:SI (match_operand:SI 1 "register_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (ashiftrt:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
-  "sar{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-(define_expand "ashrhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (ASHIFTRT, HImode, operands); DONE;")
-
-(define_insn "*ashrhi3_1_one_bit"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, HImode, operands)"
-  "sar{w}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
-
-(define_insn "*ashrhi3_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,rm")
-       (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (ASHIFTRT, HImode, operands)"
-  "@
-   sar{w}\t{%2, %0|%0, %2}
-   sar{w}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "HI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrhi3_one_bit_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:HI (match_dup 1) (match_dup 2)))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, HImode, operands)"
-  "sar{w}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
-
-(define_insn "*ashrhi3_one_bit_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, HImode, operands)"
-  "sar{w}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrhi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (ashiftrt:HI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, HImode, operands)"
-  "sar{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "HI")])
-
-(define_insn "*ashrhi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, HImode, operands)"
-  "sar{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "HI")])
-
-(define_expand "ashrqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (ASHIFTRT, QImode, operands); DONE;")
-
-(define_insn "*ashrqi3_1_one_bit"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "sar{b}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*ashrqi3_1_one_bit_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
-       (ashiftrt:QI (match_dup 0)
-                    (match_operand:QI 1 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "sar{b}\t%0"
-  [(set_attr "type" "ishift1")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*ashrqi3_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,qm")
-       (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "@
-   sar{b}\t{%2, %0|%0, %2}
-   sar{b}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "QI")])
-
-(define_insn "*ashrqi3_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,qm"))
-       (ashiftrt:QI (match_dup 0)
-                    (match_operand:QI 1 "nonmemory_operand" "I,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "@
-   sar{b}\t{%1, %0|%0, %1}
-   sar{b}\t{%b1, %0|%0, %b1}"
-  [(set_attr "type" "ishift1")
-   (set_attr "mode" "QI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrqi3_one_bit_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (ashiftrt:QI (match_dup 1) (match_dup 2)))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "sar{b}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*ashrqi3_one_bit_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "sar{b}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*ashrqi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (ashiftrt:QI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "sar{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "QI")])
-
-(define_insn "*ashrqi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFTRT, QImode, operands)"
-  "sar{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "QI")])
-
-\f
-;; Logical shift instructions
-
-;; See comment above `ashldi3' about how this works.
-
-(define_expand "lshrti3"
-  [(set (match_operand:TI 0 "register_operand" "")
-       (lshiftrt:TI (match_operand:TI 1 "register_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (LSHIFTRT, TImode, operands); DONE;")
-
-(define_insn "*lshrti3_1"
-  [(set (match_operand:TI 0 "register_operand" "=r")
-       (lshiftrt:TI (match_operand:TI 1 "register_operand" "0")
-                    (match_operand:QI 2 "nonmemory_operand" "Oc")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "#"
-  [(set_attr "type" "multi")])
-
-(define_peephole2
-  [(match_scratch:DI 3 "r")
-   (parallel [(set (match_operand:TI 0 "register_operand" "")
-                  (lshiftrt:TI (match_operand:TI 1 "register_operand" "")
-                               (match_operand:QI 2 "nonmemory_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])
-   (match_dup 3)]
-  "TARGET_64BIT"
-  [(const_int 0)]
-  "ix86_split_lshr (operands, operands[3], TImode); DONE;")
-
-(define_split
-  [(set (match_operand:TI 0 "register_operand" "")
-       (lshiftrt:TI (match_operand:TI 1 "register_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ((optimize > 0 && flag_peephole2)
