]> granicus.if.org Git - jq/commitdiff
Import a copy of David Gay's dtoa.c
authorStephen Dolan <mu@netsoc.tcd.ie>
Sat, 1 Sep 2012 15:50:25 +0000 (16:50 +0100)
committerStephen Dolan <mu@netsoc.tcd.ie>
Sat, 1 Sep 2012 15:50:25 +0000 (16:50 +0100)
c/dtoa.c [new file with mode: 0644]

diff --git a/c/dtoa.c b/c/dtoa.c
new file mode 100644 (file)
index 0000000..41ed698
--- /dev/null
+++ b/c/dtoa.c
@@ -0,0 +1,4357 @@
+/****************************************************************
+ *
+ * The author of this software is David M. Gay.
+ *
+ * Copyright (c) 1991, 2000, 2001 by Lucent Technologies.
+ *
+ * Permission to use, copy, modify, and distribute this software for any
+ * purpose without fee is hereby granted, provided that this entire notice
+ * is included in all copies of any software which is or includes a copy
+ * or modification of this software and in all copies of the supporting
+ * documentation for such software.
+ *
+ * THIS SOFTWARE IS BEING PROVIDED "AS IS", WITHOUT ANY EXPRESS OR IMPLIED
+ * WARRANTY.  IN PARTICULAR, NEITHER THE AUTHOR NOR LUCENT MAKES ANY
+ * REPRESENTATION OR WARRANTY OF ANY KIND CONCERNING THE MERCHANTABILITY
+ * OF THIS SOFTWARE OR ITS FITNESS FOR ANY PARTICULAR PURPOSE.
+ *
+ ***************************************************************/
+
+/* Please send bug reports to David M. Gay (dmg at acm dot org,
+ * with " at " changed at "@" and " dot " changed to ".").     */
+
+/* On a machine with IEEE extended-precision registers, it is
+ * necessary to specify double-precision (53-bit) rounding precision
+ * before invoking strtod or dtoa.  If the machine uses (the equivalent
+ * of) Intel 80x87 arithmetic, the call
+ *     _control87(PC_53, MCW_PC);
+ * does this with many compilers.  Whether this or another call is
+ * appropriate depends on the compiler; for this to work, it may be
+ * necessary to #include "float.h" or another system-dependent header
+ * file.
+ */
+
+/* strtod for IEEE-, VAX-, and IBM-arithmetic machines.
+ * (Note that IEEE arithmetic is disabled by gcc's -ffast-math flag.)
+ *
+ * This strtod returns a nearest machine number to the input decimal
+ * string (or sets errno to ERANGE).  With IEEE arithmetic, ties are
+ * broken by the IEEE round-even rule.  Otherwise ties are broken by
+ * biased rounding (add half and chop).
+ *
+ * Inspired loosely by William D. Clinger's paper "How to Read Floating
+ * Point Numbers Accurately" [Proc. ACM SIGPLAN '90, pp. 92-101].
+ *
+ * Modifications:
+ *
+ *     1. We only require IEEE, IBM, or VAX double-precision
+ *             arithmetic (not IEEE double-extended).
+ *     2. We get by with floating-point arithmetic in a case that
+ *             Clinger missed -- when we're computing d * 10^n
+ *             for a small integer d and the integer n is not too
+ *             much larger than 22 (the maximum integer k for which
+ *             we can represent 10^k exactly), we may be able to
+ *             compute (d*10^k) * 10^(e-k) with just one roundoff.
+ *     3. Rather than a bit-at-a-time adjustment of the binary
+ *             result in the hard case, we use floating-point
+ *             arithmetic to determine the adjustment to within
+ *             one bit; only in really hard cases do we need to
+ *             compute a second residual.
+ *     4. Because of 3., we don't need a large table of powers of 10
+ *             for ten-to-e (just some small tables, e.g. of 10^k
+ *             for 0 <= k <= 22).
+ */
+
+/*
+ * #define IEEE_8087 for IEEE-arithmetic machines where the least
+ *     significant byte has the lowest address.
+ * #define IEEE_MC68k for IEEE-arithmetic machines where the most
+ *     significant byte has the lowest address.
+ * #define Long int on machines with 32-bit ints and 64-bit longs.
+ * #define IBM for IBM mainframe-style floating-point arithmetic.
+ * #define VAX for VAX-style floating-point arithmetic (D_floating).
+ * #define No_leftright to omit left-right logic in fast floating-point
+ *     computation of dtoa.  This will cause dtoa modes 4 and 5 to be
+ *     treated the same as modes 2 and 3 for some inputs.
+ * #define Honor_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3
+ *     and strtod and dtoa should round accordingly.  Unless Trust_FLT_ROUNDS
+ *     is also #defined, fegetround() will be queried for the rounding mode.
+ *     Note that both FLT_ROUNDS and fegetround() are specified by the C99
+ *     standard (and are specified to be consistent, with fesetround()
+ *     affecting the value of FLT_ROUNDS), but that some (Linux) systems
+ *     do not work correctly in this regard, so using fegetround() is more
+ *     portable than using FLT_ROUNDS directly.
+ * #define Check_FLT_ROUNDS if FLT_ROUNDS can assume the values 2 or 3
+ *     and Honor_FLT_ROUNDS is not #defined.
+ * #define RND_PRODQUOT to use rnd_prod and rnd_quot (assembly routines
+ *     that use extended-precision instructions to compute rounded
+ *     products and quotients) with IBM.
+ * #define ROUND_BIASED for IEEE-format with biased rounding and arithmetic
+ *     that rounds toward +Infinity.
+ * #define ROUND_BIASED_without_Round_Up for IEEE-format with biased
+ *     rounding when the underlying floating-point arithmetic uses
+ *     unbiased rounding.  This prevent using ordinary floating-point
+ *     arithmetic when the result could be computed with one rounding error.
+ * #define Inaccurate_Divide for IEEE-format with correctly rounded
+ *     products but inaccurate quotients, e.g., for Intel i860.
+ * #define NO_LONG_LONG on machines that do not have a "long long"
+ *     integer type (of >= 64 bits).  On such machines, you can
+ *     #define Just_16 to store 16 bits per 32-bit Long when doing
+ *     high-precision integer arithmetic.  Whether this speeds things
+ *     up or slows things down depends on the machine and the number
+ *     being converted.  If long long is available and the name is
+ *     something other than "long long", #define Llong to be the name,
+ *     and if "unsigned Llong" does not work as an unsigned version of
+ *     Llong, #define #ULLong to be the corresponding unsigned type.
+ * #define KR_headers for old-style C function headers.
+ * #define Bad_float_h if your system lacks a float.h or if it does not
+ *     define some or all of DBL_DIG, DBL_MAX_10_EXP, DBL_MAX_EXP,
+ *     FLT_RADIX, FLT_ROUNDS, and DBL_MAX.
+ * #define MALLOC your_malloc, where your_malloc(n) acts like malloc(n)
+ *     if memory is available and otherwise does something you deem
+ *     appropriate.  If MALLOC is undefined, malloc will be invoked
+ *     directly -- and assumed always to succeed.  Similarly, if you
+ *     want something other than the system's free() to be called to
+ *     recycle memory acquired from MALLOC, #define FREE to be the
+ *     name of the alternate routine.  (FREE or free is only called in
+ *     pathological cases, e.g., in a dtoa call after a dtoa return in
+ *     mode 3 with thousands of digits requested.)
+ * #define Omit_Private_Memory to omit logic (added Jan. 1998) for making
+ *     memory allocations from a private pool of memory when possible.
+ *     When used, the private pool is PRIVATE_MEM bytes long:  2304 bytes,
+ *     unless #defined to be a different length.  This default length
+ *     suffices to get rid of MALLOC calls except for unusual cases,
+ *     such as decimal-to-binary conversion of a very long string of
+ *     digits.  The longest string dtoa can return is about 751 bytes
+ *     long.  For conversions by strtod of strings of 800 digits and
+ *     all dtoa conversions in single-threaded executions with 8-byte
+ *     pointers, PRIVATE_MEM >= 7400 appears to suffice; with 4-byte
+ *     pointers, PRIVATE_MEM >= 7112 appears adequate.
+ * #define NO_INFNAN_CHECK if you do not wish to have INFNAN_CHECK
+ *     #defined automatically on IEEE systems.  On such systems,
+ *     when INFNAN_CHECK is #defined, strtod checks
+ *     for Infinity and NaN (case insensitively).  On some systems
+ *     (e.g., some HP systems), it may be necessary to #define NAN_WORD0
+ *     appropriately -- to the most significant word of a quiet NaN.
+ *     (On HP Series 700/800 machines, -DNAN_WORD0=0x7ff40000 works.)
+ *     When INFNAN_CHECK is #defined and No_Hex_NaN is not #defined,
+ *     strtod also accepts (case insensitively) strings of the form
+ *     NaN(x), where x is a string of hexadecimal digits and spaces;
+ *     if there is only one string of hexadecimal digits, it is taken
+ *     for the 52 fraction bits of the resulting NaN; if there are two
+ *     or more strings of hex digits, the first is for the high 20 bits,
+ *     the second and subsequent for the low 32 bits, with intervening
+ *     white space ignored; but if this results in none of the 52
+ *     fraction bits being on (an IEEE Infinity symbol), then NAN_WORD0
+ *     and NAN_WORD1 are used instead.
+ * #define MULTIPLE_THREADS if the system offers preemptively scheduled
+ *     multiple threads.  In this case, you must provide (or suitably
+ *     #define) two locks, acquired by ACQUIRE_DTOA_LOCK(n) and freed
+ *     by FREE_DTOA_LOCK(n) for n = 0 or 1.  (The second lock, accessed
+ *     in pow5mult, ensures lazy evaluation of only one copy of high
+ *     powers of 5; omitting this lock would introduce a small
+ *     probability of wasting memory, but would otherwise be harmless.)
+ *     You must also invoke freedtoa(s) to free the value s returned by
+ *     dtoa.  You may do so whether or not MULTIPLE_THREADS is #defined.
+ * #define NO_IEEE_Scale to disable new (Feb. 1997) logic in strtod that
+ *     avoids underflows on inputs whose result does not underflow.
+ *     If you #define NO_IEEE_Scale on a machine that uses IEEE-format
+ *     floating-point numbers and flushes underflows to zero rather
+ *     than implementing gradual underflow, then you must also #define
+ *     Sudden_Underflow.
+ * #define USE_LOCALE to use the current locale's decimal_point value.
+ * #define SET_INEXACT if IEEE arithmetic is being used and extra
+ *     computation should be done to set the inexact flag when the
+ *     result is inexact and avoid setting inexact when the result
+ *     is exact.  In this case, dtoa.c must be compiled in
+ *     an environment, perhaps provided by #include "dtoa.c" in a
+ *     suitable wrapper, that defines two functions,
+ *             int get_inexact(void);
+ *             void clear_inexact(void);
+ *     such that get_inexact() returns a nonzero value if the
+ *     inexact bit is already set, and clear_inexact() sets the
+ *     inexact bit to 0.  When SET_INEXACT is #defined, strtod
+ *     also does extra computations to set the underflow and overflow
+ *     flags when appropriate (i.e., when the result is tiny and
+ *     inexact or when it is a numeric value rounded to +-infinity).
+ * #define NO_ERRNO if strtod should not assign errno = ERANGE when
+ *     the result overflows to +-Infinity or underflows to 0.
+ * #define NO_HEX_FP to omit recognition of hexadecimal floating-point
+ *     values by strtod.
+ * #define NO_STRTOD_BIGCOMP (on IEEE-arithmetic systems only for now)
+ *     to disable logic for "fast" testing of very long input strings
+ *     to strtod.  This testing proceeds by initially truncating the
+ *     input string, then if necessary comparing the whole string with
+ *     a decimal expansion to decide close cases. This logic is only
+ *     used for input more than STRTOD_DIGLIM digits long (default 40).
+ */
+
+#ifndef Long
+#define Long long
+#endif
+#ifndef ULong
+typedef unsigned Long ULong;
+#endif
+
+#ifdef DEBUG
+#include "stdio.h"
+#define Bug(x) {fprintf(stderr, "%s\n", x); exit(1);}
+#endif
+
+#include "stdlib.h"
+#include "string.h"
+
+#ifdef USE_LOCALE
+#include "locale.h"
+#endif
+
+#ifdef Honor_FLT_ROUNDS
+#ifndef Trust_FLT_ROUNDS
+#include <fenv.h>
+#endif
+#endif
+
+#ifdef MALLOC
+#ifdef KR_headers
+extern char *MALLOC();
+#else
+extern void *MALLOC(size_t);
+#endif
+#else
+#define MALLOC malloc
+#endif
+
+#ifndef Omit_Private_Memory
+#ifndef PRIVATE_MEM
+#define PRIVATE_MEM 2304
+#endif
+#define PRIVATE_mem ((PRIVATE_MEM+sizeof(double)-1)/sizeof(double))
+static double private_mem[PRIVATE_mem], *pmem_next = private_mem;
+#endif
+
+#undef IEEE_Arith
+#undef Avoid_Underflow
+#ifdef IEEE_MC68k
+#define IEEE_Arith
+#endif
+#ifdef IEEE_8087
+#define IEEE_Arith
+#endif
+
+#ifdef IEEE_Arith
+#ifndef NO_INFNAN_CHECK
+#undef INFNAN_CHECK
+#define INFNAN_CHECK
+#endif
+#else
+#undef INFNAN_CHECK
+#define NO_STRTOD_BIGCOMP
+#endif
+
+#include "errno.h"
+
+#ifdef Bad_float_h
+
+#ifdef IEEE_Arith
+#define DBL_DIG 15
+#define DBL_MAX_10_EXP 308
+#define DBL_MAX_EXP 1024
+#define FLT_RADIX 2
+#endif /*IEEE_Arith*/
+
+#ifdef IBM
+#define DBL_DIG 16
+#define DBL_MAX_10_EXP 75
+#define DBL_MAX_EXP 63
+#define FLT_RADIX 16
+#define DBL_MAX 7.2370055773322621e+75
+#endif
+
+#ifdef VAX
+#define DBL_DIG 16
+#define DBL_MAX_10_EXP 38
+#define DBL_MAX_EXP 127
+#define FLT_RADIX 2
+#define DBL_MAX 1.7014118346046923e+38
+#endif
+
+#ifndef LONG_MAX
+#define LONG_MAX 2147483647
+#endif
+
+#else /* ifndef Bad_float_h */
+#include "float.h"
+#endif /* Bad_float_h */
+
+#ifndef __MATH_H__
+#include "math.h"
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef CONST
+#ifdef KR_headers
+#define CONST /* blank */
+#else
+#define CONST const
+#endif
+#endif
+
+#if defined(IEEE_8087) + defined(IEEE_MC68k) + defined(VAX) + defined(IBM) != 1
+Exactly one of IEEE_8087, IEEE_MC68k, VAX, or IBM should be defined.
+#endif
+
+typedef union { double d; ULong L[2]; } U;
+
+#ifdef IEEE_8087
+#define word0(x) (x)->L[1]
+#define word1(x) (x)->L[0]
+#else
+#define word0(x) (x)->L[0]
+#define word1(x) (x)->L[1]
+#endif
+#define dval(x) (x)->d
+
+#ifndef STRTOD_DIGLIM
+#define STRTOD_DIGLIM 40
+#endif
+
+#ifdef DIGLIM_DEBUG
+extern int strtod_diglim;
+#else
+#define strtod_diglim STRTOD_DIGLIM
+#endif
+
+/* The following definition of Storeinc is appropriate for MIPS processors.
