OSDN Git Service

2012-07-04 Tristan Gingold <gingold@adacore.com>
[pf3gnuchains/gcc-fork.git] / gcc / config / i386 / i386.md
index 3b9b7f5..eba1b26 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, 2010, 2011
+;; 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
 ;; Free Software Foundation, Inc.
 ;; Mostly by William Schelter.
 ;; x86_64 support added by Jan Hubicka
@@ -53,6 +53,7 @@
 ;; d -- print duplicated register operand for AVX instruction.
 ;; D -- print condition for SSE cmp instruction.
 ;; P -- if PIC, print an @PLT suffix.
+;; p -- print raw symbol name.
 ;; 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
@@ -61,8 +62,6 @@
 ;; ; -- print a semicolon (after prefixes due to bug in older gas).
 ;; @ -- print a segment register of thread base pointer load
 
-;; UNSPEC usage:
-
 (define_c_enum "unspec" [
   ;; Relocation specifiers
   UNSPEC_GOT
@@ -81,8 +80,6 @@
   ;; Prologue support
   UNSPEC_STACK_ALLOC
   UNSPEC_SET_GOT
-  UNSPEC_REG_SAVE
-  UNSPEC_DEF_CFA
   UNSPEC_SET_RIP
   UNSPEC_SET_GOT_OFFSET
   UNSPEC_MEMORY_BLOCKAGE
   UNSPEC_LD_MPIC       ; load_macho_picbase
   UNSPEC_TRUNC_NOOP
   UNSPEC_DIV_ALREADY_SPLIT
+  UNSPEC_MS_TO_SYSV_CALL
   UNSPEC_CALL_NEEDS_VZEROUPPER
+  UNSPEC_PAUSE
 
   ;; For SSE/MMX support:
   UNSPEC_FIX_NOTRUNC
   UNSPEC_MASKMOV
   UNSPEC_MOVMSK
-  UNSPEC_MOVNT
-  UNSPEC_MOVU
   UNSPEC_RCP
   UNSPEC_RSQRT
-  UNSPEC_SFENCE
-  UNSPEC_PFRCP
-  UNSPEC_PFRCPIT1
-  UNSPEC_PFRCPIT2
-  UNSPEC_PFRSQRT
-  UNSPEC_PFRSQIT1
-  UNSPEC_MFENCE
-  UNSPEC_LFENCE
   UNSPEC_PSADBW
-  UNSPEC_LDDQU
-  UNSPEC_MS_TO_SYSV_CALL
 
   ;; Generic math support
   UNSPEC_COPYSIGN
   UNSPEC_SP_TLS_SET
   UNSPEC_SP_TLS_TEST
 
-  ;; SSSE3
-  UNSPEC_PSHUFB
-  UNSPEC_PSIGN
-  UNSPEC_PALIGNR
-
-  ;; For SSE4A support
-  UNSPEC_EXTRQI
-  UNSPEC_EXTRQ
-  UNSPEC_INSERTQI
-  UNSPEC_INSERTQ
-
-  ;; For SSE4.1 support
-  UNSPEC_BLENDV
-  UNSPEC_INSERTPS
-  UNSPEC_DP
-  UNSPEC_MOVNTDQA
-  UNSPEC_MPSADBW
-  UNSPEC_PHMINPOSUW
-  UNSPEC_PTEST
+  ;; For ROUND support
   UNSPEC_ROUND
 
-  ;; For SSE4.2 support
+  ;; For CRC32 support
   UNSPEC_CRC32
-  UNSPEC_PCMPESTR
-  UNSPEC_PCMPISTR
-
-  ;; For FMA4 support
-  UNSPEC_FMADDSUB
-  UNSPEC_XOP_UNSIGNED_CMP
-  UNSPEC_XOP_TRUEFALSE
-  UNSPEC_XOP_PERMUTE
-  UNSPEC_FRCZ
-
-  ;; For AES support
-  UNSPEC_AESENC
-  UNSPEC_AESENCLAST
-  UNSPEC_AESDEC
-  UNSPEC_AESDECLAST
-  UNSPEC_AESIMC
-  UNSPEC_AESKEYGENASSIST
-
-  ;; For PCLMUL support
-  UNSPEC_PCLMUL
-
-  ;; For AVX support
-  UNSPEC_PCMP
-  UNSPEC_VPERMIL
-  UNSPEC_VPERMIL2
-  UNSPEC_VPERMIL2F128
-  UNSPEC_CAST
-  UNSPEC_VTESTP
-  UNSPEC_VCVTPH2PS
-  UNSPEC_VCVTPS2PH
 
   ;; For BMI support
   UNSPEC_BEXTR
 
-  ;; For RDRAND support
-  UNSPEC_RDRAND
+  ;; For BMI2 support
+  UNSPEC_PDEP
+  UNSPEC_PEXT
 ])
 
 (define_c_enum "unspecv" [
   UNSPECV_BLOCKAGE
   UNSPECV_STACK_PROBE
   UNSPECV_PROBE_STACK_RANGE
-  UNSPECV_EMMS
-  UNSPECV_LDMXCSR
-  UNSPECV_STMXCSR
-  UNSPECV_FEMMS
-  UNSPECV_CLFLUSH
   UNSPECV_ALIGN
-  UNSPECV_MONITOR
-  UNSPECV_MWAIT
-  UNSPECV_CMPXCHG
-  UNSPECV_XCHG
-  UNSPECV_LOCK
   UNSPECV_PROLOGUE_USE
+  UNSPECV_SPLIT_STACK_RETURN
   UNSPECV_CLD
   UNSPECV_NOPS
-  UNSPECV_VZEROALL
-  UNSPECV_VZEROUPPER
   UNSPECV_RDTSC
   UNSPECV_RDTSCP
   UNSPECV_RDPMC
   UNSPECV_RDGSBASE
   UNSPECV_WRFSBASE
   UNSPECV_WRGSBASE
-  UNSPECV_SPLIT_STACK_RETURN
+
+  ;; For RDRAND support
+  UNSPECV_RDRAND
 ])
 
 ;; Constants to represent rounding modes in the ROUND instruction
 \f
 ;; Processor type.
 (define_attr "cpu" "none,pentium,pentiumpro,geode,k6,athlon,k8,core2,corei7,
-                   atom,generic64,amdfam10,bdver1,btver1"
+                   atom,generic64,amdfam10,bdver1,bdver2,btver1"
   (const (symbol_ref "ix86_schedule")))
 
 ;; A basic instruction type.  Refinements due to arguments to be
 (define_attr "type"
   "other,multi,
    alu,alu1,negnot,imov,imovx,lea,
-   incdec,ishift,ishift1,rotate,rotate1,imul,idiv,
+   incdec,ishift,ishiftx,ishift1,rotate,rotatex,rotate1,imul,imulx,idiv,
    icmp,test,ibr,setcc,icmov,
    push,pop,call,callv,leave,
    str,bitmanip,
 ;; The (bounding maximum) length of an instruction immediate.
 (define_attr "length_immediate" ""
   (cond [(eq_attr "type" "incdec,setcc,icmov,str,lea,other,multi,idiv,leave,
-                          bitmanip")
+                          bitmanip,imulx")
           (const_int 0)
         (eq_attr "unit" "i387,sse,mmx")
           (const_int 0)
-        (eq_attr "type" "alu,alu1,negnot,imovx,ishift,rotate,ishift1,rotate1,
-                         imul,icmp,push,pop")
-          (symbol_ref "ix86_attr_length_immediate_default(insn,1)")
+        (eq_attr "type" "alu,alu1,negnot,imovx,ishift,ishiftx,ishift1,
+                         rotate,rotatex,rotate1,imul,icmp,push,pop")
+          (symbol_ref "ix86_attr_length_immediate_default (insn, true)")
         (eq_attr "type" "imov,test")
-          (symbol_ref "ix86_attr_length_immediate_default(insn,0)")
+          (symbol_ref "ix86_attr_length_immediate_default (insn, false)")
         (eq_attr "type" "call")
           (if_then_else (match_operand 0 "constant_call_address_operand" "")
             (const_int 4)
 
 ;; Set when REX opcode prefix is used.
 (define_attr "prefix_rex" ""
-  (cond [(eq (symbol_ref "TARGET_64BIT") (const_int 0))
+  (cond [(not (match_test "TARGET_64BIT"))
           (const_int 0)
         (and (eq_attr "mode" "DI")
              (and (eq_attr "type" "!push,pop,call,callv,leave,ibr")
                   (eq_attr "unit" "!mmx")))
           (const_int 1)
         (and (eq_attr "mode" "QI")
-             (ne (symbol_ref "x86_extended_QIreg_mentioned_p (insn)")
-                 (const_int 0)))
+             (match_test "x86_extended_QIreg_mentioned_p (insn)"))
           (const_int 1)
-        (ne (symbol_ref "x86_extended_reg_mentioned_p (insn)")
-            (const_int 0))
+        (match_test "x86_extended_reg_mentioned_p (insn)")
           (const_int 1)
         (and (eq_attr "type" "imovx")
              (match_operand:QI 1 "ext_QIreg_operand" ""))
   (if_then_else (and (eq_attr "prefix_0f" "1")
                     (eq_attr "prefix_extra" "0"))
     (if_then_else (eq_attr "prefix_vex_w" "1")
-      (symbol_ref "ix86_attr_length_vex_default (insn, 1, 1)")
-      (symbol_ref "ix86_attr_length_vex_default (insn, 1, 0)"))
+      (symbol_ref "ix86_attr_length_vex_default (insn, true, true)")
+      (symbol_ref "ix86_attr_length_vex_default (insn, true, false)"))
     (if_then_else (eq_attr "prefix_vex_w" "1")
-      (symbol_ref "ix86_attr_length_vex_default (insn, 0, 1)")
-      (symbol_ref "ix86_attr_length_vex_default (insn, 0, 0)"))))
+      (symbol_ref "ix86_attr_length_vex_default (insn, false, true)")
+      (symbol_ref "ix86_attr_length_vex_default (insn, false, false)"))))
 
 ;; Set when modrm byte is used.
 (define_attr "modrm" ""
         (eq_attr "unit" "i387")
           (const_int 0)
          (and (eq_attr "type" "incdec")
-             (and (eq (symbol_ref "TARGET_64BIT") (const_int 0))
+             (and (not (match_test "TARGET_64BIT"))
                   (ior (match_operand:SI 1 "register_operand" "")
                        (match_operand:HI 1 "register_operand" ""))))
           (const_int 0)
                       (attr "length_address")))
         (ior (eq_attr "prefix" "vex")
              (and (eq_attr "prefix" "maybe_vex")
-                   (ne (symbol_ref "TARGET_AVX") (const_int 0))))
+                  (match_test "TARGET_AVX")))
           (plus (attr "length_vex")
                 (plus (attr "length_immediate")
                       (plus (attr "modrm")
              (and (match_operand 0 "memory_displacement_operand" "")
                   (match_operand 1 "immediate_operand" "")))
           (const_string "true")
-        (and (eq_attr "type" "alu,ishift,rotate,imul,idiv")
+        (and (eq_attr "type" "alu,ishift,ishiftx,rotate,rotatex,imul,idiv")
              (and (match_operand 0 "memory_displacement_operand" "")
                   (match_operand 2 "immediate_operand" "")))
           (const_string "true")
 (define_attr "movu" "0,1" (const_string "0"))
 
 ;; Used to control the "enabled" attribute on a per-instruction basis.
-(define_attr "isa" "base,noavx,avx"
+(define_attr "isa" "base,sse2,sse2_noavx,sse3,sse4,sse4_noavx,noavx,avx,bmi2"
   (const_string "base"))
 
 (define_attr "enabled" ""
-  (cond [(eq_attr "isa" "noavx") (symbol_ref "!TARGET_AVX")
+  (cond [(eq_attr "isa" "sse2") (symbol_ref "TARGET_SSE2")
+        (eq_attr "isa" "sse2_noavx")
+          (symbol_ref "TARGET_SSE2 && !TARGET_AVX")
+        (eq_attr "isa" "sse3") (symbol_ref "TARGET_SSE3")
+        (eq_attr "isa" "sse4") (symbol_ref "TARGET_SSE4_1")
+        (eq_attr "isa" "sse4_noavx")
+          (symbol_ref "TARGET_SSE4_1 && !TARGET_AVX")
         (eq_attr "isa" "avx") (symbol_ref "TARGET_AVX")
+        (eq_attr "isa" "noavx") (symbol_ref "!TARGET_AVX")
+        (eq_attr "isa" "bmi2") (symbol_ref "TARGET_BMI2")
        ]
        (const_int 1)))
 
 (define_code_attr comm [(plus "%") (ss_plus "%") (us_plus "%")
                        (minus "") (ss_minus "") (us_minus "")])
 
+;; Mapping of max and min
+(define_code_iterator maxmin [smax smin umax umin])
+
 ;; Mapping of signed max and min
 (define_code_iterator smaxmin [smax smin])
 
 ;; Base name for insn mnemonic.
 (define_code_attr logic [(and "and") (ior "or") (xor "xor")])
 
+;; Mapping of logic-shift operators
+(define_code_iterator any_lshift [ashift lshiftrt])
+
 ;; 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")])
+(define_code_attr shift_insn
+  [(ashift "ashl") (lshiftrt "lshr") (ashiftrt "ashr")])
 
 ;; Base name for insn mnemonic.
-(define_code_attr shiftrt [(lshiftrt "shr") (ashiftrt "sar")])
+(define_code_attr shift [(ashift "sll") (lshiftrt "shr") (ashiftrt "sar")])
+(define_code_attr vshift [(ashift "sll") (lshiftrt "srl") (ashiftrt "sra")])
 
 ;; Mapping of rotate operators
 (define_code_iterator any_rotate [rotate rotatert])
 ;; 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])
+;; Prefix for insn menmonic.
+(define_code_attr sgnprefix [(sign_extend "i") (zero_extend "")])
 
-;; Instruction prefix for signed and unsigned operations.
-(define_code_attr sgnprefix [(sign_extend "i") (zero_extend "")
-                            (div "i") (udiv "")])
+;; Prefix for define_insn
+(define_code_attr u [(sign_extend "") (zero_extend "u")])
+(define_code_attr s [(sign_extend "s") (zero_extend "u")])
 
-;; 64bit single word integer modes.
+;; All integer modes.
 (define_mode_iterator SWI1248x [QI HI SI DI])
 
-;; 64bit single word integer modes without QImode and HImode.
-(define_mode_iterator SWI48x [SI DI])
+;; All integer modes without QImode.
+(define_mode_iterator SWI248x [HI SI DI])
 
-;; Single word integer modes.
-(define_mode_iterator SWI [QI HI SI (DI "TARGET_64BIT")])
+;; All integer modes without QImode and HImode.
+(define_mode_iterator SWI48x [SI DI])
 
-;; Single word integer modes without SImode and DImode.
+;; All integer modes without SImode and DImode.
 (define_mode_iterator SWI12 [QI HI])
 
-;; Single word integer modes without DImode.
+;; All integer modes without DImode.
 (define_mode_iterator SWI124 [QI HI SI])
 
-;; Single word integer modes without QImode and DImode.
+;; All integer modes without QImode and DImode.
 (define_mode_iterator SWI24 [HI SI])
 
+;; Single word integer modes.
+(define_mode_iterator SWI [QI HI SI (DI "TARGET_64BIT")])
+
 ;; Single word integer modes without QImode.
 (define_mode_iterator SWI248 [HI SI (DI "TARGET_64BIT")])
 
                            (HI "TARGET_HIMODE_MATH")
                            SI (DI "TARGET_64BIT")])
 
-;; Math-dependant single word integer modes without DImode.
+;; Math-dependant integer modes without DImode.
 (define_mode_iterator SWIM124 [(QI "TARGET_QIMODE_MATH")
                               (HI "TARGET_HIMODE_MATH")
                               SI])
 (define_mode_attr r [(QI "q") (HI "r") (SI "r") (DI "r")])
 
 ;; Immediate operand constraint for integer modes.
-(define_mode_attr i [(QI "n") (HI "n") (SI "i") (DI "e")])
+(define_mode_attr i [(QI "n") (HI "n") (SI "e") (DI "e")])
 
 ;; General operand constraint for word modes.
-(define_mode_attr g [(QI "qmn") (HI "rmn") (SI "g") (DI "rme")])
+(define_mode_attr g [(QI "qmn") (HI "rmn") (SI "rme") (DI "rme")])
 
 ;; Immediate operand constraint for double integer modes.
-(define_mode_attr di [(SI "iF") (DI "e")])
+(define_mode_attr di [(SI "nF") (DI "e")])
 
 ;; Immediate operand constraint for shifts.
 (define_mode_attr S [(QI "I") (HI "I") (SI "I") (DI "J") (TI "O")])
 (define_mode_attr general_operand
        [(QI "general_operand")
         (HI "general_operand")
-        (SI "general_operand")
+        (SI "x86_64_general_operand")
         (DI "x86_64_general_operand")
         (TI "x86_64_general_operand")])
 
 (define_mode_attr general_szext_operand
        [(QI "general_operand")
         (HI "general_operand")
-        (SI "general_operand")
+        (SI "x86_64_szext_general_operand")
         (DI "x86_64_szext_general_operand")])
 
 ;; Immediate operand predicate for integer modes.
 (define_mode_attr immediate_operand
        [(QI "immediate_operand")
         (HI "immediate_operand")
-        (SI "immediate_operand")
+        (SI "x86_64_immediate_operand")
         (DI "x86_64_immediate_operand")])
 
 ;; Nonmemory operand predicate for integer modes.
 (define_mode_attr nonmemory_operand
        [(QI "nonmemory_operand")
         (HI "nonmemory_operand")
-        (SI "nonmemory_operand")
+        (SI "x86_64_nonmemory_operand")
         (DI "x86_64_nonmemory_operand")])
 
 ;; Operand predicate for shifts.
 ;; All x87 floating point modes
 (define_mode_iterator X87MODEF [SF DF XF])
 
-;; All integer modes handled by x87 fisttp operator.
-(define_mode_iterator X87MODEI [HI SI DI])
-
-;; All integer modes handled by integer x87 operators.
-(define_mode_iterator X87MODEI12 [HI SI])
-
-;; All integer modes handled by SSE cvtts?2si* operators.
-(define_mode_iterator SSEMODEI24 [SI DI])
-
 ;; SSE instruction suffix for various modes
 (define_mode_attr ssemodesuffix
   [(SF "ss") (DF "sd")
    (V8SF "ps") (V4DF "pd")
    (V4SF "ps") (V2DF "pd")
    (V16QI "b") (V8HI "w") (V4SI "d") (V2DI "q")
-   (V8SI "si")])
+   (V32QI "b") (V16HI "w") (V8SI "d") (V4DI "q")])
 
 ;; SSE vector suffix for floating point modes
 (define_mode_attr ssevecmodesuffix [(SF "ps") (DF "pd")])
 ;; 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")])
+
+;; This mode iterator allows :PTR to be used for patterns that operate on
+;; ptr_mode sized quantities.
+(define_mode_iterator PTR
+  [(SI "ptr_mode == SImode") (DI "ptr_mode == DImode")])
 \f
 ;; Scheduling descriptions
 
        UNSPEC_FNSTSW))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && GET_MODE (operands[1]) == GET_MODE (operands[2])"
-  "* return output_fp_compare (insn, operands, 0, 0);"
+  "* return output_fp_compare (insn, operands, false, false);"
   [(set_attr "type" "multi")
    (set_attr "unit" "i387")
    (set (attr "mode")
             (match_operand:XF 2 "register_operand" "f"))]
          UNSPEC_FNSTSW))]
   "TARGET_80387"
-  "* return output_fp_compare (insn, operands, 0, 0);"
+  "* return output_fp_compare (insn, operands, false, false);"
   [(set_attr "type" "multi")
    (set_attr "unit" "i387")
    (set_attr "mode" "XF")])
             (match_operand:MODEF 2 "nonimmediate_operand" "fm"))]
          UNSPEC_FNSTSW))]
   "TARGET_80387"
-  "* return output_fp_compare (insn, operands, 0, 0);"
+  "* return output_fp_compare (insn, operands, false, false);"
   [(set_attr "type" "multi")
    (set_attr "unit" "i387")
    (set_attr "mode" "<MODE>")])
          UNSPEC_FNSTSW))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && GET_MODE (operands[1]) == GET_MODE (operands[2])"
-  "* return output_fp_compare (insn, operands, 0, 1);"
+  "* return output_fp_compare (insn, operands, false, true);"
   [(set_attr "type" "multi")
    (set_attr "unit" "i387")
    (set (attr "mode")
          [(compare:CCFP
             (match_operand 1 "register_operand" "f")
             (match_operator 3 "float_operator"
-              [(match_operand:X87MODEI12 2 "memory_operand" "m")]))]
+              [(match_operand:SWI24 2 "memory_operand" "m")]))]
          UNSPEC_FNSTSW))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && (TARGET_USE_<MODE>MODE_FIOP || optimize_function_for_size_p (cfun))
    && (GET_MODE (operands [3]) == GET_MODE (operands[1]))"
-  "* return output_fp_compare (insn, operands, 0, 0);"
+  "* return output_fp_compare (insn, operands, false, false);"
   [(set_attr "type" "multi")
    (set_attr "unit" "i387")
    (set_attr "fp_int_src" "true")
        (compare:CCFP
          (match_operand 1 "register_operand" "f")
          (match_operator 3 "float_operator"
-           [(match_operand:X87MODEI12 2 "memory_operand" "m")])))
+           [(match_operand:SWI24 2 "memory_operand" "m")])))
    (clobber (match_operand:HI 0 "register_operand" "=a"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && TARGET_SAHF && !TARGET_CMOVE
        (unspec:HI [(reg:CCFP FPSR_REG)] UNSPEC_FNSTSW))]
   "TARGET_80387"
   "fnstsw\t%0"
-  [(set (attr "length") (symbol_ref "ix86_attr_length_address_default (insn) + 2"))
+  [(set (attr "length")
+       (symbol_ref "ix86_attr_length_address_default (insn) + 2"))
    (set_attr "mode" "SI")
    (set_attr "unit" "i387")])
 
    (set_attr "mode" "SI")])
 
 ;; Pentium Pro can do steps 1 through 3 in one go.
-;; comi*, ucomi*, fcomi*, ficomi*,fucomi* (i387 instructions set condition codes)
+;; comi*, ucomi*, fcomi*, ficomi*, fucomi*
+;; (these i387 instructions set flags directly)
 (define_insn "*cmpfp_i_mixed"
   [(set (reg:CCFP FLAGS_REG)
        (compare:CCFP (match_operand 0 "register_operand" "f,x")
   "TARGET_MIX_SSE_I387
    && SSE_FLOAT_MODE_P (GET_MODE (operands[0]))
    && GET_MODE (operands[0]) == GET_MODE (operands[1])"
-  "* return output_fp_compare (insn, operands, 1, 0);"
+  "* return output_fp_compare (insn, operands, true, false);"
   [(set_attr "type" "fcmp,ssecomi")
    (set_attr "prefix" "orig,maybe_vex")
    (set (attr "mode")
   "TARGET_SSE_MATH
    && SSE_FLOAT_MODE_P (GET_MODE (operands[0]))
    && GET_MODE (operands[0]) == GET_MODE (operands[1])"
-  "* return output_fp_compare (insn, operands, 1, 0);"
+  "* return output_fp_compare (insn, operands, true, false);"
   [(set_attr "type" "ssecomi")
    (set_attr "prefix" "maybe_vex")
    (set (attr "mode")
    && TARGET_CMOVE
    && !(SSE_FLOAT_MODE_P (GET_MODE (operands[0])) && TARGET_SSE_MATH)
    && GET_MODE (operands[0]) == GET_MODE (operands[1])"
-  "* return output_fp_compare (insn, operands, 1, 0);"
+  "* return output_fp_compare (insn, operands, true, false);"
   [(set_attr "type" "fcmp")
    (set (attr "mode")
      (cond [(match_operand:SF 1 "" "")
   "TARGET_MIX_SSE_I387
    && SSE_FLOAT_MODE_P (GET_MODE (operands[0]))
    && GET_MODE (operands[0]) == GET_MODE (operands[1])"
-  "* return output_fp_compare (insn, operands, 1, 1);"
+  "* return output_fp_compare (insn, operands, true, true);"
   [(set_attr "type" "fcmp,ssecomi")
    (set_attr "prefix" "orig,maybe_vex")
    (set (attr "mode")
   "TARGET_SSE_MATH
    && SSE_FLOAT_MODE_P (GET_MODE (operands[0]))
    && GET_MODE (operands[0]) == GET_MODE (operands[1])"
-  "* return output_fp_compare (insn, operands, 1, 1);"
+  "* return output_fp_compare (insn, operands, true, true);"
   [(set_attr "type" "ssecomi")
    (set_attr "prefix" "maybe_vex")
    (set (attr "mode")
    && TARGET_CMOVE
    && !(SSE_FLOAT_MODE_P (GET_MODE (operands[0])) && TARGET_SSE_MATH)
    && GET_MODE (operands[0]) == GET_MODE (operands[1])"
-  "* return output_fp_compare (insn, operands, 1, 1);"
+  "* return output_fp_compare (insn, operands, true, true);"
   [(set_attr "type" "fcmp")
    (set (attr "mode")
      (cond [(match_operand:SF 1 "" "")
 
 (define_insn "*push<mode>2"
   [(set (match_operand:DWI 0 "push_operand" "=<")
-       (match_operand:DWI 1 "general_no_elim_operand" "riF*m"))]
+       (match_operand:DWI 1 "general_no_elim_operand" "riF*o"))]
   ""
-  "#")
+  "#"
+  [(set_attr "type" "multi")
+   (set_attr "mode" "<MODE>")])
 
 (define_split
   [(set (match_operand:TI 0 "push_operand" "")
   switch (which_alternative)
     {
     case 0:
-      return "vxorps\t%0, %0, %0";
+      return standard_sse_constant_opcode (insn, operands[1]);
     case 1:
     case 2:
       if (misaligned_operand (operands[0], OImode)
     case 1:
       return "#";
     case 2:
-      if (get_attr_mode (insn) == MODE_V4SF)
-       return "%vxorps\t%0, %d0";
-      else
-       return "%vpxor\t%0, %d0";
+      return standard_sse_constant_opcode (insn, operands[1]);
     case 3:
     case 4:
       /* TDmode values are passed as TImode on the stack.  Moving them
        {
          if (get_attr_mode (insn) == MODE_V4SF)
            return "%vmovups\t{%1, %0|%0, %1}";
-        else
-          return "%vmovdqu\t{%1, %0|%0, %1}";
+         else
+           return "%vmovdqu\t{%1, %0|%0, %1}";
        }
       else
        {
          if (get_attr_mode (insn) == MODE_V4SF)
            return "%vmovaps\t{%1, %0|%0, %1}";
-        else
-          return "%vmovdqa\t{%1, %0|%0, %1}";
+         else
+           return "%vmovdqa\t{%1, %0|%0, %1}";
        }
     default:
       gcc_unreachable ();
    (set (attr "mode")
        (cond [(eq_attr "alternative" "2,3")
                 (if_then_else
-                  (ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                      (const_int 0))
+                  (match_test "optimize_function_for_size_p (cfun)")
                   (const_string "V4SF")
                   (const_string "TI"))
               (eq_attr "alternative" "4")
                 (if_then_else
-                  (ior (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
-                           (const_int 0))
-                       (ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                           (const_int 0)))
+                  (ior (match_test "TARGET_SSE_TYPELESS_STORES")
+                       (match_test "optimize_function_for_size_p (cfun)"))
                   (const_string "V4SF")
                   (const_string "TI"))]
               (const_string "DI")))])
   switch (which_alternative)
     {
     case 0:
-      if (get_attr_mode (insn) == MODE_V4SF)
-       return "%vxorps\t%0, %d0";
-      else
-       return "%vpxor\t%0, %d0";
+      return standard_sse_constant_opcode (insn, operands[1]);
     case 1:
     case 2:
       /* TDmode values are passed as TImode on the stack.  Moving them
        {
          if (get_attr_mode (insn) == MODE_V4SF)
            return "%vmovups\t{%1, %0|%0, %1}";
-        else
-          return "%vmovdqu\t{%1, %0|%0, %1}";
+         else
+           return "%vmovdqu\t{%1, %0|%0, %1}";
        }
       else
        {
          if (get_attr_mode (insn) == MODE_V4SF)
            return "%vmovaps\t{%1, %0|%0, %1}";
-        else
-          return "%vmovdqa\t{%1, %0|%0, %1}";
+         else
+           return "%vmovdqa\t{%1, %0|%0, %1}";
        }
     default:
       gcc_unreachable ();
   [(set_attr "type" "sselog1,ssemov,ssemov")
    (set_attr "prefix" "maybe_vex")
    (set (attr "mode")
-       (cond [(ior (eq (symbol_ref "TARGET_SSE2") (const_int 0))
-                   (ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                       (const_int 0)))
+       (cond [(ior (not (match_test "TARGET_SSE2"))
+                   (match_test "optimize_function_for_size_p (cfun)"))
                 (const_string "V4SF")
               (and (eq_attr "alternative" "2")
-                   (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
-                       (const_int 0)))
+                   (match_test "TARGET_SSE_TYPELESS_STORES"))
                 (const_string "V4SF")]
              (const_string "TI")))])
 
 (define_insn "*movdi_internal_rex64"
   [(set (match_operand:DI 0 "nonimmediate_operand"
-         "=r,r  ,r,m ,!m,*y,*y,?r ,m ,?*Ym,?*y,*x,*x,?r ,m,?*Yi,*x,?*x,?*Ym")
+         "=r,r  ,r,m ,!o,*y,m*y,?*y,?r ,?*Ym,*x,m ,*x,*x,?r ,?*Yi,?*x,?*Ym")
        (match_operand:DI 1 "general_operand"
-         "Z ,rem,i,re,n ,C ,*y,*Ym,*y,r   ,m  ,C ,*x,*Yi,*x,r  ,m ,*Ym,*x"))]
+         "Z ,rem,i,re,n ,C ,*y ,m  ,*Ym,r   ,C ,*x,*x,m ,*Yi,r   ,*Ym,*x"))]
   "TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
 {
   switch (get_attr_type (insn))
     case TYPE_SSEMOV:
       if (get_attr_mode (insn) == MODE_TI)
        return "%vmovdqa\t{%1, %0|%0, %1}";
-      /* Handle broken assemblers that reqire movd instead of movq.  */
+      /* Handle broken assemblers that require movd instead of movq.  */
       if (GENERAL_REG_P (operands[0]) || GENERAL_REG_P (operands[1]))
        return "%vmovd\t{%1, %0|%0, %1}";
-      return "%vmovq\t{%1, %0|%0, %1}";
+      else
+       return "%vmovq\t{%1, %0|%0, %1}";
 
     case TYPE_MMXMOV:
-      /* Handle broken assemblers that reqire movd instead of movq.  */
+      /* Handle broken assemblers that require movd instead of movq.  */
       if (GENERAL_REG_P (operands[0]) || GENERAL_REG_P (operands[1]))
        return "movd\t{%1, %0|%0, %1}";
-      return "movq\t{%1, %0|%0, %1}";
+      else
+       return "movq\t{%1, %0|%0, %1}";
 
     case TYPE_SSELOG1:
-      return "%vpxor\t%0, %d0";
+      return standard_sse_constant_opcode (insn, operands[1]);
 
     case TYPE_MMX:
       return "pxor\t%0, %0";
        return "mov{l}\t{%k1, %k0|%k0, %k1}";
       else if (which_alternative == 2)
        return "movabs{q}\t{%1, %0|%0, %1}";
+      else if (ix86_use_lea_for_mov (insn, operands))
+       return "lea{q}\t{%a1, %0|%0, %a1}";
       else
        return "mov{q}\t{%1, %0|%0, %1}";
     }
 }
   [(set (attr "type")
-     (cond [(eq_attr "alternative" "5")
+     (cond [(eq_attr "alternative" "4")
+             (const_string "multi")
+           (eq_attr "alternative" "5")
              (const_string "mmx")
-           (eq_attr "alternative" "6,7,8,9,10")
+           (eq_attr "alternative" "6,7,8,9")
              (const_string "mmxmov")
-           (eq_attr "alternative" "11")
+           (eq_attr "alternative" "10")
              (const_string "sselog1")
-           (eq_attr "alternative" "12,13,14,15,16")
+           (eq_attr "alternative" "11,12,13,14,15")
              (const_string "ssemov")
-           (eq_attr "alternative" "17,18")
+           (eq_attr "alternative" "16,17")
              (const_string "ssecvt")
-           (eq_attr "alternative" "4")
-             (const_string "multi")
-           (match_operand:DI 1 "pic_32bit_operand" "")
+           (match_operand 1 "pic_32bit_operand" "")
              (const_string "lea")
           ]
           (const_string "imov")))
        (and (eq_attr "alternative" "2") (eq_attr "type" "imov"))
         (const_string "8")
         (const_string "*")))
-   (set_attr "prefix_rex" "*,*,*,*,*,*,*,1,*,1,*,*,*,*,*,*,*,*,*")
-   (set_attr "prefix_data16" "*,*,*,*,*,*,*,*,*,*,*,*,*,*,*,1,*,*,*")
+   (set (attr "prefix_rex")
+     (if_then_else (eq_attr "alternative" "8,9")
+       (const_string "1")
+       (const_string "*")))
+   (set (attr "prefix_data16")
+     (if_then_else (eq_attr "alternative" "11")
+       (const_string "1")
+       (const_string "*")))
    (set (attr "prefix")
-     (if_then_else (eq_attr "alternative" "11,12,13,14,15,16")
+     (if_then_else (eq_attr "alternative" "10,11,12,13,14,15")
        (const_string "maybe_vex")
        (const_string "orig")))
-   (set_attr "mode" "SI,DI,DI,DI,SI,DI,DI,DI,DI,DI,DI,TI,TI,DI,DI,DI,DI,DI,DI")])
+   (set_attr "mode" "SI,DI,DI,DI,SI,DI,DI,DI,DI,DI,TI,DI,TI,DI,DI,DI,DI,DI")])
+
+;; Reload patterns to support multi-word load/store
+;; with non-offsetable address.
+(define_expand "reload_noff_store"
+  [(parallel [(match_operand 0 "memory_operand" "=m")
+              (match_operand 1 "register_operand" "r")
+              (match_operand:DI 2 "register_operand" "=&r")])]
+  "TARGET_64BIT"
+{
+  rtx mem = operands[0];
+  rtx addr = XEXP (mem, 0);
+
+  emit_move_insn (operands[2], addr);
+  mem = replace_equiv_address_nv (mem, operands[2]);
+
+  emit_insn (gen_rtx_SET (VOIDmode, mem, operands[1]));
+  DONE;
+})
+
+(define_expand "reload_noff_load"
+  [(parallel [(match_operand 0 "register_operand" "=r")
+              (match_operand 1 "memory_operand" "m")
+              (match_operand:DI 2 "register_operand" "=r")])]
+  "TARGET_64BIT"
+{
+  rtx mem = operands[1];
+  rtx addr = XEXP (mem, 0);
+
+  emit_move_insn (operands[2], addr);
+  mem = replace_equiv_address_nv (mem, operands[2]);
+
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], mem));
+  DONE;
+})
 
