daily update
[deliverable/binutils-gdb.git] / gas / config / tc-i386.c
index 6172eb97efacf1a4ea344461cbc546d8fe6afb95..b91b2bb996542c664b31410070b75017f71ed229 100644 (file)
@@ -1,6 +1,6 @@
 /* i386.c -- Assemble code for the Intel 80386
    Copyright 1989, 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
-   2000, 2001
+   2000, 2001, 2002
    Free Software Foundation, Inc.
 
    This file is part of GAS, the GNU Assembler.
@@ -26,9 +26,8 @@
    Bugs & suggestions are completely welcome.  This is free software.
    Please help us make it better.  */
 
-#include <ctype.h>
-
 #include "as.h"
+#include "safe-ctype.h"
 #include "subsegs.h"
 #include "dwarf2dbg.h"
 #include "opcode/i386.h"
 #define SCALE1_WHEN_NO_INDEX 1
 #endif
 
-#define true 1
-#define false 0
+#ifdef BFD_ASSEMBLER
+#define RELOC_ENUM enum bfd_reloc_code_real
+#else
+#define RELOC_ENUM int
+#endif
+
+#ifndef DEFAULT_ARCH
+#define DEFAULT_ARCH "i386"
+#endif
+
+#ifndef INLINE
+#if __GNUC__ >= 2
+#define INLINE __inline__
+#else
+#define INLINE
+#endif
+#endif
 
-static unsigned int mode_from_disp_size PARAMS ((unsigned int));
-static int fits_in_signed_byte PARAMS ((offsetT));
-static int fits_in_unsigned_byte PARAMS ((offsetT));
-static int fits_in_unsigned_word PARAMS ((offsetT));
-static int fits_in_signed_word PARAMS ((offsetT));
-static int fits_in_unsigned_long PARAMS ((offsetT));
-static int fits_in_signed_long PARAMS ((offsetT));
+static INLINE unsigned int mode_from_disp_size PARAMS ((unsigned int));
+static INLINE int fits_in_signed_byte PARAMS ((offsetT));
+static INLINE int fits_in_unsigned_byte PARAMS ((offsetT));
+static INLINE int fits_in_unsigned_word PARAMS ((offsetT));
+static INLINE int fits_in_signed_word PARAMS ((offsetT));
+static INLINE int fits_in_unsigned_long PARAMS ((offsetT));
+static INLINE int fits_in_signed_long PARAMS ((offsetT));
 static int smallest_imm_type PARAMS ((offsetT));
 static offsetT offset_in_range PARAMS ((offsetT, int));
 static int add_prefix PARAMS ((unsigned int));
@@ -66,19 +80,39 @@ static void set_code_flag PARAMS ((int));
 static void set_16bit_gcc_code_flag PARAMS ((int));
 static void set_intel_syntax PARAMS ((int));
 static void set_cpu_arch PARAMS ((int));
-
-#ifdef BFD_ASSEMBLER
-static bfd_reloc_code_real_type reloc
-  PARAMS ((int, int, int, bfd_reloc_code_real_type));
-#define RELOC_ENUM enum bfd_reloc_code_real
-#else
-#define RELOC_ENUM int
+static char *output_invalid PARAMS ((int c));
+static int i386_operand PARAMS ((char *operand_string));
+static int i386_intel_operand PARAMS ((char *operand_string, int got_a_float));
+static const reg_entry *parse_register PARAMS ((char *reg_string,
+                                               char **end_op));
+static char *parse_insn PARAMS ((char *, char *));
+static char *parse_operands PARAMS ((char *, const char *));
+static void swap_operands PARAMS ((void));
+static void optimize_imm PARAMS ((void));
+static void optimize_disp PARAMS ((void));
+static int match_template PARAMS ((void));
+static int check_string PARAMS ((void));
+static int process_suffix PARAMS ((void));
+static int check_byte_reg PARAMS ((void));
+static int check_long_reg PARAMS ((void));
+static int check_qword_reg PARAMS ((void));
+static int check_word_reg PARAMS ((void));
+static int finalize_imm PARAMS ((void));
+static int process_operands PARAMS ((void));
+static const seg_entry *build_modrm_byte PARAMS ((void));
+static void output_insn PARAMS ((void));
+static void output_branch PARAMS ((void));
+static void output_jump PARAMS ((void));
+static void output_interseg_jump PARAMS ((void));
+static void output_imm PARAMS ((fragS *insn_start_frag,
+                               offsetT insn_start_off));
+static void output_disp PARAMS ((fragS *insn_start_frag,
+                                offsetT insn_start_off));
+#ifndef I386COFF
+static void s_bss PARAMS ((int));
 #endif
 
-#ifndef DEFAULT_ARCH
-#define DEFAULT_ARCH "i386"
-#endif
-static char *default_arch = DEFAULT_ARCH;
+static const char *default_arch = DEFAULT_ARCH;
 
 /* 'md_assemble ()' gathers together information and puts it into a
    i386_insn.  */
@@ -150,22 +184,20 @@ typedef struct _i386_insn i386_insn;
 /* List of chars besides those in app.c:symbol_chars that can start an
    operand.  Used to prevent the scrubber eating vital white-space.  */
 #ifdef LEX_AT
-const char extra_symbol_chars[] = "*%-(@";
+const char extra_symbol_chars[] = "*%-(@[";
 #else
-const char extra_symbol_chars[] = "*%-(";
+const char extra_symbol_chars[] = "*%-([";
 #endif
 
+#if (defined (TE_I386AIX)                              \
+     || ((defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF))        \
+        && !defined (TE_LINUX)                         \
+        && !defined (TE_FreeBSD)                       \
+        && !defined (TE_NetBSD)))
 /* This array holds the chars that always start a comment.  If the
    pre-processor is disabled, these aren't very useful.  */
-#if defined (TE_I386AIX) || ((defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF)) && ! defined (TE_LINUX) && !defined(TE_FreeBSD) && !defined(TE_NetBSD))
-/* Putting '/' here makes it impossible to use the divide operator.
-   However, we need it for compatibility with SVR4 systems.  */
 const char comment_chars[] = "#/";
 #define PREFIX_SEPARATOR '\\'
-#else
-const char comment_chars[] = "#";
-#define PREFIX_SEPARATOR '/'
-#endif
 
 /* This array holds the chars that only start a comment at the beginning of
    a line.  If the line seems to have the form '# 123 filename'
@@ -175,9 +207,14 @@ const char comment_chars[] = "#";
    #NO_APP at the beginning of its output.
    Also note that comments started like this one will always work if
    '/' isn't otherwise defined.  */
-#if defined (TE_I386AIX) || ((defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF)) && ! defined (TE_LINUX) && !defined(TE_FreeBSD) && !defined(TE_NetBSD))
 const char line_comment_chars[] = "";
+
 #else
+/* Putting '/' here makes it impossible to use the divide operator.
+   However, we need it for compatibility with SVR4 systems.  */
+const char comment_chars[] = "#";
+#define PREFIX_SEPARATOR '/'
+
 const char line_comment_chars[] = "/";
 #endif
 
@@ -272,12 +309,15 @@ static int quiet_warnings = 0;
 static const char *cpu_arch_name = NULL;
 
 /* CPU feature flags.  */
-static unsigned int cpu_arch_flags = CpuUnknownFlags|CpuNo64;
+static unsigned int cpu_arch_flags = CpuUnknownFlags | CpuNo64;
 
 /* If set, conditional jumps are not automatically promoted to handle
    larger than a byte offset.  */
 static unsigned int no_cond_jump_promotion = 0;
 
+/* Pre-defined "_GLOBAL_OFFSET_TABLE_".  */
+symbolS *GOT_symbol;
+
 /* Interface to relax_segment.
    There are 3 major relax states for 386 jump insns because the
    different types of jumps add different sizes to frags when we're
@@ -291,9 +331,9 @@ static unsigned int no_cond_jump_promotion = 0;
 /* Sizes.  */
 #define CODE16 1
 #define SMALL  0
-#define SMALL16 (SMALL|CODE16)
+#define SMALL16 (SMALL | CODE16)
 #define BIG    2
-#define BIG16  (BIG|CODE16)
+#define BIG16  (BIG | CODE16)
 
 #ifndef INLINE
 #ifdef __GNUC__
@@ -374,6 +414,43 @@ static const arch_entry cpu_arch[] = {
   {NULL, 0 }
 };
 
+const pseudo_typeS md_pseudo_table[] =
+{
+#if !defined(OBJ_AOUT) && !defined(USE_ALIGN_PTWO)
+  {"align", s_align_bytes, 0},
+#else
+  {"align", s_align_ptwo, 0},
+#endif
+  {"arch", set_cpu_arch, 0},
+#ifndef I386COFF
+  {"bss", s_bss, 0},
+#endif
+  {"ffloat", float_cons, 'f'},
+  {"dfloat", float_cons, 'd'},
+  {"tfloat", float_cons, 'x'},
+  {"value", cons, 2},
+  {"noopt", s_ignore, 0},
+  {"optim", s_ignore, 0},
+  {"code16gcc", set_16bit_gcc_code_flag, CODE_16BIT},
+  {"code16", set_code_flag, CODE_16BIT},
+  {"code32", set_code_flag, CODE_32BIT},
+  {"code64", set_code_flag, CODE_64BIT},
+  {"intel_syntax", set_intel_syntax, 1},
+  {"att_syntax", set_intel_syntax, 0},
+  {"file", (void (*) PARAMS ((int))) dwarf2_directive_file, 0},
+  {"loc", dwarf2_directive_loc, 0},
+  {0, 0, 0}
+};
+
+/* For interface with expression ().  */
+extern char *input_line_pointer;
+
+/* Hash table for instruction mnemonic lookup.  */
+static struct hash_control *op_hash;
+
+/* Hash table for register lookup.  */
+static struct hash_control *reg_hash;
+\f
 void
 i386_align_code (fragP, count)
      fragS *fragP;
@@ -446,40 +523,47 @@ i386_align_code (fragP, count)
     f32_15, f32_15, f32_15, f32_15, f32_15, f32_15, f32_15
   };
 
-  /* ??? We can't use these fillers for x86_64, since they often kills the
-     upper halves.  Solve later.  */
-  if (flag_code == CODE_64BIT)
-    count = 1;
+  if (count <= 0 || count > 15)
+    return;
 
-  if (count > 0 && count <= 15)
+  /* The recommended way to pad 64bit code is to use NOPs preceded by
+     maximally four 0x66 prefixes.  Balance the size of nops.  */
+  if (flag_code == CODE_64BIT)
     {
-      if (flag_code == CODE_16BIT)
+      int i;
+      int nnops = (count + 3) / 4;
+      int len = count / nnops;
+      int remains = count - nnops * len;
+      int pos = 0;
+
+      for (i = 0; i < remains; i++)
        {
-         memcpy (fragP->fr_literal + fragP->fr_fix,
-                 f16_patt[count - 1], count);
-         if (count > 8)
-           /* Adjust jump offset.  */
-           fragP->fr_literal[fragP->fr_fix + 1] = count - 2;
+         memset (fragP->fr_literal + fragP->fr_fix + pos, 0x66, len);
+         fragP->fr_literal[fragP->fr_fix + pos + len] = 0x90;
+         pos += len + 1;
+       }
+      for (; i < nnops; i++)
+       {
+         memset (fragP->fr_literal + fragP->fr_fix + pos, 0x66, len - 1);
+         fragP->fr_literal[fragP->fr_fix + pos + len - 1] = 0x90;
+         pos += len;
        }
-      else
-       memcpy (fragP->fr_literal + fragP->fr_fix,
-               f32_patt[count - 1], count);
-      fragP->fr_var = count;
     }
+  else
+    if (flag_code == CODE_16BIT)
+      {
+       memcpy (fragP->fr_literal + fragP->fr_fix,
+               f16_patt[count - 1], count);
+       if (count > 8)
+         /* Adjust jump offset.  */
+         fragP->fr_literal[fragP->fr_fix + 1] = count - 2;
+      }
+    else
+      memcpy (fragP->fr_literal + fragP->fr_fix,
+             f32_patt[count - 1], count);
+  fragP->fr_var = count;
 }
 
-static char *output_invalid PARAMS ((int c));
-static int i386_operand PARAMS ((char *operand_string));
-static int i386_intel_operand PARAMS ((char *operand_string, int got_a_float));
-static const reg_entry *parse_register PARAMS ((char *reg_string,
-                                               char **end_op));
-
-#ifndef I386COFF
-static void s_bss PARAMS ((int));
-#endif
-
-symbolS *GOT_symbol;           /* Pre-defined "_GLOBAL_OFFSET_TABLE_".  */
-
 static INLINE unsigned int
 mode_from_disp_size (t)
      unsigned int t;
@@ -540,8 +624,7 @@ static int
 smallest_imm_type (num)
      offsetT num;
 {
-  if (cpu_arch_flags != (Cpu086 | Cpu186 | Cpu286 | Cpu386 | Cpu486 | CpuNo64)
-      && !(cpu_arch_flags & (CpuUnknown)))
+  if (cpu_arch_flags != (Cpu086 | Cpu186 | Cpu286 | Cpu386 | Cpu486 | CpuNo64))
     {
       /* This code is disabled on the 486 because all the Imm1 forms
         in the opcode table are slower on the i486.  They're the
@@ -608,7 +691,8 @@ add_prefix (prefix)
   int ret = 1;
   int q;
 
-  if (prefix >= 0x40 && prefix < 0x50 && flag_code == CODE_64BIT)
+  if (prefix >= REX_OPCODE && prefix < REX_OPCODE + 16
+      && flag_code == CODE_64BIT)
     q = REX_PREFIX;
   else
     switch (prefix)
@@ -646,7 +730,7 @@ add_prefix (prefix)
        break;
       }
 
-  if (i.prefix[q])
+  if (i.prefix[q] != 0)
     {
       as_bad (_("same type of prefix used twice"));
       return 0;
@@ -693,7 +777,7 @@ set_intel_syntax (syntax_flag)
   int ask_naked_reg = 0;
 
   SKIP_WHITESPACE ();
-  if (! is_end_of_line[(unsigned char) *input_line_pointer])
+  if (!is_end_of_line[(unsigned char) *input_line_pointer])
     {
       char *string = input_line_pointer;
       int e = get_symbol_end ();
@@ -730,7 +814,7 @@ set_cpu_arch (dummy)
 {
   SKIP_WHITESPACE ();
 
-  if (! is_end_of_line[(unsigned char) *input_line_pointer])
+  if (!is_end_of_line[(unsigned char) *input_line_pointer])
     {
       char *string = input_line_pointer;
       int e = get_symbol_end ();
@@ -756,7 +840,7 @@ set_cpu_arch (dummy)
 
   no_cond_jump_promotion = 0;
   if (*input_line_pointer == ','
-      && ! is_end_of_line[(unsigned char) input_line_pointer[1]])
+      && !is_end_of_line[(unsigned char) input_line_pointer[1]])
     {
       char *string = ++input_line_pointer;
       int e = get_symbol_end ();
@@ -774,43 +858,6 @@ set_cpu_arch (dummy)
   demand_empty_rest_of_line ();
 }
 
-const pseudo_typeS md_pseudo_table[] =
-{
-#if !defined(OBJ_AOUT) && !defined(USE_ALIGN_PTWO)
-  {"align", s_align_bytes, 0},
-#else
-  {"align", s_align_ptwo, 0},
-#endif
-  {"arch", set_cpu_arch, 0},
-#ifndef I386COFF
-  {"bss", s_bss, 0},
-#endif
-  {"ffloat", float_cons, 'f'},
-  {"dfloat", float_cons, 'd'},
-  {"tfloat", float_cons, 'x'},
-  {"value", cons, 2},
-  {"noopt", s_ignore, 0},
-  {"optim", s_ignore, 0},
-  {"code16gcc", set_16bit_gcc_code_flag, CODE_16BIT},
-  {"code16", set_code_flag, CODE_16BIT},
-  {"code32", set_code_flag, CODE_32BIT},
-  {"code64", set_code_flag, CODE_64BIT},
-  {"intel_syntax", set_intel_syntax, 1},
-  {"att_syntax", set_intel_syntax, 0},
-  {"file", dwarf2_directive_file, 0},
-  {"loc", dwarf2_directive_loc, 0},
-  {0, 0, 0}
-};
-
-/* For interface with expression ().  */
-extern char *input_line_pointer;
-
-/* Hash table for instruction mnemonic lookup.  */
-static struct hash_control *op_hash;
-
-/* Hash table for register lookup.  */
-static struct hash_control *reg_hash;
-\f
 #ifdef BFD_ASSEMBLER
 unsigned long
 i386_mach ()
@@ -833,8 +880,8 @@ md_begin ()
   op_hash = hash_new ();
 
   {
-    register const template *optab;
-    register templates *core_optab;
+    const template *optab;
+    templates *core_optab;
 
     /* Setup for loop.  */
     optab = i386_optab;
@@ -870,7 +917,7 @@ md_begin ()
   /* Initialize reg_hash hash table.  */
   reg_hash = hash_new ();
   {
-    register const reg_entry *regtab;
+    const reg_entry *regtab;
 
     for (regtab = i386_regtab;
         regtab < i386_regtab + sizeof (i386_regtab) / sizeof (i386_regtab[0]);
@@ -886,32 +933,32 @@ md_begin ()
 
   /* Fill in lexical tables:  mnemonic_chars, operand_chars.  */
   {
-    register int c;
-    register char *p;
+    int c;
+    char *p;
 
     for (c = 0; c < 256; c++)
       {
-       if (isdigit (c))
+       if (ISDIGIT (c))
          {
            digit_chars[c] = c;
            mnemonic_chars[c] = c;
            register_chars[c] = c;
            operand_chars[c] = c;
          }
-       else if (islower (c))
+       else if (ISLOWER (c))
          {
            mnemonic_chars[c] = c;
            register_chars[c] = c;
            operand_chars[c] = c;
          }
-       else if (isupper (c))
+       else if (ISUPPER (c))
          {
-           mnemonic_chars[c] = tolower (c);
+           mnemonic_chars[c] = TOLOWER (c);
            register_chars[c] = mnemonic_chars[c];
            operand_chars[c] = c;
          }
 
-       if (isalpha (c) || isdigit (c))
+       if (ISALPHA (c) || ISDIGIT (c))
          identifier_chars[c] = c;
        else if (c >= 128)
          {
@@ -976,7 +1023,10 @@ pi (line, x)
   fprintf (stdout, "  sib:  base %x  index %x  scale %x\n",
           x->sib.base, x->sib.index, x->sib.scale);
   fprintf (stdout, "  rex: 64bit %x  extX %x  extY %x  extZ %x\n",
-          x->rex.mode64, x->rex.extX, x->rex.extY, x->rex.extZ);
+          (x->rex & REX_MODE64) != 0,
+          (x->rex & REX_EXTX) != 0,
+          (x->rex & REX_EXTY) != 0,
+          (x->rex & REX_EXTZ) != 0);
   for (i = 0; i < x->operands; i++)
     {
       fprintf (stdout, "    #%d:  ", i + 1);
@@ -1051,7 +1101,7 @@ struct type_name
     char *tname;
   }
 
-type_names[] =
+static const type_names[] =
 {
   { Reg8, "r8" },
   { Reg16, "r16" },
@@ -1091,7 +1141,7 @@ static void
 pt (t)
      unsigned int t;
 {
-  register struct type_name *ty;
+  const struct type_name *ty;
 
   for (ty = type_names; ty->mask; ty++)
     if (t & ty->mask)
@@ -1101,22 +1151,9 @@ pt (t)
 
 #endif /* DEBUG386 */
 \f
-int
-tc_i386_force_relocation (fixp)
-     struct fix *fixp;
-{
-#ifdef BFD_ASSEMBLER
-  if (fixp->fx_r_type == BFD_RELOC_VTABLE_INHERIT
-      || fixp->fx_r_type == BFD_RELOC_VTABLE_ENTRY)
-    return 1;
-  return 0;
-#else
-  /* For COFF.  */
-  return fixp->fx_r_type == 7;
-#endif
-}
-
 #ifdef BFD_ASSEMBLER
+static bfd_reloc_code_real_type reloc
+  PARAMS ((int, int, int, bfd_reloc_code_real_type));
 
 static bfd_reloc_code_real_type
 reloc (size, pcrel, sign, other)
@@ -1170,46 +1207,79 @@ reloc (size, pcrel, sign, other)
 
 int
 tc_i386_fix_adjustable (fixP)
-     fixS *fixP;
+     fixS *fixP ATTRIBUTE_UNUSED;
 {
 #if defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF)
-  /* Prevent all adjustments to global symbols, or else dynamic
-     linking will not work correctly.  */
-  if (S_IS_EXTERNAL (fixP->fx_addsy)
-      || S_IS_WEAK (fixP->fx_addsy))
+  if (OUTPUT_FLAVOR != bfd_target_elf_flavour)
+    return 1;
+
+  /* Don't adjust pc-relative references to merge sections in 64-bit
+     mode.  */
+  if (use_rela_relocations
+      && (S_GET_SEGMENT (fixP->fx_addsy)->flags & SEC_MERGE) != 0
+      && fixP->fx_pcrel)
     return 0;
-#endif
+
   /* adjust_reloc_syms doesn't know about the GOT.  */
   if (fixP->fx_r_type == BFD_RELOC_386_GOTOFF
       || fixP->fx_r_type == BFD_RELOC_386_PLT32
       || fixP->fx_r_type == BFD_RELOC_386_GOT32
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_GD
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_LDM
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_LDO_32
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_IE_32
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_IE
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_GOTIE
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_LE_32
+      || fixP->fx_r_type == BFD_RELOC_386_TLS_LE
       || fixP->fx_r_type == BFD_RELOC_X86_64_PLT32
       || fixP->fx_r_type == BFD_RELOC_X86_64_GOT32
       || fixP->fx_r_type == BFD_RELOC_X86_64_GOTPCREL
+      || fixP->fx_r_type == BFD_RELOC_X86_64_TLSGD
+      || fixP->fx_r_type == BFD_RELOC_X86_64_TLSLD
+      || fixP->fx_r_type == BFD_RELOC_X86_64_DTPOFF32
+      || fixP->fx_r_type == BFD_RELOC_X86_64_GOTTPOFF
+      || fixP->fx_r_type == BFD_RELOC_X86_64_TPOFF32
       || fixP->fx_r_type == BFD_RELOC_VTABLE_INHERIT
       || fixP->fx_r_type == BFD_RELOC_VTABLE_ENTRY)
     return 0;
+#endif
   return 1;
 }
 #else
 #define reloc(SIZE,PCREL,SIGN,OTHER)   0
+#define BFD_RELOC_8                    0
 #define BFD_RELOC_16                   0
 #define BFD_RELOC_32                   0
+#define BFD_RELOC_8_PCREL              0
 #define BFD_RELOC_16_PCREL             0
 #define BFD_RELOC_32_PCREL             0
 #define BFD_RELOC_386_PLT32            0
 #define BFD_RELOC_386_GOT32            0
 #define BFD_RELOC_386_GOTOFF           0
+#define BFD_RELOC_386_TLS_GD           0
+#define BFD_RELOC_386_TLS_LDM          0
+#define BFD_RELOC_386_TLS_LDO_32       0
+#define BFD_RELOC_386_TLS_IE_32                0
+#define BFD_RELOC_386_TLS_IE           0
+#define BFD_RELOC_386_TLS_GOTIE                0
+#define BFD_RELOC_386_TLS_LE_32                0
+#define BFD_RELOC_386_TLS_LE           0
 #define BFD_RELOC_X86_64_PLT32         0
 #define BFD_RELOC_X86_64_GOT32         0
 #define BFD_RELOC_X86_64_GOTPCREL      0
+#define BFD_RELOC_X86_64_TLSGD         0
+#define BFD_RELOC_X86_64_TLSLD         0
+#define BFD_RELOC_X86_64_DTPOFF32      0
+#define BFD_RELOC_X86_64_GOTTPOFF      0
+#define BFD_RELOC_X86_64_TPOFF32       0
 #endif
 
-static int intel_float_operand PARAMS ((char *mnemonic));
+static int intel_float_operand PARAMS ((const char *mnemonic));
 
 static int
 intel_float_operand (mnemonic)
-     char *mnemonic;
+     const char *mnemonic;
 {
   if (mnemonic[0] == 'f' && mnemonic[1] == 'i')
     return 2;
@@ -1228,11 +1298,7 @@ void
 md_assemble (line)
      char *line;
 {
-  /* Points to template once we've found it.  */
-  const template *t;
-
   int j;
-
   char mnemonic[MAX_MNEM_SIZE];
 
