diff options
author | H.J. Lu <hongjiu.lu@intel.com> | 2007-09-27 17:47:23 +0000 |
---|---|---|
committer | H.J. Lu <hjl@gcc.gnu.org> | 2007-09-27 10:47:23 -0700 |
commit | b2a00c8984668b802b3331db38cd7f530f886842 (patch) | |
tree | 373fa862a43e43bb30e4f569c9c95e43119764f8 /libgcc/config/libbid/bid128_to_uint64.c | |
parent | e233ac979c0f179f858f3c86281835a4c09cece2 (diff) | |
download | gcc-b2a00c8984668b802b3331db38cd7f530f886842.tar.gz |
Makefile.in (dfp-filenames): Replace decimal_globals...
libgcc/
2007-09-27 H.J. Lu <hongjiu.lu@intel.com>
* Makefile.in (dfp-filenames): Replace decimal_globals,
decimal_data, binarydecimal and convert_data with
bid_decimal_globals, bid_decimal_data, bid_binarydecimal
and bid_convert_data, respectively.
libgcc/config/libbid/
2007-09-27 H.J. Lu <hongjiu.lu@intel.com>
* bid128_fromstring.c: Removed.
* bid_dpd.c: New from libbid 2007-09-26.
* bid128_to_int16.c: Likewise.
* bid128_to_int8.c: Likewise.
* bid128_to_uint8.c: Likewise.
* bid128_to_uint16.c: Likewise.
* bid64_to_int16.c: Likewise.
* bid64_to_int8.c: Likewise.
* bid64_to_uint16.c: Likewise.
* bid64_to_uint8.c: Likewise.
* bid128_2_str.h: Updated from libbid 2007-09-26.
* bid128_2_str_macros.h: Likewise.
* bid128_2_str_tables.c: Likewise.
* bid128_add.c: Likewise.
* bid128.c: Likewise.
* bid128_compare.c: Likewise.
* bid128_div.c: Likewise.
* bid128_fma.c: Likewise.
* bid128_logb.c: Likewise.
* bid128_minmax.c: Likewise.
* bid128_mul.c: Likewise.
* bid128_next.c: Likewise.
* bid128_noncomp.c: Likewise.
* bid128_quantize.c: Likewise.
* bid128_rem.c: Likewise.
* bid128_round_integral.c: Likewise.
* bid128_scalb.c: Likewise.
* bid128_sqrt.c: Likewise.
* bid128_string.c: Likewise.
* bid128_to_int32.c: Likewise.
* bid128_to_int64.c: Likewise.
* bid128_to_uint32.c: Likewise.
* bid128_to_uint64.c: Likewise.
* bid32_to_bid128.c: Likewise.
* bid32_to_bid64.c: Likewise.
* bid64_add.c: Likewise.
* bid64_compare.c: Likewise.
* bid64_div.c: Likewise.
* bid64_fma.c: Likewise.
* bid64_logb.c: Likewise.
* bid64_minmax.c: Likewise.
* bid64_mul.c: Likewise.
* bid64_next.c: Likewise.
* bid64_noncomp.c: Likewise.
* bid64_quantize.c: Likewise.
* bid64_rem.c: Likewise.
* bid64_round_integral.c: Likewise.
* bid64_scalb.c: Likewise.
* bid64_sqrt.c: Likewise.
* bid64_string.c: Likewise.
* bid64_to_bid128.c: Likewise.
* bid64_to_int32.c: Likewise.
* bid64_to_int64.c: Likewise.
* bid64_to_uint32.c: Likewise.
* bid64_to_uint64.c: Likewise.
* bid_b2d.h: Likewise.
* bid_binarydecimal.c: Likewise.
* bid_conf.h: Likewise.
* bid_convert_data.c: Likewise.
* bid_decimal_data.c: Likewise.
* bid_decimal_globals.c: Likewise.
* bid_div_macros.h: Likewise.
* bid_flag_operations.c: Likewise.
* bid_from_int.c: Likewise.
* bid_functions.h: Likewise.
* bid_gcc_intrinsics.h: Likewise.
* bid_inline_add.h: Likewise.
* bid_internal.h: Likewise.
* bid_round.c: Likewise.
* bid_sqrt_macros.h: Likewise.
* _addsub_dd.c: Likewise.