 ;; Convert impossible stores of immediate to existing instructions.
 ;; First try to get scratch register and go through it.  In case this
 
 (define_insn "*movdi_internal"
   [(set (match_operand:DI 0 "nonimmediate_operand"
-                       "=r  ,o  ,*y,m*y,*y,*Y2,m  ,*Y2,*Y2,*x,m ,*x,*x")
+         "=r  ,o  ,*y,m*y,*y,*x,m ,*x,*x,*x,m ,*x,*x,?*x,?*Ym")
        (match_operand:DI 1 "general_operand"
-                       "riFo,riF,C ,*y ,m ,C  ,*Y2,*Y2,m  ,C ,*x,*x,m "))]
+         "riFo,riF,C ,*y ,m ,C ,*x,*x,m ,C ,*x,*x,m ,*Ym,*x"))]
   "!TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "@
-   #
-   #
-   pxor\t%0, %0
-   movq\t{%1, %0|%0, %1}
-   movq\t{%1, %0|%0, %1}
-   %vpxor\t%0, %d0
-   %vmovq\t{%1, %0|%0, %1}
-   %vmovdqa\t{%1, %0|%0, %1}
-   %vmovq\t{%1, %0|%0, %1}
-   xorps\t%0, %0
-   movlps\t{%1, %0|%0, %1}
-   movaps\t{%1, %0|%0, %1}
-   movlps\t{%1, %0|%0, %1}"
+{
+  switch (get_attr_type (insn))
+    {
+    case TYPE_SSECVT:
+      if (SSE_REG_P (operands[0]))
+       return "movq2dq\t{%1, %0|%0, %1}";
+      else
+       return "movdq2q\t{%1, %0|%0, %1}";
+
+    case TYPE_SSEMOV:
+      switch (get_attr_mode (insn))
+       {
+       case MODE_TI:
+         return "%vmovdqa\t{%1, %0|%0, %1}";
+       case MODE_DI:
+          return "%vmovq\t{%1, %0|%0, %1}";
+       case MODE_V4SF:
+         return "movaps\t{%1, %0|%0, %1}";
+       case MODE_V2SF:
+         return "movlps\t{%1, %0|%0, %1}";
+       default:
+         gcc_unreachable ();
+       }
+
+    case TYPE_MMXMOV:
+      return "movq\t{%1, %0|%0, %1}";
+
+    case TYPE_SSELOG1:
+      return standard_sse_constant_opcode (insn, operands[1]);
+
+    case TYPE_MMX:
+      return "pxor\t%0, %0";
+
+    case TYPE_MULTI:
+      return "#";
+
+    default:
+      gcc_unreachable ();
+    }
+}
   [(set (attr "isa")
-     (if_then_else (eq_attr "alternative" "9,10,11,12")
-       (const_string "noavx")
-       (const_string "base")))
-   (set_attr "type" "*,*,mmx,mmxmov,mmxmov,sselog1,ssemov,ssemov,ssemov,sselog1,ssemov,ssemov,ssemov")
+     (cond [(eq_attr "alternative" "5,6,7,8,13,14")
+             (const_string "sse2")
+           (eq_attr "alternative" "9,10,11,12")
+             (const_string "noavx")
+          ]
+           (const_string "*")))
+   (set (attr "type")
+     (cond [(eq_attr "alternative" "0,1")
+             (const_string "multi")
+           (eq_attr "alternative" "2")
+             (const_string "mmx")
+           (eq_attr "alternative" "3,4")
+             (const_string "mmxmov")
+           (eq_attr "alternative" "5,9")
+             (const_string "sselog1")
+           (eq_attr "alternative" "13,14")
+             (const_string "ssecvt")
+          ]
+          (const_string "ssemov")))
    (set (attr "prefix")
      (if_then_else (eq_attr "alternative" "5,6,7,8")
        (const_string "maybe_vex")
        (const_string "orig")))
-   (set_attr "mode" "DI,DI,DI,DI,DI,TI,DI,TI,DI,V4SF,V2SF,V4SF,V2SF")])
+   (set_attr "mode" "DI,DI,DI,DI,DI,TI,DI,TI,DI,V4SF,V2SF,V4SF,V2SF,DI,DI")])
 
 (define_split
   [(set (match_operand:DI 0 "nonimmediate_operand" "")
   [(set (match_operand:SI 0 "nonimmediate_operand"
                        "=r,m ,*y,*y,?rm,?*y,*x,*x,?r ,m ,?*Yi,*x")
        (match_operand:SI 1 "general_operand"
-                       "g ,ri,C ,*y,*y ,rm ,C ,*x,*Yi,*x,r   ,m "))]
+                       "g ,re,C ,*y,*y ,rm ,C ,*x,*Yi,*x,r   ,m "))]
   "!(MEM_P (operands[0]) && MEM_P (operands[1]))"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_SSELOG1:
-      if (get_attr_mode (insn) == MODE_TI)
-        return "%vpxor\t%0, %d0";
-      return "%vxorps\t%0, %d0";
+      return standard_sse_constant_opcode (insn, operands[1]);
 
     case TYPE_SSEMOV:
       switch (get_attr_mode (insn))
 
     default:
       gcc_assert (!flag_pic || LEGITIMATE_PIC_OPERAND_P (operands[1]));
-      return "mov{l}\t{%1, %0|%0, %1}";
+      if (ix86_use_lea_for_mov (insn, operands))
+       return "lea{l}\t{%a1, %0|%0, %a1}";
+      else
+       return "mov{l}\t{%1, %0|%0, %1}";
     }
 }
   [(set (attr "type")
              (const_string "sselog1")
            (eq_attr "alternative" "7,8,9,10,11")
              (const_string "ssemov")
-           (match_operand:DI 1 "pic_32bit_operand" "")
+           (match_operand 1 "pic_32bit_operand" "")
              (const_string "lea")
           ]
           (const_string "imov")))
              (const_string "DI")
            (eq_attr "alternative" "6,7")
              (if_then_else
-               (eq (symbol_ref "TARGET_SSE2") (const_int 0))
+               (not (match_test "TARGET_SSE2"))
                (const_string "V4SF")
                (const_string "TI"))
            (and (eq_attr "alternative" "8,9,10,11")
-                (eq (symbol_ref "TARGET_SSE2") (const_int 0)))
+                (not (match_test "TARGET_SSE2")))
              (const_string "SF")
           ]
           (const_string "SI")))])
     }
 }
   [(set (attr "type")
-     (cond [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-               (const_int 0))
+     (cond [(match_test "optimize_function_for_size_p (cfun)")
              (const_string "imov")
            (and (eq_attr "alternative" "0")
-                (ior (eq (symbol_ref "TARGET_PARTIAL_REG_STALL")
-                         (const_int 0))
-                     (eq (symbol_ref "TARGET_HIMODE_MATH")
-                         (const_int 0))))
+                (ior (not (match_test "TARGET_PARTIAL_REG_STALL"))
+                     (not (match_test "TARGET_HIMODE_MATH"))))
              (const_string "imov")
            (and (eq_attr "alternative" "1,2")
                 (match_operand:HI 1 "aligned_operand" ""))
              (const_string "imov")
-           (and (ne (symbol_ref "TARGET_MOVX")
-                    (const_int 0))
+           (and (match_test "TARGET_MOVX")
                 (eq_attr "alternative" "0,2"))
              (const_string "imovx")
           ]
                  (match_operand:HI 1 "aligned_operand" ""))
               (const_string "SI")
             (and (eq_attr "alternative" "0")
-                 (ior (eq (symbol_ref "TARGET_PARTIAL_REG_STALL")
-                          (const_int 0))
-                      (eq (symbol_ref "TARGET_HIMODE_MATH")
-                          (const_int 0))))
+                 (ior (not (match_test "TARGET_PARTIAL_REG_STALL"))
+                      (not (match_test "TARGET_HIMODE_MATH"))))
               (const_string "SI")
            ]
            (const_string "HI")))])
      (cond [(and (eq_attr "alternative" "5")
                 (not (match_operand:QI 1 "aligned_operand" "")))
              (const_string "imovx")
-           (ne (symbol_ref "optimize_function_for_size_p (cfun)")
-               (const_int 0))
+           (match_test "optimize_function_for_size_p (cfun)")
              (const_string "imov")
            (and (eq_attr "alternative" "3")
-                (ior (eq (symbol_ref "TARGET_PARTIAL_REG_STALL")
-                         (const_int 0))
-                     (eq (symbol_ref "TARGET_QIMODE_MATH")
-                         (const_int 0))))
+                (ior (not (match_test "TARGET_PARTIAL_REG_STALL"))
+                     (not (match_test "TARGET_QIMODE_MATH"))))
              (const_string "imov")
            (eq_attr "alternative" "3,5")
              (const_string "imovx")
-           (and (ne (symbol_ref "TARGET_MOVX")
-                    (const_int 0))
+           (and (match_test "TARGET_MOVX")
                 (eq_attr "alternative" "2"))
              (const_string "imovx")
           ]
               (const_string "SI")
             (and (eq_attr "type" "imov")
                  (and (eq_attr "alternative" "0,1")
-                      (and (ne (symbol_ref "TARGET_PARTIAL_REG_DEPENDENCY")
-                               (const_int 0))
-                           (and (eq (symbol_ref "optimize_function_for_size_p (cfun)")
-                                    (const_int 0))
-                                (eq (symbol_ref "TARGET_PARTIAL_REG_STALL")
-                                    (const_int 0))))))
+                      (and (match_test "TARGET_PARTIAL_REG_DEPENDENCY")
+                           (and (not (match_test "optimize_function_for_size_p (cfun)"))
+                                (not (match_test "TARGET_PARTIAL_REG_STALL"))))))
               (const_string "SI")
             ;; Avoid partial register stalls when not using QImode arithmetic
             (and (eq_attr "type" "imov")
                  (and (eq_attr "alternative" "0,1")
-                      (and (ne (symbol_ref "TARGET_PARTIAL_REG_STALL")
-                               (const_int 0))
-                           (eq (symbol_ref "TARGET_QIMODE_MATH")
-                               (const_int 0)))))
+                      (and (match_test "TARGET_PARTIAL_REG_STALL")
+                           (not (match_test "TARGET_QIMODE_MATH")))))
               (const_string "SI")
           ]
           (const_string "QI")))])
 ;; into register when rax is not available
 (define_insn "*movabs<mode>_1"
   [(set (mem:SWI1248x (match_operand:DI 0 "x86_64_movabs_operand" "i,r"))
-       (match_operand:SWI1248x 1 "nonmemory_operand" "a,er"))]
-  "TARGET_64BIT && ix86_check_movabs (insn, 0)"
+       (match_operand:SWI1248x 1 "nonmemory_operand" "a,r<i>"))]
+  "TARGET_LP64 && ix86_check_movabs (insn, 0)"
   "@
    movabs{<imodesuffix>}\t{%1, %P0|%P0, %1}
    mov{<imodesuffix>}\t{%1, %a0|%a0, %1}"
 (define_insn "*movabs<mode>_2"
   [(set (match_operand:SWI1248x 0 "register_operand" "=a,r")
         (mem:SWI1248x (match_operand:DI 1 "x86_64_movabs_operand" "i,r")))]
-  "TARGET_64BIT && ix86_check_movabs (insn, 1)"
+  "TARGET_LP64 && ix86_check_movabs (insn, 1)"
   "@
    movabs{<imodesuffix>}\t{%P1, %0|%0, %P1}
    mov{<imodesuffix>}\t{%a1, %0|%0, %a1}"
     }
 }
   [(set (attr "type")
-     (if_then_else (and (match_operand:QI 0 "register_operand" "")
-                       (ior (not (match_operand:QI 0 "q_regs_operand" ""))
-                            (ne (symbol_ref "TARGET_MOVX")
-                                (const_int 0))))
+     (if_then_else (ior (not (match_operand:QI 0 "QIreg_operand" ""))
+                       (match_test "TARGET_MOVX"))
        (const_string "imovx")
        (const_string "imov")))
    (set (attr "mode")
 }
   [(set (attr "type")
      (if_then_else (and (match_operand:QI 0 "register_operand" "")
-                       (ior (not (match_operand:QI 0 "q_regs_operand" ""))
-                            (ne (symbol_ref "TARGET_MOVX")
-                                (const_int 0))))
+                       (ior (not (match_operand:QI 0 "QIreg_operand" ""))
+                            (match_test "TARGET_MOVX")))
        (const_string "imovx")
        (const_string "imov")))
    (set (attr "mode")
     }
 }
   [(set (attr "type")
-     (if_then_else (ior (not (match_operand:QI 0 "q_regs_operand" ""))
-                       (ne (symbol_ref "TARGET_MOVX")
-                           (const_int 0)))
+     (if_then_else (ior (not (match_operand:QI 0 "QIreg_operand" ""))
+                       (match_test "TARGET_MOVX"))
        (const_string "imovx")
        (const_string "imov")))
    (set (attr "mode")
 }
   [(set (attr "type")
      (if_then_else (and (match_operand:QI 0 "register_operand" "")
-                       (ior (not (match_operand:QI 0 "q_regs_operand" ""))
-                            (ne (symbol_ref "TARGET_MOVX")
-                                (const_int 0))))
+                       (ior (not (match_operand:QI 0 "QIreg_operand" ""))
+                            (match_test "TARGET_MOVX")))
        (const_string "imovx")
        (const_string "imov")))
    (set (attr "mode")
    (set_attr "unit" "sse,*,*")
    (set_attr "mode" "TF,SI,SI")])
 
+;; %%% Kill this when call knows how to work this out.
 (define_split
   [(set (match_operand:TF 0 "push_operand" "")
        (match_operand:TF 1 "sse_reg_operand" ""))]
   [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -16)))
    (set (mem:TF (reg:P SP_REG)) (match_dup 1))])
 
-(define_split
-  [(set (match_operand:TF 0 "push_operand" "")
-       (match_operand:TF 1 "general_operand" ""))]
-  "TARGET_SSE2 && reload_completed
-   && !SSE_REG_P (operands[1])"
-  [(const_int 0)]
-  "ix86_split_long_move (operands); DONE;")
-
 (define_insn "*pushxf"
   [(set (match_operand:XF 0 "push_operand" "=<,<")
        (match_operand:XF 1 "general_no_elim_operand" "f,ro"))]
 ;; only once, but this ought to be handled elsewhere).
 
 (define_insn "*pushxf_nointeger"
-  [(set (match_operand:XF 0 "push_operand" "=X,X,X")
-       (match_operand:XF 1 "general_no_elim_operand" "f,Fo,*r"))]
+  [(set (match_operand:XF 0 "push_operand" "=<,<")
+       (match_operand:XF 1 "general_no_elim_operand" "f,*rFo"))]
   "optimize_function_for_size_p (cfun)"
 {
   /* This insn should be already split before reg-stack.  */
   gcc_unreachable ();
 }
   [(set_attr "type" "multi")
-   (set_attr "unit" "i387,*,*")
-   (set_attr "mode" "XF,SI,SI")])
+   (set_attr "unit" "i387,*")
+   (set_attr "mode" "XF,SI")])
 
+;; %%% Kill this when call knows how to work this out.
 (define_split
   [(set (match_operand:XF 0 "push_operand" "")
        (match_operand:XF 1 "fp_register_operand" ""))]
    (set (mem:XF (reg:P SP_REG)) (match_dup 1))]
   "operands[2] = GEN_INT (-GET_MODE_SIZE (XFmode));")
 
-(define_split
-  [(set (match_operand:XF 0 "push_operand" "")
-       (match_operand:XF 1 "general_operand" ""))]
-  "reload_completed
-   && !FP_REG_P (operands[1])"
-  [(const_int 0)]
-  "ix86_split_long_move (operands); DONE;")
-
-(define_insn "*pushdf"
+(define_insn "*pushdf_rex64"
   [(set (match_operand:DF 0 "push_operand" "=<,<,<")
-       (match_operand:DF 1 "general_no_elim_operand" "f,rFo,Y2"))]
-  "TARGET_64BIT || TARGET_INTEGER_DFMODE_MOVES"
+       (match_operand:DF 1 "general_no_elim_operand" "f,Yd*rFm,x"))]
+  "TARGET_64BIT"
 {
   /* This insn should be already split before reg-stack.  */
   gcc_unreachable ();
 }
   [(set_attr "type" "multi")
    (set_attr "unit" "i387,*,*")
-   (set_attr "mode" "DF,SI,DF")])
+   (set_attr "mode" "DF,DI,DF")])
 
 ;; Size of pushdf is 3 (for sub) + 2 (for fstp) + memory operand size.
 ;; Size of pushdf using integer instructions is 2+2*memory operand size
-;; On the average, pushdf using integers can be still shorter.  Allow this
-;; pattern for optimize_size too.
+;; On the average, pushdf using integers can be still shorter.
 
-(define_insn "*pushdf_nointeger"
-  [(set (match_operand:DF 0 "push_operand" "=<,<,<,<")
-       (match_operand:DF 1 "general_no_elim_operand" "f,Fo,*r,Y2"))]
-  "!(TARGET_64BIT || TARGET_INTEGER_DFMODE_MOVES)"
+(define_insn "*pushdf"
+  [(set (match_operand:DF 0 "push_operand" "=<,<,<")
+       (match_operand:DF 1 "general_no_elim_operand" "f,Yd*rFo,x"))]
+  "!TARGET_64BIT"
 {
   /* This insn should be already split before reg-stack.  */
   gcc_unreachable ();
 }
-  [(set_attr "type" "multi")
-   (set_attr "unit" "i387,*,*,*")
-   (set_attr "mode" "DF,SI,SI,DF")])
+  [(set_attr "isa" "*,*,sse2")
+   (set_attr "type" "multi")
+   (set_attr "unit" "i387,*,*")
+   (set_attr "mode" "DF,DI,DF")])
 
 ;; %%% Kill this when call knows how to work this out.
 (define_split
   [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (const_int -8)))
    (set (mem:DF (reg:P SP_REG)) (match_dup 1))])
 
-(define_split
-  [(set (match_operand:DF 0 "push_operand" "")
-       (match_operand:DF 1 "general_operand" ""))]
-  "reload_completed
-   && !ANY_FP_REG_P (operands[1])"
-  [(const_int 0)]
-  "ix86_split_long_move (operands); DONE;")
-
 (define_insn "*pushsf_rex64"
   [(set (match_operand:SF 0 "push_operand" "=X,X,X")
        (match_operand:SF 1 "nonmemory_no_elim_operand" "f,rF,x"))]
    (set_attr "unit" "i387,*,*")
    (set_attr "mode" "SF,SI,SF")])
 
-(define_split
-  [(set (match_operand:SF 0 "push_operand" "")
-       (match_operand:SF 1 "memory_operand" ""))]
-  "reload_completed
-   && MEM_P (operands[1])
-   && (operands[2] = find_constant_src (insn))"
-  [(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" "")
   "reload_completed"
   [(set (reg:P SP_REG) (plus:P (reg:P SP_REG) (match_dup 2)))
    (set (mem:SF (reg:P SP_REG)) (match_dup 1))]
-  "operands[2] = GEN_INT (-GET_MODE_SIZE (<MODE>mode));")
+  "operands[2] = GEN_INT (-GET_MODE_SIZE (<P:MODE>mode));")
+
+(define_split
+  [(set (match_operand:SF 0 "push_operand" "")
+       (match_operand:SF 1 "memory_operand" ""))]
+  "reload_completed
+   && (operands[2] = find_constant_src (insn))"
+  [(set (match_dup 0) (match_dup 2))])
+
+(define_split
+  [(set (match_operand 0 "push_operand" "")
+       (match_operand 1 "general_operand" ""))]
+  "reload_completed
+   && (GET_MODE (operands[0]) == TFmode
+       || GET_MODE (operands[0]) == XFmode
+       || GET_MODE (operands[0]) == DFmode)
+   && !ANY_FP_REG_P (operands[1])"
+  [(const_int 0)]
+  "ix86_split_long_move (operands); DONE;")
 \f
 ;; Floating point move instructions.
 
   "ix86_expand_move (<MODE>mode, operands); DONE;")
 
 (define_insn "*movtf_internal"
-  [(set (match_operand:TF 0 "nonimmediate_operand" "=x,m,x,?r,?o")
-       (match_operand:TF 1 "general_operand" "xm,x,C,roF,Fr"))]
+  [(set (match_operand:TF 0 "nonimmediate_operand" "=x,m,x,?*r ,!o")
+       (match_operand:TF 1 "general_operand"      "xm,x,C,*roF,F*r"))]
   "TARGET_SSE2
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
+   && !(MEM_P (operands[0]) && MEM_P (operands[1]))
+   && (!can_create_pseudo_p ()
+       || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
+       || GET_CODE (operands[1]) != CONST_DOUBLE
+       || (optimize_function_for_size_p (cfun)
+          && standard_sse_constant_p (operands[1])
+          && !memory_operand (operands[0], TFmode))
+       || (!TARGET_MEMORY_MISMATCH_STALL
+          && memory_operand (operands[0], TFmode)))"
 {
   switch (which_alternative)
     {
     case 0:
     case 1:
-      if (get_attr_mode (insn) == MODE_V4SF)
-       return "%vmovaps\t{%1, %0|%0, %1}";
+      /* Handle misaligned load/store since we
+         don't have movmisaligntf pattern. */
+      if (misaligned_operand (operands[0], TFmode)
+         || misaligned_operand (operands[1], TFmode))
+       {
+         if (get_attr_mode (insn) == MODE_V4SF)
+           return "%vmovups\t{%1, %0|%0, %1}";
+         else
+           return "%vmovdqu\t{%1, %0|%0, %1}";
+       }
       else
-       return "%vmovdqa\t{%1, %0|%0, %1}";
+       {
+         if (get_attr_mode (insn) == MODE_V4SF)
+           return "%vmovaps\t{%1, %0|%0, %1}";
+         else
+           return "%vmovdqa\t{%1, %0|%0, %1}";
+       }
+
     case 2:
-      if (get_attr_mode (insn) == MODE_V4SF)
-       return "%vxorps\t%0, %d0";
-      else
-       return "%vpxor\t%0, %d0";
+      return standard_sse_constant_opcode (insn, operands[1]);
+
     case 3:
     case 4:
        return "#";
+
     default:
       gcc_unreachable ();
     }
    (set (attr "mode")
         (cond [(eq_attr "alternative" "0,2")
                 (if_then_else
-                  (ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                      (const_int 0))
+                  (match_test "optimize_function_for_size_p (cfun)")
                   (const_string "V4SF")
                   (const_string "TI"))
               (eq_attr "alternative" "1")
                 (if_then_else
-                  (ior (ne (symbol_ref "TARGET_SSE_TYPELESS_STORES")
-                           (const_int 0))
-                       (ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                           (const_int 0)))
+                  (ior (match_test "TARGET_SSE_TYPELESS_STORES")
+                       (match_test "optimize_function_for_size_p (cfun)"))
                   (const_string "V4SF")
                   (const_string "TI"))]
               (const_string "DI")))])
 
-(define_split
-  [(set (match_operand:TF 0 "nonimmediate_operand" "")
-        (match_operand:TF 1 "general_operand" ""))]
-  "reload_completed
-   && !(SSE_REG_P (operands[0]) || SSE_REG_P (operands[1]))"
-  [(const_int 0)]
-  "ix86_split_long_move (operands); DONE;")
-
+;; Possible store forwarding (partial memory) stall in alternative 4.
 (define_insn "*movxf_internal"
-  [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,r,o")
-       (match_operand:XF 1 "general_operand" "fm,f,G,roF,Fr"))]
-  "optimize_function_for_speed_p (cfun)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && (reload_in_progress || reload_completed
+  [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,?Yx*r ,!o")
+       (match_operand:XF 1 "general_operand"      "fm,f,G,Yx*roF,FYx*r"))]
+  "!(MEM_P (operands[0]) && MEM_P (operands[1]))
+   && (!can_create_pseudo_p ()
+       || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
        || GET_CODE (operands[1]) != CONST_DOUBLE
-       || memory_operand (operands[0], XFmode))"
+       || (optimize_function_for_size_p (cfun)
+          && standard_80387_constant_p (operands[1]) > 0
+          && !memory_operand (operands[0], XFmode))
+       || (!TARGET_MEMORY_MISMATCH_STALL
+          && memory_operand (operands[0], XFmode)))"
 {
   switch (which_alternative)
     {
     case 2:
       return standard_80387_constant_opcode (operands[1]);
 
-    case 3: case 4:
+    case 3:
+    case 4:
       return "#";
 
     default:
   [(set_attr "type" "fmov,fmov,fmov,multi,multi")
    (set_attr "mode" "XF,XF,XF,SI,SI")])
 
-;; Do not use integer registers when optimizing for size
-(define_insn "*movxf_internal_nointeger"
-  [(set (match_operand:XF 0 "nonimmediate_operand" "=f,m,f,*r,o")
-       (match_operand:XF 1 "general_operand" "fm,f,G,*roF,F*r"))]
-  "optimize_function_for_size_p (cfun)
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && (reload_in_progress || reload_completed
-       || standard_80387_constant_p (operands[1])
-       || GET_CODE (operands[1]) != CONST_DOUBLE
-       || memory_operand (operands[0], XFmode))"
-{
-  switch (which_alternative)
-    {
-    case 0:
-    case 1:
-      return output_387_reg_move (insn, operands);
-
-    case 2:
-      return standard_80387_constant_opcode (operands[1]);
-
-    case 3: case 4:
-      return "#";
-    default:
-      gcc_unreachable ();
-    }
-}
-  [(set_attr "type" "fmov,fmov,fmov,multi,multi")
-   (set_attr "mode" "XF,XF,XF,SI,SI")])
-
-(define_split
-  [(set (match_operand:XF 0 "nonimmediate_operand" "")
-       (match_operand:XF 1 "general_operand" ""))]
-  "reload_completed
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && ! (FP_REG_P (operands[0]) ||
-        (GET_CODE (operands[0]) == SUBREG
-         && FP_REG_P (SUBREG_REG (operands[0]))))
-   && ! (FP_REG_P (operands[1]) ||
-        (GET_CODE (operands[1]) == SUBREG
-         && FP_REG_P (SUBREG_REG (operands[1]))))"
-  [(const_int 0)]
-  "ix86_split_long_move (operands); DONE;")
-
 (define_insn "*movdf_internal_rex64"
   [(set (match_operand:DF 0 "nonimmediate_operand"
-               "=f,m,f,r ,m,!r,!m,Y2*x,Y2*x,Y2*x,m   ,Yi,r ")
+               "=f,m,f,?r,?m,?r,!o,x,x,x,m,Yi,r ")
        (match_operand:DF 1 "general_operand"
-               "fm,f,G,rm,r,F ,F ,C   ,Y2*x,m   ,Y2*x,r ,Yi"))]
+               "fm,f,G,rm,r ,F ,F ,C,x,m,x,r ,Yi"))]
   "TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && (reload_in_progress || reload_completed
+   && (!can_create_pseudo_p ()
        || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
-       || (!(TARGET_SSE2 && TARGET_SSE_MATH)
-           && optimize_function_for_size_p (cfun)
-          && standard_80387_constant_p (operands[1]))
        || GET_CODE (operands[1]) != CONST_DOUBLE
+       || (optimize_function_for_size_p (cfun)
+          && ((!(TARGET_SSE2 && TARGET_SSE_MATH)
+               && standard_80387_constant_p (operands[1]) > 0)
+              || (TARGET_SSE2 && TARGET_SSE_MATH
+                  && standard_sse_constant_p (operands[1]))))
        || memory_operand (operands[0], DFmode))"
 {
   switch (which_alternative)
       return "#";
 
     case 7:
-      switch (get_attr_mode (insn))
-       {
-       case MODE_V4SF:
-         return "%vxorps\t%0, %d0";
-       case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vxorpd\t%0, %d0";
-       case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vpxor\t%0, %d0";
-       default:
-         gcc_unreachable ();
-       }
+      return standard_sse_constant_opcode (insn, operands[1]);
+
     case 8:
     case 9:
     case 10:
       switch (get_attr_mode (insn))
        {
-       case MODE_V4SF:
-         return "%vmovaps\t{%1, %0|%0, %1}";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
+         if (!TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
            return "%vmovapd\t{%1, %0|%0, %1}";
-       case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovdqa\t{%1, %0|%0, %1}";
+       case MODE_V4SF:
+         return "%vmovaps\t{%1, %0|%0, %1}";
+
        case MODE_DI:
          return "%vmovq\t{%1, %0|%0, %1}";
        case MODE_DF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]) && REG_P (operands[1]))
-               return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovsd\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movsd\t{%1, %0|%0, %1}";
+         if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
+           return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
+         return "%vmovsd\t{%1, %0|%0, %1}";
        case MODE_V1DF:
          return "%vmovlpd\t{%1, %d0|%d0, %1}";
        case MODE_V2SF:
 
     case 11:
     case 12:
-      /* Handle broken assemblers that reqire movd instead of movq.  */
+      /* Handle broken assemblers that require movd instead of movq.  */
       return "%vmovd\t{%1, %0|%0, %1}";
 
     default:
       gcc_unreachable();
     }
 }
-  [(set_attr "type" "fmov,fmov,fmov,imov,imov,imov,multi,sselog1,ssemov,ssemov,ssemov,ssemov,ssemov")
+  [(set (attr "type")
+       (cond [(eq_attr "alternative" "0,1,2")
+                (const_string "fmov")
+              (eq_attr "alternative" "3,4,5")
+                (const_string "imov")
+              (eq_attr "alternative" "6")
+                (const_string "multi")
+              (eq_attr "alternative" "7")
+                (const_string "sselog1")
+             ]
+             (const_string "ssemov")))
    (set (attr "modrm")
      (if_then_else
        (and (eq_attr "alternative" "5") (eq_attr "type" "imov"))
               (eq_attr "alternative" "3,4,5,6,11,12")
                 (const_string "DI")
 
-              /* For SSE1, we have many fewer alternatives.  */
-              (eq (symbol_ref "TARGET_SSE2") (const_int 0))
-                (cond [(eq_attr "alternative" "7,8")
-                         (const_string "V4SF")
-                      ]
-                  (const_string "V2SF"))
-
               /* xorps is one byte shorter.  */
               (eq_attr "alternative" "7")
-                (cond [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                           (const_int 0))
+                (cond [(match_test "optimize_function_for_size_p (cfun)")
                          (const_string "V4SF")
-                       (ne (symbol_ref "TARGET_SSE_LOAD0_BY_PXOR")
-                           (const_int 0))
+                       (match_test "TARGET_SSE_LOAD0_BY_PXOR")
                          (const_string "TI")
                       ]
                       (const_string "V2DF"))
                  movaps encodes one byte shorter.  */
               (eq_attr "alternative" "8")
                 (cond
-                  [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                       (const_int 0))
+                  [(match_test "optimize_function_for_size_p (cfun)")
                      (const_string "V4SF")
-                   (ne (symbol_ref "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
-                       (const_int 0))
+                   (match_test "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
                      (const_string "V2DF")
                   ]
                   (const_string "DF"))
                  of register.  */
               (eq_attr "alternative" "9")
                 (if_then_else
-                  (ne (symbol_ref "TARGET_SSE_SPLIT_REGS")
-                      (const_int 0))
+                  (match_test "TARGET_SSE_SPLIT_REGS")
                   (const_string "V1DF")
                   (const_string "DF"))
              ]
              (const_string "DF")))])
 