   /* Initialize globals.  */
@@ -1246,1900 +1312,2161 @@ md_assemble (line)
   /* First parse an instruction mnemonic & call i386_operand for the operands.
      We assume that the scrubber has arranged it so that line[0] is the valid
      start of a (possibly prefixed) mnemonic.  */
-  {
-    char *l = line;
-    char *token_start = l;
-    char *mnem_p;
-
-    /* Non-zero if we found a prefix only acceptable with string insns.  */
-    const char *expecting_string_instruction = NULL;
-
-    while (1)
-      {
-       mnem_p = mnemonic;
-       while ((*mnem_p = mnemonic_chars[(unsigned char) *l]) != 0)
-         {
-           mnem_p++;
-           if (mnem_p >= mnemonic + sizeof (mnemonic))
-             {
-               as_bad (_("no such instruction: `%s'"), token_start);
-               return;
-             }
-           l++;
-         }
-       if (!is_space_char (*l)
-           && *l != END_OF_INSN
-           && *l != PREFIX_SEPARATOR
-           && *l != ',')
-         {
-           as_bad (_("invalid character %s in mnemonic"),
-                   output_invalid (*l));
-           return;
-         }
-       if (token_start == l)
-         {
-           if (*l == PREFIX_SEPARATOR)
-             as_bad (_("expecting prefix; got nothing"));
-           else
-             as_bad (_("expecting mnemonic; got nothing"));
-           return;
-         }
 
-       /* Look up instruction (or prefix) via hash table.  */
-       current_templates = hash_find (op_hash, mnemonic);
+  line = parse_insn (line, mnemonic);
+  if (line == NULL)
+    return;
 
-       if (*l != END_OF_INSN
-           && (! is_space_char (*l) || l[1] != END_OF_INSN)
-           && current_templates
-           && (current_templates->start->opcode_modifier & IsPrefix))
-         {
-           /* If we are in 16-bit mode, do not allow addr16 or data16.
-              Similarly, in 32-bit mode, do not allow addr32 or data32.  */
-           if ((current_templates->start->opcode_modifier & (Size16 | Size32))
-               && (((current_templates->start->opcode_modifier & Size32) != 0)
-                   ^ (flag_code == CODE_16BIT)))
-             {
-               as_bad (_("redundant %s prefix"),
-                       current_templates->start->name);
-               return;
-             }
-           /* Add prefix, checking for repeated prefixes.  */
-           switch (add_prefix (current_templates->start->base_opcode))
-             {
-             case 0:
-               return;
-             case 2:
-               expecting_string_instruction = current_templates->start->name;
-               break;
-             }
-           /* Skip past PREFIX_SEPARATOR and reset token_start.  */
-           token_start = ++l;
-         }
-       else
-         break;
-      }
+  line = parse_operands (line, mnemonic);
+  if (line == NULL)
+    return;
 
-    if (!current_templates)
-      {
-       /* See if we can get a match by trimming off a suffix.  */
-       switch (mnem_p[-1])
-         {
-         case WORD_MNEM_SUFFIX:
-         case BYTE_MNEM_SUFFIX:
-         case QWORD_MNEM_SUFFIX:
-           i.suffix = mnem_p[-1];
-           mnem_p[-1] = '\0';
-           current_templates = hash_find (op_hash, mnemonic);
-           break;
-         case SHORT_MNEM_SUFFIX:
-         case LONG_MNEM_SUFFIX:
-           if (!intel_syntax)
-             {
-               i.suffix = mnem_p[-1];
-               mnem_p[-1] = '\0';
-               current_templates = hash_find (op_hash, mnemonic);
-             }
-           break;
+  /* Now we've parsed the mnemonic into a set of templates, and have the
+     operands at hand.  */
+
+  /* All intel opcodes have reversed operands except for "bound" and
+     "enter".  We also don't reverse intersegment "jmp" and "call"
+     instructions with 2 immediate operands so that the immediate segment
+     precedes the offset, as it does when in AT&T mode.  "enter" and the
+     intersegment "jmp" and "call" instructions are the only ones that
+     have two immediate operands.  */
+  if (intel_syntax && i.operands > 1
+      && (strcmp (mnemonic, "bound") != 0)
+      && !((i.types[0] & Imm) && (i.types[1] & Imm)))
+    swap_operands ();
+
+  if (i.imm_operands)
+    optimize_imm ();
+
+  if (i.disp_operands)
+    optimize_disp ();
+
+  /* Next, we find a template that matches the given insn,
+     making sure the overlap of the given operands types is consistent
+     with the template operand types.  */
 
-         /* Intel Syntax.  */
-         case 'd':
-           if (intel_syntax)
-             {
-               if (intel_float_operand (mnemonic))
-                 i.suffix = SHORT_MNEM_SUFFIX;
-               else
-                 i.suffix = LONG_MNEM_SUFFIX;
-               mnem_p[-1] = '\0';
-               current_templates = hash_find (op_hash, mnemonic);
-             }
-           break;
-         }
-       if (!current_templates)
-         {
-           as_bad (_("no such instruction: `%s'"), token_start);
-           return;
-         }
-      }
+  if (!match_template ())
+    return;
 
-    if (current_templates->start->opcode_modifier & (Jump | JumpByte))
-      {
-       /* Check for a branch hint.  We allow ",pt" and ",pn" for
-          predict taken and predict not taken respectively.
-          I'm not sure that branch hints actually do anything on loop
-          and jcxz insns (JumpByte) for current Pentium4 chips.  They
-          may work in the future and it doesn't hurt to accept them
-          now.  */
-       if (l[0] == ',' && l[1] == 'p')
-         {
-           if (l[2] == 't')
-             {
-               if (! add_prefix (DS_PREFIX_OPCODE))
-                 return;
-               l += 3;
-             }
-           else if (l[2] == 'n')
-             {
-               if (! add_prefix (CS_PREFIX_OPCODE))
-                 return;
-               l += 3;
-             }
-         }
-      }
-    /* Any other comma loses.  */
-    if (*l == ',')
-      {
-       as_bad (_("invalid character %s in mnemonic"),
-               output_invalid (*l));
-       return;
-      }
+  if (intel_syntax)
+    {
+      /* Undo SYSV386_COMPAT brokenness when in Intel mode.  See i386.h  */
+      if (SYSV386_COMPAT
+         && (i.tm.base_opcode & 0xfffffde0) == 0xdce0)
+       i.tm.base_opcode ^= FloatR;
+
+      /* Zap movzx and movsx suffix.  The suffix may have been set from
+        "word ptr" or "byte ptr" on the source operand, but we'll use
+        the suffix later to choose the destination register.  */
+      if ((i.tm.base_opcode & ~9) == 0x0fb6)
+       i.suffix = 0;
+    }
 
-    /* Check if instruction is supported on specified architecture.  */
-    if (cpu_arch_flags != 0)
-      {
-       if ((current_templates->start->cpu_flags & ~(Cpu64 | CpuNo64))
-           & ~(cpu_arch_flags & ~(Cpu64 | CpuNo64)))
-         {
-           as_warn (_("`%s' is not supported on `%s'"),
-                    current_templates->start->name, cpu_arch_name);
-         }
-       else if ((Cpu386 & ~cpu_arch_flags) && (flag_code != CODE_16BIT))
-         {
-           as_warn (_("use .code16 to ensure correct addressing mode"));
-         }
-      }
+  if (i.tm.opcode_modifier & FWait)
+    if (!add_prefix (FWAIT_OPCODE))
+      return;
 
-    /* Check for rep/repne without a string instruction.  */
-    if (expecting_string_instruction
-       && !(current_templates->start->opcode_modifier & IsString))
-      {
-       as_bad (_("expecting string instruction after `%s'"),
-               expecting_string_instruction);
+  /* Check string instruction segment overrides.  */
+  if ((i.tm.opcode_modifier & IsString) != 0 && i.mem_operands != 0)
+    {
+      if (!check_string ())
        return;
-      }
+    }
 
-    /* There may be operands to parse.  */
-    if (*l != END_OF_INSN)
-      {
-       /* 1 if operand is pending after ','.  */
-       unsigned int expecting_operand = 0;
+  if (!process_suffix ())
+    return;
 
-       /* Non-zero if operand parens not balanced.  */
-       unsigned int paren_not_balanced;
+  /* Make still unresolved immediate matches conform to size of immediate
+     given in i.suffix.  */
+  if (!finalize_imm ())
+    return;
 
-       do
-         {
-           /* Skip optional white space before operand.  */
-           if (is_space_char (*l))
-             ++l;
-           if (!is_operand_char (*l) && *l != END_OF_INSN)
-             {
-               as_bad (_("invalid character %s before operand %d"),
-                       output_invalid (*l),
-                       i.operands + 1);
-               return;
-             }
-           token_start = l;    /* after white space */
-           paren_not_balanced = 0;
-           while (paren_not_balanced || *l != ',')
-             {
-               if (*l == END_OF_INSN)
-                 {
-                   if (paren_not_balanced)
-                     {
-                       if (!intel_syntax)
-                         as_bad (_("unbalanced parenthesis in operand %d."),
-                                 i.operands + 1);
-                       else
-                         as_bad (_("unbalanced brackets in operand %d."),
-                                 i.operands + 1);
-                       return;
-                     }
-                   else
-                     break;    /* we are done */
-                 }
-               else if (!is_operand_char (*l) && !is_space_char (*l))
-                 {
-                   as_bad (_("invalid character %s in operand %d"),
-                           output_invalid (*l),
-                           i.operands + 1);
-                   return;
-                 }
-               if (!intel_syntax)
-                 {
-                   if (*l == '(')
-                     ++paren_not_balanced;
-                   if (*l == ')')
-                     --paren_not_balanced;
-                 }
-               else
-                 {
-                   if (*l == '[')
-                     ++paren_not_balanced;
-                   if (*l == ']')
-                     --paren_not_balanced;
-                 }
-               l++;
-             }
-           if (l != token_start)
-             {                 /* Yes, we've read in another operand.  */
-               unsigned int operand_ok;
-               this_operand = i.operands++;
-               if (i.operands > MAX_OPERANDS)
-                 {
-                   as_bad (_("spurious operands; (%d operands/instruction max)"),
-                           MAX_OPERANDS);
-                   return;
-                 }
-               /* Now parse operand adding info to 'i' as we go along.  */
-               END_STRING_AND_SAVE (l);
-
-               if (intel_syntax)
-                 operand_ok =
-                   i386_intel_operand (token_start,
-                                       intel_float_operand (mnemonic));
-               else
-                 operand_ok = i386_operand (token_start);
-
-               RESTORE_END_STRING (l);
-               if (!operand_ok)
-                 return;
-             }
-           else
-             {
-               if (expecting_operand)
-                 {
-                 expecting_operand_after_comma:
-                   as_bad (_("expecting operand after ','; got nothing"));
-                   return;
-                 }
-               if (*l == ',')
-                 {
-                   as_bad (_("expecting operand before ','; got nothing"));
-                   return;
-                 }
-             }
+  if (i.types[0] & Imm1)
+    i.imm_operands = 0;        /* kludge for shift insns.  */
+  if (i.types[0] & ImplicitRegister)
+    i.reg_operands--;
+  if (i.types[1] & ImplicitRegister)
+    i.reg_operands--;
+  if (i.types[2] & ImplicitRegister)
+    i.reg_operands--;
 
-           /* Now *l must be either ',' or END_OF_INSN.  */
-           if (*l == ',')
-             {
-               if (*++l == END_OF_INSN)
-                 {
-                   /* Just skip it, if it's \n complain.  */
-                   goto expecting_operand_after_comma;
-                 }
-               expecting_operand = 1;
-             }
-         }
-       while (*l != END_OF_INSN);
-      }
-  }
+  if (i.tm.opcode_modifier & ImmExt)
+    {
+      /* These AMD 3DNow! and Intel Katmai New Instructions have an
+        opcode suffix which is coded in the same place as an 8-bit
+        immediate field would be.  Here we fake an 8-bit immediate
+        operand from the opcode suffix stored in tm.extension_opcode.  */
 
-  /* Now we've parsed the mnemonic into a set of templates, and have the
-     operands at hand.
+      expressionS *exp;
 
-     Next, we find a template that matches the given insn,
-     making sure the overlap of the given operands types is consistent
-     with the template operand types.  */
+      assert (i.imm_operands == 0 && i.operands <= 2 && 2 < MAX_OPERANDS);
 
-#define MATCH(overlap, given, template) \
-  ((overlap & ~JumpAbsolute) \
-   && ((given) & (BaseIndex|JumpAbsolute)) == ((overlap) & (BaseIndex|JumpAbsolute)))
+      exp = &im_expressions[i.imm_operands++];
+      i.op[i.operands].imms = exp;
+      i.types[i.operands++] = Imm8;
+      exp->X_op = O_constant;
+      exp->X_add_number = i.tm.extension_opcode;
+      i.tm.extension_opcode = None;
+    }
 
-  /* If given types r0 and r1 are registers they must be of the same type
-     unless the expected operand type register overlap is null.
-     Note that Acc in a template matches every size of reg.  */
-#define CONSISTENT_REGISTER_MATCH(m0, g0, t0, m1, g1, t1) \
-  ( ((g0) & Reg) == 0 || ((g1) & Reg) == 0 || \
-    ((g0) & Reg) == ((g1) & Reg) || \
-    ((((m0) & Acc) ? Reg : (t0)) & (((m1) & Acc) ? Reg : (t1)) & Reg) == 0 )
+  /* For insns with operands there are more diddles to do to the opcode.  */
+  if (i.operands)
+    {
+      if (!process_operands ())
+       return;
+    }
+  else if (!quiet_warnings && (i.tm.opcode_modifier & Ugh) != 0)
+    {
+      /* UnixWare fsub no args is alias for fsubp, fadd -> faddp, etc.  */
+      as_warn (_("translating to `%sp'"), i.tm.name);
+    }
 
-  {
-    register unsigned int overlap0, overlap1;
-    unsigned int overlap2;
-    unsigned int found_reverse_match;
-    int suffix_check;
-
-    /* All intel opcodes have reversed operands except for "bound" and
-       "enter".  We also don't reverse intersegment "jmp" and "call"
-       instructions with 2 immediate operands so that the immediate segment
-       precedes the offset, as it does when in AT&T mode.  "enter" and the
-       intersegment "jmp" and "call" instructions are the only ones that
-       have two immediate operands.  */
-    if (intel_syntax && i.operands > 1
-       && (strcmp (mnemonic, "bound") != 0)
-       && !((i.types[0] & Imm) && (i.types[1] & Imm)))
-      {
-       union i386_op temp_op;
-       unsigned int temp_type;
-       RELOC_ENUM temp_reloc;
-       int xchg1 = 0;
-       int xchg2 = 0;
+  /* Handle conversion of 'int $3' --> special int3 insn.  */
+  if (i.tm.base_opcode == INT_OPCODE && i.op[0].imms->X_add_number == 3)
+    {
+      i.tm.base_opcode = INT3_OPCODE;
+      i.imm_operands = 0;
+    }
 
-       if (i.operands == 2)
-         {
-           xchg1 = 0;
-           xchg2 = 1;
-         }
-       else if (i.operands == 3)
-         {
-           xchg1 = 0;
-           xchg2 = 2;
-         }
-       temp_type = i.types[xchg2];
-       i.types[xchg2] = i.types[xchg1];
-       i.types[xchg1] = temp_type;
-       temp_op = i.op[xchg2];
-       i.op[xchg2] = i.op[xchg1];
-       i.op[xchg1] = temp_op;
-       temp_reloc = i.reloc[xchg2];
-       i.reloc[xchg2] = i.reloc[xchg1];
-       i.reloc[xchg1] = temp_reloc;
-
-       if (i.mem_operands == 2)
-         {
-           const seg_entry *temp_seg;
-           temp_seg = i.seg[0];
-           i.seg[0] = i.seg[1];
-           i.seg[1] = temp_seg;
-         }
-      }
+  if ((i.tm.opcode_modifier & (Jump | JumpByte | JumpDword))
+      && i.op[0].disps->X_op == O_constant)
+    {
+      /* Convert "jmp constant" (and "call constant") to a jump (call) to
+        the absolute address given by the constant.  Since ix86 jumps and
+        calls are pc relative, we need to generate a reloc.  */
+      i.op[0].disps->X_add_symbol = &abs_symbol;
+      i.op[0].disps->X_op = O_symbol;
+    }
 
-    if (i.imm_operands)
-      {
-       /* Try to ensure constant immediates are represented in the smallest
-          opcode possible.  */
-       char guess_suffix = 0;
-       int op;
-
-       if (i.suffix)
-         guess_suffix = i.suffix;
-       else if (i.reg_operands)
-         {
-           /* Figure out a suffix from the last register operand specified.
-              We can't do this properly yet, ie. excluding InOutPortReg,
-              but the following works for instructions with immediates.
-              In any case, we can't set i.suffix yet.  */
-           for (op = i.operands; --op >= 0;)
-             if (i.types[op] & Reg)
-               {
-                 if (i.types[op] & Reg8)
-                   guess_suffix = BYTE_MNEM_SUFFIX;
-                 else if (i.types[op] & Reg16)
-                   guess_suffix = WORD_MNEM_SUFFIX;
-                 else if (i.types[op] & Reg32)
-                   guess_suffix = LONG_MNEM_SUFFIX;
-                 else if (i.types[op] & Reg64)
-                   guess_suffix = QWORD_MNEM_SUFFIX;
-                 break;
-               }
-         }
-       else if ((flag_code == CODE_16BIT) ^ (i.prefix[DATA_PREFIX] != 0))
-         guess_suffix = WORD_MNEM_SUFFIX;
+  if ((i.tm.opcode_modifier & Rex64) != 0)
+    i.rex |= REX_MODE64;
 
-       for (op = i.operands; --op >= 0;)
-         if (i.types[op] & Imm)
+  /* For 8 bit registers we need an empty rex prefix.  Also if the
+     instruction already has a prefix, we need to convert old
+     registers to new ones.  */
+
+  if (((i.types[0] & Reg8) != 0
+       && (i.op[0].regs->reg_flags & RegRex64) != 0)
+      || ((i.types[1] & Reg8) != 0
+         && (i.op[1].regs->reg_flags & RegRex64) != 0)
+      || (((i.types[0] & Reg8) != 0 || (i.types[1] & Reg8) != 0)
+         && i.rex != 0))
+    {
+      int x;
+
+      i.rex |= REX_OPCODE;
+      for (x = 0; x < 2; x++)
+       {
+         /* Look for 8 bit operand that uses old registers.  */
+         if ((i.types[x] & Reg8) != 0
+             && (i.op[x].regs->reg_flags & RegRex64) == 0)
            {
-             switch (i.op[op].imms->X_op)
-               {
-                 case O_constant:
-                   /* If a suffix is given, this operand may be shortened.  */
-                   switch (guess_suffix)
-                     {
-                     case LONG_MNEM_SUFFIX:
-                       i.types[op] |= Imm32 | Imm64;
-                       break;
-                     case WORD_MNEM_SUFFIX:
-                       i.types[op] |= Imm16 | Imm32S | Imm32 | Imm64;
-                       break;
-                     case BYTE_MNEM_SUFFIX:
-                       i.types[op] |= Imm16 | Imm8 | Imm8S | Imm32S | Imm32 | Imm64;
-                       break;
-                     }
+             /* In case it is "hi" register, give up.  */
+             if (i.op[x].regs->reg_num > 3)
+               as_bad (_("can't encode register '%%%s' in an instruction requiring REX prefix.\n"),
+                       i.op[x].regs->reg_name);
 
-                   /* If this operand is at most 16 bits, convert it
-                      to a signed 16 bit number before trying to see
-                      whether it will fit in an even smaller size.
-                      This allows a 16-bit operand such as $0xffe0 to
-                      be recognised as within Imm8S range.  */
-                   if ((i.types[op] & Imm16)
-                       && (i.op[op].imms->X_add_number & ~(offsetT) 0xffff) == 0)
-                     {
-                       i.op[op].imms->X_add_number =
-                         (((i.op[op].imms->X_add_number & 0xffff) ^ 0x8000) - 0x8000);
-                     }
-                   if ((i.types[op] & Imm32)
-                       && (i.op[op].imms->X_add_number & ~(((offsetT) 2 << 31) - 1)) == 0)
-                     {
-                       i.op[op].imms->X_add_number =
-                         (i.op[op].imms->X_add_number ^ ((offsetT) 1 << 31)) - ((addressT) 1 << 31);
-                     }
-                   i.types[op] |= smallest_imm_type (i.op[op].imms->X_add_number);
-                   /* We must avoid matching of Imm32 templates when 64bit only immediate is available.  */
-                   if (guess_suffix == QWORD_MNEM_SUFFIX)
-                     i.types[op] &= ~Imm32;
-                   break;
-                 case O_absent:
-                 case O_register:
-                   abort ();
-                 /* Symbols and expressions.  */
-                 default:
-                   /* Convert symbolic operand to proper sizes for matching.  */
-                   switch (guess_suffix)
-                     {
-                       case QWORD_MNEM_SUFFIX:
-                         i.types[op] = Imm64 | Imm32S;
-                         break;
-                       case LONG_MNEM_SUFFIX:
-                         i.types[op] = Imm32 | Imm64;
-                         break;
-                       case WORD_MNEM_SUFFIX:
-                         i.types[op] = Imm16 | Imm32 | Imm64;
-                         break;
-                         break;
-                       case BYTE_MNEM_SUFFIX:
-                         i.types[op] = Imm8 | Imm8S | Imm16 | Imm32S | Imm32;
-                               break;
-                         break;
-                     }
-                   break;
-               }
+             /* Otherwise it is equivalent to the extended register.
+                Since the encoding doesn't change this is merely
+                cosmetic cleanup for debug output.  */
+
+             i.op[x].regs = i.op[x].regs + 8;
            }
-      }
+       }
+    }
 
-    if (i.disp_operands)
-      {
-       /* Try to use the smallest displacement type too.  */
-       int op;
+  if (i.rex != 0)
+    add_prefix (REX_OPCODE | i.rex);
+
+  /* We are ready to output the insn.  */
+  output_insn ();
+}
+
+static char *
+parse_insn (line, mnemonic)
+     char *line;
+     char *mnemonic;
+{
+  char *l = line;
+  char *token_start = l;
+  char *mnem_p;
 
-       for (op = i.operands; --op >= 0;)
-         if ((i.types[op] & Disp)
-             && i.op[op].disps->X_op == O_constant)
+  /* Non-zero if we found a prefix only acceptable with string insns.  */
+  const char *expecting_string_instruction = NULL;
+
+  while (1)
+    {
+      mnem_p = mnemonic;
+      while ((*mnem_p = mnemonic_chars[(unsigned char) *l]) != 0)
+       {
+         mnem_p++;
+         if (mnem_p >= mnemonic + MAX_MNEM_SIZE)
            {
-             offsetT disp = i.op[op].disps->X_add_number;
+             as_bad (_("no such instruction: `%s'"), token_start);
+             return NULL;
+           }
+         l++;
+       }
+      if (!is_space_char (*l)
+         && *l != END_OF_INSN
+         && *l != PREFIX_SEPARATOR
+         && *l != ',')
+       {
+         as_bad (_("invalid character %s in mnemonic"),
+                 output_invalid (*l));
+         return NULL;
+       }
+      if (token_start == l)
+       {
+         if (*l == PREFIX_SEPARATOR)
+           as_bad (_("expecting prefix; got nothing"));
+         else
+           as_bad (_("expecting mnemonic; got nothing"));
+         return NULL;
+       }
 
-             if (i.types[op] & Disp16)
-               {
-                 /* We know this operand is at most 16 bits, so
-                    convert to a signed 16 bit number before trying
-                    to see whether it will fit in an even smaller
-                    size.  */
+      /* Look up instruction (or prefix) via hash table.  */
+      current_templates = hash_find (op_hash, mnemonic);
 
