glue_x86.h

Go to the documentation of this file.

#ifndef _NSEEL_GLUE_X86_H_
#define _NSEEL_GLUE_X86_H_
 
#define GLUE_MAX_FPSTACK_SIZE 8
 
// endOfInstruction is end of jump with relative offset, offset is offset from end of instruction to jump to
#define GLUE_JMP_SET_OFFSET(endOfInstruction,offset) (((int *)(endOfInstruction))[-1] = (offset))
 
static const unsigned char GLUE_JMP_NC[] = { 0xE9, 0,0,0,0, }; // jmp<offset>
static const unsigned char GLUE_JMP_IF_P1_Z[] = {0x85, 0xC0, 0x0F, 0x84, 0,0,0,0 }; // test eax, eax, jz
static const unsigned char GLUE_JMP_IF_P1_NZ[] = {0x85, 0xC0, 0x0F, 0x85, 0,0,0,0 }; // test eax, eax, jnz
 
#define GLUE_FUNC_ENTER_SIZE 0
#define GLUE_FUNC_LEAVE_SIZE 0
const static unsigned int GLUE_FUNC_ENTER[1];
const static unsigned int GLUE_FUNC_LEAVE[1];
 
  // x86
  // stack is 16 byte aligned
  // when pushing values to stack, alignment pushed first, then value (value is at the lower address)
  // when pushing pointers to stack, alignment pushed first, then pointer (pointer is at the lower address)
 
  static const unsigned char GLUE_PUSH_P1PTR_AS_VALUE[] = 
  { 
    0x83, 0xEC, 8, /* sub esp, 8 */   
    0xff, 0x70, 0x4, /* push dword [eax+4] */ 
    0xff, 0x30, /* push dword [eax] */
  };
 
  static int GLUE_POP_VALUE_TO_ADDR(unsigned char *buf, void *destptr)
  {    
    if (buf)
    {
      *buf++ = 0xB8; *(void **) buf = destptr; buf+=4; // mov eax, directvalue
           
      *buf++ = 0x8f; *buf++ = 0x00; // pop dword [eax]
      *buf++ = 0x8f; *buf++ = 0x40; *buf++ = 4; // pop dword [eax+4]
      
      *buf++ = 0x59; // pop ecx (alignment)
      *buf++ = 0x59; // pop ecx (alignment)
    }
    
    return 12;
  }
 
  static int GLUE_COPY_VALUE_AT_P1_TO_PTR(unsigned char *buf, void *destptr)
  {    
    if (buf)
    {
      *buf++ = 0x8B; *buf++ = 0x38; // mov edi, [eax]
      *buf++ = 0x8B; *buf++ = 0x48; *buf++ = 0x04; // mov ecx, [eax+4]
      
      
      *buf++ = 0xB8; *(void **) buf = destptr; buf+=4; // mov eax, directvalue
      *buf++ = 0x89; *buf++ = 0x38; // mov [eax], edi
      *buf++ = 0x89; *buf++ = 0x48; *buf++ = 0x04; // mov [eax+4], ecx
    }
    
    return 2 + 3 + 5 + 2 + 3;
  }
 
  static int GLUE_POP_FPSTACK_TO_PTR(unsigned char *buf, void *destptr)
  {
    if (buf)
    {
      *buf++ = 0xB8; *(void **) buf = destptr; buf+=4; // mov eax, directvalue
      *buf++ = 0xDD; *buf++ = 0x18;  // fstp qword [eax]
    }
    return 1+4+2;
  }
 
 
  #define GLUE_MOV_PX_DIRECTVALUE_SIZE 5
  #define GLUE_MOV_PX_DIRECTVALUE_TOSTACK_SIZE 6 // length when wv == -1
 
  static void GLUE_MOV_PX_DIRECTVALUE_GEN(void *b, INT_PTR v, int wv) 
  {   
    if (wv==-1)
    {
      const static unsigned char t[2] = {0xDD, 0x05};
      memcpy(b,t,2);
      b= ((unsigned char *)b)+2;
    }
    else
    {
      const static unsigned char tab[3] = {
        0xB8 /* mov eax, dv*/, 
        0xBF /* mov edi, dv */ , 
        0xB9 /* mov ecx, dv */ 
      };
      *((unsigned char *)b) = tab[wv];  // mov eax, dv
      b= ((unsigned char *)b)+1;
    }
    *(INT_PTR *)b = v; 
  }
  const static unsigned char  GLUE_PUSH_P1[4]={0x83, 0xEC, 12,   0x50}; // sub esp, 12, push eax
 
