OSDN Git Service

* config/i386/i386.md (extendsidi2 splitter): Also check for DX_REG
[pf3gnuchains/gcc-fork.git] / gcc / config / i386 / i386.md
index 2041396..b879996 100644 (file)
@@ -1,6 +1,6 @@
 ;; GCC machine description for IA-32 and x86-64.
 ;; Copyright (C) 1988, 1994, 1995, 1996, 1997, 1998, 1999, 2000,
-;; 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009
+;; 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
 ;; Free Software Foundation, Inc.
 ;; Mostly by William Schelter.
 ;; x86_64 support added by Jan Hubicka
@@ -30,7 +30,6 @@
 ;; L,W,B,Q,S,T -- print the opcode suffix for specified size of operand.
 ;; C -- print opcode suffix for set/cmov insn.
 ;; c -- like C, but print reversed condition
-;; E,e -- likewise, but for compare-and-branch fused insn.
 ;; F,f -- likewise, but for floating-point.
 ;; O -- if HAVE_AS_IX86_CMOV_SUN_SYNTAX, expand to "w.", "l." or "q.",
 ;;      otherwise nothing
@@ -57,7 +56,7 @@
 ;; X -- don't print any sort of PIC '@' suffix for a symbol.
 ;; & -- print some in-use local-dynamic symbol name.
 ;; H -- print a memory address offset by 8; used for sse high-parts
-;; Y -- print condition for SSE5 com* instruction.
+;; Y -- print condition for XOP pcom* instruction.
 ;; + -- print a branch hint as 'cs' or 'ds' prefix
 ;; ; -- print a semicolon (after prefixes due to bug in older gas).
 
    (UNSPEC_PCMPESTR            144)
    (UNSPEC_PCMPISTR            145)
 
-   ;; For SSE5
-   (UNSPEC_SSE5_INTRINSIC      150)
-   (UNSPEC_SSE5_UNSIGNED_CMP   151)
-   (UNSPEC_SSE5_TRUEFALSE      152)
-   (UNSPEC_SSE5_PERMUTE                153)
+   ; For FMA4 support
+   (UNSPEC_FMA4_INTRINSIC      150)
+   (UNSPEC_FMA4_FMADDSUB       151)
+   (UNSPEC_FMA4_FMSUBADD       152)
+   (UNSPEC_XOP_UNSIGNED_CMP    151)
+   (UNSPEC_XOP_TRUEFALSE       152)
+   (UNSPEC_XOP_PERMUTE         153)
    (UNSPEC_FRCZ                        154)
-   (UNSPEC_CVTPH2PS            155)
-   (UNSPEC_CVTPS2PH            156)
 
    ; For AES support
    (UNSPEC_AESENC              159)
    ; For AVX support
    (UNSPEC_PCMP                        166)
    (UNSPEC_VPERMIL             167)
-   (UNSPEC_VPERMIL2F128                168)
-   (UNSPEC_MASKLOAD            169)
-   (UNSPEC_MASKSTORE           170)
-   (UNSPEC_CAST                        171)
-   (UNSPEC_VTESTP              172)
+   (UNSPEC_VPERMIL2            168)
+   (UNSPEC_VPERMIL2F128                169)
+   (UNSPEC_MASKLOAD            170)
+   (UNSPEC_MASKSTORE           171)
+   (UNSPEC_CAST                        172)
+   (UNSPEC_VTESTP              173)
   ])
 
 (define_constants
    (UNSPECV_CLD                        15)
    (UNSPECV_VZEROALL           16)
    (UNSPECV_VZEROUPPER         17)
+   (UNSPECV_RDTSC              18)
+   (UNSPECV_RDTSCP             19)
+   (UNSPECV_RDPMC              20)
+   (UNSPECV_VSWAPMOV           21)
+   (UNSPECV_LLWP_INTRINSIC     22)
+   (UNSPECV_SLWP_INTRINSIC     23)
+   (UNSPECV_LWPVAL_INTRINSIC   24)
+   (UNSPECV_LWPINS_INTRINSIC   25)
   ])
 
 ;; Constants to represent pcomtrue/pcomfalse variants
    (COM_TRUE_P                 5)
   ])
 
-;; Constants used in the SSE5 pperm instruction
+;; Constants used in the XOP pperm instruction
 (define_constants
   [(PPERM_SRC                  0x00)   /* copy source */
    (PPERM_INVERT               0x20)   /* invert source */
    fmov,fop,fsgn,fmul,fdiv,fpspc,fcmov,fcmp,fxch,fistp,fisttp,frndint,
    sselog,sselog1,sseiadd,sseiadd1,sseishft,sseimul,
    sse,ssemov,sseadd,ssemul,ssecmp,ssecomi,ssecvt,ssecvt1,sseicvt,ssediv,sseins,
-   ssemuladd,sse4arg,
+   ssemuladd,sse4arg,lwp,
    mmx,mmxmov,mmxadd,mmxmul,mmxcmp,mmxcvt,mmxshft"
   (const_string "other"))
 
        ]
        (const_int 0)))
 
-;; There are also additional prefixes in 3DNOW, SSSE3 or SSE5.
+;; There are also additional prefixes in 3DNOW, SSSE3.
 ;; ssemuladd,sse4arg default to 0f24/0f25 and DREX byte,
 ;; sseiadd1,ssecvt1 to 0f7a with no DREX byte.
 ;; 3DNOW has 0f0f prefix, SSSE3 and SSE4_{1,2} 0f38/0f3a.
 ;; if the instruction is complex.
 
 (define_attr "memory" "none,load,store,both,unknown"
-  (cond [(eq_attr "type" "other,multi,str")
+  (cond [(eq_attr "type" "other,multi,str,lwp")
           (const_string "unknown")
         (eq_attr "type" "lea,fcmov,fpspc")
           (const_string "none")
    (set_attr "type" "multi")])
 
 ;; All integer comparison codes.
-(define_code_iterator int_cond [ne eq ge gt le lt geu gtu leu ltu ])
+(define_code_iterator int_cond [ne eq ge gt le lt geu gtu leu ltu])
 
 ;; All floating-point comparison codes.
 (define_code_iterator fp_cond [unordered ordered
-                              uneq unge ungt unle unlt ltgt ])
+                              uneq unge ungt unle unlt ltgt])
 
 (define_code_iterator plusminus [plus minus])
 
 (define_code_attr plusminus_mnemonic
   [(plus "add") (ss_plus "adds") (us_plus "addus")
    (minus "sub") (ss_minus "subs") (us_minus "subus")])
+(define_code_attr plusminus_carry_mnemonic
+  [(plus "adc") (minus "sbb")])
 
 ;; Mark commutative operators as such in constraints.
 (define_code_attr comm [(plus "%") (ss_plus "%") (us_plus "%")
                                 (umax "maxu") (umin "minu")])
 (define_code_attr maxminfprefix [(smax "max") (smin "min")])
 
-;; Mapping of parallel logic operators
-(define_code_iterator plogic [and ior xor])
+;; Mapping of logic operators
+(define_code_iterator any_logic [and ior xor])
+(define_code_iterator any_or [ior xor])
 
 ;; Base name for insn mnemonic.
-(define_code_attr plogicprefix [(and "and") (ior "or") (xor "xor")])
+(define_code_attr logicprefix [(and "and") (ior "or") (xor "xor")])
+
+;; Mapping of shift-right operators
+(define_code_iterator any_shiftrt [lshiftrt ashiftrt])
+
+;; Base name for define_insn
+(define_code_attr shiftrt_insn [(lshiftrt "lshr") (ashiftrt "ashr")])
+
+;; Base name for insn mnemonic.
+(define_code_attr shiftrt [(lshiftrt "shr") (ashiftrt "sar")])
+
+;; Mapping of rotate operators
+(define_code_iterator any_rotate [rotate rotatert])
+
+;; Base name for define_insn
+(define_code_attr rotate_insn [(rotate "rotl") (rotatert "rotr")])
+
+;; Base name for insn mnemonic.
+(define_code_attr rotate [(rotate "rol") (rotatert "ror")])
 
 ;; Mapping of abs neg operators
 (define_code_iterator absneg [abs neg])
 ;; Base name for x87 insn mnemonic.
 (define_code_attr absnegprefix [(abs "abs") (neg "chs")])
 
+;; Used in signed and unsigned widening multiplications.
+(define_code_iterator any_extend [sign_extend zero_extend])
+
+;; Various insn prefixes for signed and unsigned operations.
+(define_code_attr u [(sign_extend "") (zero_extend "u")
+                    (div "") (udiv "u")])
+(define_code_attr s [(sign_extend "s") (zero_extend "u")])
+
+;; Used in signed and unsigned divisions.
+(define_code_iterator any_div [div udiv])
+
+;; Instruction prefix for signed and unsigned operations.
+(define_code_attr sgnprefix [(sign_extend "i") (zero_extend "")
+                            (div "i") (udiv "")])
+
 ;; All single word integer modes.
 (define_mode_iterator SWI [QI HI SI (DI "TARGET_64BIT")])
 
+;; Single word integer modes without DImode.
+(define_mode_iterator SWI124 [QI HI SI])
+
 ;; Single word integer modes without QImode.
 (define_mode_iterator SWI248 [HI SI (DI "TARGET_64BIT")])
 
+;; Single word integer modes without QImode and HImode.
+(define_mode_iterator SWI48 [SI (DI "TARGET_64BIT")])
+
+;; All math-dependant single and double word integer modes.
+(define_mode_iterator SDWIM [(QI "TARGET_QIMODE_MATH")
+                            (HI "TARGET_HIMODE_MATH")
+                            SI DI (TI "TARGET_64BIT")])
+
+;; Math-dependant single word integer modes.
+(define_mode_iterator SWIM [(QI "TARGET_QIMODE_MATH")
+                           (HI "TARGET_HIMODE_MATH")
+                           SI (DI "TARGET_64BIT")])
+
+;; Math-dependant single word integer modes without DImode.
+(define_mode_iterator SWIM124 [(QI "TARGET_QIMODE_MATH")
+                              (HI "TARGET_HIMODE_MATH")
+                              SI])
+
+;; Math-dependant single word integer modes without QImode.
+(define_mode_iterator SWIM248 [(HI "TARGET_HIMODE_MATH")
+                              SI (DI "TARGET_64BIT")])
+
+;; Double word integer modes.
+(define_mode_iterator DWI [(DI "!TARGET_64BIT")
+                          (TI "TARGET_64BIT")])
+
+;; Double word integer modes as mode attribute.
+(define_mode_attr DWI [(SI "DI") (DI "TI")])
+(define_mode_attr dwi [(SI "di") (DI "ti")])
+
+;; Half mode for double word integer modes.
+(define_mode_iterator DWIH [(SI "!TARGET_64BIT")
+                           (DI "TARGET_64BIT")])
+
 ;; Instruction suffix for integer modes.
 (define_mode_attr imodesuffix [(QI "b") (HI "w") (SI "l") (DI "q")])
 
 ;; Immediate operand constraint for integer modes.
 (define_mode_attr i [(QI "n") (HI "n") (SI "i") (DI "e")])
 
+;; General operand constraint for word modes.
+(define_mode_attr g [(QI "qmn") (HI "rmn") (SI "g") (DI "rme")])
+
+;; Immediate operand constraint for double integer modes.
+(define_mode_attr di [(SI "iF") (DI "e")])
+
+;; Immediate operand constraint for shifts.
+(define_mode_attr S [(QI "I") (HI "I") (SI "I") (DI "J") (TI "O")])
+
 ;; General operand predicate for integer modes.
 (define_mode_attr general_operand
        [(QI "general_operand")
         (HI "general_operand")
         (SI "general_operand")
-        (DI "x86_64_general_operand")])
+        (DI "x86_64_general_operand")
+        (TI "x86_64_general_operand")])
+
+;; General sign/zero extend operand predicate for integer modes.
+(define_mode_attr general_szext_operand
+       [(QI "general_operand")
+        (HI "general_operand")
+        (SI "general_operand")
+        (DI "x86_64_szext_general_operand")])
+
+;; Operand predicate for shifts.
+(define_mode_attr shift_operand
+       [(QI "nonimmediate_operand")
+        (HI "nonimmediate_operand")
+        (SI "nonimmediate_operand")
+        (DI "shiftdi_operand")
+        (TI "register_operand")])
+
+;; Operand predicate for shift argument.
+(define_mode_attr shift_immediate_operand
+       [(QI "const_1_to_31_operand")
+        (HI "const_1_to_31_operand")
+        (SI "const_1_to_31_operand")
+        (DI "const_1_to_63_operand")])
+
+;; Input operand predicate for arithmetic left shifts.
+(define_mode_attr ashl_input_operand
+       [(QI "nonimmediate_operand")
+        (HI "nonimmediate_operand")
+        (SI "nonimmediate_operand")
+        (DI "ashldi_input_operand")
+        (TI "reg_or_pm1_operand")])
 
 ;; SSE and x87 SFmode and DFmode floating point modes
 (define_mode_iterator MODEF [SF DF])
 ;; This mode iterator allows :P to be used for patterns that operate on
 ;; pointer-sized quantities.  Exactly one of the two alternatives will match.
 (define_mode_iterator P [(SI "Pmode == SImode") (DI "Pmode == DImode")])
-
 \f
 ;; Scheduling descriptions
 
 \f
 ;; Compare and branch/compare and store instructions.
 
-(define_expand "cbranchti4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:TI 1 "nonimmediate_operand" "")
-                   (match_operand:TI 2 "x86_64_general_operand" "")))
-   (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)])
-              (label_ref (match_operand 3 "" ""))
-              (pc)))]
-  "TARGET_64BIT"
-{
-  if (MEM_P (operands[1]) && MEM_P (operands[2]))
-    operands[1] = force_reg (TImode, operands[1]);
-  ix86_compare_op0 = operands[1];
-  ix86_compare_op1 = operands[2];
-  ix86_expand_branch (GET_CODE (operands[0]), operands[3]);
-  DONE;
-})
-
-(define_expand "cbranchdi4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:DI 1 "nonimmediate_operand" "")
-                   (match_operand:DI 2 "x86_64_general_operand" "")))
-   (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)])
-              (label_ref (match_operand 3 "" ""))
-              (pc)))]
-  ""
-{
-  if (MEM_P (operands[1]) && MEM_P (operands[2]))
-    operands[1] = force_reg (DImode, operands[1]);
-  ix86_compare_op0 = operands[1];
-  ix86_compare_op1 = operands[2];
-  ix86_expand_branch (GET_CODE (operands[0]), operands[3]);
-  DONE;
-})
-
-(define_expand "cstoredi4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:DI 2 "nonimmediate_operand" "")
-                   (match_operand:DI 3 "x86_64_general_operand" "")))
-   (set (match_operand:QI 0 "register_operand" "")
-              (match_operator 1 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)]))]
-  "TARGET_64BIT"
-{
-  if (MEM_P (operands[2]) && MEM_P (operands[3]))
-    operands[2] = force_reg (DImode, operands[2]);
-  ix86_compare_op0 = operands[2];
-  ix86_compare_op1 = operands[3];
-  ix86_expand_setcc (GET_CODE (operands[1]), operands[0]);
-  DONE;
-})
-
-(define_expand "cbranchsi4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:SI 1 "cmpsi_operand" "")
-                   (match_operand:SI 2 "general_operand" "")))
-   (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)])
-              (label_ref (match_operand 3 "" ""))
-              (pc)))]
-  ""
-{
-  if (MEM_P (operands[1]) && MEM_P (operands[2]))
-    operands[1] = force_reg (SImode, operands[1]);
-  ix86_compare_op0 = operands[1];
-  ix86_compare_op1 = operands[2];
-  ix86_expand_branch (GET_CODE (operands[0]), operands[3]);
-  DONE;
-})
-
-(define_expand "cstoresi4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:SI 2 "cmpsi_operand" "")
-                   (match_operand:SI 3 "general_operand" "")))
-   (set (match_operand:QI 0 "register_operand" "")
-              (match_operator 1 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)]))]
-  ""
-{
-  if (MEM_P (operands[2]) && MEM_P (operands[3]))
-    operands[2] = force_reg (SImode, operands[2]);
-  ix86_compare_op0 = operands[2];
-  ix86_compare_op1 = operands[3];
-  ix86_expand_setcc (GET_CODE (operands[1]), operands[0]);
-  DONE;
-})
-
-(define_expand "cbranchhi4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:HI 1 "nonimmediate_operand" "")
-                   (match_operand:HI 2 "general_operand" "")))
-   (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)])
-              (label_ref (match_operand 3 "" ""))
-              (pc)))]
-  ""
-{
-  if (MEM_P (operands[1]) && MEM_P (operands[2]))
-    operands[1] = force_reg (HImode, operands[1]);
-  ix86_compare_op0 = operands[1];
-  ix86_compare_op1 = operands[2];
-  ix86_expand_branch (GET_CODE (operands[0]), operands[3]);
-  DONE;
-})
-
-(define_expand "cstorehi4"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:HI 2 "nonimmediate_operand" "")
-                   (match_operand:HI 3 "general_operand" "")))
-   (set (match_operand:QI 0 "register_operand" "")
-              (match_operator 1 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)]))]
-  ""
-{
-  if (MEM_P (operands[2]) && MEM_P (operands[3]))
-    operands[2] = force_reg (HImode, operands[2]);
-  ix86_compare_op0 = operands[2];
-  ix86_compare_op1 = operands[3];
-  ix86_expand_setcc (GET_CODE (operands[1]), operands[0]);
-  DONE;
-})
-
-
-(define_expand "cbranchqi4"
+(define_expand "cbranch<mode>4"
   [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:QI 1 "nonimmediate_operand" "")
-                   (match_operand:QI 2 "general_operand" "")))
+       (compare:CC (match_operand:SDWIM 1 "nonimmediate_operand" "")
+                   (match_operand:SDWIM 2 "<general_operand>" "")))
    (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)])
-              (label_ref (match_operand 3 "" ""))
-              (pc)))]
+              (match_operator 0 "comparison_operator"
+               [(reg:CC FLAGS_REG) (const_int 0)])
+              (label_ref (match_operand 3 "" ""))
+              (pc)))]
   ""
 {
   if (MEM_P (operands[1]) && MEM_P (operands[2]))
-    operands[1] = force_reg (QImode, operands[1]);
+    operands[1] = force_reg (<MODE>mode, operands[1]);
   ix86_compare_op0 = operands[1];
   ix86_compare_op1 = operands[2];
   ix86_expand_branch (GET_CODE (operands[0]), operands[3]);
   DONE;
 })
 
-
-(define_expand "cstoreqi4"
+(define_expand "cstore<mode>4"
   [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:QI 2 "nonimmediate_operand" "")
-                   (match_operand:QI 3 "general_operand" "")))
+       (compare:CC (match_operand:SWIM 2 "nonimmediate_operand" "")
+                   (match_operand:SWIM 3 "<general_operand>" "")))
    (set (match_operand:QI 0 "register_operand" "")
-              (match_operator 1 "comparison_operator"
-               [(reg:CC FLAGS_REG)
-                (const_int 0)]))]
+       (match_operator 1 "comparison_operator"
+         [(reg:CC FLAGS_REG) (const_int 0)]))]
   ""
 {
   if (MEM_P (operands[2]) && MEM_P (operands[3]))
-    operands[2] = force_reg (QImode, operands[2]);
+    operands[2] = force_reg (<MODE>mode, operands[2]);
   ix86_compare_op0 = operands[2];
   ix86_compare_op1 = operands[3];
   ix86_expand_setcc (GET_CODE (operands[1]), operands[0]);
   DONE;
 })
 
-
-(define_insn "cmpdi_ccno_1_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:DI 0 "nonimmediate_operand" "r,?mr")
-                (match_operand:DI 1 "const0_operand" "")))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
-  "@
-   test{q}\t%0, %0
-   cmp{q}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "test,icmp")
-   (set_attr "length_immediate" "0,1")
-   (set_attr "mode" "DI")])
-
-(define_insn "*cmpdi_minus_1_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (minus:DI (match_operand:DI 0 "nonimmediate_operand" "rm,r")
-                          (match_operand:DI 1 "x86_64_general_operand" "re,mr"))
-                (const_int 0)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCGOCmode)"
-  "cmp{q}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "DI")])
-
-(define_expand "cmpdi_1_rex64"
-  [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:DI 0 "nonimmediate_operand" "")
-                   (match_operand:DI 1 "general_operand" "")))]
-  "TARGET_64BIT"
-  "")
-
-(define_insn "cmpdi_1_insn_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:DI 0 "nonimmediate_operand" "mr,r")
-                (match_operand:DI 1 "x86_64_general_operand" "re,mr")))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCmode)"
-  "cmp{q}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "DI")])
-
-
-(define_insn "*cmpsi_ccno_1"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:SI 0 "nonimmediate_operand" "r,?mr")
-                (match_operand:SI 1 "const0_operand" "")))]
-  "ix86_match_ccmode (insn, CCNOmode)"
-  "@
-   test{l}\t%0, %0
-   cmp{l}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "test,icmp")
-   (set_attr "length_immediate" "0,1")
-   (set_attr "mode" "SI")])
-
-(define_insn "*cmpsi_minus_1"
-  [(set (reg FLAGS_REG)
-       (compare (minus:SI (match_operand:SI 0 "nonimmediate_operand" "rm,r")
-                          (match_operand:SI 1 "general_operand" "ri,mr"))
-                (const_int 0)))]
-  "ix86_match_ccmode (insn, CCGOCmode)"
-  "cmp{l}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "SI")])
-
-(define_expand "cmpsi_1"
+(define_expand "cmp<mode>_1"
   [(set (reg:CC FLAGS_REG)
-       (compare:CC (match_operand:SI 0 "nonimmediate_operand" "")
-                   (match_operand:SI 1 "general_operand" "")))]
+       (compare:CC (match_operand:SWI48 0 "nonimmediate_operand" "")
+                   (match_operand:SWI48 1 "<general_operand>" "")))]
   ""
   "")
 
-(define_insn "*cmpsi_1_insn"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:SI 0 "nonimmediate_operand" "rm,r")
-                (match_operand:SI 1 "general_operand" "ri,mr")))]
-  "!(MEM_P (operands[0]) && MEM_P (operands[1]))
-    && ix86_match_ccmode (insn, CCmode)"
-  "cmp{l}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "SI")])
-
-(define_insn "*cmphi_ccno_1"
+(define_insn "*cmp<mode>_ccno_1"
   [(set (reg FLAGS_REG)
-       (compare (match_operand:HI 0 "nonimmediate_operand" "r,?mr")
-                (match_operand:HI 1 "const0_operand" "")))]
+       (compare (match_operand:SWI 0 "nonimmediate_operand" "<r>,?m<r>")
+                (match_operand:SWI 1 "const0_operand" "")))]
   "ix86_match_ccmode (insn, CCNOmode)"
   "@
-   test{w}\t%0, %0
-   cmp{w}\t{%1, %0|%0, %1}"
+   test{<imodesuffix>}\t%0, %0
+   cmp{<imodesuffix>}\t{%1, %0|%0, %1}"
   [(set_attr "type" "test,icmp")
    (set_attr "length_immediate" "0,1")
-   (set_attr "mode" "HI")])
-
-(define_insn "*cmphi_minus_1"
-  [(set (reg FLAGS_REG)
-       (compare (minus:HI (match_operand:HI 0 "nonimmediate_operand" "rm,r")
-                          (match_operand:HI 1 "general_operand" "rn,mr"))
-                (const_int 0)))]
-  "ix86_match_ccmode (insn, CCGOCmode)"
-  "cmp{w}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "HI")])
-
-(define_insn "*cmphi_1"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:HI 0 "nonimmediate_operand" "rm,r")
-                (match_operand:HI 1 "general_operand" "rn,mr")))]
-  "!(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && ix86_match_ccmode (insn, CCmode)"
-  "cmp{w}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "HI")])
-
-(define_insn "*cmpqi_ccno_1"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:QI 0 "nonimmediate_operand" "q,?mq")
-                (match_operand:QI 1 "const0_operand" "")))]
-  "ix86_match_ccmode (insn, CCNOmode)"
-  "@
-   test{b}\t%0, %0
-   cmp{b}\t{$0, %0|%0, 0}"
-  [(set_attr "type" "test,icmp")
-   (set_attr "length_immediate" "0,1")
-   (set_attr "mode" "QI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*cmpqi_1"
+(define_insn "*cmp<mode>_1"
   [(set (reg FLAGS_REG)
-       (compare (match_operand:QI 0 "nonimmediate_operand" "qm,q")
-                (match_operand:QI 1 "general_operand" "qn,mq")))]
-  "!(MEM_P (operands[0]) && MEM_P (operands[1]))
-    && ix86_match_ccmode (insn, CCmode)"
-  "cmp{b}\t{%1, %0|%0, %1}"
+       (compare (match_operand:SWI 0 "nonimmediate_operand" "<r>m,<r>")
+                (match_operand:SWI 1 "<general_operand>" "<r><i>,<r>m")))]
+  "ix86_match_ccmode (insn, CCmode)"
+  "cmp{<imodesuffix>}\t{%1, %0|%0, %1}"
   [(set_attr "type" "icmp")
-   (set_attr "mode" "QI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*cmpqi_minus_1"
+(define_insn "*cmp<mode>_minus_1"
   [(set (reg FLAGS_REG)
-       (compare (minus:QI (match_operand:QI 0 "nonimmediate_operand" "qm,q")
-                          (match_operand:QI 1 "general_operand" "qn,mq"))
-                (const_int 0)))]
+       (compare
+         (minus:SWI (match_operand:SWI 0 "nonimmediate_operand" "<r>m,<r>")
+                    (match_operand:SWI 1 "<general_operand>" "<r><i>,<r>m"))
+         (const_int 0)))]
   "ix86_match_ccmode (insn, CCGOCmode)"
-  "cmp{b}\t{%1, %0|%0, %1}"
+  "cmp{<imodesuffix>}\t{%1, %0|%0, %1}"
   [(set_attr "type" "icmp")
-   (set_attr "mode" "QI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_insn "*cmpqi_ext_1"
   [(set (reg FLAGS_REG)
              (match_operand 0 "ext_register_operand" "")
              (const_int 8)
              (const_int 8)) 0)
-         (match_operand:QI 1 "general_operand" "")))]
+         (match_operand:QI 1 "immediate_operand" "")))]
   ""
   "")
 
-(define_insn "cmpqi_ext_3_insn"
+(define_insn "*cmpqi_ext_3_insn"
   [(set (reg FLAGS_REG)
        (compare
          (subreg:QI
    (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
 
-(define_insn "cmpqi_ext_3_insn_rex64"
+(define_insn "*cmpqi_ext_3_insn_rex64"
   [(set (reg FLAGS_REG)
        (compare
          (subreg:QI
        (compare:CC (match_operand:XF 1 "nonmemory_operand" "")
                    (match_operand:XF 2 "nonmemory_operand" "")))
    (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
+              (match_operator 0 "ix86_fp_comparison_operator"
                [(reg:CC FLAGS_REG)
                 (const_int 0)])
               (label_ref (match_operand 3 "" ""))
        (compare:CC (match_operand:XF 2 "nonmemory_operand" "")
                    (match_operand:XF 3 "nonmemory_operand" "")))
    (set (match_operand:QI 0 "register_operand" "")
-              (match_operator 1 "comparison_operator"
+              (match_operator 1 "ix86_fp_comparison_operator"
                [(reg:CC FLAGS_REG)
                 (const_int 0)]))]
   "TARGET_80387"
        (compare:CC (match_operand:MODEF 1 "cmp_fp_expander_operand" "")
                    (match_operand:MODEF 2 "cmp_fp_expander_operand" "")))
    (set (pc) (if_then_else
-              (match_operator 0 "comparison_operator"
+              (match_operator 0 "ix86_fp_comparison_operator"
                [(reg:CC FLAGS_REG)
                 (const_int 0)])
               (label_ref (match_operand 3 "" ""))
        (compare:CC (match_operand:MODEF 2 "cmp_fp_expander_operand" "")
                    (match_operand:MODEF 3 "cmp_fp_expander_operand" "")))
    (set (match_operand:QI 0 "register_operand" "")
-              (match_operator 1 "comparison_operator"
+              (match_operator 1 "ix86_fp_comparison_operator"
                [(reg:CC FLAGS_REG)
                 (const_int 0)]))]
   "TARGET_80387 || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)"
 #ifdef HAVE_AS_IX86_SAHF
   return "sahf";
 #else
-  return ".byte\t0x9e";
+  return ASM_BYTE "0x9e";
 #endif
 }
   [(set_attr "length" "1")
   [(set (match_dup 0)
        (match_dup 2))])
 
-
 ;; %%% Kill this when call knows how to work this out.
 (define_split
   [(set (match_operand:SF 0 "push_operand" "")
    && (TARGET_ZERO_EXTEND_WITH_AND
        && optimize_function_for_speed_p (cfun))
    && !reg_overlap_mentioned_p (operands[0], operands[1])"
-  [(set (match_dup 0) (const_int 0))
-   (set (strict_low_part (match_dup 2)) (match_dup 1))]
-  "operands[2] = gen_lowpart (QImode, operands[0]);")
+  [(set (strict_low_part (match_dup 2)) (match_dup 1))]
+{
+  operands[2] = gen_lowpart (QImode, operands[0]);
+  ix86_expand_clear (operands[0]);
+})
 
 ;; Rest is handled by single and.
 (define_split
   [(set_attr "type" "alu1")
    (set_attr "mode" "SI")])
 
-(define_insn "*zero_extendqisi2_movzbw_and"
+(define_insn "*zero_extendqisi2_movzbl_and"
   [(set (match_operand:SI 0 "register_operand" "=r,r")
      (zero_extend:SI (match_operand:QI 1 "nonimmediate_operand" "qm,0")))
    (clobber (reg:CC FLAGS_REG))]
   [(set_attr "type" "imovx,alu1")
    (set_attr "mode" "SI")])
 
-(define_insn "*zero_extendqisi2_movzbw"
+(define_insn "*zero_extendqisi2_movzbl"
   [(set (match_operand:SI 0 "register_operand" "=r")
      (zero_extend:SI (match_operand:QI 1 "nonimmediate_operand" "qm")))]
   "(!TARGET_ZERO_EXTEND_WITH_AND || optimize_function_for_size_p (cfun))
    && (ANY_QI_REG_P (operands[1]) || MEM_P (operands[1]))
    && (TARGET_ZERO_EXTEND_WITH_AND && optimize_function_for_speed_p (cfun))
    && !reg_overlap_mentioned_p (operands[0], operands[1])"
-  [(set (match_dup 0) (const_int 0))
-   (set (strict_low_part (match_dup 2)) (match_dup 1))]
-  "operands[2] = gen_lowpart (QImode, operands[0]);")
+  [(set (strict_low_part (match_dup 2)) (match_dup 1))]
+{
+  operands[2] = gen_lowpart (QImode, operands[0]);
+  ix86_expand_clear (operands[0]);
+})
 
 ;; Rest is handled by single and.
 (define_split
       && true_regnum (operands[1]) == AX_REG
       && true_regnum (operands[2]) == DX_REG)
     {
-      emit_insn (gen_ashrsi3_31 (operands[2], operands[1], GEN_INT (31)));
+      emit_insn (gen_ashrsi3_cvt (operands[2], operands[1], GEN_INT (31)));
     }
   else
     {
       emit_move_insn (operands[2], operands[1]);
-      emit_insn (gen_ashrsi3_31 (operands[2], operands[2], GEN_INT (31)));
+      emit_insn (gen_ashrsi3_cvt (operands[2], operands[2], GEN_INT (31)));
     }
   emit_move_insn (operands[4], operands[2]);
   DONE;
 
   /* Generate a cltd if possible and doing so it profitable.  */
   if ((optimize_function_for_size_p (cfun) || TARGET_USE_CLTD)
-      && true_regnum (operands[3]) == AX_REG)
+      && true_regnum (operands[3]) == AX_REG
+      && true_regnum (operands[4]) == DX_REG)
     {
-      emit_insn (gen_ashrsi3_31 (operands[4], operands[3], GEN_INT (31)));
+      emit_insn (gen_ashrsi3_cvt (operands[4], operands[3], GEN_INT (31)));
       DONE;
     }
 
   if (true_regnum (operands[4]) != true_regnum (operands[1]))
     emit_move_insn (operands[4], operands[1]);
 
-  emit_insn (gen_ashrsi3_31 (operands[4], operands[4], GEN_INT (31)));
+  emit_insn (gen_ashrsi3_cvt (operands[4], operands[4], GEN_INT (31)));
   DONE;
 })
 
        }
       else
        operands[3] = simplify_gen_subreg (V4SFmode, operands[1], SFmode, 0);
-      emit_insn (gen_sse_unpcklps (operands[3], operands[3], operands[3]));
+      emit_insn (gen_vec_interleave_lowv4sf (operands[3], operands[3],
+                                            operands[3]));
     }
   else
     emit_insn (gen_vec_setv4sf_0 (operands[3],
    && !((SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
         && (TARGET_64BIT || <MODE>mode != DImode))
        && TARGET_SSE_MATH)
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    && !TARGET_FISTTP
    && !(SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
         && (TARGET_64BIT || <MODE>mode != DImode))
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
   "TARGET_80387
    && (!(SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)
        || TARGET_MIX_SSE_I387)
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(parallel [(set (match_dup 0)
   "TARGET_80387
    || ((<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
        && SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH)"
-  "
 {
   if (!((<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
        && SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH)
       && !X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode))
     {
       rtx reg = gen_reg_rtx (XFmode);
+      rtx insn;
+
       emit_insn (gen_float<SSEMODEI24:mode>xf2 (reg, operands[1]));
-/* Avoid references to nonexistent function in dead code in XFmode case.  */
-#define gen_truncxfxf2 gen_truncxfdf2
-      emit_insn (gen_truncxf<X87MODEF:mode>2 (operands[0], reg));
-#undef gen_truncxfxf2
+
+      if (<X87MODEF:MODE>mode == SFmode)
+       insn = gen_truncxfsf2 (operands[0], reg);
+      else if (<X87MODEF:MODE>mode == DFmode)
+       insn = gen_truncxfdf2 (operands[0], reg);
+      else
+       gcc_unreachable ();
+
+      emit_insn (insn);
       DONE;
     }
-}")
+})
 
 ;; Pre-reload splitter to add memory clobber to the pattern.
 (define_insn_and_split "*float<SSEMODEI24:mode><X87MODEF:mode>2_1"
             && flag_trapping_math)
            || !(TARGET_INTER_UNIT_CONVERSIONS
                 || optimize_function_for_size_p (cfun)))))
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(parallel [(set (match_dup 0) (float:X87MODEF (match_dup 1)))
                              gen_rtx_SUBREG (SImode, operands[1], 0)));
   emit_insn (gen_sse2_loadld (operands[4], CONST0_RTX (V4SImode),
                              gen_rtx_SUBREG (SImode, operands[1], 4)));
-  emit_insn (gen_sse2_punpckldq (operands[3], operands[3], operands[4]));
+  emit_insn (gen_vec_interleave_lowv4si (operands[3], operands[3],
+                                        operands[4]));
 
   operands[3] = gen_rtx_REG (DImode, REGNO (operands[3]));
 })
 \f
 ;; Add instructions
 
-;; %%% splits for addditi3
-
-(define_expand "addti3"
-  [(set (match_operand:TI 0 "nonimmediate_operand" "")
-       (plus:TI (match_operand:TI 1 "nonimmediate_operand" "")
-                (match_operand:TI 2 "x86_64_general_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (PLUS, TImode, operands); DONE;")
-
-(define_insn "*addti3_1"
-  [(set (match_operand:TI 0 "nonimmediate_operand" "=r,o")
-       (plus:TI (match_operand:TI 1 "nonimmediate_operand" "%0,0")
-                (match_operand:TI 2 "x86_64_general_operand" "roe,re")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, TImode, operands)"
-  "#")
+(define_expand "add<mode>3"
+  [(set (match_operand:SDWIM 0 "nonimmediate_operand" "")
+       (plus:SDWIM (match_operand:SDWIM 1 "nonimmediate_operand" "")
+                   (match_operand:SDWIM 2 "<general_operand>" "")))]
+  ""
+  "ix86_expand_binary_operator (PLUS, <MODE>mode, operands); DONE;")
 
-(define_split
-  [(set (match_operand:TI 0 "nonimmediate_operand" "")
-       (plus:TI (match_operand:TI 1 "nonimmediate_operand" "")
-                (match_operand:TI 2 "x86_64_general_operand" "")))
+(define_insn_and_split "*add<dwi>3_doubleword"
+  [(set (match_operand:<DWI> 0 "nonimmediate_operand" "=r,o")
+       (plus:<DWI>
+         (match_operand:<DWI> 1 "nonimmediate_operand" "%0,0")
+         (match_operand:<DWI> 2 "<general_operand>" "ro<di>,r<di>")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed"
-  [(parallel [(set (reg:CC FLAGS_REG) (unspec:CC [(match_dup 1) (match_dup 2)]
-                                         UNSPEC_ADD_CARRY))
-             (set (match_dup 0) (plus:DI (match_dup 1) (match_dup 2)))])
+  "ix86_binary_operator_ok (PLUS, <DWI>mode, operands)"
+  "#"
+  "reload_completed"
+  [(parallel [(set (reg:CC FLAGS_REG)
+                  (unspec:CC [(match_dup 1) (match_dup 2)]
+                             UNSPEC_ADD_CARRY))
+             (set (match_dup 0)
+                  (plus:DWIH (match_dup 1) (match_dup 2)))])
    (parallel [(set (match_dup 3)
-                  (plus:DI (plus:DI (ltu:DI (reg:CC FLAGS_REG) (const_int 0))
-                                    (match_dup 4))
-                           (match_dup 5)))
+                  (plus:DWIH
+                    (match_dup 4)
+                    (plus:DWIH
+                      (ltu:DWIH (reg:CC FLAGS_REG) (const_int 0))
+                      (match_dup 5))))
              (clobber (reg:CC FLAGS_REG))])]
-  "split_ti (&operands[0], 3, &operands[0], &operands[3]);")
+  "split_<dwi> (&operands[0], 3, &operands[0], &operands[3]);")
 
-;; %%% splits for addsidi3
-;  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-;      (plus:DI (match_operand:DI 1 "general_operand" "")
-;               (zero_extend:DI (match_operand:SI 2 "general_operand" ""))))]
-
-(define_expand "adddi3"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (plus:DI (match_operand:DI 1 "nonimmediate_operand" "")
-                (match_operand:DI 2 "x86_64_general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (PLUS, DImode, operands); DONE;")
-
-(define_insn "*adddi3_1"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=r,o")
-       (plus:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
-                (match_operand:DI 2 "general_operand" "roiF,riF")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && ix86_binary_operator_ok (PLUS, DImode, operands)"
-  "#")
-
-(define_split
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (plus:DI (match_operand:DI 1 "nonimmediate_operand" "")
-                (match_operand:DI 2 "general_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && reload_completed"
-  [(parallel [(set (reg:CC FLAGS_REG) (unspec:CC [(match_dup 1) (match_dup 2)]
-                                         UNSPEC_ADD_CARRY))
-             (set (match_dup 0) (plus:SI (match_dup 1) (match_dup 2)))])
-   (parallel [(set (match_dup 3)
-                  (plus:SI (plus:SI (ltu:SI (reg:CC FLAGS_REG) (const_int 0))
-                                    (match_dup 4))
-                           (match_dup 5)))
-             (clobber (reg:CC FLAGS_REG))])]
-  "split_di (&operands[0], 3, &operands[0], &operands[3]);")
-
-(define_insn "adddi3_carry_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-         (plus:DI (plus:DI (match_operand:DI 3 "ix86_carry_flag_operator" "")
-                           (match_operand:DI 1 "nonimmediate_operand" "%0,0"))
-                  (match_operand:DI 2 "x86_64_general_operand" "re,rm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, DImode, operands)"
-  "adc{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*adddi3_cc_rex64"
+(define_insn "*add<mode>3_cc"
   [(set (reg:CC FLAGS_REG)
-       (unspec:CC [(match_operand:DI 1 "nonimmediate_operand" "%0,0")
-                   (match_operand:DI 2 "x86_64_general_operand" "re,rm")]
-                  UNSPEC_ADD_CARRY))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (plus:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, DImode, operands)"
-  "add{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*<plusminus_insn><mode>3_cc_overflow"
-  [(set (reg:CCC FLAGS_REG)
-       (compare:CCC
-           (plusminus:SWI
-               (match_operand:SWI 1 "nonimmediate_operand" "<comm>0,0")
-               (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m"))
-           (match_dup 1)))
-   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
-       (plusminus:SWI (match_dup 1) (match_dup 2)))]
-  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
-  "<plusminus_mnemonic>{<imodesuffix>}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "<MODE>")])
-
-(define_insn "*add<mode>3_cconly_overflow"
-  [(set (reg:CCC FLAGS_REG)
-       (compare:CCC
-               (plus:SWI (match_operand:SWI 1 "nonimmediate_operand" "%0")
-                         (match_operand:SWI 2 "<general_operand>" "<r><i>m"))
-               (match_dup 1)))
-   (clobber (match_scratch:SWI 0 "=<r>"))]
+       (unspec:CC
+         [(match_operand:SWI48 1 "nonimmediate_operand" "%0,0")
+          (match_operand:SWI48 2 "<general_operand>" "r<i>,rm")]
+         UNSPEC_ADD_CARRY))
+   (set (match_operand:SWI48 0 "nonimmediate_operand" "=rm,r")
+       (plus:SWI48 (match_dup 1) (match_dup 2)))]
   "ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
   "add{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
    (set_attr "mode" "<MODE>")])
 
-(define_insn "*sub<mode>3_cconly_overflow"
-  [(set (reg:CCC FLAGS_REG)
-       (compare:CCC
-            (minus:SWI (match_operand:SWI 0 "nonimmediate_operand" "<r>m,<r>")
-                       (match_operand:SWI 1 "<general_operand>" "<r><i>,<r>m"))
-            (match_dup 0)))]
-  ""
-  "cmp{<imodesuffix>}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "icmp")
-   (set_attr "mode" "<MODE>")])
-
-(define_insn "*<plusminus_insn>si3_zext_cc_overflow"
-  [(set (reg:CCC FLAGS_REG)
-       (compare:CCC
-           (plusminus:SI (match_operand:SI 1 "nonimmediate_operand" "<comm>0")
-                         (match_operand:SI 2 "general_operand" "g"))
-           (match_dup 1)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (plusminus:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
-  "<plusminus_mnemonic>{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_insn "addqi3_carry"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q")
-         (plus:QI (plus:QI (match_operand:QI 3 "ix86_carry_flag_operator" "")
-                           (match_operand:QI 1 "nonimmediate_operand" "%0,0"))
-                  (match_operand:QI 2 "general_operand" "qn,qm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (PLUS, QImode, operands)"
-  "adc{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "QI")])
-
-(define_insn "addhi3_carry"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-         (plus:HI (plus:HI (match_operand:HI 3 "ix86_carry_flag_operator" "")
-                           (match_operand:HI 1 "nonimmediate_operand" "%0,0"))
-                  (match_operand:HI 2 "general_operand" "rn,rm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (PLUS, HImode, operands)"
-  "adc{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "HI")])
-
-(define_insn "addsi3_carry"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-         (plus:SI (plus:SI (match_operand:SI 3 "ix86_carry_flag_operator" "")
-                           (match_operand:SI 1 "nonimmediate_operand" "%0,0"))
-                  (match_operand:SI 2 "general_operand" "ri,rm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (PLUS, SImode, operands)"
-  "adc{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*addsi3_carry_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-         (zero_extend:DI
-           (plus:SI (plus:SI (match_operand:SI 3 "ix86_carry_flag_operator" "")
-                             (match_operand:SI 1 "nonimmediate_operand" "%0"))
-                    (match_operand:SI 2 "general_operand" "g"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, SImode, operands)"
-  "adc{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*addsi3_cc"
-  [(set (reg:CC FLAGS_REG)
-       (unspec:CC [(match_operand:SI 1 "nonimmediate_operand" "%0,0")
-                   (match_operand:SI 2 "general_operand" "ri,rm")]
-                  UNSPEC_ADD_CARRY))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (plus:SI (match_dup 1) (match_dup 2)))]
-  "ix86_binary_operator_ok (PLUS, SImode, operands)"
-  "add{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
 (define_insn "addqi3_cc"
   [(set (reg:CC FLAGS_REG)
-       (unspec:CC [(match_operand:QI 1 "nonimmediate_operand" "%0,0")
-                   (match_operand:QI 2 "general_operand" "qn,qm")]
-                  UNSPEC_ADD_CARRY))
+       (unspec:CC
+         [(match_operand:QI 1 "nonimmediate_operand" "%0,0")
+          (match_operand:QI 2 "general_operand" "qn,qm")]
+         UNSPEC_ADD_CARRY))
    (set (match_operand:QI 0 "nonimmediate_operand" "=qm,q")
        (plus:QI (match_dup 1) (match_dup 2)))]
   "ix86_binary_operator_ok (PLUS, QImode, operands)"
   [(set_attr "type" "alu")
    (set_attr "mode" "QI")])
 