-                 disp = (((disp & 0xffff) ^ 0x8000) - 0x8000);
-               }
-             else if (i.types[op] & Disp32)
-               {
-                 /* We know this operand is at most 32 bits, so convert to a
-                    signed 32 bit number before trying to see whether it will
-                    fit in an even smaller size.  */
-                 disp &= (((offsetT) 2 << 31) - 1);
-                 disp = (disp ^ ((offsetT) 1 << 31)) - ((addressT) 1 << 31);
-               }
-             if (flag_code == CODE_64BIT)
+      if (*l != END_OF_INSN
+         && (!is_space_char (*l) || l[1] != END_OF_INSN)
+         && current_templates
+         && (current_templates->start->opcode_modifier & IsPrefix))
+       {
+         /* If we are in 16-bit mode, do not allow addr16 or data16.
+            Similarly, in 32-bit mode, do not allow addr32 or data32.  */
+         if ((current_templates->start->opcode_modifier & (Size16 | Size32))
+             && flag_code != CODE_64BIT
+             && (((current_templates->start->opcode_modifier & Size32) != 0)
+                 ^ (flag_code == CODE_16BIT)))
+           {
+             as_bad (_("redundant %s prefix"),
+                     current_templates->start->name);
+             return NULL;
+           }
+         /* Add prefix, checking for repeated prefixes.  */
+         switch (add_prefix (current_templates->start->base_opcode))
+           {
+           case 0:
+             return NULL;
+           case 2:
+             expecting_string_instruction = current_templates->start->name;
+             break;
+           }
+         /* Skip past PREFIX_SEPARATOR and reset token_start.  */
+         token_start = ++l;
+       }
+      else
+       break;
+    }
+
+  if (!current_templates)
+    {
+      /* See if we can get a match by trimming off a suffix.  */
+      switch (mnem_p[-1])
+       {
+       case WORD_MNEM_SUFFIX:
+       case BYTE_MNEM_SUFFIX:
+       case QWORD_MNEM_SUFFIX:
+         i.suffix = mnem_p[-1];
+         mnem_p[-1] = '\0';
+         current_templates = hash_find (op_hash, mnemonic);
+         break;
+       case SHORT_MNEM_SUFFIX:
+       case LONG_MNEM_SUFFIX:
+         if (!intel_syntax)
+           {
+             i.suffix = mnem_p[-1];
+             mnem_p[-1] = '\0';
+             current_templates = hash_find (op_hash, mnemonic);
+           }
+         break;
+
+         /* Intel Syntax.  */
+       case 'd':
+         if (intel_syntax)
+           {
+             if (intel_float_operand (mnemonic))
+               i.suffix = SHORT_MNEM_SUFFIX;
+             else
+               i.suffix = LONG_MNEM_SUFFIX;
+             mnem_p[-1] = '\0';
+             current_templates = hash_find (op_hash, mnemonic);
+           }
+         break;
+       }
+      if (!current_templates)
+       {
+         as_bad (_("no such instruction: `%s'"), token_start);
+         return NULL;
+       }
+    }
+
+  if (current_templates->start->opcode_modifier & (Jump | JumpByte))
+    {
+      /* Check for a branch hint.  We allow ",pt" and ",pn" for
+        predict taken and predict not taken respectively.
+        I'm not sure that branch hints actually do anything on loop
+        and jcxz insns (JumpByte) for current Pentium4 chips.  They
+        may work in the future and it doesn't hurt to accept them
+        now.  */
+      if (l[0] == ',' && l[1] == 'p')
+       {
+         if (l[2] == 't')
+           {
+             if (!add_prefix (DS_PREFIX_OPCODE))
+               return NULL;
+             l += 3;
+           }
+         else if (l[2] == 'n')
+           {
+             if (!add_prefix (CS_PREFIX_OPCODE))
+               return NULL;
+             l += 3;
+           }
+       }
+    }
+  /* Any other comma loses.  */
+  if (*l == ',')
+    {
+      as_bad (_("invalid character %s in mnemonic"),
+             output_invalid (*l));
+      return NULL;
+    }
+
+  /* Check if instruction is supported on specified architecture.  */
+  if ((current_templates->start->cpu_flags & ~(Cpu64 | CpuNo64))
+      & ~(cpu_arch_flags & ~(Cpu64 | CpuNo64)))
+    {
+      as_warn (_("`%s' is not supported on `%s'"),
+              current_templates->start->name, cpu_arch_name);
+    }
+  else if ((Cpu386 & ~cpu_arch_flags) && (flag_code != CODE_16BIT))
+    {
+      as_warn (_("use .code16 to ensure correct addressing mode"));
+    }
+
+  /* Check for rep/repne without a string instruction.  */
+  if (expecting_string_instruction
+      && !(current_templates->start->opcode_modifier & IsString))
+    {
+      as_bad (_("expecting string instruction after `%s'"),
+             expecting_string_instruction);
+      return NULL;
+    }
+
+  return l;
+}
+
+static char *
+parse_operands (l, mnemonic)
+     char *l;
+     const char *mnemonic;
+{
+  char *token_start;
+
+  /* 1 if operand is pending after ','.  */
+  unsigned int expecting_operand = 0;
+
+  /* Non-zero if operand parens not balanced.  */
+  unsigned int paren_not_balanced;
+
+  while (*l != END_OF_INSN)
+    {
+      /* Skip optional white space before operand.  */
+      if (is_space_char (*l))
+       ++l;
+      if (!is_operand_char (*l) && *l != END_OF_INSN)
+       {
+         as_bad (_("invalid character %s before operand %d"),
+                 output_invalid (*l),
+                 i.operands + 1);
+         return NULL;
+       }
+      token_start = l; /* after white space */
+      paren_not_balanced = 0;
+      while (paren_not_balanced || *l != ',')
+       {
+         if (*l == END_OF_INSN)
+           {
+             if (paren_not_balanced)
                {
-                 if (fits_in_signed_long (disp))
-                   i.types[op] |= Disp32S;
-                 if (fits_in_unsigned_long (disp))
-                   i.types[op] |= Disp32;
+                 if (!intel_syntax)
+                   as_bad (_("unbalanced parenthesis in operand %d."),
+                           i.operands + 1);
+                 else
+                   as_bad (_("unbalanced brackets in operand %d."),
+                           i.operands + 1);
+                 return NULL;
                }
-             if ((i.types[op] & (Disp32 | Disp32S | Disp16))
-                 && fits_in_signed_byte (disp))
-               i.types[op] |= Disp8;
+             else
+               break;  /* we are done */
            }
-      }
+         else if (!is_operand_char (*l) && !is_space_char (*l))
+           {
+             as_bad (_("invalid character %s in operand %d"),
+                     output_invalid (*l),
+                     i.operands + 1);
+             return NULL;
+           }
+         if (!intel_syntax)
+           {
+             if (*l == '(')
+               ++paren_not_balanced;
+             if (*l == ')')
+               --paren_not_balanced;
+           }
+         else
+           {
+             if (*l == '[')
+               ++paren_not_balanced;
+             if (*l == ']')
+               --paren_not_balanced;
+           }
+         l++;
+       }
+      if (l != token_start)
+       {                       /* Yes, we've read in another operand.  */
+         unsigned int operand_ok;
+         this_operand = i.operands++;
+         if (i.operands > MAX_OPERANDS)
+           {
+             as_bad (_("spurious operands; (%d operands/instruction max)"),
+                     MAX_OPERANDS);
+             return NULL;
+           }
+         /* Now parse operand adding info to 'i' as we go along.  */
+         END_STRING_AND_SAVE (l);
 
-    overlap0 = 0;
-    overlap1 = 0;
-    overlap2 = 0;
-    found_reverse_match = 0;
-    suffix_check = (i.suffix == BYTE_MNEM_SUFFIX
-                   ? No_bSuf
-                   : (i.suffix == WORD_MNEM_SUFFIX
-                      ? No_wSuf
-                      : (i.suffix == SHORT_MNEM_SUFFIX
-                         ? No_sSuf
-                         : (i.suffix == LONG_MNEM_SUFFIX
-                            ? No_lSuf
-                            : (i.suffix == QWORD_MNEM_SUFFIX
-                               ? No_qSuf
-                               : (i.suffix == LONG_DOUBLE_MNEM_SUFFIX ? No_xSuf : 0))))));
-
-    for (t = current_templates->start;
-        t < current_templates->end;
-        t++)
-      {
-       /* Must have right number of operands.  */
-       if (i.operands != t->operands)
-         continue;
+         if (intel_syntax)
+           operand_ok =
+             i386_intel_operand (token_start,
+                                 intel_float_operand (mnemonic));
+         else
+           operand_ok = i386_operand (token_start);
 
-       /* Check the suffix, except for some instructions in intel mode.  */
-       if ((t->opcode_modifier & suffix_check)
-           && !(intel_syntax
-                && (t->opcode_modifier & IgnoreSize))
-           && !(intel_syntax
-                && t->base_opcode == 0xd9
-                && (t->extension_opcode == 5        /* 0xd9,5 "fldcw"  */
-                    || t->extension_opcode == 7)))  /* 0xd9,7 "f{n}stcw"  */
-         continue;
+         RESTORE_END_STRING (l);
+         if (!operand_ok)
+           return NULL;
+       }
+      else
+       {
+         if (expecting_operand)
+           {
+           expecting_operand_after_comma:
+             as_bad (_("expecting operand after ','; got nothing"));
+             return NULL;
+           }
+         if (*l == ',')
+           {
+             as_bad (_("expecting operand before ','; got nothing"));
+             return NULL;
+           }
+       }
+
+      /* Now *l must be either ',' or END_OF_INSN.  */
+      if (*l == ',')
+       {
+         if (*++l == END_OF_INSN)
+           {
+             /* Just skip it, if it's \n complain.  */
+             goto expecting_operand_after_comma;
+           }
+         expecting_operand = 1;
+       }
+    }
+  return l;
+}
+
+static void
+swap_operands ()
+{
+  union i386_op temp_op;
+  unsigned int temp_type;
+  RELOC_ENUM temp_reloc;
+  int xchg1 = 0;
+  int xchg2 = 0;
 
-       /* Do not verify operands when there are none.  */
-       else if (!t->operands)
+  if (i.operands == 2)
+    {
+      xchg1 = 0;
+      xchg2 = 1;
+    }
+  else if (i.operands == 3)
+    {
+      xchg1 = 0;
+      xchg2 = 2;
+    }
+  temp_type = i.types[xchg2];
+  i.types[xchg2] = i.types[xchg1];
+  i.types[xchg1] = temp_type;
+  temp_op = i.op[xchg2];
+  i.op[xchg2] = i.op[xchg1];
+  i.op[xchg1] = temp_op;
+  temp_reloc = i.reloc[xchg2];
+  i.reloc[xchg2] = i.reloc[xchg1];
+  i.reloc[xchg1] = temp_reloc;
+
+  if (i.mem_operands == 2)
+    {
+      const seg_entry *temp_seg;
+      temp_seg = i.seg[0];
+      i.seg[0] = i.seg[1];
+      i.seg[1] = temp_seg;
+    }
+}
+
+/* Try to ensure constant immediates are represented in the smallest
+   opcode possible.  */
+static void
+optimize_imm ()
+{
+  char guess_suffix = 0;
+  int op;
+
+  if (i.suffix)
+    guess_suffix = i.suffix;
+  else if (i.reg_operands)
+    {
+      /* Figure out a suffix from the last register operand specified.
+        We can't do this properly yet, ie. excluding InOutPortReg,
+        but the following works for instructions with immediates.
+        In any case, we can't set i.suffix yet.  */
+      for (op = i.operands; --op >= 0;)
+       if (i.types[op] & Reg)
          {
-           if (t->cpu_flags & ~cpu_arch_flags)
-             continue;
-           /* We've found a match; break out of loop.  */
+           if (i.types[op] & Reg8)
+             guess_suffix = BYTE_MNEM_SUFFIX;
+           else if (i.types[op] & Reg16)
+             guess_suffix = WORD_MNEM_SUFFIX;
+           else if (i.types[op] & Reg32)
+             guess_suffix = LONG_MNEM_SUFFIX;
+           else if (i.types[op] & Reg64)
+             guess_suffix = QWORD_MNEM_SUFFIX;
            break;
          }
+    }
+  else if ((flag_code == CODE_16BIT) ^ (i.prefix[DATA_PREFIX] != 0))
+    guess_suffix = WORD_MNEM_SUFFIX;
 
-       overlap0 = i.types[0] & t->operand_types[0];
-       switch (t->operands)
+  for (op = i.operands; --op >= 0;)
+    if (i.types[op] & Imm)
+      {
+       switch (i.op[op].imms->X_op)
          {
-         case 1:
-           if (!MATCH (overlap0, i.types[0], t->operand_types[0]))
-             continue;
-           break;
-         case 2:
-         case 3:
-           overlap1 = i.types[1] & t->operand_types[1];
-           if (!MATCH (overlap0, i.types[0], t->operand_types[0])
-               || !MATCH (overlap1, i.types[1], t->operand_types[1])
-               || !CONSISTENT_REGISTER_MATCH (overlap0, i.types[0],
-                                              t->operand_types[0],
-                                              overlap1, i.types[1],
-                                              t->operand_types[1]))
+         case O_constant:
+           /* If a suffix is given, this operand may be shortened.  */
+           switch (guess_suffix)
              {
-               /* Check if other direction is valid ...  */
-               if ((t->opcode_modifier & (D|FloatD)) == 0)
-                 continue;
+             case LONG_MNEM_SUFFIX:
+               i.types[op] |= Imm32 | Imm64;
+               break;
+             case WORD_MNEM_SUFFIX:
+               i.types[op] |= Imm16 | Imm32S | Imm32 | Imm64;
+               break;
+             case BYTE_MNEM_SUFFIX:
+               i.types[op] |= Imm16 | Imm8 | Imm8S | Imm32S | Imm32 | Imm64;
+               break;
+             }
 
-               /* Try reversing direction of operands.  */
-               overlap0 = i.types[0] & t->operand_types[1];
-               overlap1 = i.types[1] & t->operand_types[0];
-               if (!MATCH (overlap0, i.types[0], t->operand_types[1])
-                   || !MATCH (overlap1, i.types[1], t->operand_types[0])
-                   || !CONSISTENT_REGISTER_MATCH (overlap0, i.types[0],
-                                                  t->operand_types[1],
-                                                  overlap1, i.types[1],
-                                                  t->operand_types[0]))
-                 {
-                   /* Does not match either direction.  */
-                   continue;
-                 }
-               /* found_reverse_match holds which of D or FloatDR
-                  we've found.  */
-               found_reverse_match = t->opcode_modifier & (D|FloatDR);
+           /* If this operand is at most 16 bits, convert it
+              to a signed 16 bit number before trying to see
+              whether it will fit in an even smaller size.
+              This allows a 16-bit operand such as $0xffe0 to
+              be recognised as within Imm8S range.  */
+           if ((i.types[op] & Imm16)
+               && (i.op[op].imms->X_add_number & ~(offsetT) 0xffff) == 0)
+             {
+               i.op[op].imms->X_add_number =
+                 (((i.op[op].imms->X_add_number & 0xffff) ^ 0x8000) - 0x8000);
              }
-           /* Found a forward 2 operand match here.  */
-           else if (t->operands == 3)
+           if ((i.types[op] & Imm32)
+               && ((i.op[op].imms->X_add_number & ~(((offsetT) 2 << 31) - 1))
+                   == 0))
              {
-               /* Here we make use of the fact that there are no
-                  reverse match 3 operand instructions, and all 3
-                  operand instructions only need to be checked for
-                  register consistency between operands 2 and 3.  */
-               overlap2 = i.types[2] & t->operand_types[2];
-               if (!MATCH (overlap2, i.types[2], t->operand_types[2])
-                   || !CONSISTENT_REGISTER_MATCH (overlap1, i.types[1],
-                                                  t->operand_types[1],
-                                                  overlap2, i.types[2],
-                                                  t->operand_types[2]))
+               i.op[op].imms->X_add_number = ((i.op[op].imms->X_add_number
+                                               ^ ((offsetT) 1 << 31))
+                                              - ((offsetT) 1 << 31));
+             }
+           i.types[op] |= smallest_imm_type (i.op[op].imms->X_add_number);
 
-                 continue;
+           /* We must avoid matching of Imm32 templates when 64bit
+              only immediate is available.  */
+           if (guess_suffix == QWORD_MNEM_SUFFIX)
+             i.types[op] &= ~Imm32;
+           break;
+
+         case O_absent:
+         case O_register:
+           abort ();
+
+           /* Symbols and expressions.  */
+         default:
+           /* Convert symbolic operand to proper sizes for matching.  */
+           switch (guess_suffix)
+             {
+             case QWORD_MNEM_SUFFIX:
+               i.types[op] = Imm64 | Imm32S;
+               break;
+             case LONG_MNEM_SUFFIX:
+               i.types[op] = Imm32 | Imm64;
+               break;
+             case WORD_MNEM_SUFFIX:
+               i.types[op] = Imm16 | Imm32 | Imm64;
+               break;
+               break;
+             case BYTE_MNEM_SUFFIX:
+               i.types[op] = Imm8 | Imm8S | Imm16 | Imm32S | Imm32;
+               break;
+               break;
              }
-           /* Found either forward/reverse 2 or 3 operand match here:
-              slip through to break.  */
-         }
-       if (t->cpu_flags & ~cpu_arch_flags)
-         {
-           found_reverse_match = 0;
-           continue;
+           break;
          }
-       /* We've found a match; break out of loop.  */
-       break;
-      }
-    if (t == current_templates->end)
-      {
-       /* We found no match.  */
-       as_bad (_("suffix or operands invalid for `%s'"),
-               current_templates->start->name);
-       return;
       }
+}
+
+/* Try to use the smallest displacement type too.  */
+static void
+optimize_disp ()
+{
+  int op;
 
-    if (!quiet_warnings)
+  for (op = i.operands; --op >= 0;)
+    if ((i.types[op] & Disp) && i.op[op].disps->X_op == O_constant)
       {
-       if (!intel_syntax
-           && ((i.types[0] & JumpAbsolute)
-               != (t->operand_types[0] & JumpAbsolute)))
+       offsetT disp = i.op[op].disps->X_add_number;
+
+       if (i.types[op] & Disp16)
          {
-           as_warn (_("indirect %s without `*'"), t->name);
-         }
+           /* We know this operand is at most 16 bits, so
+              convert to a signed 16 bit number before trying
+              to see whether it will fit in an even smaller
+              size.  */
 
-       if ((t->opcode_modifier & (IsPrefix|IgnoreSize))
-           == (IsPrefix|IgnoreSize))
+           disp = (((disp & 0xffff) ^ 0x8000) - 0x8000);
+         }
+       else if (i.types[op] & Disp32)
          {
-           /* Warn them that a data or address size prefix doesn't
-              affect assembly of the next line of code.  */
-           as_warn (_("stand-alone `%s' prefix"), t->name);
+           /* We know this operand is at most 32 bits, so convert to a
+              signed 32 bit number before trying to see whether it will
+              fit in an even smaller size.  */
+           disp &= (((offsetT) 2 << 31) - 1);
+           disp = (disp ^ ((offsetT) 1 << 31)) - ((addressT) 1 << 31);
          }
+       if (flag_code == CODE_64BIT)
+         {
+           if (fits_in_signed_long (disp))
+             i.types[op] |= Disp32S;
+           if (fits_in_unsigned_long (disp))
+             i.types[op] |= Disp32;
+         }
+       if ((i.types[op] & (Disp32 | Disp32S | Disp16))
+           && fits_in_signed_byte (disp))
+         i.types[op] |= Disp8;
       }
+}
 
-    /* Copy the template we found.  */
-    i.tm = *t;
-    if (found_reverse_match)
-      {
-       /* If we found a reverse match we must alter the opcode
-          direction bit.  found_reverse_match holds bits to change
-          (different for int & float insns).  */
+static int
+match_template ()
+{
+  /* Points to template once we've found it.  */
+  const template *t;
+  unsigned int overlap0, overlap1, overlap2;
+  unsigned int found_reverse_match;
+  int suffix_check;
 
-       i.tm.base_opcode ^= found_reverse_match;
+#define MATCH(overlap, given, template)                                \
+  ((overlap & ~JumpAbsolute)                                   \
+   && (((given) & (BaseIndex | JumpAbsolute))                  \
+       == ((overlap) & (BaseIndex | JumpAbsolute))))
 
-       i.tm.operand_types[0] = t->operand_types[1];
-       i.tm.operand_types[1] = t->operand_types[0];
-      }
+  /* If given types r0 and r1 are registers they must be of the same type
+     unless the expected operand type register overlap is null.
+     Note that Acc in a template matches every size of reg.  */
+#define CONSISTENT_REGISTER_MATCH(m0, g0, t0, m1, g1, t1)      \
+  (((g0) & Reg) == 0 || ((g1) & Reg) == 0                      \
+   || ((g0) & Reg) == ((g1) & Reg)                             \
+   || ((((m0) & Acc) ? Reg : (t0)) & (((m1) & Acc) ? Reg : (t1)) & Reg) == 0 )
+
+  overlap0 = 0;
+  overlap1 = 0;
+  overlap2 = 0;
+  found_reverse_match = 0;
+  suffix_check = (i.suffix == BYTE_MNEM_SUFFIX
+                 ? No_bSuf
+                 : (i.suffix == WORD_MNEM_SUFFIX
+                    ? No_wSuf
+                    : (i.suffix == SHORT_MNEM_SUFFIX
+                       ? No_sSuf
+                       : (i.suffix == LONG_MNEM_SUFFIX
+                          ? No_lSuf
+                          : (i.suffix == QWORD_MNEM_SUFFIX
+                             ? No_qSuf
+                             : (i.suffix == LONG_DOUBLE_MNEM_SUFFIX
+                                ? No_xSuf : 0))))));
+
+  for (t = current_templates->start;
+       t < current_templates->end;
+       t++)
+    {
+      /* Must have right number of operands.  */
+      if (i.operands != t->operands)
+       continue;
+
+      /* Check the suffix, except for some instructions in intel mode.  */
+      if ((t->opcode_modifier & suffix_check)
+         && !(intel_syntax
+              && (t->opcode_modifier & IgnoreSize))
+         && !(intel_syntax
+              && t->base_opcode == 0xd9
+              && (t->extension_opcode == 5          /* 0xd9,5 "fldcw"  */
+                  || t->extension_opcode == 7)))  /* 0xd9,7 "f{n}stcw"  */
+       continue;
+
+      /* Do not verify operands when there are none.  */
+      else if (!t->operands)
+       {
+         if (t->cpu_flags & ~cpu_arch_flags)
+           continue;
+         /* We've found a match; break out of loop.  */
+         break;
+       }
 
-    /* Undo SYSV386_COMPAT brokenness when in Intel mode.  See i386.h  */
-    if (SYSV386_COMPAT
-       && intel_syntax
-       && (i.tm.base_opcode & 0xfffffde0) == 0xdce0)
-      i.tm.base_opcode ^= FloatR;
+      overlap0 = i.types[0] & t->operand_types[0];
+      switch (t->operands)
+       {
+       case 1:
+         if (!MATCH (overlap0, i.types[0], t->operand_types[0]))
+           continue;
+         break;
+       case 2:
+       case 3:
+         overlap1 = i.types[1] & t->operand_types[1];
+         if (!MATCH (overlap0, i.types[0], t->operand_types[0])
+             || !MATCH (overlap1, i.types[1], t->operand_types[1])
+             || !CONSISTENT_REGISTER_MATCH (overlap0, i.types[0],
+                                            t->operand_types[0],
+                                            overlap1, i.types[1],
+                                            t->operand_types[1]))
+           {
+             /* Check if other direction is valid ...  */
+             if ((t->opcode_modifier & (D | FloatD)) == 0)
+               continue;
+
+             /* Try reversing direction of operands.  */
+             overlap0 = i.types[0] & t->operand_types[1];
+             overlap1 = i.types[1] & t->operand_types[0];
+             if (!MATCH (overlap0, i.types[0], t->operand_types[1])
+                 || !MATCH (overlap1, i.types[1], t->operand_types[0])
+                 || !CONSISTENT_REGISTER_MATCH (overlap0, i.types[0],
+                                                t->operand_types[1],
+                                                overlap1, i.types[1],
+                                                t->operand_types[0]))
+               {
+                 /* Does not match either direction.  */
+                 continue;
+               }
+             /* found_reverse_match holds which of D or FloatDR
+                we've found.  */
+             found_reverse_match = t->opcode_modifier & (D | FloatDR);
+           }
+         /* Found a forward 2 operand match here.  */
+         else if (t->operands == 3)
+           {
+             /* Here we make use of the fact that there are no
+                reverse match 3 operand instructions, and all 3
+                operand instructions only need to be checked for
+                register consistency between operands 2 and 3.  */
+             overlap2 = i.types[2] & t->operand_types[2];
+             if (!MATCH (overlap2, i.types[2], t->operand_types[2])
+                 || !CONSISTENT_REGISTER_MATCH (overlap1, i.types[1],
+                                                t->operand_types[1],
+                                                overlap2, i.types[2],
+                                                t->operand_types[2]))
+
+               continue;
+           }
+         /* Found either forward/reverse 2 or 3 operand match here:
+            slip through to break.  */
+       }
+      if (t->cpu_flags & ~cpu_arch_flags)
+       {
+         found_reverse_match = 0;
+         continue;
+       }
+      /* We've found a match; break out of loop.  */
+      break;
+    }
 
-    if (i.tm.opcode_modifier & FWait)
-      if (! add_prefix (FWAIT_OPCODE))
-       return;
+  if (t == current_templates->end)
+    {
+      /* We found no match.  */
+      as_bad (_("suffix or operands invalid for `%s'"),
+             current_templates->start->name);
+      return 0;
+    }
 
