If you are bold enough to try something different here is some code
to examine. I could butcher up something to create full implementation
outside my code. Testing showed faster than gdtoa. Low precision
routines around 0 exponent show near tie. Added code to show how I
have connected to my printf:

      case 'a':  case 'e':  case 'f':  case 'g':
      case 'A':  case 'E':  case 'F':  case 'G':
          /* gcc doesn't permit floats, it converts to doubles */
      {
        unsigned char *(*toStr)(uint32_t [5], unsigned char **,
                                                      int, int, int, 
unsigned);
        uint32_t w160[5];
        uint32_t dp;  /* 4 byte decimal point param */
        int32_t sch;
        __decimalLoad((unsigned char*)&dp, params);
        if ((ch | 0x20) == 'a')  ch -= 'A', toStr = realtohex;
        else                     ch -= 'E', toStr = realtostr;
        sch = format[-1];
          /* force 128bit long double into quad parsing */
        if (sch == 'q') {
#if (__LDBL_MANT_DIG__ != 64)
    quad_parse:
#endif
          sch = (('q' << 8) | ch);
          qtow(((vPtr == NULL) ? va_arg(arg, octle) : *(octle *)vPtr), 
w160);
        } else if (sch == 'L') {
#if (__LDBL_MANT_DIG__ != 64)
          goto quad_parse;
#endif
          sch = (('L' << 8) | ch);
          ldtow(((vPtr == NULL) ? va_arg(arg, long double) : *(long 
double *)vPtr), w160);
        } else {
          sch = (('l' << 8) | ch);
          dtow(((vPtr == NULL) ? va_arg(arg, double) : *(double 
*)vPtr), w160);
        }
        arg_flags |= (strBuf != buffer) << 7;
        strPtr = toStr(w160, &strBuf, sch, arg_prec, (int)arg_flags, 
dp);
      }
        if (strPtr == NULL)  goto error;

        val_ul = (n = (int)(strPtr - strBuf));
        strPtr = strBuf;
        arg_prec = 0;
        arg_flags &= ~ALT_PREFIX;
        if ((arg_flags & GRP_PREFIX) != 0)
          strPtr = __localizeReal(strPtr, &n), val_ul = n;
        goto L300;

This is part of my sys/types so above 'octle' doesn't look so strange:

typedef union {
#if __SIZEOF_INT128__ == 16
  uint128_t   uo;
#endif
#ifdef __BIG_ENDIAN__
  struct { uint64_t ull1, ull0; } ulls;
  struct { uint32_t ul3, ul2, ul1, ul0; } uls;
  struct { uint16_t us7, us6, us5, us4, us3, us2, us1, us0; } uss;
#else  /* __LITTLE_ENDIAN__ */
  struct { uint64_t ull0, ull1; } ulls;
  struct { uint32_t ul0, ul1, ul2, ul3; } uls;
  struct { uint16_t us0, us1, us2, us3, us4, us5, us6, us7; } uss;
#endif
} octle;

And from sysdep.h for attached ldtostr:

/* endian break down of uint64_t */
#ifdef __BIG_ENDIAN__
  typedef struct {  uint16_t w3; uint16_t w2; uint16_t w1; uint16_t w0; 
 } _ull_ws;
  typedef struct {  uint32_t hi; uint32_t lo;  } _ull_ls;
#elif defined (__LITTLE_ENDIAN__)
  typedef struct {  uint16_t w0; uint16_t w1; uint16_t w2; uint16_t w3; 
 } _ull_ws;
  typedef struct {  uint32_t lo; uint32_t hi;  } _ull_ls;
#else
  #error: undefined endianness
#endif
typedef union {  uint64_t ull;  _ull_ls  uls;  _ull_ws  uss;  } _ull_t;

Steve