+;; Possible store forwarding (partial memory) stall in alternative 4.
 (define_insn "*movdf_internal"
   [(set (match_operand:DF 0 "nonimmediate_operand"
-               "=f,m,f,r  ,o ,Y2*x,Y2*x,Y2*x,m   ")
+               "=f,m,f,?Yd*r ,!o   ,x,x,x,m,*x,*x,*x,m")
        (match_operand:DF 1 "general_operand"
-               "fm,f,G,roF,Fr,C   ,Y2*x,m   ,Y2*x"))]
+               "fm,f,G,Yd*roF,FYd*r,C,x,m,x,C ,*x,m ,*x"))]
   "!TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && optimize_function_for_speed_p (cfun)
-   && TARGET_INTEGER_DFMODE_MOVES
-   && (reload_in_progress || reload_completed
+   && (!can_create_pseudo_p ()
        || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
-       || (!(TARGET_SSE2 && TARGET_SSE_MATH)
-           && optimize_function_for_size_p (cfun)
-          && standard_80387_constant_p (operands[1]))
        || GET_CODE (operands[1]) != CONST_DOUBLE
-       || memory_operand (operands[0], DFmode))"
+       || (optimize_function_for_size_p (cfun)
+          && ((!(TARGET_SSE2 && TARGET_SSE_MATH)
+               && standard_80387_constant_p (operands[1]) > 0)
+              || (TARGET_SSE2 && TARGET_SSE_MATH
+                  && standard_sse_constant_p (operands[1])))
+          && !memory_operand (operands[0], DFmode))
+       || (!TARGET_MEMORY_MISMATCH_STALL
+          && memory_operand (operands[0], DFmode)))"
 {
   switch (which_alternative)
     {
       return "#";
 
     case 5:
-      switch (get_attr_mode (insn))
-       {
-       case MODE_V4SF:
-         return "%vxorps\t%0, %d0";
-       case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vxorpd\t%0, %d0";
-       case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vpxor\t%0, %d0";
-       default:
-         gcc_unreachable ();
-       }
+    case 9:
+      return standard_sse_constant_opcode (insn, operands[1]);
+
     case 6:
     case 7:
     case 8:
+    case 10:
+    case 11:
+    case 12:
       switch (get_attr_mode (insn))
        {
-       case MODE_V4SF:
-         return "%vmovaps\t{%1, %0|%0, %1}";
        case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
+         if (!TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
            return "%vmovapd\t{%1, %0|%0, %1}";
-       case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovdqa\t{%1, %0|%0, %1}";
+       case MODE_V4SF:
+         return "%vmovaps\t{%1, %0|%0, %1}";
+
        case MODE_DI:
          return "%vmovq\t{%1, %0|%0, %1}";
        case MODE_DF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]) && REG_P (operands[1]))
-               return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovsd\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movsd\t{%1, %0|%0, %1}";
+         if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
+           return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
+         return "%vmovsd\t{%1, %0|%0, %1}";
        case MODE_V1DF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]))
-               return "vmovlpd\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovlpd\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movlpd\t{%1, %0|%0, %1}";
+         return "%vmovlpd\t{%1, %d0|%d0, %1}";
        case MODE_V2SF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]))
-               return "vmovlps\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovlps\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movlps\t{%1, %0|%0, %1}";
+         return "%vmovlps\t{%1, %d0|%d0, %1}";
        default:
          gcc_unreachable ();
        }
       gcc_unreachable ();
     }
 }
-  [(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov")
-   (set (attr "prefix")
-     (if_then_else (eq_attr "alternative" "0,1,2,3,4")
-       (const_string "orig")
-       (const_string "maybe_vex")))
-   (set (attr "prefix_data16")
-     (if_then_else (eq_attr "mode" "V1DF")
-       (const_string "1")
+  [(set (attr "isa")
+     (if_then_else (eq_attr "alternative" "5,6,7,8")
+       (const_string "sse2")
        (const_string "*")))
-   (set (attr "mode")
-        (cond [(eq_attr "alternative" "0,1,2")
-                (const_string "DF")
+   (set (attr "type")
+       (cond [(eq_attr "alternative" "0,1,2")
+                (const_string "fmov")
               (eq_attr "alternative" "3,4")
-                (const_string "SI")
-
-              /* For SSE1, we have many fewer alternatives.  */
-              (eq (symbol_ref "TARGET_SSE2") (const_int 0))
-                (cond [(eq_attr "alternative" "5,6")
-                         (const_string "V4SF")
-                      ]
-                  (const_string "V2SF"))
-
-              /* xorps is one byte shorter.  */
-              (eq_attr "alternative" "5")
-                (cond [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                           (const_int 0))
-                         (const_string "V4SF")
-                       (ne (symbol_ref "TARGET_SSE_LOAD0_BY_PXOR")
-                           (const_int 0))
-                         (const_string "TI")
-                      ]
-                      (const_string "V2DF"))
-
-              /* For architectures resolving dependencies on
-                 whole SSE registers use APD move to break dependency
-                 chains, otherwise use short move to avoid extra work.
-
-                 movaps encodes one byte shorter.  */
-              (eq_attr "alternative" "6")
-                (cond
-                  [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                       (const_int 0))
-                     (const_string "V4SF")
-                   (ne (symbol_ref "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
-                       (const_int 0))
-                     (const_string "V2DF")
-                  ]
-                  (const_string "DF"))
-              /* For architectures resolving dependencies on register
-                 parts we may avoid extra work to zero out upper part
-                 of register.  */
-              (eq_attr "alternative" "7")
-                (if_then_else
-                  (ne (symbol_ref "TARGET_SSE_SPLIT_REGS")
-                      (const_int 0))
-                  (const_string "V1DF")
-                  (const_string "DF"))
+                (const_string "multi")
+              (eq_attr "alternative" "5,9")
+                (const_string "sselog1")
              ]
-             (const_string "DF")))])
-
-;; Moving is usually shorter when only FP registers are used. This separate
-;; movdf pattern avoids the use of integer registers for FP operations
-;; when optimizing for size.
-
-(define_insn "*movdf_internal_nointeger"
-  [(set (match_operand:DF 0 "nonimmediate_operand"
-                       "=f,m,f,*r  ,o  ,Y2*x,Y2*x,Y2*x ,m  ")
-       (match_operand:DF 1 "general_operand"
-                       "fm,f,G,*roF,*Fr,C   ,Y2*x,mY2*x,Y2*x"))]
-  "!TARGET_64BIT && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && (optimize_function_for_size_p (cfun)
-       || !TARGET_INTEGER_DFMODE_MOVES)
-   && (reload_in_progress || reload_completed
-       || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
-       || (!(TARGET_SSE2 && TARGET_SSE_MATH)
-           && optimize_function_for_size_p (cfun)
-           && !memory_operand (operands[0], DFmode)
-          && standard_80387_constant_p (operands[1]))
-       || GET_CODE (operands[1]) != CONST_DOUBLE
-       || ((optimize_function_for_size_p (cfun)
-            || !TARGET_MEMORY_MISMATCH_STALL
-           || reload_in_progress || reload_completed)
-          && memory_operand (operands[0], DFmode)))"
-{
-  switch (which_alternative)
-    {
-    case 0:
-    case 1:
-      return output_387_reg_move (insn, operands);
-
-    case 2:
-      return standard_80387_constant_opcode (operands[1]);
-
-    case 3:
-    case 4:
-      return "#";
-
-    case 5:
-      switch (get_attr_mode (insn))
-       {
-       case MODE_V4SF:
-         return "%vxorps\t%0, %d0";
-       case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vxorpd\t%0, %d0";
-       case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vxorps\t%0, %d0";
-         else
-           return "%vpxor\t%0, %d0";
-       default:
-         gcc_unreachable ();
-       }
-    case 6:
-    case 7:
-    case 8:
-      switch (get_attr_mode (insn))
-       {
-       case MODE_V4SF:
-         return "%vmovaps\t{%1, %0|%0, %1}";
-       case MODE_V2DF:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovapd\t{%1, %0|%0, %1}";
-       case MODE_TI:
-         if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-           return "%vmovaps\t{%1, %0|%0, %1}";
-         else
-           return "%vmovdqa\t{%1, %0|%0, %1}";
-       case MODE_DI:
-         return "%vmovq\t{%1, %0|%0, %1}";
-       case MODE_DF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]) && REG_P (operands[1]))
-               return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovsd\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movsd\t{%1, %0|%0, %1}";
-       case MODE_V1DF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]))
-               return "vmovlpd\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovlpd\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movlpd\t{%1, %0|%0, %1}";
-       case MODE_V2SF:
-         if (TARGET_AVX)
-           {
-             if (REG_P (operands[0]))
-               return "vmovlps\t{%1, %0, %0|%0, %0, %1}";
-             else
-               return "vmovlps\t{%1, %0|%0, %1}";
-           }
-         else
-           return "movlps\t{%1, %0|%0, %1}";
-       default:
-         gcc_unreachable ();
-       }
-
-    default:
-      gcc_unreachable ();
-    }
-}
-  [(set_attr "type" "fmov,fmov,fmov,multi,multi,sselog1,ssemov,ssemov,ssemov")
+             (const_string "ssemov")))
    (set (attr "prefix")
      (if_then_else (eq_attr "alternative" "0,1,2,3,4")
        (const_string "orig")
                 (const_string "SI")
 
               /* For SSE1, we have many fewer alternatives.  */
-              (eq (symbol_ref "TARGET_SSE2") (const_int 0))
-                (cond [(eq_attr "alternative" "5,6")
-                         (const_string "V4SF")
-                      ]
+              (not (match_test "TARGET_SSE2"))
+                (if_then_else
+                  (eq_attr "alternative" "5,6,9,10")
+                  (const_string "V4SF")
                   (const_string "V2SF"))
 
               /* xorps is one byte shorter.  */
-              (eq_attr "alternative" "5")
-                (cond [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                           (const_int 0))
+              (eq_attr "alternative" "5,9")
+                (cond [(match_test "optimize_function_for_size_p (cfun)")
                          (const_string "V4SF")
-                       (ne (symbol_ref "TARGET_SSE_LOAD0_BY_PXOR")
-                           (const_int 0))
+                       (match_test "TARGET_SSE_LOAD0_BY_PXOR")
                          (const_string "TI")
                       ]
                       (const_string "V2DF"))
                  chains, otherwise use short move to avoid extra work.
 
                  movaps encodes one byte shorter.  */
-              (eq_attr "alternative" "6")
+              (eq_attr "alternative" "6,10")
                 (cond
-                  [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
-                       (const_int 0))
+                  [(match_test "optimize_function_for_size_p (cfun)")
                      (const_string "V4SF")
-                   (ne (symbol_ref "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
-                       (const_int 0))
+                   (match_test "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
                      (const_string "V2DF")
                   ]
                   (const_string "DF"))
               /* For architectures resolving dependencies on register
                  parts we may avoid extra work to zero out upper part
                  of register.  */
-              (eq_attr "alternative" "7")
+              (eq_attr "alternative" "7,11")
                 (if_then_else
-                  (ne (symbol_ref "TARGET_SSE_SPLIT_REGS")
-                      (const_int 0))
+                  (match_test "TARGET_SSE_SPLIT_REGS")
                   (const_string "V1DF")
                   (const_string "DF"))
              ]
              (const_string "DF")))])
 
-(define_split
-  [(set (match_operand:DF 0 "nonimmediate_operand" "")
-       (match_operand:DF 1 "general_operand" ""))]
-  "reload_completed
-   && !(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && ! (ANY_FP_REG_P (operands[0]) ||
-        (GET_CODE (operands[0]) == SUBREG
-         && ANY_FP_REG_P (SUBREG_REG (operands[0]))))
-   && ! (ANY_FP_REG_P (operands[1]) ||
-        (GET_CODE (operands[1]) == SUBREG
-         && ANY_FP_REG_P (SUBREG_REG (operands[1]))))"
-  [(const_int 0)]
-  "ix86_split_long_move (operands); DONE;")
-
 (define_insn "*movsf_internal"
   [(set (match_operand:SF 0 "nonimmediate_operand"
-         "=f,m,f,r  ,m ,x,x,x ,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
+         "=f,m,f,?r ,?m,x,x,x,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
        (match_operand:SF 1 "general_operand"
-         "fm,f,G,rmF,Fr,C,x,xm,x,m  ,*y,*y ,r  ,Yi,r   ,*Ym"))]
+         "fm,f,G,rmF,Fr,C,x,m,x,m  ,*y,*y ,r  ,Yi,r   ,*Ym"))]
   "!(MEM_P (operands[0]) && MEM_P (operands[1]))
-   && (reload_in_progress || reload_completed
+   && (!can_create_pseudo_p ()
        || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
-       || (!TARGET_SSE_MATH && optimize_function_for_size_p (cfun)
-          && standard_80387_constant_p (operands[1]))
        || GET_CODE (operands[1]) != CONST_DOUBLE
+       || (optimize_function_for_size_p (cfun)
+          && ((!TARGET_SSE_MATH
+               && standard_80387_constant_p (operands[1]) > 0)
+              || (TARGET_SSE_MATH
+                  && standard_sse_constant_p (operands[1]))))
        || memory_operand (operands[0], SFmode))"
 {
   switch (which_alternative)
     case 3:
     case 4:
       return "mov{l}\t{%1, %0|%0, %1}";
+
     case 5:
-      if (get_attr_mode (insn) == MODE_TI)
-       return "%vpxor\t%0, %d0";
-      else
-       return "%vxorps\t%0, %d0";
+      return standard_sse_constant_opcode (insn, operands[1]);
+
     case 6:
       if (get_attr_mode (insn) == MODE_V4SF)
        return "%vmovaps\t{%1, %0|%0, %1}";
-      else
-       return "%vmovss\t{%1, %d0|%d0, %1}";
-    case 7:
       if (TARGET_AVX)
-        {
-         if (REG_P (operands[1]))
-           return "vmovss\t{%1, %0, %0|%0, %0, %1}";
-         else
-           return "vmovss\t{%1, %0|%0, %1}";
-       }
-      else
-       return "movss\t{%1, %0|%0, %1}";
+       return "vmovss\t{%1, %0, %0|%0, %0, %1}";
+
+    case 7:
     case 8:
       return "%vmovss\t{%1, %0|%0, %1}";
 
-    case 9: case 10: case 14: case 15:
+    case 9:
+    case 10:
+    case 14:
+    case 15:
       return "movd\t{%1, %0|%0, %1}";
 
     case 11:
       return "movq\t{%1, %0|%0, %1}";
 
-    case 12: case 13:
+    case 12:
+    case 13:
       return "%vmovd\t{%1, %0|%0, %1}";
 
     default:
       gcc_unreachable ();
     }
 }
-  [(set_attr "type" "fmov,fmov,fmov,imov,imov,sselog1,ssemov,ssemov,ssemov,mmxmov,mmxmov,mmxmov,ssemov,ssemov,mmxmov,mmxmov")
+  [(set (attr "type")
+       (cond [(eq_attr "alternative" "0,1,2")
+                (const_string "fmov")
+              (eq_attr "alternative" "3,4")
+                (const_string "multi")
+              (eq_attr "alternative" "5")
+                (const_string "sselog1")
+              (eq_attr "alternative" "9,10,11,14,15")
+                (const_string "mmxmov")
+             ]
+             (const_string "ssemov")))
    (set (attr "prefix")
      (if_then_else (eq_attr "alternative" "5,6,7,8,12,13")
        (const_string "maybe_vex")
                 (const_string "SI")
               (eq_attr "alternative" "5")
                 (if_then_else
-                  (and (and (ne (symbol_ref "TARGET_SSE_LOAD0_BY_PXOR")
-                                (const_int 0))
-                            (ne (symbol_ref "TARGET_SSE2")
-                                (const_int 0)))
-                       (eq (symbol_ref "optimize_function_for_size_p (cfun)")
-                           (const_int 0)))
+                  (and (and (match_test "TARGET_SSE_LOAD0_BY_PXOR")
+                            (match_test "TARGET_SSE2"))
+                       (not (match_test "optimize_function_for_size_p (cfun)")))
                   (const_string "TI")
                   (const_string "V4SF"))
               /* For architectures resolving dependencies on
                  to avoid problems on using packed logical operations.  */
               (eq_attr "alternative" "6")
                 (if_then_else
-                  (ior (ne (symbol_ref "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
-                           (const_int 0))
-                       (ne (symbol_ref "TARGET_SSE_SPLIT_REGS")
-                           (const_int 0)))
+                  (ior (match_test "TARGET_SSE_PARTIAL_REG_DEPENDENCY")
+                       (match_test "TARGET_SSE_SPLIT_REGS"))
                   (const_string "V4SF")
                   (const_string "SF"))
               (eq_attr "alternative" "11")
               (const_string "SF")))])
 
 (define_split
-  [(set (match_operand 0 "register_operand" "")
+  [(set (match_operand 0 "any_fp_register_operand" "")
        (match_operand 1 "memory_operand" ""))]
   "reload_completed
-   && MEM_P (operands[1])
    && (GET_MODE (operands[0]) == TFmode
        || GET_MODE (operands[0]) == XFmode
        || GET_MODE (operands[0]) == DFmode
   [(set (match_dup 0) (match_dup 2))]
 {
   rtx c = operands[2];
-  rtx r = operands[0];
+  int r = REGNO (operands[0]);
 
-  if (GET_CODE (r) == SUBREG)
-    r = SUBREG_REG (r);
-
-  if (SSE_REG_P (r))
-    {
-      if (!standard_sse_constant_p (c))
-       FAIL;
-    }
-  else if (FP_REG_P (r))
-    {
-      if (!standard_80387_constant_p (c))
-       FAIL;
-    }
-  else if (MMX_REG_P (r))
+  if ((SSE_REGNO_P (r) && !standard_sse_constant_p (c))
+      || (FP_REGNO_P (r) && standard_80387_constant_p (c) < 1))
     FAIL;
 })
 
 (define_split
-  [(set (match_operand 0 "register_operand" "")
+  [(set (match_operand 0 "any_fp_register_operand" "")
        (float_extend (match_operand 1 "memory_operand" "")))]
   "reload_completed
-   && MEM_P (operands[1])
    && (GET_MODE (operands[0]) == TFmode
        || GET_MODE (operands[0]) == XFmode
-       || GET_MODE (operands[0]) == DFmode
-       || GET_MODE (operands[0]) == SFmode)
+       || GET_MODE (operands[0]) == DFmode)
    && (operands[2] = find_constant_src (insn))"
   [(set (match_dup 0) (match_dup 2))]
 {
   rtx c = operands[2];
-  rtx r = operands[0];
+  int r = REGNO (operands[0]);
 
-  if (GET_CODE (r) == SUBREG)
-    r = SUBREG_REG (r);
-
-  if (SSE_REG_P (r))
-    {
-      if (!standard_sse_constant_p (c))
-       FAIL;
-    }
-  else if (FP_REG_P (r))
-    {
-      if (!standard_80387_constant_p (c))
-       FAIL;
-    }
-  else if (MMX_REG_P (r))
+  if ((SSE_REGNO_P (r) && !standard_sse_constant_p (c))
+      || (FP_REGNO_P (r) && standard_80387_constant_p (c) < 1))
     FAIL;
 })
 
 ;; Split the load of -0.0 or -1.0 into fldz;fchs or fld1;fchs sequence
 (define_split
-  [(set (match_operand:X87MODEF 0 "register_operand" "")
+  [(set (match_operand:X87MODEF 0 "fp_register_operand" "")
        (match_operand:X87MODEF 1 "immediate_operand" ""))]
-  "reload_completed && FP_REGNO_P (REGNO (operands[0]))
+  "reload_completed
    && (standard_80387_constant_p (operands[1]) == 8
        || standard_80387_constant_p (operands[1]) == 9)"
   [(set (match_dup 0)(match_dup 1))
     operands[1] = CONST1_RTX (<MODE>mode);
 })
 
+(define_split
+  [(set (match_operand 0 "nonimmediate_operand" "")
+        (match_operand 1 "general_operand" ""))]
+  "reload_completed
+   && (GET_MODE (operands[0]) == TFmode
+       || GET_MODE (operands[0]) == XFmode
+       || GET_MODE (operands[0]) == DFmode)
+   && !(ANY_FP_REG_P (operands[0]) || ANY_FP_REG_P (operands[1]))"
+  [(const_int 0)]
+  "ix86_split_long_move (operands); DONE;")
+
 (define_insn "swapxf"
   [(set (match_operand:XF 0 "register_operand" "+f")
        (match_operand:XF 1 "register_operand" "+f"))
 })
 
 (define_insn "*zero_extendsidi2_rex64"
-  [(set (match_operand:DI 0 "nonimmediate_operand"  "=r,o,?*Ym,?*y,?*Yi,*Y2")
+  [(set (match_operand:DI 0 "nonimmediate_operand"  "=r,o,?*Ym,?*y,?*Yi,*x")
        (zero_extend:DI
         (match_operand:SI 1 "nonimmediate_operand" "rm,0,r   ,m  ,r   ,m")))]
   "TARGET_64BIT"
   "@
-   mov\t{%k1, %k0|%k0, %k1}
+   mov{l}\t{%1, %k0|%k0, %1}
    #
    movd\t{%1, %0|%0, %1}
    movd\t{%1, %0|%0, %1}
 
 ;; %%% Kill me once multi-word ops are sane.
 (define_insn "zero_extendsidi2_1"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=r,?r,?o,?*Ym,?*y,?*Yi,*Y2")
+  [(set (match_operand:DI 0 "nonimmediate_operand" "=r,?r,?o,?*Ym,?*y,?*Yi,*x")
        (zero_extend:DI
         (match_operand:SI 1 "nonimmediate_operand" "0,rm,r ,r   ,m  ,r   ,m")))
    (clobber (reg:CC FLAGS_REG))]
    movd\t{%1, %0|%0, %1}
    %vmovd\t{%1, %0|%0, %1}
    %vmovd\t{%1, %0|%0, %1}"
-  [(set_attr "type" "multi,multi,multi,mmxmov,mmxmov,ssemov,ssemov")
+  [(set_attr "isa" "*,*,*,*,*,*,sse2")
+   (set_attr "type" "multi,multi,multi,mmxmov,mmxmov,ssemov,ssemov")
    (set_attr "prefix" "*,*,*,orig,orig,maybe_vex,maybe_vex")
    (set_attr "mode" "SI,SI,SI,DI,DI,TI,TI")])
 
    (set_attr "mode" "SF")])
 
 (define_insn "*truncdfsf_mixed"
-  [(set (match_operand:SF 0 "nonimmediate_operand"   "=m,Y2 ,?f,?x,?*r")
+  [(set (match_operand:SF 0 "nonimmediate_operand"   "=m,x ,?f,?x,?*r")
        (float_truncate:SF
-         (match_operand:DF 1 "nonimmediate_operand" "f ,Y2m,f ,f ,f")))
-   (clobber (match_operand:SF 2 "memory_operand"     "=X,X  ,m ,m ,m"))]
+         (match_operand:DF 1 "nonimmediate_operand" "f ,xm,f ,f ,f")))
+   (clobber (match_operand:SF 2 "memory_operand"     "=X,X ,m ,m ,m"))]
   "TARGET_MIX_SSE_I387"
 {
   switch (which_alternative)
       return "#";
     }
 }
-  [(set_attr "type" "fmov,ssecvt,multi,multi,multi")
+  [(set_attr "isa" "*,sse2,*,*,*")
+   (set_attr "type" "fmov,ssecvt,multi,multi,multi")
    (set_attr "unit" "*,*,i387,i387,i387")
    (set_attr "prefix" "orig,maybe_vex,orig,orig,orig")
    (set_attr "mode" "SF")])
    (set_attr "mode" "SF")])
 
 (define_insn "*truncxfdf2_mixed"
-  [(set (match_operand:DF 0 "nonimmediate_operand" "=m,?f,?Y2,?*r")
+  [(set (match_operand:DF 0 "nonimmediate_operand" "=m,?f,?x,?*r")
        (float_truncate:DF
          (match_operand:XF 1 "register_operand"   "f ,f ,f  ,f")))
    (clobber (match_operand:DF 2 "memory_operand"   "=X,m ,m  ,m"))]
   gcc_assert (!which_alternative);
   return output_387_reg_move (insn, operands);
 }
-  [(set_attr "type" "fmov,multi,multi,multi")
+  [(set_attr "isa" "*,*,sse2,*")
+   (set_attr "type" "fmov,multi,multi,multi")
    (set_attr "unit" "*,i387,i387,i387")
    (set_attr "mode" "DF")])
 
 (define_peephole2
   [(set (match_operand:MODEF 0 "register_operand" "")
        (match_operand:MODEF 1 "memory_operand" ""))
-   (set (match_operand:SSEMODEI24 2 "register_operand" "")
-       (fix:SSEMODEI24 (match_dup 0)))]
+   (set (match_operand:SWI48x 2 "register_operand" "")
+       (fix:SWI48x (match_dup 0)))]
   "TARGET_SHORTEN_X87_SSE
    && !(TARGET_AVOID_VECTOR_DECODE && optimize_insn_for_speed_p ())
    && peep2_reg_dead_p (2, operands[0])"
-  [(set (match_dup 2) (fix:SSEMODEI24 (match_dup 1)))])
+  [(set (match_dup 2) (fix:SWI48x (match_dup 1)))])
 
 ;; Avoid vector decoded forms of the instruction.
 (define_peephole2
-  [(match_scratch:DF 2 "Y2")
-   (set (match_operand:SSEMODEI24 0 "register_operand" "")
-       (fix:SSEMODEI24 (match_operand:DF 1 "memory_operand" "")))]
-  "TARGET_AVOID_VECTOR_DECODE && optimize_insn_for_speed_p ()"
+  [(match_scratch:DF 2 "x")
+   (set (match_operand:SWI48x 0 "register_operand" "")
+       (fix:SWI48x (match_operand:DF 1 "memory_operand" "")))]
+  "TARGET_SSE2 && TARGET_AVOID_VECTOR_DECODE && optimize_insn_for_speed_p ()"
   [(set (match_dup 2) (match_dup 1))
-   (set (match_dup 0) (fix:SSEMODEI24 (match_dup 2)))])
+   (set (match_dup 0) (fix:SWI48x (match_dup 2)))])
 
 (define_peephole2
   [(match_scratch:SF 2 "x")
-   (set (match_operand:SSEMODEI24 0 "register_operand" "")
-       (fix:SSEMODEI24 (match_operand:SF 1 "memory_operand" "")))]
+   (set (match_operand:SWI48x 0 "register_operand" "")
+       (fix:SWI48x (match_operand:SF 1 "memory_operand" "")))]
   "TARGET_AVOID_VECTOR_DECODE && optimize_insn_for_speed_p ()"
   [(set (match_dup 2) (match_dup 1))
-   (set (match_dup 0) (fix:SSEMODEI24 (match_dup 2)))])
+   (set (match_dup 0) (fix:SWI48x (match_dup 2)))])
 
 (define_insn_and_split "fix_trunc<mode>_fisttp_i387_1"
-  [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-       (fix:X87MODEI (match_operand 1 "register_operand" "")))]
+  [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+       (fix:SWI248x (match_operand 1 "register_operand" "")))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && TARGET_FISTTP
    && !((SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fix_trunc<mode>_i387_fisttp"
-  [(set (match_operand:X87MODEI 0 "memory_operand" "=m")
-       (fix:X87MODEI (match_operand 1 "register_operand" "f")))
+  [(set (match_operand:SWI248x 0 "memory_operand" "=m")
+       (fix:SWI248x (match_operand 1 "register_operand" "f")))
    (clobber (match_scratch:XF 2 "=&1f"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && TARGET_FISTTP
    && !((SSE_FLOAT_MODE_P (GET_MODE (operands[1]))
         && (TARGET_64BIT || <MODE>mode != DImode))
        && TARGET_SSE_MATH)"
-  "* return output_fix_trunc (insn, operands, 1);"
+  "* return output_fix_trunc (insn, operands, true);"
   [(set_attr "type" "fisttp")
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fix_trunc<mode>_i387_fisttp_with_temp"
-  [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "=m,?r")
-       (fix:X87MODEI (match_operand 1 "register_operand" "f,f")))
-   (clobber (match_operand:X87MODEI 2 "memory_operand" "=X,m"))
+  [(set (match_operand:SWI248x 0 "nonimmediate_operand" "=m,?r")
+       (fix:SWI248x (match_operand 1 "register_operand" "f,f")))
+   (clobber (match_operand:SWI248x 2 "memory_operand" "=X,m"))
    (clobber (match_scratch:XF 3 "=&1f,&1f"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && TARGET_FISTTP
    (set_attr "mode" "<MODE>")])
 
 (define_split
-  [(set (match_operand:X87MODEI 0 "register_operand" "")
-       (fix:X87MODEI (match_operand 1 "register_operand" "")))
-   (clobber (match_operand:X87MODEI 2 "memory_operand" ""))
+  [(set (match_operand:SWI248x 0 "register_operand" "")
+       (fix:SWI248x (match_operand 1 "register_operand" "")))
+   (clobber (match_operand:SWI248x 2 "memory_operand" ""))
    (clobber (match_scratch 3 ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 2) (fix:X87MODEI (match_dup 1)))
+  [(parallel [(set (match_dup 2) (fix:SWI248x (match_dup 1)))
              (clobber (match_dup 3))])
    (set (match_dup 0) (match_dup 2))])
 
 (define_split
-  [(set (match_operand:X87MODEI 0 "memory_operand" "")
-       (fix:X87MODEI (match_operand 1 "register_operand" "")))
-   (clobber (match_operand:X87MODEI 2 "memory_operand" ""))
+  [(set (match_operand:SWI248x 0 "memory_operand" "")
+       (fix:SWI248x (match_operand 1 "register_operand" "")))
+   (clobber (match_operand:SWI248x 2 "memory_operand" ""))
    (clobber (match_scratch 3 ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 0) (fix:X87MODEI (match_dup 1)))
+  [(parallel [(set (match_dup 0) (fix:SWI248x (match_dup 1)))
              (clobber (match_dup 3))])])
 
 ;; See the comments in i386.h near OPTIMIZE_MODE_SWITCHING for the description
 ;; clobbering insns can be used. Look at emit_i387_cw_initialization ()
 ;; function in i386.c.
 (define_insn_and_split "*fix_trunc<mode>_i387_1"
-  [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-       (fix:X87MODEI (match_operand 1 "register_operand" "")))
+  [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+       (fix:SWI248x (match_operand 1 "register_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && !TARGET_FISTTP
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && !TARGET_FISTTP
    && !(TARGET_64BIT && SSE_FLOAT_MODE_P (GET_MODE (operands[1])))"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fistp")
    (set_attr "i387_cw" "trunc")
    (set_attr "mode" "DI")])
              (clobber (match_dup 5))])])
 
 (define_insn "fix_trunc<mode>_i387"
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "=m")
-       (fix:X87MODEI12 (match_operand 1 "register_operand" "f")))
+  [(set (match_operand:SWI24 0 "memory_operand" "=m")
+       (fix:SWI24 (match_operand 1 "register_operand" "f")))
    (use (match_operand:HI 2 "memory_operand" "m"))
    (use (match_operand:HI 3 "memory_operand" "m"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && !TARGET_FISTTP
    && !SSE_FLOAT_MODE_P (GET_MODE (operands[1]))"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fistp")
    (set_attr "i387_cw" "trunc")
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fix_trunc<mode>_i387_with_temp"
-  [(set (match_operand:X87MODEI12 0 "nonimmediate_operand" "=m,?r")
-       (fix:X87MODEI12 (match_operand 1 "register_operand" "f,f")))
+  [(set (match_operand:SWI24 0 "nonimmediate_operand" "=m,?r")
+       (fix:SWI24 (match_operand 1 "register_operand" "f,f")))
    (use (match_operand:HI 2 "memory_operand" "m,m"))
    (use (match_operand:HI 3 "memory_operand" "m,m"))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" "=X,m"))]
+   (clobber (match_operand:SWI24 4 "memory_operand" "=X,m"))]
   "X87_FLOAT_MODE_P (GET_MODE (operands[1]))
    && !TARGET_FISTTP
    && !SSE_FLOAT_MODE_P (GET_MODE (operands[1]))"
    (set_attr "mode" "<MODE>")])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "register_operand" "")
-       (fix:X87MODEI12 (match_operand 1 "register_operand" "")))
+  [(set (match_operand:SWI24 0 "register_operand" "")
+       (fix:SWI24 (match_operand 1 "register_operand" "")))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" ""))]
+   (clobber (match_operand:SWI24 4 "memory_operand" ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 4) (fix:X87MODEI12 (match_dup 1)))
+  [(parallel [(set (match_dup 4) (fix:SWI24 (match_dup 1)))
              (use (match_dup 2))
              (use (match_dup 3))])
    (set (match_dup 0) (match_dup 4))])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "")
-       (fix:X87MODEI12 (match_operand 1 "register_operand" "")))
+  [(set (match_operand:SWI24 0 "memory_operand" "")
+       (fix:SWI24 (match_operand 1 "register_operand" "")))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" ""))]
+   (clobber (match_operand:SWI24 4 "memory_operand" ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 0) (fix:X87MODEI12 (match_dup 1)))
+  [(parallel [(set (match_dup 0) (fix:SWI24 (match_dup 1)))
              (use (match_dup 2))
              (use (match_dup 3))])])
 
 (define_insn "*floathi<mode>2_i387_with_temp"
   [(set (match_operand:X87MODEF 0 "register_operand" "=f,f")
        (float:X87MODEF (match_operand:HI 1 "nonimmediate_operand" "m,?r")))
-  (clobber (match_operand:HI 2 "memory_operand" "=m,m"))]
+  (clobber (match_operand:HI 2 "memory_operand" "=X,m"))]
   "TARGET_80387
    && (!(SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)
        || TARGET_MIX_SSE_I387)"
     && reload_completed"
   [(set (match_dup 0) (float:X87MODEF (match_dup 1)))])
 
-(define_expand "float<SSEMODEI24:mode><X87MODEF:mode>2"
+(define_expand "float<SWI48x:mode><X87MODEF:mode>2"
   [(set (match_operand:X87MODEF 0 "register_operand" "")
        (float:X87MODEF
-         (match_operand:SSEMODEI24 1 "nonimmediate_operand" "")))]
+         (match_operand:SWI48x 1 "nonimmediate_operand" "")))]
   "TARGET_80387
-   || ((<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+   || ((<SWI48x:MODE>mode != DImode || TARGET_64BIT)
        && SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH)"
 {
-  if (!((<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+  if (!((<SWI48x:MODE>mode != DImode || TARGET_64BIT)
        && SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH)
-      && !X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode))
+      && !X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SWI48x:MODE>mode))
     {
       rtx reg = gen_reg_rtx (XFmode);
       rtx (*insn)(rtx, rtx);
 
-      emit_insn (gen_float<SSEMODEI24:mode>xf2 (reg, operands[1]));
+      emit_insn (gen_float<SWI48x:mode>xf2 (reg, operands[1]));
 
       if (<X87MODEF:MODE>mode == SFmode)
        insn = gen_truncxfsf2;
 })
 
 ;; Pre-reload splitter to add memory clobber to the pattern.
-(define_insn_and_split "*float<SSEMODEI24:mode><X87MODEF:mode>2_1"
+(define_insn_and_split "*float<SWI48x:mode><X87MODEF:mode>2_1"
   [(set (match_operand:X87MODEF 0 "register_operand" "")
-       (float:X87MODEF (match_operand:SSEMODEI24 1 "register_operand" "")))]
+       (float:X87MODEF (match_operand:SWI48x 1 "register_operand" "")))]
   "((TARGET_80387
-     && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode)
-     && (!((<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+     && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SWI48x:MODE>mode)
+     && (!((<SWI48x:MODE>mode != DImode || TARGET_64BIT)
           && SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH)
         || TARGET_MIX_SSE_I387))
-    || ((<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+    || ((<SWI48x:MODE>mode != DImode || TARGET_64BIT)
        && SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH
-       && ((<SSEMODEI24:MODE>mode == SImode
+       && ((<SWI48x:MODE>mode == SImode
             && TARGET_SSE2 && TARGET_USE_VECTOR_CONVERTS
             && optimize_function_for_speed_p (cfun)
             && flag_trapping_math)
   [(parallel [(set (match_dup 0) (float:X87MODEF (match_dup 1)))
              (clobber (match_dup 2))])]
 {
-  operands[2] = assign_386_stack_local (<SSEMODEI24:MODE>mode, SLOT_TEMP);
+  operands[2] = assign_386_stack_local (<SWI48x:MODE>mode, SLOT_TEMP);
 
   /* Avoid store forwarding (partial memory) stall penalty
      by passing DImode value through XMM registers.  */
-  if (<SSEMODEI24:MODE>mode == DImode && !TARGET_64BIT
+  if (<SWI48x:MODE>mode == DImode && !TARGET_64BIT
       && TARGET_80387 && TARGET_SSE2 && TARGET_INTER_UNIT_MOVES
       && optimize_function_for_speed_p (cfun))
     {
    (set_attr "bdver1_decode" "*,direct")
    (set_attr "fp_int_src" "true")])
 
-(define_insn "*float<SSEMODEI24:mode><MODEF:mode>2_mixed_with_temp"
+(define_insn "*float<SWI48x:mode><MODEF:mode>2_mixed_with_temp"
   [(set (match_operand:MODEF 0 "register_operand" "=f,f,x,x")
        (float:MODEF
-         (match_operand:SSEMODEI24 1 "nonimmediate_operand" "m,?r,r,m")))
-  (clobber (match_operand:SSEMODEI24 2 "memory_operand" "=X,m,m,X"))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+         (match_operand:SWI48x 1 "nonimmediate_operand" "m,?r,r,m")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" "=X,m,m,X"))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_MIX_SSE_I387"
   "#"
   [(set_attr "type" "fmov,multi,sseicvt,sseicvt")
 