-    /* Check string instruction segment overrides.  */
-    if ((i.tm.opcode_modifier & IsString) != 0 && i.mem_operands != 0)
-      {
-       int mem_op = (i.types[0] & AnyMem) ? 0 : 1;
-       if ((i.tm.operand_types[mem_op] & EsSeg) != 0)
-         {
-           if (i.seg[0] != NULL && i.seg[0] != &es)
+  if (!quiet_warnings)
+    {
+      if (!intel_syntax
+         && ((i.types[0] & JumpAbsolute)
+             != (t->operand_types[0] & JumpAbsolute)))
+       {
+         as_warn (_("indirect %s without `*'"), t->name);
+       }
+
+      if ((t->opcode_modifier & (IsPrefix | IgnoreSize))
+         == (IsPrefix | IgnoreSize))
+       {
+         /* Warn them that a data or address size prefix doesn't
+            affect assembly of the next line of code.  */
+         as_warn (_("stand-alone `%s' prefix"), t->name);
+       }
+    }
+
+  /* Copy the template we found.  */
+  i.tm = *t;
+  if (found_reverse_match)
+    {
+      /* If we found a reverse match we must alter the opcode
+        direction bit.  found_reverse_match holds bits to change
+        (different for int & float insns).  */
+
+      i.tm.base_opcode ^= found_reverse_match;
+
+      i.tm.operand_types[0] = t->operand_types[1];
+      i.tm.operand_types[1] = t->operand_types[0];
+    }
+
+  return 1;
+}
+
+static int
+check_string ()
+{
+  int mem_op = (i.types[0] & AnyMem) ? 0 : 1;
+  if ((i.tm.operand_types[mem_op] & EsSeg) != 0)
+    {
+      if (i.seg[0] != NULL && i.seg[0] != &es)
+       {
+         as_bad (_("`%s' operand %d must use `%%es' segment"),
+                 i.tm.name,
+                 mem_op + 1);
+         return 0;
+       }
+      /* There's only ever one segment override allowed per instruction.
+        This instruction possibly has a legal segment override on the
+        second operand, so copy the segment to where non-string
+        instructions store it, allowing common code.  */
+      i.seg[0] = i.seg[1];
+    }
+  else if ((i.tm.operand_types[mem_op + 1] & EsSeg) != 0)
+    {
+      if (i.seg[1] != NULL && i.seg[1] != &es)
+       {
+         as_bad (_("`%s' operand %d must use `%%es' segment"),
+                 i.tm.name,
+                 mem_op + 2);
+         return 0;
+       }
+    }
+  return 1;
+}
+
+static int
+process_suffix ()
+{
+  /* If matched instruction specifies an explicit instruction mnemonic
+     suffix, use it.  */
+  if (i.tm.opcode_modifier & (Size16 | Size32 | Size64))
+    {
+      if (i.tm.opcode_modifier & Size16)
+       i.suffix = WORD_MNEM_SUFFIX;
+      else if (i.tm.opcode_modifier & Size64)
+       i.suffix = QWORD_MNEM_SUFFIX;
+      else
+       i.suffix = LONG_MNEM_SUFFIX;
+    }
+  else if (i.reg_operands)
+    {
+      /* If there's no instruction mnemonic suffix we try to invent one
+        based on register operands.  */
+      if (!i.suffix)
+       {
+         /* We take i.suffix from the last register operand specified,
+            Destination register type is more significant than source
+            register type.  */
+         int op;
+         for (op = i.operands; --op >= 0;)
+           if ((i.types[op] & Reg)
+               && !(i.tm.operand_types[op] & InOutPortReg))
              {
-               as_bad (_("`%s' operand %d must use `%%es' segment"),
-                       i.tm.name,
-                       mem_op + 1);
-               return;
+               i.suffix = ((i.types[op] & Reg8) ? BYTE_MNEM_SUFFIX :
+                           (i.types[op] & Reg16) ? WORD_MNEM_SUFFIX :
+                           (i.types[op] & Reg64) ? QWORD_MNEM_SUFFIX :
+                           LONG_MNEM_SUFFIX);
+               break;
              }
-           /* There's only ever one segment override allowed per instruction.
-              This instruction possibly has a legal segment override on the
-              second operand, so copy the segment to where non-string
-              instructions store it, allowing common code.  */
-           i.seg[0] = i.seg[1];
-         }
-       else if ((i.tm.operand_types[mem_op + 1] & EsSeg) != 0)
+       }
+      else if (i.suffix == BYTE_MNEM_SUFFIX)
+       {
+         if (!check_byte_reg ())
+           return 0;
+       }
+      else if (i.suffix == LONG_MNEM_SUFFIX)
+       {
+         if (!check_long_reg ())
+           return 0;
+       }
+      else if (i.suffix == QWORD_MNEM_SUFFIX)
+       {
+         if (!check_qword_reg ())
+           return 0;
+       }
+      else if (i.suffix == WORD_MNEM_SUFFIX)
+       {
+         if (!check_word_reg ())
+           return 0;
+       }
+      else if (intel_syntax && (i.tm.opcode_modifier & IgnoreSize))
+       /* Do nothing if the instruction is going to ignore the prefix.  */
+       ;
+      else
+       abort ();
+    }
+  else if ((i.tm.opcode_modifier & DefaultSize) && !i.suffix)
+    {
+      i.suffix = stackop_size;
+    }
+
+  /* Change the opcode based on the operand size given by i.suffix;
+     We need not change things for byte insns.  */
+
+  if (!i.suffix && (i.tm.opcode_modifier & W))
+    {
+      as_bad (_("no instruction mnemonic suffix given and no register operands; can't size instruction"));
+      return 0;
+    }
+
+  if (i.suffix && i.suffix != BYTE_MNEM_SUFFIX)
+    {
+      /* It's not a byte, select word/dword operation.  */
+      if (i.tm.opcode_modifier & W)
+       {
+         if (i.tm.opcode_modifier & ShortForm)
+           i.tm.base_opcode |= 8;
+         else
+           i.tm.base_opcode |= 1;
+       }
+
+      /* Now select between word & dword operations via the operand
+        size prefix, except for instructions that will ignore this
+        prefix anyway.  */
+      if (i.suffix != QWORD_MNEM_SUFFIX
+         && !(i.tm.opcode_modifier & IgnoreSize)
+         && ((i.suffix == LONG_MNEM_SUFFIX) == (flag_code == CODE_16BIT)
+             || (flag_code == CODE_64BIT
+                 && (i.tm.opcode_modifier & JumpByte))))
+       {
+         unsigned int prefix = DATA_PREFIX_OPCODE;
+         if (i.tm.opcode_modifier & JumpByte) /* jcxz, loop */
+           prefix = ADDR_PREFIX_OPCODE;
+
+         if (!add_prefix (prefix))
+           return 0;
+       }
+
+      /* Set mode64 for an operand.  */
+      if (i.suffix == QWORD_MNEM_SUFFIX
+         && flag_code == CODE_64BIT
+         && (i.tm.opcode_modifier & NoRex64) == 0)
+       i.rex |= REX_MODE64;
+
+      /* Size floating point instruction.  */
+      if (i.suffix == LONG_MNEM_SUFFIX)
+       {
+         if (i.tm.opcode_modifier & FloatMF)
+           i.tm.base_opcode ^= 4;
+       }
+    }
+
+  return 1;
+}
+
+static int
+check_byte_reg ()
+{
+  int op;
+  for (op = i.operands; --op >= 0;)
+    {
+      /* If this is an eight bit register, it's OK.  If it's the 16 or
+        32 bit version of an eight bit register, we will just use the
+        low portion, and that's OK too.  */
+      if (i.types[op] & Reg8)
+       continue;
+
+      /* movzx and movsx should not generate this warning.  */
+      if (intel_syntax
+         && (i.tm.base_opcode == 0xfb7
+             || i.tm.base_opcode == 0xfb6
+             || i.tm.base_opcode == 0x63
+             || i.tm.base_opcode == 0xfbe
+             || i.tm.base_opcode == 0xfbf))
+       continue;
+
+      if ((i.types[op] & WordReg) && i.op[op].regs->reg_num < 4
+#if 0
+         /* Check that the template allows eight bit regs.  This
+            kills insns such as `orb $1,%edx', which maybe should be
+            allowed.  */
+         && (i.tm.operand_types[op] & (Reg8 | InOutPortReg))
+#endif
+         )
+       {
+         /* Prohibit these changes in the 64bit mode, since the
+            lowering is more complicated.  */
+         if (flag_code == CODE_64BIT
+             && (i.tm.operand_types[op] & InOutPortReg) == 0)
+           {
+             as_bad (_("Incorrect register `%%%s' used with `%c' suffix"),
+                     i.op[op].regs->reg_name,
+                     i.suffix);
+             return 0;
+           }
+#if REGISTER_WARNINGS
+         if (!quiet_warnings
+             && (i.tm.operand_types[op] & InOutPortReg) == 0)
+           as_warn (_("using `%%%s' instead of `%%%s' due to `%c' suffix"),
+                    (i.op[op].regs + (i.types[op] & Reg16
+                                      ? REGNAM_AL - REGNAM_AX
+                                      : REGNAM_AL - REGNAM_EAX))->reg_name,
+                    i.op[op].regs->reg_name,
+                    i.suffix);
+#endif
+         continue;
+       }
+      /* Any other register is bad.  */
+      if (i.types[op] & (Reg | RegMMX | RegXMM
+                        | SReg2 | SReg3
+                        | Control | Debug | Test
+                        | FloatReg | FloatAcc))
+       {
+         as_bad (_("`%%%s' not allowed with `%s%c'"),
+                 i.op[op].regs->reg_name,
+                 i.tm.name,
+                 i.suffix);
+         return 0;
+       }
+    }
+  return 1;
+}
+
+static int
+check_long_reg ()
+{
+  int op;
+
+  for (op = i.operands; --op >= 0;)
+    /* Reject eight bit registers, except where the template requires
+       them. (eg. movzb)  */
+    if ((i.types[op] & Reg8) != 0
+       && (i.tm.operand_types[op] & (Reg16 | Reg32 | Acc)) != 0)
+      {
+       as_bad (_("`%%%s' not allowed with `%s%c'"),
+               i.op[op].regs->reg_name,
+               i.tm.name,
+               i.suffix);
+       return 0;
+      }
+  /* Warn if the e prefix on a general reg is missing.  */
+    else if ((!quiet_warnings || flag_code == CODE_64BIT)
+            && (i.types[op] & Reg16) != 0
+            && (i.tm.operand_types[op] & (Reg32 | Acc)) != 0)
+      {
+       /* Prohibit these changes in the 64bit mode, since the
+          lowering is more complicated.  */
+       if (flag_code == CODE_64BIT)
          {
-           if (i.seg[1] != NULL && i.seg[1] != &es)
-             {
-               as_bad (_("`%s' operand %d must use `%%es' segment"),
-                       i.tm.name,
-                       mem_op + 2);
-               return;
-             }
+           as_bad (_("Incorrect register `%%%s' used with `%c' suffix"),
+                   i.op[op].regs->reg_name,
+                   i.suffix);
+           return 0;
          }
+#if REGISTER_WARNINGS
+       else
+         as_warn (_("using `%%%s' instead of `%%%s' due to `%c' suffix"),
+                  (i.op[op].regs + REGNAM_EAX - REGNAM_AX)->reg_name,
+                  i.op[op].regs->reg_name,
+                  i.suffix);
+#endif
+      }
+  /* Warn if the r prefix on a general reg is missing.  */
+    else if ((i.types[op] & Reg64) != 0
+            && (i.tm.operand_types[op] & (Reg32 | Acc)) != 0)
+      {
+       as_bad (_("Incorrect register `%%%s' used with `%c' suffix"),
+               i.op[op].regs->reg_name,
+               i.suffix);
+       return 0;
+      }
+  return 1;
+}
+
+static int
+check_qword_reg ()
+{
+  int op;
+
+  for (op = i.operands; --op >= 0; )
+    /* Reject eight bit registers, except where the template requires
+       them. (eg. movzb)  */
+    if ((i.types[op] & Reg8) != 0
+       && (i.tm.operand_types[op] & (Reg16 | Reg32 | Acc)) != 0)
+      {
+       as_bad (_("`%%%s' not allowed with `%s%c'"),
+               i.op[op].regs->reg_name,
+               i.tm.name,
+               i.suffix);
+       return 0;
+      }
+  /* Warn if the e prefix on a general reg is missing.  */
+    else if (((i.types[op] & Reg16) != 0
+             || (i.types[op] & Reg32) != 0)
+            && (i.tm.operand_types[op] & (Reg32 | Acc)) != 0)
+      {
+       /* Prohibit these changes in the 64bit mode, since the
+          lowering is more complicated.  */
+       as_bad (_("Incorrect register `%%%s' used with `%c' suffix"),
+               i.op[op].regs->reg_name,
+               i.suffix);
+       return 0;
       }
+  return 1;
+}
 
-    if (i.reg_operands && flag_code < CODE_64BIT)
+static int
+check_word_reg ()
+{
+  int op;
+  for (op = i.operands; --op >= 0;)
+    /* Reject eight bit registers, except where the template requires
+       them. (eg. movzb)  */
+    if ((i.types[op] & Reg8) != 0
+       && (i.tm.operand_types[op] & (Reg16 | Reg32 | Acc)) != 0)
       {
-       int op;
-       for (op = i.operands; --op >= 0;)
-         if ((i.types[op] & Reg)
-             && (i.op[op].regs->reg_flags & (RegRex64|RegRex)))
+       as_bad (_("`%%%s' not allowed with `%s%c'"),
+               i.op[op].regs->reg_name,
+               i.tm.name,
+               i.suffix);
+       return 0;
+      }
+  /* Warn if the e prefix on a general reg is present.  */
+    else if ((!quiet_warnings || flag_code == CODE_64BIT)
+            && (i.types[op] & Reg32) != 0
+            && (i.tm.operand_types[op] & (Reg16 | Acc)) != 0)
+      {
+       /* Prohibit these changes in the 64bit mode, since the
+          lowering is more complicated.  */
+       if (flag_code == CODE_64BIT)
+         {
+           as_bad (_("Incorrect register `%%%s' used with `%c' suffix"),
+                   i.op[op].regs->reg_name,
+                   i.suffix);
+           return 0;
+         }
+       else
+#if REGISTER_WARNINGS
+         as_warn (_("using `%%%s' instead of `%%%s' due to `%c' suffix"),
+                  (i.op[op].regs + REGNAM_AX - REGNAM_EAX)->reg_name,
+                  i.op[op].regs->reg_name,
+                  i.suffix);
+#endif
+      }
+  return 1;
+}
+
+static int
+finalize_imm ()
+{
+  unsigned int overlap0, overlap1, overlap2;
+
+  overlap0 = i.types[0] & i.tm.operand_types[0];
+  if ((overlap0 & (Imm8 | Imm8S | Imm16 | Imm32 | Imm32S))
+      && overlap0 != Imm8 && overlap0 != Imm8S
+      && overlap0 != Imm16 && overlap0 != Imm32S
+      && overlap0 != Imm32 && overlap0 != Imm64)
+    {
+      if (i.suffix)
+       {
+         overlap0 &= (i.suffix == BYTE_MNEM_SUFFIX
+                      ? Imm8 | Imm8S
+                      : (i.suffix == WORD_MNEM_SUFFIX
+                         ? Imm16
+                         : (i.suffix == QWORD_MNEM_SUFFIX
+                            ? Imm64 | Imm32S
+                            : Imm32)));
+       }
+      else if (overlap0 == (Imm16 | Imm32S | Imm32)
+              || overlap0 == (Imm16 | Imm32)
+              || overlap0 == (Imm16 | Imm32S))
+       {
+         overlap0 = ((flag_code == CODE_16BIT) ^ (i.prefix[DATA_PREFIX] != 0)
+                     ? Imm16 : Imm32S);
+       }
+      if (overlap0 != Imm8 && overlap0 != Imm8S
+         && overlap0 != Imm16 && overlap0 != Imm32S
+         && overlap0 != Imm32 && overlap0 != Imm64)
+       {
+         as_bad (_("no instruction mnemonic suffix given; can't determine immediate size"));
+         return 0;
+       }
+    }
+  i.types[0] = overlap0;
+
+  overlap1 = i.types[1] & i.tm.operand_types[1];
+  if ((overlap1 & (Imm8 | Imm8S | Imm16 | Imm32S | Imm32))
+      && overlap1 != Imm8 && overlap1 != Imm8S
+      && overlap1 != Imm16 && overlap1 != Imm32S
+      && overlap1 != Imm32 && overlap1 != Imm64)
+    {
+      if (i.suffix)
+       {
+         overlap1 &= (i.suffix == BYTE_MNEM_SUFFIX
+                      ? Imm8 | Imm8S
+                      : (i.suffix == WORD_MNEM_SUFFIX
+                         ? Imm16
+                         : (i.suffix == QWORD_MNEM_SUFFIX
+                            ? Imm64 | Imm32S
+                            : Imm32)));
+       }
+      else if (overlap1 == (Imm16 | Imm32 | Imm32S)
+              || overlap1 == (Imm16 | Imm32)
+              || overlap1 == (Imm16 | Imm32S))
+       {
+         overlap1 = ((flag_code == CODE_16BIT) ^ (i.prefix[DATA_PREFIX] != 0)
+                     ? Imm16 : Imm32S);
+       }
+      if (overlap1 != Imm8 && overlap1 != Imm8S
+         && overlap1 != Imm16 && overlap1 != Imm32S
+         && overlap1 != Imm32 && overlap1 != Imm64)
+       {
+         as_bad (_("no instruction mnemonic suffix given; can't determine immediate size %x %c"),overlap1, i.suffix);
+         return 0;
+       }
+    }
+  i.types[1] = overlap1;
+
+  overlap2 = i.types[2] & i.tm.operand_types[2];
+  assert ((overlap2 & Imm) == 0);
+  i.types[2] = overlap2;
+
+  return 1;
+}
+
+static int
+process_operands ()
+{
+  /* Default segment register this instruction will use for memory
+     accesses.  0 means unknown.  This is only for optimizing out
+     unnecessary segment overrides.  */
+  const seg_entry *default_seg = 0;
+
+  /* The imul $imm, %reg instruction is converted into
+     imul $imm, %reg, %reg, and the clr %reg instruction
+     is converted into xor %reg, %reg.  */
+  if (i.tm.opcode_modifier & regKludge)
+    {
+      unsigned int first_reg_op = (i.types[0] & Reg) ? 0 : 1;
+      /* Pretend we saw the extra register operand.  */
+      assert (i.op[first_reg_op + 1].regs == 0);
+      i.op[first_reg_op + 1].regs = i.op[first_reg_op].regs;
+      i.types[first_reg_op + 1] = i.types[first_reg_op];
+      i.reg_operands = 2;
+    }
+
+  if (i.tm.opcode_modifier & ShortForm)
+    {
+      /* The register or float register operand is in operand 0 or 1.  */
+      unsigned int op = (i.types[0] & (Reg | FloatReg)) ? 0 : 1;
+      /* Register goes in low 3 bits of opcode.  */
+      i.tm.base_opcode |= i.op[op].regs->reg_num;
+      if ((i.op[op].regs->reg_flags & RegRex) != 0)
+       i.rex |= REX_EXTZ;
+      if (!quiet_warnings && (i.tm.opcode_modifier & Ugh) != 0)
+       {
+         /* Warn about some common errors, but press on regardless.
+            The first case can be generated by gcc (<= 2.8.1).  */
+         if (i.operands == 2)
            {
-             as_bad (_("Extended register `%%%s' available only in 64bit mode."),
-                     i.op[op].regs->reg_name);
-             return;
+             /* Reversed arguments on faddp, fsubp, etc.  */
+             as_warn (_("translating to `%s %%%s,%%%s'"), i.tm.name,
+                      i.op[1].regs->reg_name,
+                      i.op[0].regs->reg_name);
            }
-      }
+         else
+           {
+             /* Extraneous `l' suffix on fp insn.  */
+             as_warn (_("translating to `%s %%%s'"), i.tm.name,
+                      i.op[0].regs->reg_name);
+           }
+       }
+    }
+  else if (i.tm.opcode_modifier & Modrm)
+    {
+      /* The opcode is completed (modulo i.tm.extension_opcode which
+        must be put into the modrm byte).  Now, we make the modrm and
+        index base bytes based on all the info we've collected.  */
+
+      default_seg = build_modrm_byte ();
+    }
+  else if (i.tm.opcode_modifier & (Seg2ShortForm | Seg3ShortForm))
+    {
+      if (i.tm.base_opcode == POP_SEG_SHORT
+         && i.op[0].regs->reg_num == 1)
+       {
+         as_bad (_("you can't `pop %%cs'"));
+         return 0;
+       }
+      i.tm.base_opcode |= (i.op[0].regs->reg_num << 3);
+      if ((i.op[0].regs->reg_flags & RegRex) != 0)
+       i.rex |= REX_EXTZ;
+    }
+  else if ((i.tm.base_opcode & ~(D | W)) == MOV_AX_DISP32)
+    {
+      default_seg = &ds;
+    }
+  else if ((i.tm.opcode_modifier & IsString) != 0)
+    {
+      /* For the string instructions that allow a segment override
+        on one of their operands, the default segment is ds.  */
+      default_seg = &ds;
+    }
 
-    /* If matched instruction specifies an explicit instruction mnemonic
-       suffix, use it.  */
-    if (i.tm.opcode_modifier & (Size16 | Size32 | Size64))
-      {
-       if (i.tm.opcode_modifier & Size16)
-         i.suffix = WORD_MNEM_SUFFIX;
-       else if (i.tm.opcode_modifier & Size64)
-         i.suffix = QWORD_MNEM_SUFFIX;
-       else
-         i.suffix = LONG_MNEM_SUFFIX;
-      }
-    else if (i.reg_operands)
-      {
-       /* If there's no instruction mnemonic suffix we try to invent one
-          based on register operands.  */
-       if (!i.suffix)
-         {
-           /* We take i.suffix from the last register operand specified,
-              Destination register type is more significant than source
-              register type.  */
-           int op;
-           for (op = i.operands; --op >= 0;)
-             if ((i.types[op] & Reg)
-                 && !(i.tm.operand_types[op] & InOutPortReg))
-               {
-                 i.suffix = ((i.types[op] & Reg8) ? BYTE_MNEM_SUFFIX :
-                             (i.types[op] & Reg16) ? WORD_MNEM_SUFFIX :
-                             (i.types[op] & Reg64) ? QWORD_MNEM_SUFFIX :
-                             LONG_MNEM_SUFFIX);
-                 break;
-               }
-         }
-       else if (i.suffix == BYTE_MNEM_SUFFIX)
-         {
-           int op;
-           for (op = i.operands; --op >= 0;)
-             {
-               /* If this is an eight bit register, it's OK.  If it's
-                  the 16 or 32 bit version of an eight bit register,
-                  we will just use the low portion, and that's OK too.  */
-               if (i.types[op] & Reg8)
-                 continue;
+  if (i.tm.base_opcode == 0x8d /* lea */ && i.seg[0] && !quiet_warnings)
+    as_warn (_("segment override on `lea' is ineffectual"));
 
-               /* movzx and movsx should not generate this warning.  */
-               if (intel_syntax
-                   && (i.tm.base_opcode == 0xfb7
-                       || i.tm.base_opcode == 0xfb6
-                       || i.tm.base_opcode == 0x63
-                       || i.tm.base_opcode == 0xfbe
-                       || i.tm.base_opcode == 0xfbf))
-                 continue;
+  /* If a segment was explicitly specified, and the specified segment
+     is not the default, use an opcode prefix to select it.  If we
+     never figured out what the default segment is, then default_seg
+     will be zero at this point, and the specified segment prefix will
+     always be used.  */
+  if ((i.seg[0]) && (i.seg[0] != default_seg))
+    {
+      if (!add_prefix (i.seg[0]->seg_prefix))
+       return 0;
+    }
+  return 1;
+}
 