+ * An alternative that might be better on some machines is
+ * #define Storeinc(a,b,c) (*a++ = b << 16 | c & 0xffff)
+ */
+#if defined(IEEE_8087) + defined(VAX)
+#define Storeinc(a,b,c) (((unsigned short *)a)[1] = (unsigned short)b, \
+((unsigned short *)a)[0] = (unsigned short)c, a++)
+#else
+#define Storeinc(a,b,c) (((unsigned short *)a)[0] = (unsigned short)b, \
+((unsigned short *)a)[1] = (unsigned short)c, a++)
+#endif
+
+/* #define P DBL_MANT_DIG */
+/* Ten_pmax = floor(P*log(2)/log(5)) */
+/* Bletch = (highest power of 2 < DBL_MAX_10_EXP) / 16 */
+/* Quick_max = floor((P-1)*log(FLT_RADIX)/log(10) - 1) */
+/* Int_max = floor(P*log(FLT_RADIX)/log(10) - 1) */
+
+#ifdef IEEE_Arith
+#define Exp_shift  20
+#define Exp_shift1 20
+#define Exp_msk1    0x100000
+#define Exp_msk11   0x100000
+#define Exp_mask  0x7ff00000
+#define P 53
+#define Nbits 53
+#define Bias 1023
+#define Emax 1023
+#define Emin (-1022)
+#define Exp_1  0x3ff00000
+#define Exp_11 0x3ff00000
+#define Ebits 11
+#define Frac_mask  0xfffff
+#define Frac_mask1 0xfffff
+#define Ten_pmax 22
+#define Bletch 0x10
+#define Bndry_mask  0xfffff
+#define Bndry_mask1 0xfffff
+#define LSB 1
+#define Sign_bit 0x80000000
+#define Log2P 1
+#define Tiny0 0
+#define Tiny1 1
+#define Quick_max 14
+#define Int_max 14
+#ifndef NO_IEEE_Scale
+#define Avoid_Underflow
+#ifdef Flush_Denorm    /* debugging option */
+#undef Sudden_Underflow
+#endif
+#endif
+
+#ifndef Flt_Rounds
+#ifdef FLT_ROUNDS
+#define Flt_Rounds FLT_ROUNDS
+#else
+#define Flt_Rounds 1
+#endif
+#endif /*Flt_Rounds*/
+
+#ifdef Honor_FLT_ROUNDS
+#undef Check_FLT_ROUNDS
+#define Check_FLT_ROUNDS
+#else
+#define Rounding Flt_Rounds
+#endif
+
+#else /* ifndef IEEE_Arith */
+#undef Check_FLT_ROUNDS
+#undef Honor_FLT_ROUNDS
+#undef SET_INEXACT
+#undef  Sudden_Underflow
+#define Sudden_Underflow
+#ifdef IBM
+#undef Flt_Rounds
+#define Flt_Rounds 0
+#define Exp_shift  24
+#define Exp_shift1 24
+#define Exp_msk1   0x1000000
+#define Exp_msk11  0x1000000
+#define Exp_mask  0x7f000000
+#define P 14
+#define Nbits 56
+#define Bias 65
+#define Emax 248
+#define Emin (-260)
+#define Exp_1  0x41000000
+#define Exp_11 0x41000000
+#define Ebits 8        /* exponent has 7 bits, but 8 is the right value in b2d */
+#define Frac_mask  0xffffff
+#define Frac_mask1 0xffffff
+#define Bletch 4
+#define Ten_pmax 22
+#define Bndry_mask  0xefffff
+#define Bndry_mask1 0xffffff
+#define LSB 1
+#define Sign_bit 0x80000000
+#define Log2P 4
+#define Tiny0 0x100000
+#define Tiny1 0
+#define Quick_max 14
+#define Int_max 15
+#else /* VAX */
+#undef Flt_Rounds
+#define Flt_Rounds 1
+#define Exp_shift  23
+#define Exp_shift1 7
+#define Exp_msk1    0x80
+#define Exp_msk11   0x800000
+#define Exp_mask  0x7f80
+#define P 56
+#define Nbits 56
+#define Bias 129
+#define Emax 126
+#define Emin (-129)
+#define Exp_1  0x40800000
+#define Exp_11 0x4080
+#define Ebits 8
+#define Frac_mask  0x7fffff
+#define Frac_mask1 0xffff007f
+#define Ten_pmax 24
+#define Bletch 2
+#define Bndry_mask  0xffff007f
+#define Bndry_mask1 0xffff007f
+#define LSB 0x10000
+#define Sign_bit 0x8000
+#define Log2P 1
+#define Tiny0 0x80
+#define Tiny1 0
+#define Quick_max 15
+#define Int_max 15
+#endif /* IBM, VAX */
+#endif /* IEEE_Arith */
+
+#ifndef IEEE_Arith
+#define ROUND_BIASED
+#else
+#ifdef ROUND_BIASED_without_Round_Up
+#undef  ROUND_BIASED
+#define ROUND_BIASED
+#endif
+#endif
+
+#ifdef RND_PRODQUOT
+#define rounded_product(a,b) a = rnd_prod(a, b)
+#define rounded_quotient(a,b) a = rnd_quot(a, b)
+#ifdef KR_headers
+extern double rnd_prod(), rnd_quot();
+#else
+extern double rnd_prod(double, double), rnd_quot(double, double);
+#endif
+#else
+#define rounded_product(a,b) a *= b
+#define rounded_quotient(a,b) a /= b
+#endif
+
+#define Big0 (Frac_mask1 | Exp_msk1*(DBL_MAX_EXP+Bias-1))
+#define Big1 0xffffffff
+
+#ifndef Pack_32
+#define Pack_32
+#endif
+
+typedef struct BCinfo BCinfo;
+ struct
+BCinfo { int dp0, dp1, dplen, dsign, e0, inexact, nd, nd0, rounding, scale, uflchk; };
+
+#ifdef KR_headers
+#define FFFFFFFF ((((unsigned long)0xffff)<<16)|(unsigned long)0xffff)
+#else
+#define FFFFFFFF 0xffffffffUL
+#endif
+
+#ifdef NO_LONG_LONG
+#undef ULLong
+#ifdef Just_16
+#undef Pack_32
+/* When Pack_32 is not defined, we store 16 bits per 32-bit Long.
+ * This makes some inner loops simpler and sometimes saves work
+ * during multiplications, but it often seems to make things slightly
+ * slower.  Hence the default is now to store 32 bits per Long.
+ */
+#endif
+#else  /* long long available */
+#ifndef Llong
+#define Llong long long
+#endif
+#ifndef ULLong
+#define ULLong unsigned Llong
+#endif
+#endif /* NO_LONG_LONG */
+
+#ifndef MULTIPLE_THREADS
+#define ACQUIRE_DTOA_LOCK(n)   /*nothing*/
+#define FREE_DTOA_LOCK(n)      /*nothing*/
+#endif
+
+#define Kmax 7
+
+#ifdef __cplusplus
+extern "C" double strtod(const char *s00, char **se);
+extern "C" char *dtoa(double d, int mode, int ndigits,
+                       int *decpt, int *sign, char **rve);
+#endif
+
+ struct
+Bigint {
+       struct Bigint *next;
+       int k, maxwds, sign, wds;
+       ULong x[1];
+       };
+
+ typedef struct Bigint Bigint;
+
+ static Bigint *freelist[Kmax+1];
+
+ static Bigint *
+Balloc
+#ifdef KR_headers
+       (k) int k;
+#else
+       (int k)
+#endif
+{
+       int x;
+       Bigint *rv;
+#ifndef Omit_Private_Memory
+       unsigned int len;
+#endif
+
+       ACQUIRE_DTOA_LOCK(0);
+       /* The k > Kmax case does not need ACQUIRE_DTOA_LOCK(0), */
+       /* but this case seems very unlikely. */
+       if (k <= Kmax && (rv = freelist[k]))
+               freelist[k] = rv->next;
+       else {
+               x = 1 << k;
+#ifdef Omit_Private_Memory
+               rv = (Bigint *)MALLOC(sizeof(Bigint) + (x-1)*sizeof(ULong));
+#else
+               len = (sizeof(Bigint) + (x-1)*sizeof(ULong) + sizeof(double) - 1)
+                       /sizeof(double);
+               if (k <= Kmax && pmem_next - private_mem + len <= PRIVATE_mem) {
+                       rv = (Bigint*)pmem_next;
+                       pmem_next += len;
+                       }
+               else
+                       rv = (Bigint*)MALLOC(len*sizeof(double));
+#endif
+               rv->k = k;
+               rv->maxwds = x;
+               }
+       FREE_DTOA_LOCK(0);
+       rv->sign = rv->wds = 0;
+       return rv;
+       }
+
+ static void
+Bfree
+#ifdef KR_headers
+       (v) Bigint *v;
+#else
+       (Bigint *v)
+#endif
+{
+       if (v) {
+               if (v->k > Kmax)
+#ifdef FREE
+                       FREE((void*)v);
+#else
+                       free((void*)v);
+#endif
+               else {
+                       ACQUIRE_DTOA_LOCK(0);
+                       v->next = freelist[v->k];
+                       freelist[v->k] = v;
+                       FREE_DTOA_LOCK(0);
+                       }
+               }
+       }
+
+#define Bcopy(x,y) memcpy((char *)&x->sign, (char *)&y->sign, \
+y->wds*sizeof(Long) + 2*sizeof(int))
+
+ static Bigint *
+multadd
+#ifdef KR_headers
+       (b, m, a) Bigint *b; int m, a;
+#else
+       (Bigint *b, int m, int a)       /* multiply by m and add a */
+#endif
+{
+       int i, wds;
+#ifdef ULLong
+       ULong *x;
+       ULLong carry, y;
+#else
+       ULong carry, *x, y;
+#ifdef Pack_32
+       ULong xi, z;
+#endif
+#endif
+       Bigint *b1;
+
+       wds = b->wds;
+       x = b->x;
+       i = 0;
+       carry = a;
+       do {
+#ifdef ULLong
+               y = *x * (ULLong)m + carry;
+               carry = y >> 32;
+               *x++ = y & FFFFFFFF;
+#else
+#ifdef Pack_32
+               xi = *x;
+               y = (xi & 0xffff) * m + carry;
+               z = (xi >> 16) * m + (y >> 16);
+               carry = z >> 16;
+               *x++ = (z << 16) + (y & 0xffff);
+#else
+               y = *x * m + carry;
+               carry = y >> 16;
+               *x++ = y & 0xffff;
+#endif
+#endif
+               }
+               while(++i < wds);
+       if (carry) {
+               if (wds >= b->maxwds) {
+                       b1 = Balloc(b->k+1);
+                       Bcopy(b1, b);
+                       Bfree(b);
+                       b = b1;
+                       }
+               b->x[wds++] = carry;
+               b->wds = wds;
+               }
+       return b;
+       }
+
+ static Bigint *
+s2b
+#ifdef KR_headers
+       (s, nd0, nd, y9, dplen) CONST char *s; int nd0, nd, dplen; ULong y9;
+#else
+       (const char *s, int nd0, int nd, ULong y9, int dplen)
+#endif
+{
+       Bigint *b;
+       int i, k;
+       Long x, y;
+
+       x = (nd + 8) / 9;
+       for(k = 0, y = 1; x > y; y <<= 1, k++) ;
+#ifdef Pack_32
+       b = Balloc(k);
+       b->x[0] = y9;
+       b->wds = 1;
+#else
+       b = Balloc(k+1);
+       b->x[0] = y9 & 0xffff;
+       b->wds = (b->x[1] = y9 >> 16) ? 2 : 1;
+#endif
+
+       i = 9;
+       if (9 < nd0) {
+               s += 9;
+               do b = multadd(b, 10, *s++ - '0');
+                       while(++i < nd0);
+               s += dplen;
+               }
+       else
+               s += dplen + 9;
+       for(; i < nd; i++)
+               b = multadd(b, 10, *s++ - '0');
+       return b;
+       }
+
+ static int
+hi0bits
+#ifdef KR_headers
+       (x) ULong x;
+#else
+       (ULong x)
+#endif
+{
+       int k = 0;
+
+       if (!(x & 0xffff0000)) {
+               k = 16;
+               x <<= 16;
+               }
+       if (!(x & 0xff000000)) {
+               k += 8;
+               x <<= 8;
+               }
+       if (!(x & 0xf0000000)) {
+               k += 4;
+               x <<= 4;
+               }
+       if (!(x & 0xc0000000)) {
+               k += 2;
+               x <<= 2;
+               }
+       if (!(x & 0x80000000)) {
+               k++;
+               if (!(x & 0x40000000))
+                       return 32;
+               }
+       return k;
+       }
+
+ static int
+lo0bits
+#ifdef KR_headers
+       (y) ULong *y;
+#else
+       (ULong *y)
+#endif
+{
+       int k;
+       ULong x = *y;
+
+       if (x & 7) {
+               if (x & 1)
+                       return 0;
+               if (x & 2) {
+                       *y = x >> 1;
+                       return 1;
+                       }
+               *y = x >> 2;
+               return 2;
+               }
+       k = 0;
+       if (!(x & 0xffff)) {
+               k = 16;
+               x >>= 16;
+               }
+       if (!(x & 0xff)) {
+               k += 8;
+               x >>= 8;
+               }
+       if (!(x & 0xf)) {
+               k += 4;
+               x >>= 4;
+               }
+       if (!(x & 0x3)) {
+               k += 2;
+               x >>= 2;
+               }
+       if (!(x & 1)) {
+               k++;
+               x >>= 1;
+               if (!x)
+                       return 32;
+               }
+       *y = x;
+       return k;
+       }
+
+ static Bigint *
+i2b
+#ifdef KR_headers
+       (i) int i;
+#else
+       (int i)
+#endif
+{
+       Bigint *b;
+
+       b = Balloc(1);
+       b->x[0] = i;
+       b->wds = 1;
+       return b;
+       }
+
+ static Bigint *
+mult
+#ifdef KR_headers
+       (a, b) Bigint *a, *b;
+#else
+       (Bigint *a, Bigint *b)
+#endif
+{
+       Bigint *c;
+       int k, wa, wb, wc;
+       ULong *x, *xa, *xae, *xb, *xbe, *xc, *xc0;
+       ULong y;
+#ifdef ULLong
+       ULLong carry, z;
+#else
+       ULong carry, z;
+#ifdef Pack_32
+       ULong z2;
+#endif
+#endif
+
+       if (a->wds < b->wds) {
+               c = a;
+               a = b;
+               b = c;
+               }
+       k = a->k;
+       wa = a->wds;
+       wb = b->wds;
+       wc = wa + wb;
+       if (wc > a->maxwds)
+               k++;
+       c = Balloc(k);
+       for(x = c->x, xa = x + wc; x < xa; x++)
+               *x = 0;
+       xa = a->x;
+       xae = xa + wa;
+       xb = b->x;
+       xbe = xb + wb;
+       xc0 = c->x;
+#ifdef ULLong
+       for(; xb < xbe; xc0++) {
+               if ((y = *xb++)) {
+                       x = xa;
+                       xc = xc0;
+                       carry = 0;
+                       do {
+                               z = *x++ * (ULLong)y + *xc + carry;
+                               carry = z >> 32;
+                               *xc++ = z & FFFFFFFF;
+                               }
+                               while(x < xae);
+                       *xc = carry;
+                       }
+               }
+#else
+#ifdef Pack_32
+       for(; xb < xbe; xb++, xc0++) {
+               if (y = *xb & 0xffff) {
+                       x = xa;
+                       xc = xc0;
+                       carry = 0;
+                       do {
+                               z = (*x & 0xffff) * y + (*xc & 0xffff) + carry;
+                               carry = z >> 16;
+                               z2 = (*x++ >> 16) * y + (*xc >> 16) + carry;
+                               carry = z2 >> 16;
+                               Storeinc(xc, z2, z);
+                               }
+                               while(x < xae);
+                       *xc = carry;
+                       }
+               if (y = *xb >> 16) {
+                       x = xa;
+                       xc = xc0;
+                       carry = 0;
+                       z2 = *xc;
+                       do {
+                               z = (*x & 0xffff) * y + (*xc >> 16) + carry;
+                               carry = z >> 16;
+                               Storeinc(xc, z, z2);
+                               z2 = (*x++ >> 16) * y + (*xc & 0xffff) + carry;
+                               carry = z2 >> 16;
+                               }
+                               while(x < xae);
+                       *xc = z2;
+                       }
+               }
+#else
+       for(; xb < xbe; xc0++) {
+               if (y = *xb++) {
+                       x = xa;
+                       xc = xc0;
+                       carry = 0;
+                       do {
+                               z = *x++ * y + *xc + carry;
+                               carry = z >> 16;
+                               *xc++ = z & 0xffff;
+                               }
+                               while(x < xae);
+                       *xc = carry;
+                       }
+               }
+#endif
+#endif
+       for(xc0 = c->x, xc = xc0 + wc; wc > 0 && !*--xc; --wc) ;
+       c->wds = wc;
+       return c;
+       }
+
+ static Bigint *p5s;
+
+ static Bigint *
+pow5mult
+#ifdef KR_headers
+       (b, k) Bigint *b; int k;
+#else
+       (Bigint *b, int k)
+#endif
+{
+       Bigint *b1, *p5, *p51;
+       int i;
+       static int p05[3] = { 5, 25, 125 };
+
+       if ((i = k & 3))
+               b = multadd(b, p05[i-1], 0);
+
+       if (!(k >>= 2))
+               return b;
+       if (!(p5 = p5s)) {
+               /* first time */
+#ifdef MULTIPLE_THREADS
+               ACQUIRE_DTOA_LOCK(1);
+               if (!(p5 = p5s)) {
+                       p5 = p5s = i2b(625);
+                       p5->next = 0;
+                       }
+               FREE_DTOA_LOCK(1);
+#else
+               p5 = p5s = i2b(625);
+               p5->next = 0;
+#endif
+               }
+       for(;;) {
+               if (k & 1) {
+                       b1 = mult(b, p5);
+                       Bfree(b);
+                       b = b1;
+                       }
+               if (!(k >>= 1))
+                       break;
+               if (!(p51 = p5->next)) {
+#ifdef MULTIPLE_THREADS
+                       ACQUIRE_DTOA_LOCK(1);
+                       if (!(p51 = p5->next)) {
+                               p51 = p5->next = mult(p5,p5);
+                               p51->next = 0;
+                               }
+                       FREE_DTOA_LOCK(1);
+#else
+                       p51 = p5->next = mult(p5,p5);
+                       p51->next = 0;
+#endif
+                       }
+               p5 = p51;
+               }
+       return b;
+       }
+
+ static Bigint *
+lshift
+#ifdef KR_headers
+       (b, k) Bigint *b; int k;
+#else
+       (Bigint *b, int k)
+#endif
+{
+       int i, k1, n, n1;
+       Bigint *b1;
+       ULong *x, *x1, *xe, z;
+
+#ifdef Pack_32
+       n = k >> 5;
+#else
+       n = k >> 4;
+#endif
+       k1 = b->k;
+       n1 = n + b->wds + 1;
+       for(i = b->maxwds; n1 > i; i <<= 1)
+               k1++;
+       b1 = Balloc(k1);
+       x1 = b1->x;
+       for(i = 0; i < n; i++)
+               *x1++ = 0;
+       x = b->x;
+       xe = x + b->wds;
+#ifdef Pack_32
+       if (k &= 0x1f) {
+               k1 = 32 - k;
+               z = 0;
+               do {
+                       *x1++ = *x << k | z;
+                       z = *x++ >> k1;
+                       }
+                       while(x < xe);
+               if ((*x1 = z))
+                       ++n1;
+               }
+#else
+       if (k &= 0xf) {
+               k1 = 16 - k;
+               z = 0;
+               do {
+                       *x1++ = *x << k  & 0xffff | z;
+                       z = *x++ >> k1;
+                       }
+                       while(x < xe);
+               if (*x1 = z)
+                       ++n1;
+               }
+#endif
+       else do
+               *x1++ = *x++;
+               while(x < xe);
+       b1->wds = n1 - 1;
+       Bfree(b);
+       return b1;
+       }
+
+ static int
+cmp
+#ifdef KR_headers
+       (a, b) Bigint *a, *b;
+#else
+       (Bigint *a, Bigint *b)
+#endif
+{
+       ULong *xa, *xa0, *xb, *xb0;
+       int i, j;
+
+       i = a->wds;
+       j = b->wds;
+#ifdef DEBUG
+       if (i > 1 && !a->x[i-1])
+               Bug("cmp called with a->x[a->wds-1] == 0");
+       if (j > 1 && !b->x[j-1])
+               Bug("cmp called with b->x[b->wds-1] == 0");
+#endif
+       if (i -= j)
+               return i;
+       xa0 = a->x;
+       xa = xa0 + j;
+       xb0 = b->x;
+       xb = xb0 + j;
+       for(;;) {
+               if (*--xa != *--xb)
+                       return *xa < *xb ? -1 : 1;
+               if (xa <= xa0)
+                       break;
+               }
+       return 0;
+       }
+
+ static Bigint *
+diff
+#ifdef KR_headers
+       (a, b) Bigint *a, *b;
+#else
+       (Bigint *a, Bigint *b)
+#endif
+{
+       Bigint *c;
+       int i, wa, wb;
+       ULong *xa, *xae, *xb, *xbe, *xc;
+#ifdef ULLong
+       ULLong borrow, y;
+#else
+       ULong borrow, y;
+#ifdef Pack_32