 (define_split
   [(set (match_operand:MODEF 0 "register_operand" "")
-       (float:MODEF (match_operand:SSEMODEI24 1 "register_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+       (float:MODEF (match_operand:SWI48x 1 "register_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_MIX_SSE_I387
    && TARGET_INTER_UNIT_CONVERSIONS
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(set (match_dup 0) (float:MODEF (match_dup 1)))])
 
 (define_split
   [(set (match_operand:MODEF 0 "register_operand" "")
-       (float:MODEF (match_operand:SSEMODEI24 1 "register_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+       (float:MODEF (match_operand:SWI48x 1 "register_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_MIX_SSE_I387
    && !(TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(set (match_dup 2) (match_dup 1))
    (set (match_dup 0) (float:MODEF (match_dup 2)))])
 
-(define_insn "*float<SSEMODEI24:mode><MODEF:mode>2_mixed_interunit"
+(define_insn "*float<SWI48x:mode><MODEF:mode>2_mixed_interunit"
   [(set (match_operand:MODEF 0 "register_operand" "=f,x,x")
        (float:MODEF
-         (match_operand:SSEMODEI24 1 "nonimmediate_operand" "m,r,m")))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+         (match_operand:SWI48x 1 "nonimmediate_operand" "m,r,m")))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_MIX_SSE_I387
    && (TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))"
   "@
    fild%Z1\t%1
-   %vcvtsi2<MODEF:ssemodesuffix><SSEMODEI24:rex64suffix>\t{%1, %d0|%d0, %1}
-   %vcvtsi2<MODEF:ssemodesuffix><SSEMODEI24:rex64suffix>\t{%1, %d0|%d0, %1}"
+   %vcvtsi2<MODEF:ssemodesuffix><SWI48x:rex64suffix>\t{%1, %d0|%d0, %1}
+   %vcvtsi2<MODEF:ssemodesuffix><SWI48x:rex64suffix>\t{%1, %d0|%d0, %1}"
   [(set_attr "type" "fmov,sseicvt,sseicvt")
    (set_attr "prefix" "orig,maybe_vex,maybe_vex")
    (set_attr "mode" "<MODEF:MODE>")
    (set (attr "prefix_rex")
      (if_then_else
        (and (eq_attr "prefix" "maybe_vex")
-           (ne (symbol_ref "<SSEMODEI24:MODE>mode == DImode") (const_int 0)))
+           (match_test "<SWI48x:MODE>mode == DImode"))
        (const_string "1")
        (const_string "*")))
    (set_attr "unit" "i387,*,*")
    (set_attr "bdver1_decode" "*,double,direct")
    (set_attr "fp_int_src" "true")])
 
-(define_insn "*float<SSEMODEI24:mode><MODEF:mode>2_mixed_nointerunit"
+(define_insn "*float<SWI48x:mode><MODEF:mode>2_mixed_nointerunit"
   [(set (match_operand:MODEF 0 "register_operand" "=f,x")
        (float:MODEF
-         (match_operand:SSEMODEI24 1 "memory_operand" "m,m")))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+         (match_operand:SWI48x 1 "memory_operand" "m,m")))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_MIX_SSE_I387
    && !(TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))"
   "@
    fild%Z1\t%1
-   %vcvtsi2<MODEF:ssemodesuffix><SSEMODEI24:rex64suffix>\t{%1, %d0|%d0, %1}"
+   %vcvtsi2<MODEF:ssemodesuffix><SWI48x:rex64suffix>\t{%1, %d0|%d0, %1}"
   [(set_attr "type" "fmov,sseicvt")
    (set_attr "prefix" "orig,maybe_vex")
    (set_attr "mode" "<MODEF:MODE>")
    (set (attr "prefix_rex")
      (if_then_else
        (and (eq_attr "prefix" "maybe_vex")
-           (ne (symbol_ref "<SSEMODEI24:MODE>mode == DImode") (const_int 0)))
+           (match_test "<SWI48x:MODE>mode == DImode"))
        (const_string "1")
        (const_string "*")))
    (set_attr "athlon_decode" "*,direct")
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(const_int 0)]
 {
   rtx op1 = operands[1];
       emit_insn (gen_sse2_loadld (operands[4],
                                  CONST0_RTX (V4SImode), operands[2]));
     }
-  emit_insn
-    (gen_sse2_cvtdq2<ssevecmodesuffix> (operands[3], operands[4]));
+  if (<ssevecmode>mode == V4SFmode)
+    emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
+  else
+    emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
   DONE;
 })
 
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(const_int 0)]
 {
   operands[3] = simplify_gen_subreg (<ssevecmode>mode, operands[0],
 
   emit_insn (gen_sse2_loadld (operands[4],
                              CONST0_RTX (V4SImode), operands[1]));
-  emit_insn
-    (gen_sse2_cvtdq2<ssevecmodesuffix> (operands[3], operands[4]));
+  if (<ssevecmode>mode == V4SFmode)
+    emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
+  else
+    emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
   DONE;
 })
 
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(const_int 0)]
 {
   rtx op1 = operands[1];
   if (GET_CODE (op1) == SUBREG)
     op1 = SUBREG_REG (op1);
 
-  if (GENERAL_REG_P (op1) && TARGET_INTER_UNIT_MOVES)
+  if (GENERAL_REG_P (op1))
     {
       operands[4] = simplify_gen_subreg (V4SImode, operands[0], <MODE>mode, 0);
-      emit_insn (gen_sse2_loadld (operands[4],
-                                 CONST0_RTX (V4SImode), operands[1]));
+      if (TARGET_INTER_UNIT_MOVES)
+       emit_insn (gen_sse2_loadld (operands[4],
+                                   CONST0_RTX (V4SImode), operands[1]));
+      else
+       {
+         operands[5] = ix86_force_to_memory (GET_MODE (operands[1]),
+                                             operands[1]);
+         emit_insn (gen_sse2_loadld (operands[4],
+                                     CONST0_RTX (V4SImode), operands[5]));
+         ix86_free_from_memory (GET_MODE (operands[1]));
+       }
     }
   /* We can ignore possible trapping value in the
      high part of SSE register for non-trapping math. */
     operands[4] = simplify_gen_subreg (V4SImode, operands[1], SImode, 0);
   else
     gcc_unreachable ();
-  emit_insn
-    (gen_sse2_cvtdq2<ssevecmodesuffix> (operands[3], operands[4]));
+  if (<ssevecmode>mode == V4SFmode)
+    emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
+  else
+    emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
   DONE;
 })
 
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(const_int 0)]
 {
   operands[3] = simplify_gen_subreg (<ssevecmode>mode, operands[0],
 
   emit_insn (gen_sse2_loadld (operands[4],
                              CONST0_RTX (V4SImode), operands[1]));
-  emit_insn
-    (gen_sse2_cvtdq2<ssevecmodesuffix> (operands[3], operands[4]));
+  if (<ssevecmode>mode == V4SFmode)
+    emit_insn (gen_floatv4siv4sf2 (operands[3], operands[4]));
+  else
+    emit_insn (gen_sse2_cvtdq2pd (operands[3], operands[4]));
   DONE;
 })
 
-(define_insn "*float<SSEMODEI24:mode><MODEF:mode>2_sse_with_temp"
+(define_insn "*float<SWI48x:mode><MODEF:mode>2_sse_with_temp"
   [(set (match_operand:MODEF 0 "register_operand" "=x,x")
        (float:MODEF
-         (match_operand:SSEMODEI24 1 "nonimmediate_operand" "r,m")))
-  (clobber (match_operand:SSEMODEI24 2 "memory_operand" "=m,X"))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+         (match_operand:SWI48x 1 "nonimmediate_operand" "r,m")))
+  (clobber (match_operand:SWI48x 2 "memory_operand" "=m,X"))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH"
   "#"
   [(set_attr "type" "sseicvt")
    (set_attr "bdver1_decode" "double,direct")
    (set_attr "fp_int_src" "true")])
 
-(define_insn "*float<SSEMODEI24:mode><MODEF:mode>2_sse_interunit"
+(define_insn "*float<SWI48x:mode><MODEF:mode>2_sse_interunit"
   [(set (match_operand:MODEF 0 "register_operand" "=x,x")
        (float:MODEF
-         (match_operand:SSEMODEI24 1 "nonimmediate_operand" "r,m")))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+         (match_operand:SWI48x 1 "nonimmediate_operand" "r,m")))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && (TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))"
-  "%vcvtsi2<MODEF:ssemodesuffix><SSEMODEI24:rex64suffix>\t{%1, %d0|%d0, %1}"
+  "%vcvtsi2<MODEF:ssemodesuffix><SWI48x:rex64suffix>\t{%1, %d0|%d0, %1}"
   [(set_attr "type" "sseicvt")
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "<MODEF:MODE>")
    (set (attr "prefix_rex")
      (if_then_else
        (and (eq_attr "prefix" "maybe_vex")
-           (ne (symbol_ref "<SSEMODEI24:MODE>mode == DImode") (const_int 0)))
+           (match_test "<SWI48x:MODE>mode == DImode"))
        (const_string "1")
        (const_string "*")))
    (set_attr "athlon_decode" "double,direct")
 
 (define_split
   [(set (match_operand:MODEF 0 "register_operand" "")
-       (float:MODEF (match_operand:SSEMODEI24 1 "nonimmediate_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+       (float:MODEF (match_operand:SWI48x 1 "nonimmediate_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && (TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(set (match_dup 0) (float:MODEF (match_dup 1)))])
 
-(define_insn "*float<SSEMODEI24:mode><MODEF:mode>2_sse_nointerunit"
+(define_insn "*float<SWI48x:mode><MODEF:mode>2_sse_nointerunit"
   [(set (match_operand:MODEF 0 "register_operand" "=x")
        (float:MODEF
-         (match_operand:SSEMODEI24 1 "memory_operand" "m")))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+         (match_operand:SWI48x 1 "memory_operand" "m")))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && !(TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))"
-  "%vcvtsi2<MODEF:ssemodesuffix><SSEMODEI24:rex64suffix>\t{%1, %d0|%d0, %1}"
+  "%vcvtsi2<MODEF:ssemodesuffix><SWI48x:rex64suffix>\t{%1, %d0|%d0, %1}"
   [(set_attr "type" "sseicvt")
    (set_attr "prefix" "maybe_vex")
    (set_attr "mode" "<MODEF:MODE>")
    (set (attr "prefix_rex")
      (if_then_else
        (and (eq_attr "prefix" "maybe_vex")
-           (ne (symbol_ref "<SSEMODEI24:MODE>mode == DImode") (const_int 0)))
+           (match_test "<SWI48x:MODE>mode == DImode"))
        (const_string "1")
        (const_string "*")))
    (set_attr "athlon_decode" "direct")
 
 (define_split
   [(set (match_operand:MODEF 0 "register_operand" "")
-       (float:MODEF (match_operand:SSEMODEI24 1 "register_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+       (float:MODEF (match_operand:SWI48x 1 "register_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && !(TARGET_INTER_UNIT_CONVERSIONS || optimize_function_for_size_p (cfun))
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(set (match_dup 2) (match_dup 1))
    (set (match_dup 0) (float:MODEF (match_dup 2)))])
 
 (define_split
   [(set (match_operand:MODEF 0 "register_operand" "")
-       (float:MODEF (match_operand:SSEMODEI24 1 "memory_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
-  "(<SSEMODEI24:MODE>mode != DImode || TARGET_64BIT)
+       (float:MODEF (match_operand:SWI48x 1 "memory_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
+  "(<SWI48x:MODE>mode != DImode || TARGET_64BIT)
    && SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && reload_completed
    && (SSE_REG_P (operands[0])
        || (GET_CODE (operands[0]) == SUBREG
-          && SSE_REG_P (operands[0])))"
+          && SSE_REG_P (SUBREG_REG (operands[0]))))"
   [(set (match_dup 0) (float:MODEF (match_dup 1)))])
 
-(define_insn "*float<SSEMODEI24:mode><X87MODEF:mode>2_i387_with_temp"
+(define_insn "*float<SWI48x:mode><X87MODEF:mode>2_i387_with_temp"
   [(set (match_operand:X87MODEF 0 "register_operand" "=f,f")
        (float:X87MODEF
-         (match_operand:SSEMODEI24 1 "nonimmediate_operand" "m,?r")))
-  (clobber (match_operand:SSEMODEI24 2 "memory_operand" "=X,m"))]
+         (match_operand:SWI48x 1 "nonimmediate_operand" "m,?r")))
+  (clobber (match_operand:SWI48x 2 "memory_operand" "=X,m"))]
   "TARGET_80387
-   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode)"
+   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SWI48x:MODE>mode)"
   "@
    fild%Z1\t%1
    #"
    (set_attr "unit" "*,i387")
    (set_attr "fp_int_src" "true")])
 
-(define_insn "*float<SSEMODEI24:mode><X87MODEF:mode>2_i387"
+(define_insn "*float<SWI48x:mode><X87MODEF:mode>2_i387"
   [(set (match_operand:X87MODEF 0 "register_operand" "=f")
        (float:X87MODEF
-         (match_operand:SSEMODEI24 1 "memory_operand" "m")))]
+         (match_operand:SWI48x 1 "memory_operand" "m")))]
   "TARGET_80387
-   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode)"
+   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SWI48x:MODE>mode)"
   "fild%Z1\t%1"
   [(set_attr "type" "fmov")
    (set_attr "mode" "<X87MODEF:MODE>")
    (set_attr "fp_int_src" "true")])
 
 (define_split
-  [(set (match_operand:X87MODEF 0 "register_operand" "")
-       (float:X87MODEF (match_operand:SSEMODEI24 1 "register_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
+  [(set (match_operand:X87MODEF 0 "fp_register_operand" "")
+       (float:X87MODEF (match_operand:SWI48x 1 "register_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
   "TARGET_80387
-   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode)
-   && reload_completed
-   && FP_REG_P (operands[0])"
+   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SWI48x:MODE>mode)
+   && reload_completed"
   [(set (match_dup 2) (match_dup 1))
    (set (match_dup 0) (float:X87MODEF (match_dup 2)))])
 
 (define_split
-  [(set (match_operand:X87MODEF 0 "register_operand" "")
-       (float:X87MODEF (match_operand:SSEMODEI24 1 "memory_operand" "")))
-   (clobber (match_operand:SSEMODEI24 2 "memory_operand" ""))]
+  [(set (match_operand:X87MODEF 0 "fp_register_operand" "")
+       (float:X87MODEF (match_operand:SWI48x 1 "memory_operand" "")))
+   (clobber (match_operand:SWI48x 2 "memory_operand" ""))]
   "TARGET_80387
-   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SSEMODEI24:MODE>mode)
-   && reload_completed
-   && FP_REG_P (operands[0])"
+   && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, <SWI48x:MODE>mode)
+   && reload_completed"
   [(set (match_dup 0) (float:X87MODEF (match_dup 1)))])
 
 ;; Avoid store forwarding (partial memory) stall penalty
    (set_attr "fp_int_src" "true")])
 
 (define_split
-  [(set (match_operand:X87MODEF 0 "register_operand" "")
+  [(set (match_operand:X87MODEF 0 "fp_register_operand" "")
        (float:X87MODEF (match_operand:DI 1 "register_operand" "")))
    (clobber (match_scratch:V4SI 3 ""))
    (clobber (match_scratch:V4SI 4 ""))
   "TARGET_80387 && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, DImode)
    && TARGET_SSE2 && TARGET_INTER_UNIT_MOVES
    && !TARGET_64BIT && optimize_function_for_speed_p (cfun)
-   && reload_completed
-   && FP_REG_P (operands[0])"
+   && reload_completed"
   [(set (match_dup 2) (match_dup 3))
    (set (match_dup 0) (float:X87MODEF (match_dup 2)))]
 {
 })
 
 (define_split
-  [(set (match_operand:X87MODEF 0 "register_operand" "")
+  [(set (match_operand:X87MODEF 0 "fp_register_operand" "")
        (float:X87MODEF (match_operand:DI 1 "memory_operand" "")))
    (clobber (match_scratch:V4SI 3 ""))
    (clobber (match_scratch:V4SI 4 ""))
   "TARGET_80387 && X87_ENABLE_FLOAT (<X87MODEF:MODE>mode, DImode)
    && TARGET_SSE2 && TARGET_INTER_UNIT_MOVES
    && !TARGET_64BIT && optimize_function_for_speed_p (cfun)
-   && reload_completed
-   && FP_REG_P (operands[0])"
+   && reload_completed"
   [(set (match_dup 0) (float:X87MODEF (match_dup 1)))])
 
 ;; Avoid store forwarding (partial memory) stall penalty by extending
   [(set_attr "type" "alu")
    (set_attr "mode" "QI")])
 
-(define_insn "*lea_1"
-  [(set (match_operand:P 0 "register_operand" "=r")
-       (match_operand:P 1 "no_seg_address_operand" "p"))]
-  ""
-  "lea{<imodesuffix>}\t{%a1, %0|%0, %a1}"
-  [(set_attr "type" "lea")
-   (set_attr "mode" "<MODE>")])
-
-(define_insn "*lea_2"
+(define_insn_and_split "*lea_1"
   [(set (match_operand:SI 0 "register_operand" "=r")
-       (subreg:SI (match_operand:DI 1 "no_seg_address_operand" "p") 0))]
+       (subreg:SI (match_operand:DI 1 "lea_address_operand" "p") 0))]
   "TARGET_64BIT"
   "lea{l}\t{%a1, %0|%0, %a1}"
+  "&& reload_completed && ix86_avoid_lea_for_addr (insn, operands)"
+  [(const_int 0)]
+{
+  ix86_split_lea_for_addr (operands, SImode);
+  DONE;
+}
   [(set_attr "type" "lea")
    (set_attr "mode" "SI")])
 
-(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)))]
-  "TARGET_64BIT"
+(define_insn_and_split "*lea<mode>_2"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (match_operand:SWI48 1 "lea_address_operand" "p"))]
+  ""
+  "lea{<imodesuffix>}\t{%a1, %0|%0, %a1}"
+  "reload_completed && ix86_avoid_lea_for_addr (insn, operands)"
+  [(const_int 0)]
+{
+  ix86_split_lea_for_addr (operands, <MODE>mode);
+  DONE;
+}
+  [(set_attr "type" "lea")
+   (set_attr "mode" "<MODE>")])
+
+(define_insn "*lea_3_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (subreg:SI (match_operand:DI 1 "lea_address_operand" "j") 0)))]
+  "TARGET_64BIT"
+  "lea{l}\t{%a1, %k0|%k0, %a1}"
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
+
+(define_insn "*lea_4_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (match_operand:SI 1 "lea_address_operand" "j")))]
+  "TARGET_64BIT"
+  "lea{l}\t{%a1, %k0|%k0, %a1}"
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
+
+(define_insn "*lea_5_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (and:DI
+         (subreg:DI (match_operand:SI 1 "lea_address_operand" "p") 0)
+         (match_operand:DI 2 "const_32bit_mask" "n")))]
+  "TARGET_64BIT"
+  "lea{l}\t{%a1, %k0|%k0, %a1}"
+  [(set_attr "type" "lea")
+   (set_attr "mode" "SI")])
+
+(define_insn "*lea_6_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (and:DI
+         (match_operand:DI 1 "lea_address_operand" "p")
+         (match_operand:DI 2 "const_32bit_mask" "n")))]
+  "TARGET_64BIT"
   "lea{l}\t{%a1, %k0|%k0, %a1}"
   [(set_attr "type" "lea")
    (set_attr "mode" "SI")])
   [(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>")))
+         (match_operand:SWI48 2 "x86_64_general_operand" "rme,re,0,le")))
    (clobber (reg:CC FLAGS_REG))]
   "ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
 {
 ;; operands so proper swapping will be done in reload.  This allow
 ;; patterns constructed from addsi_1 to match.
 
-(define_insn "*addsi_1_zext"
+(define_insn "addsi_1_zext"
   [(set (match_operand:DI 0 "register_operand" "=r,r,r")
        (zero_extend:DI
          (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,r,r")
-                  (match_operand:SI 2 "general_operand" "g,0,li"))))
+                  (match_operand:SI 2 "x86_64_general_operand" "rme,0,le"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (PLUS, SImode, operands)"
 {
    (set_attr "mode" "SI")])
 
 (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_INCDEC:
-      if (operands[2] == const1_rtx)
-       return "inc{w}\t%0";
-      else
-        {
-         gcc_assert (operands[2] == constm1_rtx);
-         return "dec{w}\t%0";
-       }
-
-    default:
-      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: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_insn "*addhi_1_lea"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=r,rm,r,r")
-       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0,r,r")
-                (match_operand:HI 2 "general_operand" "rmn,rn,0,ln")))
+  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r,r,Yp")
+       (plus:HI (match_operand:HI 1 "nonimmediate_operand" "%0,0,r,Yp")
+                (match_operand:HI 2 "general_operand" "rn,rm,0,ln")))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (PLUS, HImode, operands)"
+  "ix86_binary_operator_ok (PLUS, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
        (const_string "*")))
    (set_attr "mode" "HI,HI,HI,SI")])
 
-;; %%% Potential partial reg stall on alternative 2.  What to do?
-(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:
-      if (operands[2] == const1_rtx)
-       return widen ? "inc{l}\t%k0" : "inc{b}\t%0";
-      else
-       {
-         gcc_assert (operands[2] == constm1_rtx);
-         return widen ? "dec{l}\t%k0" : "dec{b}\t%0";
-       }
-
-    default:
-      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: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" "QI,QI,SI")])
-
 ;; %%% Potential partial reg stall on alternatives 3 and 4.  What to do?
-(define_insn "*addqi_1_lea"
-  [(set (match_operand:QI 0 "nonimmediate_operand" "=q,qm,q,r,r,r")
-       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,q,0,r,r")
-                (match_operand:QI 2 "general_operand" "qmn,qn,0,rn,0,ln")))
+(define_insn "*addqi_1"
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,q,q,r,r,Yp")
+       (plus:QI (match_operand:QI 1 "nonimmediate_operand" "%0,0,q,0,r,Yp")
+                (match_operand:QI 2 "general_operand" "qn,qm,0,rn,0,ln")))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (PLUS, QImode, operands)"
+  "ix86_binary_operator_ok (PLUS, QImode, operands)"
 {
-  int widen = (which_alternative == 3 || which_alternative == 4);
+  bool widen = (which_alternative == 3 || which_alternative == 4);
 
   switch (get_attr_type (insn))
     {
 (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")))
+                (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]))"
         (const_string "none")))
    (set_attr "mode" "QI")])
 
-;; Convert lea to the lea pattern to avoid flags dependency.
+;; Split non destructive adds if we cannot use lea.
 (define_split
-  [(set (match_operand 0 "register_operand" "")
-       (plus (match_operand 1 "register_operand" "")
-              (match_operand 2 "nonmemory_operand" "")))
+  [(set (match_operand:SWI48 0 "register_operand" "")
+       (plus:SWI48 (match_operand:SWI48 1 "register_operand" "")
+              (match_operand:SWI48 2 "nonmemory_operand" "")))
+   (clobber (reg:CC FLAGS_REG))]
+  "reload_completed && ix86_avoid_lea_for_add (insn, operands)"
+  [(set (match_dup 0) (match_dup 1))
+   (parallel [(set (match_dup 0) (plus:<MODE> (match_dup 0) (match_dup 2)))
+             (clobber (reg:CC FLAGS_REG))])])
+
+;; Convert add to the lea pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:SWI 0 "register_operand" "")
+       (plus:SWI (match_operand:SWI 1 "register_operand" "")
+                 (match_operand:SWI 2 "<nonmemory_operand>" "")))
    (clobber (reg:CC FLAGS_REG))]
   "reload_completed && ix86_lea_for_add_ok (insn, operands)" 
   [(const_int 0)]
 {
+  enum machine_mode mode = <MODE>mode;
   rtx pat;
-  enum machine_mode mode = GET_MODE (operands[0]);
-
-  /* In -fPIC mode the constructs like (const (unspec [symbol_ref]))
-     may confuse gen_lowpart.  */
-  if (mode != Pmode)
-    {
-      operands[1] = gen_lowpart (Pmode, operands[1]);
-      operands[2] = gen_lowpart (Pmode, operands[2]);
-    }
-
-  pat = gen_rtx_PLUS (Pmode, operands[1], operands[2]);
 
   if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (SImode))
-    operands[0] = gen_lowpart (SImode, operands[0]);
+    { 
+      mode = SImode; 
+      operands[0] = gen_lowpart (mode, operands[0]);
+      operands[1] = gen_lowpart (mode, operands[1]);
+      operands[2] = gen_lowpart (mode, operands[2]);
+    }
 
-  if (TARGET_64BIT && mode != Pmode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  pat = gen_rtx_PLUS (mode, operands[1], operands[2]);
 
   emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
   DONE;
 })
 
-;; Convert lea to the lea pattern to avoid flags dependency.
-;; ??? This pattern handles immediate operands that do not satisfy immediate
-;; operand predicate (TARGET_LEGITIMATE_CONSTANT_P) in the previous pattern.
-(define_split
-  [(set (match_operand:DI 0 "register_operand" "")
-       (plus:DI (match_operand:DI 1 "register_operand" "")
-                (match_operand:DI 2 "x86_64_immediate_operand" "")))
-   (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed 
-   && true_regnum (operands[0]) != true_regnum (operands[1])"
-  [(set (match_dup 0)
-       (plus:DI (match_dup 1) (match_dup 2)))])
-
-;; Convert lea to the lea pattern to avoid flags dependency.
+;; Convert add 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" ""))))
+                  (match_operand:SI 2 "x86_64_nonmemory_operand" ""))))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && reload_completed
-   && ix86_lea_for_add_ok (insn, operands)"
+  "TARGET_64BIT && reload_completed && ix86_lea_for_add_ok (insn, operands)"
   [(set (match_dup 0)
-       (zero_extend:DI (subreg:SI (plus:DI (match_dup 1) (match_dup 2)) 0)))]
-{
-  operands[1] = gen_lowpart (DImode, operands[1]);
-  operands[2] = gen_lowpart (DImode, operands[2]);
-})
+       (zero_extend:DI (plus:SI (match_dup 1) (match_dup 2))))])
 
 (define_insn "*add<mode>_2"
   [(set (reg FLAGS_REG)
        (compare
          (plus:SWI
-           (match_operand:SWI 1 "nonimmediate_operand" "%0,0")
-           (match_operand:SWI 2 "<general_operand>" "<g>,<r><i>"))
+           (match_operand:SWI 1 "nonimmediate_operand" "%0,0,<r>")
+           (match_operand:SWI 2 "<general_operand>" "<g>,<r><i>,0"))
          (const_int 0)))
-   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>,<r>m")
+   (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>,<r>m,<r>")
        (plus:SWI (match_dup 1) (match_dup 2)))]
   "ix86_match_ccmode (insn, CCGOCmode)
    && ix86_binary_operator_ok (PLUS, <MODE>mode, operands)"
        }
 
     default:
+      if (which_alternative == 2)
+       {
+         rtx tmp;
+         tmp = operands[1], operands[1] = operands[2], operands[2] = tmp;
+       }
+        
+      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}";
 
 (define_insn "*addsi_2_zext"
   [(set (reg FLAGS_REG)
        (compare
-         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                  (match_operand:SI 2 "general_operand" "g"))
+         (plus:SI (match_operand:SI 1 "nonimmediate_operand" "%0,r")
+                  (match_operand:SI 2 "x86_64_general_operand" "rme,0"))
          (const_int 0)))
-   (set (match_operand:DI 0 "register_operand" "=r")
+   (set (match_operand:DI 0 "register_operand" "=r,r")
        (zero_extend:DI (plus:SI (match_dup 1) (match_dup 2))))]
   "TARGET_64BIT && ix86_match_ccmode (insn, CCGOCmode)
    && ix86_binary_operator_ok (PLUS, SImode, operands)"
        }
 
     default:
+      if (which_alternative == 1)
+       {
+         rtx tmp;
+         tmp = operands[1], operands[1] = operands[2], operands[2] = tmp;
+       }
+
       if (x86_maybe_negate_const_int (&operands[2], SImode))
         return "sub{l}\t{%2, %k0|%k0, %2}";
 
 (define_insn "*add<mode>_3"
   [(set (reg FLAGS_REG)
        (compare
-         (neg:SWI (match_operand:SWI 2 "<general_operand>" "<g>"))
-         (match_operand:SWI 1 "nonimmediate_operand" "%0")))
-   (clobber (match_scratch:SWI 0 "=<r>"))]
+         (neg:SWI (match_operand:SWI 2 "<general_operand>" "<g>,0"))
+         (match_operand:SWI 1 "nonimmediate_operand" "%0,<r>")))
+   (clobber (match_scratch:SWI 0 "=<r>,<r>"))]
   "ix86_match_ccmode (insn, CCZmode)
    && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
 {
        }
 
     default:
+      if (which_alternative == 1)
+       {
+         rtx tmp;
+         tmp = operands[1], operands[1] = operands[2], operands[2] = tmp;
+       }
+
+      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}";
 
 (define_insn "*addsi_3_zext"
   [(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")
+         (neg:SI (match_operand:SI 2 "x86_64_general_operand" "rme,0"))
+         (match_operand:SI 1 "nonimmediate_operand" "%0,r")))
+   (set (match_operand:DI 0 "register_operand" "=r,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)"
        }
 
     default:
+      if (which_alternative == 1)
+       {
+         rtx tmp;
+         tmp = operands[1], operands[1] = operands[2], operands[2] = tmp;
+       }
+
       if (x86_maybe_negate_const_int (&operands[2], SImode))
         return "sub{l}\t{%2, %k0|%k0, %2}";
 
   [(set (reg FLAGS_REG)
        (compare
          (plus:SWI
-           (match_operand:SWI 1 "nonimmediate_operand" "%0")
-           (match_operand:SWI 2 "<general_operand>" "<g>"))
+           (match_operand:SWI 1 "nonimmediate_operand" "%0,<r>")
+           (match_operand:SWI 2 "<general_operand>" "<g>,0"))
          (const_int 0)))
-   (clobber (match_scratch:SWI 0 "=<r>"))]
+   (clobber (match_scratch:SWI 0 "=<r>,<r>"))]
   "ix86_match_ccmode (insn, CCGOCmode)
    && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
 {
        }
 
     default:
+      if (which_alternative == 1)
+       {
+         rtx tmp;
+         tmp = operands[1], operands[1] = operands[2], operands[2] = tmp;
+       }
+
+      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}";
 
   [(set_attr "type" "alu")
    (set_attr "mode" "QI")])
 
-;; The lea patterns for non-Pmodes needs to be matched by
+;; The lea patterns for modes less than 32 bits need to be matched by
 ;; several insns converted to real lea by splitters.
 