-(define_expand "addsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (plus:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                (match_operand:SI 2 "general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (PLUS, SImode, operands); DONE;")
-
 (define_insn "*lea_1"
-  [(set (match_operand:SI 0 "register_operand" "=r")
-       (match_operand:SI 1 "no_seg_address_operand" "p"))]
-  "!TARGET_64BIT"
-  "lea{l}\t{%a1, %0|%0, %a1}"
+  [(set (match_operand:DWIH 0 "register_operand" "=r")
+       (match_operand:DWIH 1 "no_seg_address_operand" "p"))]
+  ""
+  "lea{<imodesuffix>}\t{%a1, %0|%0, %a1}"
   [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*lea_1_rex64"
+(define_insn "*lea_2"
   [(set (match_operand:SI 0 "register_operand" "=r")
        (subreg:SI (match_operand:DI 1 "no_seg_address_operand" "p") 0))]
   "TARGET_64BIT"
   [(set_attr "type" "lea")
    (set_attr "mode" "SI")])
 
-(define_insn "*lea_1_zext"
+(define_insn "*lea_2_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
-        (subreg:SI (match_operand:DI 1 "no_seg_address_operand" "p") 0)))]
+         (subreg:SI (match_operand:DI 1 "no_seg_address_operand" "p") 0)))]
   "TARGET_64BIT"
   "lea{l}\t{%a1, %k0|%k0, %a1}"
   [(set_attr "type" "lea")
    (set_attr "mode" "SI")])
 
-(define_insn "*lea_2_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (match_operand:DI 1 "no_seg_address_operand" "p"))]
-  "TARGET_64BIT"
-  "lea{q}\t{%a1, %0|%0, %a1}"
-  [(set_attr "type" "lea")
-   (set_attr "mode" "DI")])
+(define_insn "*add<mode>_1"
+  [(set (match_operand:SWI48 0 "nonimmediate_operand" "=r,rm,r,r")
+       (plus:SWI48
+         (match_operand:SWI48 1 "nonimmediate_operand" "%0,0,r,r")
+         (match_operand:SWI48 2 "<general_operand>" "<g>,r<i>,0,l<i>")))
+   (clobber (reg:CC FLAGS_REG))]
+  "ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
+{
+  switch (get_attr_type (insn))
+    {
+    case TYPE_LEA:
+      operands[2] = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
+      return "lea{<imodesuffix>}\t{%a2, %0|%0, %a2}";
 
-;; The lea patterns for non-Pmodes needs to be matched by several
-;; insns converted to real lea by splitters.
+    case TYPE_INCDEC:
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (operands[2] == const1_rtx)
+        return "inc{<imodesuffix>}\t%0";
+      else
+        {
+         gcc_assert (operands[2] == constm1_rtx);
+          return "dec{<imodesuffix>}\t%0";
+       }
 
-(define_insn_and_split "*lea_general_1"
-  [(set (match_operand 0 "register_operand" "=r")
-       (plus (plus (match_operand 1 "index_register_operand" "l")
-                   (match_operand 2 "register_operand" "r"))
-             (match_operand 3 "immediate_operand" "i")))]
-  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
-    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
-   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && GET_MODE (operands[0]) == GET_MODE (operands[1])
-   && GET_MODE (operands[0]) == GET_MODE (operands[2])
-   && (GET_MODE (operands[0]) == GET_MODE (operands[3])
-       || GET_MODE (operands[3]) == VOIDmode)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-{
-  rtx pat;
-  operands[0] = gen_lowpart (SImode, operands[0]);
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[2] = gen_lowpart (Pmode, operands[2]);
-  operands[3] = gen_lowpart (Pmode, operands[3]);
-  pat = gen_rtx_PLUS (Pmode, gen_rtx_PLUS (Pmode, operands[1], operands[2]),
-                     operands[3]);
-  if (Pmode != SImode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
-  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
-  DONE;
+    default:
+      /* Use add as much as possible to replace lea for AGU optimization. */
+      if (which_alternative == 2 && TARGET_OPT_AGU)
+        return "add{<imodesuffix>}\t{%1, %0|%0, %1}";
+        
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (x86_maybe_negate_const_int (&operands[2], <MODE>mode))
+        return "sub{<imodesuffix>}\t{%2, %0|%0, %2}";
+
+      return "add{<imodesuffix>}\t{%2, %0|%0, %2}";
+    }
 }
-  [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
+  [(set (attr "type")
+     (cond [(and (eq_attr "alternative" "2") 
+                 (eq (symbol_ref "TARGET_OPT_AGU") (const_int 0)))
+             (const_string "lea")
+            (eq_attr "alternative" "3")
+              (const_string "lea")
+           ; Current assemblers are broken and do not allow @GOTOFF in
+           ; ought but a memory context.
+           (match_operand:SWI48 2 "pic_symbolic_operand" "")
+             (const_string "lea")
+           (match_operand:SWI48 2 "incdec_operand" "")
+             (const_string "incdec")
+          ]
+          (const_string "alu")))
+   (set (attr "length_immediate")
+      (if_then_else
+       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
+       (const_string "1")
+       (const_string "*")))
+   (set_attr "mode" "<MODE>")])
 
-(define_insn_and_split "*lea_general_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
+;; It may seem that nonimmediate operand is proper one for operand 1.
+;; The addsi_1 pattern allows nonimmediate operand at that place and
+;; we take care in ix86_binary_operator_ok to not allow two memory
+;; operands so proper swapping will be done in reload.  This allow
+;; patterns constructed from addsi_1 to match.
+
+(define_insn "*addsi_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r,r")
        (zero_extend:DI
-         (plus:SI (plus:SI (match_operand:SI 1 "index_register_operand" "l")
-                           (match_operand:SI 2 "register_operand" "r"))
-                  (match_operand:SI 3 "immediate_operand" "i"))))]
-  "TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(set (match_dup 0)
-       (zero_extend:DI (subreg:SI (plus:DI (plus:DI (match_dup 1)
-                                                    (match_dup 2))
-                                           (match_dup 3)) 0)))]
+         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,r")
+                  (match_operand:SI 2 "general_operand" "g,li"))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, SImode, operands)"
 {
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[2] = gen_lowpart (Pmode, operands[2]);
-  operands[3] = gen_lowpart (Pmode, operands[3]);
+  switch (get_attr_type (insn))
+    {
+    case TYPE_LEA:
+      operands[2] = XEXP (SET_SRC (XVECEXP (PATTERN (insn), 0, 0)), 0);
+      return "lea{l}\t{%a2, %k0|%k0, %a2}";
+
+    case TYPE_INCDEC:
+      if (operands[2] == const1_rtx)
+        return "inc{l}\t%k0";
+      else
+        {
+         gcc_assert (operands[2] == constm1_rtx);
+          return "dec{l}\t%k0";
+       }
+
+    default:
+      if (x86_maybe_negate_const_int (&operands[2], SImode))
+        return "sub{l}\t{%2, %k0|%k0, %2}";
+
+      return "add{l}\t{%2, %k0|%k0, %2}";
+    }
 }
-  [(set_attr "type" "lea")
+  [(set (attr "type")
+     (cond [(eq_attr "alternative" "1")
+             (const_string "lea")
+           ; Current assemblers are broken and do not allow @GOTOFF in
+           ; ought but a memory context.
+           (match_operand:SI 2 "pic_symbolic_operand" "")
+             (const_string "lea")
+           (match_operand:SI 2 "incdec_operand" "")
+             (const_string "incdec")
+          ]
+          (const_string "alu")))
+   (set (attr "length_immediate")
+      (if_then_else
+       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
+       (const_string "1")
+       (const_string "*")))
    (set_attr "mode" "SI")])
 
-(define_insn_and_split "*lea_general_2"
-  [(set (match_operand 0 "register_operand" "=r")
-       (plus (mult (match_operand 1 "index_register_operand" "l")
-                   (match_operand 2 "const248_operand" "i"))
-             (match_operand 3 "nonmemory_operand" "ri")))]
-  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
-    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
-   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && GET_MODE (operands[0]) == GET_MODE (operands[1])
-   && (GET_MODE (operands[0]) == GET_MODE (operands[3])
-       || GET_MODE (operands[3]) == VOIDmode)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-{
-  rtx pat;
-  operands[0] = gen_lowpart (SImode, operands[0]);
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[3] = gen_lowpart (Pmode, operands[3]);
-  pat = gen_rtx_PLUS (Pmode, gen_rtx_MULT (Pmode, operands[1], operands[2]),
-                     operands[3]);
-  if (Pmode != SImode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
-  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
-  DONE;
-}
-  [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
-
-(define_insn_and_split "*lea_general_2_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI
-         (plus:SI (mult:SI (match_operand:SI 1 "index_register_operand" "l")
-                           (match_operand:SI 2 "const248_operand" "n"))
-                  (match_operand:SI 3 "nonmemory_operand" "ri"))))]
-  "TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(set (match_dup 0)
-       (zero_extend:DI (subreg:SI (plus:DI (mult:DI (match_dup 1)
-                                                    (match_dup 2))
-                                           (match_dup 3)) 0)))]
-{
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[3] = gen_lowpart (Pmode, operands[3]);
-}
-  [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
-
-(define_insn_and_split "*lea_general_3"
-  [(set (match_operand 0 "register_operand" "=r")
-       (plus (plus (mult (match_operand 1 "index_register_operand" "l")
-                         (match_operand 2 "const248_operand" "i"))
-                   (match_operand 3 "register_operand" "r"))
-             (match_operand 4 "immediate_operand" "i")))]
-  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
-    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
-   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && GET_MODE (operands[0]) == GET_MODE (operands[1])
-   && GET_MODE (operands[0]) == GET_MODE (operands[3])"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-{
-  rtx pat;
-  operands[0] = gen_lowpart (SImode, operands[0]);
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[3] = gen_lowpart (Pmode, operands[3]);
-  operands[4] = gen_lowpart (Pmode, operands[4]);
-  pat = gen_rtx_PLUS (Pmode,
-                     gen_rtx_PLUS (Pmode, gen_rtx_MULT (Pmode, operands[1],
-                                                        operands[2]),
-                                   operands[3]),
-                     operands[4]);
-  if (Pmode != SImode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
-  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
-  DONE;
-}
-  [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
-
-(define_insn_and_split "*lea_general_3_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI
-         (plus:SI (plus:SI (mult:SI
-                             (match_operand:SI 1 "index_register_operand" "l")
-                             (match_operand:SI 2 "const248_operand" "n"))
-                           (match_operand:SI 3 "register_operand" "r"))
-                  (match_operand:SI 4 "immediate_operand" "i"))))]
-  "TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(set (match_dup 0)
-       (zero_extend:DI (subreg:SI (plus:DI (plus:DI (mult:DI (match_dup 1)
-                                                             (match_dup 2))
-                                                    (match_dup 3))
-                                           (match_dup 4)) 0)))]
-{
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[3] = gen_lowpart (Pmode, operands[3]);
-  operands[4] = gen_lowpart (Pmode, operands[4]);
-}
-  [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
-
-(define_insn "*adddi_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=r,rm,r,r")
-       (plus:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0,r,r")
-                (match_operand:DI 2 "x86_64_general_operand" "rme,re,0,le")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, DImode, operands)"
+(define_insn "*addhi_1"
+  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
+       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
+                (match_operand:HI 2 "general_operand" "rn,rm")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (PLUS, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
-    case TYPE_LEA:
-      operands[2] = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
-      return "lea{q}\t{%a2, %0|%0, %a2}";
-
     case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-        return "inc{q}\t%0";
+       return "inc{w}\t%0";
       else
         {
          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{q}\t%0";
+         return "dec{w}\t%0";
        }
 
     default:
-      /* Use add as much as possible to replace lea for AGU optimization. */
-      if (which_alternative == 2 && TARGET_OPT_AGU)
-        return "add{q}\t{%1, %0|%0, %1}";
-        
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (x86_maybe_negate_const_int (&operands[2], HImode))
+       return "sub{w}\t{%2, %0|%0, %2}";
 
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-         /* Avoid overflows.  */
-         && ((INTVAL (operands[2]) & ((((unsigned int) 1) << 31) - 1)))
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{q}\t{%2, %0|%0, %2}";
-        }
-      return "add{q}\t{%2, %0|%0, %2}";
+      return "add{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (cond [(and (eq_attr "alternative" "2") 
-                 (eq (symbol_ref "TARGET_OPT_AGU") (const_int 0)))
-             (const_string "lea")
-            (eq_attr "alternative" "3")
-              (const_string "lea")
-           ; Current assemblers are broken and do not allow @GOTOFF in
-           ; ought but a memory context.
-           (match_operand:DI 2 "pic_symbolic_operand" "")
-             (const_string "lea")
-           (match_operand:DI 2 "incdec_operand" "")
-             (const_string "incdec")
-          ]
-          (const_string "alu")))
+     (if_then_else (match_operand:HI 2 "incdec_operand" "")
+       (const_string "incdec")
+       (const_string "alu")))
    (set (attr "length_immediate")
       (if_then_else
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "HI")])
 
-;; Convert lea to the lea pattern to avoid flags dependency.
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (plus:DI (match_operand:DI 1 "register_operand" "")
-                (match_operand:DI 2 "x86_64_nonmemory_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed 
-   && ix86_lea_for_add_ok (PLUS, insn, operands)"
-  [(set (match_dup 0)
-       (plus:DI (match_dup 1)
-                (match_dup 2)))]
-  "")
+;; %%% After Dave's SUBREG_BYTE stuff goes in, re-enable incb %ah
+;; type optimizations enabled by define-splits.  This is not important
+;; for PII, and in fact harmful because of partial register stalls.
 
-(define_insn "*adddi_2_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (plus:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
-                  (match_operand:DI 2 "x86_64_general_operand" "rme,re"))
-         (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=r,rm")
-       (plus:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (PLUS, DImode, operands)
-   /* Current assemblers are broken and do not allow @GOTOFF in
-      ought but a memory context.  */
-   && ! pic_symbolic_operand (operands[2], VOIDmode)"
+(define_insn "*addhi_1_lea"
+  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r,r")
+       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0,r")
+                (match_operand:HI 2 "general_operand" "rn,rm,ln")))
+   (clobber (reg:CC FLAGS_REG))]
+  "!TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (PLUS, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
+    case TYPE_LEA:
+      return "#";
     case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-        return "inc{q}\t%0";
+       return "inc{w}\t%0";
       else
-        {
+       {
          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{q}\t%0";
+         return "dec{w}\t%0";
        }
 
     default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* ???? We ought to handle there the 32bit case too
-        - do we need new constraint?  */
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-         /* Avoid overflows.  */
-         && ((INTVAL (operands[2]) & ((((unsigned int) 1) << 31) - 1)))
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{q}\t{%2, %0|%0, %2}";
-        }
-      return "add{q}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], HImode))
+       return "sub{w}\t{%2, %0|%0, %2}";
+
+      return "add{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:DI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
+     (if_then_else (eq_attr "alternative" "2")
+       (const_string "lea")
+       (if_then_else (match_operand:HI 2 "incdec_operand" "")
+          (const_string "incdec")
+          (const_string "alu"))))
    (set (attr "length_immediate")
       (if_then_else
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "HI,HI,SI")])
 
-(define_insn "*adddi_3_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (neg:DI (match_operand:DI 2 "x86_64_general_operand" "rme"))
-                (match_operand:DI 1 "x86_64_general_operand" "%0")))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && ix86_match_ccmode (insn, CCZmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))
-   /* Current assemblers are broken and do not allow @GOTOFF in
-      ought but a memory context.  */
-   && ! pic_symbolic_operand (operands[2], VOIDmode)"
+(define_insn "*addqi_1"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q,r")
+       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
+                (match_operand:QI 2 "general_operand" "qn,qmn,rn")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (PLUS, QImode, operands)"
 {
+  int widen = (which_alternative == 2);
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-        return "inc{q}\t%0";
+       return widen ? "inc{l}\t%k0" : "inc{b}\t%0";
       else
-        {
+       {
          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{q}\t%0";
+         return widen ? "dec{l}\t%k0" : "dec{b}\t%0";
        }
 
     default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* ???? We ought to handle there the 32bit case too
-        - do we need new constraint?  */
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-         /* Avoid overflows.  */
-         && ((INTVAL (operands[2]) & ((((unsigned int) 1) << 31) - 1)))
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{q}\t{%2, %0|%0, %2}";
-        }
-      return "add{q}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], QImode))
+       {
+         if (widen)
+           return "sub{l}\t{%2, %k0|%k0, %2}";
+         else
+           return "sub{b}\t{%2, %0|%0, %2}";
+       }
+      if (widen)
+        return "add{l}\t{%k2, %k0|%k0, %k2}";
+      else
+        return "add{b}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:DI 2 "incdec_operand" "")
+     (if_then_else (match_operand:QI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "QI,QI,SI")])
 
-; For comparisons against 1, -1 and 128, we may generate better code
-; by converting cmp to add, inc or dec as done by peephole2.  This pattern
-; is matched then.  We can't accept general immediate, because for
-; case of overflows,  the result is messed up.
-; This pattern also don't hold of 0x8000000000000000, since the value overflows
-; when negated.
-; Also carry flag is reversed compared to cmp, so this conversion is valid
-; only for comparisons not depending on it.
-(define_insn "*adddi_4_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:DI 1 "nonimmediate_operand" "0")
-                (match_operand:DI 2 "x86_64_immediate_operand" "e")))
-   (clobber (match_scratch:DI 0 "=rm"))]
-  "TARGET_64BIT
-   &&  ix86_match_ccmode (insn, CCGCmode)"
+;; %%% Potential partial reg stall on alternative 2.  What to do?
+(define_insn "*addqi_1_lea"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q,r,r")
+       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0,r")
+                (match_operand:QI 2 "general_operand" "qn,qmn,rn,ln")))
+   (clobber (reg:CC FLAGS_REG))]
+  "!TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (PLUS, QImode, operands)"
 {
+  int widen = (which_alternative == 2);
   switch (get_attr_type (insn))
     {
+    case TYPE_LEA:
+      return "#";
     case TYPE_INCDEC:
-      if (operands[2] == constm1_rtx)
-        return "inc{q}\t%0";
+      if (operands[2] == const1_rtx)
+       return widen ? "inc{l}\t%k0" : "inc{b}\t%0";
       else
-        {
-         gcc_assert (operands[2] == const1_rtx);
-          return "dec{q}\t%0";
+       {
+         gcc_assert (operands[2] == constm1_rtx);
+         return widen ? "dec{l}\t%k0" : "dec{b}\t%0";
        }
 
     default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if ((INTVAL (operands[2]) == -128
-          || (INTVAL (operands[2]) > 0
-              && INTVAL (operands[2]) != 128))
-         /* Avoid overflows.  */
-         && ((INTVAL (operands[2]) & ((((unsigned int) 1) << 31) - 1))))
-       return "sub{q}\t{%2, %0|%0, %2}";
-      operands[2] = GEN_INT (-INTVAL (operands[2]));
-      return "add{q}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (if_then_else (match_operand:DI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
-   (set (attr "length_immediate")
-      (if_then_else
-       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
-       (const_string "1")
-       (const_string "*")))
-   (set_attr "mode" "DI")])
-
-(define_insn "*adddi_5_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (plus:DI (match_operand:DI 1 "nonimmediate_operand" "%0")
-                  (match_operand:DI 2 "x86_64_general_operand" "rme"))
-         (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))
-   /* Current assemblers are broken and do not allow @GOTOFF in
-      ought but a memory context.  */
-   && ! pic_symbolic_operand (operands[2], VOIDmode)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      if (operands[2] == const1_rtx)
-        return "inc{q}\t%0";
-      else
-        {
-          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{q}\t%0";
+      if (x86_maybe_negate_const_int (&operands[2], QImode))
+       {
+         if (widen)
+           return "sub{l}\t{%2, %k0|%k0, %2}";
+         else
+           return "sub{b}\t{%2, %0|%0, %2}";
        }
-
-    default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-         /* Avoid overflows.  */
-         && ((INTVAL (operands[2]) & ((((unsigned int) 1) << 31) - 1)))
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{q}\t{%2, %0|%0, %2}";
-        }
-      return "add{q}\t{%2, %0|%0, %2}";
+      if (widen)
+        return "add{l}\t{%k2, %k0|%k0, %k2}";
+      else
+        return "add{b}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:DI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
+     (if_then_else (eq_attr "alternative" "3")
+       (const_string "lea")
+       (if_then_else (match_operand:QI 2 "incdec_operand" "")
+          (const_string "incdec")
+          (const_string "alu"))))
    (set (attr "length_immediate")
       (if_then_else
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "DI")])
-
+   (set_attr "mode" "QI,QI,SI,SI")])
 
-(define_insn "*addsi_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=r,rm,r,r")
-       (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0,r,r")
-                (match_operand:SI 2 "general_operand" "g,ri,0,li")))
+(define_insn "*addqi_1_slp"
+  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,q"))
+       (plus:QI (match_dup 0)
+                (match_operand:QI 1 "general_operand" "qn,qnm")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (PLUS, SImode, operands)"
+  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
 {
   switch (get_attr_type (insn))
     {
-    case TYPE_LEA:
-      operands[2] = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
-      return "lea{l}\t{%a2, %0|%0, %a2}";
-
     case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      if (operands[2] == const1_rtx)
-        return "inc{l}\t%0";
+      if (operands[1] == const1_rtx)
+       return "inc{b}\t%0";
       else
        {
-         gcc_assert (operands[2] == constm1_rtx);
-          return "dec{l}\t%0";
+         gcc_assert (operands[1] == constm1_rtx);
+         return "dec{b}\t%0";
        }
 
     default:
-      /* Use add as much as possible to replace lea for AGU optimization. */
-      if (which_alternative == 2 && TARGET_OPT_AGU)
-        return "add{l}\t{%1, %0|%0, %1}";
-
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %0|%0, %2}";
-        }
-      return "add{l}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (eq_attr "alternative" "2") 
-                 (eq (symbol_ref "TARGET_OPT_AGU") (const_int 0)))
-               (const_string "lea")
-            (eq_attr "alternative" "3")
-             (const_string "lea")
-           ; Current assemblers are broken and do not allow @GOTOFF in
-           ; ought but a memory context.
-           (match_operand:SI 2 "pic_symbolic_operand" "")
-             (const_string "lea")
-           (match_operand:SI 2 "incdec_operand" "")
-             (const_string "incdec")
-          ]
-          (const_string "alu")))
-   (set (attr "length_immediate")
-      (if_then_else
-       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
-       (const_string "1")
-       (const_string "*")))
-   (set_attr "mode" "SI")])
-
-;; Convert lea to the lea pattern to avoid flags dependency.
-(define_split
-  [(set (match_operand 0 "register_operand" "")
-       (plus (match_operand 1 "register_operand" "")
-              (match_operand 2 "nonmemory_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "reload_completed && ix86_lea_for_add_ok (PLUS, insn, operands)" 
-  [(const_int 0)]
-{
-  rtx pat;
-  /* In -fPIC mode the constructs like (const (unspec [symbol_ref]))
-     may confuse gen_lowpart.  */
-  if (GET_MODE (operands[0]) != Pmode)
-    {
-      operands[1] = gen_lowpart (Pmode, operands[1]);
-      operands[2] = gen_lowpart (Pmode, operands[2]);
-    }
-  operands[0] = gen_lowpart (SImode, operands[0]);
-  pat = gen_rtx_PLUS (Pmode, operands[1], operands[2]);
-  if (Pmode != SImode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
-  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
-  DONE;
-})
-
-;; It may seem that nonimmediate operand is proper one for operand 1.
-;; The addsi_1 pattern allows nonimmediate operand at that place and
-;; we take care in ix86_binary_operator_ok to not allow two memory
-;; operands so proper swapping will be done in reload.  This allow
-;; patterns constructed from addsi_1 to match.
-(define_insn "addsi_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (zero_extend:DI
-         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,r")
-                  (match_operand:SI 2 "general_operand" "g,li"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, SImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_LEA:
-      operands[2] = SET_SRC (XVECEXP (PATTERN (insn), 0, 0));
-      return "lea{l}\t{%a2, %k0|%k0, %a2}";
-
-    case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-        return "inc{l}\t%k0";
-      else
-        {
-         gcc_assert (operands[2] == constm1_rtx);
-          return "dec{l}\t%k0";
-       }
+      if (x86_maybe_negate_const_int (&operands[1], QImode))
+       return "sub{b}\t{%1, %0|%0, %1}";
 
-    default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %k0|%k0, %2}";
-        }
-      return "add{l}\t{%2, %k0|%k0, %2}";
+      return "add{b}\t{%1, %0|%0, %1}";
     }
 }
   [(set (attr "type")
-     (cond [(eq_attr "alternative" "1")
-             (const_string "lea")
-           ; Current assemblers are broken and do not allow @GOTOFF in
-           ; ought but a memory context.
-           (match_operand:SI 2 "pic_symbolic_operand" "")
-             (const_string "lea")
-           (match_operand:SI 2 "incdec_operand" "")
-             (const_string "incdec")
-          ]
-          (const_string "alu")))
-   (set (attr "length_immediate")
-      (if_then_else
-       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
-       (const_string "1")
-       (const_string "*")))
-   (set_attr "mode" "SI")])
-
-;; Convert lea to the lea pattern to avoid flags dependency.
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (zero_extend:DI
-         (plus:SI (match_operand:SI 1 "register_operand" "")
-                  (match_operand:SI 2 "nonmemory_operand" ""))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed
-   && true_regnum (operands[0]) != true_regnum (operands[1])"
-  [(set (match_dup 0)
-       (zero_extend:DI (subreg:SI (plus:DI (match_dup 1) (match_dup 2)) 0)))]
-{
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[2] = gen_lowpart (Pmode, operands[2]);
-})
+     (if_then_else (match_operand:QI 1 "incdec_operand" "")
+       (const_string "incdec")
+       (const_string "alu1")))
+   (set (attr "memory")
+     (if_then_else (match_operand 1 "memory_operand" "")
+        (const_string "load")
+        (const_string "none")))
+   (set_attr "mode" "QI")])
 
-(define_insn "*addsi_2"
+(define_insn "*add<mode>_2"
   [(set (reg FLAGS_REG)
        (compare
-         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
-                  (match_operand:SI 2 "general_operand" "g,ri"))
+         (plus:SWI48
+           (match_operand:SWI48 1 "nonimmediate_operand" "%0,0")
+           (match_operand:SWI48 2 "<general_operand>" "<g>,r<i>"))
          (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=r,rm")
-       (plus:SI (match_dup 1) (match_dup 2)))]
+   (set (match_operand:SWI48 0 "nonimmediate_operand" "=r,rm")
+       (plus:SWI48 (match_dup 1) (match_dup 2)))]
   "ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (PLUS, SImode, operands)
+   && ix86_binary_operator_ok (PLUS, <MODE>mode, operands)
    /* Current assemblers are broken and do not allow @GOTOFF in
       ought but a memory context.  */
    && ! pic_symbolic_operand (operands[2], VOIDmode)"
     case TYPE_INCDEC:
       gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-        return "inc{l}\t%0";
+        return "inc{<imodesuffix>}\t%0";
       else
         {
          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{l}\t%0";
+          return "dec{<imodesuffix>}\t%0";
        }
 
     default:
       gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %0|%0, %2}";
-        }
-      return "add{l}\t{%2, %0|%0, %2}";
+      /* ???? In DImode, we ought to handle there the 32bit case too
+        - do we need new constraint?  */
+      if (x86_maybe_negate_const_int (&operands[2], <MODE>mode))
+        return "sub{<imodesuffix>}\t{%2, %0|%0, %2}";
+
+      return "add{<imodesuffix>}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:SI 2 "incdec_operand" "")
+     (if_then_else (match_operand:SWI48 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 ;; See comment for addsi_1_zext why we do use nonimmediate_operand
 (define_insn "*addsi_2_zext"
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %k0|%k0, %2}";
-        }
+      if (x86_maybe_negate_const_int (&operands[2], SImode))
+        return "sub{l}\t{%2, %k0|%k0, %2}";
+
       return "add{l}\t{%2, %k0|%k0, %2}";
     }
 }
        (const_string "*")))
    (set_attr "mode" "SI")])
 
-(define_insn "*addsi_3"
+(define_insn "*addhi_2"
   [(set (reg FLAGS_REG)
-       (compare (neg:SI (match_operand:SI 2 "general_operand" "g"))
-                (match_operand:SI 1 "nonimmediate_operand" "%0")))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCZmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))
-   /* Current assemblers are broken and do not allow @GOTOFF in
-      ought but a memory context.  */
-   && ! pic_symbolic_operand (operands[2], VOIDmode)"
+       (compare
+         (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
+                  (match_operand:HI 2 "general_operand" "rmn,rn"))
+         (const_int 0)))
+   (set (match_operand:HI 0 "nonimmediate_operand" "=r,rm")
+       (plus:HI (match_dup 1) (match_dup 2)))]
+  "ix86_match_ccmode (insn, CCGOCmode)
+   && ix86_binary_operator_ok (PLUS, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-        return "inc{l}\t%0";
+       return "inc{w}\t%0";
       else
         {
          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{l}\t%0";
+         return "dec{w}\t%0";
        }
 
     default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %0|%0, %2}";
-        }
-      return "add{l}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], HImode))
+       return "sub{w}\t{%2, %0|%0, %2}";
+
+      return "add{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:SI 2 "incdec_operand" "")
+     (if_then_else (match_operand:HI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "HI")])
 
-;; See comment for addsi_1_zext why we do use nonimmediate_operand
-(define_insn "*addsi_3_zext"
+(define_insn "*addqi_2"
   [(set (reg FLAGS_REG)
-       (compare (neg:SI (match_operand:SI 2 "general_operand" "g"))
-                (match_operand:SI 1 "nonimmediate_operand" "%0")))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (plus:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCZmode)
-   && ix86_binary_operator_ok (PLUS, SImode, operands)
-   /* Current assemblers are broken and do not allow @GOTOFF in
-      ought but a memory context.  */
-   && ! pic_symbolic_operand (operands[2], VOIDmode)"
+       (compare
+         (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0")
+                  (match_operand:QI 2 "general_operand" "qmn,qn"))
+         (const_int 0)))
+   (set (match_operand:QI 0 "nonimmediate_operand" "=q,qm")
+       (plus:QI (match_dup 1) (match_dup 2)))]
+  "ix86_match_ccmode (insn, CCGOCmode)
+   && ix86_binary_operator_ok (PLUS, QImode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
       if (operands[2] == const1_rtx)
-        return "inc{l}\t%k0";
+       return "inc{b}\t%0";
       else
         {
-         gcc_assert (operands[2] == constm1_rtx);
-          return "dec{l}\t%k0";
+         gcc_assert (operands[2] == constm1_rtx
+                     || (CONST_INT_P (operands[2])
+                         && INTVAL (operands[2]) == 255));
+         return "dec{b}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %k0|%k0, %2}";
-        }
-      return "add{l}\t{%2, %k0|%k0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], QImode))
+       return "sub{b}\t{%2, %0|%0, %2}";
+
+      return "add{b}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:SI 2 "incdec_operand" "")
+     (if_then_else (match_operand:QI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
-   (set (attr "length_immediate")
-      (if_then_else
-       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
-       (const_string "1")
-       (const_string "*")))
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "QI")])
 
-; For comparisons against 1, -1 and 128, we may generate better code
-; by converting cmp to add, inc or dec as done by peephole2.  This pattern
-; is matched then.  We can't accept general immediate, because for
-; case of overflows,  the result is messed up.
-; This pattern also don't hold of 0x80000000, since the value overflows
-; when negated.
-; Also carry flag is reversed compared to cmp, so this conversion is valid
-; only for comparisons not depending on it.
-(define_insn "*addsi_4"
+(define_insn "*add<mode>_3"
   [(set (reg FLAGS_REG)
-       (compare (match_operand:SI 1 "nonimmediate_operand" "0")
-                (match_operand:SI 2 "const_int_operand" "n")))
-   (clobber (match_scratch:SI 0 "=rm"))]
-  "ix86_match_ccmode (insn, CCGCmode)
-   && (INTVAL (operands[2]) & 0xffffffff) != 0x80000000"
+       (compare
+         (neg:SWI48 (match_operand:SWI48 2 "<general_operand>" "<g>"))
+         (match_operand:SWI48 1 "nonimmediate_operand" "%0")))
+   (clobber (match_scratch:SWI48 0 "=r"))]
+  "ix86_match_ccmode (insn, CCZmode)
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))
+   /* Current assemblers are broken and do not allow @GOTOFF in
+      ought but a memory context.  */
+   && ! pic_symbolic_operand (operands[2], VOIDmode)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      if (operands[2] == constm1_rtx)
-        return "inc{l}\t%0";
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (operands[2] == const1_rtx)
+        return "inc{<imodesuffix>}\t%0";
       else
         {
-         gcc_assert (operands[2] == const1_rtx);
-          return "dec{l}\t%0";
+         gcc_assert (operands[2] == constm1_rtx);
+          return "dec{<imodesuffix>}\t%0";
        }
 
     default:
       gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if ((INTVAL (operands[2]) == -128
-          || (INTVAL (operands[2]) > 0
-              && INTVAL (operands[2]) != 128)))
-       return "sub{l}\t{%2, %0|%0, %2}";
-      operands[2] = GEN_INT (-INTVAL (operands[2]));
-      return "add{l}\t{%2, %0|%0, %2}";
+      /* ???? In DImode, we ought to handle there the 32bit case too
+        - do we need new constraint?  */
+      if (x86_maybe_negate_const_int (&operands[2], <MODE>mode))
+        return "sub{<imodesuffix>}\t{%2, %0|%0, %2}";
+
+      return "add{<imodesuffix>}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:SI 2 "incdec_operand" "")
+     (if_then_else (match_operand:SWI48 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*addsi_5"
+;; See comment for addsi_1_zext why we do use nonimmediate_operand
+(define_insn "*addsi_3_zext"
   [(set (reg FLAGS_REG)
        (compare
-         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                  (match_operand:SI 2 "general_operand" "g"))
-         (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))
+         (neg:SI (match_operand:SI 2 "general_operand" "g"))
+         (match_operand:SI 1 "nonimmediate_operand" "%0")))
+   (set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI (plus:SI (match_dup 1) (match_dup 2))))]
+  "TARGET_64BIT && ix86_match_ccmode (insn, CCZmode)
+   && ix86_binary_operator_ok (PLUS, SImode, operands)
    /* Current assemblers are broken and do not allow @GOTOFF in
       ought but a memory context.  */
    && ! pic_symbolic_operand (operands[2], VOIDmode)"
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-        return "inc{l}\t%0";
+        return "inc{l}\t%k0";
       else
         {
          gcc_assert (operands[2] == constm1_rtx);
-          return "dec{l}\t%0";
+          return "dec{l}\t%k0";
        }
 
     default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "sub{l}\t{%2, %0|%0, %2}";
-        }
-      return "add{l}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], SImode))
+        return "sub{l}\t{%2, %k0|%k0, %2}";
+
+      return "add{l}\t{%2, %k0|%k0, %2}";
     }
 }
   [(set (attr "type")
        (const_string "*")))
    (set_attr "mode" "SI")])
 
-(define_expand "addhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                (match_operand:HI 2 "general_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (PLUS, HImode, operands); DONE;")
-
-;; %%% After Dave's SUBREG_BYTE stuff goes in, re-enable incb %ah
-;; type optimizations enabled by define-splits.  This is not important
-;; for PII, and in fact harmful because of partial register stalls.
-
-(define_insn "*addhi_1_lea"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r,r")
-       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0,r")
-                (match_operand:HI 2 "general_operand" "rn,rm,ln")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (PLUS, HImode, operands)"
+(define_insn "*addhi_3"
+  [(set (reg FLAGS_REG)
+       (compare
+         (neg:HI (match_operand:HI 2 "general_operand" "rmn"))
+         (match_operand:HI 1 "nonimmediate_operand" "%0")))
+   (clobber (match_scratch:HI 0 "=r"))]
+  "ix86_match_ccmode (insn, CCZmode)
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
 {
   switch (get_attr_type (insn))
     {
-    case TYPE_LEA:
-      return "#";
     case TYPE_INCDEC:
       if (operands[2] == const1_rtx)
        return "inc{w}\t%0";
       else
-       {
+        {
          gcc_assert (operands[2] == constm1_rtx);
          return "dec{w}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{w}\t{%2, %0|%0, %2}";
-       }
+      if (x86_maybe_negate_const_int (&operands[2], HImode))
+       return "sub{w}\t{%2, %0|%0, %2}";
+
       return "add{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (eq_attr "alternative" "2")
-       (const_string "lea")
-       (if_then_else (match_operand:HI 2 "incdec_operand" "")
-          (const_string "incdec")
-          (const_string "alu"))))
+     (if_then_else (match_operand:HI 2 "incdec_operand" "")
+       (const_string "incdec")
+       (const_string "alu")))
    (set (attr "length_immediate")
       (if_then_else
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "HI,HI,SI")])
+   (set_attr "mode" "HI")])
 
-(define_insn "*addhi_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-                (match_operand:HI 2 "general_operand" "rn,rm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (PLUS, HImode, operands)"
+(define_insn "*addqi_3"
+  [(set (reg FLAGS_REG)
+       (compare
+         (neg:QI (match_operand:QI 2 "general_operand" "qmn"))
+         (match_operand:QI 1 "nonimmediate_operand" "%0")))
+   (clobber (match_scratch:QI 0 "=q"))]
+  "ix86_match_ccmode (insn, CCZmode)
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
       if (operands[2] == const1_rtx)
-       return "inc{w}\t%0";
+       return "inc{b}\t%0";
       else
         {
-         gcc_assert (operands[2] == constm1_rtx);
-         return "dec{w}\t%0";
+         gcc_assert (operands[2] == constm1_rtx
+                     || (CONST_INT_P (operands[2])
+                         && INTVAL (operands[2]) == 255));
+         return "dec{b}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{w}\t{%2, %0|%0, %2}";
-       }
-      return "add{w}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], QImode))
+       return "sub{b}\t{%2, %0|%0, %2}";
+
+      return "add{b}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:HI 2 "incdec_operand" "")
+     (if_then_else (match_operand:QI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
-   (set (attr "length_immediate")
-      (if_then_else
-       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
-       (const_string "1")
-       (const_string "*")))
-   (set_attr "mode" "HI")])
+   (set_attr "mode" "QI")])
 
-(define_insn "*addhi_2"
+; For comparisons against 1, -1 and 128, we may generate better code
+; by converting cmp to add, inc or dec as done by peephole2.  This pattern
+; is matched then.  We can't accept general immediate, because for
+; case of overflows,  the result is messed up.
+; Also carry flag is reversed compared to cmp, so this conversion is valid
+; only for comparisons not depending on it.
+
+(define_insn "*adddi_4"
   [(set (reg FLAGS_REG)
        (compare
-         (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-                  (match_operand:HI 2 "general_operand" "rmn,rn"))
-         (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=r,rm")
-       (plus:HI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (PLUS, HImode, operands)"
+         (match_operand:DI 1 "nonimmediate_operand" "0")
+         (match_operand:DI 2 "x86_64_immediate_operand" "e")))
+   (clobber (match_scratch:DI 0 "=rm"))]
+  "TARGET_64BIT
+   && ix86_match_ccmode (insn, CCGCmode)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{w}\t%0";
+      if (operands[2] == constm1_rtx)
+        return "inc{q}\t%0";
       else
         {
-         gcc_assert (operands[2] == constm1_rtx);
-         return "dec{w}\t%0";
+         gcc_assert (operands[2] == const1_rtx);
+          return "dec{q}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{w}\t{%2, %0|%0, %2}";
-       }
-      return "add{w}\t{%2, %0|%0, %2}";
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (x86_maybe_negate_const_int (&operands[2], DImode))
+       return "add{q}\t{%2, %0|%0, %2}";
+
+      return "sub{q}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:HI 2 "incdec_operand" "")
+     (if_then_else (match_operand:DI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "HI")])
+   (set_attr "mode" "DI")])
 
-(define_insn "*addhi_3"
+; For comparisons against 1, -1 and 128, we may generate better code
+; by converting cmp to add, inc or dec as done by peephole2.  This pattern
+; is matched then.  We can't accept general immediate, because for
+; case of overflows,  the result is messed up.
+; Also carry flag is reversed compared to cmp, so this conversion is valid
+; only for comparisons not depending on it.
+
+(define_insn "*addsi_4"
   [(set (reg FLAGS_REG)
-       (compare (neg:HI (match_operand:HI 2 "general_operand" "rmn"))
-                (match_operand:HI 1 "nonimmediate_operand" "%0")))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCZmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
+       (compare
+         (match_operand:SI 1 "nonimmediate_operand" "0")
+         (match_operand:SI 2 "const_int_operand" "n")))
+   (clobber (match_scratch:SI 0 "=rm"))]
+  "ix86_match_ccmode (insn, CCGCmode)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{w}\t%0";
+      if (operands[2] == constm1_rtx)
+        return "inc{l}\t%0";
       else
         {
-         gcc_assert (operands[2] == constm1_rtx);
-         return "dec{w}\t%0";
+         gcc_assert (operands[2] == const1_rtx);
+          return "dec{l}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{w}\t{%2, %0|%0, %2}";
-       }
-      return "add{w}\t{%2, %0|%0, %2}";
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (x86_maybe_negate_const_int (&operands[2], SImode))
+       return "add{l}\t{%2, %0|%0, %2}";
+
+      return "sub{l}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:HI 2 "incdec_operand" "")
+     (if_then_else (match_operand:SI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "HI")])
+   (set_attr "mode" "SI")])
 
 ; See comments above addsi_4 for details.
+
 (define_insn "*addhi_4"
   [(set (reg FLAGS_REG)
-       (compare (match_operand:HI 1 "nonimmediate_operand" "0")
-                (match_operand:HI 2 "const_int_operand" "n")))
+       (compare
+         (match_operand:HI 1 "nonimmediate_operand" "0")
+         (match_operand:HI 2 "const_int_operand" "n")))
    (clobber (match_scratch:HI 0 "=rm"))]
-  "ix86_match_ccmode (insn, CCGCmode)
-   && (INTVAL (operands[2]) & 0xffff) != 0x8000"
+  "ix86_match_ccmode (insn, CCGCmode)"
 {
   switch (get_attr_type (insn))
     {
 
     default:
       gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if ((INTVAL (operands[2]) == -128
-          || (INTVAL (operands[2]) > 0
-              && INTVAL (operands[2]) != 128)))
-       return "sub{w}\t{%2, %0|%0, %2}";
-      operands[2] = GEN_INT (-INTVAL (operands[2]));
-      return "add{w}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], HImode))
+       return "add{w}\t{%2, %0|%0, %2}";
+
+      return "sub{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
        (const_string "*")))
    (set_attr "mode" "HI")])
 
+; See comments above addsi_4 for details.
 