-               if ((i.types[op] & WordReg) && i.op[op].regs->reg_num < 4
-#if 0
-                   /* Check that the template allows eight bit regs
-                      This kills insns such as `orb $1,%edx', which
-                      maybe should be allowed.  */
-                   && (i.tm.operand_types[op] & (Reg8|InOutPortReg))
-#endif
-                   )
-                 {
-                   /* Prohibit these changes in the 64bit mode, since
-                      the lowering is more complicated.  */
-                   if (flag_code == CODE_64BIT
-                       && (i.tm.operand_types[op] & InOutPortReg) == 0)
-                     as_bad (_("Incorrect register `%%%s' used with`%c' suffix"),
-                             i.op[op].regs->reg_name,
-                             i.suffix);
-#if REGISTER_WARNINGS
-                   if (!quiet_warnings
-                       && (i.tm.operand_types[op] & InOutPortReg) == 0)
-                     as_warn (_("using `%%%s' instead of `%%%s' due to `%c' suffix"),
-                              (i.op[op].regs
-                               + (i.types[op] & Reg16
-                                  ? REGNAM_AL - REGNAM_AX
-                                  : REGNAM_AL - REGNAM_EAX))->reg_name,
-                              i.op[op].regs->reg_name,
-                              i.suffix);
-#endif
-                   continue;
-                 }
-               /* Any other register is bad.  */
-               if (i.types[op] & (Reg | RegMMX | RegXMM
-                                  | SReg2 | SReg3
-                                  | Control | Debug | Test
-                                  | FloatReg | FloatAcc))
-                 {
-                   as_bad (_("`%%%s' not allowed with `%s%c'"),
-                           i.op[op].regs->reg_name,
-                           i.tm.name,
-                           i.suffix);
-                   return;
-                 }
-             }
-         }
-       else if (i.suffix == LONG_MNEM_SUFFIX)
-         {
-           int op;
+static const seg_entry *
+build_modrm_byte ()
+{
+  const seg_entry *default_seg = 0;
+
+  /* i.reg_operands MUST be the number of real register operands;
+     implicit registers do not count.  */
+  if (i.reg_operands == 2)
+    {
+      unsigned int source, dest;
+      source = ((i.types[0]
+                & (Reg | RegMMX | RegXMM
+                   | SReg2 | SReg3
+                   | Control | Debug | Test))
+               ? 0 : 1);
+      dest = source + 1;
+
+      i.rm.mode = 3;
+      /* One of the register operands will be encoded in the i.tm.reg
+        field, the other in the combined i.tm.mode and i.tm.regmem
+        fields.  If no form of this instruction supports a memory
+        destination operand, then we assume the source operand may
+        sometimes be a memory operand and so we need to store the
+        destination in the i.rm.reg field.  */
+      if ((i.tm.operand_types[dest] & AnyMem) == 0)
+       {
+         i.rm.reg = i.op[dest].regs->reg_num;
+         i.rm.regmem = i.op[source].regs->reg_num;
+         if ((i.op[dest].regs->reg_flags & RegRex) != 0)
+           i.rex |= REX_EXTX;
+         if ((i.op[source].regs->reg_flags & RegRex) != 0)
+           i.rex |= REX_EXTZ;
+       }
+      else
+       {
+         i.rm.reg = i.op[source].regs->reg_num;
+         i.rm.regmem = i.op[dest].regs->reg_num;
+         if ((i.op[dest].regs->reg_flags & RegRex) != 0)
+           i.rex |= REX_EXTZ;
+         if ((i.op[source].regs->reg_flags & RegRex) != 0)
+           i.rex |= REX_EXTX;
+       }
+    }
+  else
+    {                  /* If it's not 2 reg operands...  */
+      if (i.mem_operands)
+       {
+         unsigned int fake_zero_displacement = 0;
+         unsigned int op = ((i.types[0] & AnyMem)
+                            ? 0
+                            : (i.types[1] & AnyMem) ? 1 : 2);
 
-           for (op = i.operands; --op >= 0;)
-             /* Reject eight bit registers, except where the template
-                requires them. (eg. movzb)  */
-             if ((i.types[op] & Reg8) != 0
-                 && (i.tm.operand_types[op] & (Reg16 | Reg32 | Acc)) != 0)
+         default_seg = &ds;
+
+         if (i.base_reg == 0)
+           {
+             i.rm.mode = 0;
+             if (!i.disp_operands)
+               fake_zero_displacement = 1;
+             if (i.index_reg == 0)
                {
-                 as_bad (_("`%%%s' not allowed with `%s%c'"),
-                         i.op[op].regs->reg_name,
-                         i.tm.name,
-                         i.suffix);
-                 return;
+                 /* Operand is just <disp>  */
+                 if ((flag_code == CODE_16BIT) ^ (i.prefix[ADDR_PREFIX] != 0)
+                     && (flag_code != CODE_64BIT))
+                   {
+                     i.rm.regmem = NO_BASE_REGISTER_16;
+                     i.types[op] &= ~Disp;
+                     i.types[op] |= Disp16;
+                   }
+                 else if (flag_code != CODE_64BIT
+                          || (i.prefix[ADDR_PREFIX] != 0))
+                   {
+                     i.rm.regmem = NO_BASE_REGISTER;
+                     i.types[op] &= ~Disp;
+                     i.types[op] |= Disp32;
+                   }
+                 else
+                   {
+                     /* 64bit mode overwrites the 32bit absolute
+                        addressing by RIP relative addressing and
+                        absolute addressing is encoded by one of the
+                        redundant SIB forms.  */
+                     i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
+                     i.sib.base = NO_BASE_REGISTER;
+                     i.sib.index = NO_INDEX_REGISTER;
+                     i.types[op] &= ~Disp;
+                     i.types[op] |= Disp32S;
+                   }
                }
-             /* Warn if the e prefix on a general reg is missing.  */
-             else if ((!quiet_warnings || flag_code == CODE_64BIT)
-                      && (i.types[op] & Reg16) != 0
-                      && (i.tm.operand_types[op] & (Reg32|Acc)) != 0)
+             else /* !i.base_reg && i.index_reg  */
                {
-                 /* Prohibit these changes in the 64bit mode, since
-                    the lowering is more complicated.  */
-                 if (flag_code == CODE_64BIT)
-                   as_bad (_("Incorrect register `%%%s' used with`%c' suffix"),
-                           i.op[op].regs->reg_name,
-                           i.suffix);
-#if REGISTER_WARNINGS
+                 i.sib.index = i.index_reg->reg_num;
+                 i.sib.base = NO_BASE_REGISTER;
+                 i.sib.scale = i.log2_scale_factor;
+                 i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
+                 i.types[op] &= ~Disp;
+                 if (flag_code != CODE_64BIT)
+                   i.types[op] |= Disp32;      /* Must be 32 bit */
                  else
-                   as_warn (_("using `%%%s' instead of `%%%s' due to `%c' suffix"),
-                            (i.op[op].regs + REGNAM_EAX - REGNAM_AX)->reg_name,
-                            i.op[op].regs->reg_name,
-                            i.suffix);
-#endif
+                   i.types[op] |= Disp32S;
+                 if ((i.index_reg->reg_flags & RegRex) != 0)
+                   i.rex |= REX_EXTY;
                }
-             /* Warn if the r prefix on a general reg is missing.  */
-             else if ((i.types[op] & Reg64) != 0
-                      && (i.tm.operand_types[op] & (Reg32|Acc)) != 0)
+           }
+         /* RIP addressing for 64bit mode.  */
+         else if (i.base_reg->reg_type == BaseIndex)
+           {
+             i.rm.regmem = NO_BASE_REGISTER;
+             i.types[op] &= ~Disp;
+             i.types[op] |= Disp32S;
+             i.flags[op] = Operand_PCrel;
+           }
+         else if (i.base_reg->reg_type & Reg16)
+           {
+             switch (i.base_reg->reg_num)
                {
-                 as_bad (_("Incorrect register `%%%s' used with`%c' suffix"),
-                         i.op[op].regs->reg_name,
-                         i.suffix);
+               case 3: /* (%bx)  */
+                 if (i.index_reg == 0)
+                   i.rm.regmem = 7;
+                 else /* (%bx,%si) -> 0, or (%bx,%di) -> 1  */
+                   i.rm.regmem = i.index_reg->reg_num - 6;
+                 break;
+               case 5: /* (%bp)  */
+                 default_seg = &ss;
+                 if (i.index_reg == 0)
+                   {
+                     i.rm.regmem = 6;
+                     if ((i.types[op] & Disp) == 0)
+                       {
+                         /* fake (%bp) into 0(%bp)  */
+                         i.types[op] |= Disp8;
+                         fake_zero_displacement = 1;
+                       }
+                   }
+                 else /* (%bp,%si) -> 2, or (%bp,%di) -> 3  */
+                   i.rm.regmem = i.index_reg->reg_num - 6 + 2;
+                 break;
+               default: /* (%si) -> 4 or (%di) -> 5  */
+                 i.rm.regmem = i.base_reg->reg_num - 6 + 4;
                }
-         }
-       else if (i.suffix == QWORD_MNEM_SUFFIX)
-         {
-           int op;
-
-           for (op = i.operands; --op >= 0; )
-             /* Reject eight bit registers, except where the template
-                requires them. (eg. movzb)  */
-             if ((i.types[op] & Reg8) != 0
-                 && (i.tm.operand_types[op] & (Reg16|Reg32|Acc)) != 0)
+             i.rm.mode = mode_from_disp_size (i.types[op]);
+           }
+         else /* i.base_reg and 32/64 bit mode  */
+           {
+             if (flag_code == CODE_64BIT
+                 && (i.types[op] & Disp))
                {
-                 as_bad (_("`%%%s' not allowed with `%s%c'"),
-                         i.op[op].regs->reg_name,
-                         i.tm.name,
-                         i.suffix);
-                 return;
+                 if (i.types[op] & Disp8)
+                   i.types[op] = Disp8 | Disp32S;
+                 else
+                   i.types[op] = Disp32S;
                }
-             /* Warn if the e prefix on a general reg is missing.  */
-             else if (((i.types[op] & Reg16) != 0
-                       || (i.types[op] & Reg32) != 0)
-                      && (i.tm.operand_types[op] & (Reg32|Acc)) != 0)
+             i.rm.regmem = i.base_reg->reg_num;
+             if ((i.base_reg->reg_flags & RegRex) != 0)
+               i.rex |= REX_EXTZ;
+             i.sib.base = i.base_reg->reg_num;
+             /* x86-64 ignores REX prefix bit here to avoid decoder
+                complications.  */
+             if ((i.base_reg->reg_num & 7) == EBP_REG_NUM)
                {
-                 /* Prohibit these changes in the 64bit mode, since
-                    the lowering is more complicated.  */
-                 as_bad (_("Incorrect register `%%%s' used with`%c' suffix"),
-                         i.op[op].regs->reg_name,
-                         i.suffix);
+                 default_seg = &ss;
+                 if (i.disp_operands == 0)
+                   {
+                     fake_zero_displacement = 1;
+                     i.types[op] |= Disp8;
+                   }
                }
-         }
-       else if (i.suffix == WORD_MNEM_SUFFIX)
-         {
-           int op;
-           for (op = i.operands; --op >= 0;)
-             /* Reject eight bit registers, except where the template
-                requires them. (eg. movzb)  */
-             if ((i.types[op] & Reg8) != 0
-                 && (i.tm.operand_types[op] & (Reg16|Reg32|Acc)) != 0)
+             else if (i.base_reg->reg_num == ESP_REG_NUM)
                {
-                 as_bad (_("`%%%s' not allowed with `%s%c'"),
-                         i.op[op].regs->reg_name,
-                         i.tm.name,
-                         i.suffix);
-                 return;
+                 default_seg = &ss;
                }
-             /* Warn if the e prefix on a general reg is present.  */
-             else if ((!quiet_warnings || flag_code == CODE_64BIT)
-                      && (i.types[op] & Reg32) != 0
-                      && (i.tm.operand_types[op] & (Reg16|Acc)) != 0)
+             i.sib.scale = i.log2_scale_factor;
+             if (i.index_reg == 0)
                {
-                 /* Prohibit these changes in the 64bit mode, since
-                    the lowering is more complicated.  */
-                 if (flag_code == CODE_64BIT)
-                   as_bad (_("Incorrect register `%%%s' used with`%c' suffix"),
-                           i.op[op].regs->reg_name,
-                           i.suffix);
-                 else
-#if REGISTER_WARNINGS
-                   as_warn (_("using `%%%s' instead of `%%%s' due to `%c' suffix"),
-                            (i.op[op].regs + REGNAM_AX - REGNAM_EAX)->reg_name,
-                            i.op[op].regs->reg_name,
-                            i.suffix);
+                 /* <disp>(%esp) becomes two byte modrm with no index
+                    register.  We've already stored the code for esp
+                    in i.rm.regmem ie. ESCAPE_TO_TWO_BYTE_ADDRESSING.
+                    Any base register besides %esp will not use the
+                    extra modrm byte.  */
+                 i.sib.index = NO_INDEX_REGISTER;
+#if !SCALE1_WHEN_NO_INDEX
+                 /* Another case where we force the second modrm byte.  */
+                 if (i.log2_scale_factor)
+                   i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
 #endif
                }
-         }
-       else if (intel_syntax && (i.tm.opcode_modifier & IgnoreSize))
-         /* Do nothing if the instruction is going to ignore the prefix.  */
-         ;
-       else
-         abort ();
-      }
-    else if ((i.tm.opcode_modifier & DefaultSize) && !i.suffix)
-      {
-       i.suffix = stackop_size;
-      }
-    /* Make still unresolved immediate matches conform to size of immediate
-       given in i.suffix.  Note: overlap2 cannot be an immediate!  */
-    if ((overlap0 & (Imm8 | Imm8S | Imm16 | Imm32 | Imm32S))
-       && overlap0 != Imm8 && overlap0 != Imm8S
-       && overlap0 != Imm16 && overlap0 != Imm32S
-       && overlap0 != Imm32 && overlap0 != Imm64)
-      {
-       if (i.suffix)
-         {
-           overlap0 &= (i.suffix == BYTE_MNEM_SUFFIX ? (Imm8 | Imm8S) :
-                       (i.suffix == WORD_MNEM_SUFFIX ? Imm16 :
-                       (i.suffix == QWORD_MNEM_SUFFIX ? Imm64 | Imm32S : Imm32)));
-         }
-       else if (overlap0 == (Imm16 | Imm32S | Imm32)
-                || overlap0 == (Imm16 | Imm32)
-                || overlap0 == (Imm16 | Imm32S))
-         {
-           overlap0 =
-             ((flag_code == CODE_16BIT) ^ (i.prefix[DATA_PREFIX] != 0)) ? Imm16 : Imm32S;
-         }
-       if (overlap0 != Imm8 && overlap0 != Imm8S
-           && overlap0 != Imm16 && overlap0 != Imm32S
-           && overlap0 != Imm32 && overlap0 != Imm64)
-         {
-           as_bad (_("no instruction mnemonic suffix given; can't determine immediate size"));
-           return;
-         }
-      }
-    if ((overlap1 & (Imm8 | Imm8S | Imm16 | Imm32S | Imm32))
-       && overlap1 != Imm8 && overlap1 != Imm8S
-       && overlap1 != Imm16 && overlap1 != Imm32S
-       && overlap1 != Imm32 && overlap1 != Imm64)
-      {
-       if (i.suffix)
-         {
-           overlap1 &= (i.suffix == BYTE_MNEM_SUFFIX ? (Imm8 | Imm8S) :
-                       (i.suffix == WORD_MNEM_SUFFIX ? Imm16 :
-                       (i.suffix == QWORD_MNEM_SUFFIX ? Imm64 | Imm32S : Imm32)));
-         }
-       else if (overlap1 == (Imm16 | Imm32 | Imm32S)
-                || overlap1 == (Imm16 | Imm32)
-                || overlap1 == (Imm16 | Imm32S))
-         {
-           overlap1 =
-             ((flag_code == CODE_16BIT) ^ (i.prefix[DATA_PREFIX] != 0)) ? Imm16 : Imm32S;
-         }
-       if (overlap1 != Imm8 && overlap1 != Imm8S
-           && overlap1 != Imm16 && overlap1 != Imm32S
-           && overlap1 != Imm32 && overlap1 != Imm64)
-         {
-           as_bad (_("no instruction mnemonic suffix given; can't determine immediate size %x %c"),overlap1, i.suffix);
-           return;
-         }
-      }
-    assert ((overlap2 & Imm) == 0);
-
-    i.types[0] = overlap0;
-    if (overlap0 & ImplicitRegister)
-      i.reg_operands--;
-    if (overlap0 & Imm1)
-      i.imm_operands = 0;      /* kludge for shift insns.  */
-
-    i.types[1] = overlap1;
-    if (overlap1 & ImplicitRegister)
-      i.reg_operands--;
-
-    i.types[2] = overlap2;
-    if (overlap2 & ImplicitRegister)
-      i.reg_operands--;
-
-    /* Finalize opcode.  First, we change the opcode based on the operand
-       size given by i.suffix:  We need not change things for byte insns.  */
+             else
+               {
+                 i.sib.index = i.index_reg->reg_num;
+                 i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
+                 if ((i.index_reg->reg_flags & RegRex) != 0)
+                   i.rex |= REX_EXTY;
+               }
+             i.rm.mode = mode_from_disp_size (i.types[op]);
+           }
 
-    if (!i.suffix && (i.tm.opcode_modifier & W))
-      {
-       as_bad (_("no instruction mnemonic suffix given and no register operands; can't size instruction"));
-       return;
-      }
+         if (fake_zero_displacement)
+           {
+             /* Fakes a zero displacement assuming that i.types[op]
+                holds the correct displacement size.  */
+             expressionS *exp;
+
+             assert (i.op[op].disps == 0);
+             exp = &disp_expressions[i.disp_operands++];
+             i.op[op].disps = exp;
+             exp->X_op = O_constant;
+             exp->X_add_number = 0;
+             exp->X_add_symbol = (symbolS *) 0;
+             exp->X_op_symbol = (symbolS *) 0;
+           }
+       }
 
-    /* For movzx and movsx, need to check the register type.  */
-    if (intel_syntax
-       && (i.tm.base_opcode == 0xfb6 || i.tm.base_opcode == 0xfbe))
-      if (i.suffix && i.suffix == BYTE_MNEM_SUFFIX)
+      /* Fill in i.rm.reg or i.rm.regmem field with register operand
+        (if any) based on i.tm.extension_opcode.  Again, we must be
+        careful to make sure that segment/control/debug/test/MMX
+        registers are coded into the i.rm.reg field.  */
+      if (i.reg_operands)
        {
-         unsigned int prefix = DATA_PREFIX_OPCODE;
+         unsigned int op =
+           ((i.types[0]
+             & (Reg | RegMMX | RegXMM
+                | SReg2 | SReg3
+                | Control | Debug | Test))
+            ? 0
+            : ((i.types[1]
+                & (Reg | RegMMX | RegXMM
+                   | SReg2 | SReg3
+                   | Control | Debug | Test))
+               ? 1
+               : 2));
+         /* If there is an extension opcode to put here, the register
+            number must be put into the regmem field.  */
+         if (i.tm.extension_opcode != None)
+           {
+             i.rm.regmem = i.op[op].regs->reg_num;
+             if ((i.op[op].regs->reg_flags & RegRex) != 0)
+               i.rex |= REX_EXTZ;
+           }
+         else
+           {
+             i.rm.reg = i.op[op].regs->reg_num;
+             if ((i.op[op].regs->reg_flags & RegRex) != 0)
+               i.rex |= REX_EXTX;
+           }
 
-         if ((i.op[1].regs->reg_type & Reg16) != 0)
-           if (!add_prefix (prefix))
-             return;
+         /* Now, if no memory operand has set i.rm.mode = 0, 1, 2 we
+            must set it to 3 to indicate this is a register operand
+            in the regmem field.  */
+         if (!i.mem_operands)
+           i.rm.mode = 3;
        }
 
