/*** *x10fout.c - floating point output for 10-byte long double * * Copyright (c) 1991-2001, Microsoft Corporation. All rights reserved. * *Purpose: * Support conversion of a long double into a string * *Revision History: * 07/15/91 GDP Initial version in C (ported from assembly) * 01/23/92 GDP Support MIPS encoding for NaN * *******************************************************************************/ #include #include #define STRCPY strcpy #define PUT_ZERO_FOS(fos) \ fos->exp = 0, \ fos->sign = ' ', \ fos->ManLen = 1, \ fos->man[0] = '0',\ fos->man[1] = 0; #define SNAN_STR "1#SNAN" #define SNAN_STR_LEN 6 #define QNAN_STR "1#QNAN" #define QNAN_STR_LEN 6 #define INF_STR "1#INF" #define INF_STR_LEN 5 #define IND_STR "1#IND" #define IND_STR_LEN 5 /*** *int _CALLTYPE5 * _$i10_output(_LDOUBLE ld, * int ndigits, * unsigned output_flags, * FOS *fos) - output conversion of a 10-byte _LDOUBLE * *Purpose: * Fill in a FOS structure for a given _LDOUBLE * *Entry: * _LDOUBLE ld: The long double to be converted into a string * int ndigits: number of digits allowed in the output format. * unsigned output_flags: The following flags can be used: * SO_FFORMAT: Indicates 'f' format * (default is 'e' format) * FOS *fos: the structure that i10_output will fill in * *Exit: * modifies *fos * return 1 if original number was ok, 0 otherwise (infinity, NaN, etc) * *Exceptions: * *******************************************************************************/ int _CALLTYPE5 $I10_OUTPUT(_LDOUBLE ld, int ndigits, unsigned output_flags, FOS *fos) { u_short expn; u_long manhi,manlo; u_short sign; /* useful constants (see algorithm explanation below) */ u_short const log2hi = 0x4d10; u_short const log2lo = 0x4d; u_short const log4hi = 0x9a; u_long const c = 0x134312f4; #if defined(L_END) _LDBL12 ld12_one_tenth = { {0xcc,0xcc,0xcc,0xcc,0xcc,0xcc, 0xcc,0xcc,0xcc,0xcc,0xfb,0x3f} }; #elif defined(B_END) _LDBL12 ld12_one_tenth = { {0x3f,0xfb,0xcc,0xcc,0xcc,0xcc, 0xcc,0xcc,0xcc,0xcc,0xcc,0xcc} }; #endif _LDBL12 ld12; /* space for a 12-byte long double */ _LDBL12 tmp12; u_short hh,ll; /* the bytes of the exponent grouped in 2 words*/ u_short mm; /* the two MSBytes of the mantissa */ s_long r; /* the corresponding power of 10 */ s_short ir; /* ir = floor(r) */ int retval = 1; /* assume valid number */ char round; /* an additional character at the end of the string */ char *p; int i; int ub_exp; int digcount; /* grab the components of the long double */ expn = *U_EXP_LD(&ld); manhi = *UL_MANHI_LD(&ld); manlo = *UL_MANLO_LD(&ld); sign = expn & MSB_USHORT; expn &= 0x7fff; if (sign) fos->sign = '-'; else fos->sign = ' '; if (expn==0 && manhi==0 && manlo==0) { PUT_ZERO_FOS(fos); return 1; } if (expn == 0x7fff) { fos->exp = 1; /* set a positive exponent for proper output */ /* check for special cases */ if (_IS_MAN_SNAN(sign, manhi, manlo)) { /* signaling NAN */ STRCPY(fos->man,SNAN_STR); fos->ManLen = SNAN_STR_LEN; retval = 0; } else if (_IS_MAN_IND(sign, manhi, manlo)) { /* indefinite */ STRCPY(fos->man,IND_STR); fos->ManLen = IND_STR_LEN; retval = 0; } else if (_IS_MAN_INF(sign, manhi, manlo)) { /* infinity */ STRCPY(fos->man,INF_STR); fos->ManLen = INF_STR_LEN; retval = 0; } else { /* quiet NAN */ STRCPY(fos->man,QNAN_STR); fos->ManLen = QNAN_STR_LEN; retval = 0; } } else { /* * Algorithm for the decoding of a valid real number x * * In the following INT(r) is the largest integer less than or * equal to r (i.e. r rounded toward -infinity). We want a result * r equal to 1 + log(x), because then x = mantissa * * 10^(INT(r)) so that .1 <= mantissa < 1. Unfortunately, * we cannot compute s exactly so we must alter the procedure * slightly. We will instead compute an estimate r of 1 + * log(x) which is always low. This will either result * in the correctly normalized number on the top of the stack * or perhaps a number which is a factor of 10 too large. We * will then check to see that if x is larger than one * and if so multiply x by 1/10. * * We will use a low precision (fixed point 24 bit) estimate * of of 1 + log base 10 of x. We have approximately .mm * * 2^hhll on the top of the stack where m, h, and l represent * hex digits, mm represents the high 2 hex digits of the * mantissa, hh represents the high 2 hex digits of the exponent, * and ll represents the low 2 hex digits of the exponent. Since * .mm is a truncated representation of the mantissa, using it * in this monotonically increasing polynomial approximation * of the logarithm will naturally give a low result. Let's * derive a formula for a lower bound r on 1 + log(x): * * .4D104D42H < log(2)=.30102999...(base 10) < .4D104D43H * .9A20H < log(4)=.60205999...