  #define GLUE_STORE_P1_TO_STACK_AT_OFFS_SIZE(x) 7
  static void GLUE_STORE_P1_TO_STACK_AT_OFFS(void *b, int offs)
  {
    ((unsigned char *)b)[0] = 0x89; // mov [esp+offs], eax
    ((unsigned char *)b)[1] = 0x84;
    ((unsigned char *)b)[2] = 0x24;
    *(int *)((unsigned char *)b+3) = offs;
  }
 
  #define GLUE_MOVE_PX_STACKPTR_SIZE 2
  static void GLUE_MOVE_PX_STACKPTR_GEN(void *b, int wv)
  {
    static const unsigned char tab[3][GLUE_MOVE_PX_STACKPTR_SIZE]=
    {
      { 0x89, 0xe0 }, // mov eax, esp
      { 0x89, 0xe7 }, // mov edi, esp
      { 0x89, 0xe1 }, // mov ecx, esp
    };    
    memcpy(b,tab[wv],GLUE_MOVE_PX_STACKPTR_SIZE);
  }
 
  #define GLUE_MOVE_STACK_SIZE 6
  static void GLUE_MOVE_STACK(void *b, int amt)
  {
    ((unsigned char *)b)[0] = 0x81;
    if (amt <0)
    {
      ((unsigned char *)b)[1] = 0xEC;
      *(int *)((char*)b+2) = -amt; // sub esp, -amt
    }
    else
    {
      ((unsigned char *)b)[1] = 0xc4;
      *(int *)((char*)b+2) = amt; // add esp, amt
    }
  }
 
 
  #define GLUE_POP_PX_SIZE 4
  static void GLUE_POP_PX(void *b, int wv)
  {
    static const unsigned char tab[3][GLUE_POP_PX_SIZE]=
    {
      {0x58,/*pop eax*/  0x83, 0xC4, 12 /* add esp, 12*/},
      {0x5F,/*pop edi*/  0x83, 0xC4, 12}, 
      {0x59,/*pop ecx*/  0x83, 0xC4, 12}, 
    };    
    memcpy(b,tab[wv],GLUE_POP_PX_SIZE);
  }
 
  #define GLUE_SET_PX_FROM_P1_SIZE 2
  static void GLUE_SET_PX_FROM_P1(void *b, int wv)
  {
    static const unsigned char tab[3][GLUE_SET_PX_FROM_P1_SIZE]={
      {0x90,0x90}, // should never be used! (nopnop)
      {0x89,0xC7}, // mov edi, eax
      {0x89,0xC1}, // mov ecx, eax
    };
    memcpy(b,tab[wv],GLUE_SET_PX_FROM_P1_SIZE);
  }
 
  #define GLUE_POP_FPSTACK_SIZE 2
  static const unsigned char GLUE_POP_FPSTACK[2] = { 0xDD, 0xD8 }; // fstp st0
 
  static const unsigned char GLUE_POP_FPSTACK_TOSTACK[] = {
    0x83, 0xEC, 16, // sub esp, 16
    0xDD, 0x1C, 0x24 // fstp qword (%esp)  
  };
 
  static const unsigned char GLUE_POP_STACK_TO_FPSTACK[] = {
    0xDD, 0x04, 0x24, // fld qword (%esp)
    0x83, 0xC4, 16 //  add esp, 16
  };
 
  static const unsigned char GLUE_POP_FPSTACK_TO_WTP[] = { 
      0xDD, 0x1E, /* fstp qword [esi] */
      0x83, 0xC6, 8, /* add esi, 8 */ 
  };
 