-    if (i.suffix && i.suffix != BYTE_MNEM_SUFFIX)
-      {
-       /* It's not a byte, select word/dword operation.  */
-       if (i.tm.opcode_modifier & W)
-         {
-           if (i.tm.opcode_modifier & ShortForm)
-             i.tm.base_opcode |= 8;
-           else
-             i.tm.base_opcode |= 1;
-         }
-       /* Now select between word & dword operations via the operand
-          size prefix, except for instructions that will ignore this
-          prefix anyway.  */
-       if (i.suffix != QWORD_MNEM_SUFFIX
-           && (i.suffix == LONG_MNEM_SUFFIX) == (flag_code == CODE_16BIT)
-           && !(i.tm.opcode_modifier & IgnoreSize))
-         {
-           unsigned int prefix = DATA_PREFIX_OPCODE;
-           if (i.tm.opcode_modifier & JumpByte) /* jcxz, loop */
-             prefix = ADDR_PREFIX_OPCODE;
-
-           if (! add_prefix (prefix))
-             return;
-         }
-
-       /* Set mode64 for an operand.  */
-       if (i.suffix == QWORD_MNEM_SUFFIX
-           && !(i.tm.opcode_modifier & NoRex64))
-         {
-           i.rex.mode64 = 1;
-           if (flag_code < CODE_64BIT)
-             {
-               as_bad (_("64bit operations available only in 64bit modes."));
-               return;
-             }
-         }
-
-       /* Size floating point instruction.  */
-       if (i.suffix == LONG_MNEM_SUFFIX)
-         {
-           if (i.tm.opcode_modifier & FloatMF)
-             i.tm.base_opcode ^= 4;
-         }
-      }
-
-    if (i.tm.opcode_modifier & ImmExt)
-      {
-       /* These AMD 3DNow! and Intel Katmai New Instructions have an
-          opcode suffix which is coded in the same place as an 8-bit
-          immediate field would be.  Here we fake an 8-bit immediate
-          operand from the opcode suffix stored in tm.extension_opcode.  */
-
-       expressionS *exp;
-
-       assert (i.imm_operands == 0 && i.operands <= 2 && 2 < MAX_OPERANDS);
-
-       exp = &im_expressions[i.imm_operands++];
-       i.op[i.operands].imms = exp;
-       i.types[i.operands++] = Imm8;
-       exp->X_op = O_constant;
-       exp->X_add_number = i.tm.extension_opcode;
-       i.tm.extension_opcode = None;
-      }
-
-    /* For insns with operands there are more diddles to do to the opcode.  */
-    if (i.operands)
-      {
-       /* Default segment register this instruction will use
-          for memory accesses.  0 means unknown.
-          This is only for optimizing out unnecessary segment overrides.  */
-       const seg_entry *default_seg = 0;
-
-       /* The imul $imm, %reg instruction is converted into
-          imul $imm, %reg, %reg, and the clr %reg instruction
-          is converted into xor %reg, %reg.  */
-       if (i.tm.opcode_modifier & regKludge)
-         {
-           unsigned int first_reg_op = (i.types[0] & Reg) ? 0 : 1;
-           /* Pretend we saw the extra register operand.  */
-           assert (i.op[first_reg_op + 1].regs == 0);
-           i.op[first_reg_op + 1].regs = i.op[first_reg_op].regs;
-           i.types[first_reg_op + 1] = i.types[first_reg_op];
-           i.reg_operands = 2;
-         }
-
-       if (i.tm.opcode_modifier & ShortForm)
-         {
-           /* The register or float register operand is in operand 0 or 1.  */
-           unsigned int op = (i.types[0] & (Reg | FloatReg)) ? 0 : 1;
-           /* Register goes in low 3 bits of opcode.  */
-           i.tm.base_opcode |= i.op[op].regs->reg_num;
-           if (i.op[op].regs->reg_flags & RegRex)
-             i.rex.extZ = 1;
-           if (!quiet_warnings && (i.tm.opcode_modifier & Ugh) != 0)
-             {
-               /* Warn about some common errors, but press on regardless.
-                  The first case can be generated by gcc (<= 2.8.1).  */
-               if (i.operands == 2)
-                 {
-                   /* Reversed arguments on faddp, fsubp, etc.  */
-                   as_warn (_("translating to `%s %%%s,%%%s'"), i.tm.name,
-                            i.op[1].regs->reg_name,
-                            i.op[0].regs->reg_name);
-                 }
-               else
-                 {
-                   /* Extraneous `l' suffix on fp insn.  */
-                   as_warn (_("translating to `%s %%%s'"), i.tm.name,
-                            i.op[0].regs->reg_name);
-                 }
-             }
-         }
-       else if (i.tm.opcode_modifier & Modrm)
-         {
-           /* The opcode is completed (modulo i.tm.extension_opcode which
-              must be put into the modrm byte).
-              Now, we make the modrm & index base bytes based on all the
-              info we've collected.  */
-
-           /* i.reg_operands MUST be the number of real register operands;
-              implicit registers do not count.  */
-           if (i.reg_operands == 2)
-             {
-               unsigned int source, dest;
-               source = ((i.types[0]
-                          & (Reg | RegMMX | RegXMM
-                             | SReg2 | SReg3
-                             | Control | Debug | Test))
-                         ? 0 : 1);
-               dest = source + 1;
-
-               i.rm.mode = 3;
-               /* One of the register operands will be encoded in the
-                  i.tm.reg field, the other in the combined i.tm.mode
-                  and i.tm.regmem fields.  If no form of this
-                  instruction supports a memory destination operand,
-                  then we assume the source operand may sometimes be
-                  a memory operand and so we need to store the
-                  destination in the i.rm.reg field.  */
-               if ((i.tm.operand_types[dest] & AnyMem) == 0)
-                 {
-                   i.rm.reg = i.op[dest].regs->reg_num;
-                   i.rm.regmem = i.op[source].regs->reg_num;
-                   if (i.op[dest].regs->reg_flags & RegRex)
-                     i.rex.extX = 1;
-                   if (i.op[source].regs->reg_flags & RegRex)
-                     i.rex.extZ = 1;
-                 }
-               else
-                 {
-                   i.rm.reg = i.op[source].regs->reg_num;
-                   i.rm.regmem = i.op[dest].regs->reg_num;
-                   if (i.op[dest].regs->reg_flags & RegRex)
-                     i.rex.extZ = 1;
-                   if (i.op[source].regs->reg_flags & RegRex)
-                     i.rex.extX = 1;
-                 }
-             }
-           else
-             {                 /* If it's not 2 reg operands...  */
-               if (i.mem_operands)
-                 {
-                   unsigned int fake_zero_displacement = 0;
-                   unsigned int op = ((i.types[0] & AnyMem)
-                                      ? 0
-                                      : (i.types[1] & AnyMem) ? 1 : 2);
-
-                   default_seg = &ds;
-
-                   if (! i.base_reg)
-                     {
-                       i.rm.mode = 0;
-                       if (! i.disp_operands)
-                         fake_zero_displacement = 1;
-                       if (! i.index_reg)
-                         {
-                           /* Operand is just <disp>  */
-                           if ((flag_code == CODE_16BIT) ^ (i.prefix[ADDR_PREFIX] != 0))
-                             {
-                               i.rm.regmem = NO_BASE_REGISTER_16;
-                               i.types[op] &= ~Disp;
-                               i.types[op] |= Disp16;
-                             }
-                           else if (flag_code != CODE_64BIT)
-                             {
-                               i.rm.regmem = NO_BASE_REGISTER;
-                               i.types[op] &= ~Disp;
-                               i.types[op] |= Disp32;
-                             }
-                           else
-                             {
-                               /* 64bit mode overwrites the 32bit
-                                  absolute addressing by RIP relative
-                                  addressing and absolute addressing
-                                  is encoded by one of the redundant
-                                  SIB forms.  */
-
-                               i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
-                               i.sib.base = NO_BASE_REGISTER;
-                               i.sib.index = NO_INDEX_REGISTER;
-                               i.types[op] &= ~Disp;
-                               i.types[op] |= Disp32S;
-                             }
-                         }
-                       else /* ! i.base_reg && i.index_reg  */
-                         {
-                           i.sib.index = i.index_reg->reg_num;
-                           i.sib.base = NO_BASE_REGISTER;
-                           i.sib.scale = i.log2_scale_factor;
-                           i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
-                           i.types[op] &= ~Disp;
-                           if (flag_code != CODE_64BIT)
-                             i.types[op] |= Disp32;    /* Must be 32 bit */
-                           else
-                             i.types[op] |= Disp32S;
-                           if (i.index_reg->reg_flags & RegRex)
-                             i.rex.extY = 1;
-                         }
-                     }
-                   /* RIP addressing for 64bit mode.  */
-                   else if (i.base_reg->reg_type == BaseIndex)
-                     {
-                       i.rm.regmem = NO_BASE_REGISTER;
-                       i.types[op] &= ~Disp;
-                       i.types[op] |= Disp32S;
-                       i.flags[op] = Operand_PCrel;
-                     }
-                   else if (i.base_reg->reg_type & Reg16)
-                     {
-                       switch (i.base_reg->reg_num)
-                         {
-                         case 3: /* (%bx)  */
-                           if (! i.index_reg)
-                             i.rm.regmem = 7;
-                           else /* (%bx,%si) -> 0, or (%bx,%di) -> 1  */
-                             i.rm.regmem = i.index_reg->reg_num - 6;
-                           break;
-                         case 5: /* (%bp)  */
-                           default_seg = &ss;
-                           if (! i.index_reg)
-                             {
-                               i.rm.regmem = 6;
-                               if ((i.types[op] & Disp) == 0)
-                                 {
-                                   /* fake (%bp) into 0(%bp)  */
-                                   i.types[op] |= Disp8;
-                                   fake_zero_displacement = 1;
-                                 }
-                             }
-                           else /* (%bp,%si) -> 2, or (%bp,%di) -> 3  */
-                             i.rm.regmem = i.index_reg->reg_num - 6 + 2;
-                           break;
-                         default: /* (%si) -> 4 or (%di) -> 5  */
-                           i.rm.regmem = i.base_reg->reg_num - 6 + 4;
-                         }
-                       i.rm.mode = mode_from_disp_size (i.types[op]);
-                     }
-                   else /* i.base_reg and 32/64 bit mode  */
-                     {
-                       if (flag_code == CODE_64BIT
-                           && (i.types[op] & Disp))
-                         {
-                           if (i.types[op] & Disp8)
-                             i.types[op] = Disp8 | Disp32S;
-                           else
-                             i.types[op] = Disp32S;
-                         }
-                       i.rm.regmem = i.base_reg->reg_num;
-                       if (i.base_reg->reg_flags & RegRex)
-                         i.rex.extZ = 1;
-                       i.sib.base = i.base_reg->reg_num;
-                       /* x86-64 ignores REX prefix bit here to avoid
-                          decoder complications.  */
-                       if ((i.base_reg->reg_num & 7) == EBP_REG_NUM)
-                         {
-                           default_seg = &ss;
-                           if (i.disp_operands == 0)
-                             {
-                               fake_zero_displacement = 1;
-                               i.types[op] |= Disp8;
-                             }
-                         }
-                       else if (i.base_reg->reg_num == ESP_REG_NUM)
-                         {
-                           default_seg = &ss;
-                         }
-                       i.sib.scale = i.log2_scale_factor;
-                       if (! i.index_reg)
-                         {
-                           /* <disp>(%esp) becomes two byte modrm
-                              with no index register.  We've already
-                              stored the code for esp in i.rm.regmem
-                              ie. ESCAPE_TO_TWO_BYTE_ADDRESSING.  Any
-                              base register besides %esp will not use
-                              the extra modrm byte.  */
-                           i.sib.index = NO_INDEX_REGISTER;
-#if ! SCALE1_WHEN_NO_INDEX
-                           /* Another case where we force the second
-                              modrm byte.  */
-                           if (i.log2_scale_factor)
-                             i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
-#endif
-                         }
-                       else
-                         {
-                           i.sib.index = i.index_reg->reg_num;
-                           i.rm.regmem = ESCAPE_TO_TWO_BYTE_ADDRESSING;
-                           if (i.index_reg->reg_flags & RegRex)
-                             i.rex.extY = 1;
-                         }
-                       i.rm.mode = mode_from_disp_size (i.types[op]);
-                     }
-
-                   if (fake_zero_displacement)
-                     {
-                       /* Fakes a zero displacement assuming that i.types[op]
-                          holds the correct displacement size.  */
-                       expressionS *exp;
-
-                       assert (i.op[op].disps == 0);
-                       exp = &disp_expressions[i.disp_operands++];
-                       i.op[op].disps = exp;
-                       exp->X_op = O_constant;
-                       exp->X_add_number = 0;
-                       exp->X_add_symbol = (symbolS *) 0;
-                       exp->X_op_symbol = (symbolS *) 0;
-                     }
-                 }
-
-               /* Fill in i.rm.reg or i.rm.regmem field with register
-                  operand (if any) based on i.tm.extension_opcode.
-                  Again, we must be careful to make sure that
-                  segment/control/debug/test/MMX registers are coded
-                  into the i.rm.reg field.  */
-               if (i.reg_operands)
-                 {
-                   unsigned int op =
-                     ((i.types[0]
-                       & (Reg | RegMMX | RegXMM
-                          | SReg2 | SReg3
-                          | Control | Debug | Test))
-                      ? 0
-                      : ((i.types[1]
-                          & (Reg | RegMMX | RegXMM
-                             | SReg2 | SReg3
-                             | Control | Debug | Test))
-                         ? 1
-                         : 2));
-                   /* If there is an extension opcode to put here, the
-                      register number must be put into the regmem field.  */
-                   if (i.tm.extension_opcode != None)
-                     {
-                       i.rm.regmem = i.op[op].regs->reg_num;
-                       if (i.op[op].regs->reg_flags & RegRex)
-                         i.rex.extZ = 1;
-                     }
-                   else
-                     {
-                       i.rm.reg = i.op[op].regs->reg_num;
-                       if (i.op[op].regs->reg_flags & RegRex)
-                         i.rex.extX = 1;
-                     }
-
-                   /* Now, if no memory operand has set i.rm.mode = 0, 1, 2
-                      we must set it to 3 to indicate this is a register
-                      operand in the regmem field.  */
-                   if (!i.mem_operands)
-                     i.rm.mode = 3;
-                 }
-
-               /* Fill in i.rm.reg field with extension opcode (if any).  */
-               if (i.tm.extension_opcode != None)
-                 i.rm.reg = i.tm.extension_opcode;
-             }
-         }
-       else if (i.tm.opcode_modifier & (Seg2ShortForm | Seg3ShortForm))
-         {
-           if (i.tm.base_opcode == POP_SEG_SHORT
-               && i.op[0].regs->reg_num == 1)
-             {
-               as_bad (_("you can't `pop %%cs'"));
-               return;
-             }
-           i.tm.base_opcode |= (i.op[0].regs->reg_num << 3);
-           if (i.op[0].regs->reg_flags & RegRex)
-             i.rex.extZ = 1;
-         }
-       else if ((i.tm.base_opcode & ~(D|W)) == MOV_AX_DISP32)
-         {
-           default_seg = &ds;
-         }
-       else if ((i.tm.opcode_modifier & IsString) != 0)
-         {
-           /* For the string instructions that allow a segment override
-              on one of their operands, the default segment is ds.  */
-           default_seg = &ds;
-         }
-
-       /* If a segment was explicitly specified,
-          and the specified segment is not the default,
-          use an opcode prefix to select it.
-          If we never figured out what the default segment is,
-          then default_seg will be zero at this point,
-          and the specified segment prefix will always be used.  */
-       if ((i.seg[0]) && (i.seg[0] != default_seg))
-         {
-           if (! add_prefix (i.seg[0]->seg_prefix))
-             return;
-         }
-      }
-    else if (!quiet_warnings && (i.tm.opcode_modifier & Ugh) != 0)
-      {
-       /* UnixWare fsub no args is alias for fsubp, fadd -> faddp, etc.  */
-       as_warn (_("translating to `%sp'"), i.tm.name);
-      }
-  }
+      /* Fill in i.rm.reg field with extension opcode (if any).  */
+      if (i.tm.extension_opcode != None)
+       i.rm.reg = i.tm.extension_opcode;
+    }
+  return default_seg;
+}
 
-  /* Handle conversion of 'int $3' --> special int3 insn.  */
-  if (i.tm.base_opcode == INT_OPCODE && i.op[0].imms->X_add_number == 3)
+static void
+output_branch ()
+{
+  char *p;
+  int code16;
+  int prefix;
+  relax_substateT subtype;
+  symbolS *sym;
+  offsetT off;
+
+  code16 = 0;
+  if (flag_code == CODE_16BIT)
+    code16 = CODE16;
+
+  prefix = 0;
+  if (i.prefix[DATA_PREFIX] != 0)
     {
-      i.tm.base_opcode = INT3_OPCODE;
-      i.imm_operands = 0;
+      prefix = 1;
+      i.prefixes -= 1;
+      code16 ^= CODE16;
     }
-
-  if ((i.tm.opcode_modifier & (Jump | JumpByte | JumpDword))
-      && i.op[0].disps->X_op == O_constant)
+  /* Pentium4 branch hints.  */
+  if (i.prefix[SEG_PREFIX] == CS_PREFIX_OPCODE /* not taken */
+      || i.prefix[SEG_PREFIX] == DS_PREFIX_OPCODE /* taken */)
     {
-      /* Convert "jmp constant" (and "call constant") to a jump (call) to
-        the absolute address given by the constant.  Since ix86 jumps and
-        calls are pc relative, we need to generate a reloc.  */
-      i.op[0].disps->X_add_symbol = &abs_symbol;
-      i.op[0].disps->X_op = O_symbol;
+      prefix++;
+      i.prefixes--;
+    }
+  if (i.prefix[REX_PREFIX] != 0)
+    {
+      prefix++;
+      i.prefixes--;
     }
 
-  if (i.tm.opcode_modifier & Rex64)
-    i.rex.mode64 = 1;
+  if (i.prefixes != 0 && !intel_syntax)
+    as_warn (_("skipping prefixes on this instruction"));
+
+  /* It's always a symbol;  End frag & setup for relax.
+     Make sure there is enough room in this frag for the largest
+     instruction we may generate in md_convert_frag.  This is 2
+     bytes for the opcode and room for the prefix and largest
+     displacement.  */
+  frag_grow (prefix + 2 + 4);
+  /* Prefix and 1 opcode byte go in fr_fix.  */
+  p = frag_more (prefix + 1);
+  if (i.prefix[DATA_PREFIX] != 0)
+    *p++ = DATA_PREFIX_OPCODE;
+  if (i.prefix[SEG_PREFIX] == CS_PREFIX_OPCODE
+      || i.prefix[SEG_PREFIX] == DS_PREFIX_OPCODE)
+    *p++ = i.prefix[SEG_PREFIX];
+  if (i.prefix[REX_PREFIX] != 0)
+    *p++ = i.prefix[REX_PREFIX];
+  *p = i.tm.base_opcode;
+
+  if ((unsigned char) *p == JUMP_PC_RELATIVE)
+    subtype = ENCODE_RELAX_STATE (UNCOND_JUMP, SMALL);
+  else if ((cpu_arch_flags & Cpu386) != 0)
+    subtype = ENCODE_RELAX_STATE (COND_JUMP, SMALL);
+  else
+    subtype = ENCODE_RELAX_STATE (COND_JUMP86, SMALL);
+  subtype |= code16;
 
-  /* For 8bit registers we would need an empty rex prefix.
-     Also in the case instruction is already having prefix,
-     we need to convert old registers to new ones.  */
+  sym = i.op[0].disps->X_add_symbol;
+  off = i.op[0].disps->X_add_number;
 
-  if (((i.types[0] & Reg8) && (i.op[0].regs->reg_flags & RegRex64))
-      || ((i.types[1] & Reg8) && (i.op[1].regs->reg_flags & RegRex64))
-      || ((i.rex.mode64 || i.rex.extX || i.rex.extY || i.rex.extZ || i.rex.empty)
-         && ((i.types[0] & Reg8) || (i.types[1] & Reg8))))
+  if (i.op[0].disps->X_op != O_constant
+      && i.op[0].disps->X_op != O_symbol)
     {
-      int x;
-      i.rex.empty = 1;
-      for (x = 0; x < 2; x++)
-       {
-         /* Look for 8bit operand that does use old registers.  */
-         if (i.types[x] & Reg8
-             && !(i.op[x].regs->reg_flags & RegRex64))
-           {
-             /* In case it is "hi" register, give up.  */
-             if (i.op[x].regs->reg_num > 3)
-               as_bad (_("Can't encode registers '%%%s' in the instruction requiring REX prefix.\n"),
-                       i.op[x].regs->reg_name);
+      /* Handle complex expressions.  */
+      sym = make_expr_symbol (i.op[0].disps);
+      off = 0;
+    }
 
-             /* Otherwise it is equivalent to the extended register.
-                Since the encoding don't change this is merely cosmetical
-                cleanup for debug output.  */
+  /* 1 possible extra opcode + 4 byte displacement go in var part.
+     Pass reloc in fr_var.  */
+  frag_var (rs_machine_dependent, 5, i.reloc[0], subtype, sym, off, p);
+}
 
-             i.op[x].regs = i.op[x].regs + 8;
-           }
+static void
+output_jump ()
+{
+  char *p;
+  int size;
+
+  if (i.tm.opcode_modifier & JumpByte)
+    {
+      /* This is a loop or jecxz type instruction.  */
+      size = 1;
+      if (i.prefix[ADDR_PREFIX] != 0)
+       {
+         FRAG_APPEND_1_CHAR (ADDR_PREFIX_OPCODE);
+         i.prefixes -= 1;
+       }
+      /* Pentium4 branch hints.  */
+      if (i.prefix[SEG_PREFIX] == CS_PREFIX_OPCODE /* not taken */
+         || i.prefix[SEG_PREFIX] == DS_PREFIX_OPCODE /* taken */)
+       {
+         FRAG_APPEND_1_CHAR (i.prefix[SEG_PREFIX]);
+         i.prefixes--;
        }
     }
+  else
+    {
+      int code16;
 
-  if (i.rex.mode64 || i.rex.extX || i.rex.extY || i.rex.extZ || i.rex.empty)
-    add_prefix (0x40
-               | (i.rex.mode64 ? 8 : 0)
-               | (i.rex.extX ? 4 : 0)
-               | (i.rex.extY ? 2 : 0)
-               | (i.rex.extZ ? 1 : 0));
+      code16 = 0;
+      if (flag_code == CODE_16BIT)
+       code16 = CODE16;
 
-  /* We are ready to output the insn.  */
-  {
-    register char *p;
+      if (i.prefix[DATA_PREFIX] != 0)
+       {
+         FRAG_APPEND_1_CHAR (DATA_PREFIX_OPCODE);
+         i.prefixes -= 1;
+         code16 ^= CODE16;
+       }
 
-    /* Tie dwarf2 debug info to the address at the start of the insn.
-       We can't do this after the insn has been output as the current
-       frag may have been closed off.  eg. by frag_var.  */
-    dwarf2_emit_insn (0);
+      size = 4;
+      if (code16)
+       size = 2;
+    }
 
-    /* Output jumps.  */
-    if (i.tm.opcode_modifier & Jump)
-      {
-       int code16;
-       int prefix;
+  if (i.prefix[REX_PREFIX] != 0)
+    {
+      FRAG_APPEND_1_CHAR (i.prefix[REX_PREFIX]);
+      i.prefixes -= 1;
+    }
 
-       code16 = 0;
-       if (flag_code == CODE_16BIT)
-         code16 = CODE16;
+  if (i.prefixes != 0 && !intel_syntax)
+    as_warn (_("skipping prefixes on this instruction"));
 
-       prefix = 0;
-       if (i.prefix[DATA_PREFIX])
-         {
-           prefix = 1;
-           i.prefixes -= 1;
-           code16 ^= CODE16;
-         }
-       /* Pentium4 branch hints.  */
-       if (i.prefix[SEG_PREFIX] == CS_PREFIX_OPCODE /* not taken */
-           || i.prefix[SEG_PREFIX] == DS_PREFIX_OPCODE /* taken */)
-         {
-           prefix++;
-           i.prefixes--;
-         }
-       if (i.prefix[REX_PREFIX])
-         {
-           prefix++;
-           i.prefixes--;
-         }
+  p = frag_more (1 + size);
+  *p++ = i.tm.base_opcode;
 
-       if (i.prefixes != 0 && !intel_syntax)
-         as_warn (_("skipping prefixes on this instruction"));
-
-       /* It's always a symbol;  End frag & setup for relax.
-          Make sure there is enough room in this frag for the largest
-          instruction we may generate in md_convert_frag.  This is 2
-          bytes for the opcode and room for the prefix and largest
-          displacement.  */
-       frag_grow (prefix + 2 + 4);
-       /* Prefix and 1 opcode byte go in fr_fix.  */
-       p = frag_more (prefix + 1);
-       if (i.prefix[DATA_PREFIX])
-         *p++ = DATA_PREFIX_OPCODE;
-       if (i.prefix[SEG_PREFIX] == CS_PREFIX_OPCODE
-           || i.prefix[SEG_PREFIX] == DS_PREFIX_OPCODE)
-         *p++ = i.prefix[SEG_PREFIX];
-       if (i.prefix[REX_PREFIX])
-         *p++ = i.prefix[REX_PREFIX];
-       *p = i.tm.base_opcode;
-       /* 1 possible extra opcode + displacement go in var part.
-          Pass reloc in fr_var.  */
-       frag_var (rs_machine_dependent,
-                 1 + 4,
-                 i.reloc[0],
-                 ((unsigned char) *p == JUMP_PC_RELATIVE
-                  ? ENCODE_RELAX_STATE (UNCOND_JUMP, SMALL) | code16
-                  : ((cpu_arch_flags & Cpu386) != 0
-                     ? ENCODE_RELAX_STATE (COND_JUMP, SMALL) | code16
-                     : ENCODE_RELAX_STATE (COND_JUMP86, SMALL) | code16)),
-                 i.op[0].disps->X_add_symbol,
-                 i.op[0].disps->X_add_number,
-                 p);
-      }
-    else if (i.tm.opcode_modifier & (JumpByte | JumpDword))
-      {
-       int size;
+  fix_new_exp (frag_now, p - frag_now->fr_literal, size,
+              i.op[0].disps, 1, reloc (size, 1, 1, i.reloc[0]));
+}
 
-       if (i.tm.opcode_modifier & JumpByte)
-         {
-           /* This is a loop or jecxz type instruction.  */
-           size = 1;
-           if (i.prefix[ADDR_PREFIX])
-             {
-               FRAG_APPEND_1_CHAR (ADDR_PREFIX_OPCODE);
-               i.prefixes -= 1;
-             }
-           /* Pentium4 branch hints.  */
-           if (i.prefix[SEG_PREFIX] == CS_PREFIX_OPCODE /* not taken */
-               || i.prefix[SEG_PREFIX] == DS_PREFIX_OPCODE /* taken */)
-             {
-               FRAG_APPEND_1_CHAR (i.prefix[SEG_PREFIX]);
-               i.prefixes--;
-             }
-         }
-       else
-         {
-           int code16;
+static void
+output_interseg_jump ()
+{
+  char *p;
+  int size;
+  int prefix;
+  int code16;
 
-           code16 = 0;
-           if (flag_code == CODE_16BIT)
-             code16 = CODE16;
+  code16 = 0;
+  if (flag_code == CODE_16BIT)
+    code16 = CODE16;
 
-           if (i.prefix[DATA_PREFIX])
-             {
-               FRAG_APPEND_1_CHAR (DATA_PREFIX_OPCODE);
-               i.prefixes -= 1;
-               code16 ^= CODE16;
-             }
+  prefix = 0;
+  if (i.prefix[DATA_PREFIX] != 0)
+    {
+      prefix = 1;
+      i.prefixes -= 1;
+      code16 ^= CODE16;
+    }
+  if (i.prefix[REX_PREFIX] != 0)
+    {
+      prefix++;
+      i.prefixes -= 1;
+    }
 
-           size = 4;
-           if (code16)
-             size = 2;
-         }
+  size = 4;
+  if (code16)
+    size = 2;
 
-       if (i.prefix[REX_PREFIX])
-         {
-           FRAG_APPEND_1_CHAR (i.prefix[REX_PREFIX]);
-           i.prefixes -= 1;
-         }
+  if (i.prefixes != 0 && !intel_syntax)
+    as_warn (_("skipping prefixes on this instruction"));
 
-       if (i.prefixes != 0 && !intel_syntax)
-         as_warn (_("skipping prefixes on this instruction"));
+  /* 1 opcode; 2 segment; offset  */
+  p = frag_more (prefix + 1 + 2 + size);
 
-       p = frag_more (1 + size);
-       *p++ = i.tm.base_opcode;
+  if (i.prefix[DATA_PREFIX] != 0)
+    *p++ = DATA_PREFIX_OPCODE;
 
-       fix_new_exp (frag_now, p - frag_now->fr_literal, size,
-                    i.op[0].disps, 1, reloc (size, 1, 1, i.reloc[0]));
-      }
-    else if (i.tm.opcode_modifier & JumpInterSegment)
-      {
-       int size;
-       int prefix;
-       int code16;
+  if (i.prefix[REX_PREFIX] != 0)
+    *p++ = i.prefix[REX_PREFIX];
 
-       code16 = 0;
-       if (flag_code == CODE_16BIT)
-         code16 = CODE16;
+  *p++ = i.tm.base_opcode;
+  if (i.op[1].imms->X_op == O_constant)
+    {
+      offsetT n = i.op[1].imms->X_add_number;
 
-       prefix = 0;
-       if (i.prefix[DATA_PREFIX])
-         {
-           prefix = 1;
-           i.prefixes -= 1;
-           code16 ^= CODE16;
-         }
-       if (i.prefix[REX_PREFIX])
-         {
-           prefix++;
-           i.prefixes -= 1;
-         }
+      if (size == 2
+         && !fits_in_unsigned_word (n)
+         && !fits_in_signed_word (n))
+       {
+         as_bad (_("16-bit jump out of range"));
+         return;
+       }
+      md_number_to_chars (p, n, size);
+    }
+  else
+    fix_new_exp (frag_now, p - frag_now->fr_literal, size,
+                i.op[1].imms, 0, reloc (size, 0, 0, i.reloc[1]));
+  if (i.op[0].imms->X_op != O_constant)
+    as_bad (_("can't handle non absolute segment in `%s'"),
+           i.tm.name);
+  md_number_to_chars (p + size, (valueT) i.op[0].imms->X_add_number, 2);
+}
 
-       size = 4;
-       if (code16)
-         size = 2;
 