* _addsub_sd.c: Likewise.
* _addsub_td.c: Likewise.
* _dd_to_df.c: Likewise.
* _dd_to_di.c: Likewise.
* _dd_to_sd.c: Likewise.
* _dd_to_sf.c: Likewise.
* _dd_to_si.c: Likewise.
* _dd_to_td.c: Likewise.
* _dd_to_tf.c: Likewise.
* _dd_to_udi.c: Likewise.
* _dd_to_usi.c: Likewise.
* _dd_to_xf.c: Likewise.
* _df_to_dd.c: Likewise.
* _df_to_sd.c: Likewise.
* _df_to_td.c: Likewise.
* _di_to_dd.c: Likewise.
* _di_to_sd.c: Likewise.
* _di_to_td.c: Likewise.
* _div_dd.c: Likewise.
* _div_sd.c: Likewise.
* _div_td.c: Likewise.
* _eq_dd.c: Likewise.
* _eq_sd.c: Likewise.
* _eq_td.c: Likewise.
* _ge_dd.c: Likewise.
* _ge_sd.c: Likewise.
* _ge_td.c: Likewise.
* _gt_dd.c: Likewise.
* _gt_sd.c: Likewise.
* _gt_td.c: Likewise.
* _isinfd128.c: Likewise.
* _isinfd32.c: Likewise.
* _isinfd64.c: Likewise.
* _le_dd.c: Likewise.
* _le_sd.c: Likewise.
* _le_td.c: Likewise.
* _lt_dd.c: Likewise.
* _lt_sd.c: Likewise.
* _lt_td.c: Likewise.
* _mul_dd.c: Likewise.
* _mul_sd.c: Likewise.
* _mul_td.c: Likewise.
* _ne_dd.c: Likewise.
* _ne_sd.c: Likewise.
* _ne_td.c: Likewise.
* _sd_to_dd.c: Likewise.
* _sd_to_df.c: Likewise.
* _sd_to_di.c: Likewise.
* _sd_to_sf.c: Likewise.
* _sd_to_si.c: Likewise.
* _sd_to_td.c: Likewise.
* _sd_to_tf.c: Likewise.
* _sd_to_udi.c: Likewise.
* _sd_to_usi.c: Likewise.
* _sd_to_xf.c: Likewise.
* _sf_to_dd.c: Likewise.
* _sf_to_sd.c: Likewise.
* _sf_to_td.c: Likewise.
* _si_to_dd.c: Likewise.
* _si_to_sd.c: Likewise.
* _si_to_td.c: Likewise.
* _td_to_dd.c: Likewise.
* _td_to_df.c: Likewise.
* _td_to_di.c: Likewise.
* _td_to_sd.c: Likewise.
* _td_to_sf.c: Likewise.
* _td_to_si.c: Likewise.
* _td_to_tf.c: Likewise.
* _td_to_udi.c: Likewise.
* _td_to_usi.c: Likewise.
* _td_to_xf.c: Likewise.
* _tf_to_dd.c: Likewise.
* _tf_to_sd.c: Likewise.
* _tf_to_td.c: Likewise.
* _udi_to_dd.c: Likewise.
* _udi_to_sd.c: Likewise.
* _udi_to_td.c: Likewise.
* _unord_dd.c: Likewise.
* _unord_sd.c: Likewise.
* _unord_td.c: Likewise.
* _usi_to_dd.c: Likewise.
* _usi_to_sd.c: Likewise.
* _usi_to_td.c: Likewise.
* _xf_to_dd.c: Likewise.
* _xf_to_sd.c: Likewise.
* _xf_to_td.c: Likewise.
2007-09-27 H.J. Lu <hongjiu.lu@intel.com>
* b2d.h: Renamed to ...
* bid_b2d.h: This.
* bid128_to_string.c: Renamed to ...
* bid128_string.c: This.
* bid_intrinsics.h: Renamed to ...
* bid_gcc_intrinsics.h: This.
* bid_string.c: Renamed to ...
* bid64_string.c: This.
* binarydecimal.c: Renamed to ...
* bid_decimal_globals.c: This.
* convert_data.c: Renamed to ...
* bid_convert_data.c: This.
* decimal_data.c: Renamed to ...
* bid_decimal_data.c: This.