  #define GLUE_SET_PX_FROM_WTP_SIZE 2
  static void GLUE_SET_PX_FROM_WTP(void *b, int wv)
  {
    static const unsigned char tab[3][GLUE_SET_PX_FROM_WTP_SIZE]={
      {0x89,0xF0}, // mov eax, esi
      {0x89,0xF7}, // mov edi, esi
      {0x89,0xF1}, // mov ecx, esi
    };
    memcpy(b,tab[wv],GLUE_SET_PX_FROM_WTP_SIZE);
  }
 
  #define GLUE_PUSH_VAL_AT_PX_TO_FPSTACK_SIZE 2
  static void GLUE_PUSH_VAL_AT_PX_TO_FPSTACK(void *b, int wv)
  {
    static const unsigned char tab[3][GLUE_PUSH_VAL_AT_PX_TO_FPSTACK_SIZE]={
      {0xDD,0x00}, // fld qword [eax]
      {0xDD,0x07}, // fld qword [edi]
      {0xDD,0x01}, // fld qword [ecx]
    };
    memcpy(b,tab[wv],GLUE_PUSH_VAL_AT_PX_TO_FPSTACK_SIZE);
  }
 
#define GLUE_POP_FPSTACK_TO_WTP_TO_PX_SIZE (GLUE_SET_PX_FROM_WTP_SIZE + sizeof(GLUE_POP_FPSTACK_TO_WTP))
static void GLUE_POP_FPSTACK_TO_WTP_TO_PX(unsigned char *buf, int wv)
{
  GLUE_SET_PX_FROM_WTP(buf,wv);
  memcpy(buf + GLUE_SET_PX_FROM_WTP_SIZE,GLUE_POP_FPSTACK_TO_WTP,sizeof(GLUE_POP_FPSTACK_TO_WTP));
};
 
 
const static unsigned char  GLUE_RET=0xC3;
 
static int GLUE_RESET_WTP(unsigned char *out, void *ptr)
{
  if (out)
  {
    *out++ = 0xBE; // mov esi, constant
    memcpy(out,&ptr,sizeof(void *));
    out+=sizeof(void *);
  }
  return 1+sizeof(void *);
}
 
 
#ifdef _MSC_VER
#pragma warning(push)
#pragma warning(disable: 4731)
#endif
 
#define GLUE_TABPTR_IGNORED
#define GLUE_CALL_CODE(bp, cp, rt) do { \
  if (h->compile_flags&NSEEL_CODE_COMPILE_FLAG_NOFPSTATE) eel_callcode32_fast(cp, rt); \
  else eel_callcode32(cp, rt);\
  } while(0)
 
static void eel_callcode32(INT_PTR cp, INT_PTR ramptr) 
{
  #ifndef NSEEL_EEL1_COMPAT_MODE
    short oldsw, newsw;
  #endif
  #ifdef _MSC_VER
 
    __asm
    {
#ifndef NSEEL_EEL1_COMPAT_MODE
      fnstcw [oldsw]
      mov ax, [oldsw]
      or ax, 0x23F  // 53 or 64 bit precision (depending on whether 0x100 is set), and masking all exceptions
      mov [newsw], ax
      fldcw [newsw]
#endif
      
      mov eax, cp
      mov ebx, ramptr
 
      pushad 
      mov ebp, esp
      and esp, -16
 
       // on win32, which _MSC_VER implies, we keep things aligned to 16 bytes, and if we call a win32 function,
       // the stack is 16 byte aligned before the call, meaning that if calling a function with no frame pointer,
       // the stack would be aligned to a 16 byte boundary +4, which isn't good for performance. Having said that,
       // normally we compile with frame pointers (which brings that to 16 byte + 8, which is fine), or ICC, which
       // for nontrivial functions will align the stack itself (for very short functions, it appears to weigh the 
       // cost of aligning the stack vs that of the slower misaligned double accesses).
 