-       if (i.prefixes != 0 && !intel_syntax)
-         as_warn (_("skipping prefixes on this instruction"));
+static void
+output_insn ()
+{
+  fragS *insn_start_frag;
+  offsetT insn_start_off;
+
+  /* Tie dwarf2 debug info to the address at the start of the insn.
+     We can't do this after the insn has been output as the current
+     frag may have been closed off.  eg. by frag_var.  */
+  dwarf2_emit_insn (0);
+
+  insn_start_frag = frag_now;
+  insn_start_off = frag_now_fix ();
+
+  /* Output jumps.  */
+  if (i.tm.opcode_modifier & Jump)
+    output_branch ();
+  else if (i.tm.opcode_modifier & (JumpByte | JumpDword))
+    output_jump ();
+  else if (i.tm.opcode_modifier & JumpInterSegment)
+    output_interseg_jump ();
+  else
+    {
+      /* Output normal instructions here.  */
+      char *p;
+      unsigned char *q;
+
+      /* All opcodes on i386 have either 1 or 2 bytes.  We may use third
+        byte for the SSE instructions to specify a prefix they require.  */
+      if (i.tm.base_opcode & 0xff0000)
+       add_prefix ((i.tm.base_opcode >> 16) & 0xff);
+
+      /* The prefix bytes.  */
+      for (q = i.prefix;
+          q < i.prefix + sizeof (i.prefix) / sizeof (i.prefix[0]);
+          q++)
+       {
+         if (*q)
+           {
+             p = frag_more (1);
+             md_number_to_chars (p, (valueT) *q, 1);
+           }
+       }
 
-       /* 1 opcode; 2 segment; offset  */
-       p = frag_more (prefix + 1 + 2 + size);
+      /* Now the opcode; be careful about word order here!  */
+      if (fits_in_unsigned_byte (i.tm.base_opcode))
+       {
+         FRAG_APPEND_1_CHAR (i.tm.base_opcode);
+       }
+      else
+       {
+         p = frag_more (2);
+         /* Put out high byte first: can't use md_number_to_chars!  */
+         *p++ = (i.tm.base_opcode >> 8) & 0xff;
+         *p = i.tm.base_opcode & 0xff;
+       }
 
-       if (i.prefix[DATA_PREFIX])
-         *p++ = DATA_PREFIX_OPCODE;
+      /* Now the modrm byte and sib byte (if present).  */
+      if (i.tm.opcode_modifier & Modrm)
+       {
+         p = frag_more (1);
+         md_number_to_chars (p,
+                             (valueT) (i.rm.regmem << 0
+                                       | i.rm.reg << 3
+                                       | i.rm.mode << 6),
+                             1);
+         /* If i.rm.regmem == ESP (4)
+            && i.rm.mode != (Register mode)
+            && not 16 bit
+            ==> need second modrm byte.  */
+         if (i.rm.regmem == ESCAPE_TO_TWO_BYTE_ADDRESSING
+             && i.rm.mode != 3
+             && !(i.base_reg && (i.base_reg->reg_type & Reg16) != 0))
+           {
+             p = frag_more (1);
+             md_number_to_chars (p,
+                                 (valueT) (i.sib.base << 0
+                                           | i.sib.index << 3
+                                           | i.sib.scale << 6),
+                                 1);
+           }
+       }
 
-       if (i.prefix[REX_PREFIX])
-         *p++ = i.prefix[REX_PREFIX];
+      if (i.disp_operands)
+       output_disp (insn_start_frag, insn_start_off);
 
-       *p++ = i.tm.base_opcode;
-       if (i.op[1].imms->X_op == O_constant)
-         {
-           offsetT n = i.op[1].imms->X_add_number;
+      if (i.imm_operands)
+       output_imm (insn_start_frag, insn_start_off);
+    }
 
-           if (size == 2
-               && !fits_in_unsigned_word (n)
-               && !fits_in_signed_word (n))
-             {
-               as_bad (_("16-bit jump out of range"));
-               return;
-             }
-           md_number_to_chars (p, n, size);
-         }
-       else
-         fix_new_exp (frag_now, p - frag_now->fr_literal, size,
-                      i.op[1].imms, 0, reloc (size, 0, 0, i.reloc[1]));
-       if (i.op[0].imms->X_op != O_constant)
-         as_bad (_("can't handle non absolute segment in `%s'"),
-                 i.tm.name);
-       md_number_to_chars (p + size, (valueT) i.op[0].imms->X_add_number, 2);
-      }
-    else
-      {
-       /* Output normal instructions here.  */
-       unsigned char *q;
-
-       /* All opcodes on i386 have eighter 1 or 2 bytes.  We may use third
-          byte for the SSE instructions to specify prefix they require.  */
-       if (i.tm.base_opcode & 0xff0000)
-         add_prefix ((i.tm.base_opcode >> 16) & 0xff);
-
-       /* The prefix bytes.  */
-       for (q = i.prefix;
-            q < i.prefix + sizeof (i.prefix) / sizeof (i.prefix[0]);
-            q++)
-         {
-           if (*q)
-             {
-               p = frag_more (1);
-               md_number_to_chars (p, (valueT) *q, 1);
-             }
-         }
+#ifdef DEBUG386
+  if (flag_debug)
+    {
+      pi (line, &i);
+    }
+#endif /* DEBUG386  */
+}
 
-       /* Now the opcode; be careful about word order here!  */
-       if (fits_in_unsigned_byte (i.tm.base_opcode))
-         {
-           FRAG_APPEND_1_CHAR (i.tm.base_opcode);
-         }
-       else
-         {
-           p = frag_more (2);
-           /* Put out high byte first: can't use md_number_to_chars!  */
-           *p++ = (i.tm.base_opcode >> 8) & 0xff;
-           *p = i.tm.base_opcode & 0xff;
-         }
+static void
+output_disp (insn_start_frag, insn_start_off)
+    fragS *insn_start_frag;
+    offsetT insn_start_off;
+{
+  char *p;
+  unsigned int n;
 
-       /* Now the modrm byte and sib byte (if present).  */
-       if (i.tm.opcode_modifier & Modrm)
-         {
-           p = frag_more (1);
-           md_number_to_chars (p,
-                               (valueT) (i.rm.regmem << 0
-                                         | i.rm.reg << 3
-                                         | i.rm.mode << 6),
-                               1);
-           /* If i.rm.regmem == ESP (4)
-              && i.rm.mode != (Register mode)
-              && not 16 bit
-              ==> need second modrm byte.  */
-           if (i.rm.regmem == ESCAPE_TO_TWO_BYTE_ADDRESSING
-               && i.rm.mode != 3
-               && !(i.base_reg && (i.base_reg->reg_type & Reg16) != 0))
-             {
-               p = frag_more (1);
-               md_number_to_chars (p,
-                                   (valueT) (i.sib.base << 0
-                                             | i.sib.index << 3
-                                             | i.sib.scale << 6),
-                                   1);
-             }
-         }
+  for (n = 0; n < i.operands; n++)
+    {
+      if (i.types[n] & Disp)
+       {
+         if (i.op[n].disps->X_op == O_constant)
+           {
+             int size;
+             offsetT val;
 
-       if (i.disp_operands)
-         {
-           register unsigned int n;
+             size = 4;
+             if (i.types[n] & (Disp8 | Disp16 | Disp64))
+               {
+                 size = 2;
+                 if (i.types[n] & Disp8)
+                   size = 1;
+                 if (i.types[n] & Disp64)
+                   size = 8;
+               }
+             val = offset_in_range (i.op[n].disps->X_add_number,
+                                    size);
+             p = frag_more (size);
+             md_number_to_chars (p, val, size);
+           }
+         else
+           {
+             RELOC_ENUM reloc_type;
+             int size = 4;
+             int sign = 0;
+             int pcrel = (i.flags[n] & Operand_PCrel) != 0;
+
+             /* The PC relative address is computed relative
+                to the instruction boundary, so in case immediate
+                fields follows, we need to adjust the value.  */
+             if (pcrel && i.imm_operands)
+               {
+                 int imm_size = 4;
+                 unsigned int n1;
 
-           for (n = 0; n < i.operands; n++)
-             {
-               if (i.types[n] & Disp)
-                 {
-                   if (i.op[n].disps->X_op == O_constant)
+                 for (n1 = 0; n1 < i.operands; n1++)
+                   if (i.types[n1] & Imm)
                      {
-                       int size;
-                       offsetT val;
-
-                       size = 4;
-                       if (i.types[n] & (Disp8 | Disp16 | Disp64))
+                       if (i.types[n1] & (Imm8 | Imm8S | Imm16 | Imm64))
                          {
-                           size = 2;
-                           if (i.types[n] & Disp8)
-                             size = 1;
-                           if (i.types[n] & Disp64)
-                             size = 8;
+                           imm_size = 2;
+                           if (i.types[n1] & (Imm8 | Imm8S))
+                             imm_size = 1;
+                           if (i.types[n1] & Imm64)
+                             imm_size = 8;
                          }
-                       val = offset_in_range (i.op[n].disps->X_add_number,
-                                              size);
-                       p = frag_more (size);
-                       md_number_to_chars (p, val, size);
+                       break;
                      }
-                   else
-                     {
-                       int size = 4;
-                       int sign = 0;
-                       int pcrel = (i.flags[n] & Operand_PCrel) != 0;
-
-                       /* The PC relative address is computed relative
-                          to the instruction boundary, so in case immediate
-                          fields follows, we need to adjust the value.  */
-                       if (pcrel && i.imm_operands)
-                         {
-                           int imm_size = 4;
-                           register unsigned int n1;
-
-                           for (n1 = 0; n1 < i.operands; n1++)
-                             if (i.types[n1] & Imm)
-                               {
-                                 if (i.types[n1] & (Imm8 | Imm8S | Imm16 | Imm64))
-                                   {
-                                     imm_size = 2;
-                                     if (i.types[n1] & (Imm8 | Imm8S))
-                                       imm_size = 1;
-                                     if (i.types[n1] & Imm64)
-                                       imm_size = 8;
-                                   }
-                                 break;
-                               }
-                           /* We should find the immediate.  */
-                           if (n1 == i.operands)
-                             abort ();
-                           i.op[n].disps->X_add_number -= imm_size;
-                         }
+                 /* We should find the immediate.  */
+                 if (n1 == i.operands)
+                   abort ();
+                 i.op[n].disps->X_add_number -= imm_size;
+               }
 
-                       if (i.types[n] & Disp32S)
-                         sign = 1;
+             if (i.types[n] & Disp32S)
+               sign = 1;
 
-                       if (i.types[n] & (Disp16 | Disp64))
-                         {
-                           size = 2;
-                           if (i.types[n] & Disp64)
-                             size = 8;
-                         }
+             if (i.types[n] & (Disp16 | Disp64))
+               {
+                 size = 2;
+                 if (i.types[n] & Disp64)
+                   size = 8;
+               }
 
-                       p = frag_more (size);
-                       fix_new_exp (frag_now, p - frag_now->fr_literal, size,
-                                    i.op[n].disps, pcrel,
-                                    reloc (size, pcrel, sign, i.reloc[n]));
-                     }
-                 }
-             }
-         }
+             p = frag_more (size);
+             reloc_type = reloc (size, pcrel, sign, i.reloc[n]);
+#ifdef BFD_ASSEMBLER
+             if (reloc_type == BFD_RELOC_32
+                 && GOT_symbol
+                 && GOT_symbol == i.op[n].disps->X_add_symbol
+                 && (i.op[n].disps->X_op == O_symbol
+                     || (i.op[n].disps->X_op == O_add
+                         && ((symbol_get_value_expression
+                              (i.op[n].disps->X_op_symbol)->X_op)
+                             == O_subtract))))
+               {
+                 offsetT add;
 
-       /* Output immediate.  */
-       if (i.imm_operands)
-         {
-           register unsigned int n;
+                 if (insn_start_frag == frag_now)
+                   add = (p - frag_now->fr_literal) - insn_start_off;
+                 else
+                   {
+                     fragS *fr;
 
-           for (n = 0; n < i.operands; n++)
-             {
-               if (i.types[n] & Imm)
-                 {
-                   if (i.op[n].imms->X_op == O_constant)
-                     {
-                       int size;
-                       offsetT val;
+                     add = insn_start_frag->fr_fix - insn_start_off;
+                     for (fr = insn_start_frag->fr_next;
+                          fr && fr != frag_now; fr = fr->fr_next)
+                       add += fr->fr_fix;
+                     add += p - frag_now->fr_literal;
+                   }
 
-                       size = 4;
-                       if (i.types[n] & (Imm8 | Imm8S | Imm16 | Imm64))
-                         {
-                           size = 2;
-                           if (i.types[n] & (Imm8 | Imm8S))
-                             size = 1;
-                           else if (i.types[n] & Imm64)
-                             size = 8;
-                         }
-                       val = offset_in_range (i.op[n].imms->X_add_number,
-                                              size);
-                       p = frag_more (size);
-                       md_number_to_chars (p, val, size);
-                     }
-                   else
-                     {
-                       /* Not absolute_section.
-                          Need a 32-bit fixup (don't support 8bit
-                          non-absolute imms).  Try to support other
-                          sizes ...  */
-                       RELOC_ENUM reloc_type;
-                       int size = 4;
-                       int sign = 0;
-
-                       if ((i.types[n] & (Imm32S))
-                           && i.suffix == QWORD_MNEM_SUFFIX)
-                         sign = 1;
-                       if (i.types[n] & (Imm8 | Imm8S | Imm16 | Imm64))
-                         {
-                           size = 2;
-                           if (i.types[n] & (Imm8 | Imm8S))
-                             size = 1;
-                           if (i.types[n] & Imm64)
-                             size = 8;
-                         }
+                 /* We don't support dynamic linking on x86-64 yet.  */
+                 if (flag_code == CODE_64BIT)
+                   abort ();
+                 reloc_type = BFD_RELOC_386_GOTPC;
+                 i.op[n].disps->X_add_number += add;
+               }
+#endif
+             fix_new_exp (frag_now, p - frag_now->fr_literal, size,
+                          i.op[n].disps, pcrel, reloc_type);
+           }
+       }
+    }
+}
+
+static void
+output_imm (insn_start_frag, insn_start_off)
+    fragS *insn_start_frag;
+    offsetT insn_start_off;
+{
+  char *p;
+  unsigned int n;
+
+  for (n = 0; n < i.operands; n++)
+    {
+      if (i.types[n] & Imm)
+       {
+         if (i.op[n].imms->X_op == O_constant)
+           {
+             int size;
+             offsetT val;
+
+             size = 4;
+             if (i.types[n] & (Imm8 | Imm8S | Imm16 | Imm64))
+               {
+                 size = 2;
+                 if (i.types[n] & (Imm8 | Imm8S))
+                   size = 1;
+                 else if (i.types[n] & Imm64)
+                   size = 8;
+               }
+             val = offset_in_range (i.op[n].imms->X_add_number,
+                                    size);
+             p = frag_more (size);
+             md_number_to_chars (p, val, size);
+           }
+         else
+           {
+             /* Not absolute_section.
+                Need a 32-bit fixup (don't support 8bit
+                non-absolute imms).  Try to support other
+                sizes ...  */
+             RELOC_ENUM reloc_type;
+             int size = 4;
+             int sign = 0;
+
+             if ((i.types[n] & (Imm32S))
+                 && i.suffix == QWORD_MNEM_SUFFIX)
+               sign = 1;
+             if (i.types[n] & (Imm8 | Imm8S | Imm16 | Imm64))
+               {
+                 size = 2;
+                 if (i.types[n] & (Imm8 | Imm8S))
+                   size = 1;
+                 if (i.types[n] & Imm64)
+                   size = 8;
+               }
 
-                       p = frag_more (size);
-                       reloc_type = reloc (size, 0, sign, i.reloc[n]);
+             p = frag_more (size);
+             reloc_type = reloc (size, 0, sign, i.reloc[n]);
 #ifdef BFD_ASSEMBLER
-                       if (reloc_type == BFD_RELOC_32
-                           && GOT_symbol
-                           && GOT_symbol == i.op[n].imms->X_add_symbol
-                           && (i.op[n].imms->X_op == O_symbol
-                               || (i.op[n].imms->X_op == O_add
-                                   && ((symbol_get_value_expression
-                                        (i.op[n].imms->X_op_symbol)->X_op)
-                                       == O_subtract))))
-                         {
-                           /* We don't support dynamic linking on x86-64 yet.  */
-                           if (flag_code == CODE_64BIT)
-                             abort ();
-                           reloc_type = BFD_RELOC_386_GOTPC;
-                           i.op[n].imms->X_add_number += 3;
-                         }
-#endif
-                       fix_new_exp (frag_now, p - frag_now->fr_literal, size,
-                                    i.op[n].imms, 0, reloc_type);
-                     }
-                 }
-             }
-         }
-      }
+             /*   This is tough to explain.  We end up with this one if we
+              * have operands that look like
+              * "_GLOBAL_OFFSET_TABLE_+[.-.L284]".  The goal here is to
+              * obtain the absolute address of the GOT, and it is strongly
+              * preferable from a performance point of view to avoid using
+              * a runtime relocation for this.  The actual sequence of
+              * instructions often look something like:
+              *
+              *        call    .L66
+              * .L66:
+              *        popl    %ebx
+              *        addl    $_GLOBAL_OFFSET_TABLE_+[.-.L66],%ebx
+              *
+              *   The call and pop essentially return the absolute address
+              * of the label .L66 and store it in %ebx.  The linker itself
+              * will ultimately change the first operand of the addl so
+              * that %ebx points to the GOT, but to keep things simple, the
+              * .o file must have this operand set so that it generates not
+              * the absolute address of .L66, but the absolute address of
+              * itself.  This allows the linker itself simply treat a GOTPC
+              * relocation as asking for a pcrel offset to the GOT to be
+              * added in, and the addend of the relocation is stored in the
+              * operand field for the instruction itself.
+              *
+              *   Our job here is to fix the operand so that it would add
+              * the correct offset so that %ebx would point to itself.  The
+              * thing that is tricky is that .-.L66 will point to the
+              * beginning of the instruction, so we need to further modify
+              * the operand so that it will point to itself.  There are
+              * other cases where you have something like:
+              *
+              *        .long   $_GLOBAL_OFFSET_TABLE_+[.-.L66]
+              *
+              * and here no correction would be required.  Internally in
+              * the assembler we treat operands of this form as not being
+              * pcrel since the '.' is explicitly mentioned, and I wonder
+              * whether it would simplify matters to do it this way.  Who
+              * knows.  In earlier versions of the PIC patches, the
+              * pcrel_adjust field was used to store the correction, but
+              * since the expression is not pcrel, I felt it would be
+              * confusing to do it this way.  */
+
+             if (reloc_type == BFD_RELOC_32
+                 && GOT_symbol
+                 && GOT_symbol == i.op[n].imms->X_add_symbol
+                 && (i.op[n].imms->X_op == O_symbol
+                     || (i.op[n].imms->X_op == O_add
+                         && ((symbol_get_value_expression
+                              (i.op[n].imms->X_op_symbol)->X_op)
+                             == O_subtract))))
+               {
+                 offsetT add;
 
-#ifdef DEBUG386
-    if (flag_debug)
-      {
-       pi (line, &i);
-      }
-#endif /* DEBUG386  */
-  }
+                 if (insn_start_frag == frag_now)
+                   add = (p - frag_now->fr_literal) - insn_start_off;
+                 else
+                   {
+                     fragS *fr;
+
+                     add = insn_start_frag->fr_fix - insn_start_off;
+                     for (fr = insn_start_frag->fr_next;
+                          fr && fr != frag_now; fr = fr->fr_next)
+                       add += fr->fr_fix;
+                     add += p - frag_now->fr_literal;
+                   }
+
+                 /* We don't support dynamic linking on x86-64 yet.  */
+                 if (flag_code == CODE_64BIT)
+                   abort ();
+                 reloc_type = BFD_RELOC_386_GOTPC;
+                 i.op[n].imms->X_add_number += add;
+               }
+#endif
+             fix_new_exp (frag_now, p - frag_now->fr_literal, size,
+                          i.op[n].imms, 0, reloc_type);
+           }
+       }
+    }
 }
 \f
 #ifndef LEX_AT
@@ -3164,10 +3491,19 @@ lex_got (reloc, adjust)
     const char *str;
     const RELOC_ENUM rel[NUM_FLAG_CODE];
   } gotrel[] = {
-    { "PLT",      { BFD_RELOC_386_PLT32,  0, BFD_RELOC_X86_64_PLT32    } },
-    { "GOTOFF",   { BFD_RELOC_386_GOTOFF, 0, 0                         } },
-    { "GOTPCREL", { 0,                    0, BFD_RELOC_X86_64_GOTPCREL } },
-    { "GOT",      { BFD_RELOC_386_GOT32,  0, BFD_RELOC_X86_64_GOT32    } }
+    { "PLT",      { BFD_RELOC_386_PLT32,      0, BFD_RELOC_X86_64_PLT32    } },
+    { "GOTOFF",   { BFD_RELOC_386_GOTOFF,     0, 0                         } },
+    { "GOTPCREL", { 0,                        0, BFD_RELOC_X86_64_GOTPCREL } },
+    { "TLSGD",    { BFD_RELOC_386_TLS_GD,     0, BFD_RELOC_X86_64_TLSGD    } },
+    { "TLSLDM",   { BFD_RELOC_386_TLS_LDM,    0, 0                         } },
+    { "TLSLD",    { 0,                        0, BFD_RELOC_X86_64_TLSLD    } },
+    { "GOTTPOFF", { BFD_RELOC_386_TLS_IE_32,  0, BFD_RELOC_X86_64_GOTTPOFF } },
+    { "TPOFF",    { BFD_RELOC_386_TLS_LE_32,  0, BFD_RELOC_X86_64_TPOFF32  } },
+    { "NTPOFF",   { BFD_RELOC_386_TLS_LE,     0, 0                         } },
+    { "DTPOFF",   { BFD_RELOC_386_TLS_LDO_32, 0, BFD_RELOC_X86_64_DTPOFF32 } },
+    { "GOTNTPOFF",{ BFD_RELOC_386_TLS_GOTIE,  0, 0                         } },
+    { "INDNTPOFF",{ BFD_RELOC_386_TLS_IE,     0, 0                         } },
+    { "GOT",      { BFD_RELOC_386_GOT32,      0, BFD_RELOC_X86_64_GOT32    } }
   };
   char *cp;
   unsigned int j;
@@ -3181,27 +3517,41 @@ lex_got (reloc, adjust)
       int len;
 
       len = strlen (gotrel[j].str);
-      if (strncmp (cp + 1, gotrel[j].str, len) == 0)
+      if (strncasecmp (cp + 1, gotrel[j].str, len) == 0)
        {
          if (gotrel[j].rel[(unsigned int) flag_code] != 0)
            {
-             int first;
-             char *tmpbuf;
+             int first, second;
+             char *tmpbuf, *past_reloc;
 
              *reloc = gotrel[j].rel[(unsigned int) flag_code];
+             if (adjust)
+               *adjust = len;
 
              if (GOT_symbol == NULL)
                GOT_symbol = symbol_find_or_make (GLOBAL_OFFSET_TABLE_NAME);
 
              /* Replace the relocation token with ' ', so that
                 errors like foo@GOTOFF1 will be detected.  */
+
+             /* The length of the first part of our input line.  */
              first = cp - input_line_pointer;
-             tmpbuf = xmalloc (strlen (input_line_pointer));
+
+             /* The second part goes from after the reloc token until
+                (and including) an end_of_line char.  Don't use strlen
+                here as the end_of_line char may not be a NUL.  */
+             past_reloc = cp + 1 + len;
+             for (cp = past_reloc; !is_end_of_line[(unsigned char) *cp++]; )
+               ;
+             second = cp - past_reloc;
+
+             /* Allocate and copy string.  The trailing NUL shouldn't
+                be necessary, but be safe.  */
+             tmpbuf = xmalloc (first + second + 2);
              memcpy (tmpbuf, input_line_pointer, first);
              tmpbuf[first] = ' ';
-             strcpy (tmpbuf + first + 1, cp + 1 + len);
-             if (adjust)
-               *adjust = len;
+             memcpy (tmpbuf + first + 1, past_reloc, second);
+             tmpbuf[first + second + 1] = '\0';
              return tmpbuf;
            }
 
@@ -3336,6 +3686,7 @@ i386_immediate (imm_start)
 #ifdef BFD_ASSEMBLER
           && OUTPUT_FLAVOR == bfd_target_aout_flavour
 #endif
+          && exp_seg != absolute_section
           && exp_seg != text_section
           && exp_seg != data_section
           && exp_seg != bss_section
@@ -3397,7 +3748,7 @@ i386_scale (scale)
       input_line_pointer = save;
       return NULL;
     }
-  if (i.log2_scale_factor != 0 && ! i.index_reg)
+  if (i.log2_scale_factor != 0 && i.index_reg == 0)
     {
       as_warn (_("scale factor of %d without an index register"),
               1 << i.log2_scale_factor);
@@ -3417,7 +3768,7 @@ i386_displacement (disp_start, disp_end)
      char *disp_start;
      char *disp_end;
 {
-  register expressionS *exp;
+  expressionS *exp;
   segT exp_seg = 0;
   char *save_input_line_pointer;
 #ifndef LEX_AT
@@ -3425,10 +3776,13 @@ i386_displacement (disp_start, disp_end)
 #endif
   int bigdisp = Disp32;
 