+       ULong z;
+#endif
+#endif
+
+       i = cmp(a,b);
+       if (!i) {
+               c = Balloc(0);
+               c->wds = 1;
+               c->x[0] = 0;
+               return c;
+               }
+       if (i < 0) {
+               c = a;
+               a = b;
+               b = c;
+               i = 1;
+               }
+       else
+               i = 0;
+       c = Balloc(a->k);
+       c->sign = i;
+       wa = a->wds;
+       xa = a->x;
+       xae = xa + wa;
+       wb = b->wds;
+       xb = b->x;
+       xbe = xb + wb;
+       xc = c->x;
+       borrow = 0;
+#ifdef ULLong
+       do {
+               y = (ULLong)*xa++ - *xb++ - borrow;
+               borrow = y >> 32 & (ULong)1;
+               *xc++ = y & FFFFFFFF;
+               }
+               while(xb < xbe);
+       while(xa < xae) {
+               y = *xa++ - borrow;
+               borrow = y >> 32 & (ULong)1;
+               *xc++ = y & FFFFFFFF;
+               }
+#else
+#ifdef Pack_32
+       do {
+               y = (*xa & 0xffff) - (*xb & 0xffff) - borrow;
+               borrow = (y & 0x10000) >> 16;
+               z = (*xa++ >> 16) - (*xb++ >> 16) - borrow;
+               borrow = (z & 0x10000) >> 16;
+               Storeinc(xc, z, y);
+               }
+               while(xb < xbe);
+       while(xa < xae) {
+               y = (*xa & 0xffff) - borrow;
+               borrow = (y & 0x10000) >> 16;
+               z = (*xa++ >> 16) - borrow;
+               borrow = (z & 0x10000) >> 16;
+               Storeinc(xc, z, y);
+               }
+#else
+       do {
+               y = *xa++ - *xb++ - borrow;
+               borrow = (y & 0x10000) >> 16;
+               *xc++ = y & 0xffff;
+               }
+               while(xb < xbe);
+       while(xa < xae) {
+               y = *xa++ - borrow;
+               borrow = (y & 0x10000) >> 16;
+               *xc++ = y & 0xffff;
+               }
+#endif
+#endif
+       while(!*--xc)
+               wa--;
+       c->wds = wa;
+       return c;
+       }
+
+ static double
+ulp
+#ifdef KR_headers
+       (x) U *x;
+#else
+       (U *x)
+#endif
+{
+       Long L;
+       U u;
+
+       L = (word0(x) & Exp_mask) - (P-1)*Exp_msk1;
+#ifndef Avoid_Underflow
+#ifndef Sudden_Underflow
+       if (L > 0) {
+#endif
+#endif
+#ifdef IBM
+               L |= Exp_msk1 >> 4;
+#endif
+               word0(&u) = L;
+               word1(&u) = 0;
+#ifndef Avoid_Underflow
+#ifndef Sudden_Underflow
+               }
+       else {
+               L = -L >> Exp_shift;
+               if (L < Exp_shift) {
+                       word0(&u) = 0x80000 >> L;
+                       word1(&u) = 0;
+                       }
+               else {
+                       word0(&u) = 0;
+                       L -= Exp_shift;
+                       word1(&u) = L >= 31 ? 1 : 1 << 31 - L;
+                       }
+               }
+#endif
+#endif
+       return dval(&u);
+       }
+
+ static double
+b2d
+#ifdef KR_headers
+       (a, e) Bigint *a; int *e;
+#else
+       (Bigint *a, int *e)
+#endif
+{
+       ULong *xa, *xa0, w, y, z;
+       int k;
+       U d;
+#ifdef VAX
+       ULong d0, d1;
+#else
+#define d0 word0(&d)
+#define d1 word1(&d)
+#endif
+
+       xa0 = a->x;
+       xa = xa0 + a->wds;
+       y = *--xa;
+#ifdef DEBUG
+       if (!y) Bug("zero y in b2d");
+#endif
+       k = hi0bits(y);
+       *e = 32 - k;
+#ifdef Pack_32
+       if (k < Ebits) {
+               d0 = Exp_1 | y >> (Ebits - k);
+               w = xa > xa0 ? *--xa : 0;
+               d1 = y << ((32-Ebits) + k) | w >> (Ebits - k);
+               goto ret_d;
+               }
+       z = xa > xa0 ? *--xa : 0;
+       if (k -= Ebits) {
+               d0 = Exp_1 | y << k | z >> (32 - k);
+               y = xa > xa0 ? *--xa : 0;
+               d1 = z << k | y >> (32 - k);
+               }
+       else {
+               d0 = Exp_1 | y;
+               d1 = z;
+               }
+#else
+       if (k < Ebits + 16) {
+               z = xa > xa0 ? *--xa : 0;
+               d0 = Exp_1 | y << k - Ebits | z >> Ebits + 16 - k;
+               w = xa > xa0 ? *--xa : 0;
+               y = xa > xa0 ? *--xa : 0;
+               d1 = z << k + 16 - Ebits | w << k - Ebits | y >> 16 + Ebits - k;
+               goto ret_d;
+               }
+       z = xa > xa0 ? *--xa : 0;
+       w = xa > xa0 ? *--xa : 0;
+       k -= Ebits + 16;
+       d0 = Exp_1 | y << k + 16 | z << k | w >> 16 - k;
+       y = xa > xa0 ? *--xa : 0;
+       d1 = w << k + 16 | y << k;
+#endif
+ ret_d:
+#ifdef VAX
+       word0(&d) = d0 >> 16 | d0 << 16;
+       word1(&d) = d1 >> 16 | d1 << 16;
+#else
+#undef d0
+#undef d1
+#endif
+       return dval(&d);
+       }
+
+ static Bigint *
+d2b
+#ifdef KR_headers
+       (d, e, bits) U *d; int *e, *bits;
+#else
+       (U *d, int *e, int *bits)
+#endif
+{
+       Bigint *b;
+       int de, k;
+       ULong *x, y, z;
+#ifndef Sudden_Underflow
+       int i;
+#endif
+#ifdef VAX
+       ULong d0, d1;
+       d0 = word0(d) >> 16 | word0(d) << 16;
+       d1 = word1(d) >> 16 | word1(d) << 16;
+#else
+#define d0 word0(d)
+#define d1 word1(d)
+#endif
+
+#ifdef Pack_32
+       b = Balloc(1);
+#else
+       b = Balloc(2);
+#endif
+       x = b->x;
+
+       z = d0 & Frac_mask;
+       d0 &= 0x7fffffff;       /* clear sign bit, which we ignore */
+#ifdef Sudden_Underflow
+       de = (int)(d0 >> Exp_shift);
+#ifndef IBM
+       z |= Exp_msk11;
+#endif
+#else
+       if ((de = (int)(d0 >> Exp_shift)))
+               z |= Exp_msk1;
+#endif
+#ifdef Pack_32
+       if ((y = d1)) {
+               if ((k = lo0bits(&y))) {
+                       x[0] = y | z << (32 - k);
+                       z >>= k;
+                       }
+               else
+                       x[0] = y;
+#ifndef Sudden_Underflow
+               i =
+#endif
+                   b->wds = (x[1] = z) ? 2 : 1;
+               }
+       else {
+               k = lo0bits(&z);
+               x[0] = z;
+#ifndef Sudden_Underflow
+               i =
+#endif
+                   b->wds = 1;
+               k += 32;
+               }
+#else
+       if (y = d1) {
+               if (k = lo0bits(&y))
+                       if (k >= 16) {
+                               x[0] = y | z << 32 - k & 0xffff;
+                               x[1] = z >> k - 16 & 0xffff;
+                               x[2] = z >> k;
+                               i = 2;
+                               }
+                       else {
+                               x[0] = y & 0xffff;
+                               x[1] = y >> 16 | z << 16 - k & 0xffff;
+                               x[2] = z >> k & 0xffff;
+                               x[3] = z >> k+16;
+                               i = 3;
+                               }
+               else {
+                       x[0] = y & 0xffff;
+                       x[1] = y >> 16;
+                       x[2] = z & 0xffff;
+                       x[3] = z >> 16;
+                       i = 3;
+                       }
+               }
+       else {
+#ifdef DEBUG
+               if (!z)
+                       Bug("Zero passed to d2b");
+#endif
+               k = lo0bits(&z);
+               if (k >= 16) {
+                       x[0] = z;
+                       i = 0;
+                       }
+               else {
+                       x[0] = z & 0xffff;
+                       x[1] = z >> 16;
+                       i = 1;
+                       }
+               k += 32;
+               }
+       while(!x[i])
+               --i;
+       b->wds = i + 1;
+#endif
+#ifndef Sudden_Underflow
+       if (de) {
+#endif
+#ifdef IBM
+               *e = (de - Bias - (P-1) << 2) + k;
+               *bits = 4*P + 8 - k - hi0bits(word0(d) & Frac_mask);
+#else
+               *e = de - Bias - (P-1) + k;
+               *bits = P - k;
+#endif
+#ifndef Sudden_Underflow
+               }
+       else {
+               *e = de - Bias - (P-1) + 1 + k;
+#ifdef Pack_32
+               *bits = 32*i - hi0bits(x[i-1]);
+#else
+               *bits = (i+2)*16 - hi0bits(x[i]);
+#endif
+               }
+#endif
+       return b;
+       }
+#undef d0
+#undef d1
+
+ static double
+ratio
+#ifdef KR_headers
+       (a, b) Bigint *a, *b;
+#else
+       (Bigint *a, Bigint *b)
+#endif
+{
+       U da, db;
+       int k, ka, kb;
+
+       dval(&da) = b2d(a, &ka);
+       dval(&db) = b2d(b, &kb);
+#ifdef Pack_32
+       k = ka - kb + 32*(a->wds - b->wds);
+#else
+       k = ka - kb + 16*(a->wds - b->wds);
+#endif
+#ifdef IBM
+       if (k > 0) {
+               word0(&da) += (k >> 2)*Exp_msk1;
+               if (k &= 3)
+                       dval(&da) *= 1 << k;
+               }
+       else {
+               k = -k;
+               word0(&db) += (k >> 2)*Exp_msk1;
+               if (k &= 3)
+                       dval(&db) *= 1 << k;
+               }
+#else
+       if (k > 0)
+               word0(&da) += k*Exp_msk1;
+       else {
+               k = -k;
+               word0(&db) += k*Exp_msk1;
+               }
+#endif
+       return dval(&da) / dval(&db);
+       }
+
+ static CONST double
+tens[] = {
+               1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9,
+               1e10, 1e11, 1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19,
+               1e20, 1e21, 1e22
+#ifdef VAX
+               , 1e23, 1e24
+#endif
+               };
+
+ static CONST double
+#ifdef IEEE_Arith
+bigtens[] = { 1e16, 1e32, 1e64, 1e128, 1e256 };
+static CONST double tinytens[] = { 1e-16, 1e-32, 1e-64, 1e-128,
+#ifdef Avoid_Underflow
+               9007199254740992.*9007199254740992.e-256
+               /* = 2^106 * 1e-256 */
+#else
+               1e-256
+#endif
+               };
+/* The factor of 2^53 in tinytens[4] helps us avoid setting the underflow */
+/* flag unnecessarily.  It leads to a song and dance at the end of strtod. */
+#define Scale_Bit 0x10
+#define n_bigtens 5
+#else
+#ifdef IBM
+bigtens[] = { 1e16, 1e32, 1e64 };
+static CONST double tinytens[] = { 1e-16, 1e-32, 1e-64 };
+#define n_bigtens 3
+#else
+bigtens[] = { 1e16, 1e32 };
+static CONST double tinytens[] = { 1e-16, 1e-32 };
+#define n_bigtens 2
+#endif
+#endif
+
+#undef Need_Hexdig
+#ifdef INFNAN_CHECK
+#ifndef No_Hex_NaN
+#define Need_Hexdig
+#endif
+#endif
+
+#ifndef Need_Hexdig
+#ifndef NO_HEX_FP
+#define Need_Hexdig
+#endif
+#endif
+
+#ifdef Need_Hexdig /*{*/
+static unsigned char hexdig[256];
+
+ static void
+#ifdef KR_headers
+htinit(h, s, inc) unsigned char *h; unsigned char *s; int inc;
+#else
+htinit(unsigned char *h, unsigned char *s, int inc)
+#endif
+{
+       int i, j;
+       for(i = 0; (j = s[i]) !=0; i++)
+               h[j] = i + inc;
+       }
+
+ static void
+#ifdef KR_headers
+hexdig_init()
+#else
+hexdig_init(void)
+#endif
+{
+#define USC (unsigned char *)
+       htinit(hexdig, USC "0123456789", 0x10);
+       htinit(hexdig, USC "abcdef", 0x10 + 10);
+       htinit(hexdig, USC "ABCDEF", 0x10 + 10);
+       }
+#endif /* } Need_Hexdig */
+
+#ifdef INFNAN_CHECK
+
+#ifndef NAN_WORD0
+#define NAN_WORD0 0x7ff80000
+#endif
+
+#ifndef NAN_WORD1
+#define NAN_WORD1 0
+#endif
+
+ static int
+match
+#ifdef KR_headers
+       (sp, t) char **sp, *t;
+#else
+       (const char **sp, const char *t)
+#endif
+{
+       int c, d;
+       CONST char *s = *sp;
+
+       while((d = *t++)) {
+               if ((c = *++s) >= 'A' && c <= 'Z')
+                       c += 'a' - 'A';
+               if (c != d)
+                       return 0;
+               }
+       *sp = s + 1;
+       return 1;
+       }
+
+#ifndef No_Hex_NaN
+ static void
+hexnan
+#ifdef KR_headers
+       (rvp, sp) U *rvp; CONST char **sp;
+#else
+       (U *rvp, const char **sp)
+#endif
+{
+       ULong c, x[2];
+       CONST char *s;
+       int c1, havedig, udx0, xshift;
+
+       if (!hexdig['0'])
+               hexdig_init();
+       x[0] = x[1] = 0;
+       havedig = xshift = 0;
+       udx0 = 1;
+       s = *sp;
+       /* allow optional initial 0x or 0X */
+       while((c = *(CONST unsigned char*)(s+1)) && c <= ' ')
+               ++s;
+       if (s[1] == '0' && (s[2] == 'x' || s[2] == 'X'))
+               s += 2;
+       while((c = *(CONST unsigned char*)++s)) {
+               if ((c1 = hexdig[c]))
+                       c  = c1 & 0xf;
+               else if (c <= ' ') {
+                       if (udx0 && havedig) {
+                               udx0 = 0;
+                               xshift = 1;
+                               }
+                       continue;
+                       }
+#ifdef GDTOA_NON_PEDANTIC_NANCHECK
+               else if (/*(*/ c == ')' && havedig) {
+                       *sp = s + 1;
+                       break;
+                       }
+               else
+                       return; /* invalid form: don't change *sp */
+#else
+               else {
+                       do {
+                               if (/*(*/ c == ')') {
+                                       *sp = s + 1;
+                                       break;
+                                       }
+                               } while((c = *++s));
+                       break;
+                       }
+#endif
+               havedig = 1;
+               if (xshift) {
+                       xshift = 0;
+                       x[0] = x[1];
+                       x[1] = 0;
+                       }
+               if (udx0)
+                       x[0] = (x[0] << 4) | (x[1] >> 28);
+               x[1] = (x[1] << 4) | c;
+               }
+       if ((x[0] &= 0xfffff) || x[1]) {
+               word0(rvp) = Exp_mask | x[0];
+               word1(rvp) = x[1];
+               }
+       }
+#endif /*No_Hex_NaN*/
+#endif /* INFNAN_CHECK */
+
+#ifdef Pack_32
+#define ULbits 32
+#define kshift 5
+#define kmask 31
+#else
+#define ULbits 16
+#define kshift 4
+#define kmask 15
+#endif
+
+#if !defined(NO_HEX_FP) || defined(Honor_FLT_ROUNDS) /*{*/
+ static Bigint *
+#ifdef KR_headers
+increment(b) Bigint *b;
+#else
+increment(Bigint *b)
+#endif
+{
+       ULong *x, *xe;
+       Bigint *b1;
+
+       x = b->x;
+       xe = x + b->wds;
+       do {
+               if (*x < (ULong)0xffffffffL) {
+                       ++*x;
+                       return b;
+                       }
+               *x++ = 0;
+               } while(x < xe);
+       {
+               if (b->wds >= b->maxwds) {
+                       b1 = Balloc(b->k+1);
+                       Bcopy(b1,b);
+                       Bfree(b);
+                       b = b1;
+                       }
+               b->x[b->wds++] = 1;
+               }
+       return b;
+       }
+
+#endif /*}*/
+
+#ifndef NO_HEX_FP /*{*/
+
+ static void
+#ifdef KR_headers
+rshift(b, k) Bigint *b; int k;
+#else
+rshift(Bigint *b, int k)
+#endif
+{
+       ULong *x, *x1, *xe, y;
+       int n;
+
+       x = x1 = b->x;
+       n = k >> kshift;
+       if (n < b->wds) {
+               xe = x + b->wds;
+               x += n;
+               if (k &= kmask) {
+                       n = 32 - k;
+                       y = *x++ >> k;
+                       while(x < xe) {
+                               *x1++ = (y | (*x << n)) & 0xffffffff;
+                               y = *x++ >> k;
+                               }
+                       if ((*x1 = y) !=0)
+                               x1++;
+                       }
+               else
+                       while(x < xe)
+                               *x1++ = *x++;
+               }
+       if ((b->wds = x1 - b->x) == 0)
+               b->x[0] = 0;
+       }
+
+ static ULong
+#ifdef KR_headers
+any_on(b, k) Bigint *b; int k;
+#else
+any_on(Bigint *b, int k)
+#endif
+{
+       int n, nwds;
+       ULong *x, *x0, x1, x2;
+
+       x = b->x;
+       nwds = b->wds;
+       n = k >> kshift;
+       if (n > nwds)
+               n = nwds;
+       else if (n < nwds && (k &= kmask)) {
+               x1 = x2 = x[n];
+               x1 >>= k;
+               x1 <<= k;
+               if (x1 != x2)
+                       return 1;
+               }
+       x0 = x;
+       x += n;
+       while(x > x0)
+               if (*--x)
+                       return 1;
+       return 0;
+       }
+
+enum { /* rounding values: same as FLT_ROUNDS */
+       Round_zero = 0,
+       Round_near = 1,
+       Round_up = 2,
+       Round_down = 3
+       };
+
+ void
+#ifdef KR_headers
+gethex(sp, rvp, rounding, sign)
+       CONST char **sp; U *rvp; int rounding, sign;
+#else
+gethex( CONST char **sp, U *rvp, int rounding, int sign)
+#endif
+{
+       Bigint *b;
+       CONST unsigned char *decpt, *s0, *s, *s1;
+       Long e, e1;
+       ULong L, lostbits, *x;
+       int big, denorm, esign, havedig, k, n, nbits, up, zret;
+#ifdef IBM
+       int j;
+#endif
+       enum {
+#ifdef IEEE_Arith /*{{*/
+               emax = 0x7fe - Bias - P + 1,
+               emin = Emin - P + 1
+#else /*}{*/
+               emin = Emin - P,
+#ifdef VAX
+               emax = 0x7ff - Bias - P + 1
+#endif
+#ifdef IBM
+               emax = 0x7f - Bias - P
+#endif
+#endif /*}}*/
+               };
+#ifdef USE_LOCALE
+       int i;
+#ifdef NO_LOCALE_CACHE
+       const unsigned char *decimalpoint = (unsigned char*)
+               localeconv()->decimal_point;
+#else
+       const unsigned char *decimalpoint;
+       static unsigned char *decimalpoint_cache;
+       if (!(s0 = decimalpoint_cache)) {
+               s0 = (unsigned char*)localeconv()->decimal_point;
+               if ((decimalpoint_cache = (unsigned char*)
+                               MALLOC(strlen((CONST char*)s0) + 1))) {
+                       strcpy((char*)decimalpoint_cache, (CONST char*)s0);
+                       s0 = decimalpoint_cache;
+                       }
+               }
+       decimalpoint = s0;
+#endif
+#endif
+
+       if (!hexdig['0'])
+               hexdig_init();
+       havedig = 0;
+       s0 = *(CONST unsigned char **)sp + 2;
+       while(s0[havedig] == '0')
+               havedig++;
+       s0 += havedig;
+       s = s0;
+       decpt = 0;
+       zret = 0;
+       e = 0;
+       if (hexdig[*s])
+               havedig++;
+       else {
+               zret = 1;
+#ifdef USE_LOCALE
+               for(i = 0; decimalpoint[i]; ++i) {
+                       if (s[i] != decimalpoint[i])
+                               goto pcheck;
+                       }
+               decpt = s += i;
+#else
+               if (*s != '.')