-(define_insn "*addhi_5"
+(define_insn "*addqi_4"
   [(set (reg FLAGS_REG)
        (compare
-         (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0")
-                  (match_operand:HI 2 "general_operand" "rmn"))
-         (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
+         (match_operand:QI 1 "nonimmediate_operand" "0")
+         (match_operand:QI 2 "const_int_operand" "n")))
+   (clobber (match_scratch:QI 0 "=qm"))]
+  "ix86_match_ccmode (insn, CCGCmode)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{w}\t%0";
+      if (operands[2] == constm1_rtx
+         || (CONST_INT_P (operands[2])
+             && INTVAL (operands[2]) == 255))
+        return "inc{b}\t%0";
       else
        {
-         gcc_assert (operands[2] == constm1_rtx);
-         return "dec{w}\t%0";
+         gcc_assert (operands[2] == const1_rtx);
+          return "dec{b}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{w}\t{%2, %0|%0, %2}";
-       }
-      return "add{w}\t{%2, %0|%0, %2}";
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (x86_maybe_negate_const_int (&operands[2], QImode))
+       return "add{b}\t{%2, %0|%0, %2}";
+
+      return "sub{b}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
      (if_then_else (match_operand:HI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
-   (set (attr "length_immediate")
-      (if_then_else
-       (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
-       (const_string "1")
-       (const_string "*")))
-   (set_attr "mode" "HI")])
-
-(define_expand "addqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                (match_operand:QI 2 "general_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (PLUS, QImode, operands); DONE;")
+   (set_attr "mode" "QI")])
 
-;; %%% Potential partial reg stall on alternative 2.  What to do?
-(define_insn "*addqi_1_lea"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q,r,r")
-       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0,r")
-                (match_operand:QI 2 "general_operand" "qn,qmn,rn,ln")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (PLUS, QImode, operands)"
+(define_insn "*add<mode>_5"
+  [(set (reg FLAGS_REG)
+       (compare
+         (plus:SWI48
+           (match_operand:SWI48 1 "nonimmediate_operand" "%0")
+           (match_operand:SWI48 2 "<general_operand>" "<g>"))
+         (const_int 0)))
+   (clobber (match_scratch:SWI48 0 "=r"))]
+  "ix86_match_ccmode (insn, CCGOCmode)
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))
+   /* Current assemblers are broken and do not allow @GOTOFF in
+      ought but a memory context.  */
+   && ! pic_symbolic_operand (operands[2], VOIDmode)"
 {
-  int widen = (which_alternative == 2);
   switch (get_attr_type (insn))
     {
-    case TYPE_LEA:
-      return "#";
     case TYPE_INCDEC:
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
       if (operands[2] == const1_rtx)
-       return widen ? "inc{l}\t%k0" : "inc{b}\t%0";
+        return "inc{<imodesuffix>}\t%0";
       else
-       {
-         gcc_assert (operands[2] == constm1_rtx);
-         return widen ? "dec{l}\t%k0" : "dec{b}\t%0";
+        {
+          gcc_assert (operands[2] == constm1_rtx);
+          return "dec{<imodesuffix>}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         if (widen)
-           return "sub{l}\t{%2, %k0|%k0, %2}";
-         else
-           return "sub{b}\t{%2, %0|%0, %2}";
-       }
-      if (widen)
-        return "add{l}\t{%k2, %k0|%k0, %k2}";
-      else
-        return "add{b}\t{%2, %0|%0, %2}";
+      gcc_assert (rtx_equal_p (operands[0], operands[1]));
+      if (x86_maybe_negate_const_int (&operands[2], <MODE>mode))
+        return "sub{<imodesuffix>}\t{%2, %0|%0, %2}";
+
+      return "add{<imodesuffix>}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (eq_attr "alternative" "3")
-       (const_string "lea")
-       (if_then_else (match_operand:QI 2 "incdec_operand" "")
-          (const_string "incdec")
-          (const_string "alu"))))
+     (if_then_else (match_operand:SWI48 2 "incdec_operand" "")
+       (const_string "incdec")
+       (const_string "alu")))
    (set (attr "length_immediate")
       (if_then_else
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "QI,QI,SI,SI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*addqi_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q,r")
-       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
-                (match_operand:QI 2 "general_operand" "qn,qmn,rn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (PLUS, QImode, operands)"
+(define_insn "*addhi_5"
+  [(set (reg FLAGS_REG)
+       (compare
+         (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0")
+                  (match_operand:HI 2 "general_operand" "rmn"))
+         (const_int 0)))
+   (clobber (match_scratch:HI 0 "=r"))]
+  "ix86_match_ccmode (insn, CCGOCmode)
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
 {
-  int widen = (which_alternative == 2);
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
       if (operands[2] == const1_rtx)
-       return widen ? "inc{l}\t%k0" : "inc{b}\t%0";
+       return "inc{w}\t%0";
       else
        {
          gcc_assert (operands[2] == constm1_rtx);
-         return widen ? "dec{l}\t%k0" : "dec{b}\t%0";
+         return "dec{w}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.
-        Exceptions: -128 encodes smaller than 128, so swap sign and op.  */
-      if (CONST_INT_P (operands[2])
-          && (INTVAL (operands[2]) == 128
-             || (INTVAL (operands[2]) < 0
-                 && INTVAL (operands[2]) != -128)))
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         if (widen)
-           return "sub{l}\t{%2, %k0|%k0, %2}";
-         else
-           return "sub{b}\t{%2, %0|%0, %2}";
-       }
-      if (widen)
-        return "add{l}\t{%k2, %k0|%k0, %k2}";
-      else
-        return "add{b}\t{%2, %0|%0, %2}";
+      if (x86_maybe_negate_const_int (&operands[2], HImode))
+       return "sub{w}\t{%2, %0|%0, %2}";
+
+      return "add{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:QI 2 "incdec_operand" "")
+     (if_then_else (match_operand:HI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
    (set (attr "length_immediate")
        (and (eq_attr "type" "alu") (match_operand 2 "const128_operand" ""))
        (const_string "1")
        (const_string "*")))
-   (set_attr "mode" "QI,QI,SI")])
+   (set_attr "mode" "HI")])
 
-(define_insn "*addqi_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,q"))
-       (plus:QI (match_dup 0)
-                (match_operand:QI 1 "general_operand" "qn,qnm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+(define_insn "*addqi_5"
+  [(set (reg FLAGS_REG)
+       (compare
+         (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0")
+                  (match_operand:QI 2 "general_operand" "qmn"))
+         (const_int 0)))
+   (clobber (match_scratch:QI 0 "=q"))]
+  "ix86_match_ccmode (insn, CCGOCmode)
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
-      if (operands[1] == const1_rtx)
+      if (operands[2] == const1_rtx)
        return "inc{b}\t%0";
       else
-       {
-         gcc_assert (operands[1] == constm1_rtx);
+        {
+         gcc_assert (operands[2] == constm1_rtx
+                     || (CONST_INT_P (operands[2])
+                         && INTVAL (operands[2]) == 255));
          return "dec{b}\t%0";
        }
 
     default:
-      /* Make things pretty and `subl $4,%eax' rather than `addl $-4, %eax'.  */
-      if (CONST_INT_P (operands[1])
-         && INTVAL (operands[1]) < 0)
-       {
-         operands[1] = GEN_INT (-INTVAL (operands[1]));
-         return "sub{b}\t{%1, %0|%0, %1}";
-       }
-      return "add{b}\t{%1, %0|%0, %1}";
+      if (x86_maybe_negate_const_int (&operands[2], QImode))
+       return "sub{b}\t{%2, %0|%0, %2}";
+
+      return "add{b}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
-     (if_then_else (match_operand:QI 1 "incdec_operand" "")
+     (if_then_else (match_operand:QI 2 "incdec_operand" "")
        (const_string "incdec")
-       (const_string "alu1")))
-   (set (attr "memory")
-     (if_then_else (match_operand 1 "memory_operand" "")
-        (const_string "load")
-        (const_string "none")))
+       (const_string "alu")))
    (set_attr "mode" "QI")])
 
-(define_insn "*addqi_2"
-  [(set (reg FLAGS_REG)
-       (compare
-         (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0")
-                  (match_operand:QI 2 "general_operand" "qmn,qn"))
-         (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=q,qm")
-       (plus:QI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (PLUS, QImode, operands)"
+(define_insn "*addqi_ext_1_rex64"
+  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
+                        (const_int 8)
+                        (const_int 8))
+       (plus:SI
+         (zero_extract:SI
+           (match_operand 1 "ext_register_operand" "0")
+           (const_int 8)
+           (const_int 8))
+         (match_operand:QI 2 "nonmemory_operand" "Qn")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
       if (operands[2] == const1_rtx)
-       return "inc{b}\t%0";
+       return "inc{b}\t%h0";
       else
         {
          gcc_assert (operands[2] == constm1_rtx
                      || (CONST_INT_P (operands[2])
-                         && INTVAL (operands[2]) == 255));
-         return "dec{b}\t%0";
-       }
+                         && INTVAL (operands[2]) == 255));
+          return "dec{b}\t%h0";
+        }
 
     default:
-      /* Make things pretty and `subb $4,%al' rather than `addb $-4, %al'.  */
-      if (CONST_INT_P (operands[2])
-          && INTVAL (operands[2]) < 0)
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{b}\t{%2, %0|%0, %2}";
-       }
-      return "add{b}\t{%2, %0|%0, %2}";
+      return "add{b}\t{%2, %h0|%h0, %2}";
     }
 }
   [(set (attr "type")
      (if_then_else (match_operand:QI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
+   (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
 
-(define_insn "*addqi_3"
-  [(set (reg FLAGS_REG)
-       (compare (neg:QI (match_operand:QI 2 "general_operand" "qmn"))
-                (match_operand:QI 1 "nonimmediate_operand" "%0")))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "ix86_match_ccmode (insn, CCZmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
+(define_insn "addqi_ext_1"
+  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
+                        (const_int 8)
+                        (const_int 8))
+       (plus:SI
+         (zero_extract:SI
+           (match_operand 1 "ext_register_operand" "0")
+           (const_int 8)
+           (const_int 8))
+         (match_operand:QI 2 "general_operand" "Qmn")))
+   (clobber (reg:CC FLAGS_REG))]
+  "!TARGET_64BIT"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_INCDEC:
       if (operands[2] == const1_rtx)
-       return "inc{b}\t%0";
+       return "inc{b}\t%h0";
       else
         {
          gcc_assert (operands[2] == constm1_rtx
                      || (CONST_INT_P (operands[2])
                          && INTVAL (operands[2]) == 255));
-         return "dec{b}\t%0";
+          return "dec{b}\t%h0";
        }
 
     default:
-      /* Make things pretty and `subb $4,%al' rather than `addb $-4, %al'.  */
-      if (CONST_INT_P (operands[2])
-          && INTVAL (operands[2]) < 0)
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{b}\t{%2, %0|%0, %2}";
-       }
-      return "add{b}\t{%2, %0|%0, %2}";
+      return "add{b}\t{%2, %h0|%h0, %2}";
     }
 }
   [(set (attr "type")
      (if_then_else (match_operand:QI 2 "incdec_operand" "")
        (const_string "incdec")
        (const_string "alu")))
-   (set_attr "mode" "QI")])
-
-; See comments above addsi_4 for details.
-(define_insn "*addqi_4"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:QI 1 "nonimmediate_operand" "0")
-                (match_operand:QI 2 "const_int_operand" "n")))
-   (clobber (match_scratch:QI 0 "=qm"))]
-  "ix86_match_ccmode (insn, CCGCmode)
-   && (INTVAL (operands[2]) & 0xff) != 0x80"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_INCDEC:
-      if (operands[2] == constm1_rtx
-         || (CONST_INT_P (operands[2])
-             && INTVAL (operands[2]) == 255))
-        return "inc{b}\t%0";
-      else
-       {
-         gcc_assert (operands[2] == const1_rtx);
-          return "dec{b}\t%0";
-       }
-
-    default:
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      if (INTVAL (operands[2]) < 0)
-        {
-          operands[2] = GEN_INT (-INTVAL (operands[2]));
-          return "add{b}\t{%2, %0|%0, %2}";
-        }
-      return "sub{b}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (if_then_else (match_operand:HI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
-   (set_attr "mode" "QI")])
-
-
-(define_insn "*addqi_5"
-  [(set (reg FLAGS_REG)
-       (compare
-         (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0")
-                  (match_operand:QI 2 "general_operand" "qmn"))
-         (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{b}\t%0";
-      else
-        {
-         gcc_assert (operands[2] == constm1_rtx
-                     || (CONST_INT_P (operands[2])
-                         && INTVAL (operands[2]) == 255));
-         return "dec{b}\t%0";
-       }
-
-    default:
-      /* Make things pretty and `subb $4,%al' rather than `addb $-4, %al'.  */
-      if (CONST_INT_P (operands[2])
-          && INTVAL (operands[2]) < 0)
-       {
-         operands[2] = GEN_INT (-INTVAL (operands[2]));
-         return "sub{b}\t{%2, %0|%0, %2}";
-       }
-      return "add{b}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (if_then_else (match_operand:QI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
-   (set_attr "mode" "QI")])
-
-
-(define_insn "addqi_ext_1"
-  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
-                        (const_int 8)
-                        (const_int 8))
-       (plus:SI
-         (zero_extract:SI
-           (match_operand 1 "ext_register_operand" "0")
-           (const_int 8)
-           (const_int 8))
-         (match_operand:QI 2 "general_operand" "Qmn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{b}\t%h0";
-      else
-        {
-         gcc_assert (operands[2] == constm1_rtx
-                     || (CONST_INT_P (operands[2])
-                         && INTVAL (operands[2]) == 255));
-          return "dec{b}\t%h0";
-       }
-
-    default:
-      return "add{b}\t{%2, %h0|%h0, %2}";
-    }
-}
-  [(set (attr "type")
-     (if_then_else (match_operand:QI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
-   (set_attr "modrm" "1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*addqi_ext_1_rex64"
-  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
-                        (const_int 8)
-                        (const_int 8))
-       (plus:SI
-         (zero_extract:SI
-           (match_operand 1 "ext_register_operand" "0")
-           (const_int 8)
-           (const_int 8))
-         (match_operand:QI 2 "nonmemory_operand" "Qn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{b}\t%h0";
-      else
-        {
-         gcc_assert (operands[2] == constm1_rtx
-                     || (CONST_INT_P (operands[2])
-                         && INTVAL (operands[2]) == 255));
-          return "dec{b}\t%h0";
-        }
-
-    default:
-      return "add{b}\t{%2, %h0|%h0, %2}";
-    }
-}
-  [(set (attr "type")
-     (if_then_else (match_operand:QI 2 "incdec_operand" "")
-       (const_string "incdec")
-       (const_string "alu")))
-   (set_attr "modrm" "1")
+   (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
 
 (define_insn "*addqi_ext_2"
   [(set_attr "type" "alu")
    (set_attr "mode" "QI")])
 
-;; The patterns that match these are at the end of this file.
-
-(define_expand "addxf3"
-  [(set (match_operand:XF 0 "register_operand" "")
-       (plus:XF (match_operand:XF 1 "register_operand" "")
-                (match_operand:XF 2 "register_operand" "")))]
-  "TARGET_80387"
-  "")
-
-(define_expand "add<mode>3"
-  [(set (match_operand:MODEF 0 "register_operand" "")
-       (plus:MODEF (match_operand:MODEF 1 "register_operand" "")
-                   (match_operand:MODEF 2 "nonimmediate_operand" "")))]
-  "(TARGET_80387 && X87_ENABLE_ARITH (<MODE>mode))
-    || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)"
-  "")
-\f
-;; Subtract instructions
+;; The lea patterns for non-Pmodes needs to be matched by
+;; several insns converted to real lea by splitters.
 
-;; %%% splits for subditi3
+(define_insn_and_split "*lea_general_1"
+  [(set (match_operand 0 "register_operand" "=r")
+       (plus (plus (match_operand 1 "index_register_operand" "l")
+                   (match_operand 2 "register_operand" "r"))
+             (match_operand 3 "immediate_operand" "i")))]
+  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
+    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
+   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && GET_MODE (operands[0]) == GET_MODE (operands[1])
+   && GET_MODE (operands[0]) == GET_MODE (operands[2])
+   && (GET_MODE (operands[0]) == GET_MODE (operands[3])
+       || GET_MODE (operands[3]) == VOIDmode)"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+{
+  rtx pat;
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[2] = gen_lowpart (Pmode, operands[2]);
+  operands[3] = gen_lowpart (Pmode, operands[3]);
+  pat = gen_rtx_PLUS (Pmode, gen_rtx_PLUS (Pmode, operands[1], operands[2]),
+                     operands[3]);
+  if (Pmode != SImode)
+    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
+  DONE;
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
 
-(define_expand "subti3"
-  [(set (match_operand:TI 0 "nonimmediate_operand" "")
-       (minus:TI (match_operand:TI 1 "nonimmediate_operand" "")
-                 (match_operand:TI 2 "x86_64_general_operand" "")))]
+(define_insn_and_split "*lea_general_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (plus:SI (plus:SI
+                    (match_operand:SI 1 "index_register_operand" "l")
+                    (match_operand:SI 2 "register_operand" "r"))
+                  (match_operand:SI 3 "immediate_operand" "i"))))]
   "TARGET_64BIT"
-  "ix86_expand_binary_operator (MINUS, TImode, operands); DONE;")
-
-(define_insn "*subti3_1"
-  [(set (match_operand:TI 0 "nonimmediate_operand" "=r,o")
-       (minus:TI (match_operand:TI 1 "nonimmediate_operand" "0,0")
-                 (match_operand:TI 2 "x86_64_general_operand" "roe,re")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (MINUS, TImode, operands)"
-  "#")
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (subreg:SI (plus:DI (plus:DI (match_dup 1)
+                                                    (match_dup 2))
+                                           (match_dup 3)) 0)))]
+{
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[2] = gen_lowpart (Pmode, operands[2]);
+  operands[3] = gen_lowpart (Pmode, operands[3]);
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
 
-(define_split
-  [(set (match_operand:TI 0 "nonimmediate_operand" "")
-       (minus:TI (match_operand:TI 1 "nonimmediate_operand" "")
-                 (match_operand:TI 2 "x86_64_general_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed"
-  [(parallel [(set (reg:CC FLAGS_REG) (compare:CC (match_dup 1) (match_dup 2)))
-             (set (match_dup 0) (minus:DI (match_dup 1) (match_dup 2)))])
-   (parallel [(set (match_dup 3)
-                  (minus:DI (match_dup 4)
-                            (plus:DI (ltu:DI (reg:CC FLAGS_REG) (const_int 0))
-                                     (match_dup 5))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "split_ti (&operands[0], 3, &operands[0], &operands[3]);")
+(define_insn_and_split "*lea_general_2"
+  [(set (match_operand 0 "register_operand" "=r")
+       (plus (mult (match_operand 1 "index_register_operand" "l")
+                   (match_operand 2 "const248_operand" "i"))
+             (match_operand 3 "nonmemory_operand" "ri")))]
+  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
+    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
+   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && GET_MODE (operands[0]) == GET_MODE (operands[1])
+   && (GET_MODE (operands[0]) == GET_MODE (operands[3])
+       || GET_MODE (operands[3]) == VOIDmode)"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+{
+  rtx pat;
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[3] = gen_lowpart (Pmode, operands[3]);
+  pat = gen_rtx_PLUS (Pmode, gen_rtx_MULT (Pmode, operands[1], operands[2]),
+                     operands[3]);
+  if (Pmode != SImode)
+    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
+  DONE;
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
 
-;; %%% splits for subsidi3
+(define_insn_and_split "*lea_general_2_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (plus:SI (mult:SI
+                    (match_operand:SI 1 "index_register_operand" "l")
+                    (match_operand:SI 2 "const248_operand" "n"))
+                  (match_operand:SI 3 "nonmemory_operand" "ri"))))]
+  "TARGET_64BIT"
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (subreg:SI (plus:DI (mult:DI (match_dup 1)
+                                                    (match_dup 2))
+                                           (match_dup 3)) 0)))]
+{
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[3] = gen_lowpart (Pmode, operands[3]);
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
 
-(define_expand "subdi3"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (minus:DI (match_operand:DI 1 "nonimmediate_operand" "")
-                 (match_operand:DI 2 "x86_64_general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (MINUS, DImode, operands); DONE;")
+(define_insn_and_split "*lea_general_3"
+  [(set (match_operand 0 "register_operand" "=r")
+       (plus (plus (mult (match_operand 1 "index_register_operand" "l")
+                         (match_operand 2 "const248_operand" "i"))
+                   (match_operand 3 "register_operand" "r"))
+             (match_operand 4 "immediate_operand" "i")))]
+  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
+    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
+   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && GET_MODE (operands[0]) == GET_MODE (operands[1])
+   && GET_MODE (operands[0]) == GET_MODE (operands[3])"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+{
+  rtx pat;
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[3] = gen_lowpart (Pmode, operands[3]);
+  operands[4] = gen_lowpart (Pmode, operands[4]);
+  pat = gen_rtx_PLUS (Pmode,
+                     gen_rtx_PLUS (Pmode, gen_rtx_MULT (Pmode, operands[1],
+                                                        operands[2]),
+                                   operands[3]),
+                     operands[4]);
+  if (Pmode != SImode)
+    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
+  DONE;
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
 
-(define_insn "*subdi3_1"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=r,o")
-       (minus:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                 (match_operand:DI 2 "general_operand" "roiF,riF")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && ix86_binary_operator_ok (MINUS, DImode, operands)"
-  "#")
+(define_insn_and_split "*lea_general_3_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (plus:SI (plus:SI
+                    (mult:SI
+                      (match_operand:SI 1 "index_register_operand" "l")
+                      (match_operand:SI 2 "const248_operand" "n"))
+                    (match_operand:SI 3 "register_operand" "r"))
+                  (match_operand:SI 4 "immediate_operand" "i"))))]
+  "TARGET_64BIT"
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (subreg:SI (plus:DI (plus:DI (mult:DI (match_dup 1)
+                                                             (match_dup 2))
+                                                    (match_dup 3))
+                                           (match_dup 4)) 0)))]
+{
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[3] = gen_lowpart (Pmode, operands[3]);
+  operands[4] = gen_lowpart (Pmode, operands[4]);
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
 
+;; Convert lea to the lea pattern to avoid flags dependency.
 (define_split
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (minus:DI (match_operand:DI 1 "nonimmediate_operand" "")
-                 (match_operand:DI 2 "general_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && reload_completed"
-  [(parallel [(set (reg:CC FLAGS_REG) (compare:CC (match_dup 1) (match_dup 2)))
-             (set (match_dup 0) (minus:SI (match_dup 1) (match_dup 2)))])
-   (parallel [(set (match_dup 3)
-                  (minus:SI (match_dup 4)
-                            (plus:SI (ltu:SI (reg:CC FLAGS_REG) (const_int 0))
-                                     (match_dup 5))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "split_di (&operands[0], 3, &operands[0], &operands[3]);")
-
-(define_insn "subdi3_carry_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-         (minus:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-           (plus:DI (match_operand:DI 3 "ix86_carry_flag_operator" "")
-              (match_operand:DI 2 "x86_64_general_operand" "re,rm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (MINUS, DImode, operands)"
-  "sbb{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*subdi_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (minus:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                 (match_operand:DI 2 "x86_64_general_operand" "re,rm")))
+  [(set (match_operand:DI 0 "register_operand" "")
+       (plus:DI (match_operand:DI 1 "register_operand" "")
+                (match_operand:DI 2 "x86_64_nonmemory_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (MINUS, DImode, operands)"
-  "sub{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*subdi_2_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (minus:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                   (match_operand:DI 2 "x86_64_general_operand" "re,rm"))
-         (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (minus:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (MINUS, DImode, operands)"
-  "sub{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*subdi_3_rex63"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                (match_operand:DI 2 "x86_64_general_operand" "re,rm")))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (minus:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCmode)
-   && ix86_binary_operator_ok (MINUS, SImode, operands)"
-  "sub{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
+  "TARGET_64BIT && reload_completed 
+   && ix86_lea_for_add_ok (PLUS, insn, operands)"
+  [(set (match_dup 0)
+       (plus:DI (match_dup 1)
+                (match_dup 2)))]
+  "")
 
-(define_insn "subqi3_carry"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q")
-         (minus:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
-           (plus:QI (match_operand:QI 3 "ix86_carry_flag_operator" "")
-              (match_operand:QI 2 "general_operand" "qn,qm"))))
+;; Convert lea to the lea pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand 0 "register_operand" "")
+       (plus (match_operand 1 "register_operand" "")
+              (match_operand 2 "nonmemory_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (MINUS, QImode, operands)"
-  "sbb{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "QI")])
+  "reload_completed && ix86_lea_for_add_ok (PLUS, insn, operands)" 
+  [(const_int 0)]
+{
+  rtx pat;
+  /* In -fPIC mode the constructs like (const (unspec [symbol_ref]))
+     may confuse gen_lowpart.  */
+  if (GET_MODE (operands[0]) != Pmode)
+    {
+      operands[1] = gen_lowpart (Pmode, operands[1]);
+      operands[2] = gen_lowpart (Pmode, operands[2]);
+    }
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  pat = gen_rtx_PLUS (Pmode, operands[1], operands[2]);
+  if (Pmode != SImode)
+    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
+  DONE;
+})
 
-(define_insn "subhi3_carry"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-         (minus:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
-           (plus:HI (match_operand:HI 3 "ix86_carry_flag_operator" "")
-              (match_operand:HI 2 "general_operand" "rn,rm"))))
+;; Convert lea to the lea pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (zero_extend:DI
+         (plus:SI (match_operand:SI 1 "register_operand" "")
+                  (match_operand:SI 2 "nonmemory_operand" ""))))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (MINUS, HImode, operands)"
-  "sbb{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "HI")])
+  "TARGET_64BIT && reload_completed
+   && true_regnum (operands[0]) != true_regnum (operands[1])"
+  [(set (match_dup 0)
+       (zero_extend:DI (subreg:SI (plus:DI (match_dup 1) (match_dup 2)) 0)))]
+{
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[2] = gen_lowpart (Pmode, operands[2]);
+})
+\f
+;; Subtract instructions
 
-(define_insn "subsi3_carry"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-         (minus:SI (match_operand:SI 1 "nonimmediate_operand" "0,0")
-           (plus:SI (match_operand:SI 3 "ix86_carry_flag_operator" "")
-              (match_operand:SI 2 "general_operand" "ri,rm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (MINUS, SImode, operands)"
-  "sbb{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "SI")])
+(define_expand "sub<mode>3"
+  [(set (match_operand:SDWIM 0 "nonimmediate_operand" "")
+       (minus:SDWIM (match_operand:SDWIM 1 "nonimmediate_operand" "")
+                    (match_operand:SDWIM 2 "<general_operand>" "")))]
+  ""
+  "ix86_expand_binary_operator (MINUS, <MODE>mode, operands); DONE;")
 
-(define_insn "subsi3_carry_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-         (zero_extend:DI
-           (minus:SI (match_operand:SI 1 "register_operand" "0")
-             (plus:SI (match_operand:SI 3 "ix86_carry_flag_operator" "")
-                (match_operand:SI 2 "general_operand" "g")))))
+(define_insn_and_split "*sub<dwi>3_doubleword"
+  [(set (match_operand:<DWI> 0 "nonimmediate_operand" "=r,o")
+       (minus:<DWI>
+         (match_operand:<DWI> 1 "nonimmediate_operand" "0,0")
+         (match_operand:<DWI> 2 "<general_operand>" "ro<di>,r<di>")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (MINUS, SImode, operands)"
-  "sbb{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "pent_pair" "pu")
-   (set_attr "mode" "SI")])
-
-(define_expand "subsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (minus:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                 (match_operand:SI 2 "general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (MINUS, SImode, operands); DONE;")
+  "ix86_binary_operator_ok (MINUS, <MODE>mode, operands)"
+  "#"
+  "reload_completed"
+  [(parallel [(set (reg:CC FLAGS_REG)
+                  (compare:CC (match_dup 1) (match_dup 2)))
+             (set (match_dup 0)
+                  (minus:DWIH (match_dup 1) (match_dup 2)))])
+   (parallel [(set (match_dup 3)
+                  (minus:DWIH
+                    (match_dup 4)
+                    (plus:DWIH
+                      (ltu:DWIH (reg:CC FLAGS_REG) (const_int 0))
+                      (match_dup 5))))
+             (clobber (reg:CC FLAGS_REG))])]
+  "split_<dwi> (&operands[0], 3, &operands[0], &operands[3]);")
 
-(define_insn "*subsi_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (minus:SI (match_operand:SI 1 "nonimmediate_operand" "0,0")
-                 (match_operand:SI 2 "general_operand" "ri,rm")))
+(define_insn "*sub<mode>_1"
+  [(set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
+       (minus:SWI
+         (match_operand:SWI 1 "nonimmediate_operand" "0,0")
+         (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (MINUS, SImode, operands)"
-  "sub{l}\t{%2, %0|%0, %2}"
+  "ix86_binary_operator_ok (MINUS, <MODE>mode, operands)"
+  "sub{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_insn "*subsi_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_insn "*subsi_2"
+(define_insn "*subqi_1_slp"
+  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,q"))
+       (minus:QI (match_dup 0)
+                 (match_operand:QI 1 "general_operand" "qn,qm")))
+   (clobber (reg:CC FLAGS_REG))]
+  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+  "sub{b}\t{%1, %0|%0, %1}"
+  [(set_attr "type" "alu1")
+   (set_attr "mode" "QI")])
+
+(define_insn "*sub<mode>_2"
   [(set (reg FLAGS_REG)
        (compare
-         (minus:SI (match_operand:SI 1 "nonimmediate_operand" "0,0")
-                   (match_operand:SI 2 "general_operand" "ri,rm"))
+         (minus:SWI
+           (match_operand:SWI 1 "nonimmediate_operand" "0,0")
+           (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m"))
          (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (minus:SI (match_dup 1) (match_dup 2)))]
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
+       (minus:SWI (match_dup 1) (match_dup 2)))]
   "ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (MINUS, SImode, operands)"
-  "sub{l}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (MINUS, <MODE>mode, operands)"
+  "sub{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_insn "*subsi_2_zext"
   [(set (reg FLAGS_REG)
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_insn "*subsi_3"
+(define_insn "*sub<mode>_3"
   [(set (reg FLAGS_REG)
-       (compare (match_operand:SI 1 "nonimmediate_operand" "0,0")
-                (match_operand:SI 2 "general_operand" "ri,rm")))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (minus:SI (match_dup 1) (match_dup 2)))]
+       (compare (match_operand:SWI 1 "nonimmediate_operand" "0,0")
+                (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m")))
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
+       (minus:SWI (match_dup 1) (match_dup 2)))]
   "ix86_match_ccmode (insn, CCmode)
-   && ix86_binary_operator_ok (MINUS, SImode, operands)"
-  "sub{l}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (MINUS, <MODE>mode, operands)"
+  "sub{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_insn "*subsi_3_zext"
   [(set (reg FLAGS_REG)
    && ix86_binary_operator_ok (MINUS, SImode, operands)"
   "sub{l}\t{%2, %1|%1, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "SI")])
+\f
+;; Add with carry and subtract with borrow
 
-(define_expand "subhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (minus:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                 (match_operand:HI 2 "general_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (MINUS, HImode, operands); DONE;")
+(define_expand "<plusminus_insn><mode>3_carry"
+  [(parallel
+    [(set (match_operand:SWI 0 "nonimmediate_operand" "")
+         (plusminus:SWI
+           (match_operand:SWI 1 "nonimmediate_operand" "")
+           (plus:SWI (match_operator:SWI 4 "ix86_carry_flag_operator"
+                      [(match_operand 3 "flags_reg_operand" "")
+                       (const_int 0)])
+                     (match_operand:SWI 2 "<general_operand>" ""))))
+     (clobber (reg:CC FLAGS_REG))])]
+  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "")
 
-(define_insn "*subhi_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-       (minus:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
-                 (match_operand:HI 2 "general_operand" "rn,rm")))
+(define_insn "*<plusminus_insn><mode>3_carry"
+  [(set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
+       (plusminus:SWI
+         (match_operand:SWI 1 "nonimmediate_operand" "<comm>0,0")
+         (plus:SWI
+           (match_operator 3 "ix86_carry_flag_operator"
+            [(reg FLAGS_REG) (const_int 0)])
+           (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m"))))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (MINUS, HImode, operands)"
-  "sub{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_insn "*subhi_2"
-  [(set (reg FLAGS_REG)
-       (compare
-         (minus:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
-                   (match_operand:HI 2 "general_operand" "rn,rm"))
-         (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-       (minus:HI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (MINUS, HImode, operands)"
-  "sub{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_insn "*subhi_3"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:HI 1 "nonimmediate_operand" "0,0")
-                (match_operand:HI 2 "general_operand" "rn,rm")))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-       (minus:HI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCmode)
-   && ix86_binary_operator_ok (MINUS, HImode, operands)"
-  "sub{w}\t{%2, %0|%0, %2}"
+  "ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
+  "<plusminus_carry_mnemonic>{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_expand "subqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (minus:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                 (match_operand:QI 2 "general_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (MINUS, QImode, operands); DONE;")
+   (set_attr "use_carry" "1")
+   (set_attr "pent_pair" "pu")
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*subqi_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q")
-       (minus:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
-                 (match_operand:QI 2 "general_operand" "qn,qm")))
+(define_insn "*addsi3_carry_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
+                  (plus:SI (match_operator 3 "ix86_carry_flag_operator"
+                            [(reg FLAGS_REG) (const_int 0)])
+                           (match_operand:SI 2 "general_operand" "g")))))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (MINUS, QImode, operands)"
-  "sub{b}\t{%2, %0|%0, %2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (PLUS, SImode, operands)"
+  "adc{l}\t{%2, %k0|%k0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
+   (set_attr "use_carry" "1")
+   (set_attr "pent_pair" "pu")
+   (set_attr "mode" "SI")])
 
-(define_insn "*subqi_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,q"))
-       (minus:QI (match_dup 0)
-                 (match_operand:QI 1 "general_operand" "qn,qm")))
+(define_insn "*subsi3_carry_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (minus:SI (match_operand:SI 1 "register_operand" "0")
+                   (plus:SI (match_operator 3 "ix86_carry_flag_operator"
+                             [(reg FLAGS_REG) (const_int 0)])
+                            (match_operand:SI 2 "general_operand" "g")))))
    (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "sub{b}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*subqi_2"
-  [(set (reg FLAGS_REG)
-       (compare
-         (minus:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
-                   (match_operand:QI 2 "general_operand" "qn,qm"))
-         (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm,q")
-       (minus:QI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (MINUS, QImode, operands)"
-  "sub{b}\t{%2, %0|%0, %2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (MINUS, SImode, operands)"
+  "sbb{l}\t{%2, %k0|%k0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
+   (set_attr "pent_pair" "pu")
+   (set_attr "mode" "SI")])
+\f
+;; Overflow setting add and subtract instructions
 
-(define_insn "*subqi_3"
-  [(set (reg FLAGS_REG)
-       (compare (match_operand:QI 1 "nonimmediate_operand" "0,0")
-                (match_operand:QI 2 "general_operand" "qn,qm")))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm,q")
-       (minus:QI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCmode)
-   && ix86_binary_operator_ok (MINUS, QImode, operands)"
-  "sub{b}\t{%2, %0|%0, %2}"
+(define_insn "*add<mode>3_cconly_overflow"
+  [(set (reg:CCC FLAGS_REG)
+       (compare:CCC
+         (plus:SWI
+           (match_operand:SWI 1 "nonimmediate_operand" "%0")
+           (match_operand:SWI 2 "<general_operand>" "<r><i>m"))
+         (match_dup 1)))
+   (clobber (match_scratch:SWI 0 "=<r>"))]
+  "ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
+  "add{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
+   (set_attr "mode" "<MODE>")])
 
-;; The patterns that match these are at the end of this file.
+(define_insn "*sub<mode>3_cconly_overflow"
+  [(set (reg:CCC FLAGS_REG)
+       (compare:CCC
+         (minus:SWI
+           (match_operand:SWI 0 "nonimmediate_operand" "<r>m,<r>")
+           (match_operand:SWI 1 "<general_operand>" "<r><i>,<r>m"))
+         (match_dup 0)))]
+  ""
+  "cmp{<imodesuffix>}\t{%1, %0|%0, %1}"
+  [(set_attr "type" "icmp")
+   (set_attr "mode" "<MODE>")])
+
+(define_insn "*<plusminus_insn><mode>3_cc_overflow"
+  [(set (reg:CCC FLAGS_REG)
+       (compare:CCC
+           (plusminus:SWI
+               (match_operand:SWI 1 "nonimmediate_operand" "<comm>0,0")
+               (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m"))
+           (match_dup 1)))
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
+       (plusminus:SWI (match_dup 1) (match_dup 2)))]
+  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "<plusminus_mnemonic>{<imodesuffix>}\t{%2, %0|%0, %2}"
+  [(set_attr "type" "alu")
+   (set_attr "mode" "<MODE>")])
+
+(define_insn "*<plusminus_insn>si3_zext_cc_overflow"
+  [(set (reg:CCC FLAGS_REG)
+       (compare:CCC
+         (plusminus:SI
+           (match_operand:SI 1 "nonimmediate_operand" "<comm>0")
+           (match_operand:SI 2 "general_operand" "g"))
+         (match_dup 1)))
+   (set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI (plusminus:SI (match_dup 1) (match_dup 2))))]
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+  "<plusminus_mnemonic>{l}\t{%2, %k0|%k0, %2}"
+  [(set_attr "type" "alu")
+   (set_attr "mode" "SI")])
+
+;; The patterns that match these are at the end of this file.
 