* decimal_globals.c: Renamed to ...
* bid_decimal_globals.c: This.
* div_macros.h: Renamed to ...
* bid_div_macros.h: This.
* inline_bid_add.h: Renamed to ...
* bid_inline_add.h: This.
* sqrt_macros.h: Renamed to ...
* bid_sqrt_macros.h: This.
From-SVN: r128841
Diffstat (limited to 'libgcc/config/libbid/bid128_to_uint64.c')
-rw-r--r-- | libgcc/config/libbid/bid128_to_uint64.c | 5790 |
1 files changed, 2907 insertions, 2883 deletions
diff --git a/libgcc/config/libbid/bid128_to_uint64.c b/libgcc/config/libbid/bid128_to_uint64.c index de6fd60e295..d8ec7110079 100644 --- a/libgcc/config/libbid/bid128_to_uint64.c +++ b/libgcc/config/libbid/bid128_to_uint64.c @@ -32,912 +32,1488 @@ Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA * BID128_to_uint64_rnint ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_rnint, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_rnint, x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - UINT64 tmp64; - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; + UINT64 tmp64; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; - } - exp = (x_exp >> 49) - 6176; + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n < -1/2 then n cannot be converted to uint64 with RN - // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 1/2 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 > 0x05, 1<=q<=34 - // <=> C * 10^(21-q) > 0x05, 1<=q<=34 - if (q == 21) { - // C > 5 - if (C1.w[1] != 0 || C1.w[0] > 0x05ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) > 5 is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C > 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n < -1/2 then n cannot be converted to uint64 with RN + // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 1/2 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 > 0x05, 1<=q<=34 + // <=> C * 10^(21-q) > 0x05, 1<=q<=34 + if (q == 21) { + // C > 5 + if (C1.w[1] != 0 || C1.w[0] > 0x05ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // if n > 0 and q + exp = 20 - // if n >= 2^64 - 1/2 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) - // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 - if (q == 1) { - // C * 10^20 >= 0x9fffffffffffffffb - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) >= 0x9fffffffffffffffb - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff - C.w[0] = C1.w[0] + C1.w[0]; - C.w[1] = C1.w[1] + C1.w[1]; - if (C.w[0] < C1.w[0]) - C.w[1]++; - if (C.w[1] > 0x01 || (C.w[1] == 0x01 - && C.w[0] >= 0xffffffffffffffffull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C >= 0x9fffffffffffffffb - if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 - && C1.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits - C.w[1] = 0x09; - C.w[0] = 0xfffffffffffffffbull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) > 5 is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C > 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - } - // n is not too large to be converted to int64 if -1/2 <= n < 2^64 - 1/2 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) - // return 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) - // if 0.c(0)c(1)...c(q-1) <= 0.5 <=> c(0)c(1)...c(q-1) <= 5 * 10^(q-1) - // res = 0 - // else if x > 0 - // res = +1 - // else // if x < 0 - // invalid exc - ind = q - 1; - if (ind <= 18) { // 0 <= ind <= 18 - if ((C1.w[1] == 0) && (C1.w[0] <= __bid_midpoint64[ind])) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { + } else { // if n > 0 and q + exp = 20 + // if n >= 2^64 - 1/2 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) + // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0x9fffffffffffffffb + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0x9fffffffffffffffb + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff + C.w[0] = C1.w[0] + C1.w[0]; + C.w[1] = C1.w[1] + C1.w[1]; + if (C.w[0] < C1.w[0]) + C.w[1]++; + if (C.w[1] > 0x01 || (C.w[1] == 0x01 + && C.w[0] >= 0xffffffffffffffffull)) { + // set invalid flag *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - } else { // 19 <= ind <= 33 - if ((C1.w[1] < __bid_midpoint128[ind - 19].w[1]) - || ((C1.w[1] == __bid_midpoint128[ind - 19].w[1]) - && (C1.w[0] <= __bid_midpoint128[ind - 19].w[0]))) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0x9fffffffffffffffb + if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 + && C1.