+                       goto pcheck;
+               decpt = ++s;
+#endif
+               if (!hexdig[*s])
+                       goto pcheck;
+               while(*s == '0')
+                       s++;
+               if (hexdig[*s])
+                       zret = 0;
+               havedig = 1;
+               s0 = s;
+               }
+       while(hexdig[*s])
+               s++;
+#ifdef USE_LOCALE
+       if (*s == *decimalpoint && !decpt) {
+               for(i = 1; decimalpoint[i]; ++i) {
+                       if (s[i] != decimalpoint[i])
+                               goto pcheck;
+                       }
+               decpt = s += i;
+#else
+       if (*s == '.' && !decpt) {
+               decpt = ++s;
+#endif
+               while(hexdig[*s])
+                       s++;
+               }/*}*/
+       if (decpt)
+               e = -(((Long)(s-decpt)) << 2);
+ pcheck:
+       s1 = s;
+       big = esign = 0;
+       switch(*s) {
+         case 'p':
+         case 'P':
+               switch(*++s) {
+                 case '-':
+                       esign = 1;
+                       /* no break */
+                 case '+':
+                       s++;
+                 }
+               if ((n = hexdig[*s]) == 0 || n > 0x19) {
+                       s = s1;
+                       break;
+                       }
+               e1 = n - 0x10;
+               while((n = hexdig[*++s]) !=0 && n <= 0x19) {
+                       if (e1 & 0xf8000000)
+                               big = 1;
+                       e1 = 10*e1 + n - 0x10;
+                       }
+               if (esign)
+                       e1 = -e1;
+               e += e1;
+         }
+       *sp = (char*)s;
+       if (!havedig)
+               *sp = (char*)s0 - 1;
+       if (zret)
+               goto retz1;
+       if (big) {
+               if (esign) {
+#ifdef IEEE_Arith
+                       switch(rounding) {
+                         case Round_up:
+                               if (sign)
+                                       break;
+                               goto ret_tiny;
+                         case Round_down:
+                               if (!sign)
+                                       break;
+                               goto ret_tiny;
+                         }
+#endif
+                       goto retz;
+#ifdef IEEE_Arith
+ ret_tiny:
+#ifndef NO_ERRNO
+                       errno = ERANGE;
+#endif
+                       word0(rvp) = 0;
+                       word1(rvp) = 1;
+                       return;
+#endif /* IEEE_Arith */
+                       }
+               switch(rounding) {
+                 case Round_near:
+                       goto ovfl1;
+                 case Round_up:
+                       if (!sign)
+                               goto ovfl1;
+                       goto ret_big;
+                 case Round_down:
+                       if (sign)
+                               goto ovfl1;
+                       goto ret_big;
+                 }
+ ret_big:
+               word0(rvp) = Big0;
+               word1(rvp) = Big1;
+               return;
+               }
+       n = s1 - s0 - 1;
+       for(k = 0; n > (1 << (kshift-2)) - 1; n >>= 1)
+               k++;
+       b = Balloc(k);
+       x = b->x;
+       n = 0;
+       L = 0;
+#ifdef USE_LOCALE
+       for(i = 0; decimalpoint[i+1]; ++i);
+#endif
+       while(s1 > s0) {
+#ifdef USE_LOCALE
+               if (*--s1 == decimalpoint[i]) {
+                       s1 -= i;
+                       continue;
+                       }
+#else
+               if (*--s1 == '.')
+                       continue;
+#endif
+               if (n == ULbits) {
+                       *x++ = L;
+                       L = 0;
+                       n = 0;
+                       }
+               L |= (hexdig[*s1] & 0x0f) << n;
+               n += 4;
+               }
+       *x++ = L;
+       b->wds = n = x - b->x;
+       n = ULbits*n - hi0bits(L);
+       nbits = Nbits;
+       lostbits = 0;
+       x = b->x;
+       if (n > nbits) {
+               n -= nbits;
+               if (any_on(b,n)) {
+                       lostbits = 1;
+                       k = n - 1;
+                       if (x[k>>kshift] & 1 << (k & kmask)) {
+                               lostbits = 2;
+                               if (k > 0 && any_on(b,k))
+                                       lostbits = 3;
+                               }
+                       }
+               rshift(b, n);
+               e += n;
+               }
+       else if (n < nbits) {
+               n = nbits - n;
+               b = lshift(b, n);
+               e -= n;
+               x = b->x;
+               }
+       if (e > Emax) {
+ ovfl:
+               Bfree(b);
+ ovfl1:
+#ifndef NO_ERRNO
+               errno = ERANGE;
+#endif
+               word0(rvp) = Exp_mask;
+               word1(rvp) = 0;
+               return;
+               }
+       denorm = 0;
+       if (e < emin) {
+               denorm = 1;
+               n = emin - e;
+               if (n >= nbits) {
+#ifdef IEEE_Arith /*{*/
+                       switch (rounding) {
+                         case Round_near:
+                               if (n == nbits && (n < 2 || any_on(b,n-1)))
+                                       goto ret_tiny;
+                               break;
+                         case Round_up:
+                               if (!sign)
+                                       goto ret_tiny;
+                               break;
+                         case Round_down:
+                               if (sign)
+                                       goto ret_tiny;
+                         }
+#endif /* } IEEE_Arith */
+                       Bfree(b);
+ retz:
+#ifndef NO_ERRNO
+                       errno = ERANGE;
+#endif
+ retz1:
+                       rvp->d = 0.;
+                       return;
+                       }
+               k = n - 1;
+               if (lostbits)
+                       lostbits = 1;
+               else if (k > 0)
+                       lostbits = any_on(b,k);
+               if (x[k>>kshift] & 1 << (k & kmask))
+                       lostbits |= 2;
+               nbits -= n;
+               rshift(b,n);
+               e = emin;
+               }
+       if (lostbits) {
+               up = 0;
+               switch(rounding) {
+                 case Round_zero:
+                       break;
+                 case Round_near:
+                       if (lostbits & 2
+                        && (lostbits & 1) | (x[0] & 1))
+                               up = 1;
+                       break;
+                 case Round_up:
+                       up = 1 - sign;
+                       break;
+                 case Round_down:
+                       up = sign;
+                 }
+               if (up) {
+                       k = b->wds;
+                       b = increment(b);
+                       x = b->x;
+                       if (denorm) {
+#if 0
+                               if (nbits == Nbits - 1
+                                && x[nbits >> kshift] & 1 << (nbits & kmask))
+                                       denorm = 0; /* not currently used */
+#endif
+                               }
+                       else if (b->wds > k
+                        || ((n = nbits & kmask) !=0
+                            && hi0bits(x[k-1]) < 32-n)) {
+                               rshift(b,1);
+                               if (++e > Emax)
+                                       goto ovfl;
+                               }
+                       }
+               }
+#ifdef IEEE_Arith
+       if (denorm)
+               word0(rvp) = b->wds > 1 ? b->x[1] & ~0x100000 : 0;
+       else
+               word0(rvp) = (b->x[1] & ~0x100000) | ((e + 0x3ff + 52) << 20);
+       word1(rvp) = b->x[0];
+#endif
+#ifdef IBM
+       if ((j = e & 3)) {
+               k = b->x[0] & ((1 << j) - 1);
+               rshift(b,j);
+               if (k) {
+                       switch(rounding) {
+                         case Round_up:
+                               if (!sign)
+                                       increment(b);
+                               break;
+                         case Round_down:
+                               if (sign)
+                                       increment(b);
+                               break;
+                         case Round_near:
+                               j = 1 << (j-1);
+                               if (k & j && ((k & (j-1)) | lostbits))
+                                       increment(b);
+                         }
+                       }
+               }
+       e >>= 2;
+       word0(rvp) = b->x[1] | ((e + 65 + 13) << 24);
+       word1(rvp) = b->x[0];
+#endif
+#ifdef VAX
+       /* The next two lines ignore swap of low- and high-order 2 bytes. */
+       /* word0(rvp) = (b->x[1] & ~0x800000) | ((e + 129 + 55) << 23); */
+       /* word1(rvp) = b->x[0]; */
+       word0(rvp) = ((b->x[1] & ~0x800000) >> 16) | ((e + 129 + 55) << 7) | (b->x[1] << 16);
+       word1(rvp) = (b->x[0] >> 16) | (b->x[0] << 16);
+#endif
+       Bfree(b);
+       }
+#endif /*!NO_HEX_FP}*/
+
+ static int
+#ifdef KR_headers
+dshift(b, p2) Bigint *b; int p2;
+#else
+dshift(Bigint *b, int p2)
+#endif
+{
+       int rv = hi0bits(b->x[b->wds-1]) - 4;
+       if (p2 > 0)
+               rv -= p2;
+       return rv & kmask;
+       }
+
+ static int
+quorem
+#ifdef KR_headers
+       (b, S) Bigint *b, *S;
+#else
+       (Bigint *b, Bigint *S)
+#endif
+{
+       int n;
+       ULong *bx, *bxe, q, *sx, *sxe;
+#ifdef ULLong
+       ULLong borrow, carry, y, ys;
+#else
+       ULong borrow, carry, y, ys;
+#ifdef Pack_32
+       ULong si, z, zs;
+#endif
+#endif
+
+       n = S->wds;
+#ifdef DEBUG
+       /*debug*/ if (b->wds > n)
+       /*debug*/       Bug("oversize b in quorem");
+#endif
+       if (b->wds < n)
+               return 0;
+       sx = S->x;
+       sxe = sx + --n;
+       bx = b->x;
+       bxe = bx + n;
+       q = *bxe / (*sxe + 1);  /* ensure q <= true quotient */
+#ifdef DEBUG
+#ifdef NO_STRTOD_BIGCOMP
+       /*debug*/ if (q > 9)
+#else
+       /* An oversized q is possible when quorem is called from bigcomp and */
+       /* the input is near, e.g., twice the smallest denormalized number. */
+       /*debug*/ if (q > 15)
+#endif
+       /*debug*/       Bug("oversized quotient in quorem");
+#endif
+       if (q) {
+               borrow = 0;
+               carry = 0;
+               do {
+#ifdef ULLong
+                       ys = *sx++ * (ULLong)q + carry;
+                       carry = ys >> 32;
+                       y = *bx - (ys & FFFFFFFF) - borrow;
+                       borrow = y >> 32 & (ULong)1;
+                       *bx++ = y & FFFFFFFF;
+#else
+#ifdef Pack_32
+                       si = *sx++;
+                       ys = (si & 0xffff) * q + carry;
+                       zs = (si >> 16) * q + (ys >> 16);
+                       carry = zs >> 16;
+                       y = (*bx & 0xffff) - (ys & 0xffff) - borrow;
+                       borrow = (y & 0x10000) >> 16;
+                       z = (*bx >> 16) - (zs & 0xffff) - borrow;
+                       borrow = (z & 0x10000) >> 16;
+                       Storeinc(bx, z, y);
+#else
+                       ys = *sx++ * q + carry;
+                       carry = ys >> 16;
+                       y = *bx - (ys & 0xffff) - borrow;
+                       borrow = (y & 0x10000) >> 16;
+                       *bx++ = y & 0xffff;
+#endif
+#endif
+                       }
+                       while(sx <= sxe);
+               if (!*bxe) {
+                       bx = b->x;
+                       while(--bxe > bx && !*bxe)
+                               --n;
+                       b->wds = n;
+                       }
+               }
+       if (cmp(b, S) >= 0) {
+               q++;
+               borrow = 0;
+               carry = 0;
+               bx = b->x;
+               sx = S->x;
+               do {
+#ifdef ULLong
+                       ys = *sx++ + carry;
+                       carry = ys >> 32;
+                       y = *bx - (ys & FFFFFFFF) - borrow;
+                       borrow = y >> 32 & (ULong)1;
+                       *bx++ = y & FFFFFFFF;
+#else
+#ifdef Pack_32
+                       si = *sx++;
+                       ys = (si & 0xffff) + carry;
+                       zs = (si >> 16) + (ys >> 16);
+                       carry = zs >> 16;
+                       y = (*bx & 0xffff) - (ys & 0xffff) - borrow;
+                       borrow = (y & 0x10000) >> 16;
+                       z = (*bx >> 16) - (zs & 0xffff) - borrow;
+                       borrow = (z & 0x10000) >> 16;
+                       Storeinc(bx, z, y);
+#else
+                       ys = *sx++ + carry;
+                       carry = ys >> 16;
+                       y = *bx - (ys & 0xffff) - borrow;
+                       borrow = (y & 0x10000) >> 16;
+                       *bx++ = y & 0xffff;
+#endif
+#endif
+                       }
+                       while(sx <= sxe);
+               bx = b->x;
+               bxe = bx + n;
+               if (!*bxe) {
+                       while(--bxe > bx && !*bxe)
+                               --n;
+                       b->wds = n;
+                       }
+               }
+       return q;
+       }
+
+#if defined(Avoid_Underflow) || !defined(NO_STRTOD_BIGCOMP) /*{*/
+ static double
+sulp
+#ifdef KR_headers
+       (x, bc) U *x; BCinfo *bc;
+#else
+       (U *x, BCinfo *bc)
+#endif
+{
+       U u;
+       double rv;
+       int i;
+
+       rv = ulp(x);
+       if (!bc->scale || (i = 2*P + 1 - ((word0(x) & Exp_mask) >> Exp_shift)) <= 0)
+               return rv; /* Is there an example where i <= 0 ? */
+       word0(&u) = Exp_1 + (i << Exp_shift);
+       word1(&u) = 0;
+       return rv * u.d;
+       }
+#endif /*}*/
+
+#ifndef NO_STRTOD_BIGCOMP
+ static void
+bigcomp
+#ifdef KR_headers
+       (rv, s0, bc)
+       U *rv; CONST char *s0; BCinfo *bc;
+#else
+       (U *rv, const char *s0, BCinfo *bc)
+#endif
+{
+       Bigint *b, *d;
+       int b2, bbits, d2, dd, dig, dsign, i, j, nd, nd0, p2, p5, speccase;
+
+       dsign = bc->dsign;
+       nd = bc->nd;
+       nd0 = bc->nd0;
+       p5 = nd + bc->e0 - 1;
+       speccase = 0;
+#ifndef Sudden_Underflow
+       if (rv->d == 0.) {      /* special case: value near underflow-to-zero */
+                               /* threshold was rounded to zero */
+               b = i2b(1);
+               p2 = Emin - P + 1;
+               bbits = 1;
+#ifdef Avoid_Underflow
+               word0(rv) = (P+2) << Exp_shift;
+#else
+               word1(rv) = 1;
+#endif
+               i = 0;
+#ifdef Honor_FLT_ROUNDS
+               if (bc->rounding == 1)
+#endif
+                       {
+                       speccase = 1;
+                       --p2;
+                       dsign = 0;
+                       goto have_i;
+                       }
+               }
+       else
+#endif
+               b = d2b(rv, &p2, &bbits);
+#ifdef Avoid_Underflow
+       p2 -= bc->scale;
+#endif
+       /* floor(log2(rv)) == bbits - 1 + p2 */
+       /* Check for denormal case. */
+       i = P - bbits;
+       if (i > (j = P - Emin - 1 + p2)) {
+#ifdef Sudden_Underflow
+               Bfree(b);
+               b = i2b(1);
+               p2 = Emin;
+               i = P - 1;
+#ifdef Avoid_Underflow
+               word0(rv) = (1 + bc->scale) << Exp_shift;
+#else
+               word0(rv) = Exp_msk1;
+#endif
+               word1(rv) = 0;
+#else
+               i = j;
+#endif
+               }
+#ifdef Honor_FLT_ROUNDS
+       if (bc->rounding != 1) {
+               if (i > 0)
+                       b = lshift(b, i);
+               if (dsign)
+                       b = increment(b);
+               }
+       else
+#endif
+               {
+               b = lshift(b, ++i);
+               b->x[0] |= 1;
+               }
+#ifndef Sudden_Underflow
+ have_i:
+#endif
+       p2 -= p5 + i;
+       d = i2b(1);
+       /* Arrange for convenient computation of quotients:
+        * shift left if necessary so divisor has 4 leading 0 bits.