       // it may be worthwhile (at some point) to put some logic in the code that calls out to functions
       // (generic1parm etc) to detect which alignment would be most optimal.
      sub esp, 12
      call eax
      mov esp, ebp
      popad
#ifndef NSEEL_EEL1_COMPAT_MODE
      fldcw [oldsw]
#endif
    };
 
  #else // gcc x86
    __asm__(
#ifndef NSEEL_EEL1_COMPAT_MODE
      "fnstcw %2\n"
      "movw %2, %%ax\n"
      "orw $0x23F, %%ax\n" // 53 or 64 bit precision (depending on whether 0x100 is set), and masking all exceptions
      "movw %%ax, %3\n"
      "fldcw %3\n"
#endif
          "pushl %%ebx\n"
          "movl %%ecx, %%ebx\n"
          "pushl %%ebp\n"
          "movl %%esp, %%ebp\n"
          "andl $-16, %%esp\n" // align stack to 16 bytes
          "subl $12, %%esp\n" // call will push 4 bytes on stack, align for that
          "call *%%edx\n"
          "leave\n"
          "popl %%ebx\n"
#ifndef NSEEL_EEL1_COMPAT_MODE
      "fldcw %2\n"
#endif
          ::
          "d" (cp), "c" (ramptr)
#ifndef NSEEL_EEL1_COMPAT_MODE
          , "m" (oldsw), "m" (newsw)
#endif
          : "%eax","%esi","%edi");
  #endif //gcc x86
}
 
void eel_enterfp(int s[2])
{
  #ifdef _MSC_VER
    __asm
    {
      mov ecx, s
      fnstcw [ecx]
      mov ax, [ecx]
      or ax, 0x23F  // 53 or 64 bit precision (depending on whether 0x100 is set), and masking all exceptions
      mov [ecx+4], ax
      fldcw [ecx+4]
    };
  #else
    __asm__(
      "fnstcw (%%ecx)\n"
      "movw (%%ecx), %%ax\n"
      "orw $0x23F, %%ax\n" // 53 or 64 bit precision (depending on whether 0x100 is set), and masking all exceptions
      "movw %%ax, 4(%%ecx)\n"
      "fldcw 4(%%ecx)\n"
          :: "c" (s) : "%eax");
  #endif
}
void eel_leavefp(int s[2])
{
  #ifdef _MSC_VER
    __asm
    {
      mov ecx, s
      fldcw [ecx]
    };
  #else
    __asm__(
      "fldcw (%%ecx)\n"
          :: "c" (s) : "%eax");
  #endif
}
 
static void eel_callcode32_fast(INT_PTR cp, INT_PTR ramptr) 
{
  #ifdef _MSC_VER
 
    __asm
    {
      mov eax, cp
      mov ebx, ramptr
 
      pushad 
      mov ebp, esp
      and esp, -16
 
       // on win32, which _MSC_VER implies, we keep things aligned to 16 bytes, and if we call a win32 function,
       // the stack is 16 byte aligned before the call, meaning that if calling a function with no frame pointer,
       // the stack would be aligned to a 16 byte boundary +4, which isn't good for performance. Having said that,
       // normally we compile with frame pointers (which brings that to 16 byte + 8, which is fine), or ICC, which
       // for nontrivial functions will align the stack itself (for very short functions, it appears to weigh the 
       // cost of aligning the stack vs that of the slower misaligned double accesses).
 
       // it may be worthwhile (at some point) to put some logic in the code that calls out to functions
       // (generic1parm etc) to detect which alignment would be most optimal.
      sub esp, 12
      call eax
      mov esp, ebp
      popad
    };
 
  #else // gcc x86
    __asm__(
          "pushl %%ebx\n"
          "movl %%ecx, %%ebx\n"
          "pushl %%ebp\n"
          "movl %%esp, %%ebp\n"
          "andl $-16, %%esp\n" // align stack to 16 bytes
          "subl $12, %%esp\n" // call will push 4 bytes on stack, align for that
          "call *%%edx\n"
          "leave\n"
          "popl %%ebx\n"
          ::
          "d" (cp), "c" (ramptr)
          : "%eax","%esi","%edi");
  #endif //gcc x86
}
 