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits + C.w[1] = 0x09; + C.w[0] = 0xfffffffffffffffbull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] + || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + // set invalid flag *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded - // to nearest to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 - // set invalid flag + } + } + // n is not too large to be converted to int64 if -1/2 <= n < 2^64 - 1/2 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) + // return 0 + res = 0x0000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) + // if 0.c(0)c(1)...c(q-1) <= 0.5 <=> c(0)c(1)...c(q-1) <= 5 * 10^(q-1) + // res = 0 + // else if x > 0 + // res = +1 + // else // if x < 0 + // invalid exc + ind = q - 1; + if (ind <= 18) { // 0 <= ind <= 18 + if ((C1.w[1] == 0) && (C1.w[0] <= midpoint64[ind])) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { + res = 0x8000000000000000ull; *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite + } + } else { // 19 <= ind <= 33 + if ((C1.w[1] < midpoint128[ind - 19].w[1]) + || ((C1.w[1] == midpoint128[ind - 19].w[1]) + && (C1.w[0] <= midpoint128[ind - 19].w[0]))) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { res = 0x8000000000000000ull; - BID_RETURN (res); + *pfpsf |= INVALID_EXCEPTION; } - // 1 <= x < 2^64-1/2 so x can be rounded - // to nearest to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits - tmp64 = C1.w[0]; - if (ind <= 19) { - C1.w[0] = C1.w[0] + __bid_midpoint64[ind - 1]; - } else { - C1.w[0] = C1.w[0] + __bid_midpoint128[ind - 20].w[0]; - C1.w[1] = C1.w[1] + __bid_midpoint128[ind - 20].w[1]; - } - if (C1.w[0] < tmp64) - C1.w[1]++; - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = (C1 + 1/2 * 10^x) * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // if (0 < f* < 10^(-x)) then the result is a midpoint - // if floor(C*) is even then C* = floor(C*) - logical right - // shift; C* has p decimal digits, correct by Prop. 1) - // else if floor(C*) is odd C* = floor(C*)-1 (logical right - // shift; C* has p decimal digits, correct by Pr. 1) - // else - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - // if the result was a midpoint it was rounded away from zero, so - // it will need a correction - // check for midpoints - if ((fstar.w[3] == 0) && (fstar.w[2] == 0) - && (fstar.w[1] || fstar.w[0]) - && (fstar.w[1] < __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] <= __bid_ten2mk128trunc[ind - 1].w[0]))) { - // the result is a midpoint; round to nearest - if (Cstar.w[0] & 0x01) { // Cstar.w[0] is odd; MP in [EVEN, ODD] - // if floor(C*) is odd C = floor(C*) - 1; the result >= 1 - Cstar.w[0]--; // Cstar.w[0] is now even - } // else MP in [ODD, EVEN] - } - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; + } + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded + // to nearest to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // 1 <= x < 2^64-1/2 so x can be rounded + // to nearest to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits + tmp64 = C1.w[0]; + if (ind <= 19) { + C1.w[0] = C1.w[0] + midpoint64[ind - 1]; } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + C1.w[0] = C1.w[0] + midpoint128[ind - 20].w[0]; + C1.w[1] = C1.w[1] + midpoint128[ind - 20].w[1]; + } + if (C1.w[0] < tmp64) + C1.w[1]++; + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = (C1 + 1/2 * 10^x) * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // if (0 < f* < 10^(-x)) then the result is a midpoint + // if floor(C*) is even then C* = floor(C*) - logical right + // shift; C* has p decimal digits, correct by Prop. 1) + // else if floor(C*) is odd C* = floor(C*)-1 (logical right + // shift; C* has p decimal digits, correct by Pr. 1) + // else + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + // if the result was a midpoint it was rounded away from zero, so + // it will need a correction + // check for midpoints + if ((fstar.w[3] == 0) && (fstar.w[2] == 0) + && (fstar.w[1] || fstar.w[0]) + && (fstar.w[1] < ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] <= ten2mk128trunc[ind - 1].w[0]))) { + // the result is a midpoint; round to nearest + if (Cstar.w[0] & 0x01) { // Cstar.w[0] is odd; MP in [EVEN, ODD] + // if floor(C*) is odd C = floor(C*) - 1; the result >= 1 + Cstar.w[0]--; // Cstar.w[0] is now even + } // else MP in [ODD, EVEN] } + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } /***************************************************************************** * BID128_to_uint64_xrnint ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xrnint, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_xrnint, x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - UINT64 