+        */
+       if (p5 > 0)
+               d = pow5mult(d, p5);
+       else if (p5 < 0)
+               b = pow5mult(b, -p5);
+       if (p2 > 0) {
+               b2 = p2;
+               d2 = 0;
+               }
+       else {
+               b2 = 0;
+               d2 = -p2;
+               }
+       i = dshift(d, d2);
+       if ((b2 += i) > 0)
+               b = lshift(b, b2);
+       if ((d2 += i) > 0)
+               d = lshift(d, d2);
+
+       /* Now b/d = exactly half-way between the two floating-point values */
+       /* on either side of the input string.  Compute first digit of b/d. */
+
+       if (!(dig = quorem(b,d))) {
+               b = multadd(b, 10, 0);  /* very unlikely */
+               dig = quorem(b,d);
+               }
+
+       /* Compare b/d with s0 */
+
+       for(i = 0; i < nd0; ) {
+               if ((dd = s0[i++] - '0' - dig))
+                       goto ret;
+               if (!b->x[0] && b->wds == 1) {
+                       if (i < nd)
+                               dd = 1;
+                       goto ret;
+                       }
+               b = multadd(b, 10, 0);
+               dig = quorem(b,d);
+               }
+       for(j = bc->dp1; i++ < nd;) {
+               if ((dd = s0[j++] - '0' - dig))
+                       goto ret;
+               if (!b->x[0] && b->wds == 1) {
+                       if (i < nd)
+                               dd = 1;
+                       goto ret;
+                       }
+               b = multadd(b, 10, 0);
+               dig = quorem(b,d);
+               }
+       if (dig > 0 || b->x[0] || b->wds > 1)
+               dd = -1;
+ ret:
+       Bfree(b);
+       Bfree(d);
+#ifdef Honor_FLT_ROUNDS
+       if (bc->rounding != 1) {
+               if (dd < 0) {
+                       if (bc->rounding == 0) {
+                               if (!dsign)
+                                       goto retlow1;
+                               }
+                       else if (dsign)
+                               goto rethi1;
+                       }
+               else if (dd > 0) {
+                       if (bc->rounding == 0) {
+                               if (dsign)
+                                       goto rethi1;
+                               goto ret1;
+                               }
+                       if (!dsign)
+                               goto rethi1;
+                       dval(rv) += 2.*sulp(rv,bc);
+                       }
+               else {
+                       bc->inexact = 0;
+                       if (dsign)
+                               goto rethi1;
+                       }
+               }
+       else
+#endif
+       if (speccase) {
+               if (dd <= 0)
+                       rv->d = 0.;
+               }
+       else if (dd < 0) {
+               if (!dsign)     /* does not happen for round-near */
+retlow1:
+                       dval(rv) -= sulp(rv,bc);
+               }
+       else if (dd > 0) {
+               if (dsign) {
+ rethi1:
+                       dval(rv) += sulp(rv,bc);
+                       }
+               }
+       else {
+               /* Exact half-way case:  apply round-even rule. */
+               if ((j = ((word0(rv) & Exp_mask) >> Exp_shift) - bc->scale) <= 0) {
+                       i = 1 - j;
+                       if (i <= 31) {
+                               if (word1(rv) & (0x1 << i))
+                                       goto odd;
+                               }
+                       else if (word0(rv) & (0x1 << (i-32)))
+                               goto odd;
+                       }
+               else if (word1(rv) & 1) {
+ odd:
+                       if (dsign)
+                               goto rethi1;
+                       goto retlow1;
+                       }
+               }
+
+#ifdef Honor_FLT_ROUNDS
+ ret1:
+#endif
+       return;
+       }
+#endif /* NO_STRTOD_BIGCOMP */
+
+ double
+strtod
+#ifdef KR_headers
+       (s00, se) CONST char *s00; char **se;
+#else
+       (const char *s00, char **se)
+#endif
+{
+       int bb2, bb5, bbe, bd2, bd5, bbbits, bs2, c, e, e1;
+       int esign, i, j, k, nd, nd0, nf, nz, nz0, nz1, sign;
+       CONST char *s, *s0, *s1;
+       double aadj, aadj1;
+       Long L;
+       U aadj2, adj, rv, rv0;
+       ULong y, z;
+       BCinfo bc;
+       Bigint *bb, *bb1, *bd, *bd0, *bs, *delta;
+#ifdef Avoid_Underflow
+       ULong Lsb, Lsb1;
+#endif
+#ifdef SET_INEXACT
+       int oldinexact;
+#endif
+#ifndef NO_STRTOD_BIGCOMP
+       int req_bigcomp = 0;
+#endif
+#ifdef Honor_FLT_ROUNDS /*{*/
+#ifdef Trust_FLT_ROUNDS /*{{ only define this if FLT_ROUNDS really works! */
+       bc.rounding = Flt_Rounds;
+#else /*}{*/
+       bc.rounding = 1;
+       switch(fegetround()) {
+         case FE_TOWARDZERO:   bc.rounding = 0; break;
+         case FE_UPWARD:       bc.rounding = 2; break;
+         case FE_DOWNWARD:     bc.rounding = 3;
+         }
+#endif /*}}*/
+#endif /*}*/
+#ifdef USE_LOCALE
+       CONST char *s2;
+#endif
+
+       sign = nz0 = nz1 = nz = bc.dplen = bc.uflchk = 0;
+       dval(&rv) = 0.;
+       for(s = s00;;s++) switch(*s) {
+               case '-':
+                       sign = 1;
+                       /* no break */
+               case '+':
+                       if (*++s)
+                               goto break2;
+                       /* no break */
+               case 0:
+                       goto ret0;
+               case '\t':
+               case '\n':
+               case '\v':
+               case '\f':
+               case '\r':
+               case ' ':
+                       continue;
+               default:
+                       goto break2;
+               }
+ break2:
+       if (*s == '0') {
+#ifndef NO_HEX_FP /*{*/
+               switch(s[1]) {
+                 case 'x':
+                 case 'X':
+#ifdef Honor_FLT_ROUNDS
+                       gethex(&s, &rv, bc.rounding, sign);
+#else
+                       gethex(&s, &rv, 1, sign);
+#endif
+                       goto ret;
+                 }
+#endif /*}*/
+               nz0 = 1;
+               while(*++s == '0') ;
+               if (!*s)
+                       goto ret;
+               }
+       s0 = s;
+       y = z = 0;
+       for(nd = nf = 0; (c = *s) >= '0' && c <= '9'; nd++, s++)
+               if (nd < 9)
+                       y = 10*y + c - '0';
+               else if (nd < 16)
+                       z = 10*z + c - '0';
+       nd0 = nd;
+       bc.dp0 = bc.dp1 = s - s0;
+       for(s1 = s; s1 > s0 && *--s1 == '0'; )
+               ++nz1;
+#ifdef USE_LOCALE
+       s1 = localeconv()->decimal_point;
+       if (c == *s1) {
+               c = '.';
+               if (*++s1) {
+                       s2 = s;
+                       for(;;) {
+                               if (*++s2 != *s1) {
+                                       c = 0;
+                                       break;
+                                       }
+                               if (!*++s1) {
+                                       s = s2;
+                                       break;
+                                       }
+                               }
+                       }
+               }
+#endif
+       if (c == '.') {
+               c = *++s;
+               bc.dp1 = s - s0;
+               bc.dplen = bc.dp1 - bc.dp0;
+               if (!nd) {
+                       for(; c == '0'; c = *++s)
+                               nz++;
+                       if (c > '0' && c <= '9') {
+                               bc.dp0 = s0 - s;
+                               bc.dp1 = bc.dp0 + bc.dplen;
+                               s0 = s;
+                               nf += nz;
+                               nz = 0;
+                               goto have_dig;
+                               }
+                       goto dig_done;
+                       }
+               for(; c >= '0' && c <= '9'; c = *++s) {
+ have_dig:
+                       nz++;
+                       if (c -= '0') {
+                               nf += nz;
+                               for(i = 1; i < nz; i++)
+                                       if (nd++ < 9)
+                                               y *= 10;
+                                       else if (nd <= DBL_DIG + 1)
+                                               z *= 10;
+                               if (nd++ < 9)
+                                       y = 10*y + c;
+                               else if (nd <= DBL_DIG + 1)
+                                       z = 10*z + c;
+                               nz = nz1 = 0;
+                               }
+                       }
+               }
+ dig_done:
+       e = 0;
+       if (c == 'e' || c == 'E') {
+               if (!nd && !nz && !nz0) {
+                       goto ret0;
+                       }
+               s00 = s;
+               esign = 0;
+               switch(c = *++s) {
+                       case '-':
+                               esign = 1;
+                       case '+':
+                               c = *++s;
+                       }
+               if (c >= '0' && c <= '9') {
+                       while(c == '0')
+                               c = *++s;
+                       if (c > '0' && c <= '9') {
+                               L = c - '0';
+                               s1 = s;
+                               while((c = *++s) >= '0' && c <= '9')
+                                       L = 10*L + c - '0';
+                               if (s - s1 > 8 || L > 19999)
+                                       /* Avoid confusion from exponents
+                                        * so large that e might overflow.
+                                        */
+                                       e = 19999; /* safe for 16 bit ints */
+                               else
+                                       e = (int)L;
+                               if (esign)
+                                       e = -e;
+                               }
+                       else
+                               e = 0;
+                       }
+               else
+                       s = s00;
+               }
+       if (!nd) {
+               if (!nz && !nz0) {
+#ifdef INFNAN_CHECK
+                       /* Check for Nan and Infinity */
+                       if (!bc.dplen)
+                        switch(c) {
+                         case 'i':
+                         case 'I':
+                               if (match(&s,"nf")) {
+                                       --s;
+                                       if (!match(&s,"inity"))
+                                               ++s;
+                                       word0(&rv) = 0x7ff00000;
+                                       word1(&rv) = 0;
+                                       goto ret;
+                                       }
+                               break;
+                         case 'n':
+                         case 'N':
+                               if (match(&s, "an")) {
+                                       word0(&rv) = NAN_WORD0;
+                                       word1(&rv) = NAN_WORD1;
+#ifndef No_Hex_NaN
+                                       if (*s == '(') /*)*/
+                                               hexnan(&rv, &s);
+#endif
+                                       goto ret;
+                                       }
+                         }
+#endif /* INFNAN_CHECK */
+ ret0:
+                       s = s00;
+                       sign = 0;
+                       }
+               goto ret;
+               }
+       bc.e0 = e1 = e -= nf;
+
+       /* Now we have nd0 digits, starting at s0, followed by a
+        * decimal point, followed by nd-nd0 digits.  The number we're
+        * after is the integer represented by those digits times
+        * 10**e */
+
+       if (!nd0)
+               nd0 = nd;
+       k = nd < DBL_DIG + 1 ? nd : DBL_DIG + 1;
+       dval(&rv) = y;
+       if (k > 9) {
+#ifdef SET_INEXACT
+               if (k > DBL_DIG)
+                       oldinexact = get_inexact();
+#endif
+               dval(&rv) = tens[k - 9] * dval(&rv) + z;
+               }
+       bd0 = 0;
+       if (nd <= DBL_DIG
+#ifndef RND_PRODQUOT
+#ifndef Honor_FLT_ROUNDS
+               && Flt_Rounds == 1
+#endif
+#endif
+                       ) {
+               if (!e)
+                       goto ret;
+#ifndef ROUND_BIASED_without_Round_Up
+               if (e > 0) {
+                       if (e <= Ten_pmax) {
+#ifdef VAX
+                               goto vax_ovfl_check;
+#else
+#ifdef Honor_FLT_ROUNDS
+                               /* round correctly FLT_ROUNDS = 2 or 3 */
+                               if (sign) {
+                                       rv.d = -rv.d;
+                                       sign = 0;
+                                       }
+#endif
+                               /* rv = */ rounded_product(dval(&rv), tens[e]);
+                               goto ret;
+#endif
+                               }
+                       i = DBL_DIG - nd;
+                       if (e <= Ten_pmax + i) {
+                               /* A fancier test would sometimes let us do
+                                * this for larger i values.
+                                */
+#ifdef Honor_FLT_ROUNDS
+                               /* round correctly FLT_ROUNDS = 2 or 3 */
+                               if (sign) {
+                                       rv.d = -rv.d;
+                                       sign = 0;
+                                       }
+#endif
+                               e -= i;
+                               dval(&rv) *= tens[i];
+#ifdef VAX
+                               /* VAX exponent range is so narrow we must
+                                * worry about overflow here...