-                   ? epilogue_completed : reload_completed)"
-  [(const_int 0)]
-  "ix86_split_lshr (operands, NULL_RTX, TImode); DONE;")
-
-(define_expand "lshrdi3"
-  [(set (match_operand:DI 0 "shiftdi_operand" "")
-       (lshiftrt:DI (match_operand:DI 1 "shiftdi_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (LSHIFTRT, DImode, operands); DONE;")
-
-(define_insn "*lshrdi3_1_one_bit_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{q}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-(define_insn "*lshrdi3_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,rm")
-       (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "J,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "@
-   shr{q}\t{%2, %0|%0, %2}
-   shr{q}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "DI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrdi3_cmp_one_bit_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{q}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-(define_insn "*lshrdi3_cconly_one_bit_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{q}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrdi3_cmp_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_63_operand" "J"))
-         (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "DI")])
-
-(define_insn "*lshrdi3_cconly_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_63_operand" "J"))
-         (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "DI")])
-
-(define_insn "*lshrdi3_1"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (lshiftrt:DI (match_operand:DI 1 "register_operand" "0")
-                    (match_operand:QI 2 "nonmemory_operand" "Jc")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT"
-  "#"
-  [(set_attr "type" "multi")])
-
-;; By default we don't ask for a scratch register, because when DImode
-;; values are manipulated, registers are already at a premium.  But if
-;; we have one handy, we won't turn it away.
-(define_peephole2
-  [(match_scratch:SI 3 "r")
-   (parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (lshiftrt:DI (match_operand:DI 1 "register_operand" "")
-                               (match_operand:QI 2 "nonmemory_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])
-   (match_dup 3)]
-  "!TARGET_64BIT && TARGET_CMOVE"
-  [(const_int 0)]
-  "ix86_split_lshr (operands, operands[3], DImode); DONE;")
-
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (lshiftrt:DI (match_operand:DI 1 "register_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && ((optimize > 0 && flag_peephole2)
-                    ? epilogue_completed : reload_completed)"
-  [(const_int 0)]
-  "ix86_split_lshr (operands, NULL_RTX, DImode); DONE;")
-
-(define_expand "lshrsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (LSHIFTRT, SImode, operands); DONE;")
-
-(define_insn "*lshrsi3_1_one_bit"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*lshrsi3_1_one_bit_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (lshiftrt:DI (zero_extend:DI (match_operand:SI 1 "register_operand" "0"))
-                    (match_operand:QI 2 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%k0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*lshrsi3_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,rm")
-       (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "@
-   shr{l}\t{%2, %0|%0, %2}
-   shr{l}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-(define_insn "*lshrsi3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (zero_extend:DI
-         (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0,0")
-                      (match_operand:QI 2 "nonmemory_operand" "I,c"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "@
-   shr{l}\t{%2, %k0|%k0, %2}
-   shr{l}\t{%b2, %k0|%k0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrsi3_one_bit_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:SI (match_dup 1) (match_dup 2)))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*lshrsi3_one_bit_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*lshrsi3_cmp_one_bit_zext"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:SI (match_operand:SI 1 "register_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (lshiftrt:DI (zero_extend:DI (match_dup 1)) (match_dup 2)))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%k0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrsi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:SI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
+             (clobber (reg:CC FLAGS_REG))])
+   (match_dup 3)]
+  "TARGET_CMOVE"
+  [(const_int 0)]
+  "ix86_split_<shiftrt_insn> (operands, operands[3], <DWI>mode); DONE;")
 