 (define_insn_and_split "*lea_general_1"
        (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))
+  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode)
    && (!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])
   "&& reload_completed"
   [(const_int 0)]
 {
+  enum machine_mode mode = SImode;
   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[0] = gen_lowpart (mode, operands[0]);
+  operands[1] = gen_lowpart (mode, operands[1]);
+  operands[2] = gen_lowpart (mode, operands[2]);
+  operands[3] = gen_lowpart (mode, operands[3]);
+
+  pat = gen_rtx_PLUS (mode, gen_rtx_PLUS (mode, 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_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"
-  "#"
-  "&& 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_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 2 "const248_operand" "n"))
              (match_operand 3 "nonmemory_operand" "ri")))]
-  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode
-    || (TARGET_64BIT && GET_MODE (operands[0]) == SImode))
+  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode)
    && (!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)]
 {
+  enum machine_mode mode = SImode;
   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);
+
+  operands[0] = gen_lowpart (mode, operands[0]);
+  operands[1] = gen_lowpart (mode, operands[1]);
+  operands[3] = gen_lowpart (mode, operands[3]);
+
+  pat = gen_rtx_PLUS (mode, gen_rtx_MULT (mode, operands[1], operands[2]),
+                     operands[3]);
+
   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 2 "const248_operand" "n"))
                    (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))
+  "(GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode)
    && (!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)]
 {
+  enum machine_mode mode = SImode;
   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[0] = gen_lowpart (mode, operands[0]);
+  operands[1] = gen_lowpart (mode, operands[1]);
+  operands[3] = gen_lowpart (mode, operands[3]);
+  operands[4] = gen_lowpart (mode, operands[4]);
+
+  pat = gen_rtx_PLUS (mode,
+                     gen_rtx_PLUS (mode,
+                                   gen_rtx_MULT (mode, 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"
+(define_insn_and_split "*lea_general_4"
+  [(set (match_operand 0 "register_operand" "=r")
+       (any_or (ashift
+                 (match_operand 1 "index_register_operand" "l")
+                 (match_operand 2 "const_int_operand" "n"))
+               (match_operand 3 "const_int_operand" "n")))]
+  "(((GET_MODE (operands[0]) == QImode || GET_MODE (operands[0]) == HImode)
+      && (!TARGET_PARTIAL_REG_STALL || optimize_function_for_size_p (cfun)))
+    || GET_MODE (operands[0]) == SImode
+    || (TARGET_64BIT && GET_MODE (operands[0]) == DImode))
+   && GET_MODE (operands[0]) == GET_MODE (operands[1])
+   && ((unsigned HOST_WIDE_INT) INTVAL (operands[2])) - 1 < 3
+   && ((unsigned HOST_WIDE_INT) INTVAL (operands[3])
+       < ((unsigned HOST_WIDE_INT) 1 << INTVAL (operands[2])))"
   "#"
   "&& 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]);
+  [(const_int 0)]
+{
+  enum machine_mode mode = GET_MODE (operands[0]);
+  rtx pat;
+
+  if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (SImode))
+    { 
+      mode = SImode; 
+      operands[0] = gen_lowpart (mode, operands[0]);
+      operands[1] = gen_lowpart (mode, operands[1]);
+    }
+
+  operands[2] = GEN_INT (1 << INTVAL (operands[2]));
+
+  pat = plus_constant (gen_rtx_MULT (mode, operands[1], operands[2]),
+                      INTVAL (operands[3]));
+
+  emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
+  DONE;
 }
   [(set_attr "type" "lea")
-   (set_attr "mode" "SI")])
+   (set (attr "mode")
+      (if_then_else (match_operand:DI 0 "" "")
+       (const_string "DI")
+       (const_string "SI")))])
 \f
 ;; Subtract instructions
 
   [(set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
          (minus:SI (match_operand:SI 1 "register_operand" "0")
-                   (match_operand:SI 2 "general_operand" "g"))))
+                   (match_operand:SI 2 "x86_64_general_operand" "rme"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (MINUS, SImode, operands)"
   "sub{l}\t{%2, %k0|%k0, %2}"
   [(set (reg FLAGS_REG)
        (compare
          (minus:SI (match_operand:SI 1 "register_operand" "0")
-                   (match_operand:SI 2 "general_operand" "g"))
+                   (match_operand:SI 2 "x86_64_general_operand" "rme"))
          (const_int 0)))
    (set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
 (define_insn "*subsi_3_zext"
   [(set (reg FLAGS_REG)
        (compare (match_operand:SI 1 "register_operand" "0")
-                (match_operand:SI 2 "general_operand" "g")))
+                (match_operand:SI 2 "x86_64_general_operand" "rme")))
    (set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
          (minus:SI (match_dup 1)
          (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")))))
+                           (match_operand:SI 2 "x86_64_general_operand" "rme")))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (PLUS, SImode, operands)"
   "adc{l}\t{%2, %k0|%k0, %2}"
          (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")))))
+                            (match_operand:SI 2 "x86_64_general_operand" "rme")))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (MINUS, SImode, operands)"
   "sbb{l}\t{%2, %k0|%k0, %2}"
        (compare:CCC
          (plusminus:SI
            (match_operand:SI 1 "nonimmediate_operand" "<comm>0")
-           (match_operand:SI 2 "general_operand" "g"))
+           (match_operand:SI 2 "x86_64_general_operand" "rme"))
          (match_dup 1)))
    (set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI (plusminus:SI (match_dup 1) (match_dup 2))))]
   [(set (match_operand:DI 0 "register_operand" "=r,r,r")
        (zero_extend:DI
          (mult:SI (match_operand:SI 1 "nonimmediate_operand" "%rm,rm,0")
-                  (match_operand:SI 2 "general_operand" "K,i,mr"))))
+                  (match_operand:SI 2 "x86_64_general_operand" "K,e,mr"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT
    && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
              (clobber (reg:CC FLAGS_REG))])]
   "TARGET_QIMODE_MATH")
 
-(define_insn "*<u>mul<mode><dwi>3_1"
-  [(set (match_operand:<DWI> 0 "register_operand" "=A")
+(define_insn "*bmi2_umulditi3_1"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (mult:DI
+         (match_operand:DI 2 "nonimmediate_operand" "%d")
+         (match_operand:DI 3 "nonimmediate_operand" "rm")))
+   (set (match_operand:DI 1 "register_operand" "=r")
+       (truncate:DI
+         (lshiftrt:TI
+           (mult:TI (zero_extend:TI (match_dup 2))
+                    (zero_extend:TI (match_dup 3)))
+           (const_int 64))))]
+  "TARGET_64BIT && TARGET_BMI2
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
+  "mulx\t{%3, %0, %1|%1, %0, %3}"
+  [(set_attr "type" "imulx")
+   (set_attr "prefix" "vex")
+   (set_attr "mode" "DI")])
+
+(define_insn "*bmi2_umulsidi3_1"
+  [(set (match_operand:SI 0 "register_operand" "=r")
+       (mult:SI
+         (match_operand:SI 2 "nonimmediate_operand" "%d")
+         (match_operand:SI 3 "nonimmediate_operand" "rm")))
+   (set (match_operand:SI 1 "register_operand" "=r")
+       (truncate:SI
+         (lshiftrt:DI
+           (mult:DI (zero_extend:DI (match_dup 2))
+                    (zero_extend:DI (match_dup 3)))
+           (const_int 32))))]
+  "!TARGET_64BIT && TARGET_BMI2
+   && !(MEM_P (operands[1]) && MEM_P (operands[2]))"
+  "mulx\t{%3, %0, %1|%1, %0, %3}"
+  [(set_attr "type" "imulx")
+   (set_attr "prefix" "vex")
+   (set_attr "mode" "SI")])
+
+(define_insn "*umul<mode><dwi>3_1"
+  [(set (match_operand:<DWI> 0 "register_operand" "=A,r")
        (mult:<DWI>
-         (any_extend:<DWI>
-           (match_operand:DWIH 1 "nonimmediate_operand" "%0"))
-         (any_extend:<DWI>
-           (match_operand:DWIH 2 "nonimmediate_operand" "rm"))))
+         (zero_extend:<DWI>
+           (match_operand:DWIH 1 "nonimmediate_operand" "%0,d"))
+         (zero_extend:<DWI>
+           (match_operand:DWIH 2 "nonimmediate_operand" "rm,rm"))))
    (clobber (reg:CC FLAGS_REG))]
   "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
-  "<sgnprefix>mul{<imodesuffix>}\t%2"
-  [(set_attr "type" "imul")
-   (set_attr "length_immediate" "0")
+  "@
+   mul{<imodesuffix>}\t%2
+   #"
+  [(set_attr "isa" "*,bmi2")
+   (set_attr "type" "imul,imulx")
+   (set_attr "length_immediate" "0,*")
    (set (attr "athlon_decode")
-     (if_then_else (eq_attr "cpu" "athlon")
-        (const_string "vector")
-        (const_string "double")))
+       (cond [(eq_attr "alternative" "0")
+                (if_then_else (eq_attr "cpu" "athlon")
+                  (const_string "vector")
+                  (const_string "double"))]
+             (const_string "*")))
+   (set_attr "amdfam10_decode" "double,*")
+   (set_attr "bdver1_decode" "direct,*")
+   (set_attr "prefix" "orig,vex")
+   (set_attr "mode" "<MODE>")])
+
+;; Convert mul to the mulx pattern to avoid flags dependency.
+(define_split
+ [(set (match_operand:<DWI> 0 "register_operand" "")
+       (mult:<DWI>
+        (zero_extend:<DWI>
+          (match_operand:DWIH 1 "register_operand" ""))
+        (zero_extend:<DWI>
+          (match_operand:DWIH 2 "nonimmediate_operand" ""))))
+  (clobber (reg:CC FLAGS_REG))]
+ "TARGET_BMI2 && reload_completed
+  && true_regnum (operands[1]) == DX_REG"
+  [(parallel [(set (match_dup 3)
+                  (mult:DWIH (match_dup 1) (match_dup 2)))
+             (set (match_dup 4)
+                  (truncate:DWIH
+                    (lshiftrt:<DWI>
+                      (mult:<DWI> (zero_extend:<DWI> (match_dup 1))
+                                  (zero_extend:<DWI> (match_dup 2)))
+                      (match_dup 5))))])]
+{
+  split_double_mode (<DWI>mode, &operands[0], 1, &operands[3], &operands[4]);
+
+  operands[5] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode));
+})
+
+(define_insn "*mul<mode><dwi>3_1"
+  [(set (match_operand:<DWI> 0 "register_operand" "=A")
+       (mult:<DWI>
+         (sign_extend:<DWI>
+           (match_operand:DWIH 1 "nonimmediate_operand" "%0"))
+         (sign_extend:<DWI>
+           (match_operand:DWIH 2 "nonimmediate_operand" "rm"))))
+   (clobber (reg:CC FLAGS_REG))]
+  "!(MEM_P (operands[1]) && MEM_P (operands[2]))"
+  "imul{<imodesuffix>}\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 "bdver1_decode" "direct")
    (set_attr "mode" "<MODE>")])
   "(TARGET_80387 && X87_ENABLE_ARITH (SFmode))
     || TARGET_SSE_MATH"
 {
-  if (TARGET_SSE_MATH && TARGET_RECIP && optimize_insn_for_speed_p ()
+  if (TARGET_SSE_MATH
+      && TARGET_RECIP_DIV
+      && optimize_insn_for_speed_p ()
       && flag_finite_math_only && !flag_trapping_math
       && flag_unsafe_math_optimizations)
     {
   [(set (reg:CCNO FLAGS_REG)
        (compare:CCNO
          (and:SI (match_operand:SI 0 "nonimmediate_operand" "")
-                 (match_operand:SI 1 "nonmemory_operand" ""))
+                 (match_operand:SI 1 "x86_64_nonmemory_operand" ""))
          (const_int 0)))])
 
 (define_expand "testqi_ccz_1"
        (compare
         (and:SWI124
          (match_operand:SWI124 0 "nonimmediate_operand" "%!*a,<r>,<r>m")
-         (match_operand:SWI124 1 "general_operand" "<i>,<i>,<r><i>"))
+         (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]))"
          [(and:SI (zero_extract:SI (match_dup 2) (const_int 8) (const_int 8))
                   (match_dup 3))
           (const_int 0)]))]
-  "operands[2] = gen_lowpart (SImode, operands[2]);
-   operands[3] = gen_int_mode (INTVAL (operands[3]) >> 8, SImode);")
+{
+  operands[2] = gen_lowpart (SImode, operands[2]);
+  operands[3] = gen_int_mode (INTVAL (operands[3]) >> 8, SImode);
+})
 
 (define_split
   [(set (match_operand 0 "flags_reg_operand" "")
   [(set (match_dup 0)
        (match_op_dup 1 [(and:QI (match_dup 2) (match_dup 3))
                         (const_int 0)]))]
-  "operands[2] = gen_lowpart (QImode, operands[2]);
-   operands[3] = gen_lowpart (QImode, operands[3]);")
+{
+  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,
        enum machine_mode mode;
 
        gcc_assert (CONST_INT_P (operands[2]));
-        if (INTVAL (operands[2]) == 0xff)
-         mode = QImode;
+       if (INTVAL (operands[2]) == (HOST_WIDE_INT) 0xffffffff)
+         mode = SImode;
+       else if (INTVAL (operands[2]) == 0xffff)
+         mode = HImode;
        else
          {
-           gcc_assert (INTVAL (operands[2]) == 0xffff);
-           mode = HImode;
+           gcc_assert (INTVAL (operands[2]) == 0xff);
+           mode = QImode;
          }
 
        operands[1] = gen_lowpart (mode, operands[1]);
-       if (mode == QImode)
-         return "movz{bl|x}\t{%1, %k0|%k0, %1}";
-       else
+       if (mode == SImode)
+         return "mov{l}\t{%1, %k0|%k0, %1}";
+       else if (mode == HImode)
          return "movz{wl|x}\t{%1, %k0|%k0, %1}";
+       else
+         return "movz{bl|x}\t{%1, %k0|%k0, %1}";
       }
 
     default:
    (set (attr "prefix_rex")
      (if_then_else
        (and (eq_attr "type" "imovx")
-           (and (ne (symbol_ref "INTVAL (operands[2]) == 0xff") (const_int 0))
+           (and (match_test "INTVAL (operands[2]) == 0xff")
                 (match_operand 1 "ext_QIreg_operand" "")))
        (const_string "1")
        (const_string "*")))
 (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")
-               (match_operand:SI 2 "general_operand" "ri,rm,L")))
+               (match_operand:SI 2 "x86_64_general_operand" "re,rm,L")))
    (clobber (reg:CC FLAGS_REG))]
   "ix86_binary_operator_ok (AND, SImode, operands)"
 {
        enum machine_mode mode;
 
        gcc_assert (CONST_INT_P (operands[2]));
-        if (INTVAL (operands[2]) == 0xff)
-         mode = QImode;
+        if (INTVAL (operands[2]) == 0xffff)
+         mode = HImode;
        else
          {
-           gcc_assert (INTVAL (operands[2]) == 0xffff);
-           mode = HImode;
+           gcc_assert (INTVAL (operands[2]) == 0xff);
+           mode = QImode;
          }
 
        operands[1] = gen_lowpart (mode, operands[1]);
-       if (mode == QImode)
-         return "movz{bl|x}\t{%1, %0|%0, %1}";
-       else
+       if (mode == HImode)
          return "movz{wl|x}\t{%1, %0|%0, %1}";
+       else
+         return "movz{bl|x}\t{%1, %0|%0, %1}";
       }
 
     default:
    (set (attr "prefix_rex")
      (if_then_else
        (and (eq_attr "type" "imovx")
-           (and (ne (symbol_ref "INTVAL (operands[2]) == 0xff") (const_int 0))
+           (and (match_test "INTVAL (operands[2]) == 0xff")
                 (match_operand 1 "ext_QIreg_operand" "")))
        (const_string "1")
        (const_string "*")))
   [(set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
          (and:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                 (match_operand:SI 2 "general_operand" "g"))))
+                 (match_operand:SI 2 "x86_64_general_operand" "rme"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (AND, SImode, operands)"
   "and{l}\t{%2, %k0|%k0, %2}"
   [(set (reg FLAGS_REG)
        (compare (and:SWI124
                  (match_operand:SWI124 1 "nonimmediate_operand" "%0,0")
-                 (match_operand:SWI124 2 "general_operand" "<g>,<r><i>"))
+                 (match_operand:SWI124 2 "<general_operand>" "<g>,<r><i>"))
                 (const_int 0)))
    (set (match_operand:SWI124 0 "nonimmediate_operand" "=<r>,<r>m")
        (and:SWI124 (match_dup 1) (match_dup 2)))]
   [(set (reg FLAGS_REG)
        (compare (and:SI
                  (match_operand:SI 1 "nonimmediate_operand" "%0")
-                 (match_operand:SI 2 "general_operand" "g"))
+                 (match_operand:SI 2 "x86_64_general_operand" "rme"))
                 (const_int 0)))
    (set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI (and:SI (match_dup 1) (match_dup 2))))]
                                            (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);")
+{
+  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 AND can be encoded with sign extended immediate, this is only
 ;; profitable when 7th bit is not set.
                   (and: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]);")
+{
+  operands[0] = gen_lowpart (QImode, operands[0]);
+  operands[1] = gen_lowpart (QImode, operands[1]);
+  operands[2] = gen_lowpart (QImode, operands[2]);
+})
 \f
 ;; Logical inclusive and exclusive OR instructions
 
   [(set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI
         (any_or:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                   (match_operand:SI 2 "general_operand" "g"))))
+                   (match_operand:SI 2 "x86_64_general_operand" "rme"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
   "<logic>{l}\t{%2, %k0|%k0, %2}"
 (define_insn "*<code>si_2_zext"
   [(set (reg FLAGS_REG)
        (compare (any_or:SI (match_operand:SI 1 "nonimmediate_operand" "%0")
-                           (match_operand:SI 2 "general_operand" "g"))
+                           (match_operand:SI 2 "x86_64_general_operand" "rme"))
                 (const_int 0)))
    (set (match_operand:DI 0 "register_operand" "=r")
        (zero_extend:DI (any_or:SI (match_dup 1) (match_dup 2))))]
                                               (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);")
+{
+  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 OR can be encoded with sign extended immediate, this is only
 ;; profitable when 7th bit is set.
                   (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]);")
+{
+  operands[0] = gen_lowpart (QImode, operands[0]);
+  operands[1] = gen_lowpart (QImode, operands[1]);
+  operands[2] = gen_lowpart (QImode, operands[2]);
+})
 
 (define_expand "xorqi_cc_ext_1"
   [(parallel [
                            (match_dup 0)))
    (set (match_dup 1)
        (if_then_else:SWI48 (ne (reg:CCZ FLAGS_REG) (const_int 0))
-                           (match_operand:SWI48 3 "register_operand" "r")
+                           (match_operand:SWI48 3 "register_operand" "")
                            (match_dup 1)))]
   "TARGET_CMOVE"
   "operands[4] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode));")
   [(set_attr "type" "ishift")
    (set_attr "mode" "<MODE>")])
 
+(define_insn "*bmi2_ashl<mode>3_1"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (ashift:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "rm")
+                     (match_operand:SWI48 2 "register_operand" "r")))]
+  "TARGET_BMI2"
+  "shlx\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "ishiftx")
+   (set_attr "mode" "<MODE>")])
+
 (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")))
+  [(set (match_operand:SWI48 0 "nonimmediate_operand" "=rm,r,r")
+       (ashift:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "0,l,rm")
+                     (match_operand:QI 2 "nonmemory_operand" "c<S>,M,r")))
    (clobber (reg:CC FLAGS_REG))]
   "ix86_binary_operator_ok (ASHIFT, <MODE>mode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_LEA:
+    case TYPE_ISHIFTX:
       return "#";
 
     case TYPE_ALU:
        return "sal{<imodesuffix>}\t{%2, %0|%0, %2}";
     }
 }
-  [(set (attr "type")
+  [(set_attr "isa" "*,*,bmi2")
+   (set (attr "type")
      (cond [(eq_attr "alternative" "1")
              (const_string "lea")
-            (and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
+           (eq_attr "alternative" "2")
+             (const_string "ishiftx")
+            (and (and (match_test "TARGET_DOUBLE_WITH_ADD")
                      (match_operand 0 "register_operand" ""))
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
 
+;; Convert shift to the shiftx pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:SWI48 0 "register_operand" "")
+       (ashift:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "")
+                     (match_operand:QI 2 "register_operand" "")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (ashift:SWI48 (match_dup 1) (match_dup 2)))]
+  "operands[2] = gen_lowpart (<MODE>mode, operands[2]);")
+
+(define_insn "*bmi2_ashlsi3_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (ashift:SI (match_operand:SI 1 "nonimmediate_operand" "rm")
+                    (match_operand:SI 2 "register_operand" "r"))))]
+  "TARGET_64BIT && TARGET_BMI2"
+  "shlx\t{%2, %1, %k0|%k0, %1, %2}"
+  [(set_attr "type" "ishiftx")
+   (set_attr "mode" "SI")])
+
 (define_insn "*ashlsi3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r,r")
+  [(set (match_operand:DI 0 "register_operand" "=r,r,r")
        (zero_extend:DI
-         (ashift:SI (match_operand:SI 1 "register_operand" "0,l")
-                    (match_operand:QI 2 "nonmemory_operand" "cI,M"))))
+         (ashift:SI (match_operand:SI 1 "nonimmediate_operand" "0,l,rm")
+                    (match_operand:QI 2 "nonmemory_operand" "cI,M,r"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (ASHIFT, SImode, operands)"
 {
   switch (get_attr_type (insn))
     {
     case TYPE_LEA:
+    case TYPE_ISHIFTX:
       return "#";
 
     case TYPE_ALU:
        return "sal{l}\t{%2, %k0|%k0, %2}";
     }
 }
-  [(set (attr "type")
+  [(set_attr "isa" "*,*,bmi2")
+   (set (attr "type")
      (cond [(eq_attr "alternative" "1")
              (const_string "lea")
-            (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                    (const_int 0))
+           (eq_attr "alternative" "2")
+             (const_string "ishiftx")
+            (and (match_test "TARGET_DOUBLE_WITH_ADD")
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
           ]
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "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")))
+;; Convert shift to the shiftx pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (zero_extend:DI
+         (ashift:SI (match_operand:SI 1 "nonimmediate_operand" "")
+                    (match_operand:QI 2 "register_operand" ""))))
    (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 (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")])
+  "TARGET_64BIT && TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (ashift:SI (match_dup 1) (match_dup 2))))]
+  "operands[2] = gen_lowpart (SImode, operands[2]);")
 
-(define_insn "*ashlhi3_1_lea"
-  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,r")
+(define_insn "*ashlhi3_1"
+  [(set (match_operand:HI 0 "nonimmediate_operand" "=rm,Yp")
        (ashift:HI (match_operand:HI 1 "nonimmediate_operand" "0,l")
                   (match_operand:QI 2 "nonmemory_operand" "cI,M")))
    (clobber (reg:CC FLAGS_REG))]
-  "!TARGET_PARTIAL_REG_STALL
-   && ix86_binary_operator_ok (ASHIFT, HImode, operands)"
+  "ix86_binary_operator_ok (ASHIFT, HImode, operands)"
 {
   switch (get_attr_type (insn))
     {
   [(set (attr "type")
      (cond [(eq_attr "alternative" "1")
              (const_string "lea")
-            (and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
+            (and (and (match_test "TARGET_DOUBLE_WITH_ADD")
                      (match_operand 0 "register_operand" ""))
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "HI,SI")])
 
+;; %%% Potential partial reg stall on alternative 1.  What to do?
 (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 (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 [(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")])
-
-;; %%% 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")
+  [(set (match_operand:QI 0 "nonimmediate_operand" "=qm,r,Yp")
        (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)"
+  "ix86_binary_operator_ok (ASHIFT, QImode, operands)"
 {
   switch (get_attr_type (insn))
     {
   [(set (attr "type")
      (cond [(eq_attr "alternative" "2")
              (const_string "lea")
-            (and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
+            (and (and (match_test "TARGET_DOUBLE_WITH_ADD")
                      (match_operand 0 "register_operand" ""))
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "QI,SI,SI")])
     }
 }
   [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
+     (cond [(and (and (match_test "TARGET_DOUBLE_WITH_ADD")
                      (match_operand 0 "register_operand" ""))
                 (match_operand 1 "const1_operand" ""))
              (const_string "alu")
        (ior (eq_attr "type" "alu")
            (and (eq_attr "type" "ishift1")
                 (and (match_operand 1 "const1_operand" "")
-                     (ne (symbol_ref "TARGET_SHIFT1 || optimize_function_for_size_p (cfun)")
-                         (const_int 0)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "QI")])
 
-;; Convert lea to the lea pattern to avoid flags dependency.
+;; Convert ashift 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
+  "GET_MODE (operands[0]) == GET_MODE (operands[1])
+   && reload_completed
    && true_regnum (operands[0]) != true_regnum (operands[1])"
   [(const_int 0)]
 {
-  rtx pat;
   enum machine_mode mode = GET_MODE (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]);
+  rtx pat;
 
   if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (SImode))
-    operands[0] = gen_lowpart (SImode, operands[0]);
+    { 
+      mode = SImode; 
+      operands[0] = gen_lowpart (mode, operands[0]);
+      operands[1] = gen_lowpart (mode, operands[1]);
+    }
+
+  operands[2] = gen_int_mode (1 << INTVAL (operands[2]), mode);
 
-  if (TARGET_64BIT && mode != Pmode)
-    pat = gen_rtx_SUBREG (SImode, pat, 0);
+  pat = gen_rtx_MULT (mode, operands[1], operands[2]);
 
   emit_insn (gen_rtx_SET (VOIDmode, operands[0], pat));
   DONE;
 })
 
-;; Convert lea to the lea pattern to avoid flags dependency.
+;; Convert ashift to the lea pattern to avoid flags dependency.
 (define_split
   [(set (match_operand:DI 0 "register_operand" "")
        (zero_extend:DI
     }
 }
   [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
+     (cond [(and (and (match_test "TARGET_DOUBLE_WITH_ADD")
                      (match_operand 0 "register_operand" ""))
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
     }
 }
   [(set (attr "type")
-     (cond [(and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                    (const_int 0))
+     (cond [(and (match_test "TARGET_DOUBLE_WITH_ADD")
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
           ]
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "SI")])
     }
 }
   [(set (attr "type")
-     (cond [(and (and (ne (symbol_ref "TARGET_DOUBLE_WITH_ADD")
-                         (const_int 0))
+     (cond [(and (and (match_test "TARGET_DOUBLE_WITH_ADD")
                      (match_operand 0 "register_operand" ""))
                 (match_operand 2 "const1_operand" ""))
              (const_string "alu")
        (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)))))
+                     (ior (match_test "TARGET_SHIFT1")
+                          (match_test "optimize_function_for_size_p (cfun)")))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
 
 ;; See comment above `ashl<mode>3' about how this works.
 
-(define_expand "<shiftrt_insn><mode>3"
+(define_expand "<shift_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;")
 
 ;; Avoid useless masking of count operand.
-(define_insn_and_split "*<shiftrt_insn><mode>3_mask"
+(define_insn_and_split "*<shift_insn><mode>3_mask"
   [(set (match_operand:SWI48 0 "nonimmediate_operand" "=rm")
        (any_shiftrt:SWI48
          (match_operand:SWI48 1 "nonimmediate_operand" "0")
   [(set_attr "type" "ishift")
    (set_attr "mode" "<MODE>")])
 
-(define_insn_and_split "*<shiftrt_insn><mode>3_doubleword"
+(define_insn_and_split "*<shift_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")))
   "#"
   "(optimize && flag_peephole2) ? epilogue_completed : reload_completed"
   [(const_int 0)]
-  "ix86_split_<shiftrt_insn> (operands, NULL_RTX, <MODE>mode); DONE;"
+  "ix86_split_<shift_insn> (operands, NULL_RTX, <MODE>mode); DONE;"
   [(set_attr "type" "multi")])
 
 ;; By default we don't ask for a scratch register, because when DWImode
    (match_dup 3)]
   "TARGET_CMOVE"
   [(const_int 0)]
-  "ix86_split_<shiftrt_insn> (operands, operands[3], <DWI>mode); DONE;")
+  "ix86_split_<shift_insn> (operands, operands[3], <DWI>mode); DONE;")
 
 (define_insn "x86_64_shrd"
   [(set (match_operand:DI 0 "nonimmediate_operand" "+r*m")
   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>")))
+(define_insn "*bmi2_<shift_insn><mode>3_1"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (any_shiftrt:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "rm")
+                          (match_operand:SWI48 2 "register_operand" "r")))]
+  "TARGET_BMI2"
+  "<shift>x\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "ishiftx")
+   (set_attr "mode" "<MODE>")])
+
+(define_insn "*<shift_insn><mode>3_1"
+  [(set (match_operand:SWI48 0 "nonimmediate_operand" "=rm,r")
+       (any_shiftrt:SWI48
+         (match_operand:SWI48 1 "nonimmediate_operand" "0,rm")
+         (match_operand:QI 2 "nonmemory_operand" "c<S>,r")))
    (clobber (reg:CC FLAGS_REG))]
   "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}";
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ISHIFTX:
+      return "#";
+
+    default:
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "<shift>{<imodesuffix>}\t%0";
+      else
+       return "<shift>{<imodesuffix>}\t{%2, %0|%0, %2}";
+    }
 }
-  [(set_attr "type" "ishift")
+  [(set_attr "isa" "*,bmi2")
+   (set_attr "type" "ishift,ishiftx")
    (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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
 
-(define_insn "*<shiftrt_insn>si3_1_zext"
+;; Convert shift to the shiftx pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:SWI48 0 "register_operand" "")
+       (any_shiftrt:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "")
+                          (match_operand:QI 2 "register_operand" "")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (any_shiftrt:SWI48 (match_dup 1) (match_dup 2)))]
+  "operands[2] = gen_lowpart (<MODE>mode, operands[2]);")
+
+(define_insn "*bmi2_<shift_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"))))
+         (any_shiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "rm")
+                         (match_operand:SI 2 "register_operand" "r"))))]
+  "TARGET_64BIT && TARGET_BMI2"
+  "<shift>x\t{%2, %1, %k0|%k0, %1, %2}"
+  [(set_attr "type" "ishiftx")
+   (set_attr "mode" "SI")])
+
+(define_insn "*<shift_insn>si3_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r,r")
+       (zero_extend:DI
+         (any_shiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "0,rm")
+                         (match_operand:QI 2 "nonmemory_operand" "cI,r"))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
 {
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ISHIFTX:
+      return "#";
+
+    default:
+      if (operands[2] == const1_rtx
+         && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
+       return "<shift>{l}\t%k0";
+      else
+       return "<shift>{l}\t{%2, %k0|%k0, %2}";
+    }
+}
+  [(set_attr "isa" "*,bmi2")
+   (set_attr "type" "ishift,ishiftx")
+   (set (attr "length_immediate")
+     (if_then_else
+       (and (match_operand 2 "const1_operand" "")
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
+       (const_string "0")
+       (const_string "*")))
+   (set_attr "mode" "SI")])
+
+;; Convert shift to the shiftx pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (zero_extend:DI
+         (any_shiftrt:SI (match_operand:SI 1 "nonimmediate_operand" "")
+                         (match_operand:QI 2 "register_operand" ""))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT && TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (any_shiftrt:SI (match_dup 1) (match_dup 2))))]
+  "operands[2] = gen_lowpart (SImode, operands[2]);")
+
+(define_insn "*<shift_insn><mode>3_1"
+  [(set (match_operand:SWI12 0 "nonimmediate_operand" "=<r>m")
+       (any_shiftrt:SWI12
+         (match_operand:SWI12 1 "nonimmediate_operand" "0")
+         (match_operand:QI 2 "nonmemory_operand" "c<S>")))
+   (clobber (reg:CC FLAGS_REG))]
+  "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
+{
   if (operands[2] == const1_rtx
       && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-    return "<shiftrt>{l}\t%k0";
+    return "<shift>{<imodesuffix>}\t%0";
   else
-    return "<shiftrt>{l}\t{%2, %k0|%k0, %2}";
+    return "<shift>{<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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (const_string "0")
        (const_string "*")))
-   (set_attr "mode" "SI")])
+   (set_attr "mode" "<MODE>")])
 
-(define_insn "*<shiftrt_insn>qi3_1_slp"
+(define_insn "*<shift_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")))
 {
   if (operands[1] == const1_rtx
       && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-    return "<shiftrt>{b}\t%0";
+    return "<shift>{b}\t%0";
   else
-    return "<shiftrt>{b}\t{%1, %0|%0, %1}";
+    return "<shift>{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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (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 "*<shiftrt_insn><mode>3_cmp"
+(define_insn "*<shift_insn><mode>3_cmp"
   [(set (reg FLAGS_REG)
        (compare
          (any_shiftrt:SWI
 {
   if (operands[2] == const1_rtx
       && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-    return "<shiftrt>{<imodesuffix>}\t%0";
+    return "<shift>{<imodesuffix>}\t%0";
   else
-    return "<shiftrt>{<imodesuffix>}\t{%2, %0|%0, %2}";
+    return "<shift>{<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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
 
-(define_insn "*<shiftrt_insn>si3_cmp_zext"
+(define_insn "*<shift_insn>si3_cmp_zext"
   [(set (reg FLAGS_REG)
        (compare
          (any_shiftrt:SI (match_operand:SI 1 "register_operand" "0")
 {
   if (operands[2] == const1_rtx
       && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-    return "<shiftrt>{l}\t%k0";
+    return "<shift>{l}\t%k0";
   else
-    return "<shiftrt>{l}\t{%2, %k0|%k0, %2}";
+    return "<shift>{l}\t{%2, %k0|%k0, %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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "SI")])
 
-(define_insn "*<shiftrt_insn><mode>3_cconly"
+(define_insn "*<shift_insn><mode>3_cconly"
   [(set (reg FLAGS_REG)
        (compare
          (any_shiftrt:SWI
 {
   if (operands[2] == const1_rtx
       && (TARGET_SHIFT1 || optimize_function_for_size_p (cfun)))
-    return "<shiftrt>{<imodesuffix>}\t%0";
+    return "<shift>{<imodesuffix>}\t%0";
   else
-    return "<shiftrt>{<imodesuffix>}\t{%2, %0|%0, %2}";
+    return "<shift>{<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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
   split_double_mode (<DWI>mode, &operands[0], 1, &operands[4], &operands[5]);
 })
 
+(define_insn "*bmi2_rorx<mode>3_1"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (rotatert:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "rm")
+                       (match_operand:QI 2 "immediate_operand" "<S>")))]
+  "TARGET_BMI2"
+  "rorx\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "rotatex")
+   (set_attr "mode" "<MODE>")])
+
 (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>")))
+  [(set (match_operand:SWI48 0 "nonimmediate_operand" "=rm,r")
+       (any_rotate:SWI48
+         (match_operand:SWI48 1 "nonimmediate_operand" "0,rm")
+         (match_operand:QI 2 "nonmemory_operand" "c<S>,<S>")))
    (clobber (reg:CC FLAGS_REG))]
   "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}";
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ROTATEX:
+      return "#";
+
+    default:
+      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 "isa" "*,bmi2")
+   (set_attr "type" "rotate,rotatex")
    (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)))
+       (and (eq_attr "type" "rotate")
+           (and (match_operand 2 "const1_operand" "")
+                (ior (match_test "TARGET_SHIFT1")
+                     (match_test "optimize_function_for_size_p (cfun)"))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "<MODE>")])
 
-(define_insn "*<rotate_insn>si3_1_zext"
-  [(set (match_operand:DI 0 "register_operand" "=r")
-       (zero_extend:DI
-         (any_rotate:SI (match_operand:SI 1 "register_operand" "0")
-                        (match_operand:QI 2 "nonmemory_operand" "cI"))))
+;; Convert rotate to the rotatex pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:SWI48 0 "register_operand" "")
+       (rotate:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "")
+                     (match_operand:QI 2 "immediate_operand" "")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+  "TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (rotatert:SWI48 (match_dup 1) (match_dup 2)))]
 {
-    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")
+  operands[2]
+    = GEN_INT (GET_MODE_BITSIZE (<MODE>mode) - INTVAL (operands[2]));
+})
+
+(define_split
+  [(set (match_operand:SWI48 0 "register_operand" "")
+       (rotatert:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "")
+                       (match_operand:QI 2 "immediate_operand" "")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (rotatert:SWI48 (match_dup 1) (match_dup 2)))])
+
+(define_insn "*bmi2_rorxsi3_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (rotatert:SI (match_operand:SI 1 "nonimmediate_operand" "rm")
+                      (match_operand:QI 2 "immediate_operand" "I"))))]
+  "TARGET_64BIT && TARGET_BMI2"
+  "rorx\t{%2, %1, %k0|%k0, %1, %2}"
+  [(set_attr "type" "rotatex")
+   (set_attr "mode" "SI")])
+
+(define_insn "*<rotate_insn>si3_1_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r,r")
+       (zero_extend:DI
+         (any_rotate:SI (match_operand:SI 1 "nonimmediate_operand" "0,rm")
+                        (match_operand:QI 2 "nonmemory_operand" "cI,I"))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT && ix86_binary_operator_ok (<CODE>, SImode, operands)"
+{
+  switch (get_attr_type (insn))
+    {
+    case TYPE_ROTATEX:
+      return "#";
+
+    default:
+      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 "isa" "*,bmi2")
+   (set_attr "type" "rotate,rotatex")
+   (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)))
+       (and (eq_attr "type" "rotate")
+           (and (match_operand 2 "const1_operand" "")
+                (ior (match_test "TARGET_SHIFT1")
+                     (match_test "optimize_function_for_size_p (cfun)"))))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "SI")])
 