+                                */
+ vax_ovfl_check:
+                               word0(&rv) -= P*Exp_msk1;
+                               /* rv = */ rounded_product(dval(&rv), tens[e]);
+                               if ((word0(&rv) & Exp_mask)
+                                > Exp_msk1*(DBL_MAX_EXP+Bias-1-P))
+                                       goto ovfl;
+                               word0(&rv) += P*Exp_msk1;
+#else
+                               /* rv = */ rounded_product(dval(&rv), tens[e]);
+#endif
+                               goto ret;
+                               }
+                       }
+#ifndef Inaccurate_Divide
+               else if (e >= -Ten_pmax) {
+#ifdef Honor_FLT_ROUNDS
+                       /* round correctly FLT_ROUNDS = 2 or 3 */
+                       if (sign) {
+                               rv.d = -rv.d;
+                               sign = 0;
+                               }
+#endif
+                       /* rv = */ rounded_quotient(dval(&rv), tens[-e]);
+                       goto ret;
+                       }
+#endif
+#endif /* ROUND_BIASED_without_Round_Up */
+               }
+       e1 += nd - k;
+
+#ifdef IEEE_Arith
+#ifdef SET_INEXACT
+       bc.inexact = 1;
+       if (k <= DBL_DIG)
+               oldinexact = get_inexact();
+#endif
+#ifdef Avoid_Underflow
+       bc.scale = 0;
+#endif
+#ifdef Honor_FLT_ROUNDS
+       if (bc.rounding >= 2) {
+               if (sign)
+                       bc.rounding = bc.rounding == 2 ? 0 : 2;
+               else
+                       if (bc.rounding != 2)
+                               bc.rounding = 0;
+               }
+#endif
+#endif /*IEEE_Arith*/
+
+       /* Get starting approximation = rv * 10**e1 */
+
+       if (e1 > 0) {
+               if ((i = e1 & 15))
+                       dval(&rv) *= tens[i];
+               if (e1 &= ~15) {
+                       if (e1 > DBL_MAX_10_EXP) {
+ ovfl:
+                               /* Can't trust HUGE_VAL */
+#ifdef IEEE_Arith
+#ifdef Honor_FLT_ROUNDS
+                               switch(bc.rounding) {
+                                 case 0: /* toward 0 */
+                                 case 3: /* toward -infinity */
+                                       word0(&rv) = Big0;
+                                       word1(&rv) = Big1;
+                                       break;
+                                 default:
+                                       word0(&rv) = Exp_mask;
+                                       word1(&rv) = 0;
+                                 }
+#else /*Honor_FLT_ROUNDS*/
+                               word0(&rv) = Exp_mask;
+                               word1(&rv) = 0;
+#endif /*Honor_FLT_ROUNDS*/
+#ifdef SET_INEXACT
+                               /* set overflow bit */
+                               dval(&rv0) = 1e300;
+                               dval(&rv0) *= dval(&rv0);
+#endif
+#else /*IEEE_Arith*/
+                               word0(&rv) = Big0;
+                               word1(&rv) = Big1;
+#endif /*IEEE_Arith*/
+ range_err:
+                               if (bd0) {
+                                       Bfree(bb);
+                                       Bfree(bd);
+                                       Bfree(bs);
+                                       Bfree(bd0);
+                                       Bfree(delta);
+                                       }
+#ifndef NO_ERRNO
+                               errno = ERANGE;
+#endif
+                               goto ret;
+                               }
+                       e1 >>= 4;
+                       for(j = 0; e1 > 1; j++, e1 >>= 1)
+                               if (e1 & 1)
+                                       dval(&rv) *= bigtens[j];
+               /* The last multiplication could overflow. */
+                       word0(&rv) -= P*Exp_msk1;
+                       dval(&rv) *= bigtens[j];
+                       if ((z = word0(&rv) & Exp_mask)
+                        > Exp_msk1*(DBL_MAX_EXP+Bias-P))
+                               goto ovfl;
+                       if (z > Exp_msk1*(DBL_MAX_EXP+Bias-1-P)) {
+                               /* set to largest number */
+                               /* (Can't trust DBL_MAX) */
+                               word0(&rv) = Big0;
+                               word1(&rv) = Big1;
+                               }
+                       else
+                               word0(&rv) += P*Exp_msk1;
+                       }
+               }
+       else if (e1 < 0) {
+               e1 = -e1;
+               if ((i = e1 & 15))
+                       dval(&rv) /= tens[i];
+               if (e1 >>= 4) {
+                       if (e1 >= 1 << n_bigtens)
+                               goto undfl;
+#ifdef Avoid_Underflow
+                       if (e1 & Scale_Bit)
+                               bc.scale = 2*P;
+                       for(j = 0; e1 > 0; j++, e1 >>= 1)
+                               if (e1 & 1)
+                                       dval(&rv) *= tinytens[j];
+                       if (bc.scale && (j = 2*P + 1 - ((word0(&rv) & Exp_mask)
+                                               >> Exp_shift)) > 0) {
+                               /* scaled rv is denormal; clear j low bits */
+                               if (j >= 32) {
+                                       if (j > 54)
+                                               goto undfl;
+                                       word1(&rv) = 0;
+                                       if (j >= 53)
+                                        word0(&rv) = (P+2)*Exp_msk1;
+                                       else
+                                        word0(&rv) &= 0xffffffff << (j-32);
+                                       }
+                               else
+                                       word1(&rv) &= 0xffffffff << j;
+                               }
+#else
+                       for(j = 0; e1 > 1; j++, e1 >>= 1)
+                               if (e1 & 1)
+                                       dval(&rv) *= tinytens[j];
+                       /* The last multiplication could underflow. */
+                       dval(&rv0) = dval(&rv);
+                       dval(&rv) *= tinytens[j];
+                       if (!dval(&rv)) {
+                               dval(&rv) = 2.*dval(&rv0);
+                               dval(&rv) *= tinytens[j];
+#endif
+                               if (!dval(&rv)) {
+ undfl:
+                                       dval(&rv) = 0.;
+                                       goto range_err;
+                                       }
+#ifndef Avoid_Underflow
+                               word0(&rv) = Tiny0;
+                               word1(&rv) = Tiny1;
+                               /* The refinement below will clean
+                                * this approximation up.
+                                */
+                               }
+#endif
+                       }
+               }
+
+       /* Now the hard part -- adjusting rv to the correct value.*/
+
+       /* Put digits into bd: true value = bd * 10^e */
+
+       bc.nd = nd - nz1;
+#ifndef NO_STRTOD_BIGCOMP
+       bc.nd0 = nd0;   /* Only needed if nd > strtod_diglim, but done here */
+                       /* to silence an erroneous warning about bc.nd0 */
+                       /* possibly not being initialized. */
+       if (nd > strtod_diglim) {
+               /* ASSERT(strtod_diglim >= 18); 18 == one more than the */
+               /* minimum number of decimal digits to distinguish double values */
+               /* in IEEE arithmetic. */
+               i = j = 18;
+               if (i > nd0)
+                       j += bc.dplen;
+               for(;;) {
+                       if (--j < bc.dp1 && j >= bc.dp0)
+                               j = bc.dp0 - 1;
+                       if (s0[j] != '0')
+                               break;
+                       --i;
+                       }
+               e += nd - i;
+               nd = i;
+               if (nd0 > nd)
+                       nd0 = nd;
+               if (nd < 9) { /* must recompute y */
+                       y = 0;
+                       for(i = 0; i < nd0; ++i)
+                               y = 10*y + s0[i] - '0';
+                       for(j = bc.dp1; i < nd; ++i)
+                               y = 10*y + s0[j++] - '0';
+                       }
+               }
+#endif
+       bd0 = s2b(s0, nd0, nd, y, bc.dplen);
+
+       for(;;) {
+               bd = Balloc(bd0->k);
+               Bcopy(bd, bd0);
+               bb = d2b(&rv, &bbe, &bbbits);   /* rv = bb * 2^bbe */
+               bs = i2b(1);
+
+               if (e >= 0) {
+                       bb2 = bb5 = 0;
+                       bd2 = bd5 = e;
+                       }
+               else {
+                       bb2 = bb5 = -e;
+                       bd2 = bd5 = 0;
+                       }
+               if (bbe >= 0)
+                       bb2 += bbe;
+               else
+                       bd2 -= bbe;
+               bs2 = bb2;
+#ifdef Honor_FLT_ROUNDS
+               if (bc.rounding != 1)
+                       bs2++;
+#endif
+#ifdef Avoid_Underflow
+               Lsb = LSB;
+               Lsb1 = 0;
+               j = bbe - bc.scale;
+               i = j + bbbits - 1;     /* logb(rv) */
+               j = P + 1 - bbbits;
+               if (i < Emin) { /* denormal */
+                       i = Emin - i;
+                       j -= i;
+                       if (i < 32)
+                               Lsb <<= i;
+                       else if (i < 52)
+                               Lsb1 = Lsb << (i-32);
+                       else
+                               Lsb1 = Exp_mask;
+                       }
+#else /*Avoid_Underflow*/
+#ifdef Sudden_Underflow
+#ifdef IBM
+               j = 1 + 4*P - 3 - bbbits + ((bbe + bbbits - 1) & 3);
+#else
+               j = P + 1 - bbbits;
+#endif
+#else /*Sudden_Underflow*/
+               j = bbe;
+               i = j + bbbits - 1;     /* logb(rv) */
+               if (i < Emin)   /* denormal */
+                       j += P - Emin;
+               else
+                       j = P + 1 - bbbits;
+#endif /*Sudden_Underflow*/
+#endif /*Avoid_Underflow*/
+               bb2 += j;
+               bd2 += j;
+#ifdef Avoid_Underflow
+               bd2 += bc.scale;
+#endif
+               i = bb2 < bd2 ? bb2 : bd2;
+               if (i > bs2)
+                       i = bs2;
+               if (i > 0) {
+                       bb2 -= i;
+                       bd2 -= i;
+                       bs2 -= i;
+                       }
+               if (bb5 > 0) {
+                       bs = pow5mult(bs, bb5);
+                       bb1 = mult(bs, bb);
+                       Bfree(bb);
+                       bb = bb1;
+                       }
+               if (bb2 > 0)
+                       bb = lshift(bb, bb2);
+               if (bd5 > 0)
+                       bd = pow5mult(bd, bd5);
+               if (bd2 > 0)
+                       bd = lshift(bd, bd2);
+               if (bs2 > 0)
+                       bs = lshift(bs, bs2);
+               delta = diff(bb, bd);
+               bc.dsign = delta->sign;
+               delta->sign = 0;
+               i = cmp(delta, bs);
+#ifndef NO_STRTOD_BIGCOMP /*{*/
+               if (bc.nd > nd && i <= 0) {
+                       if (bc.dsign) {
+                               /* Must use bigcomp(). */
+                               req_bigcomp = 1;
+                               break;
+                               }
+#ifdef Honor_FLT_ROUNDS
+                       if (bc.rounding != 1) {
+                               if (i < 0) {
+                                       req_bigcomp = 1;
+                                       break;
+                                       }
+                               }
+                       else
+#endif
+                               i = -1; /* Discarded digits make delta smaller. */
+                       }
+#endif /*}*/
+#ifdef Honor_FLT_ROUNDS /*{*/
+               if (bc.rounding != 1) {
+                       if (i < 0) {
+                               /* Error is less than an ulp */
+                               if (!delta->x[0] && delta->wds <= 1) {
+                                       /* exact */
+#ifdef SET_INEXACT
+                                       bc.inexact = 0;
+#endif
+                                       break;
+                                       }
+                               if (bc.rounding) {
+                                       if (bc.dsign) {
+                                               adj.d = 1.;
+                                               goto apply_adj;
+                                               }
+                                       }
+                               else if (!bc.dsign) {
+                                       adj.d = -1.;
+                                       if (!word1(&rv)
+                                        && !(word0(&rv) & Frac_mask)) {
+                                               y = word0(&rv) & Exp_mask;
+#ifdef Avoid_Underflow
+                                               if (!bc.scale || y > 2*P*Exp_msk1)
+#else
+                                               if (y)
+#endif
+                                                 {
+                                                 delta = lshift(delta,Log2P);
+                                                 if (cmp(delta, bs) <= 0)
+                                                       adj.d = -0.5;
+                                                 }
+                                               }
+ apply_adj:
+#ifdef Avoid_Underflow /*{*/
+                                       if (bc.scale && (y = word0(&rv) & Exp_mask)
+                                               <= 2*P*Exp_msk1)
+                                         word0(&adj) += (2*P+1)*Exp_msk1 - y;
+#else
+#ifdef Sudden_Underflow
+                                       if ((word0(&rv) & Exp_mask) <=
+                                                       P*Exp_msk1) {
+                                               word0(&rv) += P*Exp_msk1;
+                                               dval(&rv) += adj.d*ulp(dval(&rv));
+                                               word0(&rv) -= P*Exp_msk1;
+                                               }
+                                       else
+#endif /*Sudden_Underflow*/
+#endif /*Avoid_Underflow}*/
+                                       dval(&rv) += adj.d*ulp(&rv);
+                                       }
+                               break;
+                               }
+                       adj.d = ratio(delta, bs);
+                       if (adj.d < 1.)
+                               adj.d = 1.;
+                       if (adj.d <= 0x7ffffffe) {
+                               /* adj = rounding ? ceil(adj) : floor(adj); */
+                               y = adj.d;
+                               if (y != adj.d) {
+                                       if (!((bc.rounding>>1) ^ bc.dsign))
+                                               y++;
+                                       adj.d = y;
+                                       }
+                               }
+#ifdef Avoid_Underflow /*{*/
+                       if (bc.scale && (y = word0(&rv) & Exp_mask) <= 2*P*Exp_msk1)
+                               word0(&adj) += (2*P+1)*Exp_msk1 - y;
+#else
+#ifdef Sudden_Underflow
+                       if ((word0(&rv) & Exp_mask) <= P*Exp_msk1) {
+                               word0(&rv) += P*Exp_msk1;
+                               adj.d *= ulp(dval(&rv));
+                               if (bc.dsign)
+                                       dval(&rv) += adj.d;
+                               else
+                                       dval(&rv) -= adj.d;
+                               word0(&rv) -= P*Exp_msk1;
+                               goto cont;
+                               }
+#endif /*Sudden_Underflow*/
+#endif /*Avoid_Underflow}*/
+                       adj.d *= ulp(&rv);
+                       if (bc.dsign) {
+                               if (word0(&rv) == Big0 && word1(&rv) == Big1)
+                                       goto ovfl;
+                               dval(&rv) += adj.d;
+                               }
+                       else
+                               dval(&rv) -= adj.d;
+                       goto cont;
+                       }
+#endif /*}Honor_FLT_ROUNDS*/
+
+               if (i < 0) {
+                       /* Error is less than half an ulp -- check for
+                        * special case of mantissa a power of two.
+                        */
+                       if (bc.dsign || word1(&rv) || word0(&rv) & Bndry_mask
+#ifdef IEEE_Arith /*{*/
+#ifdef Avoid_Underflow
+                        || (word0(&rv) & Exp_mask) <= (2*P+1)*Exp_msk1
+#else
+                        || (word0(&rv) & Exp_mask) <= Exp_msk1
+#endif
+#endif /*}*/
+                               ) {
+#ifdef SET_INEXACT
+                               if (!delta->x[0] && delta->wds <= 1)
+                                       bc.inexact = 0;
+#endif
+                               break;
+                               }
+                       if (!delta->x[0] && delta->wds <= 1) {
+                               /* exact result */
+#ifdef SET_INEXACT
+                               bc.inexact = 0;
+#endif
+                               break;
+                               }
+                       delta = lshift(delta,Log2P);
+                       if (cmp(delta, bs) > 0)
+                               goto drop_down;
+                       break;
+                       }
+               if (i == 0) {
+                       /* exactly half-way between */
+                       if (bc.dsign) {
+                               if ((word0(&rv) & Bndry_mask1) == Bndry_mask1
+                                &&  word1(&rv) == (
+#ifdef Avoid_Underflow
+                       (bc.scale && (y = word0(&rv) & Exp_mask) <= 2*P*Exp_msk1)
+               ? (0xffffffff & (0xffffffff << (2*P+1-(y>>Exp_shift)))) :
+#endif
+                                                  0xffffffff)) {
+                                       /*boundary case -- increment exponent*/
+                                       if (word0(&rv) == Big0 && word1(&rv) == Big1)
+                                               goto ovfl;
+                                       word0(&rv) = (word0(&rv) & Exp_mask)
+                                               + Exp_msk1
+#ifdef IBM
+                                               | Exp_msk1 >> 4
+#endif
+                                               ;
+                                       word1(&rv) = 0;
+#ifdef Avoid_Underflow
+                                       bc.dsign = 0;
+#endif
+                                       break;
+                                       }
+                               }
+                       else if (!(word0(&rv) & Bndry_mask) && !word1(&rv)) {
+ drop_down:
+                               /* boundary case -- decrement exponent */
+#ifdef Sudden_Underflow /*{{*/
+                               L = word0(&rv) & Exp_mask;
+#ifdef IBM
+                               if (L <  Exp_msk1)
+#else
+#ifdef Avoid_Underflow
+                               if (L <= (bc.scale ? (2*P+1)*Exp_msk1 : Exp_msk1))
+#else
+                               if (L <= Exp_msk1)
+#endif /*Avoid_Underflow*/
+#endif /*IBM*/
+                                       {
+                                       if (bc.nd >nd) {
+                                               bc.uflchk = 1;
+                                               break;
+                                               }
+                                       goto undfl;
+                                       }
+                               L -= Exp_msk1;
+#else /*Sudden_Underflow}{*/
+#ifdef Avoid_Underflow
+                               if (bc.scale) {
+                                       L = word0(&rv) & Exp_mask;
+                                       if (L <= (2*P+1)*Exp_msk1) {
+                                               if (L > (P+2)*Exp_msk1)
+                                                       /* round even ==> */
+                                                       /* accept rv */
+                                                       break;
+                                               /* rv = smallest denormal */
+                                               if (bc.nd >nd) {
+                                                       bc.uflchk = 1;
+                                                       break;
+                                                       }
+                                               goto undfl;
+                                               }
+                                       }
+#endif /*Avoid_Underflow*/
+                               L = (word0(&rv) & Exp_mask) - Exp_msk1;
+#endif /*Sudden_Underflow}}*/
+                               word0(&rv) = L | Bndry_mask1;
+                               word1(&rv) = 0xffffffff;
+#ifdef IBM
+                               goto cont;
+#else
+#ifndef NO_STRTOD_BIGCOMP
+                               if (bc.nd > nd)
+                                       goto cont;
+#endif
+                               break;
+#endif
+                               }
+#ifndef ROUND_BIASED
+#ifdef Avoid_Underflow
+                       if (Lsb1) {
+                               if (!(word0(&rv) & Lsb1))
+                                       break;
+                               }
+                       else if (!(word1(&rv) & Lsb))
+                               break;
+#else
+                       if (!(word1(&rv) & LSB))
+                               break;
+#endif
+#endif
+                       if (bc.dsign)
+#ifdef Avoid_Underflow
+                               dval(&rv) += sulp(&rv, &bc);
+#else
+                               dval(&rv) += ulp(&rv);
+#endif
+#ifndef ROUND_BIASED
+                       else {
+#ifdef Avoid_Underflow
+                               dval(&rv) -= sulp(&rv, &bc);
+#else
+                               dval(&rv) -= ulp(&rv);
+#endif
+#ifndef Sudden_Underflow
+                               if (!dval(&rv)) {
+                                       if (bc.nd >nd) {
+                                               bc.uflchk = 1;
+                                               break;
+                                               }
+                                       goto undfl;
+                                       }
+#endif
+                               }
+#ifdef Avoid_Underflow
+                       bc.dsign = 1 - bc.dsign;
+#endif
+#endif
+                       break;
+                       }
+               if ((aadj = ratio(delta, bs)) <= 2.) {
+                       if (bc.dsign)
+                               aadj = aadj1 = 1.;
+                       else if (word1(&rv) || word0(&rv) & Bndry_mask) {
+#ifndef Sudden_Underflow
+                               if (word1(&rv) == Tiny1 && !word0(&rv)) {
+                                       if (bc.nd >nd) {
+                                               bc.uflchk = 1;
+                                               break;
+                                               }
+                                       goto undfl;
+                                       }
+#endif
+                               aadj = 1.;
+                               aadj1 = -1.;
+                               }
+                       else {
+                               /* special case -- power of FLT_RADIX to be */
+                               /* rounded down... */
+
+                               if (aadj < 2./FLT_RADIX)
+                                       aadj = 1./FLT_RADIX;
+                               else
+                                       aadj *= 0.5;
+                               aadj1 = -aadj;
+                               }
+                       }
+               else {
+                       aadj *= 0.5;
+                       aadj1 = bc.dsign ? aadj : -aadj;
+#ifdef Check_FLT_ROUNDS
+                       switch(bc.rounding) {
+                               case 2: /* towards +infinity */
+                                       aadj1 -= 0.5;
+                                       break;
+                               case 0: /* towards 0 */
+                               case 3: /* towards -infinity */
+                                       aadj1 += 0.5;
+                               }
+#else
+                       if (Flt_Rounds == 0)
+                               aadj1 += 0.5;
+#endif /*Check_FLT_ROUNDS*/
+                       }
+               y = word0(&rv) & Exp_mask;
+
+               /* Check for overflow */
+
+               if (y == Exp_msk1*(DBL_MAX_EXP+Bias-1)) {
+                       dval(&rv0) = dval(&rv);
+                       word0(&rv) -= P*Exp_msk1;
+                       adj.d = aadj1 * ulp(&rv);
+                       dval(&rv) += adj.d;
+                       if ((word0(&rv) & Exp_mask) >=
+                                       Exp_msk1*(DBL_MAX_EXP+Bias-P)) {
+                               if (word0(&rv0) == Big0 && word1(&rv0) == Big1)
+                                       goto ovfl;
+                               word0(&rv) = Big0;
+                               word1(&rv) = Big1;
+                               goto cont;
+                               }
+                       else
+                               word0(&rv) += P*Exp_msk1;
+                       }
+               else {
+#ifdef Avoid_Underflow
+                       if (bc.scale && y <= 2*P*Exp_msk1) {
+                               if (aadj <= 0x7fffffff) {
+                                       if ((z = aadj) <= 0)
+                                               z = 1;
+                                       aadj = z;
+                                       aadj1 = bc.dsign ? aadj : -aadj;
+                                       }
+                               dval(&aadj2) = aadj1;
+                               word0(&aadj2) += (2*P+1)*Exp_msk1 - y;
+                               aadj1 = dval(&aadj2);
+                               adj.d = aadj1 * ulp(&rv);
+                               dval(&rv) += adj.d;
+                               if (rv.d == 0.)