-(define_insn "*lshrsi3_cconly"
-  [(set (reg FLAGS_REG)
-      (compare
-       (lshiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const_1_to_31_operand" "I"))
-        (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t{%2, %0|%0, %2}"
+(define_insn "x86_64_shrd"
+  [(set (match_operand:DI 0 "nonimmediate_operand" "+r*m")
+        (ior:DI (ashiftrt:DI (match_dup 0)
+                 (match_operand:QI 2 "nonmemory_operand" "Jc"))
+               (ashift:DI (match_operand:DI 1 "register_operand" "r")
+                 (minus:QI (const_int 64) (match_dup 2)))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT"
+  "shrd{q}\t{%s2%1, %0|%0, %1, %2}"
   [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
+   (set_attr "prefix_0f" "1")
+   (set_attr "mode" "DI")
+   (set_attr "athlon_decode" "vector")
+   (set_attr "amdfam10_decode" "vector")])
 
-(define_insn "*lshrsi3_cmp_zext"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:SI (match_operand:SI 1 "register_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (lshiftrt:DI (zero_extend:DI (match_dup 1)) (match_dup 2)))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t{%2, %k0|%k0, %2}"
+(define_insn "x86_shrd"
+  [(set (match_operand:SI 0 "nonimmediate_operand" "+r*m")
+        (ior:SI (ashiftrt:SI (match_dup 0)
+                 (match_operand:QI 2 "nonmemory_operand" "Ic"))
+               (ashift:SI (match_operand:SI 1 "register_operand" "r")
+                 (minus:QI (const_int 32) (match_dup 2)))))
+   (clobber (reg:CC FLAGS_REG))]
+  ""
+  "shrd{l}\t{%s2%1, %0|%0, %1, %2}"
   [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
+   (set_attr "prefix_0f" "1")
+   (set_attr "mode" "SI")
+   (set_attr "pent_pair" "np")
+   (set_attr "athlon_decode" "vector")
+   (set_attr "amdfam10_decode" "vector")])
 
-(define_expand "lshrhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (LSHIFTRT, HImode, operands); DONE;")
+(define_insn "ashrdi3_cvt"
+  [(set (match_operand:DI 0 "nonimmediate_operand" "=*d,rm")
+       (ashiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "*a,0")
+                    (match_operand:QI 2 "const_int_operand" "")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT && INTVAL (operands[2]) == 63
+   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
+   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
+  "@
+   {cqto|cqo}
+   sar{q}\t{%2, %0|%0, %2}"
+  [(set_attr "type" "imovx,ishift")
+   (set_attr "prefix_0f" "0,*")
+   (set_attr "length_immediate" "0,*")
+   (set_attr "modrm" "0,1")
+   (set_attr "mode" "DI")])
 
-(define_insn "*lshrhi3_1_one_bit"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
+(define_insn "ashrsi3_cvt"
+  [(set (match_operand:SI 0 "nonimmediate_operand" "=*d,rm")
+       (ashiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "*a,0")
+                    (match_operand:QI 2 "const_int_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{w}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
+  "INTVAL (operands[2]) == 31
+   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
+   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
+  "@
+   {cltd|cdq}
+   sar{l}\t{%2, %0|%0, %2}"
+  [(set_attr "type" "imovx,ishift")
+   (set_attr "prefix_0f" "0,*")
+   (set_attr "length_immediate" "0,*")
+   (set_attr "modrm" "0,1")
+   (set_attr "mode" "SI")])
 
-(define_insn "*lshrhi3_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,rm")
-       (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
+(define_insn "*ashrsi3_cvt_zext"
+  [(set (match_operand:DI 0 "register_operand" "=*d,r")
+       (zero_extend:DI
+         (ashiftrt:SI (match_operand:SI 1 "register_operand" "*a,0")
+                      (match_operand:QI 2 "const_int_operand" ""))))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
+  "TARGET_64BIT && INTVAL (operands[2]) == 31
+   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
+   && ix86_binary_operator_ok (ASHIFTRT, SImode, operands)"
   "@
-   shr{w}\t{%2, %0|%0, %2}
-   shr{w}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "HI")])
+   {cltd|cdq}
+   sar{l}\t{%2, %k0|%k0, %2}"
+  [(set_attr "type" "imovx,ishift")
+   (set_attr "prefix_0f" "0,*")
+   (set_attr "length_immediate" "0,*")
+   (set_attr "modrm" "0,1")
+   (set_attr "mode" "SI")])
 
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrhi3_one_bit_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:HI (match_dup 1) (match_dup 2)))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{w}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
+(define_expand "x86_shift<mode>_adj_3"
+  [(use (match_operand:SWI48 0 "register_operand" ""))
+   (use (match_operand:SWI48 1 "register_operand" ""))
+   (use (match_operand:QI 2 "register_operand" ""))]
+  ""
+{
+  rtx label = gen_label_rtx ();
+  rtx tmp;
 
-(define_insn "*lshrhi3_one_bit_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{w}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
+  emit_insn (gen_testqi_ccz_1 (operands[2],
+                              GEN_INT (GET_MODE_BITSIZE (<MODE>mode))));
 