-(define_expand "subxf3"
+(define_expand "<plusminus_insn>xf3"
   [(set (match_operand:XF 0 "register_operand" "")
-       (minus:XF (match_operand:XF 1 "register_operand" "")
-                 (match_operand:XF 2 "register_operand" "")))]
+       (plusminus:XF
+         (match_operand:XF 1 "register_operand" "")
+         (match_operand:XF 2 "register_operand" "")))]
   "TARGET_80387"
   "")
 
-(define_expand "sub<mode>3"
+(define_expand "<plusminus_insn><mode>3"
   [(set (match_operand:MODEF 0 "register_operand" "")
-       (minus:MODEF (match_operand:MODEF 1 "register_operand" "")
-                    (match_operand:MODEF 2 "nonimmediate_operand" "")))]
+       (plusminus:MODEF
+         (match_operand:MODEF 1 "register_operand" "")
+         (match_operand:MODEF 2 "nonimmediate_operand" "")))]
   "(TARGET_80387 && X87_ENABLE_ARITH (<MODE>mode))
     || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)"
   "")
 \f
 ;; Multiply instructions
 
-(define_expand "muldi3"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (mult:DI (match_operand:DI 1 "register_operand" "")
-                           (match_operand:DI 2 "x86_64_general_operand" "")))
+(define_expand "mul<mode>3"
+  [(parallel [(set (match_operand:SWIM248 0 "register_operand" "")
+                  (mult:SWIM248
+                    (match_operand:SWIM248 1 "register_operand" "")
+                    (match_operand:SWIM248 2 "<general_operand>" "")))
              (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_64BIT"
+  ""
   "")
 
-;; On AMDFAM10
-;; IMUL reg64, reg64, imm8     Direct
-;; IMUL reg64, mem64, imm8     VectorPath
-;; IMUL reg64, reg64, imm32    Direct
-;; IMUL reg64, mem64, imm32    VectorPath
-;; IMUL reg64, reg64           Direct
-;; IMUL reg64, mem64           Direct
-
-(define_insn "*muldi3_1_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=r,r,r")
-       (mult:DI (match_operand:DI 1 "nonimmediate_operand" "%rm,rm,0")
-                (match_operand:DI 2 "x86_64_general_operand" "K,e,mr")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "@
-   imul{q}\t{%2, %1, %0|%0, %1, %2}
-   imul{q}\t{%2, %1, %0|%0, %1, %2}
-   imul{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "imul")
-   (set_attr "prefix_0f" "0,0,1")
-   (set (attr "athlon_decode")
-       (cond [(eq_attr "cpu" "athlon")
-                 (const_string "vector")
-              (eq_attr "alternative" "1")
-                 (const_string "vector")
-              (and (eq_attr "alternative" "2")
-                   (match_operand 1 "memory_operand" ""))
-                 (const_string "vector")]
-             (const_string "direct")))
-   (set (attr "amdfam10_decode")
-       (cond [(and (eq_attr "alternative" "0,1")
-                   (match_operand 1 "memory_operand" ""))
-                 (const_string "vector")]
-             (const_string "direct")))
-   (set_attr "mode" "DI")])
-
-(define_expand "mulsi3"
-  [(parallel [(set (match_operand:SI 0 "register_operand" "")
-                  (mult:SI (match_operand:SI 1 "register_operand" "")
-                           (match_operand:SI 2 "general_operand" "")))
+(define_expand "mulqi3"
+  [(parallel [(set (match_operand:QI 0 "register_operand" "")
+                  (mult:QI
+                    (match_operand:QI 1 "register_operand" "")
+                    (match_operand:QI 2 "nonimmediate_operand" "")))
              (clobber (reg:CC FLAGS_REG))])]
-  ""
+  "TARGET_QIMODE_MATH"
   "")
 
 ;; On AMDFAM10
-;; IMUL reg32, reg32, imm8     Direct
-;; IMUL reg32, mem32, imm8     VectorPath
-;; IMUL reg32, reg32, imm32    Direct
-;; IMUL reg32, mem32, imm32    VectorPath
-;; IMUL reg32, reg32           Direct
-;; IMUL reg32, mem32           Direct
-
-(define_insn "*mulsi3_1"
-  [(set (match_operand:SI 0 "register_operand" "=r,r,r")
-       (mult:SI (match_operand:SI 1 "nonimmediate_operand" "%rm,rm,0")
-                (match_operand:SI 2 "general_operand" "K,i,mr")))
+;; IMUL reg32/64, reg32/64, imm8       Direct
+;; IMUL reg32/64, mem32/64, imm8       VectorPath
+;; IMUL reg32/64, reg32/64, imm32      Direct
+;; IMUL reg32/64, mem32/64, imm32      VectorPath
+;; IMUL reg32/64, reg32/64             Direct
+;; IMUL reg32/64, mem32/64             Direct
+
+(define_insn "*mul<mode>3_1"
+  [(set (match_operand:SWI48 0 "register_operand" "=r,r,r")
+       (mult:SWI48
+         (match_operand:SWI48 1 "nonimmediate_operand" "%rm,rm,0")
+         (match_operand:SWI48 2 "<general_operand>" "K,<i>,mr")))
    (clobber (reg:CC FLAGS_REG))]
   "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
   "@
-   imul{l}\t{%2, %1, %0|%0, %1, %2}
-   imul{l}\t{%2, %1, %0|%0, %1, %2}
-   imul{l}\t{%2, %0|%0, %2}"
+   imul{<imodesuffix>}\t{%2, %1, %0|%0, %1, %2}
+   imul{<imodesuffix>}\t{%2, %1, %0|%0, %1, %2}
+   imul{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "imul")
    (set_attr "prefix_0f" "0,0,1")
    (set (attr "athlon_decode")
                    (match_operand 1 "memory_operand" ""))
                  (const_string "vector")]
              (const_string "direct")))
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_insn "*mulsi3_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r,r,r")
              (const_string "direct")))
    (set_attr "mode" "SI")])
 
-(define_expand "mulhi3"
-  [(parallel [(set (match_operand:HI 0 "register_operand" "")
-                  (mult:HI (match_operand:HI 1 "register_operand" "")
-                           (match_operand:HI 2 "general_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_HIMODE_MATH"
-  "")
-
 ;; On AMDFAM10
 ;; IMUL reg16, reg16, imm8     VectorPath
 ;; IMUL reg16, mem16, imm8     VectorPath
 ;; IMUL reg16, mem16, imm16    VectorPath
 ;; IMUL reg16, reg16           Direct
 ;; IMUL reg16, mem16           Direct
+
 (define_insn "*mulhi3_1"
   [(set (match_operand:HI 0 "register_operand" "=r,r,r")
        (mult:HI (match_operand:HI 1 "nonimmediate_operand" "%rm,rm,0")
                 (match_operand:HI 2 "general_operand" "K,n,mr")))
    (clobber (reg:CC FLAGS_REG))]
-  "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
+  "TARGET_HIMODE_MATH
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
   "@
    imul{w}\t{%2, %1, %0|%0, %1, %2}
    imul{w}\t{%2, %1, %0|%0, %1, %2}
              (const_string "direct")))
    (set_attr "mode" "HI")])
 
-(define_expand "mulqi3"
-  [(parallel [(set (match_operand:QI 0 "register_operand" "")
-                  (mult:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                           (match_operand:QI 2 "register_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_QIMODE_MATH"
-  "")
-
 ;;On AMDFAM10
 ;; MUL reg8    Direct
 ;; MUL mem8    Direct
    (set_attr "amdfam10_decode" "direct")
    (set_attr "mode" "QI")])
 
-(define_expand "umulqihi3"
-  [(parallel [(set (match_operand:HI 0 "register_operand" "")
-                  (mult:HI (zero_extend:HI
-                             (match_operand:QI 1 "nonimmediate_operand" ""))
-                           (zero_extend:HI
-                             (match_operand:QI 2 "register_operand" ""))))
+(define_expand "<u>mul<mode><dwi>3"
+  [(parallel [(set (match_operand:<DWI> 0 "register_operand" "")
+                  (mult:<DWI>
+                    (any_extend:<DWI>
+                      (match_operand:DWIH 1 "nonimmediate_operand" ""))
+                    (any_extend:<DWI>
+                      (match_operand:DWIH 2 "register_operand" ""))))
              (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_QIMODE_MATH"
+  ""
   "")
 
-(define_insn "*umulqihi3_1"
-  [(set (match_operand:HI 0 "register_operand" "=a")
-       (mult:HI (zero_extend:HI (match_operand:QI 1 "nonimmediate_operand" "%0"))
-                (zero_extend:HI (match_operand:QI 2 "nonimmediate_operand" "qm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_QIMODE_MATH
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "mul{b}\t%2"
-  [(set_attr "type" "imul")
-   (set_attr "length_immediate" "0")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "direct")))
-   (set_attr "amdfam10_decode" "direct")
-   (set_attr "mode" "QI")])
-
-(define_expand "mulqihi3"
+(define_expand "<u>mulqihi3"
   [(parallel [(set (match_operand:HI 0 "register_operand" "")
-                  (mult:HI (sign_extend:HI (match_operand:QI 1 "nonimmediate_operand" ""))
-                           (sign_extend:HI (match_operand:QI 2 "register_operand" ""))))
+                  (mult:HI
+                    (any_extend:HI
+                      (match_operand:QI 1 "nonimmediate_operand" ""))
+                    (any_extend:HI
+                      (match_operand:QI 2 "register_operand" ""))))
              (clobber (reg:CC FLAGS_REG))])]
   "TARGET_QIMODE_MATH"
   "")
 
-(define_insn "*mulqihi3_insn"
-  [(set (match_operand:HI 0 "register_operand" "=a")
-       (mult:HI (sign_extend:HI (match_operand:QI 1 "nonimmediate_operand" "%0"))
-                (sign_extend:HI (match_operand:QI 2 "nonimmediate_operand" "qm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_QIMODE_MATH
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "imul{b}\t%2"
-  [(set_attr "type" "imul")
-   (set_attr "length_immediate" "0")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "direct")))
-   (set_attr "amdfam10_decode" "direct")
-   (set_attr "mode" "QI")])
-
-(define_expand "umulditi3"
-  [(parallel [(set (match_operand:TI 0 "register_operand" "")
-                  (mult:TI (zero_extend:TI
-                             (match_operand:DI 1 "nonimmediate_operand" ""))
-                           (zero_extend:TI
-                             (match_operand:DI 2 "register_operand" ""))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_64BIT"
-  "")
-
-(define_insn "*umulditi3_insn"
-  [(set (match_operand:TI 0 "register_operand" "=A")
-       (mult:TI (zero_extend:TI (match_operand:DI 1 "nonimmediate_operand" "%0"))
-                (zero_extend:TI (match_operand:DI 2 "nonimmediate_operand" "rm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "mul{q}\t%2"
-  [(set_attr "type" "imul")
-   (set_attr "length_immediate" "0")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "double")))
-   (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "DI")])
-
-;; We can't use this pattern in 64bit mode, since it results in two separate 32bit registers
-(define_expand "umulsidi3"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (mult:DI (zero_extend:DI
-                             (match_operand:SI 1 "nonimmediate_operand" ""))
-                           (zero_extend:DI
-                             (match_operand:SI 2 "register_operand" ""))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "!TARGET_64BIT"
-  "")
-
-(define_insn "*umulsidi3_insn"
-  [(set (match_operand:DI 0 "register_operand" "=A")
-       (mult:DI (zero_extend:DI (match_operand:SI 1 "nonimmediate_operand" "%0"))
-                (zero_extend:DI (match_operand:SI 2 "nonimmediate_operand" "rm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "mul{l}\t%2"
-  [(set_attr "type" "imul")
-   (set_attr "length_immediate" "0")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "double")))
-   (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "SI")])
-
-(define_expand "mulditi3"
-  [(parallel [(set (match_operand:TI 0 "register_operand" "")
-                  (mult:TI (sign_extend:TI
-                             (match_operand:DI 1 "nonimmediate_operand" ""))
-                           (sign_extend:TI
-                             (match_operand:DI 2 "register_operand" ""))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_64BIT"
-  "")
-
-(define_insn "*mulditi3_insn"
-  [(set (match_operand:TI 0 "register_operand" "=A")
-       (mult:TI (sign_extend:TI (match_operand:DI 1 "nonimmediate_operand" "%0"))
-                (sign_extend:TI (match_operand:DI 2 "nonimmediate_operand" "rm"))))
+(define_insn "*<u>mul<mode><dwi>3_1"
+  [(set (match_operand:<DWI> 0 "register_operand" "=A")
+       (mult:<DWI>
+         (any_extend:<DWI>
+           (match_operand:DWIH 1 "nonimmediate_operand" "%0"))
+         (any_extend:<DWI>
+           (match_operand:DWIH 2 "nonimmediate_operand" "rm"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "imul{q}\t%2"
+  "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
+  "<sgnprefix>mul{<imodesuffix>}\t%2"
   [(set_attr "type" "imul")
    (set_attr "length_immediate" "0")
    (set (attr "athlon_decode")
         (const_string "vector")
         (const_string "double")))
    (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "DI")])
-
-(define_expand "mulsidi3"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (mult:DI (sign_extend:DI
-                             (match_operand:SI 1 "nonimmediate_operand" ""))
-                           (sign_extend:DI
-                             (match_operand:SI 2 "register_operand" ""))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "!TARGET_64BIT"
-  "")
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*mulsidi3_insn"
-  [(set (match_operand:DI 0 "register_operand" "=A")
-       (mult:DI (sign_extend:DI (match_operand:SI 1 "nonimmediate_operand" "%0"))
-                (sign_extend:DI (match_operand:SI 2 "nonimmediate_operand" "rm"))))
+(define_insn "*<u>mulqihi3_1"
+  [(set (match_operand:HI 0 "register_operand" "=a")
+       (mult:HI
+         (any_extend:HI
+           (match_operand:QI 1 "nonimmediate_operand" "%0"))
+         (any_extend:HI
+           (match_operand:QI 2 "nonimmediate_operand" "qm"))))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT
+  "TARGET_QIMODE_MATH
    && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "imul{l}\t%2"
+  "<sgnprefix>mul{b}\t%2"
   [(set_attr "type" "imul")
    (set_attr "length_immediate" "0")
    (set (attr "athlon_decode")
      (if_then_else (eq_attr "cpu" "athlon")
         (const_string "vector")
-        (const_string "double")))
-   (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "SI")])
+        (const_string "direct")))
+   (set_attr "amdfam10_decode" "direct")
+   (set_attr "mode" "QI")])
 
-(define_expand "umuldi3_highpart"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (truncate:DI
-                    (lshiftrt:TI
-                      (mult:TI (zero_extend:TI
-                                 (match_operand:DI 1 "nonimmediate_operand" ""))
-                               (zero_extend:TI
-                                 (match_operand:DI 2 "register_operand" "")))
-                      (const_int 64))))
-             (clobber (match_scratch:DI 3 ""))
+(define_expand "<s>mul<mode>3_highpart"
+  [(parallel [(set (match_operand:SWI48 0 "register_operand" "")
+                  (truncate:SWI48
+                    (lshiftrt:<DWI>
+                      (mult:<DWI>
+                        (any_extend:<DWI>
+                          (match_operand:SWI48 1 "nonimmediate_operand" ""))
+                        (any_extend:<DWI>
+                          (match_operand:SWI48 2 "register_operand" "")))
+                      (match_dup 4))))
+             (clobber (match_scratch:SWI48 3 ""))
              (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_64BIT"
-  "")
+  ""
+  "operands[4] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode));")
 
-(define_insn "*umuldi3_highpart_rex64"
+(define_insn "*<s>muldi3_highpart_1"
   [(set (match_operand:DI 0 "register_operand" "=d")
        (truncate:DI
          (lshiftrt:TI
-           (mult:TI (zero_extend:TI
-                      (match_operand:DI 1 "nonimmediate_operand" "%a"))
-                    (zero_extend:TI
-                      (match_operand:DI 2 "nonimmediate_operand" "rm")))
+           (mult:TI
+             (any_extend:TI
+               (match_operand:DI 1 "nonimmediate_operand" "%a"))
+             (any_extend:TI
+               (match_operand:DI 2 "nonimmediate_operand" "rm")))
            (const_int 64))))
    (clobber (match_scratch:DI 3 "=1"))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT
    && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "mul{q}\t%2"
+  "<sgnprefix>mul{q}\t%2"
   [(set_attr "type" "imul")
    (set_attr "length_immediate" "0")
    (set (attr "athlon_decode")
    (set_attr "amdfam10_decode" "double")
    (set_attr "mode" "DI")])
 
-(define_expand "umulsi3_highpart"
-  [(parallel [(set (match_operand:SI 0 "register_operand" "")
-                  (truncate:SI
-                    (lshiftrt:DI
-                      (mult:DI (zero_extend:DI
-                                 (match_operand:SI 1 "nonimmediate_operand" ""))
-                               (zero_extend:DI
-                                 (match_operand:SI 2 "register_operand" "")))
-                      (const_int 32))))
-             (clobber (match_scratch:SI 3 ""))
-             (clobber (reg:CC FLAGS_REG))])]
-  ""
-  "")
-
-(define_insn "*umulsi3_highpart_insn"
+(define_insn "*<s>mulsi3_highpart_1"
   [(set (match_operand:SI 0 "register_operand" "=d")
        (truncate:SI
          (lshiftrt:DI
-           (mult:DI (zero_extend:DI
-                      (match_operand:SI 1 "nonimmediate_operand" "%a"))
-                    (zero_extend:DI
-                      (match_operand:SI 2 "nonimmediate_operand" "rm")))
+           (mult:DI
+             (any_extend:DI
+               (match_operand:SI 1 "nonimmediate_operand" "%a"))
+             (any_extend:DI
+               (match_operand:SI 2 "nonimmediate_operand" "rm")))
            (const_int 32))))
    (clobber (match_scratch:SI 3 "=1"))
    (clobber (reg:CC FLAGS_REG))]
   "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "mul{l}\t%2"
+  "<sgnprefix>mul{l}\t%2"
   [(set_attr "type" "imul")
    (set_attr "length_immediate" "0")
    (set (attr "athlon_decode")
    (set_attr "amdfam10_decode" "double")
    (set_attr "mode" "SI")])
 
-(define_insn "*umulsi3_highpart_zext"
+(define_insn "*<s>mulsi3_highpart_zext"
   [(set (match_operand:DI 0 "register_operand" "=d")
        (zero_extend:DI (truncate:SI
          (lshiftrt:DI
-           (mult:DI (zero_extend:DI
+           (mult:DI (any_extend:DI
                       (match_operand:SI 1 "nonimmediate_operand" "%a"))
-                    (zero_extend:DI
+                    (any_extend:DI
                       (match_operand:SI 2 "nonimmediate_operand" "rm")))
            (const_int 32)))))
    (clobber (match_scratch:SI 3 "=1"))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT
    && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "mul{l}\t%2"
+  "<sgnprefix>mul{l}\t%2"
   [(set_attr "type" "imul")
    (set_attr "length_immediate" "0")
    (set (attr "athlon_decode")
    (set_attr "amdfam10_decode" "double")
    (set_attr "mode" "SI")])
 
-(define_expand "smuldi3_highpart"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (truncate:DI
-                    (lshiftrt:TI
-                      (mult:TI (sign_extend:TI
-                                 (match_operand:DI 1 "nonimmediate_operand" ""))
-                               (sign_extend:TI
-                                 (match_operand:DI 2 "register_operand" "")))
-                      (const_int 64))))
-             (clobber (match_scratch:DI 3 ""))
-             (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_64BIT"
-  "")
-
-(define_insn "*smuldi3_highpart_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=d")
-       (truncate:DI
-         (lshiftrt:TI
-           (mult:TI (sign_extend:TI
-                      (match_operand:DI 1 "nonimmediate_operand" "%a"))
-                    (sign_extend:TI
-                      (match_operand:DI 2 "nonimmediate_operand" "rm")))
-           (const_int 64))))
-   (clobber (match_scratch:DI 3 "=1"))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "imul{q}\t%2"
-  [(set_attr "type" "imul")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "double")))
-   (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "DI")])
-
-(define_expand "smulsi3_highpart"
-  [(parallel [(set (match_operand:SI 0 "register_operand" "")
-                  (truncate:SI
-                    (lshiftrt:DI
-                      (mult:DI (sign_extend:DI
-                                 (match_operand:SI 1 "nonimmediate_operand" ""))
-                               (sign_extend:DI
-                                 (match_operand:SI 2 "register_operand" "")))
-                      (const_int 32))))
-             (clobber (match_scratch:SI 3 ""))
-             (clobber (reg:CC FLAGS_REG))])]
-  ""
-  "")
-
-(define_insn "*smulsi3_highpart_insn"
-  [(set (match_operand:SI 0 "register_operand" "=d")
-       (truncate:SI
-         (lshiftrt:DI
-           (mult:DI (sign_extend:DI
-                      (match_operand:SI 1 "nonimmediate_operand" "%a"))
-                    (sign_extend:DI
-                      (match_operand:SI 2 "nonimmediate_operand" "rm")))
-           (const_int 32))))
-   (clobber (match_scratch:SI 3 "=1"))
-   (clobber (reg:CC FLAGS_REG))]
-  "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "imul{l}\t%2"
-  [(set_attr "type" "imul")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "double")))
-   (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "SI")])
-
-(define_insn "*smulsi3_highpart_zext"
-  [(set (match_operand:DI 0 "register_operand" "=d")
-       (zero_extend:DI (truncate:SI
-         (lshiftrt:DI
-           (mult:DI (sign_extend:DI
-                      (match_operand:SI 1 "nonimmediate_operand" "%a"))
-                    (sign_extend:DI
-                      (match_operand:SI 2 "nonimmediate_operand" "rm")))
-           (const_int 32)))))
-   (clobber (match_scratch:SI 3 "=1"))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "imul{l}\t%2"
-  [(set_attr "type" "imul")
-   (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "double")))
-   (set_attr "amdfam10_decode" "double")
-   (set_attr "mode" "SI")])
-
 ;; The patterns that match these are at the end of this file.
 
 (define_expand "mulxf3"
   "(TARGET_80387 && X87_ENABLE_ARITH (<MODE>mode))
     || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)"
   "")
-
-;; SSE5 scalar multiply/add instructions are defined in sse.md.
-
 \f
 ;; Divide instructions
 
-(define_insn "divqi3"
+(define_insn "<u>divqi3"
   [(set (match_operand:QI 0 "register_operand" "=a")
-       (div:QI (match_operand:HI 1 "register_operand" "0")
-               (match_operand:QI 2 "nonimmediate_operand" "qm")))
+       (any_div:QI
+         (match_operand:HI 1 "register_operand" "0")
+         (match_operand:QI 2 "nonimmediate_operand" "qm")))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_QIMODE_MATH"
-  "idiv{b}\t%2"
-  [(set_attr "type" "idiv")
-   (set_attr "mode" "QI")])
-
-(define_insn "udivqi3"
-  [(set (match_operand:QI 0 "register_operand" "=a")
-       (udiv:QI (match_operand:HI 1 "register_operand" "0")
-                (match_operand:QI 2 "nonimmediate_operand" "qm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_QIMODE_MATH"
-  "div{b}\t%2"
+  "<sgnprefix>div{b}\t%2"
   [(set_attr "type" "idiv")
    (set_attr "mode" "QI")])
 
     }
 })
 \f
-;; Remainder instructions.
-
-(define_expand "divmoddi4"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (div:DI (match_operand:DI 1 "register_operand" "")
-                          (match_operand:DI 2 "nonimmediate_operand" "")))
-             (set (match_operand:DI 3 "register_operand" "")
-                  (mod:DI (match_dup 1) (match_dup 2)))
-             (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_64BIT"
-  "")
-
-;; Allow to come the parameter in eax or edx to avoid extra moves.
-;; Penalize eax case slightly because it results in worse scheduling
-;; of code.
-(define_insn "*divmoddi4_nocltd_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=&a,?a")
-       (div:DI (match_operand:DI 2 "register_operand" "1,0")
-               (match_operand:DI 3 "nonimmediate_operand" "rm,rm")))
-   (set (match_operand:DI 1 "register_operand" "=&d,&d")
-       (mod:DI (match_dup 2) (match_dup 3)))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && optimize_function_for_speed_p (cfun) && !TARGET_USE_CLTD"
-  "#"
-  [(set_attr "type" "multi")])
-
-(define_insn "*divmoddi4_cltd_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=a")
-       (div:DI (match_operand:DI 2 "register_operand" "a")
-               (match_operand:DI 3 "nonimmediate_operand" "rm")))
-   (set (match_operand:DI 1 "register_operand" "=&d")
-       (mod:DI (match_dup 2) (match_dup 3)))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && (optimize_function_for_size_p (cfun) || TARGET_USE_CLTD)"
-  "#"
-  [(set_attr "type" "multi")])
-
-(define_insn "*divmoddi_noext_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=a")
-       (div:DI (match_operand:DI 1 "register_operand" "0")
-               (match_operand:DI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:DI 3 "register_operand" "=d")
-       (mod:DI (match_dup 1) (match_dup 2)))
-   (use (match_operand:DI 4 "register_operand" "3"))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "idiv{q}\t%2"
-  [(set_attr "type" "idiv")
-   (set_attr "mode" "DI")])
-
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (div:DI (match_operand:DI 1 "register_operand" "")
-               (match_operand:DI 2 "nonimmediate_operand" "")))
-   (set (match_operand:DI 3 "register_operand" "")
-       (mod:DI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed"
-  [(parallel [(set (match_dup 3)
-                  (ashiftrt:DI (match_dup 4) (const_int 63)))
-             (clobber (reg:CC FLAGS_REG))])
-   (parallel [(set (match_dup 0)
-                  (div:DI (reg:DI 0) (match_dup 2)))
-             (set (match_dup 3)
-                  (mod:DI (reg:DI 0) (match_dup 2)))
-             (use (match_dup 3))
-             (clobber (reg:CC FLAGS_REG))])]
-{
-  /* Avoid use of cltd in favor of a mov+shift.  */
-  if (!TARGET_USE_CLTD && optimize_function_for_speed_p (cfun))
-    {
-      if (true_regnum (operands[1]))
-        emit_move_insn (operands[0], operands[1]);
-      else
-       emit_move_insn (operands[3], operands[1]);
-      operands[4] = operands[3];
-    }
-  else
-    {
-      gcc_assert (!true_regnum (operands[1]));
-      operands[4] = operands[1];
-    }
-})
-
-
-(define_expand "divmodsi4"
-  [(parallel [(set (match_operand:SI 0 "register_operand" "")
-                  (div:SI (match_operand:SI 1 "register_operand" "")
-                          (match_operand:SI 2 "nonimmediate_operand" "")))
-             (set (match_operand:SI 3 "register_operand" "")
-                  (mod:SI (match_dup 1) (match_dup 2)))
+;; Divmod instructions.
+
+(define_expand "divmod<mode>4"
+  [(parallel [(set (match_operand:SWIM248 0 "register_operand" "")
+                  (div:SWIM248
+                    (match_operand:SWIM248 1 "register_operand" "")
+                    (match_operand:SWIM248 2 "nonimmediate_operand" "")))
+             (set (match_operand:SWIM248 3 "register_operand" "")
+                  (mod:SWIM248 (match_dup 1) (match_dup 2)))
              (clobber (reg:CC FLAGS_REG))])]
   ""
   "")
 
-;; Allow to come the parameter in eax or edx to avoid extra moves.
-;; Penalize eax case slightly because it results in worse scheduling
-;; of code.
-(define_insn "*divmodsi4_nocltd"
-  [(set (match_operand:SI 0 "register_operand" "=&a,?a")
-       (div:SI (match_operand:SI 2 "register_operand" "1,0")
-               (match_operand:SI 3 "nonimmediate_operand" "rm,rm")))
-   (set (match_operand:SI 1 "register_operand" "=&d,&d")
-       (mod:SI (match_dup 2) (match_dup 3)))
-   (clobber (reg:CC FLAGS_REG))]
-  "optimize_function_for_speed_p (cfun) && !TARGET_USE_CLTD"
-  "#"
-  [(set_attr "type" "multi")])
-
-(define_insn "*divmodsi4_cltd"
-  [(set (match_operand:SI 0 "register_operand" "=a")
-       (div:SI (match_operand:SI 2 "register_operand" "a")
-               (match_operand:SI 3 "nonimmediate_operand" "rm")))
-   (set (match_operand:SI 1 "register_operand" "=&d")
-       (mod:SI (match_dup 2) (match_dup 3)))
-   (clobber (reg:CC FLAGS_REG))]
-  "optimize_function_for_size_p (cfun) || TARGET_USE_CLTD"
-  "#"
-  [(set_attr "type" "multi")])
-
-(define_insn "*divmodsi_noext"
-  [(set (match_operand:SI 0 "register_operand" "=a")
-       (div:SI (match_operand:SI 1 "register_operand" "0")
-               (match_operand:SI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:SI 3 "register_operand" "=d")
-       (mod:SI (match_dup 1) (match_dup 2)))
-   (use (match_operand:SI 4 "register_operand" "3"))
+(define_insn_and_split "*divmod<mode>4"
+  [(set (match_operand:SWIM248 0 "register_operand" "=a")
+       (div:SWIM248 (match_operand:SWIM248 2 "register_operand" "0")
+                   (match_operand:SWIM248 3 "nonimmediate_operand" "rm")))
+   (set (match_operand:SWIM248 1 "register_operand" "=&d")
+       (mod:SWIM248 (match_dup 2) (match_dup 3)))
    (clobber (reg:CC FLAGS_REG))]
   ""
-  "idiv{l}\t%2"
-  [(set_attr "type" "idiv")
-   (set_attr "mode" "SI")])
-
-(define_split
-  [(set (match_operand:SI 0 "register_operand" "")
-       (div:SI (match_operand:SI 1 "register_operand" "")
-               (match_operand:SI 2 "nonimmediate_operand" "")))
-   (set (match_operand:SI 3 "register_operand" "")
-       (mod:SI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
-  "reload_completed"
-  [(parallel [(set (match_dup 3)
-                  (ashiftrt:SI (match_dup 4) (const_int 31)))
+  "#"
+  "&& reload_completed"
+  [(parallel [(set (match_dup 1)
+                  (ashiftrt:SWIM248 (match_dup 4) (match_dup 5)))
              (clobber (reg:CC FLAGS_REG))])
    (parallel [(set (match_dup 0)
-                  (div:SI (reg:SI 0) (match_dup 2)))
-             (set (match_dup 3)
-                  (mod:SI (reg:SI 0) (match_dup 2)))
-             (use (match_dup 3))
+                  (div:SWIM248 (match_dup 2) (match_dup 3)))
+             (set (match_dup 1)
+                  (mod:SWIM248 (match_dup 2) (match_dup 3)))
+             (use (match_dup 1))
              (clobber (reg:CC FLAGS_REG))])]
 {
-  /* Avoid use of cltd in favor of a mov+shift.  */
-  if (!TARGET_USE_CLTD && optimize_function_for_speed_p (cfun))
-    {
-      if (true_regnum (operands[1]))
-        emit_move_insn (operands[0], operands[1]);
-      else
-       emit_move_insn (operands[3], operands[1]);
-      operands[4] = operands[3];
-    }
+  operands[5] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode) - 1);
+
+  if (<MODE>mode != HImode
+      && (optimize_function_for_size_p (cfun) || TARGET_USE_CLTD))
+    operands[4] = operands[2];
   else
     {
-      gcc_assert (!true_regnum (operands[1]));
+      /* Avoid use of cltd in favor of a mov+shift.  */
+      emit_move_insn (operands[1], operands[2]);
       operands[4] = operands[1];
     }
-})
-;; %%% Split me.
-(define_insn "divmodhi4"
-  [(set (match_operand:HI 0 "register_operand" "=a")
-       (div:HI (match_operand:HI 1 "register_operand" "0")
-               (match_operand:HI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:HI 3 "register_operand" "=&d")
-       (mod:HI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_HIMODE_MATH"
-  "cwtd\;idiv{w}\t%2"
-  [(set_attr "type" "multi")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "udivmoddi4"
-  [(set (match_operand:DI 0 "register_operand" "=a")
-       (udiv:DI (match_operand:DI 1 "register_operand" "0")
-                (match_operand:DI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:DI 3 "register_operand" "=&d")
-       (umod:DI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "xor{q}\t%3, %3\;div{q}\t%2"
+}
   [(set_attr "type" "multi")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*udivmoddi4_noext"
-  [(set (match_operand:DI 0 "register_operand" "=a")
-       (udiv:DI (match_operand:DI 1 "register_operand" "0")
-                (match_operand:DI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:DI 3 "register_operand" "=d")
-       (umod:DI (match_dup 1) (match_dup 2)))
-   (use (match_dup 3))
+(define_insn "*divmod<mode>4_noext"
+  [(set (match_operand:SWIM248 0 "register_operand" "=a")
+       (div:SWIM248 (match_operand:SWIM248 2 "register_operand" "0")
+                   (match_operand:SWIM248 3 "nonimmediate_operand" "rm")))
+   (set (match_operand:SWIM248 1 "register_operand" "=d")
+       (mod:SWIM248 (match_dup 2) (match_dup 3)))
+   (use (match_operand:SWIM248 4 "register_operand" "1"))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "div{q}\t%2"
+  ""
+  "idiv{<imodesuffix>}\t%3"
   [(set_attr "type" "idiv")
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (udiv:DI (match_operand:DI 1 "register_operand" "")
-                (match_operand:DI 2 "nonimmediate_operand" "")))
-   (set (match_operand:DI 3 "register_operand" "")
-       (umod:DI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed"
-  [(set (match_dup 3) (const_int 0))
-   (parallel [(set (match_dup 0)
-                  (udiv:DI (match_dup 1) (match_dup 2)))
-             (set (match_dup 3)
-                  (umod:DI (match_dup 1) (match_dup 2)))
-             (use (match_dup 3))
+(define_expand "udivmod<mode>4"
+  [(parallel [(set (match_operand:SWIM248 0 "register_operand" "")
+                  (udiv:SWIM248
+                    (match_operand:SWIM248 1 "register_operand" "")
+                    (match_operand:SWIM248 2 "nonimmediate_operand" "")))
+             (set (match_operand:SWIM248 3 "register_operand" "")
+                  (umod:SWIM248 (match_dup 1) (match_dup 2)))
              (clobber (reg:CC FLAGS_REG))])]
-  "")
-
-(define_insn "udivmodsi4"
-  [(set (match_operand:SI 0 "register_operand" "=a")
-       (udiv:SI (match_operand:SI 1 "register_operand" "0")
-                (match_operand:SI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:SI 3 "register_operand" "=&d")
-       (umod:SI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
   ""
-  "xor{l}\t%3, %3\;div{l}\t%2"
-  [(set_attr "type" "multi")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
+  "")
 
-(define_insn "*udivmodsi4_noext"
-  [(set (match_operand:SI 0 "register_operand" "=a")
-       (udiv:SI (match_operand:SI 1 "register_operand" "0")
-                (match_operand:SI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:SI 3 "register_operand" "=d")
-       (umod:SI (match_dup 1) (match_dup 2)))
-   (use (match_dup 3))
+(define_insn_and_split "*udivmod<mode>4"
+  [(set (match_operand:SWIM248 0 "register_operand" "=a")
+       (udiv:SWIM248 (match_operand:SWIM248 2 "register_operand" "0")
+                     (match_operand:SWIM248 3 "nonimmediate_operand" "rm")))
+   (set (match_operand:SWIM248 1 "register_operand" "=&d")
+       (umod:SWIM248 (match_dup 2) (match_dup 3)))
    (clobber (reg:CC FLAGS_REG))]
   ""
-  "div{l}\t%2"
-  [(set_attr "type" "idiv")
-   (set_attr "mode" "SI")])
-
-(define_split
-  [(set (match_operand:SI 0 "register_operand" "")
-       (udiv:SI (match_operand:SI 1 "register_operand" "")
-                (match_operand:SI 2 "nonimmediate_operand" "")))
-   (set (match_operand:SI 3 "register_operand" "")
-       (umod:SI (match_dup 1) (match_dup 2)))
-   (clobber (reg:CC FLAGS_REG))]
-  "reload_completed"
-  [(set (match_dup 3) (const_int 0))
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 1) (const_int 0))
    (parallel [(set (match_dup 0)
-                  (udiv:SI (match_dup 1) (match_dup 2)))
-             (set (match_dup 3)
-                  (umod:SI (match_dup 1) (match_dup 2)))
-             (use (match_dup 3))
-             (clobber (reg:CC FLAGS_REG))])]
-  "")
-
-(define_expand "udivmodhi4"
-  [(set (match_dup 4) (const_int 0))
-   (parallel [(set (match_operand:HI 0 "register_operand" "")
-                  (udiv:HI (match_operand:HI 1 "register_operand" "")
-                           (match_operand:HI 2 "nonimmediate_operand" "")))
-             (set (match_operand:HI 3 "register_operand" "")
-                  (umod:HI (match_dup 1) (match_dup 2)))
-             (use (match_dup 4))
+                  (udiv:SWIM248 (match_dup 2) (match_dup 3)))
+             (set (match_dup 1)
+                  (umod:SWIM248 (match_dup 2) (match_dup 3)))
+             (use (match_dup 1))
              (clobber (reg:CC FLAGS_REG))])]
-  "TARGET_HIMODE_MATH"
-  "operands[4] = gen_reg_rtx (HImode);")
+  ""
+  [(set_attr "type" "multi")
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*udivmodhi_noext"
-  [(set (match_operand:HI 0 "register_operand" "=a")
-       (udiv:HI (match_operand:HI 1 "register_operand" "0")
-                (match_operand:HI 2 "nonimmediate_operand" "rm")))
-   (set (match_operand:HI 3 "register_operand" "=d")
-       (umod:HI (match_dup 1) (match_dup 2)))
-   (use (match_operand:HI 4 "register_operand" "3"))
+(define_insn "*udivmod<mode>4_noext"
+  [(set (match_operand:SWIM248 0 "register_operand" "=a")
+       (udiv:SWIM248 (match_operand:SWIM248 2 "register_operand" "0")
+                     (match_operand:SWIM248 3 "nonimmediate_operand" "rm")))
+   (set (match_operand:SWIM248 1 "register_operand" "=d")
+       (umod:SWIM248 (match_dup 2) (match_dup 3)))
+   (use (match_operand:SWIM248 4 "register_operand" "1"))
    (clobber (reg:CC FLAGS_REG))]
   ""
-  "div{w}\t%2"
+  "div{<imodesuffix>}\t%3"
   [(set_attr "type" "idiv")
-   (set_attr "mode" "HI")])
+   (set_attr "mode" "<MODE>")])
 
 ;; We cannot use div/idiv for double division, because it causes
 ;; "division by zero" on the overflow and that's not what we expect
 ;; On Pentium, "test imm, reg" is pairable only with eax, ax, and al.
 ;; Note that this excludes ah.
 
-(define_insn "*testdi_1_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (and:DI (match_operand:DI 0 "nonimmediate_operand" "%!*a,r,!*a,r,rm")
-                 (match_operand:DI 1 "x86_64_szext_general_operand" "Z,Z,e,e,re"))
-         (const_int 0)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "@
-   test{l}\t{%k1, %k0|%k0, %k1}
-   test{l}\t{%k1, %k0|%k0, %k1}
-   test{q}\t{%1, %0|%0, %1}
-   test{q}\t{%1, %0|%0, %1}
-   test{q}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "test")
-   (set_attr "modrm" "0,1,0,1,1")
-   (set_attr "mode" "SI,SI,DI,DI,DI")
-   (set_attr "pent_pair" "uv,np,uv,np,uv")])
-
-(define_insn "testsi_1"
-  [(set (reg FLAGS_REG)
-       (compare
-         (and:SI (match_operand:SI 0 "nonimmediate_operand" "%!*a,r,rm")
-                 (match_operand:SI 1 "general_operand" "i,i,ri"))
-         (const_int 0)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "test{l}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "test")
-   (set_attr "modrm" "0,1,1")
-   (set_attr "mode" "SI")
-   (set_attr "pent_pair" "uv,np,uv")])
-
 (define_expand "testsi_ccno_1"
   [(set (reg:CCNO FLAGS_REG)
        (compare:CCNO
   ""
   "")
 
-(define_insn "*testhi_1"
-  [(set (reg FLAGS_REG)
-        (compare (and:HI (match_operand:HI 0 "nonimmediate_operand" "%!*a,r,rm")
-                        (match_operand:HI 1 "general_operand" "n,n,rn"))
-                (const_int 0)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "test{w}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "test")
-   (set_attr "modrm" "0,1,1")
-   (set_attr "mode" "HI")
-   (set_attr "pent_pair" "uv,np,uv")])
-
 (define_expand "testqi_ccz_1"
   [(set (reg:CCZ FLAGS_REG)
         (compare:CCZ (and:QI (match_operand:QI 0 "nonimmediate_operand" "")
   ""
   "")
 
+(define_insn "*testdi_1"
+  [(set (reg FLAGS_REG)
+       (compare
+        (and:DI
+         (match_operand:DI 0 "nonimmediate_operand" "%!*a,r,!*a,r,rm")
+         (match_operand:DI 1 "x86_64_szext_general_operand" "Z,Z,e,e,re"))
+        (const_int 0)))]
+  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
+   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+  "@
+   test{l}\t{%k1, %k0|%k0, %k1}
+   test{l}\t{%k1, %k0|%k0, %k1}
+   test{q}\t{%1, %0|%0, %1}
+   test{q}\t{%1, %0|%0, %1}
+   test{q}\t{%1, %0|%0, %1}"
+  [(set_attr "type" "test")
+   (set_attr "modrm" "0,1,0,1,1")
+   (set_attr "mode" "SI,SI,DI,DI,DI")])
+
 (define_insn "*testqi_1_maybe_si"
   [(set (reg FLAGS_REG)
         (compare
    (set_attr "mode" "QI,QI,QI,SI")
    (set_attr "pent_pair" "uv,np,uv,np")])
 
-(define_insn "*testqi_1"
+(define_insn "*test<mode>_1"
   [(set (reg FLAGS_REG)
-        (compare
-         (and:QI
-           (match_operand:QI 0 "nonimmediate_operand" "%!*a,q,qm")
-           (match_operand:QI 1 "general_operand" "n,n,qn"))
-         (const_int 0)))]
-  "!(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && ix86_match_ccmode (insn, CCNOmode)"
-  "test{b}\t{%1, %0|%0, %1}"
+       (compare
+        (and:SWI124
+         (match_operand:SWI124 0 "nonimmediate_operand" "%!*a,<r>,<r>m")
+         (match_operand:SWI124 1 "general_operand" "<i>,<i>,<r><i>"))
+        (const_int 0)))]
+  "ix86_match_ccmode (insn, CCNOmode)
+   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+  "test{<imodesuffix>}\t{%1, %0|%0, %1}"
   [(set_attr "type" "test")
    (set_attr "modrm" "0,1,1")
-   (set_attr "mode" "QI")
+   (set_attr "mode" "<MODE>")
    (set_attr "pent_pair" "uv,np,uv")])
 
 (define_expand "testqi_ext_ccno_0"
    (set_attr "modrm" "1")
    (set_attr "pent_pair" "np")])
 
-(define_insn "*testqi_ext_1"
+(define_insn "*testqi_ext_1_rex64"
   [(set (reg FLAGS_REG)
        (compare
          (and:SI
              (const_int 8)
              (const_int 8))
            (zero_extend:SI
-             (match_operand:QI 1 "general_operand" "Qm")))
+             (match_operand:QI 1 "register_operand" "Q")))
          (const_int 0)))]
-  "!TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
   "test{b}\t{%1, %h0|%h0, %1}"
   [(set_attr "type" "test")
    (set_attr "mode" "QI")])
 
-(define_insn "*testqi_ext_1_rex64"
+(define_insn "*testqi_ext_1"
   [(set (reg FLAGS_REG)
        (compare
          (and:SI
              (const_int 8)
              (const_int 8))
            (zero_extend:SI
-             (match_operand:QI 1 "register_operand" "Q")))
+             (match_operand:QI 1 "general_operand" "Qm")))
          (const_int 0)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
+  "!TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
   "test{b}\t{%1, %h0|%h0, %1}"
   [(set_attr "type" "test")
    (set_attr "mode" "QI")])
   [(set_attr "type" "test")
    (set_attr "mode" "QI")])
 
-;; Combine likes to form bit extractions for some tests.  Humor it.
-(define_insn "*testqi_ext_3"
-  [(set (reg FLAGS_REG)
-        (compare (zero_extract:SI
-                  (match_operand 0 "nonimmediate_operand" "rm")
-                  (match_operand:SI 1 "const_int_operand" "")
-                  (match_operand:SI 2 "const_int_operand" ""))
-                (const_int 0)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && INTVAL (operands[1]) > 0
-   && INTVAL (operands[2]) >= 0
-   && INTVAL (operands[1]) + INTVAL (operands[2]) <= 32
-   && (GET_MODE (operands[0]) == SImode
-       || (TARGET_64BIT && GET_MODE (operands[0]) == DImode)
-       || GET_MODE (operands[0]) == HImode
-       || GET_MODE (operands[0]) == QImode)"
-  "#")
-
 (define_insn "*testqi_ext_3_rex64"
   [(set (reg FLAGS_REG)
         (compare (zero_extract:DI
        || GET_MODE (operands[0]) == QImode)"
   "#")
 
+;; Combine likes to form bit extractions for some tests.  Humor it.
+(define_insn "*testqi_ext_3"
+  [(set (reg FLAGS_REG)
+        (compare (zero_extract:SI
+                  (match_operand 0 "nonimmediate_operand" "rm")
+                  (match_operand:SI 1 "const_int_operand" "")
+                  (match_operand:SI 2 "const_int_operand" ""))
+                (const_int 0)))]
+  "ix86_match_ccmode (insn, CCNOmode)
+   && INTVAL (operands[1]) > 0
+   && INTVAL (operands[2]) >= 0
+   && INTVAL (operands[1]) + INTVAL (operands[2]) <= 32
+   && (GET_MODE (operands[0]) == SImode
+       || (TARGET_64BIT && GET_MODE (operands[0]) == DImode)
+       || GET_MODE (operands[0]) == HImode
+       || GET_MODE (operands[0]) == QImode)"
+  "#")
+
 (define_split
   [(set (match_operand 0 "flags_reg_operand" "")
         (match_operator 1 "compare_operator"
   else if (GET_CODE (val) == SUBREG
           && (submode = GET_MODE (SUBREG_REG (val)),
               GET_MODE_BITSIZE (mode) > GET_MODE_BITSIZE (submode))
-          && pos + len <= GET_MODE_BITSIZE (submode))
+          && pos + len <= GET_MODE_BITSIZE (submode)
+          && GET_MODE_CLASS (submode) == MODE_INT)
     {
       /* Narrow a paradoxical subreg to prevent partial register stalls.  */
       mode = submode;
   "operands[2] = gen_lowpart (QImode, operands[2]);
    operands[3] = gen_lowpart (QImode, operands[3]);")
 
-
 ;; %%% This used to optimize known byte-wide and operations to memory,
 ;; and sometimes to QImode registers.  If this is considered useful,
 ;; it should be done with splitters.
 