+#ifdef NO_STRTOD_BIGCOMP
+                                       goto undfl;
+#else
+                                       {
+                                       if (bc.nd > nd)
+                                               bc.dsign = 1;
+                                       break;
+                                       }
+#endif
+                               }
+                       else {
+                               adj.d = aadj1 * ulp(&rv);
+                               dval(&rv) += adj.d;
+                               }
+#else
+#ifdef Sudden_Underflow
+                       if ((word0(&rv) & Exp_mask) <= P*Exp_msk1) {
+                               dval(&rv0) = dval(&rv);
+                               word0(&rv) += P*Exp_msk1;
+                               adj.d = aadj1 * ulp(&rv);
+                               dval(&rv) += adj.d;
+#ifdef IBM
+                               if ((word0(&rv) & Exp_mask) <  P*Exp_msk1)
+#else
+                               if ((word0(&rv) & Exp_mask) <= P*Exp_msk1)
+#endif
+                                       {
+                                       if (word0(&rv0) == Tiny0
+                                        && word1(&rv0) == Tiny1) {
+                                               if (bc.nd >nd) {
+                                                       bc.uflchk = 1;
+                                                       break;
+                                                       }
+                                               goto undfl;
+                                               }
+                                       word0(&rv) = Tiny0;
+                                       word1(&rv) = Tiny1;
+                                       goto cont;
+                                       }
+                               else
+                                       word0(&rv) -= P*Exp_msk1;
+                               }
+                       else {
+                               adj.d = aadj1 * ulp(&rv);
+                               dval(&rv) += adj.d;
+                               }
+#else /*Sudden_Underflow*/
+                       /* Compute adj so that the IEEE rounding rules will
+                        * correctly round rv + adj in some half-way cases.
+                        * If rv * ulp(rv) is denormalized (i.e.,
+                        * y <= (P-1)*Exp_msk1), we must adjust aadj to avoid
+                        * trouble from bits lost to denormalization;
+                        * example: 1.2e-307 .
+                        */
+                       if (y <= (P-1)*Exp_msk1 && aadj > 1.) {
+                               aadj1 = (double)(int)(aadj + 0.5);
+                               if (!bc.dsign)
+                                       aadj1 = -aadj1;
+                               }
+                       adj.d = aadj1 * ulp(&rv);
+                       dval(&rv) += adj.d;
+#endif /*Sudden_Underflow*/
+#endif /*Avoid_Underflow*/
+                       }
+               z = word0(&rv) & Exp_mask;
+#ifndef SET_INEXACT
+               if (bc.nd == nd) {
+#ifdef Avoid_Underflow
+               if (!bc.scale)
+#endif
+               if (y == z) {
+                       /* Can we stop now? */
+                       L = (Long)aadj;
+                       aadj -= L;
+                       /* The tolerances below are conservative. */
+                       if (bc.dsign || word1(&rv) || word0(&rv) & Bndry_mask) {
+                               if (aadj < .4999999 || aadj > .5000001)
+                                       break;
+                               }
+                       else if (aadj < .4999999/FLT_RADIX)
+                               break;
+                       }
+               }
+#endif
+ cont:
+               Bfree(bb);
+               Bfree(bd);
+               Bfree(bs);
+               Bfree(delta);
+               }
+       Bfree(bb);
+       Bfree(bd);
+       Bfree(bs);
+       Bfree(bd0);
+       Bfree(delta);
+#ifndef NO_STRTOD_BIGCOMP
+       if (req_bigcomp) {
+               bd0 = 0;
+               bc.e0 += nz1;
+               bigcomp(&rv, s0, &bc);
+               y = word0(&rv) & Exp_mask;
+               if (y == Exp_mask)
+                       goto ovfl;
+               if (y == 0 && rv.d == 0.)
+                       goto undfl;
+               }
+#endif
+#ifdef SET_INEXACT
+       if (bc.inexact) {
+               if (!oldinexact) {
+                       word0(&rv0) = Exp_1 + (70 << Exp_shift);
+                       word1(&rv0) = 0;
+                       dval(&rv0) += 1.;
+                       }
+               }
+       else if (!oldinexact)
+               clear_inexact();
+#endif
+#ifdef Avoid_Underflow
+       if (bc.scale) {
+               word0(&rv0) = Exp_1 - 2*P*Exp_msk1;
+               word1(&rv0) = 0;
+               dval(&rv) *= dval(&rv0);
+#ifndef NO_ERRNO
+               /* try to avoid the bug of testing an 8087 register value */
+#ifdef IEEE_Arith
+               if (!(word0(&rv) & Exp_mask))
+#else
+               if (word0(&rv) == 0 && word1(&rv) == 0)
+#endif
+                       errno = ERANGE;
+#endif
+               }
+#endif /* Avoid_Underflow */
+#ifdef SET_INEXACT
+       if (bc.inexact && !(word0(&rv) & Exp_mask)) {
+               /* set underflow bit */
+               dval(&rv0) = 1e-300;
+               dval(&rv0) *= dval(&rv0);
+               }
+#endif
+ ret:
+       if (se)
+               *se = (char *)s;
+       return sign ? -dval(&rv) : dval(&rv);
+       }
+
+#ifndef MULTIPLE_THREADS
+ static char *dtoa_result;
+#endif
+
+ static char *
+#ifdef KR_headers
+rv_alloc(i) int i;
+#else
+rv_alloc(int i)
+#endif
+{
+       int j, k, *r;
+
+       j = sizeof(ULong);
+       for(k = 0;
+               sizeof(Bigint) - sizeof(ULong) - sizeof(int) + j <= i;
+               j <<= 1)
+                       k++;
+       r = (int*)Balloc(k);
+       *r = k;
+       return
+#ifndef MULTIPLE_THREADS
+       dtoa_result =
+#endif
+               (char *)(r+1);
+       }
+
+ static char *
+#ifdef KR_headers
+nrv_alloc(s, rve, n) char *s, **rve; int n;
+#else
+nrv_alloc(const char *s, char **rve, int n)
+#endif
+{
+       char *rv, *t;
+
+       t = rv = rv_alloc(n);
+       while((*t = *s++)) t++;
+       if (rve)
+               *rve = t;
+       return rv;
+       }
+
+/* freedtoa(s) must be used to free values s returned by dtoa
+ * when MULTIPLE_THREADS is #defined.  It should be used in all cases,
+ * but for consistency with earlier versions of dtoa, it is optional
+ * when MULTIPLE_THREADS is not defined.
+ */
+
+ void
+#ifdef KR_headers
+freedtoa(s) char *s;
+#else
+freedtoa(char *s)
+#endif
+{
+       Bigint *b = (Bigint *)((int *)s - 1);
+       b->maxwds = 1 << (b->k = *(int*)b);
+       Bfree(b);
+#ifndef MULTIPLE_THREADS
+       if (s == dtoa_result)
+               dtoa_result = 0;
+#endif
+       }
+
+/* dtoa for IEEE arithmetic (dmg): convert double to ASCII string.
+ *
+ * Inspired by "How to Print Floating-Point Numbers Accurately" by
+ * Guy L. Steele, Jr. and Jon L. White [Proc. ACM SIGPLAN '90, pp. 112-126].
+ *
+ * Modifications:
+ *     1. Rather than iterating, we use a simple numeric overestimate
+ *        to determine k = floor(log10(d)).  We scale relevant
+ *        quantities using O(log2(k)) rather than O(k) multiplications.
+ *     2. For some modes > 2 (corresponding to ecvt and fcvt), we don't
+ *        try to generate digits strictly left to right.  Instead, we
+ *        compute with fewer bits and propagate the carry if necessary
+ *        when rounding the final digit up.  This is often faster.
+ *     3. Under the assumption that input will be rounded nearest,
+ *        mode 0 renders 1e23 as 1e23 rather than 9.999999999999999e22.
+ *        That is, we allow equality in stopping tests when the
+ *        round-nearest rule will give the same floating-point value
+ *        as would satisfaction of the stopping test with strict
+ *        inequality.
+ *     4. We remove common factors of powers of 2 from relevant
+ *        quantities.
+ *     5. When converting floating-point integers less than 1e16,
+ *        we use floating-point arithmetic rather than resorting
+ *        to multiple-precision integers.
+ *     6. When asked to produce fewer than 15 digits, we first try
+ *        to get by with floating-point arithmetic; we resort to
+ *        multiple-precision integer arithmetic only if we cannot
+ *        guarantee that the floating-point calculation has given
+ *        the correctly rounded result.  For k requested digits and
+ *        "uniformly" distributed input, the probability is
+ *        something like 10^(k-15) that we must resort to the Long
+ *        calculation.
+ */
+
+ char *
+dtoa
+#ifdef KR_headers
+       (dd, mode, ndigits, decpt, sign, rve)
+       double dd; int mode, ndigits, *decpt, *sign; char **rve;
+#else
+       (double dd, int mode, int ndigits, int *decpt, int *sign, char **rve)
+#endif
+{
+ /*    Arguments ndigits, decpt, sign are similar to those
+       of ecvt and fcvt; trailing zeros are suppressed from
+       the returned string.  If not null, *rve is set to point
+       to the end of the return value.  If d is +-Infinity or NaN,
+       then *decpt is set to 9999.
+
+       mode:
+               0 ==> shortest string that yields d when read in
+                       and rounded to nearest.
+               1 ==> like 0, but with Steele & White stopping rule;
+                       e.g. with IEEE P754 arithmetic , mode 0 gives
+                       1e23 whereas mode 1 gives 9.999999999999999e22.
+               2 ==> max(1,ndigits) significant digits.  This gives a
+                       return value similar to that of ecvt, except
+                       that trailing zeros are suppressed.
+               3 ==> through ndigits past the decimal point.  This
+                       gives a return value similar to that from fcvt,
+                       except that trailing zeros are suppressed, and
+                       ndigits can be negative.
+               4,5 ==> similar to 2 and 3, respectively, but (in
+                       round-nearest mode) with the tests of mode 0 to
+                       possibly return a shorter string that rounds to d.
+                       With IEEE arithmetic and compilation with
+                       -DHonor_FLT_ROUNDS, modes 4 and 5 behave the same
+                       as modes 2 and 3 when FLT_ROUNDS != 1.
+               6-9 ==> Debugging modes similar to mode - 4:  don't try
+                       fast floating-point estimate (if applicable).
+
+               Values of mode other than 0-9 are treated as mode 0.
+
+               Sufficient space is allocated to the return value
+               to hold the suppressed trailing zeros.
+       */
+
+       int bbits, b2, b5, be, dig, i, ieps, ilim, ilim0, ilim1,
+               j, j1, k, k0, k_check, leftright, m2, m5, s2, s5,
+               spec_case, try_quick;
+       Long L;
+#ifndef Sudden_Underflow
+       int denorm;
+       ULong x;
+#endif
+       Bigint *b, *b1, *delta, *mlo, *mhi, *S;
+       U d2, eps, u;
+       double ds;
+       char *s, *s0;
+#ifndef No_leftright
+#ifdef IEEE_Arith
+       U eps1;
+#endif
+#endif
+#ifdef SET_INEXACT
+       int inexact, oldinexact;
+#endif
+#ifdef Honor_FLT_ROUNDS /*{*/
+       int Rounding;
+#ifdef Trust_FLT_ROUNDS /*{{ only define this if FLT_ROUNDS really works! */
+       Rounding = Flt_Rounds;
+#else /*}{*/
+       Rounding = 1;
+       switch(fegetround()) {
+         case FE_TOWARDZERO:   Rounding = 0; break;
+         case FE_UPWARD:       Rounding = 2; break;
+         case FE_DOWNWARD:     Rounding = 3;
+         }
+#endif /*}}*/
+#endif /*}*/
+
+#ifndef MULTIPLE_THREADS
+       if (dtoa_result) {
+               freedtoa(dtoa_result);
+               dtoa_result = 0;
+               }
+#endif
+
+       u.d = dd;
+       if (word0(&u) & Sign_bit) {
+               /* set sign for everything, including 0's and NaNs */
+               *sign = 1;
+               word0(&u) &= ~Sign_bit; /* clear sign bit */
+               }
+       else
+               *sign = 0;
+
+#if defined(IEEE_Arith) + defined(VAX)
+#ifdef IEEE_Arith
+       if ((word0(&u) & Exp_mask) == Exp_mask)
+#else
+       if (word0(&u)  == 0x8000)
+#endif
+               {
+               /* Infinity or NaN */
+               *decpt = 9999;
+#ifdef IEEE_Arith
+               if (!word1(&u) && !(word0(&u) & 0xfffff))
+                       return nrv_alloc("Infinity", rve, 8);
+#endif
+               return nrv_alloc("NaN", rve, 3);
+               }
+#endif
+#ifdef IBM
+       dval(&u) += 0; /* normalize */
+#endif
+       if (!dval(&u)) {
+               *decpt = 1;
+               return nrv_alloc("0", rve, 1);
+               }
+
+#ifdef SET_INEXACT
+       try_quick = oldinexact = get_inexact();
+       inexact = 1;
+#endif
+#ifdef Honor_FLT_ROUNDS
+       if (Rounding >= 2) {
+               if (*sign)
+                       Rounding = Rounding == 2 ? 0 : 2;
+               else
+                       if (Rounding != 2)
+                               Rounding = 0;
+               }
+#endif
+
+       b = d2b(&u, &be, &bbits);
+#ifdef Sudden_Underflow
+       i = (int)(word0(&u) >> Exp_shift1 & (Exp_mask>>Exp_shift1));
+#else
+       if ((i = (int)(word0(&u) >> Exp_shift1 & (Exp_mask>>Exp_shift1)))) {
+#endif
+               dval(&d2) = dval(&u);
+               word0(&d2) &= Frac_mask1;
+               word0(&d2) |= Exp_11;
+#ifdef IBM
+               if (j = 11 - hi0bits(word0(&d2) & Frac_mask))
+                       dval(&d2) /= 1 << j;
+#endif
+
+               /* log(x)       ~=~ log(1.5) + (x-1.5)/1.5
+                * log10(x)      =  log(x) / log(10)
+                *              ~=~ log(1.5)/log(10) + (x-1.5)/(1.5*log(10))
+                * log10(d) = (i-Bias)*log(2)/log(10) + log10(d2)
+                *
+                * This suggests computing an approximation k to log10(d) by
+                *
+                * k = (i - Bias)*0.301029995663981
+                *      + ( (d2-1.5)*0.289529654602168 + 0.176091259055681 );
+                *
+                * We want k to be too large rather than too small.