#ifdef _MSC_VER
#pragma warning(pop)
#endif
 
 
static unsigned char *EEL_GLUE_set_immediate(void *_p, INT_PTR newv)
{
  char *p=(char*)_p;
  INT_PTR scan = 0xFEFEFEFE;
  while (*(INT_PTR *)p != scan) p++;
  *(INT_PTR *)p = newv;
  return (unsigned char *) (((INT_PTR*)p)+1);
}
 
#define INT_TO_LECHARS(x) ((x)&0xff),(((x)>>8)&0xff), (((x)>>16)&0xff), (((x)>>24)&0xff)
 
 
#define GLUE_INLINE_LOOPS
 
#define GLUE_LOOP_LOADCNT_SIZE (nseel_has_sse3() ? sizeof(GLUE_LOOP_LOADCNT_SSE3) : sizeof(GLUE_LOOP_LOADCNT_NOSSE3))
#define GLUE_LOOP_LOADCNT (nseel_has_sse3() ? GLUE_LOOP_LOADCNT_SSE3 : GLUE_LOOP_LOADCNT_NOSSE3)
static const unsigned char GLUE_LOOP_LOADCNT_SSE3[]={
        0xdb, 0x0e, // fisttp dword [esi]
        0x8B, 0x0E,           // mov ecx, [esi]
        0x81, 0xf9, 1,0,0,0,  // cmp ecx, 1
        0x0F, 0x8C, 0,0,0,0,  // JL <skipptr>
};
 
static const unsigned char GLUE_LOOP_LOADCNT_NOSSE3[]={
        0xd9, 0x7e, 0x04,       // fnstcw [esi+4]
        0x66, 0x8b, 0x46, 0x04, // mov ax, [esi+4]
        0x66, 0x0d, 0x00, 0x0c, // or ax, 0xC00
        0x66, 0x89, 0x46, 0x08, // mov [esi+8], ax
        0xd9, 0x6e, 0x08,       // fldcw [esi+8]
        0xDB, 0x1E,           // fistp dword [esi]
        0xd9, 0x6e, 0x04,     // fldcw [esi+4]
        0x8B, 0x0E,           // mov ecx, [esi]
        0x81, 0xf9, 1,0,0,0,  // cmp ecx, 1
        0x0F, 0x8C, 0,0,0,0,  // JL <skipptr>
};
 
#if NSEEL_LOOPFUNC_SUPPORT_MAXLEN > 0
#define GLUE_LOOP_CLAMPCNT_SIZE sizeof(GLUE_LOOP_CLAMPCNT)
static const unsigned char GLUE_LOOP_CLAMPCNT[]={
        0x81, 0xf9, INT_TO_LECHARS(NSEEL_LOOPFUNC_SUPPORT_MAXLEN), // cmp ecx, NSEEL_LOOPFUNC_SUPPORT_MAXLEN
        0x0F, 0x8C, 5,0,0,0,  // JL over-the-mov
        0xB9, INT_TO_LECHARS(NSEEL_LOOPFUNC_SUPPORT_MAXLEN), // mov ecx, NSEEL_LOOPFUNC_SUPPORT_MAXLEN
};
#else
 
#define GLUE_LOOP_CLAMPCNT_SIZE 0
#define GLUE_LOOP_CLAMPCNT ""
 
#endif
 
#define GLUE_LOOP_BEGIN_SIZE sizeof(GLUE_LOOP_BEGIN)
static const unsigned char GLUE_LOOP_BEGIN[]={ 
  0x56, //push esi
  0x51, // push ecx
  0x81, 0xEC, 0x08, 0,0,0, // sub esp, 8
};
static const unsigned char GLUE_LOOP_END[]={ 
  0x81, 0xC4, 0x08, 0,0,0, // add esp, 8
  0x59, //pop ecx
  0x5E, // pop esi
  0x49, // dec ecx
  0x0f, 0x85, 0,0,0,0, // jnz ...
};
 