-(define_expand "anddi3"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (and:DI (match_operand:DI 1 "nonimmediate_operand" "")
-               (match_operand:DI 2 "x86_64_szext_general_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (AND, DImode, operands); DONE;")
+(define_expand "and<mode>3"
+  [(set (match_operand:SWIM 0 "nonimmediate_operand" "")
+       (and:SWIM (match_operand:SWIM 1 "nonimmediate_operand" "")
+                 (match_operand:SWIM 2 "<general_szext_operand>" "")))]
+  ""
+  "ix86_expand_binary_operator (AND, <MODE>mode, operands); DONE;")
 
-(define_insn "*anddi_1_rex64"
+(define_insn "*anddi_1"
   [(set (match_operand:DI 0 "nonimmediate_operand" "=r,rm,r,r")
-       (and:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0,0,qm")
-               (match_operand:DI 2 "x86_64_szext_general_operand" "Z,re,rm,L")))
+       (and:DI
+        (match_operand:DI 1 "nonimmediate_operand" "%0,0,0,qm")
+        (match_operand:DI 2 "x86_64_szext_general_operand" "Z,re,rm,L")))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (AND, DImode, operands)"
 {
        (const_string "*")))
    (set_attr "mode" "SI,DI,DI,SI")])
 
-(define_insn "*anddi_2"
-  [(set (reg FLAGS_REG)
-       (compare (and:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0,0")
-                        (match_operand:DI 2 "x86_64_szext_general_operand" "Z,rem,re"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=r,r,rm")
-       (and:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (AND, DImode, operands)"
-  "@
-   and{l}\t{%k2, %k0|%k0, %k2}
-   and{q}\t{%2, %0|%0, %2}
-   and{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI,DI,DI")])
-
-(define_expand "andsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (and:SI (match_operand:SI 1 "nonimmediate_operand" "")
-               (match_operand:SI 2 "general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (AND, SImode, operands); DONE;")
-
 (define_insn "*andsi_1"
   [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r,r")
        (and:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0,qm")
    (set_attr "length_immediate" "*,*,0")
    (set_attr "mode" "SI")])
 
-(define_split
-  [(set (match_operand 0 "register_operand" "")
-       (and (match_dup 0)
-            (const_int -65536)))
-   (clobber (reg:CC FLAGS_REG))]
-  "optimize_function_for_size_p (cfun) || (TARGET_FAST_PREFIX && !TARGET_PARTIAL_REG_STALL)"
-  [(set (strict_low_part (match_dup 1)) (const_int 0))]
-  "operands[1] = gen_lowpart (HImode, operands[0]);")
-
-(define_split
-  [(set (match_operand 0 "ext_register_operand" "")
-       (and (match_dup 0)
-            (const_int -256)))
-   (clobber (reg:CC FLAGS_REG))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_REG_STALL) && reload_completed"
-  [(set (strict_low_part (match_dup 1)) (const_int 0))]
-  "operands[1] = gen_lowpart (QImode, operands[0]);")
-
-(define_split
-  [(set (match_operand 0 "ext_register_operand" "")
-       (and (match_dup 0)
-            (const_int -65281)))
-   (clobber (reg:CC FLAGS_REG))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_REG_STALL) && reload_completed"
-  [(parallel [(set (zero_extract:SI (match_dup 0)
-                                   (const_int 8)
-                                   (const_int 8))
-                  (xor:SI
-                    (zero_extract:SI (match_dup 0)
-                                     (const_int 8)
-                                     (const_int 8))
-                    (zero_extract:SI (match_dup 0)
-                                     (const_int 8)
-                                     (const_int 8))))
-             (clobber (reg:CC FLAGS_REG))])]
-  "operands[0] = gen_lowpart (SImode, operands[0]);")
-
 ;; See comment for addsi_1_zext why we do use nonimmediate_operand
 (define_insn "*andsi_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_insn "*andsi_2"
-  [(set (reg FLAGS_REG)
-       (compare (and:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:SI 2 "general_operand" "g,ri"))
-                (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=r,rm")
-       (and:SI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (AND, SImode, operands)"
-  "and{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-;; See comment for addsi_1_zext why we do use nonimmediate_operand
-(define_insn "*andsi_2_zext"
-  [(set (reg FLAGS_REG)
-       (compare (and:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand:SI 2 "general_operand" "g"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (and:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (AND, SImode, operands)"
-  "and{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_expand "andhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (and:HI (match_operand:HI 1 "nonimmediate_operand" "")
-               (match_operand:HI 2 "general_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (AND, HImode, operands); DONE;")
-
 (define_insn "*andhi_1"
   [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r,r")
        (and:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0,qm")
        (const_string "*")))
    (set_attr "mode" "HI,HI,SI")])
 
-(define_insn "*andhi_2"
-  [(set (reg FLAGS_REG)
-       (compare (and:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:HI 2 "general_operand" "rmn,rn"))
-                (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=r,rm")
-       (and:HI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (AND, HImode, operands)"
-  "and{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_expand "andqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (and:QI (match_operand:QI 1 "nonimmediate_operand" "")
-               (match_operand:QI 2 "general_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (AND, QImode, operands); DONE;")
-
 ;; %%% Potential partial reg stall on alternative 2.  What to do?
 (define_insn "*andqi_1"
   [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q,r")
        (and:QI (match_dup 0)
                (match_operand:QI 1 "general_operand" "qn,qmn")))
    (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
    && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
   "and{b}\t{%1, %0|%0, %1}"
   [(set_attr "type" "alu1")
    (set_attr "mode" "QI")])
 
+(define_split
+  [(set (match_operand 0 "register_operand" "")
+       (and (match_dup 0)
+            (const_int -65536)))
+   (clobber (reg:CC FLAGS_REG))]
+  "(TARGET_FAST_PREFIX && !TARGET_PARTIAL_REG_STALL)
+    || optimize_function_for_size_p (cfun)"
+  [(set (strict_low_part (match_dup 1)) (const_int 0))]
+  "operands[1] = gen_lowpart (HImode, operands[0]);")
+
+(define_split
+  [(set (match_operand 0 "ext_register_operand" "")
+       (and (match_dup 0)
+            (const_int -256)))
+   (clobber (reg:CC FLAGS_REG))]
+  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && reload_completed"
+  [(set (strict_low_part (match_dup 1)) (const_int 0))]
+  "operands[1] = gen_lowpart (QImode, operands[0]);")
+
+(define_split
+  [(set (match_operand 0 "ext_register_operand" "")
+       (and (match_dup 0)
+            (const_int -65281)))
+   (clobber (reg:CC FLAGS_REG))]
+  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && reload_completed"
+  [(parallel [(set (zero_extract:SI (match_dup 0)
+                                   (const_int 8)
+                                   (const_int 8))
+                  (xor:SI
+                    (zero_extract:SI (match_dup 0)
+                                     (const_int 8)
+                                     (const_int 8))
+                    (zero_extract:SI (match_dup 0)
+                                     (const_int 8)
+                                     (const_int 8))))
+             (clobber (reg:CC FLAGS_REG))])]
+  "operands[0] = gen_lowpart (SImode, operands[0]);")
+
+(define_insn "*anddi_2"
+  [(set (reg FLAGS_REG)
+       (compare
+        (and:DI
+         (match_operand:DI 1 "nonimmediate_operand" "%0,0,0")
+         (match_operand:DI 2 "x86_64_szext_general_operand" "Z,rem,re"))
+        (const_int 0)))
+   (set (match_operand:DI 0 "nonimmediate_operand" "=r,r,rm")
+       (and:DI (match_dup 1) (match_dup 2)))]
+  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
+   && ix86_binary_operator_ok (AND, DImode, operands)"
+  "@
+   and{l}\t{%k2, %k0|%k0, %k2}
+   and{q}\t{%2, %0|%0, %2}
+   and{q}\t{%2, %0|%0, %2}"
+  [(set_attr "type" "alu")
+   (set_attr "mode" "SI,DI,DI")])
+
 (define_insn "*andqi_2_maybe_si"
   [(set (reg FLAGS_REG)
        (compare (and:QI
-                     (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
-                     (match_operand:QI 2 "general_operand" "qmn,qn,n"))
+                 (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
+                 (match_operand:QI 2 "general_operand" "qmn,qn,n"))
                 (const_int 0)))
    (set (match_operand:QI 0 "nonimmediate_operand" "=q,qm,*r")
        (and:QI (match_dup 1) (match_dup 2)))]
   [(set_attr "type" "alu")
    (set_attr "mode" "QI,QI,SI")])
 
-(define_insn "*andqi_2"
+(define_insn "*and<mode>_2"
   [(set (reg FLAGS_REG)
-       (compare (and:QI
-                  (match_operand:QI 1 "nonimmediate_operand" "%0,0")
-                  (match_operand:QI 2 "general_operand" "qmn,qn"))
+       (compare (and:SWI124
+                 (match_operand:SWI124 1 "nonimmediate_operand" "%0,0")
+                 (match_operand:SWI124 2 "general_operand" "<g>,<r><i>"))
                 (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=q,qm")
-       (and:QI (match_dup 1) (match_dup 2)))]
+   (set (match_operand:SWI124 0 "nonimmediate_operand" "=<r>,<r>m")
+       (and:SWI124 (match_dup 1) (match_dup 2)))]
   "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (AND, QImode, operands)"
-  "and{b}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (AND, <MODE>mode, operands)"
+  "and{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
+   (set_attr "mode" "<MODE>")])
+
+;; See comment for addsi_1_zext why we do use nonimmediate_operand
+(define_insn "*andsi_2_zext"
+  [(set (reg FLAGS_REG)
+       (compare (and:SI
+                 (match_operand:SI 1 "nonimmediate_operand" "%0")
+                 (match_operand:SI 2 "general_operand" "g"))
+                (const_int 0)))
+   (set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI (and:SI (match_dup 1) (match_dup 2))))]
+  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
+   && ix86_binary_operator_ok (AND, SImode, operands)"
+  "and{l}\t{%2, %k0|%k0, %2}"
+  [(set_attr "type" "alu")
+   (set_attr "mode" "SI")])
 
 (define_insn "*andqi_2_slp"
   [(set (reg FLAGS_REG)
                 (const_int 0)))
    (set (strict_low_part (match_dup 0))
        (and:QI (match_dup 0) (match_dup 1)))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
    && ix86_match_ccmode (insn, CCNOmode)
    && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
   "and{b}\t{%1, %0|%0, %1}"
 ;; ??? A bug in recog prevents it from recognizing a const_int as an
 ;; operand to zero_extend in andqi_ext_1.  It was checking explicitly
 ;; for a QImode operand, which of course failed.
-
 (define_insn "andqi_ext_0"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
 
 ;; Generated by peephole translating test to and.  This shows up
 ;; often in fp comparisons.
-
 (define_insn "*andqi_ext_0_cc"
   [(set (reg FLAGS_REG)
        (compare
    (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
 
-(define_insn "*andqi_ext_1"
+(define_insn "*andqi_ext_1_rex64"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
            (const_int 8)
            (const_int 8))
          (zero_extend:SI
-           (match_operand:QI 2 "general_operand" "Qm"))))
+           (match_operand 2 "ext_register_operand" "Q"))))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT"
+  "TARGET_64BIT"
   "and{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "length_immediate" "0")
    (set_attr "mode" "QI")])
 
-(define_insn "*andqi_ext_1_rex64"
+(define_insn "*andqi_ext_1"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
            (const_int 8)
            (const_int 8))
          (zero_extend:SI
-           (match_operand 2 "ext_register_operand" "Q"))))
+           (match_operand:QI 2 "general_operand" "Qm"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
+  "!TARGET_64BIT"
   "and{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "length_immediate" "0")
    operands[1] = gen_lowpart (QImode, operands[1]);
    operands[2] = gen_lowpart (QImode, operands[2]);")
 \f
-;; Logical inclusive OR instructions
+;; Logical inclusive and exclusive OR instructions
 
 ;; %%% This used to optimize known byte-wide and operations to memory.
 ;; If this is considered useful, it should be done with splitters.
 
-(define_expand "iordi3"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (ior:DI (match_operand:DI 1 "nonimmediate_operand" "")
-               (match_operand:DI 2 "x86_64_general_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (IOR, DImode, operands); DONE;")
+(define_expand "<code><mode>3"
+  [(set (match_operand:SWIM 0 "nonimmediate_operand" "")
+       (any_or:SWIM (match_operand:SWIM 1 "nonimmediate_operand" "")
+                    (match_operand:SWIM 2 "<general_operand>" "")))]
+  ""
+  "ix86_expand_binary_operator (<CODE>, <MODE>mode, operands); DONE;")
 
-(define_insn "*iordi_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (ior:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
-               (match_operand:DI 2 "x86_64_general_operand" "re,rme")))
+(define_insn "*<code><mode>_1"
+  [(set (match_operand:SWI248 0 "nonimmediate_operand" "=r,rm")
+       (any_or:SWI248
+        (match_operand:SWI248 1 "nonimmediate_operand" "%0,0")
+        (match_operand:SWI248 2 "<general_operand>" "<g>,r<i>")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && ix86_binary_operator_ok (IOR, DImode, operands)"
-  "or{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*iordi_2_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (ior:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:DI 2 "x86_64_general_operand" "rem,re"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=r,rm")
-       (ior:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, DImode, operands)"
-  "or{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*iordi_3_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (ior:DI (match_operand:DI 1 "nonimmediate_operand" "%0")
-                        (match_operand:DI 2 "x86_64_general_operand" "rem"))
-                (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, DImode, operands)"
-  "or{q}\t{%2, %0|%0, %2}"
+  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "<logicprefix>{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-
-(define_expand "iorsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (ior:SI (match_operand:SI 1 "nonimmediate_operand" "")
-               (match_operand:SI 2 "general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (IOR, SImode, operands); DONE;")
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*iorsi_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
-               (match_operand:SI 2 "general_operand" "ri,g")))
+;; %%% Potential partial reg stall on alternative 2.  What to do?
+(define_insn "*<code>qi_1"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=q,m,r")
+       (any_or:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
+                  (match_operand:QI 2 "general_operand" "qmn,qn,rn")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (IOR, SImode, operands)"
-  "or{l}\t{%2, %0|%0, %2}"
+  "ix86_binary_operator_ok (<CODE>, QImode, operands)"
+  "@
+   <logicprefix>{b}\t{%2, %0|%0, %2}
+   <logicprefix>{b}\t{%2, %0|%0, %2}
+   <logicprefix>{l}\t{%k2, %k0|%k0, %k2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "QI,QI,SI")])
 
 ;; See comment for addsi_1_zext why we do use nonimmediate_operand
-(define_insn "*iorsi_1_zext"
+(define_insn "*<code>si_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
-         (ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                 (match_operand:SI 2 "general_operand" "g"))))
+        (any_or:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
+                   (match_operand:SI 2 "general_operand" "g"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (IOR, SImode, operands)"
-  "or{l}\t{%2, %k0|%k0, %2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+  "<logicprefix>{l}\t{%2, %k0|%k0, %2}"
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_insn "*iorsi_1_zext_imm"
+(define_insn "*<code>si_1_zext_imm"
   [(set (match_operand:DI 0 "register_operand" "=r")
-       (ior:DI (zero_extend:DI (match_operand:SI 1 "register_operand" "%0"))
-               (match_operand:DI 2 "x86_64_zext_immediate_operand" "Z")))
+       (any_or:DI
+        (zero_extend:DI (match_operand:SI 1 "register_operand" "%0"))
+        (match_operand:DI 2 "x86_64_zext_immediate_operand" "Z")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "or{l}\t{%2, %k0|%k0, %2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+  "<logicprefix>{l}\t{%2, %k0|%k0, %2}"
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_insn "*iorsi_2"
+(define_insn "*<code>qi_1_slp"
+  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+q,m"))
+       (any_or:QI (match_dup 0)
+                  (match_operand:QI 1 "general_operand" "qmn,qn")))
+   (clobber (reg:CC FLAGS_REG))]
+  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+  "<logicprefix>{b}\t{%1, %0|%0, %1}"
+  [(set_attr "type" "alu1")
+   (set_attr "mode" "QI")])
+
+(define_insn "*<code><mode>_2"
   [(set (reg FLAGS_REG)
-       (compare (ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:SI 2 "general_operand" "g,ri"))
+       (compare (any_or:SWI
+                 (match_operand:SWI 1 "nonimmediate_operand" "%0,0")
+                 (match_operand:SWI 2 "<general_operand>" "<g>,<r><i>"))
                 (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=r,rm")
-       (ior:SI (match_dup 1) (match_dup 2)))]
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>,<r>m")
+       (any_or:SWI (match_dup 1) (match_dup 2)))]
   "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, SImode, operands)"
-  "or{l}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "<logicprefix>{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 ;; See comment for addsi_1_zext why we do use nonimmediate_operand
 ;; ??? Special case for immediate operand is missing - it is tricky.
-(define_insn "*iorsi_2_zext"
+(define_insn "*<code>si_2_zext"
   [(set (reg FLAGS_REG)
-       (compare (ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand:SI 2 "general_operand" "g"))
+       (compare (any_or:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
+                           (match_operand:SI 2 "general_operand" "g"))
                 (const_int 0)))
    (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (ior:SI (match_dup 1) (match_dup 2))))]
+       (zero_extend:DI (any_or:SI (match_dup 1) (match_dup 2))))]
   "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, SImode, operands)"
-  "or{l}\t{%2, %k0|%k0, %2}"
+   && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+  "<logicprefix>{l}\t{%2, %k0|%k0, %2}"
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_insn "*iorsi_2_zext_imm"
+(define_insn "*<code>si_2_zext_imm"
   [(set (reg FLAGS_REG)
-       (compare (ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand 2 "x86_64_zext_immediate_operand" "Z"))
+       (compare (any_or:SI
+                 (match_operand:SI 1 "nonimmediate_operand" "%0")
+                 (match_operand:SI 2 "x86_64_zext_immediate_operand" "Z"))
                 (const_int 0)))
    (set (match_operand:DI 0 "register_operand" "=r")
-       (ior:DI (zero_extend:DI (match_dup 1)) (match_dup 2)))]
+       (any_or:DI (zero_extend:DI (match_dup 1)) (match_dup 2)))]
   "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, SImode, operands)"
-  "or{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*iorsi_3"
-  [(set (reg FLAGS_REG)
-       (compare (ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand:SI 2 "general_operand" "g"))
-                (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "or{l}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+  "<logicprefix>{l}\t{%2, %k0|%k0, %2}"
   [(set_attr "type" "alu")
    (set_attr "mode" "SI")])
 
-(define_expand "iorhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (ior:HI (match_operand:HI 1 "nonimmediate_operand" "")
-               (match_operand:HI 2 "general_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (IOR, HImode, operands); DONE;")
-
-(define_insn "*iorhi_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=r,m")
-       (ior:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-               (match_operand:HI 2 "general_operand" "rmn,rn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (IOR, HImode, operands)"
-  "or{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_insn "*iorhi_2"
-  [(set (reg FLAGS_REG)
-       (compare (ior:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:HI 2 "general_operand" "rmn,rn"))
-                (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=r,rm")
-       (ior:HI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, HImode, operands)"
-  "or{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_insn "*iorhi_3"
-  [(set (reg FLAGS_REG)
-       (compare (ior:HI (match_operand:HI 1 "nonimmediate_operand" "%0")
-                        (match_operand:HI 2 "general_operand" "rmn"))
-                (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "or{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_expand "iorqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (ior:QI (match_operand:QI 1 "nonimmediate_operand" "")
-               (match_operand:QI 2 "general_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (IOR, QImode, operands); DONE;")
-
-;; %%% Potential partial reg stall on alternative 2.  What to do?
-(define_insn "*iorqi_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=q,m,r")
-       (ior:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
-               (match_operand:QI 2 "general_operand" "qmn,qn,rn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (IOR, QImode, operands)"
-  "@
-   or{b}\t{%2, %0|%0, %2}
-   or{b}\t{%2, %0|%0, %2}
-   or{l}\t{%k2, %k0|%k0, %k2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "QI,QI,SI")])
-
-(define_insn "*iorqi_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+q,m"))
-       (ior:QI (match_dup 0)
-               (match_operand:QI 1 "general_operand" "qmn,qn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "or{b}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*iorqi_2"
-  [(set (reg FLAGS_REG)
-       (compare (ior:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:QI 2 "general_operand" "qmn,qn"))
-                (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=q,qm")
-       (ior:QI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (IOR, QImode, operands)"
-  "or{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
-
-(define_insn "*iorqi_2_slp"
+(define_insn "*<code>qi_2_slp"
   [(set (reg FLAGS_REG)
-       (compare (ior:QI (match_operand:QI 0 "nonimmediate_operand" "+q,qm")
-                        (match_operand:QI 1 "general_operand" "qmn,qn"))
+       (compare (any_or:QI (match_operand:QI 0 "nonimmediate_operand" "+q,qm")
+                           (match_operand:QI 1 "general_operand" "qmn,qn"))
                 (const_int 0)))
    (set (strict_low_part (match_dup 0))
-       (ior:QI (match_dup 0) (match_dup 1)))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
+       (any_or:QI (match_dup 0) (match_dup 1)))]
+  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
    && ix86_match_ccmode (insn, CCNOmode)
    && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "or{b}\t{%1, %0|%0, %1}"
+  "<logicprefix>{b}\t{%1, %0|%0, %1}"
   [(set_attr "type" "alu1")
    (set_attr "mode" "QI")])
 
-(define_insn "*iorqi_3"
+(define_insn "*<code><mode>_3"
   [(set (reg FLAGS_REG)
-       (compare (ior:QI (match_operand:QI 1 "nonimmediate_operand" "%0")
-                        (match_operand:QI 2 "general_operand" "qmn"))
+       (compare (any_or:SWI
+                 (match_operand:SWI 1 "nonimmediate_operand" "%0")
+                 (match_operand:SWI 2 "<general_operand>" "<g>"))
                 (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
+   (clobber (match_scratch:SWI 0 "=<r>"))]
   "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "or{b}\t{%2, %0|%0, %2}"
+   && ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+  "<logicprefix>{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*iorqi_ext_0"
+(define_insn "*<code>qi_ext_0"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
-       (ior:SI
+       (any_or:SI
          (zero_extract:SI
            (match_operand 1 "ext_register_operand" "0")
            (const_int 8)
            (const_int 8))
          (match_operand 2 "const_int_operand" "n")))
    (clobber (reg:CC FLAGS_REG))]
-  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "or{b}\t{%2, %h0|%h0, %2}"
+  "!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun)"
+  "<logicprefix>{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "length_immediate" "1")
    (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
 
-(define_insn "*iorqi_ext_1"
+(define_insn "*<code>qi_ext_1_rex64"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
-       (ior:SI
+       (any_or:SI
          (zero_extract:SI
            (match_operand 1 "ext_register_operand" "0")
            (const_int 8)
            (const_int 8))
          (zero_extend:SI
-           (match_operand:QI 2 "general_operand" "Qm"))))
+           (match_operand 2 "ext_register_operand" "Q"))))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT
+  "TARGET_64BIT
    && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "or{b}\t{%2, %h0|%h0, %2}"
+  "<logicprefix>{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "length_immediate" "0")
    (set_attr "mode" "QI")])
 
-(define_insn "*iorqi_ext_1_rex64"
+(define_insn "*<code>qi_ext_1"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
-       (ior:SI
+       (any_or:SI
          (zero_extract:SI
            (match_operand 1 "ext_register_operand" "0")
            (const_int 8)
            (const_int 8))
          (zero_extend:SI
-           (match_operand 2 "ext_register_operand" "Q"))))
+           (match_operand:QI 2 "general_operand" "Qm"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
+  "!TARGET_64BIT
    && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "or{b}\t{%2, %h0|%h0, %2}"
+  "<logicprefix>{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "length_immediate" "0")
    (set_attr "mode" "QI")])
 
-(define_insn "*iorqi_ext_2"
+(define_insn "*<code>qi_ext_2"
   [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
-       (ior:SI
+       (any_or:SI
          (zero_extract:SI (match_operand 1 "ext_register_operand" "0")
                           (const_int 8)
                           (const_int 8))
                           (const_int 8)
                           (const_int 8))))
    (clobber (reg:CC FLAGS_REG))]
-  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "ior{b}\t{%h2, %h0|%h0, %h2}"
+  "!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun)"
+  "<logicprefix>{b}\t{%h2, %h0|%h0, %h2}"
   [(set_attr "type" "alu")
    (set_attr "length_immediate" "0")
    (set_attr "mode" "QI")])
 
 (define_split
   [(set (match_operand 0 "register_operand" "")
-       (ior (match_operand 1 "register_operand" "")
-            (match_operand 2 "const_int_operand" "")))
+       (any_or (match_operand 1 "register_operand" "")
+               (match_operand 2 "const_int_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
    "reload_completed
     && QI_REG_P (operands[0])
     && !(INTVAL (operands[2]) & ~(255 << 8))
     && GET_MODE (operands[0]) != QImode"
   [(parallel [(set (zero_extract:SI (match_dup 0) (const_int 8) (const_int 8))
-                  (ior:SI (zero_extract:SI (match_dup 1)
-                                           (const_int 8) (const_int 8))
-                          (match_dup 2)))
+                  (any_or:SI (zero_extract:SI (match_dup 1)
+                                              (const_int 8) (const_int 8))
+                             (match_dup 2)))
              (clobber (reg:CC FLAGS_REG))])]
   "operands[0] = gen_lowpart (SImode, operands[0]);
    operands[1] = gen_lowpart (SImode, operands[1]);
 ;; profitable when 7th bit is set.
 (define_split
   [(set (match_operand 0 "register_operand" "")
-       (ior (match_operand 1 "general_operand" "")
-            (match_operand 2 "const_int_operand" "")))
+       (any_or (match_operand 1 "general_operand" "")
+               (match_operand 2 "const_int_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
    "reload_completed
     && ANY_QI_REG_P (operands[0])
     && (INTVAL (operands[2]) & 128)
     && GET_MODE (operands[0]) != QImode"
   [(parallel [(set (strict_low_part (match_dup 0))
-                  (ior:QI (match_dup 1)
-                          (match_dup 2)))
+                  (any_or:QI (match_dup 1)
+                             (match_dup 2)))
              (clobber (reg:CC FLAGS_REG))])]
   "operands[0] = gen_lowpart (QImode, operands[0]);
    operands[1] = gen_lowpart (QImode, operands[1]);
    operands[2] = gen_lowpart (QImode, operands[2]);")
-\f
-;; Logical XOR instructions
 
-;; %%% This used to optimize known byte-wide and operations to memory.
-;; If this is considered useful, it should be done with splitters.
-
-(define_expand "xordi3"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (xor:DI (match_operand:DI 1 "nonimmediate_operand" "")
-               (match_operand:DI 2 "x86_64_general_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (XOR, DImode, operands); DONE;")
-
-(define_insn "*xordi_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (xor:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
-               (match_operand:DI 2 "x86_64_general_operand" "re,rm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && ix86_binary_operator_ok (XOR, DImode, operands)"
-  "xor{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*xordi_2_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (xor:DI (match_operand:DI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:DI 2 "x86_64_general_operand" "rem,re"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=r,rm")
-       (xor:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, DImode, operands)"
-  "xor{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_insn "*xordi_3_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (xor:DI (match_operand:DI 1 "nonimmediate_operand" "%0")
-                        (match_operand:DI 2 "x86_64_general_operand" "rem"))
-                (const_int 0)))
-   (clobber (match_scratch:DI 0 "=r"))]
-  "TARGET_64BIT
-   && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, DImode, operands)"
-  "xor{q}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "DI")])
-
-(define_expand "xorsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (xor:SI (match_operand:SI 1 "nonimmediate_operand" "")
-               (match_operand:SI 2 "general_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (XOR, SImode, operands); DONE;")
-
-(define_insn "*xorsi_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
-               (match_operand:SI 2 "general_operand" "ri,rm")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (XOR, SImode, operands)"
-  "xor{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-;; See comment for addsi_1_zext why we do use nonimmediate_operand
-;; Add speccase for immediates
-(define_insn "*xorsi_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI
-         (xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                 (match_operand:SI 2 "general_operand" "g"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (XOR, SImode, operands)"
-  "xor{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*xorsi_1_zext_imm"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (xor:DI (zero_extend:DI (match_operand:SI 1 "register_operand" "%0"))
-               (match_operand:DI 2 "x86_64_zext_immediate_operand" "Z")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (XOR, SImode, operands)"
-  "xor{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*xorsi_2"
-  [(set (reg FLAGS_REG)
-       (compare (xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:SI 2 "general_operand" "g,ri"))
-                (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=r,rm")
-       (xor:SI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, SImode, operands)"
-  "xor{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-;; See comment for addsi_1_zext why we do use nonimmediate_operand
-;; ??? Special case for immediate operand is missing - it is tricky.
-(define_insn "*xorsi_2_zext"
-  [(set (reg FLAGS_REG)
-       (compare (xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand:SI 2 "general_operand" "g"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (xor:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, SImode, operands)"
-  "xor{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*xorsi_2_zext_imm"
-  [(set (reg FLAGS_REG)
-       (compare (xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand 2 "x86_64_zext_immediate_operand" "Z"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
-       (xor:DI (zero_extend:DI (match_dup 1)) (match_dup 2)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, SImode, operands)"
-  "xor{l}\t{%2, %k0|%k0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_insn "*xorsi_3"
-  [(set (reg FLAGS_REG)
-       (compare (xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                        (match_operand:SI 2 "general_operand" "g"))
-                (const_int 0)))
-   (clobber (match_scratch:SI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "xor{l}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "SI")])
-
-(define_expand "xorhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (xor:HI (match_operand:HI 1 "nonimmediate_operand" "")
-               (match_operand:HI 2 "general_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (XOR, HImode, operands); DONE;")
-
-(define_insn "*xorhi_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=r,m")
-       (xor:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-               (match_operand:HI 2 "general_operand" "rmn,rn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (XOR, HImode, operands)"
-  "xor{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_insn "*xorhi_2"
-  [(set (reg FLAGS_REG)
-       (compare (xor:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0")
-                        (match_operand:HI 2 "general_operand" "rmn,rn"))
-                (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=r,rm")
-       (xor:HI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, HImode, operands)"
-  "xor{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_insn "*xorhi_3"
-  [(set (reg FLAGS_REG)
-       (compare (xor:HI (match_operand:HI 1 "nonimmediate_operand" "%0")
-                        (match_operand:HI 2 "general_operand" "rmn"))
-                (const_int 0)))
-   (clobber (match_scratch:HI 0 "=r"))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "xor{w}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "HI")])
-
-(define_expand "xorqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (xor:QI (match_operand:QI 1 "nonimmediate_operand" "")
-               (match_operand:QI 2 "general_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (XOR, QImode, operands); DONE;")
-
-;; %%% Potential partial reg stall on alternative 2.  What to do?
-(define_insn "*xorqi_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=q,m,r")
-       (xor:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,0")
-               (match_operand:QI 2 "general_operand" "qmn,qn,rn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (XOR, QImode, operands)"
-  "@
-   xor{b}\t{%2, %0|%0, %2}
-   xor{b}\t{%2, %0|%0, %2}
-   xor{l}\t{%k2, %k0|%k0, %k2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "QI,QI,SI")])
-
-(define_insn "*xorqi_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,q"))
-       (xor:QI (match_dup 0)
-               (match_operand:QI 1 "general_operand" "qn,qmn")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "xor{b}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_ext_0"
-  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
-                        (const_int 8)
-                        (const_int 8))
-       (xor:SI
-         (zero_extract:SI
-           (match_operand 1 "ext_register_operand" "0")
-           (const_int 8)
-           (const_int 8))
-         (match_operand 2 "const_int_operand" "n")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "xor{b}\t{%2, %h0|%h0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "length_immediate" "1")
-   (set_attr "modrm" "1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_ext_1"
-  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
-                        (const_int 8)
-                        (const_int 8))
-       (xor:SI
-         (zero_extract:SI
-           (match_operand 1 "ext_register_operand" "0")
-           (const_int 8)
-           (const_int 8))
-         (zero_extend:SI
-           (match_operand:QI 2 "general_operand" "Qm"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT
-   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "xor{b}\t{%2, %h0|%h0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_ext_1_rex64"
-  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
-                        (const_int 8)
-                        (const_int 8))
-       (xor:SI
-         (zero_extract:SI
-           (match_operand 1 "ext_register_operand" "0")
-           (const_int 8)
-           (const_int 8))
-         (zero_extend:SI
-           (match_operand 2 "ext_register_operand" "Q"))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "xor{b}\t{%2, %h0|%h0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_ext_2"
-  [(set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
-                        (const_int 8)
-                        (const_int 8))
-       (xor:SI
-         (zero_extract:SI (match_operand 1 "ext_register_operand" "0")
-                          (const_int 8)
+(define_expand "xorqi_cc_ext_1"
+  [(parallel [
+     (set (reg:CCNO FLAGS_REG)
+         (compare:CCNO
+           (xor:SI
+             (zero_extract:SI
+               (match_operand 1 "ext_register_operand" "")
+               (const_int 8)
+               (const_int 8))
+             (match_operand:QI 2 "general_operand" ""))
+           (const_int 0)))
+     (set (zero_extract:SI (match_operand 0 "ext_register_operand" "")
+                          (const_int 8)
                           (const_int 8))
-         (zero_extract:SI (match_operand 2 "ext_register_operand" "Q")
-                          (const_int 8)
-                          (const_int 8))))
-   (clobber (reg:CC FLAGS_REG))]
-  "(!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))"
-  "xor{b}\t{%h2, %h0|%h0, %h2}"
-  [(set_attr "type" "alu")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_cc_1"
-  [(set (reg FLAGS_REG)
-       (compare
-         (xor:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0")
-                 (match_operand:QI 2 "general_operand" "qmn,qn"))
-         (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=q,qm")
-       (xor:QI (match_dup 1) (match_dup 2)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_binary_operator_ok (XOR, QImode, operands)"
-  "xor{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_2_slp"
-  [(set (reg FLAGS_REG)
-       (compare (xor:QI (match_operand:QI 0 "nonimmediate_operand" "+q,qm")
-                        (match_operand:QI 1 "general_operand" "qmn,qn"))
-                (const_int 0)))
-   (set (strict_low_part (match_dup 0))
-       (xor:QI (match_dup 0) (match_dup 1)))]
-  "(! TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-   && ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "xor{b}\t{%1, %0|%0, %1}"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*xorqi_cc_2"
-  [(set (reg FLAGS_REG)
-       (compare
-         (xor:QI (match_operand:QI 1 "nonimmediate_operand" "%0")
-                 (match_operand:QI 2 "general_operand" "qmn"))
-         (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "xor{b}\t{%2, %0|%0, %2}"
-  [(set_attr "type" "alu")
-   (set_attr "mode" "QI")])
+         (xor:SI
+           (zero_extract:SI
+            (match_dup 1)
+            (const_int 8)
+            (const_int 8))
+           (match_dup 2)))])]
+  ""
+  "")
 
-(define_insn "*xorqi_cc_ext_1"
+(define_insn "*xorqi_cc_ext_1_rex64"
   [(set (reg FLAGS_REG)
        (compare
          (xor:SI
              (match_operand 1 "ext_register_operand" "0")
              (const_int 8)
              (const_int 8))
-           (match_operand:QI 2 "general_operand" "qmn"))
+           (match_operand:QI 2 "nonmemory_operand" "Qn"))
          (const_int 0)))
-   (set (zero_extract:SI (match_operand 0 "ext_register_operand" "=q")
+   (set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
                         (const_int 8)
                         (const_int 8))
        (xor:SI
-         (zero_extract:SI (match_dup 1) (const_int 8) (const_int 8))
+         (zero_extract:SI
+          (match_dup 1)
+          (const_int 8)
+          (const_int 8))
          (match_dup 2)))]
-  "!TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
+  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
   "xor{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
 
-(define_insn "*xorqi_cc_ext_1_rex64"
+(define_insn "*xorqi_cc_ext_1"
   [(set (reg FLAGS_REG)
        (compare
          (xor:SI
              (match_operand 1 "ext_register_operand" "0")
              (const_int 8)
              (const_int 8))
-           (match_operand:QI 2 "nonmemory_operand" "Qn"))
+           (match_operand:QI 2 "general_operand" "qmn"))
          (const_int 0)))
-   (set (zero_extract:SI (match_operand 0 "ext_register_operand" "=Q")
+   (set (zero_extract:SI (match_operand 0 "ext_register_operand" "=q")
                         (const_int 8)
                         (const_int 8))
        (xor:SI
-         (zero_extract:SI (match_dup 1) (const_int 8) (const_int 8))
+         (zero_extract:SI
+          (match_dup 1)
+          (const_int 8)
+          (const_int 8))
          (match_dup 2)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
+  "!TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
   "xor{b}\t{%2, %h0|%h0, %2}"
   [(set_attr "type" "alu")
    (set_attr "modrm" "1")
    (set_attr "mode" "QI")])
+\f
+;; Negation instructions
 
-(define_expand "xorqi_cc_ext_1"
-  [(parallel [
-     (set (reg:CCNO FLAGS_REG)
-         (compare:CCNO
-           (xor:SI
-             (zero_extract:SI
-               (match_operand 1 "ext_register_operand" "")
-               (const_int 8)
-               (const_int 8))
-             (match_operand:QI 2 "general_operand" ""))
-           (const_int 0)))
-     (set (zero_extract:SI (match_operand 0 "ext_register_operand" "")
-                          (const_int 8)
-                          (const_int 8))
-         (xor:SI
-           (zero_extract:SI (match_dup 1) (const_int 8) (const_int 8))
-           (match_dup 2)))])]
-  ""
-  "")
-
-(define_split
-  [(set (match_operand 0 "register_operand" "")
-       (xor (match_operand 1 "register_operand" "")
-            (match_operand 2 "const_int_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-   "reload_completed
-    && QI_REG_P (operands[0])
-    && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-    && !(INTVAL (operands[2]) & ~(255 << 8))
-    && GET_MODE (operands[0]) != QImode"
-  [(parallel [(set (zero_extract:SI (match_dup 0) (const_int 8) (const_int 8))
-                  (xor:SI (zero_extract:SI (match_dup 1)
-                                           (const_int 8) (const_int 8))
-                          (match_dup 2)))
-             (clobber (reg:CC FLAGS_REG))])]
-  "operands[0] = gen_lowpart (SImode, operands[0]);
-   operands[1] = gen_lowpart (SImode, operands[1]);
-   operands[2] = gen_int_mode ((INTVAL (operands[2]) >> 8) & 0xff, SImode);")
-
-;; Since XOR can be encoded with sign extended immediate, this is only
-;; profitable when 7th bit is set.
-(define_split
-  [(set (match_operand 0 "register_operand" "")
-       (xor (match_operand 1 "general_operand" "")
-            (match_operand 2 "const_int_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-   "reload_completed
-    && ANY_QI_REG_P (operands[0])
-    && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun))
-    && !(INTVAL (operands[2]) & ~255)
-    && (INTVAL (operands[2]) & 128)
-    && GET_MODE (operands[0]) != QImode"
-  [(parallel [(set (strict_low_part (match_dup 0))
-                  (xor:QI (match_dup 1)
-                          (match_dup 2)))
-             (clobber (reg:CC FLAGS_REG))])]
-  "operands[0] = gen_lowpart (QImode, operands[0]);
-   operands[1] = gen_lowpart (QImode, operands[1]);
-   operands[2] = gen_lowpart (QImode, operands[2]);")
-\f
-;; Negation instructions
-
-(define_expand "negti2"
-  [(set (match_operand:TI 0 "nonimmediate_operand" "")
-       (neg:TI (match_operand:TI 1 "nonimmediate_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_unary_operator (NEG, TImode, operands); DONE;")
-
-(define_insn "*negti2_1"
-  [(set (match_operand:TI 0 "nonimmediate_operand" "=ro")
-       (neg:TI (match_operand:TI 1 "nonimmediate_operand" "0")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT
-   && ix86_unary_operator_ok (NEG, TImode, operands)"
-  "#")
-
-(define_split
-  [(set (match_operand:TI 0 "nonimmediate_operand" "")
-       (neg:TI (match_operand:TI 1 "nonimmediate_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed"
-  [(parallel
-    [(set (reg:CCZ FLAGS_REG)
-         (compare:CCZ (neg:DI (match_dup 1)) (const_int 0)))
-     (set (match_dup 0) (neg:DI (match_dup 1)))])
-   (parallel
-    [(set (match_dup 2)
-         (plus:DI (plus:DI (ltu:DI (reg:CC FLAGS_REG) (const_int 0))
-                           (match_dup 3))
-                  (const_int 0)))
-     (clobber (reg:CC FLAGS_REG))])
-   (parallel
-    [(set (match_dup 2)
-         (neg:DI (match_dup 2)))
-     (clobber (reg:CC FLAGS_REG))])]
-  "split_ti (&operands[0], 2, &operands[0], &operands[2]);")
-
-(define_expand "negdi2"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (neg:DI (match_operand:DI 1 "nonimmediate_operand" "")))]
+(define_expand "neg<mode>2"
+  [(set (match_operand:SDWIM 0 "nonimmediate_operand" "")
+       (neg:SDWIM (match_operand:SDWIM 1 "nonimmediate_operand" "")))]
   ""
-  "ix86_expand_unary_operator (NEG, DImode, operands); DONE;")
-
-(define_insn "*negdi2_1"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=ro")
-       (neg:DI (match_operand:DI 1 "general_operand" "0")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT
-   && ix86_unary_operator_ok (NEG, DImode, operands)"
-  "#")
+  "ix86_expand_unary_operator (NEG, <MODE>mode, operands); DONE;")
 
-(define_split
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (neg:DI (match_operand:DI 1 "general_operand" "")))
+(define_insn_and_split "*neg<dwi>2_doubleword"
+  [(set (match_operand:<DWI> 0 "nonimmediate_operand" "=ro")
+       (neg:<DWI> (match_operand:<DWI> 1 "nonimmediate_operand" "0")))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && reload_completed"
+  "ix86_unary_operator_ok (NEG, <DWI>mode, operands)"
+  "#"
+  "reload_completed"
   [(parallel
     [(set (reg:CCZ FLAGS_REG)
-         (compare:CCZ (neg:SI (match_dup 1)) (const_int 0)))
-     (set (match_dup 0) (neg:SI (match_dup 1)))])
+         (compare:CCZ (neg:DWIH (match_dup 1)) (const_int 0)))
+     (set (match_dup 0) (neg:DWIH (match_dup 1)))])
    (parallel
     [(set (match_dup 2)
-         (plus:SI (plus:SI (ltu:SI (reg:CC FLAGS_REG) (const_int 0))
-                           (match_dup 3))
-                  (const_int 0)))
+         (plus:DWIH (match_dup 3)
+                    (plus:DWIH (ltu:DWIH (reg:CC FLAGS_REG) (const_int 0))
+                               (const_int 0))))
      (clobber (reg:CC FLAGS_REG))])
    (parallel
     [(set (match_dup 2)
-         (neg:SI (match_dup 2)))
+         (neg:DWIH (match_dup 2)))
      (clobber (reg:CC FLAGS_REG))])]
-  "split_di (&operands[0], 2, &operands[0], &operands[2]);");
-
-(define_insn "*negdi2_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (neg:DI (match_operand:DI 1 "nonimmediate_operand" "0")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_unary_operator_ok (NEG, DImode, operands)"
-  "neg{q}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "DI")])
-
-;; The problem with neg is that it does not perform (compare x 0),
-;; it really performs (compare 0 x), which leaves us with the zero
-;; flag being the only useful item.
-
-(define_insn "*negdi2_cmpz_rex64"
-  [(set (reg:CCZ FLAGS_REG)
-       (compare:CCZ (neg:DI (match_operand:DI 1 "nonimmediate_operand" "0"))
-                    (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (neg:DI (match_dup 1)))]
-  "TARGET_64BIT && ix86_unary_operator_ok (NEG, DImode, operands)"
-  "neg{q}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "DI")])
-
-
-(define_expand "negsi2"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (neg:SI (match_operand:SI 1 "nonimmediate_operand" "")))]
-  ""
-  "ix86_expand_unary_operator (NEG, SImode, operands); DONE;")
+  "split_<dwi> (&operands[0], 2, &operands[0], &operands[2]);")
 
-(define_insn "*negsi2_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (neg:SI (match_operand:SI 1 "nonimmediate_operand" "0")))
+(define_insn "*neg<mode>2_1"
+  [(set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (neg:SWI (match_operand:SWI 1 "nonimmediate_operand" "0")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_unary_operator_ok (NEG, SImode, operands)"
-  "neg{l}\t%0"
+  "ix86_unary_operator_ok (NEG, <MODE>mode, operands)"
+  "neg{<imodesuffix>}\t%0"
   [(set_attr "type" "negnot")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 ;; Combine is quite creative about this pattern.
 (define_insn "*negsi2_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
-       (lshiftrt:DI (neg:DI (ashift:DI (match_operand:DI 1 "register_operand" "0")
-                                       (const_int 32)))
-                    (const_int 32)))
+       (lshiftrt:DI
+         (neg:DI (ashift:DI (match_operand:DI 1 "register_operand" "0")
+                            (const_int 32)))
+       (const_int 32)))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_unary_operator_ok (NEG, SImode, operands)"
   "neg{l}\t%k0"
 ;; it really performs (compare 0 x), which leaves us with the zero
 ;; flag being the only useful item.
 