+                * The error in the first-order Taylor series approximation
+                * is in our favor, so we just round up the constant enough
+                * to compensate for any error in the multiplication of
+                * (i - Bias) by 0.301029995663981; since |i - Bias| <= 1077,
+                * and 1077 * 0.30103 * 2^-52 ~=~ 7.2e-14,
+                * adding 1e-13 to the constant term more than suffices.
+                * Hence we adjust the constant term to 0.1760912590558.
+                * (We could get a more accurate k by invoking log10,
+                *  but this is probably not worthwhile.)
+                */
+
+               i -= Bias;
+#ifdef IBM
+               i <<= 2;
+               i += j;
+#endif
+#ifndef Sudden_Underflow
+               denorm = 0;
+               }
+       else {
+               /* d is denormalized */
+
+               i = bbits + be + (Bias + (P-1) - 1);
+               x = i > 32  ? word0(&u) << (64 - i) | word1(&u) >> (i - 32)
+                           : word1(&u) << (32 - i);
+               dval(&d2) = x;
+               word0(&d2) -= 31*Exp_msk1; /* adjust exponent */
+               i -= (Bias + (P-1) - 1) + 1;
+               denorm = 1;
+               }
+#endif
+       ds = (dval(&d2)-1.5)*0.289529654602168 + 0.1760912590558 + i*0.301029995663981;
+       k = (int)ds;
+       if (ds < 0. && ds != k)
+               k--;    /* want k = floor(ds) */
+       k_check = 1;
+       if (k >= 0 && k <= Ten_pmax) {
+               if (dval(&u) < tens[k])
+                       k--;
+               k_check = 0;
+               }
+       j = bbits - i - 1;
+       if (j >= 0) {
+               b2 = 0;
+               s2 = j;
+               }
+       else {
+               b2 = -j;
+               s2 = 0;
+               }
+       if (k >= 0) {
+               b5 = 0;
+               s5 = k;
+               s2 += k;
+               }
+       else {
+               b2 -= k;
+               b5 = -k;
+               s5 = 0;
+               }
+       if (mode < 0 || mode > 9)
+               mode = 0;
+
+#ifndef SET_INEXACT
+#ifdef Check_FLT_ROUNDS
+       try_quick = Rounding == 1;
+#else
+       try_quick = 1;
+#endif
+#endif /*SET_INEXACT*/
+
+       if (mode > 5) {
+               mode -= 4;
+               try_quick = 0;
+               }
+       leftright = 1;
+       ilim = ilim1 = -1;      /* Values for cases 0 and 1; done here to */
+                               /* silence erroneous "gcc -Wall" warning. */
+       switch(mode) {
+               case 0:
+               case 1:
+                       i = 18;
+                       ndigits = 0;
+                       break;
+               case 2:
+                       leftright = 0;
+                       /* no break */
+               case 4:
+                       if (ndigits <= 0)
+                               ndigits = 1;
+                       ilim = ilim1 = i = ndigits;
+                       break;
+               case 3:
+                       leftright = 0;
+                       /* no break */
+               case 5:
+                       i = ndigits + k + 1;
+                       ilim = i;
+                       ilim1 = i - 1;
+                       if (i <= 0)
+                               i = 1;
+               }
+       s = s0 = rv_alloc(i);
+
+#ifdef Honor_FLT_ROUNDS
+       if (mode > 1 && Rounding != 1)
+               leftright = 0;
+#endif
+
+       if (ilim >= 0 && ilim <= Quick_max && try_quick) {
+
+               /* Try to get by with floating-point arithmetic. */
+
+               i = 0;
+               dval(&d2) = dval(&u);
+               k0 = k;
+               ilim0 = ilim;
+               ieps = 2; /* conservative */
+               if (k > 0) {
+                       ds = tens[k&0xf];
+                       j = k >> 4;
+                       if (j & Bletch) {
+                               /* prevent overflows */
+                               j &= Bletch - 1;
+                               dval(&u) /= bigtens[n_bigtens-1];
+                               ieps++;
+                               }
+                       for(; j; j >>= 1, i++)
+                               if (j & 1) {
+                                       ieps++;
+                                       ds *= bigtens[i];
+                                       }
+                       dval(&u) /= ds;
+                       }
+               else if ((j1 = -k)) {
+                       dval(&u) *= tens[j1 & 0xf];
+                       for(j = j1 >> 4; j; j >>= 1, i++)
+                               if (j & 1) {
+                                       ieps++;
+                                       dval(&u) *= bigtens[i];
+                                       }
+                       }
+               if (k_check && dval(&u) < 1. && ilim > 0) {
+                       if (ilim1 <= 0)
+                               goto fast_failed;
+                       ilim = ilim1;
+                       k--;
+                       dval(&u) *= 10.;
+                       ieps++;
+                       }
+               dval(&eps) = ieps*dval(&u) + 7.;
+               word0(&eps) -= (P-1)*Exp_msk1;
+               if (ilim == 0) {
+                       S = mhi = 0;
+                       dval(&u) -= 5.;
+                       if (dval(&u) > dval(&eps))
+                               goto one_digit;
+                       if (dval(&u) < -dval(&eps))
+                               goto no_digits;
+                       goto fast_failed;
+                       }
+#ifndef No_leftright
+               if (leftright) {
+                       /* Use Steele & White method of only
+                        * generating digits needed.
+                        */
+                       dval(&eps) = 0.5/tens[ilim-1] - dval(&eps);
+#ifdef IEEE_Arith
+                       if (k0 < 0 && j1 >= 307) {
+                               eps1.d = 1.01e256; /* 1.01 allows roundoff in the next few lines */
+                               word0(&eps1) -= Exp_msk1 * (Bias+P-1);
+                               dval(&eps1) *= tens[j1 & 0xf];
+                               for(i = 0, j = (j1-256) >> 4; j; j >>= 1, i++)
+                                       if (j & 1)
+                                               dval(&eps1) *= bigtens[i];
+                               if (eps.d < eps1.d)
+                                       eps.d = eps1.d;
+                               }
+#endif
+                       for(i = 0;;) {
+                               L = dval(&u);
+                               dval(&u) -= L;
+                               *s++ = '0' + (int)L;
+                               if (1. - dval(&u) < dval(&eps))
+                                       goto bump_up;
+                               if (dval(&u) < dval(&eps))
+                                       goto ret1;
+                               if (++i >= ilim)
+                                       break;
+                               dval(&eps) *= 10.;
+                               dval(&u) *= 10.;
+                               }
+                       }
+               else {
+#endif
+                       /* Generate ilim digits, then fix them up. */
+                       dval(&eps) *= tens[ilim-1];
+                       for(i = 1;; i++, dval(&u) *= 10.) {
+                               L = (Long)(dval(&u));
+                               if (!(dval(&u) -= L))
+                                       ilim = i;
+                               *s++ = '0' + (int)L;
+                               if (i == ilim) {
+                                       if (dval(&u) > 0.5 + dval(&eps))
+                                               goto bump_up;
+                                       else if (dval(&u) < 0.5 - dval(&eps)) {
+                                               while(*--s == '0');
+                                               s++;
+                                               goto ret1;
+                                               }
+                                       break;
+                                       }
+                               }
+#ifndef No_leftright
+                       }
+#endif
+ fast_failed:
+               s = s0;
+               dval(&u) = dval(&d2);
+               k = k0;
+               ilim = ilim0;
+               }
+
+       /* Do we have a "small" integer? */
+
+       if (be >= 0 && k <= Int_max) {
+               /* Yes. */
+               ds = tens[k];
+               if (ndigits < 0 && ilim <= 0) {
+                       S = mhi = 0;
+                       if (ilim < 0 || dval(&u) <= 5*ds)
+                               goto no_digits;
+                       goto one_digit;
+                       }
+               for(i = 1;; i++, dval(&u) *= 10.) {
+                       L = (Long)(dval(&u) / ds);
+                       dval(&u) -= L*ds;
+#ifdef Check_FLT_ROUNDS
+                       /* If FLT_ROUNDS == 2, L will usually be high by 1 */
+                       if (dval(&u) < 0) {
+                               L--;
+                               dval(&u) += ds;
+                               }
+#endif
+                       *s++ = '0' + (int)L;
+                       if (!dval(&u)) {
+#ifdef SET_INEXACT
+                               inexact = 0;
+#endif
+                               break;
+                               }
+                       if (i == ilim) {
+#ifdef Honor_FLT_ROUNDS
+                               if (mode > 1)
+                               switch(Rounding) {
+                                 case 0: goto ret1;
+                                 case 2: goto bump_up;
+                                 }
+#endif
+                               dval(&u) += dval(&u);
+#ifdef ROUND_BIASED
+                               if (dval(&u) >= ds)
+#else
+                               if (dval(&u) > ds || (dval(&u) == ds && L & 1))
+#endif
+                                       {
+ bump_up:
+                                       while(*--s == '9')
+                                               if (s == s0) {
+                                                       k++;
+                                                       *s = '0';
+                                                       break;
+                                                       }
+                                       ++*s++;
+                                       }
+                               break;
+                               }
+                       }
+               goto ret1;
+               }
+
+       m2 = b2;
+       m5 = b5;
+       mhi = mlo = 0;
+       if (leftright) {
+               i =
+#ifndef Sudden_Underflow
+                       denorm ? be + (Bias + (P-1) - 1 + 1) :
+#endif
+#ifdef IBM
+                       1 + 4*P - 3 - bbits + ((bbits + be - 1) & 3);
+#else
+                       1 + P - bbits;
+#endif
+               b2 += i;
+               s2 += i;
+               mhi = i2b(1);
+               }
+       if (m2 > 0 && s2 > 0) {
+               i = m2 < s2 ? m2 : s2;
+               b2 -= i;
+               m2 -= i;
+               s2 -= i;
+               }
+       if (b5 > 0) {
+               if (leftright) {
+                       if (m5 > 0) {
+                               mhi = pow5mult(mhi, m5);
+                               b1 = mult(mhi, b);
+                               Bfree(b);
+                               b = b1;
+                               }
+                       if ((j = b5 - m5))
+                               b = pow5mult(b, j);
+                       }
+               else
+                       b = pow5mult(b, b5);
+               }
+       S = i2b(1);
+       if (s5 > 0)
+               S = pow5mult(S, s5);
+
+       /* Check for special case that d is a normalized power of 2. */
+
+       spec_case = 0;
+       if ((mode < 2 || leftright)
+#ifdef Honor_FLT_ROUNDS
+                       && Rounding == 1
+#endif
+                               ) {
+               if (!word1(&u) && !(word0(&u) & Bndry_mask)
+#ifndef Sudden_Underflow
+                && word0(&u) & (Exp_mask & ~Exp_msk1)
+#endif
+                               ) {
+                       /* The special case */
+                       b2 += Log2P;
+                       s2 += Log2P;
+                       spec_case = 1;
+                       }
+               }
+
+       /* Arrange for convenient computation of quotients:
+        * shift left if necessary so divisor has 4 leading 0 bits.
+        *
+        * Perhaps we should just compute leading 28 bits of S once
+        * and for all and pass them and a shift to quorem, so it
+        * can do shifts and ors to compute the numerator for q.
+        */
+       i = dshift(S, s2);
+       b2 += i;
+       m2 += i;
+       s2 += i;
+       if (b2 > 0)
+               b = lshift(b, b2);
+       if (s2 > 0)
+               S = lshift(S, s2);
+       if (k_check) {
+               if (cmp(b,S) < 0) {
+                       k--;
+                       b = multadd(b, 10, 0);  /* we botched the k estimate */
+                       if (leftright)
+                               mhi = multadd(mhi, 10, 0);
+                       ilim = ilim1;
+                       }
+               }
+       if (ilim <= 0 && (mode == 3 || mode == 5)) {
+               if (ilim < 0 || cmp(b,S = multadd(S,5,0)) <= 0) {
+                       /* no digits, fcvt style */
+ no_digits:
+                       k = -1 - ndigits;
+                       goto ret;
+                       }
+ one_digit:
+               *s++ = '1';
+               k++;
+               goto ret;
+               }
+       if (leftright) {
+               if (m2 > 0)
+                       mhi = lshift(mhi, m2);
+
+               /* Compute mlo -- check for special case
+                * that d is a normalized power of 2.
+                */
+
+               mlo = mhi;
+               if (spec_case) {
+                       mhi = Balloc(mhi->k);
+                       Bcopy(mhi, mlo);
+                       mhi = lshift(mhi, Log2P);
+                       }
+
+               for(i = 1;;i++) {
+                       dig = quorem(b,S) + '0';
+                       /* Do we yet have the shortest decimal string
+                        * that will round to d?
+                        */
+                       j = cmp(b, mlo);
+                       delta = diff(S, mhi);
+                       j1 = delta->sign ? 1 : cmp(b, delta);
+                       Bfree(delta);
+#ifndef ROUND_BIASED
+                       if (j1 == 0 && mode != 1 && !(word1(&u) & 1)
+#ifdef Honor_FLT_ROUNDS
+                               && Rounding >= 1
+#endif
+                                                                  ) {
+                               if (dig == '9')
+                                       goto round_9_up;
+                               if (j > 0)
+                                       dig++;
+#ifdef SET_INEXACT
+                               else if (!b->x[0] && b->wds <= 1)
+                                       inexact = 0;
+#endif
+                               *s++ = dig;
+                               goto ret;
+                               }
+#endif
+                       if (j < 0 || (j == 0 && mode != 1
+#ifndef ROUND_BIASED
+                                                       && !(word1(&u) & 1)
+#endif
+                                       )) {
+                               if (!b->x[0] && b->wds <= 1) {
+#ifdef SET_INEXACT
+                                       inexact = 0;
+#endif
+                                       goto accept_dig;
+                                       }
+#ifdef Honor_FLT_ROUNDS
+                               if (mode > 1)
+                                switch(Rounding) {
+                                 case 0: goto accept_dig;
+                                 case 2: goto keep_dig;
+                                 }
+#endif /*Honor_FLT_ROUNDS*/
+                               if (j1 > 0) {
+                                       b = lshift(b, 1);
+                                       j1 = cmp(b, S);
+#ifdef ROUND_BIASED
+                                       if (j1 >= 0 /*)*/
+#else
+                                       if ((j1 > 0 || (j1 == 0 && dig & 1))
+#endif
+                                       && dig++ == '9')
+                                               goto round_9_up;
+                                       }
+ accept_dig:
+                               *s++ = dig;
+                               goto ret;
+                               }
+                       if (j1 > 0) {
+#ifdef Honor_FLT_ROUNDS
+                               if (!Rounding)
+                                       goto accept_dig;
+#endif
+                               if (dig == '9') { /* possible if i == 1 */
+ round_9_up:
+                                       *s++ = '9';
+                                       goto roundoff;
+                                       }
+                               *s++ = dig + 1;
+                               goto ret;
+                               }
+#ifdef Honor_FLT_ROUNDS
+ keep_dig:
+#endif
+                       *s++ = dig;
+                       if (i == ilim)
+                               break;
+                       b = multadd(b, 10, 0);
+                       if (mlo == mhi)
+                               mlo = mhi = multadd(mhi, 10, 0);
+                       else {
+                               mlo = multadd(mlo, 10, 0);
+                               mhi = multadd(mhi, 10, 0);
+                               }
+                       }
+               }
+       else
+               for(i = 1;; i++) {
+                       *s++ = dig = quorem(b,S) + '0';
+                       if (!b->x[0] && b->wds <= 1) {
+#ifdef SET_INEXACT
+                               inexact = 0;
+#endif
+                               goto ret;
+                               }
+                       if (i >= ilim)
+                               break;
+                       b = multadd(b, 10, 0);
+                       }
+
+       /* Round off last digit */
+
+#ifdef Honor_FLT_ROUNDS
+       switch(Rounding) {
+         case 0: goto trimzeros;
+         case 2: goto roundoff;
+         }
+#endif
+       b = lshift(b, 1);
+       j = cmp(b, S);
+#ifdef ROUND_BIASED
+       if (j >= 0)
+#else
+       if (j > 0 || (j == 0 && dig & 1))
+#endif
+               {
+ roundoff:
+               while(*--s == '9')
+                       if (s == s0) {
+                               k++;
+                               *s++ = '1';
+                               goto ret;
+                               }
+               ++*s++;
+               }
+       else {
+#ifdef Honor_FLT_ROUNDS
+ trimzeros:
+#endif
+               while(*--s == '0');
+               s++;
+               }
+ ret:
+       Bfree(S);
+       if (mhi) {
+               if (mlo && mlo != mhi)
+                       Bfree(mlo);
+               Bfree(mhi);
+               }
+ ret1:
+#ifdef SET_INEXACT
+       if (inexact) {
+               if (!oldinexact) {
+                       word0(&u) = Exp_1 + (70 << Exp_shift);
+                       word1(&u) = 0;
+                       dval(&u) += 1.;
+                       }
+               }
+       else if (!oldinexact)
+               clear_inexact();
+#endif
+       Bfree(b);
+       *s = 0;
+       *decpt = k + 1;
+       if (rve)
+               *rve = s;
+       return s0;
+       }
+#ifdef __cplusplus
+}
+#endif