+;; Convert rotate to the rotatex pattern to avoid flags dependency.
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (zero_extend:DI
+         (rotate:SI (match_operand:SI 1 "nonimmediate_operand" "")
+                    (match_operand:QI 2 "immediate_operand" ""))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT && TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (rotatert:SI (match_dup 1) (match_dup 2))))]
+{
+  operands[2]
+    = GEN_INT (GET_MODE_BITSIZE (SImode) - INTVAL (operands[2]));
+})
+
+(define_split
+  [(set (match_operand:DI 0 "register_operand" "")
+       (zero_extend:DI
+         (rotatert:SI (match_operand:SI 1 "nonimmediate_operand" "")
+                      (match_operand:QI 2 "immediate_operand" ""))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_64BIT && TARGET_BMI2 && reload_completed"
+  [(set (match_dup 0)
+       (zero_extend:DI (rotatert:SI (match_dup 1) (match_dup 2))))])
+
+(define_insn "*<rotate_insn><mode>3_1"
+  [(set (match_operand:SWI12 0 "nonimmediate_operand" "=<r>m")
+       (any_rotate:SWI12 (match_operand:SWI12 1 "nonimmediate_operand" "0")
+                         (match_operand:QI 2 "nonmemory_operand" "c<S>")))
+   (clobber (reg:CC FLAGS_REG))]
+  "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")
+     (if_then_else
+       (and (match_operand 2 "const1_operand" "")
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
+       (const_string "0")
+       (const_string "*")))
+   (set_attr "mode" "<MODE>")])
+
 (define_insn "*<rotate_insn>qi3_1_slp"
   [(set (strict_low_part (match_operand:QI 0 "nonimmediate_operand" "+qm"))
        (any_rotate:QI (match_dup 0)
    (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)))
+           (ior (match_test "TARGET_SHIFT1")
+                (match_test "optimize_function_for_size_p (cfun)")))
        (const_string "0")
        (const_string "*")))
    (set_attr "mode" "QI")])
          (zero_extract:SWI48
            (match_operand:SWI48 0 "register_operand" "r")
            (const_int 1)
-           (match_operand:SWI48 1 "nonmemory_operand" "rN"))
+           (match_operand:SWI48 1 "x86_64_nonmemory_operand" "rN"))
          (const_int 0)))]
   "TARGET_USE_BT || optimize_function_for_size_p (cfun)"
   "bt{<imodesuffix>}\t{%1, %0|%0, %1}"
        (if_then_else
          (match_operator 0 "ix86_swapped_fp_comparison_operator"
            [(match_operator 1 "float_operator"
-             [(match_operand:X87MODEI12 2 "nonimmediate_operand" "m,?r")])
+             [(match_operand:SWI24 2 "nonimmediate_operand" "m,?r")])
             (match_operand 3 "register_operand" "f,f")])
          (label_ref (match_operand 4 "" ""))
          (pc)))
        (if_then_else
          (match_operator 0 "ix86_swapped_fp_comparison_operator"
            [(match_operator 1 "float_operator"
-             [(match_operand:X87MODEI12 2 "memory_operand" "")])
+             [(match_operand:SWI24 2 "memory_operand" "")])
             (match_operand 3 "register_operand" "")])
          (match_operand 4 "" "")
          (match_operand 5 "" "")))
        (if_then_else
          (match_operator 0 "ix86_swapped_fp_comparison_operator"
            [(match_operator 1 "float_operator"
-             [(match_operand:X87MODEI12 2 "register_operand" "")])
+             [(match_operand:SWI24 2 "register_operand" "")])
             (match_operand 3 "register_operand" "")])
          (match_operand 4 "" "")
          (match_operand 5 "" "")))
    (set_attr "modrm" "0")])
 
 (define_expand "indirect_jump"
-  [(set (pc) (match_operand 0 "nonimmediate_operand" ""))]
-  ""
-  "")
+  [(set (pc) (match_operand 0 "indirect_branch_operand" ""))])
 
 (define_insn "*indirect_jump"
-  [(set (pc) (match_operand:P 0 "nonimmediate_operand" "rm"))]
+  [(set (pc) (match_operand:P 0 "indirect_branch_operand" "rw"))]
   ""
   "jmp\t%A0"
   [(set_attr "type" "ibr")
    (set_attr "length_immediate" "0")])
 
 (define_expand "tablejump"
-  [(parallel [(set (pc) (match_operand 0 "nonimmediate_operand" ""))
+  [(parallel [(set (pc) (match_operand 0 "indirect_branch_operand" ""))
              (use (label_ref (match_operand 1 "" "")))])]
   ""
 {
       operands[0] = expand_simple_binop (Pmode, code, op0, op1, NULL_RTX, 0,
                                         OPTAB_DIRECT);
     }
+  else if (TARGET_X32)
+    operands[0] = convert_memory_address (Pmode, operands[0]);
 })
 
 (define_insn "*tablejump_1"
-  [(set (pc) (match_operand:P 0 "nonimmediate_operand" "rm"))
+  [(set (pc) (match_operand:P 0 "indirect_branch_operand" "rw"))
    (use (label_ref (match_operand 1 "" "")))]
   ""
   "jmp\t%A0"
 ;; 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"
-  [(parallel [(call (match_operand:QI 0 "" "")
-                   (match_operand:SI 1 "" ""))
-             (set (reg:SI SP_REG)
-                  (plus:SI (reg:SI SP_REG)
-                           (match_operand:SI 3 "" "")))])]
-  "!TARGET_64BIT"
-{
-  ix86_expand_call (NULL, operands[0], operands[1],
-                   operands[2], operands[3], 0);
-  DONE;
-})
-
-(define_insn_and_split "*call_pop_0_vzeroupper"
-  [(parallel
-    [(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" "")))])
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "call")])
-
-(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" "")))]
-  "!TARGET_64BIT"
-{
-  if (SIBLING_CALL_P (insn))
-    return "jmp\t%P0";
-  else
-    return "call\t%P0";
-}
-  [(set_attr "type" "call")])
-
-(define_insn_and_split "*call_pop_1_vzeroupper"
-  [(parallel
-    [(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")))])
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "call")])
-
-(define_insn "*call_pop_1"
-  [(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 && !SIBLING_CALL_P (insn)"
-{
-  if (constant_call_address_operand (operands[0], Pmode))
-    return "call\t%P0";
-  return "call\t%A0";
-}
-  [(set_attr "type" "call")])
-
-(define_insn_and_split "*sibcall_pop_1_vzeroupper"
- [(parallel
-   [(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")))])
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "call")])
+;; Register constraint for call instruction.
+(define_mode_attr c [(SI "l") (DI "r")])
 
-(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")])
+;; Call subroutine returning no value.
 
 (define_expand "call"
   [(call (match_operand:QI 0 "" "")
    (use (match_operand 2 "" ""))]
   ""
 {
-  ix86_expand_call (NULL, operands[0], operands[1], operands[2], NULL, 0);
+  ix86_expand_call (NULL, operands[0], operands[1],
+                   operands[2], NULL, false);
   DONE;
 })
 
    (use (match_operand 2 "" ""))]
   ""
 {
-  ix86_expand_call (NULL, operands[0], operands[1], operands[2], NULL, 1);
+  ix86_expand_call (NULL, operands[0], operands[1],
+                   operands[2], NULL, true);
   DONE;
 })
 
-(define_insn_and_split "*call_0_vzeroupper"
-  [(call (mem:QI (match_operand 0 "constant_call_address_operand" ""))
+(define_insn_and_split "*call_vzeroupper"
+  [(call (mem:QI (match_operand:P 0 "call_insn_operand" "<c>zw"))
         (match_operand 1 "" ""))
    (unspec [(match_operand 2 "const_int_operand" "")]
           UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER"
+  "TARGET_VZEROUPPER && !SIBLING_CALL_P (insn)"
   "#"
   "&& reload_completed"
   [(const_int 0)]
   "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
   [(set_attr "type" "call")])
 
-(define_insn "*call_0"
-  [(call (mem:QI (match_operand 0 "constant_call_address_operand" ""))
+(define_insn "*call"
+  [(call (mem:QI (match_operand:P 0 "call_insn_operand" "<c>zw"))
         (match_operand 1 "" ""))]
-  ""
-  { return ix86_output_call_insn (insn, operands[0], 0); }
+  "!SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[0]);"
   [(set_attr "type" "call")])
 
-(define_insn_and_split "*call_1_vzeroupper"
-  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "lsm"))
+(define_insn_and_split "*call_rex64_ms_sysv_vzeroupper"
+  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rzw"))
         (match_operand 1 "" ""))
+   (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
+   (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))
    (unspec [(match_operand 2 "const_int_operand" "")]
           UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && !SIBLING_CALL_P (insn)"
+  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)"
   "#"
   "&& reload_completed"
   [(const_int 0)]
   "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
   [(set_attr "type" "call")])
 
-(define_insn "*call_1"
-  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "lsm"))
-        (match_operand 1 "" ""))]
-  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[0], 0); }
+(define_insn "*call_rex64_ms_sysv"
+  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rzw"))
+        (match_operand 1 "" ""))
+   (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
+   (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))]
+  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[0]);"
   [(set_attr "type" "call")])
 
-(define_insn_and_split "*sibcall_1_vzeroupper"
-  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "s,U"))
+(define_insn_and_split "*sibcall_vzeroupper"
+  [(call (mem:QI (match_operand:P 0 "sibcall_insn_operand" "Uz"))
         (match_operand 1 "" ""))
    (unspec [(match_operand 2 "const_int_operand" "")]
           UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "TARGET_VZEROUPPER && SIBLING_CALL_P (insn)"
   "#"
   "&& reload_completed"
   [(const_int 0)]
   "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
   [(set_attr "type" "call")])
 
-(define_insn "*sibcall_1"
-  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "s,U"))
+(define_insn "*sibcall"
+  [(call (mem:QI (match_operand:P 0 "sibcall_insn_operand" "Uz"))
         (match_operand 1 "" ""))]
-  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[0], 0); }
+  "SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[0]);"
   [(set_attr "type" "call")])
 
-(define_insn_and_split "*call_1_rex64_vzeroupper"
-  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rsm"))
-        (match_operand 1 "" ""))
-   (unspec [(match_operand 2 "const_int_operand" "")]
+(define_expand "call_pop"
+  [(parallel [(call (match_operand:QI 0 "" "")
+                   (match_operand:SI 1 "" ""))
+             (set (reg:SI SP_REG)
+                  (plus:SI (reg:SI SP_REG)
+                           (match_operand:SI 3 "" "")))])]
+  "!TARGET_64BIT"
+{
+  ix86_expand_call (NULL, operands[0], operands[1],
+                   operands[2], operands[3], false);
+  DONE;
+})
+
+(define_insn_and_split "*call_pop_vzeroupper"
+  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "lzm"))
+        (match_operand:SI 1 "" ""))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "i")))
+   (unspec [(match_operand 3 "const_int_operand" "")]
           UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)
-   && ix86_cmodel != CM_LARGE && ix86_cmodel != CM_LARGE_PIC"
+  "TARGET_VZEROUPPER && !TARGET_64BIT && !SIBLING_CALL_P (insn)"
   "#"
   "&& reload_completed"
   [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
+  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
   [(set_attr "type" "call")])
 
-(define_insn "*call_1_rex64"
-  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rsm"))
-        (match_operand 1 "" ""))]
-  "TARGET_64BIT && !SIBLING_CALL_P (insn)
-   && ix86_cmodel != CM_LARGE && ix86_cmodel != CM_LARGE_PIC"
-  { return ix86_output_call_insn (insn, operands[0], 0); }
+(define_insn "*call_pop"
+  [(call (mem:QI (match_operand:SI 0 "call_insn_operand" "lzm"))
+        (match_operand 1 "" ""))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "i")))]
+  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[0]);"
   [(set_attr "type" "call")])
 
-(define_insn_and_split "*call_1_rex64_ms_sysv_vzeroupper"
-  [(parallel
-    [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rsm"))
-          (match_operand 1 "" ""))
-     (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
-     (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))])
-   (unspec [(match_operand 2 "const_int_operand" "")]
+(define_insn_and_split "*sibcall_pop_vzeroupper"
+  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "Uz"))
+        (match_operand 1 "" ""))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "i")))
+   (unspec [(match_operand 3 "const_int_operand" "")]
           UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)"
+  "TARGET_VZEROUPPER && !TARGET_64BIT && SIBLING_CALL_P (insn)"
   "#"
   "&& reload_completed"
   [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
+  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
   [(set_attr "type" "call")])
 
-(define_insn "*call_1_rex64_ms_sysv"
-  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rsm"))
+(define_insn "*sibcall_pop"
+  [(call (mem:QI (match_operand:SI 0 "sibcall_insn_operand" "Uz"))
         (match_operand 1 "" ""))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 2 "immediate_operand" "i")))]
+  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[0]);"
+  [(set_attr "type" "call")])
+
+;; Call subroutine, returning value in operand 0
+
+(define_expand "call_value"
+  [(set (match_operand 0 "" "")
+       (call (match_operand:QI 1 "" "")
+             (match_operand 2 "" "")))
+   (use (match_operand 3 "" ""))]
+  ""
+{
+  ix86_expand_call (operands[0], operands[1], operands[2],
+                   operands[3], NULL, false);
+  DONE;
+})
+
+(define_expand "sibcall_value"
+  [(set (match_operand 0 "" "")
+       (call (match_operand:QI 1 "" "")
+             (match_operand 2 "" "")))
+   (use (match_operand 3 "" ""))]
+  ""
+{
+  ix86_expand_call (operands[0], operands[1], operands[2],
+                   operands[3], NULL, true);
+  DONE;
+})
+
+(define_insn_and_split "*call_value_vzeroupper"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:P 1 "call_insn_operand" "<c>zw"))
+             (match_operand 2 "" "")))
+   (unspec [(match_operand 3 "const_int_operand" "")]
+          UNSPEC_CALL_NEEDS_VZEROUPPER)]
+  "TARGET_VZEROUPPER && !SIBLING_CALL_P (insn)"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
+  [(set_attr "type" "callv")])
+
+(define_insn "*call_value"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:P 1 "call_insn_operand" "<c>zw"))
+             (match_operand 2 "" "")))]
+  "!SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[1]);"
+  [(set_attr "type" "callv")])
+
+(define_insn_and_split "*sibcall_value_vzeroupper"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:P 1 "sibcall_insn_operand" "Uz"))
+             (match_operand 2 "" "")))
+   (unspec [(match_operand 3 "const_int_operand" "")]
+          UNSPEC_CALL_NEEDS_VZEROUPPER)]
+  "TARGET_VZEROUPPER && SIBLING_CALL_P (insn)"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
+  [(set_attr "type" "callv")])
+
+(define_insn "*sibcall_value"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:P 1 "sibcall_insn_operand" "Uz"))
+             (match_operand 2 "" "")))]
+  "SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[1]);"
+  [(set_attr "type" "callv")])
+
+(define_insn_and_split "*call_value_rex64_ms_sysv_vzeroupper"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rzw"))
+             (match_operand 2 "" "")))
    (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
    (clobber (reg:TI XMM6_REG))
    (clobber (reg:TI XMM7_REG))
    (clobber (reg:TI XMM14_REG))
    (clobber (reg:TI XMM15_REG))
    (clobber (reg:DI SI_REG))
-   (clobber (reg:DI DI_REG))]
-  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[0], 0); }
-  [(set_attr "type" "call")])
-
-(define_insn_and_split "*call_1_rex64_large_vzeroupper"
-  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rm"))
-        (match_operand 1 "" ""))
-   (unspec [(match_operand 2 "const_int_operand" "")]
+   (clobber (reg:DI DI_REG))
+   (unspec [(match_operand 3 "const_int_operand" "")]
           UNSPEC_CALL_NEEDS_VZEROUPPER)]
   "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)"
   "#"
   "&& reload_completed"
   [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
-  [(set_attr "type" "call")])
+  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
+  [(set_attr "type" "callv")])
 
-(define_insn "*call_1_rex64_large"
-  [(call (mem:QI (match_operand:DI 0 "call_insn_operand" "rm"))
-        (match_operand 1 "" ""))]
+(define_insn "*call_value_rex64_ms_sysv"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rzw"))
+             (match_operand 2 "" "")))
+   (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
+   (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))]
   "TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[0], 0); }
-  [(set_attr "type" "call")])
-
-(define_insn_and_split "*sibcall_1_rex64_vzeroupper"
-  [(call (mem:QI (match_operand:DI 0 "sibcall_insn_operand" "s,U"))
-        (match_operand 1 "" ""))
-   (unspec [(match_operand 2 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[2]); DONE;"
-  [(set_attr "type" "call")])
-
-(define_insn "*sibcall_1_rex64"
-  [(call (mem:QI (match_operand:DI 0 "sibcall_insn_operand" "s,U"))
-        (match_operand 1 "" ""))]
-  "TARGET_64BIT && SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[0], 0); }
-  [(set_attr "type" "call")])
+  "* return ix86_output_call_insn (insn, operands[1]);"
+  [(set_attr "type" "callv")])
 
-;; Call subroutine, returning value in operand 0
 (define_expand "call_value_pop"
   [(parallel [(set (match_operand 0 "" "")
                   (call (match_operand:QI 1 "" "")
   "!TARGET_64BIT"
 {
   ix86_expand_call (operands[0], operands[1], operands[2],
-                   operands[3], operands[4], 0);
+                   operands[3], operands[4], false);
   DONE;
 })
 
-(define_expand "call_value"
+(define_insn_and_split "*call_value_pop_vzeroupper"
   [(set (match_operand 0 "" "")
-       (call (match_operand:QI 1 "" "")
-             (match_operand:SI 2 "" "")))
-   (use (match_operand:SI 3 "" ""))]
-  ;; Operand 3 is not used on the i386.
-  ""
-{
-  ix86_expand_call (operands[0], operands[1], operands[2],
-                   operands[3], NULL, 0);
-  DONE;
-})
+       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "lzm"))
+             (match_operand 2 "" "")))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "i")))
+   (unspec [(match_operand 4 "const_int_operand" "")]
+          UNSPEC_CALL_NEEDS_VZEROUPPER)]
+  "TARGET_VZEROUPPER && !TARGET_64BIT && !SIBLING_CALL_P (insn)"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+  "ix86_split_call_vzeroupper (curr_insn, operands[4]); DONE;"
+  [(set_attr "type" "callv")])
 
-(define_expand "sibcall_value"
+(define_insn "*call_value_pop"
   [(set (match_operand 0 "" "")
-       (call (match_operand:QI 1 "" "")
-             (match_operand:SI 2 "" "")))
-   (use (match_operand:SI 3 "" ""))]
-  ;; Operand 3 is not used on the i386.
-  ""
-{
-  ix86_expand_call (operands[0], operands[1], operands[2],
-                   operands[3], NULL, 1);
-  DONE;
-})
+       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "lzm"))
+             (match_operand 2 "" "")))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "i")))]
+  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[1]);"
+  [(set_attr "type" "callv")])
+
+(define_insn_and_split "*sibcall_value_pop_vzeroupper"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "Uz"))
+             (match_operand 2 "" "")))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "i")))
+   (unspec [(match_operand 4 "const_int_operand" "")]
+          UNSPEC_CALL_NEEDS_VZEROUPPER)]
+  "TARGET_VZEROUPPER && !TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "#"
+  "&& reload_completed"
+  [(const_int 0)]
+  "ix86_split_call_vzeroupper (curr_insn, operands[4]); DONE;"
+  [(set_attr "type" "callv")])
+
+(define_insn "*sibcall_value_pop"
+  [(set (match_operand 0 "" "")
+       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "Uz"))
+             (match_operand 2 "" "")))
+   (set (reg:SI SP_REG)
+       (plus:SI (reg:SI SP_REG)
+                (match_operand:SI 3 "immediate_operand" "i")))]
+  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
+  "* return ix86_output_call_insn (insn, operands[1]);"
+  [(set_attr "type" "callv")])
 
 ;; Call subroutine returning any type.
 
                                 : X86_64_MS_SSE_REGPARM_MAX)
                              : X86_32_SSE_REGPARM_MAX)
                             - 1),
-                   NULL, 0);
+                   NULL, false);
 
   for (i = 0; i < XVECLEN (operands[2], 0); i++)
     {
 ;; See comments for ix86_can_use_return_insn_p in i386.c.
 
 (define_expand "return"
-  [(return)]
+  [(simple_return)]
   "ix86_can_use_return_insn_p ()"
 {
+  ix86_maybe_emit_epilogue_vzeroupper ();
+  if (crtl->args.pops_args)
+    {
+      rtx popc = GEN_INT (crtl->args.pops_args);
+      emit_jump_insn (gen_simple_return_pop_internal (popc));
+      DONE;
+    }
+})
+
+;; We need to disable this for TARGET_SEH, as otherwise
+;; shrink-wrapped prologue gets enabled too.  This might exceed
+;; the maximum size of prologue in unwind information.
+
+(define_expand "simple_return"
+  [(simple_return)]
+  "!TARGET_SEH"
+{
+  ix86_maybe_emit_epilogue_vzeroupper ();
   if (crtl->args.pops_args)
     {
       rtx popc = GEN_INT (crtl->args.pops_args);
-      emit_jump_insn (gen_return_pop_internal (popc));
+      emit_jump_insn (gen_simple_return_pop_internal (popc));
       DONE;
     }
 })
 
-(define_insn "return_internal"
-  [(return)]
+(define_insn "simple_return_internal"
+  [(simple_return)]
   "reload_completed"
   "ret"
   [(set_attr "length" "1")
 ;; Used by x86_machine_dependent_reorg to avoid penalty on single byte RET
 ;; instruction Athlon and K8 have.
 
-(define_insn "return_internal_long"
-  [(return)
+(define_insn "simple_return_internal_long"
+  [(simple_return)
    (unspec [(const_int 0)] UNSPEC_REP)]
   "reload_completed"
   "rep\;ret"
    (set_attr "prefix_rep" "1")
    (set_attr "modrm" "0")])
 
-(define_insn "return_pop_internal"
-  [(return)
+(define_insn "simple_return_pop_internal"
+  [(simple_return)
    (use (match_operand:SI 0 "const_int_operand" ""))]
   "reload_completed"
   "ret\t%0"
    (set_attr "length_immediate" "2")
    (set_attr "modrm" "0")])
 
-(define_insn "return_indirect_internal"
-  [(return)
+(define_insn "simple_return_indirect_internal"
+  [(simple_return)
    (use (match_operand:SI 0 "register_operand" "r"))]
   "reload_completed"
   "jmp\t%A0"
        (unspec:DI
          [(label_ref (match_operand 1 "" ""))]
          UNSPEC_SET_GOT_OFFSET))]
-  "TARGET_64BIT"
+  "TARGET_LP64"
   "movabs{q}\t{$_GLOBAL_OFFSET_TABLE_-%l1, %0|%0, OFFSET FLAT:_GLOBAL_OFFSET_TABLE_-%l1}"
   [(set_attr "type" "imov")
    (set_attr "length_immediate" "0")
       (clobber (reg:CC FLAGS_REG))])]
   ""
 {
-  if (TARGET_ABM)
+  if (TARGET_LZCNT)
     {
-      emit_insn (gen_clz<mode>2_abm (operands[0], operands[1]));
+      emit_insn (gen_clz<mode>2_lzcnt (operands[0], operands[1]));
       DONE;
     }
   operands[2] = GEN_INT (GET_MODE_BITSIZE (<MODE>mode)-1);
 })
 
-(define_insn "clz<mode>2_abm"
+(define_insn "clz<mode>2_lzcnt"
   [(set (match_operand:SWI248 0 "register_operand" "=r")
        (clz:SWI248 (match_operand:SWI248 1 "nonimmediate_operand" "rm")))
    (clobber (reg:CC FLAGS_REG))]
-  "TARGET_ABM || TARGET_BMI"
+  "TARGET_LZCNT"
   "lzcnt{<imodesuffix>}\t{%1, %0|%0, %1}"
   [(set_attr "prefix_rep" "1")
    (set_attr "type" "bitmanip")
 
 (define_insn "bmi_bextr_<mode>"
   [(set (match_operand:SWI48 0 "register_operand" "=r")
-        (unspec:SWI48 [(match_operand:SWI48 1 "nonimmediate_operand" "rm")
-                       (match_operand:SWI48 2 "register_operand" "r")]
+        (unspec:SWI48 [(match_operand:SWI48 1 "register_operand" "r")
+                       (match_operand:SWI48 2 "nonimmediate_operand" "rm")]
                        UNSPEC_BEXTR))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_BMI"
   [(set_attr "type" "bitmanip")
    (set_attr "mode" "<MODE>")])
 
+;; BMI2 instructions.
+(define_insn "bmi2_bzhi_<mode>3"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+       (and:SWI48 (match_operand:SWI48 1 "register_operand" "r")
+                  (lshiftrt:SWI48 (const_int -1)
+                                  (match_operand:SWI48 2 "nonimmediate_operand" "rm"))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_BMI2"
+  "bzhi\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "bitmanip")
+   (set_attr "prefix" "vex")
+   (set_attr "mode" "<MODE>")])
+
+(define_insn "bmi2_pdep_<mode>3"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+        (unspec:SWI48 [(match_operand:SWI48 1 "register_operand" "r")
+                       (match_operand:SWI48 2 "nonimmediate_operand" "rm")]
+                       UNSPEC_PDEP))]
+  "TARGET_BMI2"
+  "pdep\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "bitmanip")
+   (set_attr "prefix" "vex")
+   (set_attr "mode" "<MODE>")])
+
+(define_insn "bmi2_pext_<mode>3"
+  [(set (match_operand:SWI48 0 "register_operand" "=r")
+        (unspec:SWI48 [(match_operand:SWI48 1 "register_operand" "r")
+                       (match_operand:SWI48 2 "nonimmediate_operand" "rm")]
+                       UNSPEC_PEXT))]
+  "TARGET_BMI2"
+  "pext\t{%2, %1, %0|%0, %1, %2}"
+  [(set_attr "type" "bitmanip")
+   (set_attr "prefix" "vex")
+   (set_attr "mode" "<MODE>")])
+
 ;; TBM instructions.
 (define_insn "tbm_bextri_<mode>"
   [(set (match_operand:SWI48 0 "register_operand" "=r")
   "xor{b}\t{%h0, %b0|%b0, %h0}"
   [(set_attr "length" "2")
    (set_attr "mode" "HI")])
+
 \f
 ;; Thread-local storage patterns for ELF.
 ;;
 
 (define_insn "*tls_global_dynamic_32_gnu"
   [(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))
+       (unspec:SI
+        [(match_operand:SI 1 "register_operand" "b")
+         (match_operand:SI 2 "tls_symbolic_operand" "")
+         (match_operand:SI 3 "constant_call_address_operand" "z")]
+        UNSPEC_TLS_GD))
    (clobber (match_scratch:SI 4 "=d"))
    (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"
+{
+  output_asm_insn
+    ("lea{l}\t{%a2@tlsgd(,%1,1), %0|%0, %a2@tlsgd[%1*1]}", operands);
+  if (TARGET_SUN_TLS)
+#ifdef HAVE_AS_IX86_TLSGDPLT
+    return "call\t%a2@tlsgdplt";
+#else
+    return "call\t%p3@plt";
+#endif
+  return "call\t%P3";
+}
   [(set_attr "type" "multi")
    (set_attr "length" "12")])
 
 (define_expand "tls_global_dynamic_32"
-  [(parallel [(set (match_operand:SI 0 "register_operand" "")
-                  (unspec:SI
-                   [(match_dup 2)
-                    (match_operand:SI 1 "tls_symbolic_operand" "")
-                    (match_dup 3)]
-                   UNSPEC_TLS_GD))
-             (clobber (match_scratch:SI 4 ""))
-             (clobber (match_scratch:SI 5 ""))
-             (clobber (reg:CC FLAGS_REG))])]
-  ""
-{
-  if (flag_pic)
-    operands[2] = pic_offset_table_rtx;
-  else
-    {
-      operands[2] = gen_reg_rtx (Pmode);
-      emit_insn (gen_set_got (operands[2]));
-    }
-  if (TARGET_GNU2_TLS)
-    {
-       emit_insn (gen_tls_dynamic_gnu2_32
-                 (operands[0], operands[1], operands[2]));
-       DONE;
-    }
-  operands[3] = ix86_tls_get_addr ();
-})
+  [(parallel
+    [(set (match_operand:SI 0 "register_operand" "")
+         (unspec:SI [(match_operand:SI 2 "register_operand" "")
+                     (match_operand:SI 1 "tls_symbolic_operand" "")
+                     (match_operand:SI 3 "constant_call_address_operand" "")]
+                    UNSPEC_TLS_GD))
+     (clobber (match_scratch:SI 4 ""))
+     (clobber (match_scratch:SI 5 ""))
+     (clobber (reg:CC FLAGS_REG))])])
 
 (define_insn "*tls_global_dynamic_64"
   [(set (match_operand:DI 0 "register_operand" "=a")
-       (call:DI (mem:QI (match_operand:DI 2 "call_insn_operand" ""))
-                (match_operand:DI 3 "" "")))
-   (unspec:DI [(match_operand:DI 1 "tls_symbolic_operand" "")]
+       (call:DI
+        (mem:QI (match_operand:DI 2 "constant_call_address_operand" "z"))
+        (match_operand:DI 3 "" "")))
+   (unspec:DI [(match_operand 1 "tls_symbolic_operand" "")]
              UNSPEC_TLS_GD)]
   "TARGET_64BIT"
-  { 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"; }
+{
+  if (!TARGET_X32)
+    fputs (ASM_BYTE "0x66\n", asm_out_file);
+  output_asm_insn
+    ("lea{q}\t{%a1@tlsgd(%%rip), %%rdi|rdi, %a1@tlsgd[rip]}", operands);
+  fputs (ASM_SHORT "0x6666\n", asm_out_file);
+  fputs ("\trex64\n", asm_out_file);
+  if (TARGET_SUN_TLS)
+    return "call\t%p2@plt";
+  return "call\t%P2";
+}
   [(set_attr "type" "multi")
-   (set_attr "length" "16")])
+   (set (attr "length")
+       (symbol_ref "TARGET_X32 ? 15 : 16"))])
 
 (define_expand "tls_global_dynamic_64"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (call:DI (mem:QI (match_dup 2)) (const_int 0)))
-             (unspec:DI [(match_operand:DI 1 "tls_symbolic_operand" "")]
-                        UNSPEC_TLS_GD)])]
-  ""
-{
-  if (TARGET_GNU2_TLS)
-    {
-       emit_insn (gen_tls_dynamic_gnu2_64
-                 (operands[0], operands[1]));
-       DONE;
-    }
-  operands[2] = ix86_tls_get_addr ();
-})
+  [(parallel
+    [(set (match_operand:DI 0 "register_operand" "")
+         (call:DI
+          (mem:QI (match_operand:DI 2 "constant_call_address_operand" ""))
+          (const_int 0)))
+     (unspec:DI [(match_operand 1 "tls_symbolic_operand" "")]
+               UNSPEC_TLS_GD)])])
 
 (define_insn "*tls_local_dynamic_base_32_gnu"
   [(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))
+       (unspec:SI
+        [(match_operand:SI 1 "register_operand" "b")
+         (match_operand:SI 2 "constant_call_address_operand" "z")]
+        UNSPEC_TLS_LD_BASE))
    (clobber (match_scratch:SI 3 "=d"))
    (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"
+{
+  output_asm_insn
+    ("lea{l}\t{%&@tlsldm(%1), %0|%0, %&@tlsldm[%1]}", operands);
+  if (TARGET_SUN_TLS)
+#ifdef HAVE_AS_IX86_TLSLDMPLT
+    return "call\t%&@tlsldmplt";
+#else
+    return "call\t%p2@plt";
+#endif
+  return "call\t%P2";
+}
   [(set_attr "type" "multi")
    (set_attr "length" "11")])
 
 (define_expand "tls_local_dynamic_base_32"
-  [(parallel [(set (match_operand:SI 0 "register_operand" "")
-                  (unspec:SI [(match_dup 1) (match_dup 2)]
-                             UNSPEC_TLS_LD_BASE))
-             (clobber (match_scratch:SI 3 ""))
-             (clobber (match_scratch:SI 4 ""))
-             (clobber (reg:CC FLAGS_REG))])]
-  ""
-{
-  if (flag_pic)
-    operands[1] = pic_offset_table_rtx;
-  else
-    {
-      operands[1] = gen_reg_rtx (Pmode);
-      emit_insn (gen_set_got (operands[1]));
-    }
-  if (TARGET_GNU2_TLS)
-    {
-       emit_insn (gen_tls_dynamic_gnu2_32
-                 (operands[0], ix86_tls_module_base (), operands[1]));
-       DONE;
-    }
-  operands[2] = ix86_tls_get_addr ();
-})
+  [(parallel
+     [(set (match_operand:SI 0 "register_operand" "")
+          (unspec:SI
+           [(match_operand:SI 1 "register_operand" "")
+            (match_operand:SI 2 "constant_call_address_operand" "")]
+           UNSPEC_TLS_LD_BASE))
+      (clobber (match_scratch:SI 3 ""))
+      (clobber (match_scratch:SI 4 ""))
+      (clobber (reg:CC FLAGS_REG))])])
 
 (define_insn "*tls_local_dynamic_base_64"
   [(set (match_operand:DI 0 "register_operand" "=a")
-       (call:DI (mem:QI (match_operand:DI 1 "call_insn_operand" ""))
-                (match_operand:DI 2 "" "")))
+       (call:DI
+        (mem:QI (match_operand:DI 1 "constant_call_address_operand" "z"))
+        (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"
+{
+  output_asm_insn
+    ("lea{q}\t{%&@tlsld(%%rip), %%rdi|rdi, %&@tlsld[rip]}", operands);
+  if (TARGET_SUN_TLS)
+    return "call\t%p1@plt";
+  return "call\t%P1";
+}
   [(set_attr "type" "multi")
    (set_attr "length" "12")])
 
 (define_expand "tls_local_dynamic_base_64"
-  [(parallel [(set (match_operand:DI 0 "register_operand" "")
-                  (call:DI (mem:QI (match_dup 1)) (const_int 0)))
-             (unspec:DI [(const_int 0)] UNSPEC_TLS_LD_BASE)])]
-  ""
-{
-  if (TARGET_GNU2_TLS)
-    {
-       emit_insn (gen_tls_dynamic_gnu2_64
-                 (operands[0], ix86_tls_module_base ()));
-       DONE;
-    }
-  operands[1] = ix86_tls_get_addr ();
-})
+  [(parallel
+     [(set (match_operand:DI 0 "register_operand" "")
+          (call:DI
+           (mem:QI (match_operand:DI 1 "constant_call_address_operand" ""))
+           (const_int 0)))
+      (unspec:DI [(const_int 0)] UNSPEC_TLS_LD_BASE)])])
 
 ;; Local dynamic of a single variable is a lose.  Show combine how
 ;; to convert that back to global dynamic.
 