-(define_insn "*negsi2_cmpz"
+(define_insn "*neg<mode>2_cmpz"
   [(set (reg:CCZ FLAGS_REG)
-       (compare:CCZ (neg:SI (match_operand:SI 1 "nonimmediate_operand" "0"))
-                    (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (neg:SI (match_dup 1)))]
-  "ix86_unary_operator_ok (NEG, SImode, operands)"
-  "neg{l}\t%0"
+       (compare:CCZ
+         (neg:SWI (match_operand:SWI 1 "nonimmediate_operand" "0"))
+                  (const_int 0)))
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (neg:SWI (match_dup 1)))]
+  "ix86_unary_operator_ok (NEG, <MODE>mode, operands)"
+  "neg{<imodesuffix>}\t%0"
   [(set_attr "type" "negnot")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_insn "*negsi2_cmpz_zext"
   [(set (reg:CCZ FLAGS_REG)
-       (compare:CCZ (lshiftrt:DI
-                      (neg:DI (ashift:DI
-                                (match_operand:DI 1 "register_operand" "0")
-                                (const_int 32)))
-                      (const_int 32))
-                    (const_int 0)))
+       (compare:CCZ
+         (lshiftrt:DI
+           (neg:DI (ashift:DI
+                     (match_operand:DI 1 "register_operand" "0")
+                     (const_int 32)))
+           (const_int 32))
+         (const_int 0)))
    (set (match_operand:DI 0 "register_operand" "=r")
        (lshiftrt:DI (neg:DI (ashift:DI (match_dup 1)
                                        (const_int 32)))
   [(set_attr "type" "negnot")
    (set_attr "mode" "SI")])
 
-(define_expand "neghi2"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (neg:HI (match_operand:HI 1 "nonimmediate_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_unary_operator (NEG, HImode, operands); DONE;")
-
-(define_insn "*neghi2_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (neg:HI (match_operand:HI 1 "nonimmediate_operand" "0")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_unary_operator_ok (NEG, HImode, operands)"
-  "neg{w}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "HI")])
-
-(define_insn "*neghi2_cmpz"
-  [(set (reg:CCZ FLAGS_REG)
-       (compare:CCZ (neg:HI (match_operand:HI 1 "nonimmediate_operand" "0"))
-                    (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (neg:HI (match_dup 1)))]
-  "ix86_unary_operator_ok (NEG, HImode, operands)"
-  "neg{w}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "HI")])
-
-(define_expand "negqi2"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (neg:QI (match_operand:QI 1 "nonimmediate_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_unary_operator (NEG, QImode, operands); DONE;")
-
-(define_insn "*negqi2_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (neg:QI (match_operand:QI 1 "nonimmediate_operand" "0")))
-   (clobber (reg:CC FLAGS_REG))]
-  "ix86_unary_operator_ok (NEG, QImode, operands)"
-  "neg{b}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "QI")])
-
-(define_insn "*negqi2_cmpz"
-  [(set (reg:CCZ FLAGS_REG)
-       (compare:CCZ (neg:QI (match_operand:QI 1 "nonimmediate_operand" "0"))
-                    (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (neg:QI (match_dup 1)))]
-  "ix86_unary_operator_ok (NEG, QImode, operands)"
-  "neg{b}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "QI")])
-
 ;; Changing of sign for FP values is doable using integer unit too.
 
 (define_expand "<code><mode>2"
 \f
 ;; One complement instructions
 
-(define_expand "one_cmpldi2"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "")
-       (not:DI (match_operand:DI 1 "nonimmediate_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_unary_operator (NOT, DImode, operands); DONE;")
+(define_expand "one_cmpl<mode>2"
+  [(set (match_operand:SWIM 0 "nonimmediate_operand" "")
+       (not:SWIM (match_operand:SWIM 1 "nonimmediate_operand" "")))]
+  ""
+  "ix86_expand_unary_operator (NOT, <MODE>mode, operands); DONE;")
 
-(define_insn "*one_cmpldi2_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (not:DI (match_operand:DI 1 "nonimmediate_operand" "0")))]
-  "TARGET_64BIT && ix86_unary_operator_ok (NOT, DImode, operands)"
-  "not{q}\t%0"
+(define_insn "*one_cmpl<mode>2_1"
+  [(set (match_operand:SWI248 0 "nonimmediate_operand" "=rm")
+       (not:SWI248 (match_operand:SWI248 1 "nonimmediate_operand" "0")))]
+  "ix86_unary_operator_ok (NOT, <MODE>mode, operands)"
+  "not{<imodesuffix>}\t%0"
   [(set_attr "type" "negnot")
-   (set_attr "mode" "DI")])
-
-(define_insn "*one_cmpldi2_2_rex64"
-  [(set (reg FLAGS_REG)
-       (compare (not:DI (match_operand:DI 1 "nonimmediate_operand" "0"))
-                (const_int 0)))
-   (set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (not:DI (match_dup 1)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)
-   && ix86_unary_operator_ok (NOT, DImode, operands)"
-  "#"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "DI")])
-
-(define_split
-  [(set (match_operand 0 "flags_reg_operand" "")
-       (match_operator 2 "compare_operator"
-         [(not:DI (match_operand:DI 3 "nonimmediate_operand" ""))
-          (const_int 0)]))
-   (set (match_operand:DI 1 "nonimmediate_operand" "")
-       (not:DI (match_dup 3)))]
-  "TARGET_64BIT && ix86_match_ccmode (insn, CCNOmode)"
-  [(parallel [(set (match_dup 0)
-                  (match_op_dup 2
-                    [(xor:DI (match_dup 3) (const_int -1))
-                     (const_int 0)]))
-             (set (match_dup 1)
-                  (xor:DI (match_dup 3) (const_int -1)))])]
-  "")
-
-(define_expand "one_cmplsi2"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (not:SI (match_operand:SI 1 "nonimmediate_operand" "")))]
-  ""
-  "ix86_expand_unary_operator (NOT, SImode, operands); DONE;")
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*one_cmplsi2_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (not:SI (match_operand:SI 1 "nonimmediate_operand" "0")))]
-  "ix86_unary_operator_ok (NOT, SImode, operands)"
-  "not{l}\t%0"
+;; %%% Potential partial reg stall on alternative 1.  What to do?
+(define_insn "*one_cmplqi2_1"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r")
+       (not:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")))]
+  "ix86_unary_operator_ok (NOT, QImode, operands)"
+  "@
+   not{b}\t%0
+   not{l}\t%k0"
   [(set_attr "type" "negnot")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "QI,SI")])
 
 ;; ??? Currently never generated - xor is used instead.
 (define_insn "*one_cmplsi2_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI (not:SI (match_operand:SI 1 "register_operand" "0"))))]
+       (zero_extend:DI
+         (not:SI (match_operand:SI 1 "register_operand" "0"))))]
   "TARGET_64BIT && ix86_unary_operator_ok (NOT, SImode, operands)"
   "not{l}\t%k0"
   [(set_attr "type" "negnot")
    (set_attr "mode" "SI")])
 
-(define_insn "*one_cmplsi2_2"
+(define_insn "*one_cmpl<mode>2_2"
   [(set (reg FLAGS_REG)
-       (compare (not:SI (match_operand:SI 1 "nonimmediate_operand" "0"))
+       (compare (not:SWI (match_operand:SWI 1 "nonimmediate_operand" "0"))
                 (const_int 0)))
-   (set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (not:SI (match_dup 1)))]
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (not:SWI (match_dup 1)))]
   "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_unary_operator_ok (NOT, SImode, operands)"
+   && ix86_unary_operator_ok (NOT, <MODE>mode, operands)"
   "#"
   [(set_attr "type" "alu1")
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
 (define_split
   [(set (match_operand 0 "flags_reg_operand" "")
        (match_operator 2 "compare_operator"
-         [(not:SI (match_operand:SI 3 "nonimmediate_operand" ""))
+         [(not:SWI (match_operand:SWI 3 "nonimmediate_operand" ""))
           (const_int 0)]))
-   (set (match_operand:SI 1 "nonimmediate_operand" "")
-       (not:SI (match_dup 3)))]
+   (set (match_operand:SWI 1 "nonimmediate_operand" "")
+       (not:SWI (match_dup 3)))]
   "ix86_match_ccmode (insn, CCNOmode)"
   [(parallel [(set (match_dup 0)
-                  (match_op_dup 2 [(xor:SI (match_dup 3) (const_int -1))
+                  (match_op_dup 2 [(xor:SWI (match_dup 3) (const_int -1))
                                    (const_int 0)]))
              (set (match_dup 1)
-                  (xor:SI (match_dup 3) (const_int -1)))])]
+                  (xor:SWI (match_dup 3) (const_int -1)))])]
   "")
 
 ;; ??? Currently never generated - xor is used instead.
              (set (match_dup 1)
                   (zero_extend:DI (xor:SI (match_dup 3) (const_int -1))))])]
   "")
-
-(define_expand "one_cmplhi2"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (not:HI (match_operand:HI 1 "nonimmediate_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_unary_operator (NOT, HImode, operands); DONE;")
-
-(define_insn "*one_cmplhi2_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (not:HI (match_operand:HI 1 "nonimmediate_operand" "0")))]
-  "ix86_unary_operator_ok (NOT, HImode, operands)"
-  "not{w}\t%0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "HI")])
-
-(define_insn "*one_cmplhi2_2"
-  [(set (reg FLAGS_REG)
-       (compare (not:HI (match_operand:HI 1 "nonimmediate_operand" "0"))
-                (const_int 0)))
-   (set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (not:HI (match_dup 1)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_unary_operator_ok (NEG, HImode, operands)"
-  "#"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "HI")])
-
-(define_split
-  [(set (match_operand 0 "flags_reg_operand" "")
-       (match_operator 2 "compare_operator"
-         [(not:HI (match_operand:HI 3 "nonimmediate_operand" ""))
-          (const_int 0)]))
-   (set (match_operand:HI 1 "nonimmediate_operand" "")
-       (not:HI (match_dup 3)))]
-  "ix86_match_ccmode (insn, CCNOmode)"
-  [(parallel [(set (match_dup 0)
-                  (match_op_dup 2 [(xor:HI (match_dup 3) (const_int -1))
-                                   (const_int 0)]))
-             (set (match_dup 1)
-                  (xor:HI (match_dup 3) (const_int -1)))])]
-  "")
-
-;; %%% Potential partial reg stall on alternative 1.  What to do?
-(define_expand "one_cmplqi2"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (not:QI (match_operand:QI 1 "nonimmediate_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_unary_operator (NOT, QImode, operands); DONE;")
-
-(define_insn "*one_cmplqi2_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r")
-       (not:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")))]
-  "ix86_unary_operator_ok (NOT, QImode, operands)"
-  "@
-   not{b}\t%0
-   not{l}\t%k0"
-  [(set_attr "type" "negnot")
-   (set_attr "mode" "QI,SI")])
-
-(define_insn "*one_cmplqi2_2"
-  [(set (reg FLAGS_REG)
-       (compare (not:QI (match_operand:QI 1 "nonimmediate_operand" "0"))
-                (const_int 0)))
-   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (not:QI (match_dup 1)))]
-  "ix86_match_ccmode (insn, CCNOmode)
-   && ix86_unary_operator_ok (NOT, QImode, operands)"
-  "#"
-  [(set_attr "type" "alu1")
-   (set_attr "mode" "QI")])
-
-(define_split
-  [(set (match_operand 0 "flags_reg_operand" "")
-       (match_operator 2 "compare_operator"
-         [(not:QI (match_operand:QI 3 "nonimmediate_operand" ""))
-          (const_int 0)]))
-   (set (match_operand:QI 1 "nonimmediate_operand" "")
-       (not:QI (match_dup 3)))]
-  "ix86_match_ccmode (insn, CCNOmode)"
-  [(parallel [(set (match_dup 0)
-                  (match_op_dup 2 [(xor:QI (match_dup 3) (const_int -1))
-                                   (const_int 0)]))
-             (set (match_dup 1)
-                  (xor:QI (match_dup 3) (const_int -1)))])]
-  "")
 \f
-;; Arithmetic shift instructions
+;; Shift instructions
 
 ;; DImode shifts are implemented using the i386 "shift double" opcode,
 ;; which is written as "sh[lr]d[lw] imm,reg,reg/mem".  If the shift count
 ;; shift pair, instead using moves and sign extension for counts greater
 ;; than 31.
 
-(define_expand "ashlti3"
-  [(set (match_operand:TI 0 "register_operand" "")
-       (ashift:TI (match_operand:TI 1 "reg_or_pm1_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_64BIT"
-  "ix86_expand_binary_operator (ASHIFT, TImode, operands); DONE;")
-
-;; This pattern must be defined before *ashlti3_1 to prevent
-;; combine pass from converting sse2_ashlti3 to *ashlti3_1.
-
-(define_insn "*avx_ashlti3"
-  [(set (match_operand:TI 0 "register_operand" "=x")
-       (ashift:TI (match_operand:TI 1 "register_operand" "x")
-                  (match_operand:SI 2 "const_0_to_255_mul_8_operand" "n")))]
-  "TARGET_AVX"
-{
-  operands[2] = GEN_INT (INTVAL (operands[2]) / 8);
-  return "vpslldq\t{%2, %1, %0|%0, %1, %2}";
-}
-  [(set_attr "type" "sseishft")
-   (set_attr "prefix" "vex")
-   (set_attr "length_immediate" "1")
-   (set_attr "mode" "TI")])
-
-(define_insn "sse2_ashlti3"
-  [(set (match_operand:TI 0 "register_operand" "=x")
-       (ashift:TI (match_operand:TI 1 "register_operand" "0")
-                  (match_operand:SI 2 "const_0_to_255_mul_8_operand" "n")))]
-  "TARGET_SSE2"
-{
-  operands[2] = GEN_INT (INTVAL (operands[2]) / 8);
-  return "pslldq\t{%2, %0|%0, %2}";
-}
-  [(set_attr "type" "sseishft")
-   (set_attr "prefix_data16" "1")
-   (set_attr "length_immediate" "1")
-   (set_attr "mode" "TI")])
+(define_expand "ashl<mode>3"
+  [(set (match_operand:SDWIM 0 "<shift_operand>" "")
+       (ashift:SDWIM (match_operand:SDWIM 1 "<ashl_input_operand>" "")
+                     (match_operand:QI 2 "nonmemory_operand" "")))]
+  ""
+  "ix86_expand_binary_operator (ASHIFT, <MODE>mode, operands); DONE;")
 
-(define_insn "*ashlti3_1"
-  [(set (match_operand:TI 0 "register_operand" "=&r,r")
-       (ashift:TI (match_operand:TI 1 "reg_or_pm1_operand" "n,0")
-                  (match_operand:QI 2 "nonmemory_operand" "Oc,Oc")))
+(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))]
-  "TARGET_64BIT"
+  ""
   "#"
   [(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;")
+
+;; 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:DI 3 "r")
-   (parallel [(set (match_operand:TI 0 "register_operand" "")
-                  (ashift:TI (match_operand:TI 1 "nonmemory_operand" "")
-                             (match_operand:QI 2 "nonmemory_operand" "")))
+  [(match_scratch:DWIH 3 "r")
+   (parallel [(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))])
    (match_dup 3)]
-  "TARGET_64BIT"
+  "TARGET_CMOVE"
   [(const_int 0)]
-  "ix86_split_ashl (operands, operands[3], TImode); DONE;")
-
-(define_split
-  [(set (match_operand:TI 0 "register_operand" "")
-       (ashift:TI (match_operand:TI 1 "nonmemory_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_ashl (operands, NULL_RTX, TImode); DONE;")
+  "ix86_split_ashl (operands, operands[3], <DWI>mode); DONE;")
 
 (define_insn "x86_64_shld"
   [(set (match_operand:DI 0 "nonimmediate_operand" "+r*m")
    (set_attr "athlon_decode" "vector")
    (set_attr "amdfam10_decode" "vector")])
 
-(define_expand "x86_64_shift_adj_1"
+(define_insn "x86_shld"
+  [(set (match_operand:SI 0 "nonimmediate_operand" "+r*m")
+        (ior:SI (ashift:SI (match_dup 0)
+                 (match_operand:QI 2 "nonmemory_operand" "Ic"))
+               (lshiftrt:SI (match_operand:SI 1 "register_operand" "r")
+                 (minus:QI (const_int 32) (match_dup 2)))))
+   (clobber (reg:CC FLAGS_REG))]
+  ""
+  "shld{l}\t{%s2%1, %0|%0, %1, %2}"
+  [(set_attr "type" "ishift")
+   (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 "x86_shift<mode>_adj_1"
   [(set (reg:CCZ FLAGS_REG)
        (compare:CCZ (and:QI (match_operand:QI 2 "register_operand" "")
-                            (const_int 64))
+                            (match_dup 4))
                     (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "")
-        (if_then_else:DI (ne (reg:CCZ FLAGS_REG) (const_int 0))
-                        (match_operand:DI 1 "register_operand" "")
-                        (match_dup 0)))
+   (set (match_operand:SWI48 0 "register_operand" "")
+        (if_then_else:SWI48 (ne (reg:CCZ FLAGS_REG) (const_int 0))
+                           (match_operand:SWI48 1 "register_operand" "")
+                           (match_dup 0)))
    (set (match_dup 1)
-       (if_then_else:DI (ne (reg:CCZ FLAGS_REG) (const_int 0))
-                        (match_operand:DI 3 "register_operand" "r")
-                        (match_dup 1)))]
-  "TARGET_64BIT"
-  "")
+       (if_then_else:SWI48 (ne (reg:CCZ FLAGS_REG) (const_int 0))
+                           (match_operand:SWI48 3 "register_operand" "r")
+                           (match_dup 1)))]
+  "TARGET_CMOVE"
+  "operands[4] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode));")
 
-(define_expand "x86_64_shift_adj_2"
-  [(use (match_operand:DI 0 "register_operand" ""))
-   (use (match_operand:DI 1 "register_operand" ""))
+(define_expand "x86_shift<mode>_adj_2"
+  [(use (match_operand:SWI48 0 "register_operand" ""))
+   (use (match_operand:SWI48 1 "register_operand" ""))
    (use (match_operand:QI 2 "register_operand" ""))]
-  "TARGET_64BIT"
+  ""
 {
   rtx label = gen_label_rtx ();
   rtx tmp;
 
-  emit_insn (gen_testqi_ccz_1 (operands[2], GEN_INT (64)));
+  emit_insn (gen_testqi_ccz_1 (operands[2],
+                              GEN_INT (GET_MODE_BITSIZE (<MODE>mode))));
 
   tmp = gen_rtx_REG (CCZmode, FLAGS_REG);
   tmp = gen_rtx_EQ (VOIDmode, tmp, const0_rtx);
   DONE;
 })
 
-(define_expand "ashldi3"
-  [(set (match_operand:DI 0 "shiftdi_operand" "")
-       (ashift:DI (match_operand:DI 1 "ashldi_input_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (ASHIFT, DImode, operands); DONE;")
-
-(define_insn "*ashldi3_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,r")
-       (ashift:DI (match_operand:DI 1 "nonimmediate_operand" "0,l")
-                  (match_operand:QI 2 "nonmemory_operand" "cJ,M")))
+(define_insn "*ashl<mode>3_1"
+  [(set (match_operand:SWI48 0 "nonimmediate_operand" "=rm,r")
+       (ashift:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "0,l")
+                     (match_operand:QI 2 "nonmemory_operand" "c<S>,M")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (ASHIFT, DImode, operands)"
+  "ix86_binary_operator_ok (ASHIFT, <MODE>mode, operands)"
 {
   switch (get_attr_type (insn))
     {
+    case TYPE_LEA:
+      return "#";
+
     case TYPE_ALU:
       gcc_assert (operands[2] == const1_rtx);
       gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      return "add{q}\t%0, %0";
-
-    case TYPE_LEA:
-      gcc_assert (CONST_INT_P (operands[2]));
-      gcc_assert ((unsigned HOST_WIDE_INT) INTVAL (operands[2]) <= 3);
-      operands[1] = gen_rtx_MULT (DImode, operands[1],
-                                 GEN_INT (1 << INTVAL (operands[2])));
-      return "lea{q}\t{%a1, %0|%0, %a1}";
+      return "add{<imodesuffix>}\t%0, %0";
 
     default:
-      if (REG_P (operands[2]))
-       return "sal{q}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{q}\t%0";
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "sal{<imodesuffix>}\t%0";
       else
-       return "sal{q}\t{%2, %0|%0, %2}";
+       return "sal{<imodesuffix>}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
                          (const_int 0)))))
        (const_string "0")
        (const_string "*")))
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "<MODE>")])
 
-;; Convert lea to the lea pattern to avoid flags dependency.
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (ashift:DI (match_operand:DI 1 "index_register_operand" "")
-                  (match_operand:QI 2 "immediate_operand" "")))
+(define_insn "*ashlsi3_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r,r")
+       (zero_extend:DI
+         (ashift:SI (match_operand:SI 1 "register_operand" "0,l")
+                    (match_operand:QI 2 "nonmemory_operand" "cI,M"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed
-   && true_regnum (operands[0]) != true_regnum (operands[1])"
-  [(set (match_dup 0)
-       (mult:DI (match_dup 1)
-                (match_dup 2)))]
-  "operands[2] = gen_int_mode (1 << INTVAL (operands[2]), DImode);")
-
-;; 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 "*ashldi3_cmp_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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")
-       (ashift:DI (match_dup 1) (match_dup 2)))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun)
-       || !TARGET_PARTIAL_FLAG_REG_STALL
-       || (operands[2] == const1_rtx
-          && (TARGET_SHIFT1
-              || (TARGET_DOUBLE_WITH_ADD && REG_P (operands[0])))))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, DImode, operands)"
+  "TARGET_64BIT && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
 {
   switch (get_attr_type (insn))
     {
+    case TYPE_LEA:
+      return "#";
+
     case TYPE_ALU:
       gcc_assert (operands[2] == const1_rtx);
-      return "add{q}\t%0, %0";
+      return "add{l}\t%k0, %k0";
 
     default:
-      if (REG_P (operands[2]))
-       return "sal{q}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{q}\t%0";
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "sal{l}\t%k0";
       else
-       return "sal{q}\t{%2, %0|%0, %2}";
+       return "sal{l}\t{%2, %k0|%k0, %2}";
     }
 }
   [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
+     (cond [(eq_attr "alternative" "1")
+             (const_string "lea")
+            (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
+                    (const_int 0))
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
           ]
                          (const_int 0)))))
        (const_string "0")
        (const_string "*")))
-   (set_attr "mode" "DI")])
+   (set_attr "mode" "SI")])
 
-(define_insn "*ashldi3_cconly_rex64"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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
-       || (operands[2] == const1_rtx
-          && (TARGET_SHIFT1
-              || TARGET_DOUBLE_WITH_ADD)))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, DImode, operands)"
+(define_insn "*ashlhi3_1"
+  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
+       (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "0")
+                  (match_operand:QI 2 "nonmemory_operand" "cI")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_ALU:
       gcc_assert (operands[2] == const1_rtx);
-      return "add{q}\t%0, %0";
+      return "add{w}\t%0, %0";
 
     default:
-      if (REG_P (operands[2]))
-       return "sal{q}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{q}\t%0";
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "sal{w}\t%0";
       else
-       return "sal{q}\t{%2, %0|%0, %2}";
+       return "sal{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
                          (const_int 0)))))
        (const_string "0")
        (const_string "*")))
-   (set_attr "mode" "DI")])
-
-(define_insn "*ashldi3_1"
-  [(set (match_operand:DI 0 "register_operand" "=&r,r")
-       (ashift:DI (match_operand:DI 1 "reg_or_pm1_operand" "n,0")
-                  (match_operand:QI 2 "nonmemory_operand" "Jc,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" "")
-                  (ashift:DI (match_operand:DI 1 "nonmemory_operand" "")
-                             (match_operand:QI 2 "nonmemory_operand" "")))
-             (clobber (reg:CC FLAGS_REG))])
-   (match_dup 3)]
-  "!TARGET_64BIT && TARGET_CMOVE"
-  [(const_int 0)]
-  "ix86_split_ashl (operands, operands[3], DImode); DONE;")
-
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (ashift:DI (match_operand:DI 1 "nonmemory_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_ashl (operands, NULL_RTX, DImode); DONE;")
-
-(define_insn "x86_shld"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "+r*m")
-        (ior:SI (ashift:SI (match_dup 0)
-                 (match_operand:QI 2 "nonmemory_operand" "Ic"))
-               (lshiftrt:SI (match_operand:SI 1 "register_operand" "r")
-                 (minus:QI (const_int 32) (match_dup 2)))))
-   (clobber (reg:CC FLAGS_REG))]
-  ""
-  "shld{l}\t{%s2%1, %0|%0, %1, %2}"
-  [(set_attr "type" "ishift")
-   (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 "x86_shift_adj_1"
-  [(set (reg:CCZ FLAGS_REG)
-       (compare:CCZ (and:QI (match_operand:QI 2 "register_operand" "")
-                            (const_int 32))
-                    (const_int 0)))
-   (set (match_operand:SI 0 "register_operand" "")
-        (if_then_else:SI (ne (reg:CCZ FLAGS_REG) (const_int 0))
-                        (match_operand:SI 1 "register_operand" "")
-                        (match_dup 0)))
-   (set (match_dup 1)
-       (if_then_else:SI (ne (reg:CCZ FLAGS_REG) (const_int 0))
-                        (match_operand:SI 3 "register_operand" "r")
-                        (match_dup 1)))]
-  "TARGET_CMOVE"
-  "")
-
-(define_expand "x86_shift_adj_2"
-  [(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]);
-  ix86_expand_clear (operands[1]);
-
-  emit_label (label);
-  LABEL_NUSES (label) = 1;
-
-  DONE;
-})
-
-(define_expand "ashlsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (ashift:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (ASHIFT, SImode, operands); DONE;")
+   (set_attr "mode" "HI")])
 
-(define_insn "*ashlsi3_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
-       (ashift:SI (match_operand:SI 1 "nonimmediate_operand" "0,l")
+(define_insn "*ashlhi3_1_lea"
+  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
+       (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "0,l")
                   (match_operand:QI 2 "nonmemory_operand" "cI,M")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (ASHIFT, SImode, operands)"
+  "!TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      gcc_assert (rtx_equal_p (operands[0], operands[1]));
-      return "add{l}\t%0, %0";
-
     case TYPE_LEA:
       return "#";
 
+    case TYPE_ALU:
+      gcc_assert (operands[2] == const1_rtx);
+      return "add{w}\t%0, %0";
+
     default:
-      if (REG_P (operands[2]))
-       return "sal{l}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{l}\t%0";
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "sal{w}\t%0";
       else
-       return "sal{l}\t{%2, %0|%0, %2}";
+       return "sal{w}\t{%2, %0|%0, %2}";
     }
 }
   [(set (attr "type")
                          (const_int 0)))))
        (const_string "0")
        (const_string "*")))
-   (set_attr "mode" "SI")])
-
-;; Convert lea to the lea pattern to avoid flags dependency.
-(define_split
-  [(set (match_operand 0 "register_operand" "")
-       (ashift (match_operand 1 "index_register_operand" "")
-                (match_operand:QI 2 "const_int_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "reload_completed
-   && true_regnum (operands[0]) != true_regnum (operands[1])
-   && GET_MODE_SIZE (GET_MODE (operands[0])) <= 4"
-  [(const_int 0)]
-{
-  rtx pat;
-  enum machine_mode mode = GET_MODE (operands[0]);
-
-  if (GET_MODE_SIZE (mode) < 4)
-    operands[0] = gen_lowpart (SImode, operands[0]);
-  if (mode != Pmode)
-    operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[2] = gen_int_mode (1 << INTVAL (operands[2]), Pmode);
-
-  pat = gen_rtx_MULT (Pmode, operands[1], operands[2]);
-  if (Pmode != SImode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
-  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
-  DONE;
-})
-
-;; Rare case of shifting RSP is handled by generating move and shift
-(define_split
-  [(set (match_operand 0 "register_operand" "")
-       (ashift (match_operand 1 "register_operand" "")
-                (match_operand:QI 2 "const_int_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "reload_completed
-   && true_regnum (operands[0]) != true_regnum (operands[1])"
-  [(const_int 0)]
-{
-  rtx pat, clob;
-  emit_move_insn (operands[0], operands[1]);
-  pat = gen_rtx_SET (VOIDmode, operands[0],
-                    gen_rtx_ASHIFT (GET_MODE (operands[0]),
-                                    operands[0], operands[2]));
-  clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
-  emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, pat, clob)));
-  DONE;
-})
+   (set_attr "mode" "HI,SI")])
 
-(define_insn "*ashlsi3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (zero_extend:DI (ashift:SI (match_operand:SI 1 "register_operand" "0,l")
-                       (match_operand:QI 2 "nonmemory_operand" "cI,M"))))
+(define_insn "*ashlqi3_1"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r")
+       (ashift:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
+                  (match_operand:QI 2 "nonmemory_operand" "cI,cI")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
+  "TARGET_PARTIAL_REG_STALL
+   && ix86_binary_operator_ok (ASHIFT, QImode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_ALU:
       gcc_assert (operands[2] == const1_rtx);
-      return "add{l}\t%k0, %k0";
-
-    case TYPE_LEA:
-      return "#";
+      if (REG_P (operands[1]) && !ANY_QI_REG_P (operands[1]))
+        return "add{l}\t%k0, %k0";
+      else
+        return "add{b}\t%0, %0";
 
     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)))
-       return "sal{l}\t%k0";
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       {
+         if (get_attr_mode (insn) == MODE_SI)
+           return "sal{l}\t%k0";
+         else
+           return "sal{b}\t%0";
+       }
       else
-       return "sal{l}\t{%2, %k0|%k0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(eq_attr "alternative" "1")
-             (const_string "lea")
-            (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                    (const_int 0))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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")])
-
-;; Convert lea to the lea pattern to avoid flags dependency.
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (zero_extend:DI (ashift (match_operand 1 "register_operand" "")
-                               (match_operand:QI 2 "const_int_operand" ""))))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed
-   && true_regnum (operands[0]) != true_regnum (operands[1])"
-  [(set (match_dup 0) (zero_extend:DI
-                       (subreg:SI (mult:SI (match_dup 1)
-                                           (match_dup 2)) 0)))]
-{
-  operands[1] = gen_lowpart (Pmode, operands[1]);
-  operands[2] = gen_int_mode (1 << INTVAL (operands[2]), Pmode);
-})
-
-;; 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 "*ashlsi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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")
-       (ashift:SI (match_dup 1) (match_dup 2)))]
-   "(optimize_function_for_size_p (cfun)
-     || !TARGET_PARTIAL_FLAG_REG_STALL
-     || (operands[2] == const1_rtx
-        && (TARGET_SHIFT1
-            || (TARGET_DOUBLE_WITH_ADD && REG_P (operands[0])))))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      return "add{l}\t%0, %0";
-
-    default:
-      if (REG_P (operands[2]))
-       return "sal{l}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{l}\t%0";
-      else
-       return "sal{l}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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 "*ashlsi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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
-    || (operands[2] == const1_rtx
-       && (TARGET_SHIFT1
-           || TARGET_DOUBLE_WITH_ADD)))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      return "add{l}\t%0, %0";
-
-    default:
-      if (REG_P (operands[2]))
-       return "sal{l}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{l}\t%0";
-      else
-       return "sal{l}\t{%2, %0|%0, %2}";
+       {
+         if (get_attr_mode (insn) == MODE_SI)
+           return "sal{l}\t{%2, %k0|%k0, %2}";
+         else
+           return "sal{b}\t{%2, %0|%0, %2}";
+       }
     }
 }
   [(set (attr "type")
                          (const_int 0)))))
        (const_string "0")
        (const_string "*")))
-   (set_attr "mode" "SI")])
-
-(define_insn "*ashlsi3_cmp_zext"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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 (ashift:SI (match_dup 1) (match_dup 2))))]
-  "TARGET_64BIT
-   && (optimize_function_for_size_p (cfun)
-       || !TARGET_PARTIAL_FLAG_REG_STALL
-       || (operands[2] == const1_rtx
-          && (TARGET_SHIFT1
-              || TARGET_DOUBLE_WITH_ADD)))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      return "add{l}\t%k0, %k0";
-
-    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)))
-       return "sal{l}\t%k0";
-      else
-       return "sal{l}\t{%2, %k0|%k0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                    (const_int 0))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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_expand "ashlhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (ASHIFT, HImode, operands); DONE;")
+   (set_attr "mode" "QI,SI")])
 
-(define_insn "*ashlhi3_1_lea"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
-       (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "0,l")
-                  (match_operand:QI 2 "nonmemory_operand" "cI,M")))
+;; %%% Potential partial reg stall on alternative 2.  What to do?
+(define_insn "*ashlqi3_1_lea"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r,r")
+       (ashift:QI (match_operand:QI 1 "nonimmediate_operand" "0,0,l")
+                  (match_operand:QI 2 "nonmemory_operand" "cI,cI,M")))
    (clobber (reg:CC FLAGS_REG))]
   "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
+   && ix86_binary_operator_ok (ASHIFT, QImode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_LEA:
       return "#";
-    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)))
-       return "sal{w}\t%0";
-      else
-       return "sal{w}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(eq_attr "alternative" "1")
-             (const_string "lea")
-            (and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "HI,SI")])
-
-(define_insn "*ashlhi3_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "0")
-                  (match_operand:QI 2 "nonmemory_operand" "cI")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    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)))
-       return "sal{w}\t%0";
-      else
-       return "sal{w}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "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 "*ashlhi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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")
-       (ashift:HI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun)
-    || !TARGET_PARTIAL_FLAG_REG_STALL
-    || (operands[2] == const1_rtx
-       && (TARGET_SHIFT1
-           || (TARGET_DOUBLE_WITH_ADD && REG_P (operands[0])))))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
     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)))
-       return "sal{w}\t%0";
-      else
-       return "sal{w}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "HI")])
-
-(define_insn "*ashlhi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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
-    || (operands[2] == const1_rtx
-       && (TARGET_SHIFT1
-           || TARGET_DOUBLE_WITH_ADD)))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    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)))
-       return "sal{w}\t%0";
-      else
-       return "sal{w}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "HI")])
-
-(define_expand "ashlqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (ashift:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (ASHIFT, QImode, operands); DONE;")
-
-;; %%% Potential partial reg stall on alternative 2.  What to do?
-
-(define_insn "*ashlqi3_1_lea"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r,r")
-       (ashift:QI (match_operand:QI 1 "nonimmediate_operand" "0,0,l")
-                  (match_operand:QI 2 "nonmemory_operand" "cI,cI,M")))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (ASHIFT, QImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    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{l}\t%k0, %k0";
-      else
-        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 (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t%0";
-         else
-           return "sal{b}\t%0";
-       }
-      else
-       {
-         if (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t{%2, %k0|%k0, %2}";
-         else
-           return "sal{b}\t{%2, %0|%0, %2}";
-       }
-    }
-}
-  [(set (attr "type")
-     (cond [(eq_attr "alternative" "2")
-             (const_string "lea")
-            (and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "QI,SI,SI")])
-
-(define_insn "*ashlqi3_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r")
-       (ashift:QI (match_operand:QI 1 "nonimmediate_operand" "0,0")
-                  (match_operand:QI 2 "nonmemory_operand" "cI,cI")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (ASHIFT, QImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      if (REG_P (operands[1]) && !ANY_QI_REG_P (operands[1]))
-        return "add{l}\t%k0, %k0";
-      else
-        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 (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t%0";
-         else
-           return "sal{b}\t%0";
-       }
-      else
-       {
-         if (get_attr_mode (insn) == MODE_SI)
-           return "sal{l}\t{%2, %k0|%k0, %2}";
-         else
-           return "sal{b}\t{%2, %0|%0, %2}";
-       }
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "QI,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 "*ashlqi3_cmp"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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")
-       (ashift:QI (match_dup 1) (match_dup 2)))]
-  "(optimize_function_for_size_p (cfun)
-    || !TARGET_PARTIAL_FLAG_REG_STALL
-    || (operands[2] == const1_rtx
-       && (TARGET_SHIFT1
-           || (TARGET_DOUBLE_WITH_ADD && REG_P (operands[0])))))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, QImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      return "add{b}\t%0, %0";
-
-    default:
-      if (REG_P (operands[2]))
-       return "sal{b}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{b}\t%0";
-      else
-       return "sal{b}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "QI")])
-
-(define_insn "*ashlqi3_cconly"
-  [(set (reg FLAGS_REG)
-       (compare
-         (ashift: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
-    || (operands[2] == const1_rtx
-       && (TARGET_SHIFT1
-           || TARGET_DOUBLE_WITH_ADD)))
-   && ix86_match_ccmode (insn, CCGOCmode)
-   && ix86_binary_operator_ok (ASHIFT, QImode, operands)"
-{
-  switch (get_attr_type (insn))
-    {
-    case TYPE_ALU:
-      gcc_assert (operands[2] == const1_rtx);
-      return "add{b}\t%0, %0";
-
-    default:
-      if (REG_P (operands[2]))
-       return "sal{b}\t{%b2, %0|%0, %b2}";
-      else if (operands[2] == const1_rtx
-              && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-       return "sal{b}\t%0";
-      else
-       return "sal{b}\t{%2, %0|%0, %2}";
-    }
-}
-  [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
-                     (match_operand 0 "register_operand" ""))
-                (match_operand 2 "const1_operand" ""))
-             (const_string "alu")
-          ]
-          (const_string "ishift")))
-   (set (attr "length_immediate")
-     (if_then_else
-       (ior (eq_attr "type" "alu")
-           (and (eq_attr "type" "ishift")
-                (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" "QI")])
-
-;; 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;")
-
-(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" ""))]
-  ""
-{
-  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")])
-
-(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")))
-   (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" "")
-                  (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" "")
-                    (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")])
+      gcc_assert (operands[2] == const1_rtx);
+      if (REG_P (operands[1]) && !ANY_QI_REG_P (operands[1]))
+        return "add{l}\t%k0, %k0";
+      else
+        return "add{b}\t%0, %0";
 
-(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;")
+    default:
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       {
+         if (get_attr_mode (insn) == MODE_SI)
+           return "sal{l}\t%k0";
+         else
+           return "sal{b}\t%0";
+       }
+      else
+       {
+         if (get_attr_mode (insn) == MODE_SI)
+           return "sal{l}\t{%2, %k0|%k0, %2}";
+         else
+           return "sal{b}\t{%2, %0|%0, %2}";
+       }
+    }
+}
+  [(set (attr "type")
+     (cond [(eq_attr "alternative" "2")
+             (const_string "lea")
+            (and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
+                         (const_int 0))
+                     (match_operand 0 "register_operand" ""))
+                (match_operand 2 "const1_operand" ""))
+             (const_string "alu")
+          ]
+          (const_string "ishift")))
+   (set (attr "length_immediate")
+     (if_then_else
+       (ior (eq_attr "type" "alu")
+           (and (eq_attr "type" "ishift")
+                (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" "QI,SI,SI")])
 
-(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" "")))
+;; Convert lea to the lea pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (ashift:DI (match_operand:DI 1 "index_register_operand" "")
+                  (match_operand:QI 2 "const_int_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")])
+  "TARGET_64BIT && reload_completed
+   && true_regnum (operands[0]) != true_regnum (operands[1])"
+  [(set (match_dup 0)
+       (mult:DI (match_dup 1)
+                (match_dup 2)))]
+  "operands[2] = gen_int_mode (1 << INTVAL (operands[2]), DImode);")
 
-(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" "")))
+;; Convert lea to the lea pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand 0 "register_operand" "")
+       (ashift (match_operand 1 "index_register_operand" "")
+                (match_operand:QI 2 "const_int_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")])
+  "reload_completed
+   && true_regnum (operands[0]) != true_regnum (operands[1])
+   && GET_MODE_SIZE (GET_MODE (operands[0])) <= 4"
+  [(const_int 0)]
+{
+  rtx pat;
+  enum machine_mode mode = GET_MODE (operands[0]);
+
+  if (GET_MODE_SIZE (mode) < 4)
+    operands[0] = gen_lowpart (SImode, operands[0]);
+  if (mode != Pmode)
+    operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[2] = gen_int_mode (1 << INTVAL (operands[2]), Pmode);
+
+  pat = gen_rtx_MULT (Pmode, operands[1], operands[2]);
+  if (Pmode != SImode)
+    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
+  DONE;
+})
 
-(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")))
+;; Rare case of shifting RSP is handled by generating move and shift
+(define_split
+  [(set (match_operand 0 "register_operand" "")
+       (ashift (match_operand 1 "register_operand" "")
+                (match_operand:QI 2 "const_int_operand" "")))
    (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")])
+  "reload_completed
+   && true_regnum (operands[0]) != true_regnum (operands[1])"
+  [(const_int 0)]
+{
+  rtx pat, clob;
+  emit_move_insn (operands[0], operands[1]);
+  pat = gen_rtx_SET (VOIDmode, operands[0],
+                    gen_rtx_ASHIFT (GET_MODE (operands[0]),
+                                    operands[0], operands[2]));
+  clob = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (CCmode, FLAGS_REG));
+  emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, pat, clob)));
+  DONE;
+})
 
-(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")))
+;; Convert lea to the lea pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (zero_extend:DI
+         (ashift:SI (match_operand:SI 1 "register_operand" "")
+                    (match_operand:QI 2 "const_int_operand" ""))))
    (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")])
+  "TARGET_64BIT && reload_completed
+   && true_regnum (operands[0]) != true_regnum (operands[1])"
+  [(set (match_dup 0)
+       (zero_extend:DI (subreg:SI (mult:SI (match_dup 1) (match_dup 2)) 0)))]
+{
+  operands[1] = gen_lowpart (Pmode, operands[1]);
+  operands[2] = gen_int_mode (1 << INTVAL (operands[2]), Pmode);
+})
 
 ;; 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"
+(define_insn "*ashl<mode>3_cmp"
   [(set (reg FLAGS_REG)
        (compare
-         (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
+         (ashift:SWI (match_operand:SWI 1 "nonimmediate_operand" "0")
+                     (match_operand:QI 2 "<shift_immediate_operand>" "<S>"))
          (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (ashift:SWI (match_dup 1) (match_dup 2)))]
+  "(optimize_function_for_size_p (cfun)
+    || !TARGET_PARTIAL_FLAG_REG_STALL
+    || (operands[2] == const1_rtx
+       && (TARGET_SHIFT1
+           || (TARGET_DOUBLE_WITH_ADD && REG_P (operands[0])))))
    && 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")])
+   && ix86_binary_operator_ok (ASHIFT, <MODE>mode, operands)"
+{
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ALU:
+      gcc_assert (operands[2] == const1_rtx);
+      return "add{<imodesuffix>}\t%0, %0";
 
-;; 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")])
+    default:
+      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}";
+    }
+}
+  [(set (attr "type")
+     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
+                         (const_int 0))
+                     (match_operand 0 "register_operand" ""))
+                (match_operand 2 "const1_operand" ""))
+             (const_string "alu")
+          ]
+          (const_string "ishift")))
+   (set (attr "length_immediate")
+     (if_then_else
+       (ior (eq_attr "type" "alu")
+           (and (eq_attr "type" "ishift")
+                (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 "*ashrqi3_cconly"
+(define_insn "*ashlsi3_cmp_zext"
   [(set (reg FLAGS_REG)
        (compare
-         (ashiftrt:QI (match_operand:QI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const_1_to_31_operand" "I"))
+         (ashift:SI (match_operand:SI 1 "register_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)
+   (set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI (ashift:SI (match_dup 1) (match_dup 2))))]
+  "TARGET_64BIT
+   && (optimize_function_for_size_p (cfun)
+       || !TARGET_PARTIAL_FLAG_REG_STALL
+       || (operands[2] == const1_rtx
+          && (TARGET_SHIFT1
+              || TARGET_DOUBLE_WITH_ADD)))
    && 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;")
-
-;; This pattern must be defined before *lshrti3_1 to prevent
-;; combine pass from converting sse2_lshrti3 to *lshrti3_1.
-
-(define_insn "*avx_lshrti3"
-  [(set (match_operand:TI 0 "register_operand" "=x")
-       (lshiftrt:TI (match_operand:TI 1 "register_operand" "x")
-                    (match_operand:SI 2 "const_0_to_255_mul_8_operand" "n")))]
-  "TARGET_AVX"
+   && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
 {
-  operands[2] = GEN_INT (INTVAL (operands[2]) / 8);
-  return "vpsrldq\t{%2, %1, %0|%0, %1, %2}";
-}
-  [(set_attr "type" "sseishft")
-   (set_attr "prefix" "vex")
-   (set_attr "length_immediate" "1")
-   (set_attr "mode" "TI")])
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ALU:
+      gcc_assert (operands[2] == const1_rtx);
+      return "add{l}\t%k0, %k0";
 
-(define_insn "sse2_lshrti3"
-  [(set (match_operand:TI 0 "register_operand" "=x")
-       (lshiftrt:TI (match_operand:TI 1 "register_operand" "0")
-                    (match_operand:SI 2 "const_0_to_255_mul_8_operand" "n")))]
-  "TARGET_SSE2"
-{
-  operands[2] = GEN_INT (INTVAL (operands[2]) / 8);
-  return "psrldq\t{%2, %0|%0, %2}";
+    default:
+      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}";
+    }
 }
-  [(set_attr "type" "sseishft")
-   (set_attr "prefix_data16" "1")
-   (set_attr "length_immediate" "1")
-   (set_attr "mode" "TI")])
-
-(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")])
+  [(set (attr "type")
+     (cond [(and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
+                    (const_int 0))
+                (match_operand 2 "const1_operand" ""))
+             (const_string "alu")
+          ]
+          (const_string "ishift")))
+   (set (attr "length_immediate")
+     (if_then_else
+       (ior (eq_attr "type" "alu")
+           (and (eq_attr "type" "ishift")
+                (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")])
 
-;; 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"
+(define_insn "*ashl<mode>3_cconly"
   [(set (reg FLAGS_REG)
        (compare
-         (lshiftrt:DI (match_operand:DI 1 "nonimmediate_operand" "0")
-                      (match_operand:QI 2 "const1_operand" ""))
+         (ashift:SWI (match_operand:SWI 1 "nonimmediate_operand" "0")
+                     (match_operand:QI 2 "<shift_immediate_operand>" "<S>"))
          (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))
+   (clobber (match_scratch:SWI 0 "=<r>"))]
+  "(optimize_function_for_size_p (cfun)
+    || !TARGET_PARTIAL_FLAG_REG_STALL
+    || (operands[2] == const1_rtx
+       && (TARGET_SHIFT1
+           || TARGET_DOUBLE_WITH_ADD)))
    && 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")])
+   && ix86_binary_operator_ok (ASHIFT, <MODE>mode, operands)"
+{
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ALU:
+      gcc_assert (operands[2] == const1_rtx);
+      return "add{<imodesuffix>}\t%0, %0";
 
-(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")])
+    default:
+      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}";
+    }
+}
+  [(set (attr "type")
+     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
+                         (const_int 0))
+                     (match_operand 0 "register_operand" ""))
+                (match_operand 2 "const1_operand" ""))
+             (const_string "alu")
+          ]
+          (const_string "ishift")))
+   (set (attr "length_immediate")
+     (if_then_else
+       (ior (eq_attr "type" "alu")
+           (and (eq_attr "type" "ishift")
+                (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>")])
 
-;; 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")])
+;; See comment above `ashl<mode>3' about how this works.
 