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrhi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (lshiftrt:HI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "HI")])
+  tmp = gen_rtx_REG (CCZmode, FLAGS_REG);
+  tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
+  tmp = gen_rtx_IF_THEN_ELSE (VOIDmode, tmp,
+                             gen_rtx_LABEL_REF (VOIDmode, label),
+                             pc_rtx);
+  tmp = emit_jump_insn (gen_rtx_SET (VOIDmode, pc_rtx, tmp));
+  JUMP_LABEL (tmp) = label;
 
-(define_insn "*lshrhi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
-         (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "HI")])
+  emit_move_insn (operands[0], operands[1]);
+  emit_insn (gen_ashr<mode>3_cvt (operands[1], operands[1],
+                                 GEN_INT (GET_MODE_BITSIZE (<MODE>mode)-1)));
+  emit_label (label);
+  LABEL_NUSES (label) = 1;
 
-(define_expand "lshrqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (lshiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (LSHIFTRT, QImode, operands); DONE;")
+  DONE;
+})
 
-(define_insn "*lshrqi3_1_one_bit"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (lshiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
+(define_insn "*<shiftrt_insn><mode>3_1"
+  [(set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (any_shiftrt:SWI (match_operand:SWI 1 "nonimmediate_operand" "0")
+                        (match_operand:QI 2 "nonmemory_operand" "c<S>")))
    (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, QImode, operands)"
-  "shr{b}\t%0"
+  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+{
+  if (operands[2] == const1_rtx
+      && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+    return "<shiftrt>{<imodesuffix>}\t%0";
+  else
+    return "<shiftrt>{<imodesuffix>}\t{%2, %0|%0, %2}";
+}
   [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*lshrqi3_1_one_bit_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
-       (lshiftrt:QI (match_dup 0)
-                    (match_operand:QI 1 "const1_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))"
-  "shr{b}\t%0"
-  [(set_attr "type" "ishift1")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
+   (set (attr "length_immediate")
+     (if_then_else
+       (and (match_operand 2 "const1_operand" "")
+           (ne (symbol_ref "TARGET_SHIFT1 || optimize_function_for_size_p (cfun)")
+               (const_int 0)))
+       (const_string "0")
+       (const_string "*")))
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*lshrqi3_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,qm")
-       (lshiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
+(define_insn "*<shiftrt_insn>si3_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (any_shiftrt:SI (match_operand:SI 1 "register_operand" "0")
+                         (match_operand:QI 2 "nonmemory_operand" "cI"))))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (LSHIFTRT, QImode, operands)"
-  "@
-   shr{b}\t{%2, %0|%0, %2}
-   shr{b}\t{%b2, %0|%0, %b2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+{
+  if (operands[2] == const1_rtx
+      && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+    return "<shiftrt>{l}\t%k0";
+  else
+    return "<shiftrt>{l}\t{%2, %k0|%k0, %2}";
+}
   [(set_attr "type" "ishift")
-   (set_attr "mode" "QI")])
+   (set (attr "length_immediate")
+     (if_then_else
+       (and (match_operand 2 "const1_operand" "")
+           (ne (symbol_ref "TARGET_SHIFT1 || optimize_function_for_size_p (cfun)")
+               (const_int 0)))
+       (const_string "0")
+       (const_string "*")))
+   (set_attr "mode" "SI")])
 