 
#if NSEEL_LOOPFUNC_SUPPORT_MAXLEN > 0
#define GLUE_WHILE_SETUP_SIZE sizeof(GLUE_WHILE_SETUP)
static const unsigned char GLUE_WHILE_SETUP[]={
        0xB9, INT_TO_LECHARS(NSEEL_LOOPFUNC_SUPPORT_MAXLEN), // mov ecx, NSEEL_LOOPFUNC_SUPPORT_MAXLEN
};
static const unsigned char GLUE_WHILE_BEGIN[]={ 
  0x56, //push esi
  0x51, // push ecx
  0x81, 0xEC, 0x08, 0,0,0, // sub esp, 8
};
static const unsigned char GLUE_WHILE_END[]={ 
  0x81, 0xC4, 0x08, 0,0,0, // add esp, 8
  0x59, //pop ecx
  0x5E, // pop esi
 
 
  0x49, // dec ecx
  0x0f, 0x84, 0,0,0,0,  // jz endpt
};
 
 
#else
 
#define GLUE_WHILE_SETUP_SIZE  0
#define GLUE_WHILE_SETUP ""
#define GLUE_WHILE_END_NOJUMP
static const unsigned char GLUE_WHILE_BEGIN[]={ 
  0x56, //push esi
  0x81, 0xEC, 12, 0,0,0, // sub esp, 12
};
static const unsigned char GLUE_WHILE_END[]={ 
  0x81, 0xC4, 12, 0,0,0, // add esp, 12
  0x5E, // pop esi
};
 
#endif
 
static const unsigned char GLUE_WHILE_CHECK_RV[] = {
  0x85, 0xC0, // test eax, eax
  0x0F, 0x85, 0,0,0,0 // jnz  looppt
};
 
static const unsigned char GLUE_SET_P1_Z[] = { 0x29, 0xC0 }; // sub eax, eax
static const unsigned char GLUE_SET_P1_NZ[] = { 0xb0, 0x01 }; // mov al, 1
 
static const unsigned char GLUE_FXCH[] = {0xd9, 0xc9};
 
#define GLUE_HAS_FLDZ
static const unsigned char GLUE_FLDZ[] = {0xd9, 0xee};
#define GLUE_HAS_FLD1
static const unsigned char GLUE_FLD1[] = {0xd9, 0xe8};
 
static EEL_F negativezeropointfive=-0.5f;
static EEL_F onepointfive=1.5f;
#define GLUE_INVSQRT_NEEDREPL &negativezeropointfive, &onepointfive,
 
 
#define GLUE_HAS_NATIVE_TRIGSQRTLOG
 
 
void nseel_asm_or(void);
void nseel_asm_or0(void);
void nseel_asm_or_op(void);
void nseel_asm_and(void);
void nseel_asm_and_op(void);
void nseel_asm_xor(void);
void nseel_asm_xor_op(void);
void nseel_asm_shl(void);
void nseel_asm_shr(void);
void nseel_asm_mod(void);
void nseel_asm_mod_op(void);
void nseel_asm_stack_peek(void);
void _asm_gmegabuf(void);
void _asm_megabuf(void);
 
static struct roundinftab {
  void *fn;
  char istores; // number of fistp's
  char flag; // 0=fistpll, 1=fistpl, 2=fistpl 4(%esp)
  int newsz;
  void *newfn;
} s_round_fixes[] = {
  { nseel_asm_or, 2, },
  { nseel_asm_or_op, 2, },
  { nseel_asm_or0, 1, },
  { nseel_asm_and, 2, },
  { nseel_asm_and_op, 2, },
  { nseel_asm_xor, 2, },
  { nseel_asm_xor_op, 2, },
  { nseel_asm_shl, 2, 1, },
  { nseel_asm_shr, 2, 1, },
  { nseel_asm_mod, 2, 1, },
  { nseel_asm_mod_op, 2, 1, },
  { nseel_asm_stack_peek, 1, 1, },
  { _asm_megabuf, 1, 1, },
  { _asm_gmegabuf, 1, 2, },
};
 