-(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_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" "")))]
+  ""
+  "ix86_expand_binary_operator (<CODE>, <MODE>mode, operands); DONE;")
 
-(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")))
+(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" "")
-                  (lshiftrt:DI (match_operand:DI 1 "register_operand" "")
-                               (match_operand:QI 2 "nonmemory_operand" "")))
+  [(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))])
    (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)"
+  "TARGET_CMOVE"
   [(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;")
+  "ix86_split_<shiftrt_insn> (operands, operands[3], <DWI>mode); 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" "")))
+(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_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%0"
+  "TARGET_64BIT"
+  "shrd{q}\t{%s2%1, %0|%0, %1, %2}"
   [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (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_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" "")))
+(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))]
-  "TARGET_64BIT
-   && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
-  "shr{l}\t%k0"
+  ""
+  "shrd{l}\t{%s2%1, %0|%0, %1, %2}"
   [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
+   (set_attr "prefix_0f" "1")
+   (set_attr "pent_pair" "np")
    (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")))
+(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))]
-  "ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
+  "TARGET_64BIT && INTVAL (operands[2]) == 63
+   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
+   && ix86_binary_operator_ok (ASHIFTRT, DImode, operands)"
   "@
-   shr{l}\t{%2, %0|%0, %2}
-   shr{l}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
+   {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 "*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"))))
+(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_64BIT && ix86_binary_operator_ok (LSHIFTRT, HImode, operands)"
+  "INTVAL (operands[2]) == 31
+   && (TARGET_USE_CLTD || optimize_function_for_size_p (cfun))
+   && ix86_binary_operator_ok (ASHIFTRT, SImode, 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")])
-
-(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}"
-  [(set_attr "type" "ishift")
-   (set_attr "mode" "SI")])
-
-(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}"
-  [(set_attr "type" "ishift")
+   {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_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 "*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" "")))
-   (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")])
-
-(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")])
-
-;; 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_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")])
+   {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_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")])
+(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_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_insn (gen_testqi_ccz_1 (operands[2],
+                              GEN_INT (GET_MODE_BITSIZE (<MODE>mode))));
 
-(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;")
+  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 "*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" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
-   && ix86_binary_operator_ok (LSHIFTRT, QImode, operands)"
-  "shr{b}\t%0"
-  [(set_attr "type" "ishift")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
+  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_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" "")))
+  DONE;
+})
+
+(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_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")])
+  "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
+       (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"
+(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 "const1_operand" ""))
+         (any_shiftrt:SWI
+           (match_operand:SWI 1 "nonimmediate_operand" "0")
+           (match_operand:QI 2 "<shift_immediate_operand>" "<S>"))
          (const_int 0)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "(TARGET_SHIFT1 || optimize_function_for_size_p (cfun))
+   (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%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")])
+   (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>")])
 
-;; 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>si3_cmp_zext"
   [(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:SI (match_operand:SI 1 "register_operand" "0")
+                         (match_operand:QI 2 "const_1_to_31_operand" "I"))
          (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:DI 0 "register_operand" "=r")
+       (zero_extend:DI (any_shiftrt:SI (match_dup 1) (match_dup 2))))]
+  "TARGET_64BIT
+   && (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>, 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 "*lshrqi2_cconly"
+(define_insn "*<shiftrt_insn><mode>3_cconly"
   [(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)))
-   (clobber (match_scratch:QI 0 "=q"))]
-  "(optimize_function_for_size_p (cfun) || !TARGET_PARTIAL_FLAG_REG_STALL)
+   (clobber (match_scratch:SWI 0 "=<r>"))]
+  "(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 "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>")])
 \f
 ;; Rotate instructions
 
-(define_expand "rotldi3"
+(define_expand "<rotate_insn>ti3"
+  [(set (match_operand:TI 0 "register_operand" "")
+       (any_rotate:TI (match_operand:TI 1 "register_operand" "")
+                      (match_operand:QI 2 "nonmemory_operand" "")))]
+  "TARGET_64BIT"
+{
+  if (const_1_to_63_operand (operands[2], VOIDmode))
+    emit_insn (gen_ix86_<rotate_insn>ti3_doubleword
+               (operands[0], operands[1], operands[2]));
+  else
+    FAIL;
+
+  DONE;
+})
+
+(define_expand "<rotate_insn>di3"
   [(set (match_operand:DI 0 "shiftdi_operand" "")
-       (rotate:DI (match_operand:DI 1 "shiftdi_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
+       (any_rotate:DI (match_operand:DI 1 "shiftdi_operand" "")
+                      (match_operand:QI 2 "nonmemory_operand" "")))]
  ""
 {
   if (TARGET_64BIT)
-    {
-      ix86_expand_binary_operator (ROTATE, DImode, operands);
-      DONE;
-    }
-  if (!const_1_to_31_operand (operands[2], VOIDmode))
+    ix86_expand_binary_operator (<CODE>, DImode, operands);
+  else if (const_1_to_31_operand (operands[2], VOIDmode))
+    emit_insn (gen_ix86_<rotate_insn>di3_doubleword
+               (operands[0], operands[1], operands[2]));
+  else
     FAIL;
-  emit_insn (gen_ix86_rotldi3 (operands[0], operands[1], operands[2]));
+
   DONE;
 })
 
-;; Implement rotation using two double-precision shift instructions
-;; and a scratch register.
-(define_insn_and_split "ix86_rotldi3"
- [(set (match_operand:DI 0 "register_operand" "=r")
-       (rotate:DI (match_operand:DI 1 "register_operand" "0")
-                  (match_operand:QI 2 "const_1_to_31_operand" "I")))
+(define_expand "<rotate_insn><mode>3"
+  [(set (match_operand:SWIM124 0 "nonimmediate_operand" "")
+       (any_rotate:SWIM124 (match_operand:SWIM124 1 "nonimmediate_operand" "")
+                           (match_operand:QI 2 "nonmemory_operand" "")))]
+  ""
+  "ix86_expand_binary_operator (<CODE>, <MODE>mode, operands); DONE;")
+
+;; Implement rotation using two double-precision
+;; shift instructions and a scratch register.
+
+(define_insn_and_split "ix86_rotl<dwi>3_doubleword"
+ [(set (match_operand:<DWI> 0 "register_operand" "=r")
+       (rotate:<DWI> (match_operand:<DWI> 1 "register_operand" "0")
+                    (match_operand:QI 2 "<shift_immediate_operand>" "<S>")))
   (clobber (reg:CC FLAGS_REG))
-  (clobber (match_scratch:SI 3 "=&r"))]
- "!TARGET_64BIT"
+  (clobber (match_scratch:DWIH 3 "=&r"))]
  ""
- "&& reload_completed"
+ "#"
+ "reload_completed"
  [(set (match_dup 3) (match_dup 4))
   (parallel
    [(set (match_dup 4)
-         (ior:SI (ashift:SI (match_dup 4) (match_dup 2))
-                 (lshiftrt:SI (match_dup 5)
-                              (minus:QI (const_int 32) (match_dup 2)))))
+        (ior:DWIH (ashift:DWIH (match_dup 4) (match_dup 2))
+                  (lshiftrt:DWIH (match_dup 5)
+                                 (minus:QI (match_dup 6) (match_dup 2)))))
     (clobber (reg:CC FLAGS_REG))])
   (parallel
    [(set (match_dup 5)
-         (ior:SI (ashift:SI (match_dup 5) (match_dup 2))
-                 (lshiftrt:SI (match_dup 3)
-                              (minus:QI (const_int 32) (match_dup 2)))))
+        (ior:DWIH (ashift:DWIH (match_dup 5) (match_dup 2))
+                  (lshiftrt:DWIH (match_dup 3)
+                                 (minus:QI (match_dup 6) (match_dup 2)))))
     (clobber (reg:CC FLAGS_REG))])]
- "split_di (&operands[0], 1, &operands[4], &operands[5]);")
-
-(define_insn "*rotlsi3_1_one_bit_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (rotate: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 (ROTATE, DImode, operands)"
-  "rol{q}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-(define_insn "*rotldi3_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,rm")
-       (rotate:DI (match_operand:DI 1 "nonimmediate_operand" "0,0")
-                  (match_operand:QI 2 "nonmemory_operand" "e,c")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (ROTATE, DImode, operands)"
-  "@
-   rol{q}\t{%2, %0|%0, %2}
-   rol{q}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "DI")])
-
-(define_expand "rotlsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (rotate:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (ROTATE, SImode, operands); DONE;")
-
-(define_insn "*rotlsi3_1_one_bit"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (rotate: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 (ROTATE, SImode, operands)"
-  "rol{l}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*rotlsi3_1_one_bit_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI
-         (rotate: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 (ROTATE, SImode, operands)"
-  "rol{l}\t%k0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*rotlsi3_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,rm")
-       (rotate: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 (ROTATE, SImode, operands)"
-  "@
-   rol{l}\t{%2, %0|%0, %2}
-   rol{l}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "SI")])
-
-(define_insn "*rotlsi3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (zero_extend:DI
-         (rotate: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 (ROTATE, SImode, operands)"
-  "@
-   rol{l}\t{%2, %k0|%k0, %2}
-   rol{l}\t{%b2, %k0|%k0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "SI")])
-
-(define_expand "rotlhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (rotate:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (ROTATE, HImode, operands); DONE;")
-
-(define_insn "*rotlhi3_1_one_bit"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (rotate: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 (ROTATE, HImode, operands)"
-  "rol{w}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
-
-(define_insn "*rotlhi3_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,rm")
-       (rotate: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 (ROTATE, HImode, operands)"
-  "@
-   rol{w}\t{%2, %0|%0, %2}
-   rol{w}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "HI")])
-
-(define_split
- [(set (match_operand:HI 0 "register_operand" "")
-       (rotate:HI (match_dup 0) (const_int 8)))
-  (clobber (reg:CC FLAGS_REG))]
- "reload_completed"
- [(parallel [(set (strict_low_part (match_dup 0))
-                 (bswap:HI (match_dup 0)))
-            (clobber (reg:CC FLAGS_REG))])]
- "")
-
-(define_expand "rotlqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (rotate:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (ROTATE, QImode, operands); DONE;")
-
-(define_insn "*rotlqi3_1_one_bit_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
-       (rotate: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))"
-  "rol{b}\t%0"
-  [(set_attr "type" "rotate1")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*rotlqi3_1_one_bit"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (rotate: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 (ROTATE, QImode, operands)"
-  "rol{b}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*rotlqi3_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,qm"))
-       (rotate: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]))"
-  "@
-   rol{b}\t{%1, %0|%0, %1}
-   rol{b}\t{%b1, %0|%0, %b1}"
-  [(set_attr "type" "rotate1")
-   (set_attr "mode" "QI")])
-
-(define_insn "*rotlqi3_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,qm")
-       (rotate: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 (ROTATE, QImode, operands)"
-  "@
-   rol{b}\t{%2, %0|%0, %2}
-   rol{b}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "QI")])
-
-(define_expand "rotrdi3"
-  [(set (match_operand:DI 0 "shiftdi_operand" "")
-       (rotate:DI (match_operand:DI 1 "shiftdi_operand" "")
-                  (match_operand:QI 2 "nonmemory_operand" "")))]
- ""
 {
-  if (TARGET_64BIT)
-    {
-      ix86_expand_binary_operator (ROTATERT, DImode, operands);
-      DONE;
-    }
-  if (!const_1_to_31_operand (operands[2], VOIDmode))
-    FAIL;
-  emit_insn (gen_ix86_rotrdi3 (operands[0], operands[1], operands[2]));
-  DONE;
+  operands[6] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode));
+
+  split_<dwi> (&operands[0], 1, &operands[4], &operands[5]);
 })
 
-;; Implement rotation using two double-precision shift instructions
-;; and a scratch register.
-(define_insn_and_split "ix86_rotrdi3"
- [(set (match_operand:DI 0 "register_operand" "=r")
-       (rotatert:DI (match_operand:DI 1 "register_operand" "0")
-                    (match_operand:QI 2 "const_1_to_31_operand" "I")))
+(define_insn_and_split "ix86_rotr<dwi>3_doubleword"
+ [(set (match_operand:<DWI> 0 "register_operand" "=r")
+       (rotatert:<DWI> (match_operand:<DWI> 1 "register_operand" "0")
+                      (match_operand:QI 2 "<shift_immediate_operand>" "<S>")))
   (clobber (reg:CC FLAGS_REG))
-  (clobber (match_scratch:SI 3 "=&r"))]
- "!TARGET_64BIT"
+  (clobber (match_scratch:DWIH 3 "=&r"))]
  ""
- "&& reload_completed"
+ "#"
+ "reload_completed"
  [(set (match_dup 3) (match_dup 4))
   (parallel
    [(set (match_dup 4)
-         (ior:SI (ashiftrt:SI (match_dup 4) (match_dup 2))
-                 (ashift:SI (match_dup 5)
-                            (minus:QI (const_int 32) (match_dup 2)))))
+        (ior:DWIH (ashiftrt:DWIH (match_dup 4) (match_dup 2))
+                  (ashift:DWIH (match_dup 5)
+                               (minus:QI (match_dup 6) (match_dup 2)))))
     (clobber (reg:CC FLAGS_REG))])
   (parallel
    [(set (match_dup 5)
-         (ior:SI (ashiftrt:SI (match_dup 5) (match_dup 2))
-                 (ashift:SI (match_dup 3)
-                            (minus:QI (const_int 32) (match_dup 2)))))
+        (ior:DWIH (ashiftrt:DWIH (match_dup 5) (match_dup 2))
+                  (ashift:DWIH (match_dup 3)
+                               (minus:QI (match_dup 6) (match_dup 2)))))
     (clobber (reg:CC FLAGS_REG))])]
- "split_di (&operands[0], 1, &operands[4], &operands[5]);")
-
-(define_insn "*rotrdi3_1_one_bit_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm")
-       (rotatert: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 (ROTATERT, DImode, operands)"
-  "ror{q}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "DI")])
-
-(define_insn "*rotrdi3_1_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=rm,rm")
-       (rotatert: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 (ROTATERT, DImode, operands)"
-  "@
-   ror{q}\t{%2, %0|%0, %2}
-   ror{q}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "DI")])
+{
+  operands[6] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode));
 
-(define_expand "rotrsi3"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "")
-       (rotatert:SI (match_operand:SI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  ""
-  "ix86_expand_binary_operator (ROTATERT, SImode, operands); DONE;")
+  split_<dwi> (&operands[0], 1, &operands[4], &operands[5]);
+})
 
-(define_insn "*rotrsi3_1_one_bit"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm")
-       (rotatert:SI (match_operand:SI 1 "nonimmediate_operand" "0")
-                    (match_operand:QI 2 "const1_operand" "")))
+(define_insn "*<rotate_insn><mode>3_1"
+  [(set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m")
+       (any_rotate: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 (ROTATERT, SImode, operands)"
-  "ror{l}\t%0"
+  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+{
+  if (operands[2] == const1_rtx
+      && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+    return "<rotate>{<imodesuffix>}\t%0";
+  else
+    return "<rotate>{<imodesuffix>}\t{%2, %0|%0, %2}";
+}
   [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
+   (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 "*rotrsi3_1_one_bit_zext"
+(define_insn "*<rotate_insn>si3_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
-         (rotatert: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 (ROTATERT, SImode, operands)"
-  "ror{l}\t%k0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "SI")])
-
-(define_insn "*rotrsi3_1"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,rm")
-       (rotatert: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 (ROTATERT, SImode, operands)"
-  "@
-   ror{l}\t{%2, %0|%0, %2}
-   ror{l}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "SI")])
-
-(define_insn "*rotrsi3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (zero_extend:DI
-         (rotatert:SI (match_operand:SI 1 "register_operand" "0,0")
-                      (match_operand:QI 2 "nonmemory_operand" "I,c"))))
+         (any_rotate:SI (match_operand:SI 1 "register_operand" "0")
+                        (match_operand:QI 2 "nonmemory_operand" "cI"))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (ROTATERT, SImode, operands)"
-  "@
-   ror{l}\t{%2, %k0|%k0, %2}
-   ror{l}\t{%b2, %k0|%k0, %b2}"
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+{
+    if (operands[2] == const1_rtx
+       && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+    return "<rotate>{l}\t%k0";
+  else
+    return "<rotate>{l}\t{%2, %k0|%k0, %2}";
+}
   [(set_attr "type" "rotate")
+   (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_expand "rotrhi3"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "")
-       (rotatert:HI (match_operand:HI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "ix86_expand_binary_operator (ROTATERT, HImode, operands); DONE;")
-
-(define_insn "*rotrhi3_one_bit"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm")
-       (rotatert: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 (ROTATERT, HImode, operands)"
-  "ror{w}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "HI")])
-
-(define_insn "*rotrhi3_1"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,rm")
-       (rotatert:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
-                    (match_operand:QI 2 "nonmemory_operand" "I,c")))
+(define_insn "*<rotate_insn>qi3_1_slp"
+  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
+       (any_rotate:QI (match_dup 0)
+                      (match_operand:QI 1 "nonmemory_operand" "cI")))
    (clobber (reg:CC FLAGS_REG))]
-  "ix86_binary_operator_ok (ROTATERT, HImode, operands)"
-  "@
-   ror{w}\t{%2, %0|%0, %2}
-   ror{w}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "HI")])
+  "(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 "<rotate>{b}\t%0";
+  else
+    return "<rotate>{b}\t{%1, %0|%0, %1}";
+}
+  [(set_attr "type" "rotate1")
+   (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")])
 
 (define_split
  [(set (match_operand:HI 0 "register_operand" "")
-       (rotatert:HI (match_dup 0) (const_int 8)))
+       (any_rotate:HI (match_dup 0) (const_int 8)))
   (clobber (reg:CC FLAGS_REG))]
- "reload_completed"
+ "reload_completed
+  && (TARGET_USE_XCHGB || optimize_function_for_size_p (cfun))"
  [(parallel [(set (strict_low_part (match_dup 0))
                  (bswap:HI (match_dup 0)))
             (clobber (reg:CC FLAGS_REG))])]
  "")
-
-(define_expand "rotrqi3"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "")
-       (rotatert:QI (match_operand:QI 1 "nonimmediate_operand" "")
-                    (match_operand:QI 2 "nonmemory_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "ix86_expand_binary_operator (ROTATERT, QImode, operands); DONE;")
-
-(define_insn "*rotrqi3_1_one_bit"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm")
-       (rotatert: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 (ROTATERT, QImode, operands)"
-  "ror{b}\t%0"
-  [(set_attr "type" "rotate")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*rotrqi3_1_one_bit_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
-       (rotatert: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))"
-  "ror{b}\t%0"
-  [(set_attr "type" "rotate1")
-   (set_attr "length_immediate" "0")
-   (set_attr "mode" "QI")])
-
-(define_insn "*rotrqi3_1"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,qm")
-       (rotatert: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 (ROTATERT, QImode, operands)"
-  "@
-   ror{b}\t{%2, %0|%0, %2}
-   ror{b}\t{%b2, %0|%0, %b2}"
-  [(set_attr "type" "rotate")
-   (set_attr "mode" "QI")])
-
-(define_insn "*rotrqi3_1_slp"
-  [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm,qm"))
-       (rotatert: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]))"
-  "@
-   ror{b}\t{%1, %0|%0, %1}
-   ror{b}\t{%b1, %0|%0, %b1}"
-  [(set_attr "type" "rotate1")
-   (set_attr "mode" "QI")])
 \f
 ;; Bit set / bit test instructions
 
 ;; For all sCOND expanders, also expand the compare or test insn that
 ;; generates cc0.  Generate an equality comparison if `seq' or `sne'.
 
-;; %%% Do the expansion to SImode.  If PII, do things the xor+setcc way
-;; to avoid partial register stalls.  Otherwise do things the setcc+movzx
-;; way, which can later delete the movzx if only QImode is needed.
+(define_insn_and_split "*setcc_di_1"
+  [(set (match_operand:DI 0 "register_operand" "=q")
+       (match_operator:DI 1 "ix86_comparison_operator"
+         [(reg FLAGS_REG) (const_int 0)]))]
+  "TARGET_64BIT && !TARGET_PARTIAL_REG_STALL"
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 2) (match_dup 1))
+   (set (match_dup 0) (zero_extend:DI (match_dup 2)))]
+{
+  PUT_MODE (operands[1], QImode);
+  operands[2] = gen_lowpart (QImode, operands[0]);
+})
+
+(define_insn_and_split "*setcc_si_1_and"
+  [(set (match_operand:SI 0 "register_operand" "=q")
+       (match_operator:SI 1 "ix86_comparison_operator"
+         [(reg FLAGS_REG) (const_int 0)]))
+   (clobber (reg:CC FLAGS_REG))]
+  "!TARGET_PARTIAL_REG_STALL
+   && TARGET_ZERO_EXTEND_WITH_AND && optimize_function_for_speed_p (cfun)"
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 2) (match_dup 1))
+   (parallel [(set (match_dup 0) (zero_extend:SI (match_dup 2)))
+             (clobber (reg:CC FLAGS_REG))])]
+{
+  PUT_MODE (operands[1], QImode);
+  operands[2] = gen_lowpart (QImode, operands[0]);
+})
+
+(define_insn_and_split "*setcc_si_1_movzbl"
+  [(set (match_operand:SI 0 "register_operand" "=q")
+       (match_operator:SI 1 "ix86_comparison_operator"
+         [(reg FLAGS_REG) (const_int 0)]))]
+  "!TARGET_PARTIAL_REG_STALL
+   && (!TARGET_ZERO_EXTEND_WITH_AND || optimize_function_for_size_p (cfun))"
+  "#"
+  "&& reload_completed"
+  [(set (match_dup 2) (match_dup 1))
+   (set (match_dup 0) (zero_extend:SI (match_dup 2)))]
+{
+  PUT_MODE (operands[1], QImode);
+  operands[2] = gen_lowpart (QImode, operands[0]);
+})
 
-(define_insn "*setcc_1"
+(define_insn "*setcc_qi"
   [(set (match_operand:QI 0 "nonimmediate_operand" "=qm")
        (match_operator:QI 1 "ix86_comparison_operator"
          [(reg FLAGS_REG) (const_int 0)]))]
   [(set_attr "type" "setcc")
    (set_attr "mode" "QI")])
 
-(define_insn "*setcc_2"
+(define_insn "*setcc_qi_slp"
   [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
        (match_operator:QI 1 "ix86_comparison_operator"
          [(reg FLAGS_REG) (const_int 0)]))]
        (match_operator:MODEF 1 "sse_comparison_operator"
          [(match_operand:MODEF 2 "register_operand" "0")
           (match_operand:MODEF 3 "nonimmediate_operand" "xm")]))]
-  "SSE_FLOAT_MODE_P (<MODE>mode) && !TARGET_SSE5"
+  "SSE_FLOAT_MODE_P (<MODE>mode)"
   "cmp%D1s<ssemodefsuffix>\t{%3, %0|%0, %3}"
   [(set_attr "type" "ssecmp")
    (set_attr "length_immediate" "1")
    (set_attr "mode" "<MODE>")])
-
-(define_insn "*sse5_setcc<mode>"
-  [(set (match_operand:MODEF 0 "register_operand" "=x")
-       (match_operator:MODEF 1 "sse5_comparison_float_operator"
-         [(match_operand:MODEF 2 "register_operand" "x")
-          (match_operand:MODEF 3 "nonimmediate_operand" "xm")]))]
-  "TARGET_SSE5"
-  "com%Y1s<ssemodefsuffix>\t{%3, %2, %0|%0, %2, %3}"
-  [(set_attr "type" "sse4arg")
-   (set_attr "length_immediate" "1")
-   (set_attr "mode" "<MODE>")])
-
 \f
 ;; Basic conditional jump instructions.
 ;; We ignore the overflow flag for signed branch instructions.
                      (pc)))]
   "PUT_CODE (operands[0], reverse_condition (GET_CODE (operands[0])));")
 
-;; Define combination compare-and-branch fp compare instructions to use
-;; during early optimization.  Splitting the operation apart early makes
-;; for bad code when we want to reverse the operation.
-
-(define_insn "*fp_jcc_1_mixed"
-  [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
-                       [(match_operand 1 "register_operand" "f,x")
-                        (match_operand 2 "nonimmediate_operand" "f,xm")])
-         (label_ref (match_operand 3 "" ""))
-         (pc)))
-   (clobber (reg:CCFP FPSR_REG))
-   (clobber (reg:CCFP FLAGS_REG))]
-  "TARGET_MIX_SSE_I387
-   && SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
-   && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
-  "#")
-
-(define_insn "*fp_jcc_1_sse"
-  [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
-                       [(match_operand 1 "register_operand" "x")
-                        (match_operand 2 "nonimmediate_operand" "xm")])
-         (label_ref (match_operand 3 "" ""))
-         (pc)))
-   (clobber (reg:CCFP FPSR_REG))
-   (clobber (reg:CCFP FLAGS_REG))]
-  "TARGET_SSE_MATH
-   && SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
-   && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
-  "#")
-
-(define_insn "*fp_jcc_1_387"
-  [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
-                       [(match_operand 1 "register_operand" "f")
-                        (match_operand 2 "register_operand" "f")])
-         (label_ref (match_operand 3 "" ""))
-         (pc)))
-   (clobber (reg:CCFP FPSR_REG))
-   (clobber (reg:CCFP FLAGS_REG))]
-  "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
-   && TARGET_CMOVE
-   && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
-  "#")
-
-(define_insn "*fp_jcc_2_mixed"
-  [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
-                       [(match_operand 1 "register_operand" "f,x")
-                        (match_operand 2 "nonimmediate_operand" "f,xm")])
-         (pc)
-         (label_ref (match_operand 3 "" ""))))
-   (clobber (reg:CCFP FPSR_REG))
-   (clobber (reg:CCFP FLAGS_REG))]
-  "TARGET_MIX_SSE_I387
-   && SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
-   && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
-  "#")
-
-(define_insn "*fp_jcc_2_sse"
-  [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
-                       [(match_operand 1 "register_operand" "x")
-                        (match_operand 2 "nonimmediate_operand" "xm")])
-         (pc)
-         (label_ref (match_operand 3 "" ""))))
-   (clobber (reg:CCFP FPSR_REG))
-   (clobber (reg:CCFP FLAGS_REG))]
-  "TARGET_SSE_MATH
-   && SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
-   && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
-  "#")
-
-(define_insn "*fp_jcc_2_387"
-  [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
-                       [(match_operand 1 "register_operand" "f")
-                        (match_operand 2 "register_operand" "f")])
-         (pc)
-         (label_ref (match_operand 3 "" ""))))
-   (clobber (reg:CCFP FPSR_REG))
-   (clobber (reg:CCFP FLAGS_REG))]
-  "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
-   && TARGET_CMOVE
-   && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
-  "#")
+;; Define combination compare-and-branch fp compare instructions to help
+;; combine.
 
 (define_insn "*fp_jcc_3_387"
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "f")
                         (match_operand 2 "nonimmediate_operand" "fm")])
          (label_ref (match_operand 3 "" ""))
   "TARGET_80387
    && (GET_MODE (operands[1]) == SFmode || GET_MODE (operands[1]) == DFmode)
    && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && !ix86_use_fcomi_compare (GET_CODE (operands[0]))
    && SELECT_CC_MODE (GET_CODE (operands[0]),
                      operands[1], operands[2]) == CCFPmode
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
+   && !TARGET_CMOVE"
   "#")
 
 (define_insn "*fp_jcc_4_387"
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "f")
                         (match_operand 2 "nonimmediate_operand" "fm")])
          (pc)
   "TARGET_80387
    && (GET_MODE (operands[1]) == SFmode || GET_MODE (operands[1]) == DFmode)
    && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && !ix86_use_fcomi_compare (GET_CODE (operands[0]))
    && SELECT_CC_MODE (GET_CODE (operands[0]),
                      operands[1], operands[2]) == CCFPmode
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
+   && !TARGET_CMOVE"
   "#")
 
 (define_insn "*fp_jcc_5_387"
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "f")
                         (match_operand 2 "register_operand" "f")])
          (label_ref (match_operand 3 "" ""))
    (clobber (match_scratch:HI 4 "=a"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
+   && !TARGET_CMOVE"
   "#")
 
 (define_insn "*fp_jcc_6_387"
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "f")
                         (match_operand 2 "register_operand" "f")])
          (pc)
    (clobber (match_scratch:HI 4 "=a"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
+   && !TARGET_CMOVE"
   "#")
 
 (define_insn "*fp_jcc_7_387"
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "f")
                         (match_operand 2 "const0_operand" "")])
          (label_ref (match_operand 3 "" ""))
    (clobber (match_scratch:HI 4 "=a"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && GET_MODE (operands[1]) == GET_MODE (operands[2])
-   && !ix86_use_fcomi_compare (GET_CODE (operands[0]))
    && SELECT_CC_MODE (GET_CODE (operands[0]),
                      operands[1], operands[2]) == CCFPmode
-   && ix86_fp_jump_nontrivial_p (GET_CODE (operands[0]))"
+   && !TARGET_CMOVE"
   "#")
 
 ;; The order of operands in *fp_jcc_8_387 is forced by combine in
 
 (define_insn "*fp_jcc_8<mode>_387"
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operator 1 "float_operator"
                           [(match_operand:X87MODEI12 2 "nonimmediate_operand" "m,?r")])
                           (match_operand 3 "register_operand" "f,f")])
   "X87_FLOAT_MODE_P (GET_MODE (operands[3]))
    && (TARGET_USE_<MODE>MODE_FIOP || optimize_function_for_size_p (cfun))
    && GET_MODE (operands[1]) == GET_MODE (operands[3])
-   && !ix86_use_fcomi_compare (swap_condition (GET_CODE (operands[0])))
    && ix86_fp_compare_mode (swap_condition (GET_CODE (operands[0]))) == CCFPmode
-   && ix86_fp_jump_nontrivial_p (swap_condition (GET_CODE (operands[0])))"
+   && !TARGET_CMOVE"
   "#")
 
 (define_split
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "")
                         (match_operand 2 "nonimmediate_operand" "")])
          (match_operand 3 "" "")
 
 (define_split
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operand 1 "register_operand" "")
                         (match_operand 2 "general_operand" "")])
          (match_operand 3 "" "")
 
 (define_split
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operator 1 "float_operator"
                           [(match_operand:X87MODEI12 2 "memory_operand" "")])
                           (match_operand 3 "register_operand" "")])
 ;; %%% Kill this when reload knows how to do it.
 (define_split
   [(set (pc)
-       (if_then_else (match_operator 0 "comparison_operator"
+       (if_then_else (match_operator 0 "ix86_fp_comparison_operator"
                        [(match_operator 1 "float_operator"
                           [(match_operand:X87MODEI12 2 "register_operand" "")])
                           (match_operand 3 "register_operand" "")])
 ;; checked for calls.  This is a bug in the generic code, but it isn't that
 ;; easy to fix.  Ignore it for now and be prepared to fix things up.
 
+;; P6 processors will jump to the address after the decrement when %esp
+;; is used as a call operand, so they will execute return address as a code.
+;; See Pentium Pro errata 70, Pentium 2 errata A33 and Pentium 3 errata E17.
 ;; Call subroutine returning no value.
 
 (define_expand "call_pop"
                            (match_operand:SI 3 "" "")))])]
   "!TARGET_64BIT"
 {
-  ix86_expand_call (NULL, operands[0], operands[1], operands[2], operands[3], 0);
+  ix86_expand_call (NULL, operands[0], operands[1],
+                   operands[2], operands[3], 0);
   DONE;
 })
 
 (define_insn "*call_pop_0"
   [(call (mem:QI (match_operand:SI 0 "constant_call_address_operand" ""))
         (match_operand:SI 1 "" ""))
-   (set (reg:SI SP_REG) (plus:SI (reg:SI SP_REG)
-                           (match_operand:SI 2 "immediate_operand" "")))]
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "")))]
   "!TARGET_64BIT"
 {
   if (SIBLING_CALL_P (insn))
   [(set_attr "type" "call")])
 
 (define_insn "*call_pop_1"
-  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "rsm"))
+  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "lsm"))
         (match_operand:SI 1 "" ""))
-   (set (reg:SI SP_REG) (plus:SI (reg:SI SP_REG)
-                           (match_operand:SI 2 "immediate_operand" "i")))]
-  "!TARGET_64BIT"
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "i")))]
+  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (constant_call_address_operand (operands[0], Pmode))
-    {
-      if (SIBLING_CALL_P (insn))
-       return "jmp\t%P0";
-      else
-       return "call\t%P0";
-    }
-  if (SIBLING_CALL_P (insn))
-    return "jmp\t%A0";
-  else
-    return "call\t%A0";
+    return "call\t%P0";
+  return "call\t%A0";
 }
   [(set_attr "type" "call")])
 
+(define_insn "*sibcall_pop_1"
+  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "s,U"))
+        (match_operand:SI 1 "" ""))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "i,i")))]
+  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "@
+   jmp\t%P0
+   jmp\t%A0"
+  [(set_attr "type" "call")])
+
 (define_expand "call"
   [(call (match_operand:QI 0 "" "")
         (match_operand 1 "" ""))
   [(set_attr "type" "call")])
 
 (define_insn "*call_1"
-  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "rsm"))
+  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "lsm"))
         (match_operand 1 "" ""))]
-  "!SIBLING_CALL_P (insn) && !TARGET_64BIT"
+  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (constant_call_address_operand (operands[0], Pmode))
     return "call\t%P0";
   [(set_attr "type" "call")])
 
 (define_insn "*sibcall_1"
-  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "s,c,d,a"))
+  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "s,U"))
         (match_operand 1 "" ""))]
-  "SIBLING_CALL_P (insn) && !TARGET_64BIT"
-{
-  if (constant_call_address_operand (operands[0], Pmode))
-    return "jmp\t%P0";
-  return "jmp\t%A0";
-}
+  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "@
+   jmp\t%P0
+   jmp\t%A0"
   [(set_attr "type" "call")])
 
 (define_insn "*call_1_rex64"
   [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rsm"))
         (match_operand 1 "" ""))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)
    && ix86_cmodel != CM_LARGE && ix86_cmodel != CM_LARGE_PIC"
 {
   if (constant_call_address_operand (operands[0], Pmode))
    (clobber (reg:TI XMM15_REG))
    (clobber (reg:DI SI_REG))
    (clobber (reg:DI DI_REG))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT"
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (constant_call_address_operand (operands[0], Pmode))
     return "call\t%P0";
 (define_insn "*call_1_rex64_large"
   [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rm"))
         (match_operand 1 "" ""))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT"
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
   "call\t%A0"
   [(set_attr "type" "call")])
 
 (define_insn "*sibcall_1_rex64"
-  [(call (mem:QI (match_operand:DI 0 "constant_call_address_operand" ""))
+  [(call (mem:QI (match_operand:DI 0 "sibcall_insn_operand" "s,U"))
         (match_operand 1 "" ""))]
-  "SIBLING_CALL_P (insn) && TARGET_64BIT"
-  "jmp\t%P0"
-  [(set_attr "type" "call")])
-
-(define_insn "*sibcall_1_rex64_v"
-  [(call (mem:QI (reg:DI R11_REG))
-        (match_operand 0 "" ""))]
-  "SIBLING_CALL_P (insn) && TARGET_64BIT"
-  "jmp\t{*%%}r11"
+  "TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "@
+   jmp\t%P0
+   jmp\t%A0"
   [(set_attr "type" "call")])
 
-
 ;; Call subroutine, returning value in operand 0
-
 (define_expand "call_value_pop"
   [(parallel [(set (match_operand 0 "" "")
                   (call (match_operand:QI 1 "" "")
        (call (match_operand:QI 1 "" "")
              (match_operand:SI 2 "" "")))
    (use (match_operand:SI 3 "" ""))]
-  ;; Operand 2 not used on the i386.
+  ;; Operand 3 is not used on the i386.
   ""
 {
-  ix86_expand_call (operands[0], operands[1], operands[2], operands[3], NULL, 0);
+  ix86_expand_call (operands[0], operands[1], operands[2],
+                   operands[3], NULL, 0);
   DONE;
 })
 
        (call (match_operand:QI 1 "" "")
              (match_operand:SI 2 "" "")))
    (use (match_operand:SI 3 "" ""))]
-  ;; Operand 2 not used on the i386.
+  ;; Operand 3 is not used on the i386.
   ""
 {
-  ix86_expand_call (operands[0], operands[1], operands[2], operands[3], NULL, 1);
+  ix86_expand_call (operands[0], operands[1], operands[2],
+                   operands[3], NULL, 1);
   DONE;
 })
 
                    GEN_INT ((TARGET_64BIT
                              ? (ix86_abi == SYSV_ABI
                                 ? X86_64_SSE_REGPARM_MAX
-                                : X64_SSE_REGPARM_MAX)
+                                : X86_64_MS_SSE_REGPARM_MAX)
                              : X86_32_SSE_REGPARM_MAX)
                             - 1),
                    NULL, 0);
    (set_attr "length_immediate" "0")
    (set_attr "modrm" "0")])
 
+(define_insn "vswapmov"
+  [(set (match_operand:SI 0 "register_operand" "=r")
+        (match_operand:SI 1 "register_operand" "r"))
+   (unspec_volatile [(const_int 0)] UNSPECV_VSWAPMOV)]
+  ""
+  "movl.s\t{%1, %0|%0, %1}"
+  [(set_attr "length" "2")
+   (set_attr "length_immediate" "0")
+   (set_attr "modrm" "0")])
+
 ;; Pad to 16-byte boundary, max skip in op0.  Used to avoid
 ;; branch prediction penalty for the third jump in a 16-byte
 ;; block on K8.
    (set_attr "type" "bitmanip")
    (set_attr "mode" "SI")])
 
-(define_insn "*bsr"
+(define_insn "bsr"
   [(set (match_operand:SI 0 "register_operand" "=r")
        (minus:SI (const_int 31)
                  (clz:SI (match_operand:SI 1 "nonimmediate_operand" "rm"))))
    (set_attr "type" "bitmanip")
    (set_attr "mode" "DI")])
 
-(define_insn "*bsr_rex64"
+(define_insn "bsr_rex64"
   [(set (match_operand:DI 0 "register_operand" "=r")
        (minus:DI (const_int 63)
                  (clz:DI (match_operand:DI 1 "nonimmediate_operand" "rm"))))
    (clobber (match_scratch:SI 5 "=c"))
    (clobber (reg:CC FLAGS_REG))]
   "!TARGET_64BIT && TARGET_GNU_TLS"
-  "lea{l}\t{%a2@TLSGD(,%1,1), %0|%0, %a2@TLSGD[%1*1]}\;call\t%P3"
+  "lea{l}\t{%a2@tlsgd(,%1,1), %0|%0, %a2@tlsgd[%1*1]}\;call\t%P3"
   [(set_attr "type" "multi")
    (set_attr "length" "12")])
 