(base 10) < .9A21H * * 1/2 <= .mm < 1 * ==> log(.mm) >= .mm * log(4) - log(4) * * Substituting in truncated hex constants in the formula above * gives r = 1 + .4D104DH * hhll. + .9AH * .mm - .9A21H. Now * multiplication of hex digits 5 and 6 of log(2) by ll has an * insignificant effect on the first 24 bits of the result so * it will not be calculated. This gives the expression r = * 1 + .4D10H * hhll. + .4DH * .hh + .9A * .mm - .9A21H. * Finally we must add terms to our formula to subtract out the * effect of the exponent bias. We obtain the following formula: * * (implied decimal point) * < >.< > * |3|3|2|2|2|2|2|2|2|2|2|2|1|1|1|1|1|1|1|1|1|1|0|0|0|0|0|0|0|0|0|0| * |1|0|9|8|7|6|5|4|3|2|1|0|9|8|7|6|5|4|3|2|1|0|9|8|7|6|5|4|3|2|1|0| * + < 1 > * + < .4D10H * hhll. > * + < .00004DH * hh00. > * + < .9AH * .mm > * - < .9A21H > * - < .4D10H * 3FFEH > * - < .00004DH * 3F00H > * * ==> r = .4D10H * hhll. + .4DH * .hh + .9AH * .mm - 1343.12F4H * * The difference between the lower bound r and the upper bound * s is calculated as follows: * * .937EH < 1/ln(10)-log(1/ln(4))=.57614993...(base 10) < .937FH * * 1/2 <= .mm < 1 * ==> log(.mm) <= .mm * log(4) - [1/ln(10) - log(1/ln(4))] * * so tenatively s = r + log(4) - [1/ln(10) - log(1/ln(4))], * but we must also add in terms to ensure we will have an upper * bound even after the truncation of various values. Because * log(2) * hh00. is truncated to .4D104DH * hh00. we must * add .0043H, because log(2) * ll. is truncated to .4D10H * * ll. we must add .0005H, because * log(4) is * truncated to .mm * .9AH we must add .009AH and .0021H. * * Thus s = r - .937EH + .9A21H + .0043H + .0005H + .009AH + .0021H * = r + .07A6H * ==> s = .4D10H * hhll. + .4DH * .hh + .9AH * .mm - 1343.0B4EH * * r is equal to 1 + log(x) more than (10000H - 7A6H) / * 10000H = 97% of the time. * * In the above formula, a u_long is use to accomodate r, and * there is an implied decimal point in the middle. */ hh = expn >> 8; ll = expn & (u_short)0xff; mm = (u_short) (manhi >> 24); r = (s_long)log2hi*(s_long)expn + log2lo*hh + log4hi*mm - c; ir = (s_short)(r >> 16); /* * * We stated that we wanted to normalize x so that * * .1 <= x < 1 * * This was a slight oversimplification. Actually we want a * number which when rounded to 16 significant digits is in the * desired range. To do this we must normalize x so that * * .1 - 5*10^(-18) <= x < 1 - 5*10^(-17) * * and then round. * * If we had f = INT(1+log(x)) we could multiply by 10^(-f) * to get x into the desired range. We do not quite have * f but we do have INT(r) from the last step which is equal * to f 97% of the time and 1 less than f the rest of the time. * We can multiply by 10^-[INT(r)] and if the result is greater * than 1 - 5*10^(-17) we can then multiply by 1/10. This final * result will lie in the proper range. */ /* convert _LDOUBLE to _LDBL12) */ *U_EXP_12(&ld12) = expn; *UL_MANHI_12(&ld12) = manhi; *UL_MANLO_12(&ld12) = manlo; *U_XT_12(&ld12) = 0; /* multiply by 10^(-ir) */ __multtenpow12(&ld12,-ir,1); /* if ld12 >= 1.0 then divide by 10.0 */ if (*U_EXP_12(&ld12) >= 0x3fff) { ir++; __ld12mul(&ld12,&ld12_one_tenth); } fos->exp = ir; if (output_flags & SO_FFORMAT){ /* 'f' format, add exponent to ndigits */ ndigits += ir; if (ndigits <= 0) { /* return 0 */ PUT_ZERO_FOS(fos); return 1; } } if (ndigits > MAX_MAN_DIGITS) ndigits = MAX_MAN_DIGITS; ub_exp = *U_EXP_12(&ld12) - 0x3ffe; /* unbias exponent */ *U_EXP_12(&ld12) = 0; /* * Now the mantissa has to be converted to fixed point. * Then we will use the MSB of ld12 for generating * the decimal digits. The next 11 bytes will hold * the mantissa (after it has been converted to * fixed point). */ for (i=0;i<8;i++) __shl_12(&ld12); /* make space for an extra byte, in case we shift right later */ if (ub_exp < 0) { int shift_count = (-ub_exp) & 0xff; for (;shift_count>0;shift_count--) __shr_12(&ld12); } p = fos->man; for(digcount=ndigits+1;digcount>0;digcount--) { tmp12 = ld12; __shl_12(&ld12); __shl_12(&ld12); __add_12(&ld12,&tmp12); __shl_12(&ld12); /* ld12 *= 10 */ /* Now we have the first decimal digit in the msbyte of exponent */ *p++ = (char) (*UCHAR_12(&ld12,11) + '0'); *UCHAR_12(&ld12,11) = 0; } round = *(--p); p--; /* p points now to the last character of the string excluding the rounding digit */ if (round >= '5') { /* look for a non-9 digit starting from the end of string */ for (;p>=fos->man && *p=='9';p--) { *p = '0'; } if (p < fos->man){ p++; fos->exp ++; } (*p)++; } else { /* remove zeros */ for (;p>=fos->man && *p=='0';p--); if (p < fos->man) { /* return 0 */ PUT_ZERO_FOS(fos); return 1; } } fos->ManLen = (char) (p - fos->man + 1); fos->man[fos->ManLen] = '\0'; } return retval; }