-(define_insn "*lshrqi3_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,qm"))
-       (lshiftrt:QI (match_dup 0)
-                    (match_operand:QI 1 "nonmemory_operand" "I,c")))
+(define_insn "*<shiftrt_insn>qi3_1_slp"
+  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
+       (any_shiftrt:QI (match_dup 0)
+                       (match_operand:QI 1 "nonmemory_operand" "cI")))
    (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "@
-   shr{b}\t{%1, %0|%0, %1}
-   shr{b}\t{%b1, %0|%0, %b1}"
+  "(optimize_function_for_size_p (cfun)
+    || !TARGET_PARTIAL_REG_STALL
+    || (operands[1] == const1_rtx
+       && TARGET_SHIFT1))"
+{
+  if (operands[1] == const1_rtx
+      && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+    return "<shiftrt>{b}\t%0";
+  else
+    return "<shiftrt>{b}\t{%1, %0|%0, %1}";
+}
   [(set_attr "type" "ishift1")
+   (set (attr "length_immediate")
+     (if_then_else
+       (and (match_operand 1 "const1_operand" "")
+           (ne (symbol_ref "TARGET_SHIFT1 || optimize_function_for_size_p (cfun)")
+               (const_int 0)))
+       (const_string "0")
+       (const_string "*")))
    (set_attr "mode" "QI")])
 
 ;; This pattern can't accept a variable shift count, since shifts by
 ;; zero don't affect the flags.  We assume that shifts by constant
 ;; zero are optimized away.
-(define_insn "*lshrqi2_one_bit_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (lshiftrt:QI (match_dup 1) (match_dup 2)))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, QImode, operands)"
-  "shr{b}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*lshrqi2_one_bit_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (lshiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
-         (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, QImode, operands)"
-  "shr{b}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-;; This pattern can't accept a variable shift count, since shifts by
-;; zero don't affect the flags.  We assume that shifts by constant
-;; zero are optimized away.
-(define_insn "*lshrqi2_cmp"
+(define_insn "*<shiftrt_insn><mode>3_cmp"
   [(set (reg FLAGS_REG)
        (compare
-         (lshiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
+         (any_shiftrt:SWI
+           (match_operand:SWI 1 "nonimmediate_operand" "0")
+           (match_operand:QI 2 "<shift_immediate_operand>" "<S>"))
          (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (lshiftrt:QI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (any_shiftrt:SWI (match_dup 1) (match_dup 2)))]
+  "(optimize_function_for_size_p (cfun)
+    || !TARGET_PARTIAL_FLAG_REG_STALL
+    || (operands[2] == const1_rtx
+       && TARGET_SHIFT1))
    && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (LSHIFTRT, QImode, operands)"
-  "shr{b}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+{
+  if (operands[2] == const1_rtx
+      && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+    return "<shiftrt>{<imodesuffix>}\t%0";
+  else
+    return "<shiftrt>{<imodesuffix>}\t{%2, %0|%0, %2}";
+}
   [(set_attr "type" "ishift")
    (set (attr "length_immediate")
      (if_then_else
   [(const_int 0)]
 {
   operands[7] = gen_rtx_FLOAT (GET_MODE (operands[1]), operands[2]);
-
   ix86_split_fp_branch (swap_condition (GET_CODE (operands[0])),
                        operands[3], operands[7],
                        operands[4], operands[5], operands[6], NULL_RTX);
 {
   operands[7] = ix86_force_to_memory (GET_MODE (operands[2]), operands[2]);
   operands[7] = gen_rtx_FLOAT (GET_MODE (operands[1]), operands[7]);
-
   ix86_split_fp_branch (swap_condition (GET_CODE (operands[0])),
                        operands[3], operands[7],
                        operands[4], operands[5], operands[6], operands[2]);
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
   "TARGET_64BIT"
-  "rep{%;} movsq"
+  "rep movsq"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "both")
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
   "!TARGET_64BIT"
-  "rep{%;} movs{l|d}"
+  "rep movs{l|d}"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "both")
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
   "TARGET_64BIT"
-  "rep{%;} movs{l|d}"
+  "rep movs{l|d}"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "both")
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
   "!TARGET_64BIT"