 (define_insn_and_split "*tls_local_dynamic_32_once"
   [(set (match_operand:SI 0 "register_operand" "=a")
-       (plus:SI (unspec:SI [(match_operand:SI 1 "register_operand" "b")
-                            (match_operand:SI 2 "call_insn_operand" "")]
-                           UNSPEC_TLS_LD_BASE)
-                (const:SI (unspec:SI
-                           [(match_operand:SI 3 "tls_symbolic_operand" "")]
-                           UNSPEC_DTPOFF))))
+       (plus:SI
+        (unspec:SI [(match_operand:SI 1 "register_operand" "b")
+                    (match_operand:SI 2 "constant_call_address_operand" "z")]
+                   UNSPEC_TLS_LD_BASE)
+        (const:SI (unspec:SI
+                   [(match_operand:SI 3 "tls_symbolic_operand" "")]
+                   UNSPEC_DTPOFF))))
    (clobber (match_scratch:SI 4 "=d"))
    (clobber (match_scratch:SI 5 "=c"))
    (clobber (reg:CC FLAGS_REG))]
   ""
   "#"
   ""
-  [(parallel [(set (match_dup 0)
-                  (unspec:SI [(match_dup 1) (match_dup 3) (match_dup 2)]
-                             UNSPEC_TLS_GD))
-             (clobber (match_dup 4))
-             (clobber (match_dup 5))
-             (clobber (reg:CC FLAGS_REG))])])
+  [(parallel
+     [(set (match_dup 0)
+          (unspec:SI [(match_dup 1) (match_dup 3) (match_dup 2)]
+                     UNSPEC_TLS_GD))
+      (clobber (match_dup 4))
+      (clobber (match_dup 5))
+      (clobber (reg:CC FLAGS_REG))])])
 
 ;; Segment register for the thread base ptr load
 (define_mode_attr tp_seg [(SI "gs") (DI "fs")])
 
-;; Load and add the thread base pointer from %gs:0.
+;; Load and add the thread base pointer from %<tp_seg>:0.
+(define_insn "*load_tp_x32"
+  [(set (match_operand:SI 0 "register_operand" "=r")
+       (unspec:SI [(const_int 0)] UNSPEC_TP))]
+  "TARGET_X32"
+  "mov{l}\t{%%fs:0, %0|%0, DWORD PTR fs:0}"
+  [(set_attr "type" "imov")
+   (set_attr "modrm" "0")
+   (set_attr "length" "7")
+   (set_attr "memory" "load")
+   (set_attr "imm_disp" "false")])
+
+(define_insn "*load_tp_x32_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI (unspec:SI [(const_int 0)] UNSPEC_TP)))]
+  "TARGET_X32"
+  "mov{l}\t{%%fs:0, %k0|%k0, DWORD PTR fs:0}"
+  [(set_attr "type" "imov")
+   (set_attr "modrm" "0")
+   (set_attr "length" "7")
+   (set_attr "memory" "load")
+   (set_attr "imm_disp" "false")])
+
 (define_insn "*load_tp_<mode>"
   [(set (match_operand:P 0 "register_operand" "=r")
        (unspec:P [(const_int 0)] UNSPEC_TP))]
-  ""
+  "!TARGET_X32"
   "mov{<imodesuffix>}\t{%%<tp_seg>:0, %0|%0, <iptrsize> PTR <tp_seg>:0}"
   [(set_attr "type" "imov")
    (set_attr "modrm" "0")
    (set_attr "memory" "load")
    (set_attr "imm_disp" "false")])
 
+(define_insn "*add_tp_x32"
+  [(set (match_operand:SI 0 "register_operand" "=r")
+       (plus:SI (unspec:SI [(const_int 0)] UNSPEC_TP)
+                (match_operand:SI 1 "register_operand" "0")))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_X32"
+  "add{l}\t{%%fs:0, %0|%0, DWORD PTR fs:0}"
+  [(set_attr "type" "alu")
+   (set_attr "modrm" "0")
+   (set_attr "length" "7")
+   (set_attr "memory" "load")
+   (set_attr "imm_disp" "false")])
+
+(define_insn "*add_tp_x32_zext"
+  [(set (match_operand:DI 0 "register_operand" "=r")
+       (zero_extend:DI
+         (plus:SI (unspec:SI [(const_int 0)] UNSPEC_TP)
+                  (match_operand:SI 1 "register_operand" "0"))))
+   (clobber (reg:CC FLAGS_REG))]
+  "TARGET_X32"
+  "add{l}\t{%%fs:0, %k0|%k0, DWORD PTR fs:0}"
+  [(set_attr "type" "alu")
+   (set_attr "modrm" "0")
+   (set_attr "length" "7")
+   (set_attr "memory" "load")
+   (set_attr "imm_disp" "false")])
+
 (define_insn "*add_tp_<mode>"
   [(set (match_operand:P 0 "register_operand" "=r")
        (plus:P (unspec:P [(const_int 0)] UNSPEC_TP)
                (match_operand:P 1 "register_operand" "0")))
    (clobber (reg:CC FLAGS_REG))]
-  ""
+  "!TARGET_X32"
   "add{<imodesuffix>}\t{%%<tp_seg>:0, %0|%0, <iptrsize> PTR <tp_seg>:0}"
   [(set_attr "type" "alu")
    (set_attr "modrm" "0")
         UNSPEC_TLS_IE_SUN))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && TARGET_SUN_TLS"
-  "mov{q}\t{%%fs:0, %0|%0, QWORD PTR fs:0}\n\tadd{q}\t{%a1@gottpoff(%%rip), %0|%0, %a1@gottpoff[rip]}"
+{
+  output_asm_insn
+    ("mov{q}\t{%%fs:0, %0|%0, QWORD PTR fs:0}", operands);
+  return "add{q}\t{%a1@gottpoff(%%rip), %0|%0, %a1@gottpoff[rip]}";
+}
   [(set_attr "type" "multi")])
 
 ;; GNU2 TLS patterns can be split.
   ix86_tls_descriptor_calls_expanded_in_cfun = true;
 })
 
-(define_insn "*tls_dynamic_lea_32"
+(define_insn "*tls_dynamic_gnu2_lea_32"
   [(set (match_operand:SI 0 "register_operand" "=r")
        (plus:SI (match_operand:SI 1 "register_operand" "b")
                 (const:SI
    (set_attr "length" "6")
    (set_attr "length_address" "4")])
 
-(define_insn "*tls_dynamic_call_32"
+(define_insn "*tls_dynamic_gnu2_call_32"
   [(set (match_operand:SI 0 "register_operand" "=a")
        (unspec:SI [(match_operand:SI 1 "tls_symbolic_operand" "")
                    (match_operand:SI 2 "register_operand" "0")
 
 (define_expand "tls_dynamic_gnu2_64"
   [(set (match_dup 2)
-       (unspec:DI [(match_operand:DI 1 "tls_symbolic_operand" "")]
+       (unspec:DI [(match_operand 1 "tls_symbolic_operand" "")]
                   UNSPEC_TLSDESC))
    (parallel
     [(set (match_operand:DI 0 "register_operand" "")
   ix86_tls_descriptor_calls_expanded_in_cfun = true;
 })
 
-(define_insn "*tls_dynamic_lea_64"
+(define_insn "*tls_dynamic_gnu2_lea_64"
   [(set (match_operand:DI 0 "register_operand" "=r")
-       (unspec:DI [(match_operand:DI 1 "tls_symbolic_operand" "")]
+       (unspec:DI [(match_operand 1 "tls_symbolic_operand" "")]
                   UNSPEC_TLSDESC))]
   "TARGET_64BIT && TARGET_GNU2_TLS"
   "lea{q}\t{%a1@TLSDESC(%%rip), %0|%0, %a1@TLSDESC[rip]}"
    (set_attr "length" "7")
    (set_attr "length_address" "4")])
 
-(define_insn "*tls_dynamic_call_64"
+(define_insn "*tls_dynamic_gnu2_call_64"
   [(set (match_operand:DI 0 "register_operand" "=a")
-       (unspec:DI [(match_operand:DI 1 "tls_symbolic_operand" "")
+       (unspec:DI [(match_operand 1 "tls_symbolic_operand" "")
                    (match_operand:DI 2 "register_operand" "0")
                    (reg:DI SP_REG)]
                   UNSPEC_TLSDESC))
                     (reg:DI SP_REG)]
                    UNSPEC_TLSDESC)
         (const:DI (unspec:DI
-                   [(match_operand:DI 1 "tls_symbolic_operand" "")]
+                   [(match_operand 1 "tls_symbolic_operand" "")]
                    UNSPEC_DTPOFF))))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_64BIT && TARGET_GNU2_TLS"
           (if_then_else (match_operand:MODEF 3 "mult_operator" "")
              (const_string "fmul")
              (const_string "fop"))))
-   (set_attr "isa" "base,noavx,avx")
+   (set_attr "isa" "*,noavx,avx")
    (set_attr "prefix" "orig,orig,vex")
    (set_attr "mode" "<MODE>")])
 
                  (const_string "fdiv")
               ]
               (const_string "fop")))
-   (set_attr "isa" "base,base,noavx,avx")
+   (set_attr "isa" "*,*,noavx,avx")
    (set_attr "prefix" "orig,orig,orig,vex")
    (set_attr "mode" "<MODE>")])
 
   [(set (match_operand:MODEF 0 "register_operand" "=f,f")
        (match_operator:MODEF 3 "binary_fp_operator"
          [(float:MODEF
-            (match_operand:X87MODEI12 1 "nonimmediate_operand" "m,?r"))
+            (match_operand:SWI24 1 "nonimmediate_operand" "m,?r"))
           (match_operand:MODEF 2 "register_operand" "0,0")]))]
-  "TARGET_80387 && X87_ENABLE_FLOAT (<MODEF:MODE>mode, <X87MODEI12:MODE>mode)
+  "TARGET_80387 && X87_ENABLE_FLOAT (<MODEF:MODE>mode, <SWI24:MODE>mode)
    && !(SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH)
-   && (TARGET_USE_<X87MODEI12:MODE>MODE_FIOP || optimize_function_for_size_p (cfun))"
+   && (TARGET_USE_<SWI24:MODE>MODE_FIOP || optimize_function_for_size_p (cfun))"
   "* return which_alternative ? \"#\" : output_387_binary_op (insn, operands);"
   [(set (attr "type")
         (cond [(match_operand:MODEF 3 "mult_operator" "")
               ]
               (const_string "fop")))
    (set_attr "fp_int_src" "true")
-   (set_attr "mode" "<X87MODEI12:MODE>")])
+   (set_attr "mode" "<SWI24:MODE>")])
 
 (define_insn "*fop_<MODEF:mode>_3_i387"
   [(set (match_operand:MODEF 0 "register_operand" "=f,f")
        (match_operator:MODEF 3 "binary_fp_operator"
          [(match_operand:MODEF 1 "register_operand" "0,0")
           (float:MODEF
-            (match_operand:X87MODEI12 2 "nonimmediate_operand" "m,?r"))]))]
-  "TARGET_80387 && X87_ENABLE_FLOAT (<MODEF:MODE>mode, <X87MODEI12:MODE>mode)
+            (match_operand:SWI24 2 "nonimmediate_operand" "m,?r"))]))]
+  "TARGET_80387 && X87_ENABLE_FLOAT (<MODEF:MODE>mode, <SWI24:MODE>mode)
    && !(SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH)
-   && (TARGET_USE_<X87MODEI12:MODE>MODE_FIOP || optimize_function_for_size_p (cfun))"
+   && (TARGET_USE_<SWI24:MODE>MODE_FIOP || optimize_function_for_size_p (cfun))"
   "* return which_alternative ? \"#\" : output_387_binary_op (insn, operands);"
   [(set (attr "type")
         (cond [(match_operand:MODEF 3 "mult_operator" "")
   [(set (match_operand:XF 0 "register_operand" "=f,f")
        (match_operator:XF 3 "binary_fp_operator"
          [(float:XF
-            (match_operand:X87MODEI12 1 "nonimmediate_operand" "m,?r"))
+            (match_operand:SWI24 1 "nonimmediate_operand" "m,?r"))
           (match_operand:XF 2 "register_operand" "0,0")]))]
   "TARGET_80387 && (TARGET_USE_<MODE>MODE_FIOP || optimize_function_for_size_p (cfun))"
   "* return which_alternative ? \"#\" : output_387_binary_op (insn, operands);"
        (match_operator:XF 3 "binary_fp_operator"
          [(match_operand:XF 1 "register_operand" "0,0")
           (float:XF
-            (match_operand:X87MODEI12 2 "nonimmediate_operand" "m,?r"))]))]
+            (match_operand:SWI24 2 "nonimmediate_operand" "m,?r"))]))]
   "TARGET_80387 && (TARGET_USE_<MODE>MODE_FIOP || optimize_function_for_size_p (cfun))"
   "* return which_alternative ? \"#\" : output_387_binary_op (insn, operands);"
   [(set (attr "type")
 (define_split
   [(set (match_operand 0 "register_operand" "")
        (match_operator 3 "binary_fp_operator"
-          [(float (match_operand:X87MODEI12 1 "register_operand" ""))
+          [(float (match_operand:SWI24 1 "register_operand" ""))
            (match_operand 2 "register_operand" "")]))]
   "reload_completed
    && X87_FLOAT_MODE_P (GET_MODE (operands[0]))
   [(set (match_operand 0 "register_operand" "")
        (match_operator 3 "binary_fp_operator"
           [(match_operand 1 "register_operand" "")
-           (float (match_operand:X87MODEI12 2 "register_operand" ""))]))]
+           (float (match_operand:SWI24 2 "register_operand" ""))]))]
   "reload_completed
    && X87_FLOAT_MODE_P (GET_MODE (operands[0]))
    && X87_ENABLE_FLOAT (GET_MODE (operands[0]), GET_MODE (operands[2]))"
    || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)"
 {
   if (<MODE>mode == SFmode
-      && TARGET_SSE_MATH && TARGET_RECIP && !optimize_function_for_size_p (cfun)
+      && TARGET_SSE_MATH
+      && TARGET_RECIP_SQRT
+      && !optimize_function_for_size_p (cfun)
       && flag_finite_math_only && !flag_trapping_math
       && flag_unsafe_math_optimizations)
     {
    (set (match_operand:XF 1 "register_operand" "")
        (unspec:XF [(match_dup 2)] UNSPEC_SINCOS_SIN))]
   "find_regno_note (insn, REG_UNUSED, REGNO (operands[0]))
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   [(set (match_dup 1) (unspec:XF [(match_dup 2)] UNSPEC_SIN))])
 
 (define_split
    (set (match_operand:XF 1 "register_operand" "")
        (unspec:XF [(match_dup 2)] UNSPEC_SINCOS_SIN))]
   "find_regno_note (insn, REG_UNUSED, REGNO (operands[1]))
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   [(set (match_dup 0) (unspec:XF [(match_dup 2)] UNSPEC_COS))])
 
 (define_insn "sincos_extend<mode>xf3_i387"
    (set (match_operand:XF 1 "register_operand" "")
        (unspec:XF [(float_extend:XF (match_dup 2))] UNSPEC_SINCOS_SIN))]
   "find_regno_note (insn, REG_UNUSED, REGNO (operands[0]))
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   [(set (match_dup 1)
        (unspec:XF [(float_extend:XF (match_dup 2))] UNSPEC_SIN))])
 
    (set (match_operand:XF 1 "register_operand" "")
        (unspec:XF [(float_extend:XF (match_dup 2))] UNSPEC_SINCOS_SIN))]
   "find_regno_note (insn, REG_UNUSED, REGNO (operands[1]))
-   && !(reload_completed || reload_in_progress)"
+   && can_create_pseudo_p ()"
   [(set (match_dup 0)
        (unspec:XF [(float_extend:XF (match_dup 2))] UNSPEC_COS))])
 
   if (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
       && !flag_trapping_math)
     {
-      if (!TARGET_ROUND && optimize_insn_for_size_p ())
-       FAIL;
       if (TARGET_ROUND)
        emit_insn (gen_sse4_1_round<mode>2
                   (operands[0], operands[1], GEN_INT (ROUND_MXCSR)));
+      else if (optimize_insn_for_size_p ())
+        FAIL;
       else
-       ix86_expand_rint (operand0, operand1);
+       ix86_expand_rint (operands[0], operands[1]);
     }
   else
     {
 })
 
 (define_expand "round<mode>2"
-  [(match_operand:MODEF 0 "register_operand" "")
-   (match_operand:MODEF 1 "nonimmediate_operand" "")]
-  "SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
-   && !flag_trapping_math && !flag_rounding_math"
+  [(match_operand:X87MODEF 0 "register_operand" "")
+   (match_operand:X87MODEF 1 "nonimmediate_operand" "")]
+  "(TARGET_USE_FANCY_MATH_387
+    && (!(SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH)
+       || TARGET_MIX_SSE_I387)
+    && flag_unsafe_math_optimizations)
+   || (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
+       && !flag_trapping_math && !flag_rounding_math)"
 {
   if (optimize_insn_for_size_p ())
     FAIL;
-  if (TARGET_64BIT || (<MODE>mode != DFmode))
-    ix86_expand_round (operand0, operand1);
+
+  if (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
+      && !flag_trapping_math && !flag_rounding_math)
+    {
+      if (TARGET_ROUND)
+        {
+         operands[1] = force_reg (<MODE>mode, operands[1]);
+         ix86_expand_round_sse4 (operands[0], operands[1]);
+       }
+      else if (TARGET_64BIT || (<MODE>mode != DFmode))
+       ix86_expand_round (operands[0], operands[1]);
+      else
+       ix86_expand_rounddf_32 (operands[0], operands[1]);
+    }
   else
-    ix86_expand_rounddf_32 (operand0, operand1);
+    {
+      operands[1] = force_reg (<MODE>mode, operands[1]);
+      ix86_emit_i387_round (operands[0], operands[1]);
+    }
   DONE;
 })
 
                   UNSPEC_FIST))
    (clobber (match_scratch:XF 2 "=&1f"))]
   "TARGET_USE_FANCY_MATH_387"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fpspc")
    (set_attr "mode" "DI")])
 
              (clobber (match_dup 3))])])
 
 (define_insn_and_split "*fist<mode>2_1"
-  [(set (match_operand:X87MODEI12 0 "register_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-                          UNSPEC_FIST))]
+  [(set (match_operand:SWI24 0 "register_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST))]
   "TARGET_USE_FANCY_MATH_387
    && can_create_pseudo_p ()"
   "#"
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fist<mode>2"
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "=m")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "f")]
-                          UNSPEC_FIST))]
+  [(set (match_operand:SWI24 0 "memory_operand" "=m")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "f")]
+                     UNSPEC_FIST))]
   "TARGET_USE_FANCY_MATH_387"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fpspc")
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fist<mode>2_with_temp"
-  [(set (match_operand:X87MODEI12 0 "register_operand" "=r")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "f")]
-                          UNSPEC_FIST))
-   (clobber (match_operand:X87MODEI12 2 "memory_operand" "=m"))]
+  [(set (match_operand:SWI24 0 "register_operand" "=r")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "f")]
+                     UNSPEC_FIST))
+   (clobber (match_operand:SWI24 2 "memory_operand" "=m"))]
   "TARGET_USE_FANCY_MATH_387"
   "#"
   [(set_attr "type" "fpspc")
    (set_attr "mode" "<MODE>")])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "register_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-                          UNSPEC_FIST))
-   (clobber (match_operand:X87MODEI12 2 "memory_operand" ""))]
+  [(set (match_operand:SWI24 0 "register_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST))
+   (clobber (match_operand:SWI24 2 "memory_operand" ""))]
   "reload_completed"
-  [(set (match_dup 2) (unspec:X87MODEI12 [(match_dup 1)] UNSPEC_FIST))
+  [(set (match_dup 2) (unspec:SWI24 [(match_dup 1)] UNSPEC_FIST))
    (set (match_dup 0) (match_dup 2))])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-                          UNSPEC_FIST))
-   (clobber (match_operand:X87MODEI12 2 "memory_operand" ""))]
+  [(set (match_operand:SWI24 0 "memory_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST))
+   (clobber (match_operand:SWI24 2 "memory_operand" ""))]
   "reload_completed"
-  [(set (match_dup 0) (unspec:X87MODEI12 [(match_dup 1)] UNSPEC_FIST))])
+  [(set (match_dup 0) (unspec:SWI24 [(match_dup 1)] UNSPEC_FIST))])
 
 (define_expand "lrintxf<mode>2"
-  [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-     (unspec:X87MODEI [(match_operand:XF 1 "register_operand" "")]
-                     UNSPEC_FIST))]
+  [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+     (unspec:SWI248x [(match_operand:XF 1 "register_operand" "")]
+                    UNSPEC_FIST))]
   "TARGET_USE_FANCY_MATH_387")
 
-(define_expand "lrint<MODEF:mode><SSEMODEI24:mode>2"
-  [(set (match_operand:SSEMODEI24 0 "nonimmediate_operand" "")
-     (unspec:SSEMODEI24 [(match_operand:MODEF 1 "register_operand" "")]
+(define_expand "lrint<MODEF:mode><SWI48x:mode>2"
+  [(set (match_operand:SWI48x 0 "nonimmediate_operand" "")
+     (unspec:SWI48x [(match_operand:MODEF 1 "register_operand" "")]
                        UNSPEC_FIX_NOTRUNC))]
   "SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
-   && ((<SSEMODEI24:MODE>mode != DImode) || TARGET_64BIT)")
+   && ((<SWI48x:MODE>mode != DImode) || TARGET_64BIT)")
 
-(define_expand "lround<MODEF:mode><SSEMODEI24:mode>2"
-  [(match_operand:SSEMODEI24 0 "nonimmediate_operand" "")
-   (match_operand:MODEF 1 "register_operand" "")]
-  "SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
-   && ((<SSEMODEI24:MODE>mode != DImode) || TARGET_64BIT)
-   && !flag_trapping_math && !flag_rounding_math"
+(define_expand "lround<X87MODEF:mode><SWI248x:mode>2"
+  [(match_operand:SWI248x 0 "nonimmediate_operand" "")
+   (match_operand:X87MODEF 1 "register_operand" "")]
+  "(TARGET_USE_FANCY_MATH_387
+    && (!(SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH)
+       || TARGET_MIX_SSE_I387)
+    && flag_unsafe_math_optimizations)
+   || (SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH
+       && <SWI248x:MODE>mode != HImode 
+       && ((<SWI248x:MODE>mode != DImode) || TARGET_64BIT)
+       && !flag_trapping_math && !flag_rounding_math)"
 {
   if (optimize_insn_for_size_p ())
     FAIL;
-  ix86_expand_lround (operand0, operand1);
+
+  if (SSE_FLOAT_MODE_P (<X87MODEF:MODE>mode) && TARGET_SSE_MATH
+      && <SWI248x:MODE>mode != HImode
+      && ((<SWI248x:MODE>mode != DImode) || TARGET_64BIT)
+      && !flag_trapping_math && !flag_rounding_math)
+    ix86_expand_lround (operands[0], operands[1]);
+  else
+    ix86_emit_i387_round (operands[0], operands[1]);
   DONE;
 })
 
        && !flag_trapping_math)"
 {
   if (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
-      && !flag_trapping_math
-      && (TARGET_ROUND || optimize_insn_for_speed_p ()))
+      && !flag_trapping_math)
     {
-      if (!TARGET_ROUND && optimize_insn_for_size_p ())
-       FAIL;
       if (TARGET_ROUND)
        emit_insn (gen_sse4_1_round<mode>2
                   (operands[0], operands[1], GEN_INT (ROUND_FLOOR)));
+      else if (optimize_insn_for_size_p ())
+        FAIL;
       else if (TARGET_64BIT || (<MODE>mode != DFmode))
-       ix86_expand_floorceil (operand0, operand1, true);
+       ix86_expand_floorceil (operands[0], operands[1], true);
       else
-       ix86_expand_floorceildf_32 (operand0, operand1, true);
+       ix86_expand_floorceildf_32 (operands[0], operands[1], true);
     }
   else
     {
 })
 
 (define_insn_and_split "*fist<mode>2_floor_1"
-  [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-       (unspec:X87MODEI [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_FLOOR))
+  [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+       (unspec:SWI248x [(match_operand:XF 1 "register_operand" "")]
+                       UNSPEC_FIST_FLOOR))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
 (define_insn "fistdi2_floor"
   [(set (match_operand:DI 0 "memory_operand" "=m")
        (unspec:DI [(match_operand:XF 1 "register_operand" "f")]
-        UNSPEC_FIST_FLOOR))
+                  UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" "m"))
    (use (match_operand:HI 3 "memory_operand" "m"))
    (clobber (match_scratch:XF 4 "=&1f"))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fistp")
    (set_attr "i387_cw" "floor")
    (set_attr "mode" "DI")])
 (define_insn "fistdi2_floor_with_temp"
   [(set (match_operand:DI 0 "nonimmediate_operand" "=m,?r")
        (unspec:DI [(match_operand:XF 1 "register_operand" "f,f")]
-        UNSPEC_FIST_FLOOR))
+                  UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" "m,m"))
    (use (match_operand:HI 3 "memory_operand" "m,m"))
    (clobber (match_operand:DI 4 "memory_operand" "=X,m"))
 (define_split
   [(set (match_operand:DI 0 "register_operand" "")
        (unspec:DI [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_FLOOR))
+                  UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
    (clobber (match_operand:DI 4 "memory_operand" ""))
    (clobber (match_scratch 5 ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 4) (unspec:DI [(match_dup 1)] UNSPEC_FIST_FLOOR))
+  [(parallel [(set (match_dup 4)
+                  (unspec:DI [(match_dup 1)] UNSPEC_FIST_FLOOR))
              (use (match_dup 2))
              (use (match_dup 3))
              (clobber (match_dup 5))])
 (define_split
   [(set (match_operand:DI 0 "memory_operand" "")
        (unspec:DI [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_FLOOR))
+                  UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
    (clobber (match_operand:DI 4 "memory_operand" ""))
    (clobber (match_scratch 5 ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 0) (unspec:DI [(match_dup 1)] UNSPEC_FIST_FLOOR))
+  [(parallel [(set (match_dup 0)
+                  (unspec:DI [(match_dup 1)] UNSPEC_FIST_FLOOR))
              (use (match_dup 2))
              (use (match_dup 3))
              (clobber (match_dup 5))])])
 
 (define_insn "fist<mode>2_floor"
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "=m")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "f")]
-        UNSPEC_FIST_FLOOR))
+  [(set (match_operand:SWI24 0 "memory_operand" "=m")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "f")]
+                     UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" "m"))
    (use (match_operand:HI 3 "memory_operand" "m"))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fistp")
    (set_attr "i387_cw" "floor")
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fist<mode>2_floor_with_temp"
-  [(set (match_operand:X87MODEI12 0 "nonimmediate_operand" "=m,?r")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "f,f")]
-        UNSPEC_FIST_FLOOR))
+  [(set (match_operand:SWI24 0 "nonimmediate_operand" "=m,?r")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "f,f")]
+                     UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" "m,m"))
    (use (match_operand:HI 3 "memory_operand" "m,m"))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" "=X,m"))]
+   (clobber (match_operand:SWI24 4 "memory_operand" "=X,m"))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations"
   "#"
    (set_attr "mode" "<MODE>")])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "register_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_FLOOR))
+  [(set (match_operand:SWI24 0 "register_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" ""))]
+   (clobber (match_operand:SWI24 4 "memory_operand" ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 4) (unspec:X87MODEI12 [(match_dup 1)]
-                                 UNSPEC_FIST_FLOOR))
+  [(parallel [(set (match_dup 4)
+                  (unspec:SWI24 [(match_dup 1)] UNSPEC_FIST_FLOOR))
              (use (match_dup 2))
              (use (match_dup 3))])
    (set (match_dup 0) (match_dup 4))])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_FLOOR))
+  [(set (match_operand:SWI24 0 "memory_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST_FLOOR))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" ""))]
+   (clobber (match_operand:SWI24 4 "memory_operand" ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 0) (unspec:X87MODEI12 [(match_dup 1)]
-                                 UNSPEC_FIST_FLOOR))
+  [(parallel [(set (match_dup 0)
+                  (unspec:SWI24 [(match_dup 1)] UNSPEC_FIST_FLOOR))
              (use (match_dup 2))
              (use (match_dup 3))])])
 
 (define_expand "lfloorxf<mode>2"
-  [(parallel [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-                  (unspec:X87MODEI [(match_operand:XF 1 "register_operand" "")]
-                   UNSPEC_FIST_FLOOR))
+  [(parallel [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+                  (unspec:SWI248x [(match_operand:XF 1 "register_operand" "")]
+                                  UNSPEC_FIST_FLOOR))
              (clobber (reg:CC FLAGS_REG))])]
   "TARGET_USE_FANCY_MATH_387
    && (!TARGET_SSE_MATH || TARGET_MIX_SSE_I387)
 {
   if (TARGET_64BIT && optimize_insn_for_size_p ())
     FAIL;
-  ix86_expand_lfloorceil (operand0, operand1, true);
+  ix86_expand_lfloorceil (operands[0], operands[1], true);
   DONE;
 })
 
        && !flag_trapping_math)"
 {
   if (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
-      && !flag_trapping_math
-      && (TARGET_ROUND || optimize_insn_for_speed_p ()))
+      && !flag_trapping_math)
     {
       if (TARGET_ROUND)
        emit_insn (gen_sse4_1_round<mode>2
       else if (optimize_insn_for_size_p ())
        FAIL;
       else if (TARGET_64BIT || (<MODE>mode != DFmode))
-       ix86_expand_floorceil (operand0, operand1, false);
+       ix86_expand_floorceil (operands[0], operands[1], false);
       else
-       ix86_expand_floorceildf_32 (operand0, operand1, false);
+       ix86_expand_floorceildf_32 (operands[0], operands[1], false);
     }
   else
     {
 })
 
 (define_insn_and_split "*fist<mode>2_ceil_1"
-  [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-       (unspec:X87MODEI [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_CEIL))
+  [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+       (unspec:SWI248x [(match_operand:XF 1 "register_operand" "")]
+                       UNSPEC_FIST_CEIL))
    (clobber (reg:CC FLAGS_REG))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations
 (define_insn "fistdi2_ceil"
   [(set (match_operand:DI 0 "memory_operand" "=m")
        (unspec:DI [(match_operand:XF 1 "register_operand" "f")]
-        UNSPEC_FIST_CEIL))
+                  UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" "m"))
    (use (match_operand:HI 3 "memory_operand" "m"))
    (clobber (match_scratch:XF 4 "=&1f"))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fistp")
    (set_attr "i387_cw" "ceil")
    (set_attr "mode" "DI")])
 (define_insn "fistdi2_ceil_with_temp"
   [(set (match_operand:DI 0 "nonimmediate_operand" "=m,?r")
        (unspec:DI [(match_operand:XF 1 "register_operand" "f,f")]
-        UNSPEC_FIST_CEIL))
+                  UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" "m,m"))
    (use (match_operand:HI 3 "memory_operand" "m,m"))
    (clobber (match_operand:DI 4 "memory_operand" "=X,m"))
 (define_split
   [(set (match_operand:DI 0 "register_operand" "")
        (unspec:DI [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_CEIL))
+                  UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
    (clobber (match_operand:DI 4 "memory_operand" ""))
    (clobber (match_scratch 5 ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 4) (unspec:DI [(match_dup 1)] UNSPEC_FIST_CEIL))
+  [(parallel [(set (match_dup 4)
+                  (unspec:DI [(match_dup 1)] UNSPEC_FIST_CEIL))
              (use (match_dup 2))
              (use (match_dup 3))
              (clobber (match_dup 5))])
 (define_split
   [(set (match_operand:DI 0 "memory_operand" "")
        (unspec:DI [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_CEIL))
+                  UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
    (clobber (match_operand:DI 4 "memory_operand" ""))
    (clobber (match_scratch 5 ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 0) (unspec:DI [(match_dup 1)] UNSPEC_FIST_CEIL))
+  [(parallel [(set (match_dup 0)
+                  (unspec:DI [(match_dup 1)] UNSPEC_FIST_CEIL))
              (use (match_dup 2))
              (use (match_dup 3))
              (clobber (match_dup 5))])])
 
 (define_insn "fist<mode>2_ceil"
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "=m")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "f")]
-        UNSPEC_FIST_CEIL))
+  [(set (match_operand:SWI24 0 "memory_operand" "=m")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "f")]
+                     UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" "m"))
    (use (match_operand:HI 3 "memory_operand" "m"))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations"
-  "* return output_fix_trunc (insn, operands, 0);"
+  "* return output_fix_trunc (insn, operands, false);"
   [(set_attr "type" "fistp")
    (set_attr "i387_cw" "ceil")
    (set_attr "mode" "<MODE>")])
 
 (define_insn "fist<mode>2_ceil_with_temp"
-  [(set (match_operand:X87MODEI12 0 "nonimmediate_operand" "=m,?r")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "f,f")]
-        UNSPEC_FIST_CEIL))
+  [(set (match_operand:SWI24 0 "nonimmediate_operand" "=m,?r")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "f,f")]
+                     UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" "m,m"))
    (use (match_operand:HI 3 "memory_operand" "m,m"))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" "=X,m"))]
+   (clobber (match_operand:SWI24 4 "memory_operand" "=X,m"))]
   "TARGET_USE_FANCY_MATH_387
    && flag_unsafe_math_optimizations"
   "#"
    (set_attr "mode" "<MODE>")])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "register_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_CEIL))
+  [(set (match_operand:SWI24 0 "register_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" ""))]
+   (clobber (match_operand:SWI24 4 "memory_operand" ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 4) (unspec:X87MODEI12 [(match_dup 1)]
-                                 UNSPEC_FIST_CEIL))
+  [(parallel [(set (match_dup 4)
+                  (unspec:SWI24 [(match_dup 1)] UNSPEC_FIST_CEIL))
              (use (match_dup 2))
              (use (match_dup 3))])
    (set (match_dup 0) (match_dup 4))])
 
 (define_split
-  [(set (match_operand:X87MODEI12 0 "memory_operand" "")
-       (unspec:X87MODEI12 [(match_operand:XF 1 "register_operand" "")]
-        UNSPEC_FIST_CEIL))
+  [(set (match_operand:SWI24 0 "memory_operand" "")
+       (unspec:SWI24 [(match_operand:XF 1 "register_operand" "")]
+                     UNSPEC_FIST_CEIL))
    (use (match_operand:HI 2 "memory_operand" ""))
    (use (match_operand:HI 3 "memory_operand" ""))
-   (clobber (match_operand:X87MODEI12 4 "memory_operand" ""))]
+   (clobber (match_operand:SWI24 4 "memory_operand" ""))]
   "reload_completed"
-  [(parallel [(set (match_dup 0) (unspec:X87MODEI12 [(match_dup 1)]
-                                 UNSPEC_FIST_CEIL))
+  [(parallel [(set (match_dup 0)
+                  (unspec:SWI24 [(match_dup 1)] UNSPEC_FIST_CEIL))
              (use (match_dup 2))
              (use (match_dup 3))])])
 
 (define_expand "lceilxf<mode>2"
-  [(parallel [(set (match_operand:X87MODEI 0 "nonimmediate_operand" "")
-                  (unspec:X87MODEI [(match_operand:XF 1 "register_operand" "")]
-                   UNSPEC_FIST_CEIL))
+  [(parallel [(set (match_operand:SWI248x 0 "nonimmediate_operand" "")
+                  (unspec:SWI248x [(match_operand:XF 1 "register_operand" "")]
+                                  UNSPEC_FIST_CEIL))
              (clobber (reg:CC FLAGS_REG))])]
   "TARGET_USE_FANCY_MATH_387
    && (!TARGET_SSE_MATH || TARGET_MIX_SSE_I387)
   "SSE_FLOAT_MODE_P (<MODEF:MODE>mode) && TARGET_SSE_MATH
    && !flag_trapping_math"
 {
-  ix86_expand_lfloorceil (operand0, operand1, false);
+  ix86_expand_lfloorceil (operands[0], operands[1], false);
   DONE;
 })
 