static void eel_fixup_sse3(unsigned char *p, unsigned char *endp, int np, int flag)
{
  const int isz = flag == 2 ? 3 : 2;
  while (p+isz <= endp && np > 0)
  {
    if (flag == 0 && p[0] == 0xdf && (p[1]&0xbe) == 0x3e)
    {
      *p++ = 0xdd;
      *p -= 0x30;
      if (*p & 0x40) p++;
      np--;
    }
    else if (flag == 1 && p[0] == 0xdb && (p[1]&0xbe) == 0x1e)
    {
      *++p -= 0x10;
      if (*p & 0x40) p++;
      np--;
    }
    else if (flag == 2 && p[0] == 0xdb && (p[1]&0xbf) == 0x1c && p[2] == 0x24)
    {
      *++p -= 0x10;
      if (*p & 0x40) p++;
      p++;
      np--;
    }
    p++;
  }
  WDL_ASSERT(np == 0);
}
 
static int nseel_has_sse3()
{
  static char c;
  if (!c)
  {
    int features = 1;
  #ifdef _MSC_VER
    __asm {
      mov eax, 1
      cpuid
      mov [features], ecx
    };
  #else
    __asm__(
        "movl $1, %%eax\n"
        "pushl %%ebx\n"
        "pushl %%edx\n"
        "cpuid\n"
        "popl %%edx\n"
        "popl %%ebx\n"
        : "=c" (features) : : "%eax");
  #endif
    c=(features&1) ? 1 : -1;
  }
  return c>0;
}
static void *GLUE_realAddress(void *fn, int *size)
{
  static const unsigned char sig[12] = { 0x89, 0x90, 0x90, 0x90, 0x90, 0x90, 0x90, 0x90, 0x90, 0x90, 0x90, 0x90 };
  unsigned char *p = (unsigned char *)fn;
 
  size_t rmatch;
  const size_t nmatch = sizeof(s_round_fixes) / sizeof(s_round_fixes[0]);
  for (rmatch = 0; rmatch < nmatch && s_round_fixes[rmatch].fn != fn; rmatch++);
  if (rmatch < nmatch && s_round_fixes[rmatch].newfn && s_round_fixes[rmatch].newsz)
  {
    *size = s_round_fixes[rmatch].newsz;
    return s_round_fixes[rmatch].newfn;
  }
 
  #if defined(_DEBUG) && defined(_MSC_VER)
    if (*p == 0xE9) // this means jump to the following address (debug stub)
    {
      p += 5 + *(int *)(p+1);
    }
  #endif
 
  while (memcmp(p,sig,sizeof(sig))) p++;
  p+=sizeof(sig);
  fn = p;
 
  while (memcmp(p,sig,sizeof(sig))) p++;
  *size = p - (unsigned char *)fn;
 
  if (rmatch < nmatch)
  {
    static const unsigned char prefix[] = {
        0xd9, 0x7e, 0x10,       // fnstcw [esi+16]
        0x66, 0x8b, 0x46, 0x10, // mov ax, [esi+16]
        0x66, 0x0d, 0x00, 0x0c, // or ax, 0xC00
        0x66, 0x89, 0x46, 0x14, // mov [esi+20], ax
        0xd9, 0x6e, 0x14,       // fldcw [esi+20]
    };
    static const unsigned char postfix[] = {
        0xd9, 0x6e, 0x10,       // fldcw [esi+16]
    };
 
    const int has_sse = nseel_has_sse3();
 
    unsigned char *tmp = (unsigned char *) malloc(*size + (has_sse ? 0 : sizeof(prefix) + sizeof(postfix)));
    if (tmp)
    {
      if (has_sse)
      {
        memcpy(tmp,fn,*size);
        eel_fixup_sse3(tmp,tmp + *size,s_round_fixes[rmatch].istores, s_round_fixes[rmatch].flag);
      }
      else
      {
        memcpy(tmp,prefix,sizeof(prefix));
        memcpy(tmp+sizeof(prefix),fn,*size);
        memcpy(tmp+sizeof(prefix)+*size,postfix,sizeof(postfix));
 
        *size += sizeof(prefix) + sizeof(postfix);
      }
      fn = tmp;
      s_round_fixes[rmatch].newsz = *size;
      s_round_fixes[rmatch].newfn = fn;
    }
  }
 
  return fn;
}
 
#endif