-  if ((flag_code == CODE_16BIT) ^ (i.prefix[ADDR_PREFIX] != 0))
-    bigdisp = Disp16;
   if (flag_code == CODE_64BIT)
-    bigdisp = Disp64;
+    {
+      if (i.prefix[ADDR_PREFIX] == 0)
+       bigdisp = Disp64;
+    }
+  else if ((flag_code == CODE_16BIT) ^ (i.prefix[ADDR_PREFIX] != 0))
+    bigdisp = Disp16;
   i.types[this_operand] |= bigdisp;
 
   exp = &disp_expressions[i.disp_operands];
@@ -3525,9 +3879,9 @@ i386_displacement (disp_start, disp_end)
       exp->X_op = O_subtract;
       exp->X_op_symbol = GOT_symbol;
       if (i.reloc[this_operand] == BFD_RELOC_X86_64_GOTPCREL)
-        i.reloc[this_operand] = BFD_RELOC_32_PCREL;
+       i.reloc[this_operand] = BFD_RELOC_32_PCREL;
       else
-        i.reloc[this_operand] = BFD_RELOC_32;
+       i.reloc[this_operand] = BFD_RELOC_32;
     }
 #endif
 
@@ -3547,10 +3901,15 @@ i386_displacement (disp_start, disp_end)
 #ifdef BFD_ASSEMBLER
       && OUTPUT_FLAVOR == bfd_target_aout_flavour
 #endif
+      && exp_seg != absolute_section
       && exp_seg != text_section
       && exp_seg != data_section
       && exp_seg != bss_section
-      && exp_seg != undefined_section)
+      && exp_seg != undefined_section
+#ifdef BFD_ASSEMBLER
+      && !bfd_is_com_section (exp_seg)
+#endif
+      )
     {
 #ifdef BFD_ASSEMBLER
       as_bad (_("unimplemented segment %s in operand"), exp_seg->name);
@@ -3583,15 +3942,28 @@ i386_index_check (operand_string)
   ok = 1;
   if (flag_code == CODE_64BIT)
     {
-      /* 64bit checks.  */
-      if ((i.base_reg
-          && ((i.base_reg->reg_type & Reg64) == 0)
-              && (i.base_reg->reg_type != BaseIndex
-                  || i.index_reg))
-         || (i.index_reg
-             && ((i.index_reg->reg_type & (Reg64|BaseIndex))
-                 != (Reg64|BaseIndex))))
-       ok = 0;
+      if (i.prefix[ADDR_PREFIX] == 0)
+       {
+         /* 64bit checks.  */
+         if ((i.base_reg
+              && ((i.base_reg->reg_type & Reg64) == 0)
+                  && (i.base_reg->reg_type != BaseIndex
+                      || i.index_reg))
+             || (i.index_reg
+                 && ((i.index_reg->reg_type & (Reg64 | BaseIndex))
+                     != (Reg64 | BaseIndex))))
+           ok = 0;
+       }
+      else
+       {
+         /* 32bit checks.  */
+         if ((i.base_reg
+              && (i.base_reg->reg_type & (Reg32 | RegRex)) != Reg32)
+             || (i.index_reg
+                 && ((i.index_reg->reg_type & (Reg32 | BaseIndex | RegRex))
+                     != (Reg32 | BaseIndex))))
+           ok = 0;
+       }
     }
   else
     {
@@ -3599,15 +3971,15 @@ i386_index_check (operand_string)
        {
          /* 16bit checks.  */
          if ((i.base_reg
-              && ((i.base_reg->reg_type & (Reg16|BaseIndex|RegRex))
-                  != (Reg16|BaseIndex)))
+              && ((i.base_reg->reg_type & (Reg16 | BaseIndex | RegRex))
+                  != (Reg16 | BaseIndex)))
              || (i.index_reg
-                 && (((i.index_reg->reg_type & (Reg16|BaseIndex))
-                      != (Reg16|BaseIndex))
-                     || ! (i.base_reg
-                           && i.base_reg->reg_num < 6
-                           && i.index_reg->reg_num >= 6
-                           && i.log2_scale_factor == 0))))
+                 && (((i.index_reg->reg_type & (Reg16 | BaseIndex))
+                      != (Reg16 | BaseIndex))
+                     || !(i.base_reg
+                          && i.base_reg->reg_num < 6
+                          && i.index_reg->reg_num >= 6
+                          && i.log2_scale_factor == 0))))
            ok = 0;
        }
       else
@@ -3616,8 +3988,8 @@ i386_index_check (operand_string)
          if ((i.base_reg
               && (i.base_reg->reg_type & (Reg32 | RegRex)) != Reg32)
              || (i.index_reg
-                 && ((i.index_reg->reg_type & (Reg32|BaseIndex|RegRex))
-                     != (Reg32|BaseIndex))))
+                 && ((i.index_reg->reg_type & (Reg32 | BaseIndex | RegRex))
+                     != (Reg32 | BaseIndex))))
            ok = 0;
        }
     }
@@ -3634,8 +4006,8 @@ i386_index_check (operand_string)
             FIXME.  There doesn't seem to be any real need for separate
             Disp16 and Disp32 flags.  The same goes for Imm16 and Imm32.
             Removing them would probably clean up the code quite a lot.  */
-         if (i.types[this_operand] & (Disp16|Disp32))
-            i.types[this_operand] ^= (Disp16|Disp32);
+         if (i.types[this_operand] & (Disp16 | Disp32))
+            i.types[this_operand] ^= (Disp16 | Disp32);
          fudged = 1;
          goto tryprefix;
        }
@@ -3954,8 +4326,8 @@ i386_operand (operand_string)
 
 int
 md_estimate_size_before_relax (fragP, segment)
-     register fragS *fragP;
-     register segT segment;
+     fragS *fragP;
+     segT segment;
 {
   /* We've already got fragP->fr_subtype right;  all we have to do is
      check for un-relaxable symbols.  On an ELF system, we can't relax
@@ -3963,8 +4335,9 @@ md_estimate_size_before_relax (fragP, segment)
      shared library.  */
   if (S_GET_SEGMENT (fragP->fr_symbol) != segment
 #if defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF)
-      || S_IS_EXTERNAL (fragP->fr_symbol)
-      || S_IS_WEAK (fragP->fr_symbol)
+      || (OUTPUT_FLAVOR == bfd_target_elf_flavour
+         && (S_IS_EXTERNAL (fragP->fr_symbol)
+             || S_IS_WEAK (fragP->fr_symbol)))
 #endif
       )
     {
@@ -3998,10 +4371,8 @@ md_estimate_size_before_relax (fragP, segment)
          break;
 
        case COND_JUMP86:
-         if (no_cond_jump_promotion)
-           goto relax_guess;
-
-         if (size == 2)
+         if (size == 2
+             && (!no_cond_jump_promotion || fragP->fr_var != NO_RELOC))
            {
              /* Negate the condition, and branch past an
                 unconditional jump.  */
@@ -4021,8 +4392,15 @@ md_estimate_size_before_relax (fragP, segment)
          /* Fall through.  */
 
        case COND_JUMP:
-         if (no_cond_jump_promotion)
-           goto relax_guess;
+         if (no_cond_jump_promotion && fragP->fr_var == NO_RELOC)
+           {
+             fragP->fr_fix += 1;
+             fix_new (fragP, old_fr_fix, 1,
+                      fragP->fr_symbol,
+                      fragP->fr_offset, 1,
+                      BFD_RELOC_8_PCREL);
+             break;
+           }
 
          /* This changes the byte-displacement jump 0x7N
             to the (d)word-displacement jump 0x0f,0x8N.  */
@@ -4044,7 +4422,6 @@ md_estimate_size_before_relax (fragP, segment)
       return fragP->fr_fix - old_fr_fix;
     }
 
- relax_guess:
   /* Guess size depending on current relax state.  Initially the relax
      state will correspond to a short jump and we return 1, because
      the variable part of the frag (the branch offset) is one byte
@@ -4068,16 +4445,16 @@ void
 md_convert_frag (headers, sec, fragP)
      object_headers *headers ATTRIBUTE_UNUSED;
      segT sec ATTRIBUTE_UNUSED;
-     register fragS *fragP;
+     fragS *fragP;
 #else
 void
 md_convert_frag (abfd, sec, fragP)
      bfd *abfd ATTRIBUTE_UNUSED;
      segT sec ATTRIBUTE_UNUSED;
-     register fragS *fragP;
+     fragS *fragP;
 #endif
 {
-  register unsigned char *opcode;
+  unsigned char *opcode;
   unsigned char *where_to_put_displacement = NULL;
   offsetT target_address;
   offsetT opcode_address;
@@ -4088,15 +4465,6 @@ md_convert_frag (abfd, sec, fragP)
 
   /* Address we want to reach in file space.  */
   target_address = S_GET_VALUE (fragP->fr_symbol) + fragP->fr_offset;
-#ifdef BFD_ASSEMBLER
-  /* Not needed otherwise?  */
-  {
-    /* Local symbols which have already been resolved have a NULL frag.  */
-    fragS *sym_frag = symbol_get_frag (fragP->fr_symbol);
-    if (sym_frag)
-      target_address += sym_frag->fr_address;
-  }
-#endif
 
   /* Address opcode resides at in file space.  */
   opcode_address = fragP->fr_address + fragP->fr_fix;
@@ -4211,19 +4579,17 @@ md_create_long_jump (ptr, from_addr, to_addr, frag, to_symbol)
    the same (little-endian) format, so we don't need to care about which
    we are handling.  */
 
-int
-md_apply_fix3 (fixP, valp, seg)
+void
+md_apply_fix3 (fixP, valP, seg)
      /* The fix we're to put in.  */
      fixS *fixP;
-
      /* Pointer to the value of the bits.  */
-     valueT *valp;
-
+     valueT *valP;
      /* Segment fix is from.  */
      segT seg ATTRIBUTE_UNUSED;
 {
-  register char *p = fixP->fx_where + fixP->fx_frag->fr_literal;
-  valueT value = *valp;
+  char *p = fixP->fx_where + fixP->fx_frag->fr_literal;
+  valueT value = *valP;
 
 #if defined (BFD_ASSEMBLER) && !defined (TE_Mach)
   if (fixP->fx_pcrel)
@@ -4245,15 +4611,16 @@ md_apply_fix3 (fixP, valp, seg)
        }
     }
 
-  /* This is a hack.  There should be a better way to handle this.
-     This covers for the fact that bfd_install_relocation will
-     subtract the current location (for partial_inplace, PC relative
-     relocations); see more below.  */
-  if ((fixP->fx_r_type == BFD_RELOC_32_PCREL
-       || fixP->fx_r_type == BFD_RELOC_16_PCREL
-       || fixP->fx_r_type == BFD_RELOC_8_PCREL)
-      && fixP->fx_addsy && !use_rela_relocations)
+  if (fixP->fx_addsy != NULL
+      && (fixP->fx_r_type == BFD_RELOC_32_PCREL
+         || fixP->fx_r_type == BFD_RELOC_16_PCREL
+         || fixP->fx_r_type == BFD_RELOC_8_PCREL)
+      && !use_rela_relocations)
     {
+      /* This is a hack.  There should be a better way to handle this.
+        This covers for the fact that bfd_install_relocation will
+        subtract the current location (for partial_inplace, PC relative
+        relocations); see more below.  */
 #ifndef OBJ_AOUT
       if (OUTPUT_FLAVOR == bfd_target_elf_flavour
 #ifdef TE_PE
@@ -4265,19 +4632,16 @@ md_apply_fix3 (fixP, valp, seg)
 #if defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF)
       if (OUTPUT_FLAVOR == bfd_target_elf_flavour)
        {
-         segT fseg = S_GET_SEGMENT (fixP->fx_addsy);
+         segT sym_seg = S_GET_SEGMENT (fixP->fx_addsy);
 
-         if ((fseg == seg
+         if ((sym_seg == seg
               || (symbol_section_p (fixP->fx_addsy)
-                  && fseg != absolute_section))
-             && ! S_IS_EXTERNAL (fixP->fx_addsy)
-             && ! S_IS_WEAK (fixP->fx_addsy)
-             && S_IS_DEFINED (fixP->fx_addsy)
-             && ! S_IS_COMMON (fixP->fx_addsy))
+                  && sym_seg != absolute_section))
+             && !S_FORCE_RELOC (fixP->fx_addsy))
            {
              /* Yes, we add the values in twice.  This is because
-                bfd_perform_relocation subtracts them out again.  I think
-                bfd_perform_relocation is broken, but I don't dare change
+                bfd_install_relocation subtracts them out again.  I think
+                bfd_install_relocation is broken, but I don't dare change
                 it.  FIXME.  */
              value += fixP->fx_where + fixP->fx_frag->fr_address;
            }
@@ -4304,83 +4668,45 @@ md_apply_fix3 (fixP, valp, seg)
           runtime we merely add the offset to the actual PLT entry.  */
        value = -4;
        break;
-      case BFD_RELOC_386_GOTPC:
-
-/*   This is tough to explain.  We end up with this one if we have
- * operands that look like "_GLOBAL_OFFSET_TABLE_+[.-.L284]".  The goal
- * here is to obtain the absolute address of the GOT, and it is strongly
- * preferable from a performance point of view to avoid using a runtime
- * relocation for this.  The actual sequence of instructions often look
- * something like:
- *
- *     call    .L66
- * .L66:
- *     popl    %ebx
- *     addl    $_GLOBAL_OFFSET_TABLE_+[.-.L66],%ebx
- *
- *   The call and pop essentially return the absolute address of
- * the label .L66 and store it in %ebx.  The linker itself will
- * ultimately change the first operand of the addl so that %ebx points to
- * the GOT, but to keep things simple, the .o file must have this operand
- * set so that it generates not the absolute address of .L66, but the
- * absolute address of itself.  This allows the linker itself simply
- * treat a GOTPC relocation as asking for a pcrel offset to the GOT to be
- * added in, and the addend of the relocation is stored in the operand
- * field for the instruction itself.
- *
- *   Our job here is to fix the operand so that it would add the correct
- * offset so that %ebx would point to itself.  The thing that is tricky is
- * that .-.L66 will point to the beginning of the instruction, so we need
- * to further modify the operand so that it will point to itself.
- * There are other cases where you have something like:
- *
- *     .long   $_GLOBAL_OFFSET_TABLE_+[.-.L66]
- *
- * and here no correction would be required.  Internally in the assembler
- * we treat operands of this form as not being pcrel since the '.' is
- * explicitly mentioned, and I wonder whether it would simplify matters
- * to do it this way.  Who knows.  In earlier versions of the PIC patches,
- * the pcrel_adjust field was used to store the correction, but since the
- * expression is not pcrel, I felt it would be confusing to do it this
- * way.  */
-
-       value -= 1;
-       break;
+
       case BFD_RELOC_386_GOT32:
+      case BFD_RELOC_386_TLS_GD:
+      case BFD_RELOC_386_TLS_LDM:
+      case BFD_RELOC_386_TLS_IE_32:
+      case BFD_RELOC_386_TLS_IE:
+      case BFD_RELOC_386_TLS_GOTIE:
       case BFD_RELOC_X86_64_GOT32:
+      case BFD_RELOC_X86_64_TLSGD:
+      case BFD_RELOC_X86_64_TLSLD:
+      case BFD_RELOC_X86_64_GOTTPOFF:
        value = 0; /* Fully resolved at runtime.  No addend.  */
        break;
-      case BFD_RELOC_386_GOTOFF:
-      case BFD_RELOC_X86_64_GOTPCREL:
-       break;
 
       case BFD_RELOC_VTABLE_INHERIT:
       case BFD_RELOC_VTABLE_ENTRY:
        fixP->fx_done = 0;
-       return 1;
+       return;
 
       default:
        break;
       }
 #endif /* defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF)  */
-  *valp = value;
+  *valP = value;
 #endif /* defined (BFD_ASSEMBLER) && !defined (TE_Mach)  */
 
-#ifndef BFD_ASSEMBLER
-  md_number_to_chars (p, value, fixP->fx_size);
-#else
   /* Are we finished with this relocation now?  */
-  if (fixP->fx_addsy == 0 && fixP->fx_pcrel == 0)
+  if (fixP->fx_addsy == NULL)
     fixP->fx_done = 1;
+#ifdef BFD_ASSEMBLER
   else if (use_rela_relocations)
     {
       fixP->fx_no_overflow = 1;
+      /* Remember value for tc_gen_reloc.  */
+      fixP->fx_addnumber = value;
       value = 0;
     }
-  md_number_to_chars (p, value, fixP->fx_size);
 #endif
-
-  return 1;
+  md_number_to_chars (p, value, fixP->fx_size);
 }
 \f
 #define MAX_LITTLENUMS 6
@@ -4443,7 +4769,7 @@ static char *
 output_invalid (c)
      int c;
 {
-  if (isprint (c))
+  if (ISPRINT (c))
     sprintf (output_invalid_buf, "'%c'", c);
   else
     sprintf (output_invalid_buf, "(0x%x)", (unsigned) c);
@@ -4514,6 +4840,13 @@ parse_register (reg_string, end_op)
        }
     }
 
+  if (r != NULL
+      && (r->reg_flags & (RegRex64 | RegRex)) != 0
+      && flag_code != CODE_64BIT)
+    {
+      return (const reg_entry *) NULL;
+    }
+
   return r;
 }
 \f
@@ -4562,7 +4895,7 @@ md_parse_option (c, arg)
 
     case 's':
       /* -s: On i386 Solaris, this tells the native assembler to use
-         .stab instead of .stab.excl.  We always use .stab anyhow.  */
+        .stab instead of .stab.excl.  We always use .stab anyhow.  */
       break;
 
     case OPTION_64:
@@ -4640,7 +4973,7 @@ i386_target_format ()
       {
        if (flag_code == CODE_64BIT)
          use_rela_relocations = 1;
-       return flag_code == CODE_64BIT ? "elf64-x86-64" : "elf32-i386";
+       return flag_code == CODE_64BIT ? "elf64-x86-64" : ELF_TARGET_FORMAT;
       }
 #endif
     default:
@@ -4650,6 +4983,48 @@ i386_target_format ()
 }
 
 #endif /* OBJ_MAYBE_ more than one  */
+
+#if (defined (OBJ_ELF) || defined (OBJ_MAYBE_ELF))
+void i386_elf_emit_arch_note ()
+{
+  if (OUTPUT_FLAVOR == bfd_target_elf_flavour
+      && cpu_arch_name != NULL)
+    {
+      char *p;
+      asection *seg = now_seg;
+      subsegT subseg = now_subseg;
+      Elf_Internal_Note i_note;
+      Elf_External_Note e_note;
+      asection *note_secp;
+      int len;
+
+      /* Create the .note section.  */
+      note_secp = subseg_new (".note", 0);
+      bfd_set_section_flags (stdoutput,
+                            note_secp,
+                            SEC_HAS_CONTENTS | SEC_READONLY);
+
+      /* Process the arch string.  */
+      len = strlen (cpu_arch_name);
+
+      i_note.namesz = len + 1;
+      i_note.descsz = 0;
+      i_note.type = NT_ARCH;
+      p = frag_more (sizeof (e_note.namesz));
+      md_number_to_chars (p, (valueT) i_note.namesz, sizeof (e_note.namesz));
+      p = frag_more (sizeof (e_note.descsz));
+      md_number_to_chars (p, (valueT) i_note.descsz, sizeof (e_note.descsz));
+      p = frag_more (sizeof (e_note.type));
+      md_number_to_chars (p, (valueT) i_note.type, sizeof (e_note.type));
+      p = frag_more (len + 1);
+      strcpy (p, cpu_arch_name);
+
+      frag_align (2, 0, 0);
+
+      subseg_set (seg, subseg);
+    }
+}
+#endif
 #endif /* BFD_ASSEMBLER  */
 \f
 symbolS *
@@ -4717,7 +5092,7 @@ static void
 s_bss (ignore)
      int ignore ATTRIBUTE_UNUSED;
 {
-  register int temp;
+  int temp;
 
   temp = get_absolute_expression ();
   subseg_set (bss_section, (subsegT) temp);
@@ -4751,6 +5126,17 @@ i386_validate_fix (fixp)
     }
 }
 
+boolean
+i386_force_relocation (fixp)
+     fixS *fixp;
+{
+  if (fixp->fx_r_type == BFD_RELOC_VTABLE_INHERIT
+      || fixp->fx_r_type == BFD_RELOC_VTABLE_ENTRY)
+    return 1;
+
+  return S_FORCE_RELOC (fixp->fx_addsy);
+}
+
 arelent *
 tc_gen_reloc (section, fixp)
      asection *section ATTRIBUTE_UNUSED;
@@ -4768,7 +5154,20 @@ tc_gen_reloc (section, fixp)
     case BFD_RELOC_386_GOT32:
     case BFD_RELOC_386_GOTOFF:
     case BFD_RELOC_386_GOTPC:
+    case BFD_RELOC_386_TLS_GD:
+    case BFD_RELOC_386_TLS_LDM:
+    case BFD_RELOC_386_TLS_LDO_32:
+    case BFD_RELOC_386_TLS_IE_32:
+    case BFD_RELOC_386_TLS_IE:
+    case BFD_RELOC_386_TLS_GOTIE:
+    case BFD_RELOC_386_TLS_LE_32:
+    case BFD_RELOC_386_TLS_LE:
     case BFD_RELOC_X86_64_32S:
+    case BFD_RELOC_X86_64_TLSGD:
+    case BFD_RELOC_X86_64_TLSLD:
+    case BFD_RELOC_X86_64_DTPOFF32:
+    case BFD_RELOC_X86_64_GOTTPOFF:
+    case BFD_RELOC_X86_64_TPOFF32:
     case BFD_RELOC_RVA:
     case BFD_RELOC_VTABLE_ENTRY:
     case BFD_RELOC_VTABLE_INHERIT:
@@ -4803,7 +5202,9 @@ tc_gen_reloc (section, fixp)
            case 1: code = BFD_RELOC_8;  break;
            case 2: code = BFD_RELOC_16; break;
            case 4: code = BFD_RELOC_32; break;
+#ifdef BFD64
            case 8: code = BFD_RELOC_64; break;
+#endif
            }
        }
       break;
@@ -4831,17 +5232,31 @@ tc_gen_reloc (section, fixp)
       if (fixp->fx_r_type == BFD_RELOC_VTABLE_ENTRY)
        rel->address = fixp->fx_offset;
 
-      if (fixp->fx_pcrel)
-       rel->addend = fixp->fx_addnumber;
-      else
-       rel->addend = 0;
+      rel->addend = 0;
     }
   /* Use the rela in 64bit mode.  */
   else
     {
-      rel->addend = fixp->fx_offset;
-      if (fixp->fx_pcrel)
-       rel->addend -= fixp->fx_size;
+      if (!fixp->fx_pcrel)
+       rel->addend = fixp->fx_offset;
+      else
+       switch (code)
+         {
+         case BFD_RELOC_X86_64_PLT32:
+         case BFD_RELOC_X86_64_GOT32:
+         case BFD_RELOC_X86_64_GOTPCREL:
+         case BFD_RELOC_X86_64_TLSGD:
+         case BFD_RELOC_X86_64_TLSLD:
+         case BFD_RELOC_X86_64_GOTTPOFF:
+           rel->addend = fixp->fx_offset - fixp->fx_size;
+           break;
+         default:
+           rel->addend = (section->vma
+                          - fixp->fx_size
+                          + fixp->fx_addnumber
+                          + md_pcrel_from (fixp));
+           break;
+         }
     }
 
   rel->howto = bfd_reloc_type_lookup (stdoutput, code);
@@ -4858,7 +5273,7 @@ tc_gen_reloc (section, fixp)
   return rel;
 }
 
-#else /* ! BFD_ASSEMBLER  */
+#else /* !BFD_ASSEMBLER  */
 
 #if (defined(OBJ_AOUT) | defined(OBJ_BOUT))
 void
@@ -4924,7 +5339,7 @@ tc_coff_sizemachdep (frag)
 
 #endif /* I386COFF  */
 
-#endif /* ! BFD_ASSEMBLER  */
+#endif /* !BFD_ASSEMBLER  */
 \f
 /* Parse operands using Intel syntax. This implements a recursive descent
    parser based on the BNF grammar published in Appendix B of the MASM 6.1
This page took 0.081751 seconds and 4 git commands to generate.