        && !flag_trapping_math)"
 {
   if (SSE_FLOAT_MODE_P (<MODE>mode) && TARGET_SSE_MATH
-      && !flag_trapping_math
-      && (TARGET_ROUND || optimize_insn_for_speed_p ()))
+      && !flag_trapping_math)
     {
       if (TARGET_ROUND)
        emit_insn (gen_sse4_1_round<mode>2
       else if (optimize_insn_for_size_p ())
        FAIL;
       else if (TARGET_64BIT || (<MODE>mode != DFmode))
-       ix86_expand_trunc (operand0, operand1);
+       ix86_expand_trunc (operands[0], operands[1]);
       else
-       ix86_expand_truncdf_32 (operand0, operand1);
+       ix86_expand_truncdf_32 (operands[0], operands[1]);
     }
   else
     {
    (set (match_operand:DI 1 "register_operand" "=S")
        (plus:DI (match_dup 3)
                 (const_int 8)))]
-  "TARGET_64BIT"
+  "TARGET_64BIT
+   && !(fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "movsq"
   [(set_attr "type" "str")
    (set_attr "memory" "both")
    (set (match_operand:P 1 "register_operand" "=S")
        (plus:P (match_dup 3)
                (const_int 4)))]
-  ""
+  "!(fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "movs{l|d}"
   [(set_attr "type" "str")
    (set_attr "memory" "both")
    (set (match_operand:P 1 "register_operand" "=S")
        (plus:P (match_dup 3)
                (const_int 2)))]
-  ""
+  "!(fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "movsw"
   [(set_attr "type" "str")
    (set_attr "memory" "both")
    (set (match_operand:P 1 "register_operand" "=S")
        (plus:P (match_dup 3)
                (const_int 1)))]
-  ""
+  "!(fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "movsb"
   [(set_attr "type" "str")
    (set_attr "memory" "both")
    (set (attr "prefix_rex")
        (if_then_else
-         (ne (symbol_ref "<P:MODE>mode == DImode") (const_int 0))
+         (match_test "<P:MODE>mode == DImode")
          (const_string "0")
          (const_string "*")))
    (set_attr "mode" "QI")])
    (set (mem:BLK (match_dup 3))
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
-  "TARGET_64BIT"
+  "TARGET_64BIT
+   && !(fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "rep{%;} movsq"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set (mem:BLK (match_dup 3))
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
-  ""
+  "!(fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "rep{%;} movs{l|d}"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set (mem:BLK (match_dup 3))
        (mem:BLK (match_dup 4)))
    (use (match_dup 5))]
-  ""
+  "!(fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "rep{%;} movsb"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
   operands[3] = gen_rtx_PLUS (Pmode, operands[0],
                              GEN_INT (GET_MODE_SIZE (GET_MODE
                                                      (operands[2]))));
-  if (TARGET_SINGLE_STRINGOP || optimize_insn_for_size_p ())
+  /* Can't use this if the user has appropriated eax or edi.  */
+  if ((TARGET_SINGLE_STRINGOP || optimize_insn_for_size_p ())
+      && !(fixed_regs[AX_REG] || fixed_regs[DI_REG]))
     {
       emit_insn (gen_strset_singleop (operands[0], operands[1], operands[2],
                                      operands[3]));
    (set (match_operand:DI 0 "register_operand" "=D")
        (plus:DI (match_dup 1)
                 (const_int 8)))]
-  "TARGET_64BIT"
+  "TARGET_64BIT
+   && !(fixed_regs[AX_REG] || fixed_regs[DI_REG])"
   "stosq"
   [(set_attr "type" "str")
    (set_attr "memory" "store")
    (set (match_operand:P 0 "register_operand" "=D")
        (plus:P (match_dup 1)
                (const_int 4)))]
-  ""
+  "!(fixed_regs[AX_REG] || fixed_regs[DI_REG])"
   "stos{l|d}"
   [(set_attr "type" "str")
    (set_attr "memory" "store")
    (set (match_operand:P 0 "register_operand" "=D")
        (plus:P (match_dup 1)
                (const_int 2)))]
-  ""
+  "!(fixed_regs[AX_REG] || fixed_regs[DI_REG])"
   "stosw"
   [(set_attr "type" "str")
    (set_attr "memory" "store")
    (set (match_operand:P 0 "register_operand" "=D")
        (plus:P (match_dup 1)
                (const_int 1)))]
-  ""
+  "!(fixed_regs[AX_REG] || fixed_regs[DI_REG])"
   "stosb"
   [(set_attr "type" "str")
    (set_attr "memory" "store")
    (set (attr "prefix_rex")
        (if_then_else
-         (ne (symbol_ref "<P:MODE>mode == DImode") (const_int 0))
+         (match_test "<P:MODE>mode == DImode")
          (const_string "0")
          (const_string "*")))
    (set_attr "mode" "QI")])
        (const_int 0))
    (use (match_operand:DI 2 "register_operand" "a"))
    (use (match_dup 4))]
-  "TARGET_64BIT"
+  "TARGET_64BIT
+   && !(fixed_regs[AX_REG] || fixed_regs[CX_REG] || fixed_regs[DI_REG])"
   "rep{%;} stosq"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
        (const_int 0))
    (use (match_operand:SI 2 "register_operand" "a"))
    (use (match_dup 4))]
-  ""
+  "!(fixed_regs[AX_REG] || fixed_regs[CX_REG] || fixed_regs[DI_REG])"
   "rep{%;} stos{l|d}"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
        (const_int 0))
    (use (match_operand:QI 2 "register_operand" "a"))
    (use (match_dup 4))]
-  ""
+  "!(fixed_regs[AX_REG] || fixed_regs[CX_REG] || fixed_regs[DI_REG])"
   "rep{%;} stosb"
   [(set_attr "type" "str")
    (set_attr "prefix_rep" "1")
    (set_attr "memory" "store")
    (set (attr "prefix_rex")
        (if_then_else
-         (ne (symbol_ref "<P:MODE>mode == DImode") (const_int 0))
+         (match_test "<P:MODE>mode == DImode")
          (const_string "0")
          (const_string "*")))
    (set_attr "mode" "QI")])
   if (optimize_insn_for_size_p () && !TARGET_INLINE_ALL_STRINGOPS)
     FAIL;
 
-  /* Can't use this if the user has appropriated esi or edi.  */
-  if (fixed_regs[SI_REG] || fixed_regs[DI_REG])
+  /* Can't use this if the user has appropriated ecx, esi or edi.  */
+  if (fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])
     FAIL;
 
   out = operands[0];
    (clobber (match_operand:P 0 "register_operand" "=S"))
    (clobber (match_operand:P 1 "register_operand" "=D"))
    (clobber (match_operand:P 2 "register_operand" "=c"))]
-  ""
+  "!(fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "repz{%;} cmpsb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set (attr "prefix_rex")
        (if_then_else
-         (ne (symbol_ref "<P:MODE>mode == DImode") (const_int 0))
+         (match_test "<P:MODE>mode == DImode")
          (const_string "0")
          (const_string "*")))
    (set_attr "prefix_rep" "1")])
    (clobber (match_operand:P 0 "register_operand" "=S"))
    (clobber (match_operand:P 1 "register_operand" "=D"))
    (clobber (match_operand:P 2 "register_operand" "=c"))]
-  ""
+  "!(fixed_regs[CX_REG] || fixed_regs[SI_REG] || fixed_regs[DI_REG])"
   "repz{%;} cmpsb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set (attr "prefix_rex")
        (if_then_else
-         (ne (symbol_ref "<P:MODE>mode == DImode") (const_int 0))
+         (match_test "<P:MODE>mode == DImode")
          (const_string "0")
          (const_string "*")))
    (set_attr "prefix_rep" "1")])
 
 (define_expand "strlen<mode>"
-  [(set (match_operand:SWI48x 0 "register_operand" "")
-       (unspec:SWI48x [(match_operand:BLK 1 "general_operand" "")
-                       (match_operand:QI 2 "immediate_operand" "")
-                       (match_operand 3 "immediate_operand" "")]
-                      UNSPEC_SCAS))]
+  [(set (match_operand:P 0 "register_operand" "")
+       (unspec:P [(match_operand:BLK 1 "general_operand" "")
+                  (match_operand:QI 2 "immediate_operand" "")
+                  (match_operand 3 "immediate_operand" "")]
+                 UNSPEC_SCAS))]
   ""
 {
  if (ix86_expand_strlen (operands[0], operands[1], operands[2], operands[3]))
                   (match_operand:P 4 "register_operand" "0")] UNSPEC_SCAS))
    (clobber (match_operand:P 1 "register_operand" "=D"))
    (clobber (reg:CC FLAGS_REG))]
-  ""
+  "!(fixed_regs[AX_REG] || fixed_regs[CX_REG] || fixed_regs[DI_REG])"
   "repnz{%;} scasb"
   [(set_attr "type" "str")
    (set_attr "mode" "QI")
    (set (attr "prefix_rex")
        (if_then_else
-         (ne (symbol_ref "<P:MODE>mode == DImode") (const_int 0))
+         (match_test "<P:MODE>mode == DImode")
          (const_string "0")
          (const_string "*")))
    (set_attr "prefix_rep" "1")])
 (define_expand "mov<mode>cc"
   [(set (match_operand:SWIM 0 "register_operand" "")
        (if_then_else:SWIM (match_operand 1 "ordered_comparison_operator" "")
-                          (match_operand:SWIM 2 "general_operand" "")
-                          (match_operand:SWIM 3 "general_operand" "")))]
+                          (match_operand:SWIM 2 "<general_operand>" "")
+                          (match_operand:SWIM 3 "<general_operand>" "")))]
   ""
   "if (ix86_expand_int_movcc (operands)) DONE; else FAIL;")
 
 (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)])))]
+                   [(reg FLAGS_REG) (const_int 0)])))
+   (clobber (reg:CC FLAGS_REG))]
   ""
   "sbb{<imodesuffix>}\t%0, %0"
   [(set_attr "type" "alu")
 
   ;; The % modifier is not operational anymore in peephole2's, so we have to
   ;; swap the operands manually in the case of addition and multiplication.
-  "if (COMMUTATIVE_ARITH_P (operands[2]))
-     operands[4] = gen_rtx_fmt_ee (GET_CODE (operands[2]),
-                                  GET_MODE (operands[2]),
-                                  operands[0], operands[1]);
-   else
-     operands[4] = gen_rtx_fmt_ee (GET_CODE (operands[2]),
-                                  GET_MODE (operands[2]),
-                                  operands[1], operands[0]);")
+{
+  rtx op0, op1;
+
+  if (COMMUTATIVE_ARITH_P (operands[2]))
+    op0 = operands[0], op1 = operands[1];
+  else
+    op0 = operands[1], op1 = operands[0];
+
+  operands[4] = gen_rtx_fmt_ee (GET_CODE (operands[2]),
+                               GET_MODE (operands[2]),
+                               op0, op1);
+})
 
 ;; Conditional addition patterns
 (define_expand "add<mode>cc"
 }
   [(set (attr "type")
        (cond [(and (eq_attr "alternative" "0")
-                   (eq (symbol_ref "TARGET_OPT_AGU") (const_int 0)))
+                   (not (match_test "TARGET_OPT_AGU")))
                 (const_string "alu")
               (match_operand:<MODE> 2 "const0_operand" "")
                 (const_string "imov")
   [(parallel [(set (match_dup 0)
                   (match_op_dup 3 [(match_dup 1) (match_dup 2)]))
              (clobber (reg:CC FLAGS_REG))])]
-  "operands[0] = gen_lowpart (SImode, operands[0]);
-   operands[1] = gen_lowpart (SImode, operands[1]);
-   if (GET_CODE (operands[3]) != ASHIFT)
-     operands[2] = gen_lowpart (SImode, operands[2]);
-   PUT_MODE (operands[3], SImode);")
+{
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[1] = gen_lowpart (SImode, operands[1]);
+  if (GET_CODE (operands[3]) != ASHIFT)
+    operands[2] = gen_lowpart (SImode, operands[2]);
+  PUT_MODE (operands[3], SImode);
+})
 
 ; Promote the QImode tests, as i386 has encoding of the AND
 ; instruction with 32-bit sign-extended immediate and thus the
   [(parallel [(set (match_dup 0)
                   (neg:SI (match_dup 1)))
              (clobber (reg:CC FLAGS_REG))])]
-  "operands[0] = gen_lowpart (SImode, operands[0]);
-   operands[1] = gen_lowpart (SImode, operands[1]);")
+{
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[1] = gen_lowpart (SImode, operands[1]);
+})
 
 (define_split
   [(set (match_operand 0 "register_operand" "")
               || optimize_insn_for_size_p ())))"
   [(set (match_dup 0)
        (not:SI (match_dup 1)))]
-  "operands[0] = gen_lowpart (SImode, operands[0]);
-   operands[1] = gen_lowpart (SImode, operands[1]);")
+{
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[1] = gen_lowpart (SImode, operands[1]);
+})
 
 (define_split
   [(set (match_operand 0 "register_operand" "")
               || optimize_insn_for_size_p ())))"
   [(set (match_dup 0)
        (if_then_else:SI (match_dup 1) (match_dup 2) (match_dup 3)))]
-  "operands[0] = gen_lowpart (SImode, operands[0]);
-   operands[2] = gen_lowpart (SImode, operands[2]);
-   operands[3] = gen_lowpart (SImode, operands[3]);")
+{
+  operands[0] = gen_lowpart (SImode, operands[0]);
+  operands[2] = gen_lowpart (SImode, operands[2]);
+  operands[3] = gen_lowpart (SImode, operands[3]);
+})
 \f
 ;; RTL Peephole optimizations, run before sched2.  These primarily look to
 ;; transform a complex memory operation into two memory to register operations.
   [(set (match_operand:SWI 0 "push_operand" "")
        (match_operand:SWI 1 "memory_operand" ""))
    (match_scratch:SWI 2 "<r>")]
-  "optimize_insn_for_speed_p () && !TARGET_PUSH_MEMORY
+  "!(TARGET_PUSH_MEMORY || optimize_insn_for_size_p ())
    && !RTX_FRAME_RELATED_P (peep2_next_insn (0))"
   [(set (match_dup 2) (match_dup 1))
    (set (match_dup 0) (match_dup 2))])
   [(set (match_operand:SF 0 "push_operand" "")
        (match_operand:SF 1 "memory_operand" ""))
    (match_scratch:SF 2 "r")]
-  "optimize_insn_for_speed_p () && !TARGET_PUSH_MEMORY
+  "!(TARGET_PUSH_MEMORY || optimize_insn_for_size_p ())
    && !RTX_FRAME_RELATED_P (peep2_next_insn (0))"
   [(set (match_dup 2) (match_dup 1))
    (set (match_dup 0) (match_dup 2))])
                      [(match_dup 0)
                       (match_operand:SI 1 "memory_operand" "")]))
               (clobber (reg:CC FLAGS_REG))])]
-  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY"
+  "!(TARGET_READ_MODIFY || optimize_insn_for_size_p ())"
   [(set (match_dup 2) (match_dup 1))
    (parallel [(set (match_dup 0)
                    (match_op_dup 3 [(match_dup 0) (match_dup 2)]))
                      [(match_operand:SI 1 "memory_operand" "")
                       (match_dup 0)]))
               (clobber (reg:CC FLAGS_REG))])]
-  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY"
+  "!(TARGET_READ_MODIFY || optimize_insn_for_size_p ())"
   [(set (match_dup 2) (match_dup 1))
    (parallel [(set (match_dup 0)
                    (match_op_dup 3 [(match_dup 2) (match_dup 0)]))
                      [(match_dup 0)
                       (match_operand:SI 1 "nonmemory_operand" "")]))
               (clobber (reg:CC FLAGS_REG))])]
-  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY_WRITE
+  "!(TARGET_READ_MODIFY_WRITE || optimize_insn_for_size_p ())
    /* Do not split stack checking probes.  */
    && GET_CODE (operands[3]) != IOR && operands[1] != const0_rtx"
   [(set (match_dup 2) (match_dup 0))
                      [(match_operand:SI 1 "nonmemory_operand" "")
                       (match_dup 0)]))
               (clobber (reg:CC FLAGS_REG))])]
-  "optimize_insn_for_speed_p () && ! TARGET_READ_MODIFY_WRITE
+  "!(TARGET_READ_MODIFY_WRITE || optimize_insn_for_size_p ())
    /* Do not split stack checking probes.  */
    && GET_CODE (operands[3]) != IOR && operands[1] != const0_rtx"
   [(set (match_dup 2) (match_dup 0))
               (clobber (reg:CC FLAGS_REG))])
    (set (match_dup 0) (match_dup 2))])
 
+;; Attempt to use arith or logical operations with memory outputs with
+;; setting of flags.
+(define_peephole2
+  [(set (match_operand:SWI 0 "register_operand" "")
+       (match_operand:SWI 1 "memory_operand" ""))
+   (parallel [(set (match_dup 0)
+                  (match_operator:SWI 3 "plusminuslogic_operator"
+                    [(match_dup 0)
+                     (match_operand:SWI 2 "<nonmemory_operand>" "")]))
+             (clobber (reg:CC FLAGS_REG))])
+   (set (match_dup 1) (match_dup 0))
+   (set (reg FLAGS_REG) (compare (match_dup 0) (const_int 0)))]
+  "(TARGET_READ_MODIFY_WRITE || optimize_insn_for_size_p ())
+   && peep2_reg_dead_p (4, operands[0])
+   && !reg_overlap_mentioned_p (operands[0], operands[1])
+   && (<MODE>mode != QImode
+       || immediate_operand (operands[2], QImode)
+       || q_regs_operand (operands[2], QImode))
+   && ix86_match_ccmode (peep2_next_insn (3),
+                        (GET_CODE (operands[3]) == PLUS
+                         || GET_CODE (operands[3]) == MINUS)
+                        ? CCGOCmode : CCNOmode)"
+  [(parallel [(set (match_dup 4) (match_dup 5))
+             (set (match_dup 1) (match_op_dup 3 [(match_dup 1)
+                                                 (match_dup 2)]))])]
+{
+  operands[4] = SET_DEST (PATTERN (peep2_next_insn (3)));
+  operands[5] = gen_rtx_fmt_ee (GET_CODE (operands[3]), <MODE>mode,
+                               copy_rtx (operands[1]),
+                               copy_rtx (operands[2]));
+  operands[5] = gen_rtx_COMPARE (GET_MODE (operands[4]),
+                                operands[5], const0_rtx);
+})
+
+(define_peephole2
+  [(parallel [(set (match_operand:SWI 0 "register_operand" "")
+                  (match_operator:SWI 2 "plusminuslogic_operator"
+                    [(match_dup 0)
+                     (match_operand:SWI 1 "memory_operand" "")]))
+             (clobber (reg:CC FLAGS_REG))])
+   (set (match_dup 1) (match_dup 0))
+   (set (reg FLAGS_REG) (compare (match_dup 0) (const_int 0)))]
+  "(TARGET_READ_MODIFY_WRITE || optimize_insn_for_size_p ())
+   && GET_CODE (operands[2]) != MINUS
+   && peep2_reg_dead_p (3, operands[0])
+   && !reg_overlap_mentioned_p (operands[0], operands[1])
+   && ix86_match_ccmode (peep2_next_insn (2),
+                        GET_CODE (operands[2]) == PLUS
+                        ? CCGOCmode : CCNOmode)"
+  [(parallel [(set (match_dup 3) (match_dup 4))
+             (set (match_dup 1) (match_op_dup 2 [(match_dup 1)
+                                                 (match_dup 0)]))])]
+{
+  operands[3] = SET_DEST (PATTERN (peep2_next_insn (2)));
+  operands[4] = gen_rtx_fmt_ee (GET_CODE (operands[2]), <MODE>mode,
+                               copy_rtx (operands[1]),
+                               copy_rtx (operands[0]));
+  operands[4] = gen_rtx_COMPARE (GET_MODE (operands[3]),
+                                operands[4], const0_rtx);
+})
+
+(define_peephole2
+  [(set (match_operand:SWI12 0 "register_operand" "")
+       (match_operand:SWI12 1 "memory_operand" ""))
+   (parallel [(set (match_operand:SI 4 "register_operand" "")
+                  (match_operator:SI 3 "plusminuslogic_operator"
+                    [(match_dup 4)
+                     (match_operand:SI 2 "nonmemory_operand" "")]))
+             (clobber (reg:CC FLAGS_REG))])
+   (set (match_dup 1) (match_dup 0))
+   (set (reg FLAGS_REG) (compare (match_dup 0) (const_int 0)))]
+  "(TARGET_READ_MODIFY_WRITE || optimize_insn_for_size_p ())
+   && REG_P (operands[0]) && REG_P (operands[4])
+   && REGNO (operands[0]) == REGNO (operands[4])
+   && peep2_reg_dead_p (4, operands[0])
+   && (<MODE>mode != QImode
+       || immediate_operand (operands[2], SImode)
+       || q_regs_operand (operands[2], SImode))
+   && !reg_overlap_mentioned_p (operands[0], operands[1])
+   && ix86_match_ccmode (peep2_next_insn (3),
+                        (GET_CODE (operands[3]) == PLUS
+                         || GET_CODE (operands[3]) == MINUS)
+                        ? CCGOCmode : CCNOmode)"
+  [(parallel [(set (match_dup 4) (match_dup 5))
+             (set (match_dup 1) (match_dup 6))])]
+{
+  operands[2] = gen_lowpart (<MODE>mode, operands[2]);
+  operands[4] = SET_DEST (PATTERN (peep2_next_insn (3)));
+  operands[5] = gen_rtx_fmt_ee (GET_CODE (operands[3]), <MODE>mode,
+                               copy_rtx (operands[1]), operands[2]);
+  operands[5] = gen_rtx_COMPARE (GET_MODE (operands[4]),
+                                operands[5], const0_rtx);
+  operands[6] = gen_rtx_fmt_ee (GET_CODE (operands[3]), <MODE>mode,
+                               copy_rtx (operands[1]),
+                               copy_rtx (operands[2]));
+})
+
 ;; Attempt to always use XOR for zeroing registers.
 (define_peephole2
   [(set (match_operand 0 "register_operand" "")
   [(parallel
     [(set (match_operand:SWI48 0 "register_operand" "")
          (mult:SWI48 (match_operand:SWI48 1 "register_operand" "")
-                     (match_operand:SWI48 2 "const_int_operand" "")))
+                     (match_operand:SWI48 2 "const359_operand" "")))
      (clobber (reg:CC FLAGS_REG))])]
-  "INTVAL (operands[2]) == 3
-   || INTVAL (operands[2]) == 5
-   || INTVAL (operands[2]) == 9"
+  "!TARGET_PARTIAL_REG_STALL
+   || <MODE>mode == SImode
+   || optimize_function_for_size_p (cfun)"
   [(set (match_dup 0)
        (plus:SWI48 (mult:SWI48 (match_dup 1) (match_dup 2))
                    (match_dup 1)))]
   [(parallel
     [(set (match_operand:SWI48 0 "register_operand" "")
          (mult:SWI48 (match_operand:SWI48 1 "nonimmediate_operand" "")
-                     (match_operand:SWI48 2 "const_int_operand" "")))
+                     (match_operand:SWI48 2 "const359_operand" "")))
      (clobber (reg:CC FLAGS_REG))])]
   "optimize_insn_for_speed_p ()
-   && (INTVAL (operands[2]) == 3
-       || INTVAL (operands[2]) == 5
-       || INTVAL (operands[2]) == 9)"
+   && (!TARGET_PARTIAL_REG_STALL || <MODE>mode == SImode)"
   [(set (match_dup 0) (match_dup 1))
    (set (match_dup 0)
        (plus:SWI48 (mult:SWI48 (match_dup 0) (match_dup 2))
   operands[0] = dest;
 })
 \f
-;; Call-value patterns last so that the wildcard operand does not
-;; disrupt insn-recog's switch tables.
-
-(define_insn_and_split "*call_value_pop_0_vzeroupper"
-  [(parallel
-    [(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" "")))])
-   (unspec [(match_operand 4 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[4]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_pop_0"
-  [(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" "")))]
-  "!TARGET_64BIT"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_pop_1_vzeroupper"
-  [(parallel
-    [(set (match_operand 0 "" "")
-         (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")))])
-   (unspec [(match_operand 4 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[4]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_pop_1"
-  [(set (match_operand 0 "" "")
-       (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 && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*sibcall_value_pop_1_vzeroupper"
- [(parallel
-   [(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")))])
-   (unspec [(match_operand 4 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[4]); DONE;"
-  [(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)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_0_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "constant_call_address_operand" ""))
-             (match_operand:SI 2 "" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_0"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "constant_call_address_operand" ""))
-             (match_operand:SI 2 "" "")))]
-  "!TARGET_64BIT"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_0_rex64_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "constant_call_address_operand" ""))
-             (match_operand:DI 2 "const_int_operand" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_0_rex64"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "constant_call_address_operand" ""))
-             (match_operand:DI 2 "const_int_operand" "")))]
-  "TARGET_64BIT"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_0_rex64_ms_sysv_vzeroupper"
-  [(parallel
-    [(set (match_operand 0 "" "")
-         (call (mem:QI (match_operand:DI 1 "constant_call_address_operand" ""))
-               (match_operand:DI 2 "const_int_operand" "")))
-     (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
-     (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))])
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_0_rex64_ms_sysv"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "constant_call_address_operand" ""))
-             (match_operand:DI 2 "const_int_operand" "")))
-   (unspec [(const_int 0)] UNSPEC_MS_TO_SYSV_CALL)
-   (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))]
-  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_1_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "lsm"))
-             (match_operand:SI 2 "" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_1"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "call_insn_operand" "lsm"))
-             (match_operand:SI 2 "" "")))]
-  "!TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*sibcall_value_1_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "s,U"))
-             (match_operand:SI 2 "" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && !TARGET_64BIT && SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*sibcall_value_1"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:SI 1 "sibcall_insn_operand" "s,U"))
-             (match_operand:SI 2 "" "")))]
-  "!TARGET_64BIT && SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_1_rex64_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rsm"))
-             (match_operand:DI 2 "" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)
-   && ix86_cmodel != CM_LARGE && ix86_cmodel != CM_LARGE_PIC"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(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 "" "")))]
-  "TARGET_64BIT && !SIBLING_CALL_P (insn)
-   && ix86_cmodel != CM_LARGE && ix86_cmodel != CM_LARGE_PIC"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_1_rex64_ms_sysv_vzeroupper"
-  [(parallel
-    [(set (match_operand 0 "" "")
-         (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 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))])
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_1_rex64_ms_sysv"
-  [(set (match_operand 0 "" "")
-       (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 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))]
-  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*call_value_1_rex64_large_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rm"))
-             (match_operand:DI 2 "" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*call_value_1_rex64_large"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "call_insn_operand" "rm"))
-             (match_operand:DI 2 "" "")))]
-  "TARGET_64BIT && !SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-
-(define_insn_and_split "*sibcall_value_1_rex64_vzeroupper"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "sibcall_insn_operand" "s,U"))
-             (match_operand:DI 2 "" "")))
-   (unspec [(match_operand 3 "const_int_operand" "")]
-          UNSPEC_CALL_NEEDS_VZEROUPPER)]
-  "TARGET_VZEROUPPER && TARGET_64BIT && SIBLING_CALL_P (insn)"
-  "#"
-  "&& reload_completed"
-  [(const_int 0)]
-  "ix86_split_call_vzeroupper (curr_insn, operands[3]); DONE;"
-  [(set_attr "type" "callv")])
-
-(define_insn "*sibcall_value_1_rex64"
-  [(set (match_operand 0 "" "")
-       (call (mem:QI (match_operand:DI 1 "sibcall_insn_operand" "s,U"))
-             (match_operand:DI 2 "" "")))]
-  "TARGET_64BIT && SIBLING_CALL_P (insn)"
-  { return ix86_output_call_insn (insn, operands[1], 1); }
-  [(set_attr "type" "callv")])
-\f
 ;; We used to use "int $5", in honor of #BR which maps to interrupt vector 5.
 ;; That, however, is usually mapped by the OS to SIGSEGV, which is often
 ;; caught for use by garbage collectors and the like.  Using an insn that
 
 #ifdef TARGET_THREAD_SSP_OFFSET
   operands[1] = GEN_INT (TARGET_THREAD_SSP_OFFSET);
-  insn = (TARGET_64BIT
+  insn = (TARGET_LP64
          ? gen_stack_tls_protect_set_di
          : gen_stack_tls_protect_set_si);
 #else
-  insn = (TARGET_64BIT
+  insn = (TARGET_LP64
          ? gen_stack_protect_set_di
          : gen_stack_protect_set_si);
 #endif
 })
 
 (define_insn "stack_protect_set_<mode>"
-  [(set (match_operand:P 0 "memory_operand" "=m")
-       (unspec:P [(match_operand:P 1 "memory_operand" "m")] UNSPEC_SP_SET))
-   (set (match_scratch:P 2 "=&r") (const_int 0))
+  [(set (match_operand:PTR 0 "memory_operand" "=m")
+       (unspec:PTR [(match_operand:PTR 1 "memory_operand" "m")]
+                   UNSPEC_SP_SET))
+   (set (match_scratch:PTR 2 "=&r") (const_int 0))
    (clobber (reg:CC FLAGS_REG))]
   ""
   "mov{<imodesuffix>}\t{%1, %2|%2, %1}\;mov{<imodesuffix>}\t{%2, %0|%0, %2}\;xor{l}\t%k2, %k2"
   [(set_attr "type" "multi")])
 
 (define_insn "stack_tls_protect_set_<mode>"
-  [(set (match_operand:P 0 "memory_operand" "=m")
-       (unspec:P [(match_operand:P 1 "const_int_operand" "i")]
-                 UNSPEC_SP_TLS_SET))
-   (set (match_scratch:P 2 "=&r") (const_int 0))
+  [(set (match_operand:PTR 0 "memory_operand" "=m")
+       (unspec:PTR [(match_operand:PTR 1 "const_int_operand" "i")]
+                   UNSPEC_SP_TLS_SET))
+   (set (match_scratch:PTR 2 "=&r") (const_int 0))
    (clobber (reg:CC FLAGS_REG))]
   ""
   "mov{<imodesuffix>}\t{%@:%P1, %2|%2, <iptrsize> PTR %@:%P1}\;mov{<imodesuffix>}\t{%2, %0|%0, %2}\;xor{l}\t%k2, %k2"
 
 #ifdef TARGET_THREAD_SSP_OFFSET
   operands[1] = GEN_INT (TARGET_THREAD_SSP_OFFSET);
-  insn = (TARGET_64BIT
+  insn = (TARGET_LP64
          ? gen_stack_tls_protect_test_di
          : gen_stack_tls_protect_test_si);
 #else
-  insn = (TARGET_64BIT
+  insn = (TARGET_LP64
          ? gen_stack_protect_test_di
          : gen_stack_protect_test_si);
 #endif
 
 (define_insn "stack_protect_test_<mode>"
   [(set (match_operand:CCZ 0 "flags_reg_operand" "")
-       (unspec:CCZ [(match_operand:P 1 "memory_operand" "m")
-                    (match_operand:P 2 "memory_operand" "m")]
+       (unspec:CCZ [(match_operand:PTR 1 "memory_operand" "m")
+                    (match_operand:PTR 2 "memory_operand" "m")]
                    UNSPEC_SP_TEST))
-   (clobber (match_scratch:P 3 "=&r"))]
+   (clobber (match_scratch:PTR 3 "=&r"))]
   ""
   "mov{<imodesuffix>}\t{%1, %3|%3, %1}\;xor{<imodesuffix>}\t{%2, %3|%3, %2}"
   [(set_attr "type" "multi")])
 
 (define_insn "stack_tls_protect_test_<mode>"
   [(set (match_operand:CCZ 0 "flags_reg_operand" "")
-       (unspec:CCZ [(match_operand:P 1 "memory_operand" "m")
-                    (match_operand:P 2 "const_int_operand" "i")]
+       (unspec:CCZ [(match_operand:PTR 1 "memory_operand" "m")
+                    (match_operand:PTR 2 "const_int_operand" "i")]
                    UNSPEC_SP_TLS_TEST))
-   (clobber (match_scratch:P 3 "=r"))]
+   (clobber (match_scratch:PTR 3 "=r"))]
   ""
   "mov{<imodesuffix>}\t{%1, %3|%3, %1}\;xor{<imodesuffix>}\t{%@:%P2, %3|%3, <iptrsize> PTR %@:%P2}"
   [(set_attr "type" "multi")])
                     (match_operand:SI 3 "const_int_operand" "i")]
                    UNSPECV_LWPVAL_INTRINSIC)]
   "TARGET_LWP"
-  "/* Avoid unused variable warning.  */
-   (void) operand0;")
+  ;; Avoid unused variable warning.
+  "(void) operands[0];")
 
 (define_insn "*lwp_lwpval<mode>3_1"
   [(unspec_volatile [(match_operand:SWI48 0 "register_operand" "r")
 
 (define_insn "rdrand<mode>_1"
   [(set (match_operand:SWI248 0 "register_operand" "=r")
-       (unspec:SWI248 [(const_int 0)] UNSPEC_RDRAND))
+       (unspec_volatile:SWI248 [(const_int 0)] UNSPECV_RDRAND))
    (set (reg:CCC FLAGS_REG)
-       (unspec:CCC [(const_int 0)] UNSPEC_RDRAND))]
+       (unspec_volatile:CCC [(const_int 0)] UNSPECV_RDRAND))]
   "TARGET_RDRND"
   "rdrand\t%0"
   [(set_attr "type" "other")
    (set_attr "prefix_extra" "1")])
 
+(define_expand "pause"
+  [(set (match_dup 0)
+       (unspec:BLK [(match_dup 0)] UNSPEC_PAUSE))]
+  ""
+{
+  operands[0] = gen_rtx_MEM (BLKmode, gen_rtx_SCRATCH (Pmode));
+  MEM_VOLATILE_P (operands[0]) = 1;
+})
+
+;; Use "rep; nop", instead of "pause", to support older assemblers.
+;; They have the same encoding.
+(define_insn "*pause"
+  [(set (match_operand:BLK 0 "" "")
+       (unspec:BLK [(match_dup 0)] UNSPEC_PAUSE))]
+  ""
+  "rep; nop"
+  [(set_attr "length" "2")
+   (set_attr "memory" "unknown")])
+
 (include "mmx.md")
 (include "sse.md")
 (include "sync.md")