-(define_insn "*tls_global_dynamic_32_sun"
-  [(set (match_operand:SI 0 "register_operand" "=a")
-       (unspec:SI [(match_operand:SI 1 "register_operand" "b")
-                   (match_operand:SI 2 "tls_symbolic_operand" "")
-                   (match_operand:SI 3 "call_insn_operand" "")]
-                   UNSPEC_TLS_GD))
-   (clobber (match_scratch:SI 4 "=d"))
-   (clobber (match_scratch:SI 5 "=c"))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && TARGET_SUN_TLS"
-  "lea{l}\t{%a2@DTLNDX(%1), %4|%4, %a2@DTLNDX[%1]}
-       push{l}\t%4\;call\t%a2@TLSPLT\;pop{l}\t%4\;nop"
-  [(set_attr "type" "multi")
-   (set_attr "length" "14")])
-
 (define_expand "tls_global_dynamic_32"
   [(parallel [(set (match_operand:SI 0 "register_operand" "")
                   (unspec:SI
    (unspec:DI [(match_operand:DI 1 "tls_symbolic_operand" "")]
              UNSPEC_TLS_GD)]
   "TARGET_64BIT"
-  ".byte\t0x66\;lea{q}\t{%a1@TLSGD(%%rip), %%rdi|rdi, %a1@TLSGD[rip]}\;.word\t0x6666\;rex64\;call\t%P2"
+  { return ASM_BYTE "0x66\n\tlea{q}\t{%a1@tlsgd(%%rip), %%rdi|rdi, %a1@tlsgd[rip]}\n" ASM_SHORT "0x6666\n\trex64\n\tcall\t%P2"; }
   [(set_attr "type" "multi")
    (set_attr "length" "16")])
 
    (clobber (match_scratch:SI 4 "=c"))
    (clobber (reg:CC FLAGS_REG))]
   "!TARGET_64BIT && TARGET_GNU_TLS"
-  "lea{l}\t{%&@TLSLDM(%1), %0|%0, %&@TLSLDM[%1]}\;call\t%P2"
+  "lea{l}\t{%&@tlsldm(%1), %0|%0, %&@tlsldm[%1]}\;call\t%P2"
   [(set_attr "type" "multi")
    (set_attr "length" "11")])
 
-(define_insn "*tls_local_dynamic_base_32_sun"
-  [(set (match_operand:SI 0 "register_operand" "=a")
-       (unspec:SI [(match_operand:SI 1 "register_operand" "b")
-                    (match_operand:SI 2 "call_insn_operand" "")]
-                  UNSPEC_TLS_LD_BASE))
-   (clobber (match_scratch:SI 3 "=d"))
-   (clobber (match_scratch:SI 4 "=c"))
-   (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_64BIT && TARGET_SUN_TLS"
-  "lea{l}\t{%&@TMDNX(%1), %3|%3, %&@TMDNX[%1]}
-       push{l}\t%3\;call\t%&@TLSPLT\;pop{l}\t%3"
-  [(set_attr "type" "multi")
-   (set_attr "length" "13")])
-
 (define_expand "tls_local_dynamic_base_32"
   [(parallel [(set (match_operand:SI 0 "register_operand" "")
                   (unspec:SI [(match_dup 1) (match_dup 2)]
                 (match_operand:DI 2 "" "")))
    (unspec:DI [(const_int 0)] UNSPEC_TLS_LD_BASE)]
   "TARGET_64BIT"
-  "lea{q}\t{%&@TLSLD(%%rip), %%rdi|rdi, %&@TLSLD[rip]}\;call\t%P1"
+  "lea{q}\t{%&@tlsld(%%rip), %%rdi|rdi, %&@tlsld[rip]}\;call\t%P1"
   [(set_attr "type" "multi")
    (set_attr "length" "12")])
 
      (clobber (reg:CC FLAGS_REG))])]
   "!TARGET_64BIT && TARGET_GNU2_TLS"
 {
-  operands[3] = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode);
+  operands[3] = can_create_pseudo_p () ? gen_reg_rtx (Pmode) : operands[0];
   ix86_tls_descriptor_calls_expanded_in_cfun = true;
 })
 
   ""
   [(set (match_dup 0) (match_dup 5))]
 {
-  operands[5] = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode);
+  operands[5] = can_create_pseudo_p () ? gen_reg_rtx (Pmode) : operands[0];
   emit_insn (gen_tls_dynamic_gnu2_32 (operands[5], operands[1], operands[2]));
 })
 
      (clobber (reg:CC FLAGS_REG))])]
   "TARGET_64BIT && TARGET_GNU2_TLS"
 {
-  operands[2] = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode);
+  operands[2] = can_create_pseudo_p () ? gen_reg_rtx (Pmode) : operands[0];
   ix86_tls_descriptor_calls_expanded_in_cfun = true;
 })
 
   ""
   [(set (match_dup 0) (match_dup 4))]
 {
-  operands[4] = !can_create_pseudo_p () ? operands[0] : gen_reg_rtx (Pmode);
+  operands[4] = can_create_pseudo_p () ? gen_reg_rtx (Pmode) : operands[0];
   emit_insn (gen_tls_dynamic_gnu2_64 (operands[4], operands[1]));
 })
 
 (define_expand "scalb<mode>3"
   [(use (match_operand:MODEF 0 "register_operand" ""))
    (use (match_operand:MODEF 1 "general_operand" ""))
-   (use (match_operand:MODEF 2 "register_operand" ""))]
+   (use (match_operand:MODEF 2 "general_operand" ""))]
  "TARGET_USE_FANCY_MATH_387
    && (!(SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)
        || TARGET_MIX_SSE_I387)
   op1 = gen_reg_rtx (XFmode);
   op2 = gen_reg_rtx (XFmode);
 
-  emit_insn (gen_extend<mode>xf2 (op1, operands[1]));
-  emit_insn (gen_extend<mode>xf2 (op2, operands[2]));
-  emit_insn (gen_scalbxf3 (op0, op1, op2));
+  emit_insn (gen_extend<mode>xf2 (op1, operands[1]));
+  emit_insn (gen_extend<mode>xf2 (op2, operands[2]));
+  emit_insn (gen_scalbxf3 (op0, op1, op2));
+  emit_insn (gen_truncxf<mode>2_i387_noop (operands[0], op0));
+  DONE;
+})
+
+(define_expand "significandxf2"
+  [(parallel [(set (match_operand:XF 0 "register_operand" "")
+                  (unspec:XF [(match_operand:XF 1 "register_operand" "")]
+                             UNSPEC_XTRACT_FRACT))
+             (set (match_dup 2)
+                  (unspec:XF [(match_dup 1)] UNSPEC_XTRACT_EXP))])]
+  "TARGET_USE_FANCY_MATH_387
+   && flag_unsafe_math_optimizations"
+{
+  operands[2] = gen_reg_rtx (XFmode);
+})
+
+(define_expand "significand<mode>2"
+  [(use (match_operand:MODEF 0 "register_operand" ""))
+   (use (match_operand:MODEF 1 "register_operand" ""))]
+  "TARGET_USE_FANCY_MATH_387
+   && (!(SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)
+       || TARGET_MIX_SSE_I387)
+   && flag_unsafe_math_optimizations"
+{
+  rtx op0 = gen_reg_rtx (XFmode);
+  rtx op1 = gen_reg_rtx (XFmode);
+
+  emit_insn (gen_fxtract_extend<mode>xf3_i387 (op0, op1, operands[1]));
   emit_insn (gen_truncxf<mode>2_i387_noop (operands[0], op0));
   DONE;
 })
        (unspec:DI [(match_operand:XF 1 "register_operand" "")]
                   UNSPEC_FIST))]
   "TARGET_USE_FANCY_MATH_387
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
        (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
                           UNSPEC_FIST))]
   "TARGET_USE_FANCY_MATH_387
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    && flag_unsafe_math_optimizations"
   "")
 
-(define_expand "lfloor<mode>di2"
-  [(match_operand:DI 0 "nonimmediate_operand" "")
-   (match_operand:MODEF 1 "register_operand" "")]
-  "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH && TARGET_64BIT
-   && !flag_trapping_math"
-{
-  if (optimize_insn_for_size_p ())
-    FAIL;
-  ix86_expand_lfloorceil (operand0, operand1, true);
-  DONE;
-})
-
-(define_expand "lfloor<mode>si2"
-  [(match_operand:SI 0 "nonimmediate_operand" "")
+(define_expand "lfloor<MODEF:mode><SWI48:mode>2"
+  [(match_operand:SWI48 0 "nonimmediate_operand" "")
    (match_operand:MODEF 1 "register_operand" "")]
-  "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
+  "SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && !flag_trapping_math"
 {
-  if (optimize_insn_for_size_p () && TARGET_64BIT)
+  if (TARGET_64BIT && optimize_insn_for_size_p ())
     FAIL;
   ix86_expand_lfloorceil (operand0, operand1, true);
   DONE;
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    && flag_unsafe_math_optimizations"
   "")
 
-(define_expand "lceil<mode>di2"
-  [(match_operand:DI 0 "nonimmediate_operand" "")
-   (match_operand:MODEF 1 "register_operand" "")]
-  "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH && TARGET_64BIT
-   && !flag_trapping_math"
-{
-  ix86_expand_lfloorceil (operand0, operand1, false);
-  DONE;
-})
-
-(define_expand "lceil<mode>si2"
-  [(match_operand:SI 0 "nonimmediate_operand" "")
+(define_expand "lceil<MODEF:mode><SWI48:mode>2"
+  [(match_operand:SWI48 0 "nonimmediate_operand" "")
    (match_operand:MODEF 1 "register_operand" "")]
-  "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
+  "SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && !flag_trapping_math"
 {
   ix86_expand_lfloorceil (operand0, operand1, false);
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(const_int 0)]
          [(match_operand:MODEF 1 "memory_operand" "")]
          UNSPEC_FXAM_MEM))]
   "TARGET_USE_FANCY_MATH_387
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   "#"
   "&& 1"
   [(set (match_dup 2)(match_dup 1))
     }
   else
     {
+      rtx (*cmp_insn)(rtx, rtx);
+
       if (TARGET_64BIT)
-       emit_insn (gen_cmpdi_1_rex64 (countreg, countreg));
+       cmp_insn = gen_cmpdi_1;
       else
-       emit_insn (gen_cmpsi_1 (countreg, countreg));
+       cmp_insn = gen_cmpsi_1;
+      emit_insn (cmp_insn (countreg, countreg));
       emit_insn (gen_cmpstrnqi_1 (addr1, addr2, countreg, align,
                                  operands[1], operands[2]));
     }
 \f
 ;; Conditional move instructions.
 
-(define_expand "movdicc"
-  [(set (match_operand:DI 0 "register_operand" "")
-       (if_then_else:DI (match_operand 1 "comparison_operator" "")
-                        (match_operand:DI 2 "general_operand" "")
-                        (match_operand:DI 3 "general_operand" "")))]
-  "TARGET_64BIT"
+(define_expand "mov<mode>cc"
+  [(set (match_operand:SWIM 0 "register_operand" "")
+       (if_then_else:SWIM (match_operand 1 "comparison_operator" "")
+                          (match_operand:SWIM 2 "general_operand" "")
+                          (match_operand:SWIM 3 "general_operand" "")))]
+  ""
   "if (ix86_expand_int_movcc (operands)) DONE; else FAIL;")
 
-(define_insn "x86_movdicc_0_m1_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (if_then_else:DI (match_operand 1 "ix86_carry_flag_operator" "")
+;; Data flow gets confused by our desire for `sbbl reg,reg', and clearing
+;; the register first winds up with `sbbl $0,reg', which is also weird.
+;; So just document what we're doing explicitly.
+
+(define_expand "x86_mov<mode>cc_0_m1"
+  [(parallel
+    [(set (match_operand:SWI48 0 "register_operand" "")
+         (if_then_else:SWI48
+           (match_operator:SWI48 2 "ix86_carry_flag_operator"
+            [(match_operand 1 "flags_reg_operand" "")
+             (const_int 0)])
+           (const_int -1)
+           (const_int 0)))
+     (clobber (reg:CC FLAGS_REG))])]
+  ""
+  "")
+
+(define_insn "*x86_mov<mode>cc_0_m1"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (if_then_else:SWI48 (match_operator 1 "ix86_carry_flag_operator"
+                            [(reg FLAGS_REG) (const_int 0)])
          (const_int -1)
          (const_int 0)))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT"
-  "sbb{q}\t%0, %0"
+  ""
+  "sbb{<imodesuffix>}\t%0, %0"
   ; Since we don't have the proper number of operands for an alu insn,
   ; fill in all the blanks.
   [(set_attr "type" "alu")
    (set_attr "pent_pair" "pu")
    (set_attr "memory" "none")
    (set_attr "imm_disp" "false")
-   (set_attr "mode" "DI")
-   (set_attr "length_immediate" "0")])
-
-(define_insn "*x86_movdicc_0_m1_se"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (sign_extract:DI (match_operand 1 "ix86_carry_flag_operator" "")
-                        (const_int 1)
-                        (const_int 0)))
-   (clobber (reg:CC FLAGS_REG))]
-  ""
-  "sbb{q}\t%0, %0"
-  [(set_attr "type" "alu")
-   (set_attr "use_carry" "1")
-   (set_attr "pent_pair" "pu")
-   (set_attr "memory" "none")
-   (set_attr "imm_disp" "false")
-   (set_attr "mode" "DI")
+   (set_attr "mode" "<MODE>")
    (set_attr "length_immediate" "0")])
 
-(define_insn "*movdicc_c_rex64"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
-       (if_then_else:DI (match_operator 1 "ix86_comparison_operator"
-                               [(reg FLAGS_REG) (const_int 0)])
-                     (match_operand:DI 2 "nonimmediate_operand" "rm,0")
-                     (match_operand:DI 3 "nonimmediate_operand" "0,rm")))]
-  "TARGET_64BIT && TARGET_CMOVE
-   && !(MEM_P (operands[2]) && MEM_P (operands[3]))"
-  "@
-   cmov%O2%C1\t{%2, %0|%0, %2}
-   cmov%O2%c1\t{%3, %0|%0, %3}"
-  [(set_attr "type" "icmov")
-   (set_attr "mode" "DI")])
-
-(define_expand "movsicc"
-  [(set (match_operand:SI 0 "register_operand" "")
-       (if_then_else:SI (match_operand 1 "comparison_operator" "")
-                        (match_operand:SI 2 "general_operand" "")
-                        (match_operand:SI 3 "general_operand" "")))]
-  ""
-  "if (ix86_expand_int_movcc (operands)) DONE; else FAIL;")
-
-;; Data flow gets confused by our desire for `sbbl reg,reg', and clearing
-;; the register first winds up with `sbbl $0,reg', which is also weird.
-;; So just document what we're doing explicitly.
-
-(define_insn "x86_movsicc_0_m1"
-  [(set (match_operand:SI 0 "register_operand" "=r")
-       (if_then_else:SI (match_operand 1 "ix86_carry_flag_operator" "")
-         (const_int -1)
-         (const_int 0)))
+(define_insn "*x86_mov<mode>cc_0_m1_se"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (sign_extract:SWI48 (match_operator 1 "ix86_carry_flag_operator"
+                            [(reg FLAGS_REG) (const_int 0)])
+                           (const_int 1)
+                           (const_int 0)))
    (clobber (reg:CC FLAGS_REG))]
   ""
-  "sbb{l}\t%0, %0"
-  ; Since we don't have the proper number of operands for an alu insn,
-  ; fill in all the blanks.
+  "sbb{<imodesuffix>}\t%0, %0"
   [(set_attr "type" "alu")
    (set_attr "use_carry" "1")
    (set_attr "pent_pair" "pu")
    (set_attr "memory" "none")
    (set_attr "imm_disp" "false")
-   (set_attr "mode" "SI")
+   (set_attr "mode" "<MODE>")
    (set_attr "length_immediate" "0")])
 
-(define_insn "*x86_movsicc_0_m1_se"
-  [(set (match_operand:SI 0 "register_operand" "=r")
-       (sign_extract:SI (match_operand 1 "ix86_carry_flag_operator" "")
-                        (const_int 1)
-                        (const_int 0)))
-   (clobber (reg:CC FLAGS_REG))]
+(define_insn "*x86_mov<mode>cc_0_m1_neg"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (neg:SWI48 (match_operator 1 "ix86_carry_flag_operator"
+                   [(reg FLAGS_REG) (const_int 0)])))]
   ""
-  "sbb{l}\t%0, %0"
+  "sbb{<imodesuffix>}\t%0, %0"
   [(set_attr "type" "alu")
    (set_attr "use_carry" "1")
    (set_attr "pent_pair" "pu")
    (set_attr "memory" "none")
    (set_attr "imm_disp" "false")
-   (set_attr "mode" "SI")
+   (set_attr "mode" "<MODE>")
    (set_attr "length_immediate" "0")])
 
-(define_insn "*movsicc_noc"
-  [(set (match_operand:SI 0 "register_operand" "=r,r")
-       (if_then_else:SI (match_operator 1 "ix86_comparison_operator"
-                               [(reg FLAGS_REG) (const_int 0)])
-                     (match_operand:SI 2 "nonimmediate_operand" "rm,0")
-                     (match_operand:SI 3 "nonimmediate_operand" "0,rm")))]
-  "TARGET_CMOVE
-   && !(MEM_P (operands[2]) && MEM_P (operands[3]))"
-  "@
-   cmov%O2%C1\t{%2, %0|%0, %2}
-   cmov%O2%c1\t{%3, %0|%0, %3}"
-  [(set_attr "type" "icmov")
-   (set_attr "mode" "SI")])
-
-(define_expand "movhicc"
-  [(set (match_operand:HI 0 "register_operand" "")
-       (if_then_else:HI (match_operand 1 "comparison_operator" "")
-                        (match_operand:HI 2 "general_operand" "")
-                        (match_operand:HI 3 "general_operand" "")))]
-  "TARGET_HIMODE_MATH"
-  "if (ix86_expand_int_movcc (operands)) DONE; else FAIL;")
-
-(define_insn "*movhicc_noc"
-  [(set (match_operand:HI 0 "register_operand" "=r,r")
-       (if_then_else:HI (match_operator 1 "ix86_comparison_operator"
-                               [(reg FLAGS_REG) (const_int 0)])
-                     (match_operand:HI 2 "nonimmediate_operand" "rm,0")
-                     (match_operand:HI 3 "nonimmediate_operand" "0,rm")))]
-  "TARGET_CMOVE
-   && !(MEM_P (operands[2]) && MEM_P (operands[3]))"
+(define_insn "*mov<mode>cc_noc"
+  [(set (match_operand:SWI248 0 "register_operand" "=r,r")
+       (if_then_else:SWI248 (match_operator 1 "ix86_comparison_operator"
+                              [(reg FLAGS_REG) (const_int 0)])
+         (match_operand:SWI248 2 "nonimmediate_operand" "rm,0")
+         (match_operand:SWI248 3 "nonimmediate_operand" "0,rm")))]
+  "TARGET_CMOVE && !(MEM_P (operands[2]) && MEM_P (operands[3]))"
   "@
    cmov%O2%C1\t{%2, %0|%0, %2}
    cmov%O2%c1\t{%3, %0|%0, %3}"
   [(set_attr "type" "icmov")
-   (set_attr "mode" "HI")])
-
-(define_expand "movqicc"
-  [(set (match_operand:QI 0 "register_operand" "")
-       (if_then_else:QI (match_operand 1 "comparison_operator" "")
-                        (match_operand:QI 2 "general_operand" "")
-                        (match_operand:QI 3 "general_operand" "")))]
-  "TARGET_QIMODE_MATH"
-  "if (ix86_expand_int_movcc (operands)) DONE; else FAIL;")
+   (set_attr "mode" "<MODE>")])
 
 (define_insn_and_split "*movqicc_noc"
   [(set (match_operand:QI 0 "register_operand" "=r,r")
        (if_then_else:QI (match_operator 1 "ix86_comparison_operator"
-                               [(match_operand 4 "flags_reg_operand" "")
-                                (const_int 0)])
+                          [(match_operand 4 "flags_reg_operand" "")
+                           (const_int 0)])
                      (match_operand:QI 2 "register_operand" "r,0")
                      (match_operand:QI 3 "register_operand" "0,r")))]
   "TARGET_CMOVE && !TARGET_PARTIAL_REG_STALL"
 (define_expand "mov<mode>cc"
   [(set (match_operand:X87MODEF 0 "register_operand" "")
        (if_then_else:X87MODEF
-         (match_operand 1 "comparison_operator" "")
+         (match_operand 1 "ix86_fp_comparison_operator" "")
          (match_operand:X87MODEF 2 "register_operand" "")
          (match_operand:X87MODEF 3 "register_operand" "")))]
   "(TARGET_80387 && TARGET_CMOVE)
   [(set_attr "type" "fcmov")
    (set_attr "mode" "XF")])
 
-;; All moves in SSE5 pcmov instructions are 128 bits and hence we restrict
+;; All moves in XOP pcmov instructions are 128 bits and hence we restrict
 ;; the scalar versions to have only XMM registers as operands.
 
-;; SSE5 conditional move
-(define_insn "*sse5_pcmov_<mode>"
-  [(set (match_operand:MODEF 0 "register_operand" "=x,x")
+;; XOP conditional move
+(define_insn "*xop_pcmov_<mode>"
+  [(set (match_operand:MODEF 0 "register_operand" "=x")
        (if_then_else:MODEF
-         (match_operand:MODEF 1 "register_operand" "x,0")
-         (match_operand:MODEF 2 "register_operand" "0,x")
-         (match_operand:MODEF 3 "register_operand" "x,x")))]
-  "TARGET_SSE5 && ix86_sse5_valid_op_p (operands, insn, 4, true, 1, false)"
-  "pcmov\t{%1, %3, %2, %0|%0, %2, %3, %1}"
+         (match_operand:MODEF 1 "register_operand" "x")
+         (match_operand:MODEF 2 "register_operand" "x")
+         (match_operand:MODEF 3 "register_operand" "x")))]
+  "TARGET_XOP"
+  "vpcmov\t{%1, %3, %2, %0|%0, %2, %3, %1}"
   [(set_attr "type" "sse4arg")])
 
 ;; These versions of the min/max patterns are intentionally ignorant of
   DONE;
 })
 
+;; Use IOR for stack probes, this is shorter.
+(define_expand "probe_stack"
+  [(match_operand 0 "memory_operand" "")]
+  ""
+{
+  if (GET_MODE (operands[0]) == DImode)
+    emit_insn (gen_iordi3 (operands[0], operands[0], const0_rtx));
+  else
+    emit_insn (gen_iorsi3 (operands[0], operands[0], const0_rtx));
+  DONE;
+})
+
 (define_expand "builtin_setjmp_receiver"
   [(label_ref (match_operand 0 "" ""))]
   "!TARGET_64BIT && flag_pic"
                      [(match_dup 0)
                       (match_operand:SI 1 "nonmemory_operand" "")]))
               (clobber (reg:CC FLAGS_REG))])]
-  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY_WRITE"
+  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY_WRITE
+   /* Do not split stack checking probes.  */
+   && GET_CODE (operands[3]) != IOR && operands[1] != const0_rtx"
   [(set (match_dup 2) (match_dup 0))
    (parallel [(set (match_dup 2)
                    (match_op_dup 3 [(match_dup 2) (match_dup 1)]))
                      [(match_operand:SI 1 "nonmemory_operand" "")
                       (match_dup 0)]))
               (clobber (reg:CC FLAGS_REG))])]
-  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY_WRITE"
+  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY_WRITE
+   /* Do not split stack checking probes.  */
+   && GET_CODE (operands[3]) != IOR && operands[1] != const0_rtx"
   [(set (match_dup 2) (match_dup 0))
    (parallel [(set (match_dup 2)
                    (match_op_dup 3 [(match_dup 1) (match_dup 2)]))
   [(set (match_operand 0 "" "")
        (call (mem:QI (match_operand:SI 1 "constant_call_address_operand" ""))
              (match_operand:SI 2 "" "")))
-   (set (reg:SI SP_REG) (plus:SI (reg:SI SP_REG)
-                           (match_operand:SI 3 "immediate_operand" "")))]
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "")))]
   "!TARGET_64BIT"
 {
   if (SIBLING_CALL_P (insn))
 
 (define_insn "*call_value_pop_1"
   [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "rsm"))
+       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "lsm"))
              (match_operand:SI 2 "" "")))
-   (set (reg:SI SP_REG) (plus:SI (reg:SI SP_REG)
-                           (match_operand:SI 3 "immediate_operand" "i")))]
-  "!TARGET_64BIT"
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "i")))]
+  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (constant_call_address_operand (operands[1], Pmode))
-    {
-      if (SIBLING_CALL_P (insn))
-       return "jmp\t%P1";
-      else
-       return "call\t%P1";
-    }
-  if (SIBLING_CALL_P (insn))
-    return "jmp\t%A1";
-  else
-    return "call\t%A1";
+    return "call\t%P1";
+  return "call\t%A1";
 }
   [(set_attr "type" "callv")])
 
+(define_insn "*sibcall_value_pop_1"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "s,U"))
+             (match_operand:SI 2 "" "")))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "i,i")))]
+  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "@
+   jmp\t%P1
+   jmp\t%A1"
+  [(set_attr "type" "callv")])
+
 (define_insn "*call_value_0"
   [(set (match_operand 0 "" "")
        (call (mem:QI (match_operand:SI 1 "constant_call_address_operand" ""))
    (clobber (reg:TI XMM15_REG))
    (clobber (reg:DI SI_REG))
    (clobber (reg:DI DI_REG))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT"
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (SIBLING_CALL_P (insn))
     return "jmp\t%P1";
 
 (define_insn "*call_value_1"
   [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "rsm"))
+       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "lsm"))
              (match_operand:SI 2 "" "")))]
-  "!SIBLING_CALL_P (insn) && !TARGET_64BIT"
+  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (constant_call_address_operand (operands[1], Pmode))
     return "call\t%P1";
 
 (define_insn "*sibcall_value_1"
   [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "s,c,d,a"))
+       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "s,U"))
              (match_operand:SI 2 "" "")))]
-  "SIBLING_CALL_P (insn) && !TARGET_64BIT"
-{
-  if (constant_call_address_operand (operands[1], Pmode))
-    return "jmp\t%P1";
-  return "jmp\t%A1";
-}
+  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "@
+   jmp\t%P1
+   jmp\t%A1"
   [(set_attr "type" "callv")])
 
 (define_insn "*call_value_1_rex64"
   [(set (match_operand 0 "" "")
        (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rsm"))
              (match_operand:DI 2 "" "")))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)
    && ix86_cmodel != CM_LARGE && ix86_cmodel != CM_LARGE_PIC"
 {
   if (constant_call_address_operand (operands[1], Pmode))
        (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rsm"))
              (match_operand:DI 2 "" "")))
    (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
-   (clobber (reg:TI 27))
-   (clobber (reg:TI 28))
-   (clobber (reg:TI 45))
-   (clobber (reg:TI 46))
-   (clobber (reg:TI 47))
-   (clobber (reg:TI 48))
-   (clobber (reg:TI 49))
-   (clobber (reg:TI 50))
-   (clobber (reg:TI 51))
-   (clobber (reg:TI 52))
+   (clobber (reg:TI XMM6_REG))
+   (clobber (reg:TI XMM7_REG))
+   (clobber (reg:TI XMM8_REG))
+   (clobber (reg:TI XMM9_REG))
+   (clobber (reg:TI XMM10_REG))
+   (clobber (reg:TI XMM11_REG))
+   (clobber (reg:TI XMM12_REG))
+   (clobber (reg:TI XMM13_REG))
+   (clobber (reg:TI XMM14_REG))
+   (clobber (reg:TI XMM15_REG))
    (clobber (reg:DI SI_REG))
    (clobber (reg:DI DI_REG))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT"
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
 {
   if (constant_call_address_operand (operands[1], Pmode))
     return "call\t%P1";
   [(set (match_operand 0 "" "")
        (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rm"))
              (match_operand:DI 2 "" "")))]
-  "!SIBLING_CALL_P (insn) && TARGET_64BIT"
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
   "call\t%A1"
   [(set_attr "type" "callv")])
 
 (define_insn "*sibcall_value_1_rex64"
   [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "constant_call_address_operand" ""))
+       (call (mem:QI (match_operand:DI 1 "sibcall_insn_operand" "s,U"))
              (match_operand:DI 2 "" "")))]
-  "SIBLING_CALL_P (insn) && TARGET_64BIT"
-  "jmp\t%P1"
-  [(set_attr "type" "callv")])
-
-(define_insn "*sibcall_value_1_rex64_v"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (reg:DI R11_REG))
-             (match_operand:DI 1 "" "")))]
-  "SIBLING_CALL_P (insn) && TARGET_64BIT"
-  "jmp\t{*%%}r11"
+  "TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "@
+   jmp\t%P1
+   jmp\t%A1"
   [(set_attr "type" "callv")])
 \f
 ;; We used to use "int $5", in honor of #BR which maps to interrupt vector 5.
 
 (define_expand "sse_prologue_save"
   [(parallel [(set (match_operand:BLK 0 "" "")
-                  (unspec:BLK [(reg:DI 21)
-                               (reg:DI 22)
-                               (reg:DI 23)
-                               (reg:DI 24)
-                               (reg:DI 25)
-                               (reg:DI 26)
-                               (reg:DI 27)
-                               (reg:DI 28)] UNSPEC_SSE_PROLOGUE_SAVE))
+                  (unspec:BLK [(reg:DI XMM0_REG)
+                               (reg:DI XMM1_REG)
+                               (reg:DI XMM2_REG)
+                               (reg:DI XMM3_REG)
+                               (reg:DI XMM4_REG)
+                               (reg:DI XMM5_REG)
+                               (reg:DI XMM6_REG)
+                               (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE))
              (use (match_operand:DI 1 "register_operand" ""))
              (use (match_operand:DI 2 "immediate_operand" ""))
              (use (label_ref:DI (match_operand 3 "" "")))])]
 (define_insn "*sse_prologue_save_insn"
   [(set (mem:BLK (plus:DI (match_operand:DI 0 "register_operand" "R")
                          (match_operand:DI 4 "const_int_operand" "n")))
-       (unspec:BLK [(reg:DI 21)
-                    (reg:DI 22)
-                    (reg:DI 23)
-                    (reg:DI 24)
-                    (reg:DI 25)
-                    (reg:DI 26)
-                    (reg:DI 27)
-                    (reg:DI 28)] UNSPEC_SSE_PROLOGUE_SAVE))
+       (unspec:BLK [(reg:DI XMM0_REG)
+                    (reg:DI XMM1_REG)
+                    (reg:DI XMM2_REG)
+                    (reg:DI XMM3_REG)
+                    (reg:DI XMM4_REG)
+                    (reg:DI XMM5_REG)
+                    (reg:DI XMM6_REG)
+                    (reg:DI XMM7_REG)] UNSPEC_SSE_PROLOGUE_SAVE))
    (use (match_operand:DI 1 "register_operand" "r"))
    (use (match_operand:DI 2 "const_int_operand" "i"))
    (use (label_ref:DI (match_operand 3 "" "X")))]
   }
   [(set_attr "type" "multi")])
 
-(define_mode_iterator CRC32MODE [QI HI SI])
-(define_mode_attr crc32modesuffix [(QI "b") (HI "w") (SI "l")])
-(define_mode_attr crc32modeconstraint [(QI "qm") (HI "rm") (SI "rm")])
-
 (define_insn "sse4_2_crc32<mode>"
   [(set (match_operand:SI 0 "register_operand" "=r")
        (unspec:SI
          [(match_operand:SI 1 "register_operand" "0")
-          (match_operand:CRC32MODE 2 "nonimmediate_operand" "<crc32modeconstraint>")]
+          (match_operand:SWI124 2 "nonimmediate_operand" "<r>m")]
          UNSPEC_CRC32))]
-  "TARGET_SSE4_2"
-  "crc32<crc32modesuffix>\t{%2, %0|%0, %2}"
+  "TARGET_SSE4_2 || TARGET_CRC32"
+  "crc32{<imodesuffix>}\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog1")
    (set_attr "prefix_rep" "1")
    (set_attr "prefix_extra" "1")
          [(match_operand:DI 1 "register_operand" "0")
           (match_operand:DI 2 "nonimmediate_operand" "rm")]
          UNSPEC_CRC32))]
-  "TARGET_SSE4_2 && TARGET_64BIT"
-  "crc32q\t{%2, %0|%0, %2}"
+  "TARGET_64BIT && (TARGET_SSE4_2 || TARGET_CRC32)"
+  "crc32{q}\t{%2, %0|%0, %2}"
   [(set_attr "type" "sselog1")
    (set_attr "prefix_rep" "1")
    (set_attr "prefix_extra" "1")
    (set_attr "mode" "DI")])
 
+(define_expand "rdpmc"
+  [(match_operand:DI 0 "register_operand" "")
+   (match_operand:SI 1 "register_operand" "")]
+  ""
+{
+  rtx reg = gen_reg_rtx (DImode);
+  rtx si;
+
+  /* Force operand 1 into ECX.  */
+  rtx ecx = gen_rtx_REG (SImode, CX_REG);
+  emit_insn (gen_rtx_SET (VOIDmode, ecx, operands[1]));
+  si = gen_rtx_UNSPEC_VOLATILE (DImode, gen_rtvec (1, ecx),
+                               UNSPECV_RDPMC);
+
+  if (TARGET_64BIT)
+    {
+      rtvec vec = rtvec_alloc (2);
+      rtx load = gen_rtx_PARALLEL (VOIDmode, vec);
+      rtx upper = gen_reg_rtx (DImode);
+      rtx di = gen_rtx_UNSPEC_VOLATILE (DImode,
+                                       gen_rtvec (1, const0_rtx),
+                                       UNSPECV_RDPMC);
+      RTVEC_ELT (vec, 0) = gen_rtx_SET (VOIDmode, reg, si);
+      RTVEC_ELT (vec, 1) = gen_rtx_SET (VOIDmode, upper, di);
+      emit_insn (load);
+      upper = expand_simple_binop (DImode, ASHIFT, upper, GEN_INT (32),
+                                  NULL, 1, OPTAB_DIRECT);
+      reg = expand_simple_binop (DImode, IOR, reg, upper, reg, 1,
+                                OPTAB_DIRECT);
+    }
+  else
+    emit_insn (gen_rtx_SET (VOIDmode, reg, si));
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], reg));
+  DONE;
+})
+
+(define_insn "*rdpmc"
+  [(set (match_operand:DI 0 "register_operand" "=A")
+       (unspec_volatile:DI [(match_operand:SI 1 "register_operand" "c")]
+                           UNSPECV_RDPMC))]
+  "!TARGET_64BIT"
+  "rdpmc"
+  [(set_attr "type" "other")
+   (set_attr "length" "2")])
+
+(define_insn "*rdpmc_rex64"
+  [(set (match_operand:DI 0 "register_operand" "=a")
+       (unspec_volatile:DI [(match_operand:SI 2 "register_operand" "c")]
+                           UNSPECV_RDPMC))
+  (set (match_operand:DI 1 "register_operand" "=d")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDPMC))]
+  "TARGET_64BIT"
+  "rdpmc"
+  [(set_attr "type" "other")
+   (set_attr "length" "2")])
+
+(define_expand "rdtsc"
+  [(set (match_operand:DI 0 "register_operand" "")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSC))]
+  ""
+{
+  if (TARGET_64BIT)
+    {
+      rtvec vec = rtvec_alloc (2);
+      rtx load = gen_rtx_PARALLEL (VOIDmode, vec);
+      rtx upper = gen_reg_rtx (DImode);
+      rtx lower = gen_reg_rtx (DImode);
+      rtx src = gen_rtx_UNSPEC_VOLATILE (DImode,
+                                        gen_rtvec (1, const0_rtx),
+                                        UNSPECV_RDTSC);
+      RTVEC_ELT (vec, 0) = gen_rtx_SET (VOIDmode, lower, src);
+      RTVEC_ELT (vec, 1) = gen_rtx_SET (VOIDmode, upper, src);
+      emit_insn (load);
+      upper = expand_simple_binop (DImode, ASHIFT, upper, GEN_INT (32),
+                                  NULL, 1, OPTAB_DIRECT);
+      lower = expand_simple_binop (DImode, IOR, lower, upper, lower, 1,
+                                  OPTAB_DIRECT);
+      emit_insn (gen_rtx_SET (VOIDmode, operands[0], lower));
+      DONE;
+    }
+})
+
+(define_insn "*rdtsc"
+  [(set (match_operand:DI 0 "register_operand" "=A")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSC))]
+  "!TARGET_64BIT"
+  "rdtsc"
+  [(set_attr "type" "other")
+   (set_attr "length" "2")])
+
+(define_insn "*rdtsc_rex64"
+  [(set (match_operand:DI 0 "register_operand" "=a")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSC))
+   (set (match_operand:DI 1 "register_operand" "=d")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSC))]
+  "TARGET_64BIT"
+  "rdtsc"
+  [(set_attr "type" "other")
+   (set_attr "length" "2")])
+
+(define_expand "rdtscp"
+  [(match_operand:DI 0 "register_operand" "")
+   (match_operand:SI 1 "memory_operand" "")]
+  ""
+{
+  rtx di = gen_rtx_UNSPEC_VOLATILE (DImode,
+                                   gen_rtvec (1, const0_rtx),
+                                   UNSPECV_RDTSCP);
+  rtx si = gen_rtx_UNSPEC_VOLATILE (SImode,
+                                   gen_rtvec (1, const0_rtx),
+                                   UNSPECV_RDTSCP);
+  rtx reg = gen_reg_rtx (DImode);
+  rtx tmp = gen_reg_rtx (SImode);
+
+  if (TARGET_64BIT)
+    {
+      rtvec vec = rtvec_alloc (3);
+      rtx load = gen_rtx_PARALLEL (VOIDmode, vec);
+      rtx upper = gen_reg_rtx (DImode);
+      RTVEC_ELT (vec, 0) = gen_rtx_SET (VOIDmode, reg, di);
+      RTVEC_ELT (vec, 1) = gen_rtx_SET (VOIDmode, upper, di);
+      RTVEC_ELT (vec, 2) = gen_rtx_SET (VOIDmode, tmp, si);
+      emit_insn (load);
+      upper = expand_simple_binop (DImode, ASHIFT, upper, GEN_INT (32),
+                                  NULL, 1, OPTAB_DIRECT);
+      reg = expand_simple_binop (DImode, IOR, reg, upper, reg, 1,
+                                OPTAB_DIRECT);
+    }
+  else
+    {
+      rtvec vec = rtvec_alloc (2);
+      rtx load = gen_rtx_PARALLEL (VOIDmode, vec);
+      RTVEC_ELT (vec, 0) = gen_rtx_SET (VOIDmode, reg, di);
+      RTVEC_ELT (vec, 1) = gen_rtx_SET (VOIDmode, tmp, si);
+      emit_insn (load);
+    }
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], reg));
+  emit_insn (gen_rtx_SET (VOIDmode, operands[1], tmp));
+  DONE;
+})
+
+(define_insn "*rdtscp"
+  [(set (match_operand:DI 0 "register_operand" "=A")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSCP))
+   (set (match_operand:SI 1 "register_operand" "=c")
+       (unspec_volatile:SI [(const_int 0)] UNSPECV_RDTSCP))]
+  "!TARGET_64BIT"
+  "rdtscp"
+  [(set_attr "type" "other")
+   (set_attr "length" "3")])
+
+(define_insn "*rdtscp_rex64"
+  [(set (match_operand:DI 0 "register_operand" "=a")
+       (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSCP))
+   (set (match_operand:DI 1 "register_operand" "=d")
+        (unspec_volatile:DI [(const_int 0)] UNSPECV_RDTSCP))
+   (set (match_operand:SI 2 "register_operand" "=c")
+       (unspec_volatile:SI [(const_int 0)] UNSPECV_RDTSCP))]
+  "TARGET_64BIT"
+  "rdtscp"
+  [(set_attr "type" "other")
+   (set_attr "length" "3")])
+
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+;;
+;; LWP instructions
+;;
+;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(define_expand "lwp_llwpcb"
+  [(unspec_volatile [(match_operand 0 "register_operand" "r")]
+                   UNSPECV_LLWP_INTRINSIC)]
+  "TARGET_LWP"
+  "")
+
+(define_insn "*lwp_llwpcb<mode>1"
+  [(unspec_volatile [(match_operand:P 0 "register_operand" "r")]
+                   UNSPECV_LLWP_INTRINSIC)]
+  "TARGET_LWP"
+  "llwpcb\t%0"
+  [(set_attr "type" "lwp")
+   (set_attr "mode" "<MODE>")
+   (set_attr "length" "5")])
+
+(define_expand "lwp_slwpcb"
+  [(set (match_operand 0 "register_operand" "=r")
+       (unspec_volatile [(const_int 0)] UNSPECV_SLWP_INTRINSIC))]
+  "TARGET_LWP"
+  {
+    if (TARGET_64BIT)
+      emit_insn (gen_lwp_slwpcbdi (operands[0]));
+    else
+      emit_insn (gen_lwp_slwpcbsi (operands[0]));
+    DONE;
+  })
+
+(define_insn "lwp_slwpcb<mode>"
+  [(set (match_operand:P 0 "register_operand" "=r")
+       (unspec_volatile:P [(const_int 0)] UNSPECV_SLWP_INTRINSIC))]
+  "TARGET_LWP"
+  "slwpcb\t%0"
+  [(set_attr "type" "lwp")
+   (set_attr "mode" "<MODE>")
+   (set_attr "length" "5")])
+
+(define_expand "lwp_lwpval<mode>3"
+  [(unspec_volatile [(match_operand:SWI48 1 "register_operand" "r")
+                    (match_operand:SI 2 "nonimmediate_operand" "rm")
+                    (match_operand:SI 3 "const_int_operand" "i")]
+                   UNSPECV_LWPVAL_INTRINSIC)]
+  "TARGET_LWP"
+  "/* Avoid unused variable warning.  */
+   (void) operand0;")
+
+(define_insn "*lwp_lwpval<mode>3_1"
+  [(unspec_volatile [(match_operand:SWI48 0 "register_operand" "r")
+                    (match_operand:SI 1 "nonimmediate_operand" "rm")
+                    (match_operand:SI 2 "const_int_operand" "i")]
+                   UNSPECV_LWPVAL_INTRINSIC)]
+  "TARGET_LWP"
+  "lwpval\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "lwp")
+   (set_attr "mode" "<MODE>")
+   (set (attr "length")
+        (symbol_ref "ix86_attr_length_address_default (insn) + 9"))])
+
+(define_expand "lwp_lwpins<mode>3"
+  [(set (reg:CCC FLAGS_REG)
+       (unspec_volatile:CCC [(match_operand:SWI48 1 "register_operand" "r")
+                             (match_operand:SI 2 "nonimmediate_operand" "rm")
+                             (match_operand:SI 3 "const_int_operand" "i")]
+                            UNSPECV_LWPINS_INTRINSIC))
+   (set (match_operand:QI 0 "nonimmediate_operand" "=qm")
+       (eq:QI (reg:CCC FLAGS_REG) (const_int 0)))]
+  "TARGET_LWP"
+  "")
+
+(define_insn "*lwp_lwpins<mode>3_1"
+  [(set (reg:CCC FLAGS_REG)
+       (unspec_volatile:CCC [(match_operand:SWI48 0 "register_operand" "r")
+                             (match_operand:SI 1 "nonimmediate_operand" "rm")
+                             (match_operand:SI 2 "const_int_operand" "i")]
+                            UNSPECV_LWPINS_INTRINSIC))]
+  "TARGET_LWP"
+  "lwpins\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "lwp")
+   (set_attr "mode" "<MODE>")
+   (set (attr "length")
+        (symbol_ref "ix86_attr_length_address_default (insn) + 9"))])
+
 (include "mmx.md")
 (include "sse.md")
 (include "sync.md")