-  "rep{%;} movsb"
+  "rep movsb"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "both")
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
   "TARGET_64BIT"
-  "rep{%;} movsb"
+  "rep movsb"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "both")
    (use (match_operand:DI 2 "register_operand" "a"))
    (use (match_dup 4))]
   "TARGET_64BIT"
-  "rep{%;} stosq"
+  "rep stosq"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "store")
    (use (match_operand:SI 2 "register_operand" "a"))
    (use (match_dup 4))]
   "!TARGET_64BIT"
-  "rep{%;} stos{l|d}"
+  "rep stos{l|d}"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "store")
    (use (match_operand:SI 2 "register_operand" "a"))
    (use (match_dup 4))]
   "TARGET_64BIT"
-  "rep{%;} stos{l|d}"
+  "rep stos{l|d}"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "store")
    (use (match_operand:QI 2 "register_operand" "a"))
    (use (match_dup 4))]
   "!TARGET_64BIT"
-  "rep{%;} stosb"
+  "rep stosb"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "store")
    (use (match_operand:QI 2 "register_operand" "a"))
    (use (match_dup 4))]
   "TARGET_64BIT"
-  "rep{%;} stosb"
+  "rep stosb"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "store")
    (clobber (match_operand:SI 1 "register_operand" "=D"))
    (clobber (match_operand:SI 2 "register_operand" "=c"))]
   "!TARGET_64BIT"
-  "repz{%;} cmpsb"
+  "repz cmpsb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set_attr "prefix_rep" "1")])
    (clobber (match_operand:DI 1 "register_operand" "=D"))
    (clobber (match_operand:DI 2 "register_operand" "=c"))]
   "TARGET_64BIT"
-  "repz{%;} cmpsb"
+  "repz cmpsb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set_attr "prefix_rex" "0")
    (clobber (match_operand:SI 1 "register_operand" "=D"))
    (clobber (match_operand:SI 2 "register_operand" "=c"))]
   "!TARGET_64BIT"
-  "repz{%;} cmpsb"
+  "repz cmpsb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set_attr "prefix_rep" "1")])
    (clobber (match_operand:DI 1 "register_operand" "=D"))
    (clobber (match_operand:DI 2 "register_operand" "=c"))]
   "TARGET_64BIT"
-  "repz{%;} cmpsb"
+  "repz cmpsb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set_attr "prefix_rex" "0")
    (clobber (match_operand:SI 1 "register_operand" "=D"))
    (clobber (reg:CC FLAGS_REG))]
   "!TARGET_64BIT"
-  "repnz{%;} scasb"
+  "repnz scasb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set_attr "prefix_rep" "1")])
    (clobber (match_operand:DI 1 "register_operand" "=D"))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT"
-  "repnz{%;} scasb"
+  "repnz scasb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set_attr "prefix_rex" "0")
        (if_then_else:SI (match_op_dup 1 [(match_dup 4) (const_int 0)])
                      (match_dup 7)
                      (match_dup 8)))]
-{
-  split_di (&operands[2], 2, &operands[5], &operands[7]);
-  split_di (&operands[0], 1, &operands[2], &operands[3]);
-})
+  "split_di (&operands[2], 2, &operands[5], &operands[7]);
+   split_di (&operands[0], 1, &operands[2], &operands[3]);")
 
 (define_insn "*movxfcc_1"
   [(set (match_operand:XF 0 "register_operand" "=f,f")
                                (reg:DI XMM5_REG)
                                (reg:DI XMM6_REG)
                                (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE))
-             (clobber (reg:CC FLAGS_REG))
              (clobber (match_operand:DI 1 "register_operand" ""))
              (use (match_operand:DI 2 "immediate_operand" ""))
              (use (label_ref:DI (match_operand 3 "" "")))
                     (reg:DI XMM5_REG)
                     (reg:DI XMM6_REG)
                     (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE))
-   (clobber (reg:CC FLAGS_REG))
    (clobber (match_operand:DI 1 "register_operand" "=r"))
    (use (match_operand:DI 2 "const_int_operand" "i"))
    (use (label_ref:DI (match_operand 3 "" "X")))
                                 (reg:DI XMM5_REG)
                                 (reg:DI XMM6_REG)
                                 (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE))
-              (clobber (reg:CC FLAGS_REG))
               (clobber (match_operand:DI 1 "register_operand" ""))
               (use (match_operand:DI 2 "const_int_operand" ""))
               (use (match_operand 3 "" ""))