tmp64, tmp64A; - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; + UINT64 tmp64, tmp64A; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - exp = (x_exp >> 49) - 6176; + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n < -1/2 then n cannot be converted to uint64 with RN - // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 1/2 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 > 0x05, 1<=q<=34 - // <=> C * 10^(21-q) > 0x05, 1<=q<=34 - if (q == 21) { - // C > 5 - if (C1.w[1] != 0 || C1.w[0] > 0x05ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) > 5 is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C > 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n < -1/2 then n cannot be converted to uint64 with RN + // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 1/2 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 > 0x05, 1<=q<=34 + // <=> C * 10^(21-q) > 0x05, 1<=q<=34 + if (q == 21) { + // C > 5 + if (C1.w[1] != 0 || C1.w[0] > 0x05ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // if n > 0 and q + exp = 20 - // if n >= 2^64 - 1/2 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) - // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 - if (q == 1) { - // C * 10^20 >= 0x9fffffffffffffffb - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) >= 0x9fffffffffffffffb - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff - C.w[0] = C1.w[0] + C1.w[0]; - C.w[1] = C1.w[1] + C1.w[1]; - if (C.w[0] < C1.w[0]) - C.w[1]++; - if (C.w[1] > 0x01 || (C.w[1] == 0x01 - && C.w[0] >= 0xffffffffffffffffull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C >= 0x9fffffffffffffffb - if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 - && C1.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits - C.w[1] = 0x09; - C.w[0] = 0xfffffffffffffffbull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) > 5 is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C > 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - } - // n is not too large to be converted to int64 if -1/2 <= n < 2^64 - 1/2 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - // return 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) - // if 0.c(0)c(1)...c(q-1) <= 0.5 <=> c(0)c(1)...c(q-1) <= 5 * 10^(q-1) - // res = 0 - // else if x > 0 - // res = +1 - // else // if x < 0 - // invalid exc - ind = q - 1; - if (ind <= 18) { // 0 <= ind <= 18 - if ((C1.w[1] == 0) && (C1.w[0] <= __bid_midpoint64[ind])) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { + } else { // if n > 0 and q + exp = 20 + // if n >= 2^64 - 1/2 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) + // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0x9fffffffffffffffb + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0x9fffffffffffffffb + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // 19 <= ind <= 33 - if ((C1.w[1] < __bid_midpoint128[ind - 19].w[1]) - || ((C1.w[1] == __bid_midpoint128[ind - 19].w[1]) - && (C1.w[0] <= __bid_midpoint128[ind - 19].w[0]))) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff + C.w[0] = C1.w[0] + C1.w[0]; + C.w[1] = C1.w[1] + C1.w[1]; + if (C.w[0] < C1.w[0]) + C.w[1]++; + if (C.w[1] > 0x01 || (C.w[1] == 0x01 + && C.w[0] >= 0xffffffffffffffffull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0x9fffffffffffffffb + if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 + && C1.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits + C.w[1] = 0x09; + C.w[0] = 0xfffffffffffffffbull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] + || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + // set invalid flag *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; BID_RETURN (res); } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded - // to nearest to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 - // set invalid flag + } + } + // n is not too large to be converted to int64 if -1/2 <= n < 2^64 - 1/2 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + // return 0 + res = 0x0000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) + // if 0.c(0)c(1)...c(q-1) <= 0.5 <=> c(0)c(1)...c(q-1) <= 5 * 10^(q-1) + // res = 0 + // else if x > 0 + // res = +1 + // else // if x < 0 + // invalid exc + ind = q - 1; + if (ind <= 18) { // 0 <= ind <= 18 + if ((C1.w[1] == 0) && (C1.w[0] <= midpoint64[ind])) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { + res = 0x8000000000000000ull; *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite + BID_RETURN (res); + } + } else { // 19 <= ind <= 33 + if ((C1.w[1] < midpoint128[ind - 19].w[1]) + || ((C1.w[1] == midpoint128[ind - 19].w[1]) + && (C1.w[0] <= midpoint128[ind - 19].w[0]))) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { res = 0x8000000000000000ull; + *pfpsf |= INVALID_EXCEPTION; BID_RETURN (res); } - // 1 <= x < 2^64-1/2 so x can be rounded - // to nearest to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits - tmp64 = C1.w[0]; - if (ind <= 19) { - C1.w[0] = C1.w[0] + __bid_midpoint64[ind - 1]; - } else { - C1.w[0] = C1.w[0] + __bid_midpoint128[ind - 20].w[0]; - C1.w[1] = C1.w[1] + __bid_midpoint128[ind - 20].w[1]; - } - if (C1.w[0] < tmp64) - C1.w[1]++; - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = (C1 + 1/2 * 10^x) * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // if (0 < f* < 10^(-x)) then the result is a midpoint - // if floor(C*) is even then C* = floor(C*) - logical right - // shift; C* has p decimal digits, correct by Prop. 1) - // else if floor(C*) is odd C* = floor(C*)-1 (logical right - // shift; C* has p decimal digits, correct by Pr. 1) - // else - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - // determine inexactness of the rounding of C* - // if (0 < f* - 1/2 < 10^(-x)) then - // the result is exact - // else // if (f* - 1/2 > T*) then - // the result is inexact - if (ind - 1 <= 2) { - if (fstar.w[1] > 0x8000000000000000ull || - ((fstar.w[1] == 0x8000000000000000ull && fstar.w[0] > 0x0ull))) { - // f* > 1/2 and the result may be exact - tmp64 = fstar.w[1] - 0x8000000000000000ull; // f* - 1/2 - if ((tmp64 > __bid_ten2mk128trunc[ind - 1].w[1] - || (tmp64 == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] >= __bid_ten2mk128trunc[ind - 1].w[0]))) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // the result is inexact; f2* <= 1/2 + } + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded + // to nearest to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // 1 <= x < 2^64-1/2 so x can be rounded + // to nearest to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits + tmp64 = C1.w[0]; + if (ind <= 19) { + C1.w[0] = C1.w[0] + midpoint64[ind - 1]; + } else { + C1.w[0] = C1.w[0] + midpoint128[ind - 20].w[0]; + C1.w[1] = C1.w[1] + midpoint128[ind - 20].w[1]; + } + if (C1.w[0] < tmp64) + C1.w[1]++; + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = (C1 + 1/2 * 10^x) * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // if (0 < f* < 10^(-x)) then the result is a midpoint + // if floor(C*) is even then C* = floor(C*) - logical right + // shift; C* has p decimal digits, correct by Prop. 1) + // else if floor(C*) is odd C* = floor(C*)-1 (logical right + // shift; C* has p decimal digits, correct by Pr. 1) + // else + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + // determine inexactness of the rounding of C* + // if (0 < f* - 1/2 < 10^(-x)) then + // the result is exact + // else // if (f* - 1/2 > T*) then + // the result is inexact + if (ind - 1 <= 2) { + if (fstar.w[1] > 0x8000000000000000ull || + (fstar.w[1] == 0x8000000000000000ull + && fstar.w[0] > 0x0ull)) { + // f* > 1/2 and the result may be exact + tmp64 = fstar.w[1] - 0x8000000000000000ull; // f* - 1/2 + if (tmp64 > ten2mk128trunc[ind - 1].w[1] + || (tmp64 == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] >= ten2mk128trunc[ind - 1].w[0])) { // set the inexact flag *pfpsf |= INEXACT_EXCEPTION; - } - } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 - if (fstar.w[3] > 0x0 || (fstar.w[3] == 0x0 - && fstar.w[2] > __bid_one_half128[ind - 1]) || (fstar.w[3] == 0x0 - && fstar.w[2] == __bid_one_half128[ind - 1] && (fstar.w[1] - || fstar.w[0]))) { - // f2* > 1/2 and the result may be exact - // Calculate f2* - 1/2 - tmp64 = fstar.w[2] - __bid_one_half128[ind - 1]; - tmp64A = fstar.w[3]; - if (tmp64 > fstar.w[2]) - tmp64A--; - if (tmp64A || tmp64 - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // the result is inexact; f2* <= 1/2 + } // else the result is exact + } else { // the result is inexact; f2* <= 1/2 + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } + } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 + if (fstar.w[3] > 0x0 || + (fstar.w[3] == 0x0 && fstar.w[2] > onehalf128[ind - 1]) || + (fstar.w[3] == 0x0 && fstar.w[2] == onehalf128[ind - 1] && + (fstar.w[1] || fstar.w[0]))) { + // f2* > 1/2 and the result may be exact + // Calculate f2* - 1/2 + tmp64 = fstar.w[2] - onehalf128[ind - 1]; + tmp64A = fstar.w[3]; + if (tmp64 > fstar.w[2]) + tmp64A--; + if (tmp64A || tmp64 + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { // set the inexact flag *pfpsf |= INEXACT_EXCEPTION; - } - } else { // if 22 <= ind <= 33 - if (fstar.w[3] > __bid_one_half128[ind - 1] - || (fstar.w[3] == __bid_one_half128[ind - 1] && (fstar.w[2] - || fstar.w[1] - || fstar.w[0]))) { - // f2* > 1/2 and the result may be exact - // Calculate f2* - 1/2 - tmp64 = fstar.w[3] - __bid_one_half128[ind - 1]; - if (tmp64 || fstar.w[2] - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // the result is inexact; f2* <= 1/2 + } // else the result is exact + } else { // the result is inexact; f2* <= 1/2 + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } + } else { // if 22 <= ind <= 33 + if (fstar.w[3] > onehalf128[ind - 1] || + (fstar.w[3] == onehalf128[ind - 1] && + (fstar.w[2] || fstar.w[1] || fstar.w[0]))) { + // f2* > 1/2 and the result may be exact + // Calculate f2* - 1/2 + tmp64 = fstar.w[3] - onehalf128[ind - 1]; + if (tmp64 || fstar.w[2] + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { // set the inexact flag *pfpsf |= INEXACT_EXCEPTION; - } + } // else the result is exact + } else { // the result is inexact; f2* <= 1/2 + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; } + } - // if the result was a midpoint it was rounded away from zero, so - // it will need a correction - // check for midpoints - if ((fstar.w[3] == 0) && (fstar.w[2] == 0) - && (fstar.w[1] || fstar.w[0]) - && (fstar.w[1] < __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] <= __bid_ten2mk128trunc[ind - 1].w[0]))) { - // the result is a midpoint; round to nearest - if (Cstar.w[0] & 0x01) { // Cstar.w[0] is odd; MP in [EVEN, ODD] - // if floor(C*) is odd C = floor(C*) - 1; the result >= 1 - Cstar.w[0]--; // Cstar.w[0] is now even - } // else MP in [ODD, EVEN] - } - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + // if the result was a midpoint it was rounded away from zero, so + // it will need a correction + // check for midpoints + if ((fstar.w[3] == 0) && (fstar.w[2] == 0) + && (fstar.w[1] || fstar.w[0]) + && (fstar.w[1] < ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] <= ten2mk128trunc[ind - 1].w[0]))) { + // the result is a midpoint; round to nearest + if (Cstar.w[0] & 0x01) { // Cstar.w[0] is odd; MP in [EVEN, ODD] + // if floor(C*) is odd C = floor(C*) - 1; the result >= 1 + Cstar.w[0]--; // Cstar.w[0] is now even + } // else MP in [ODD, EVEN] } + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } /***************************************************************************** * BID128_to_uint64_floor ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_floor, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_floor, x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 P256; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} + // check for non-canonical values (after the check for special values) +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // if n < 0 then n cannot be converted to uint64 with RM + if (x_sign) { // if n < 0 and q + exp = 20 + // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 0 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + // if n > 0 and q + exp = 20 + // if n >= 2^64 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*2^65 + // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0xa0000000000000000 + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] >= 0x0a) { + // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; - } else { // x is QNaN + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0xa0000000000000000 + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] >= 0x0a) { + // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C >= 0x10000000000000000 + if (C1.w[1] >= 0x01) { + // actually C1.w[1] == 0x01 && C1.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0xa0000000000000000 + if (C1.w[1] >= 0x0a) { + // actually C1.w[1] == 0x0a && C1.w[0] >= 0x0000000000000000ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; - } else { // x is -inf + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits + C.w[1] = 0x0a; + C.w[0] = 0x0000000000000000ull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); } - BID_RETURN (res); + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } } - // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + // n is not too large to be converted to int64 if 0 <= n < 2^64 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) <= 0) { // n = +0.[0...0]c(0)c(1)...c(q-1) + // return 0 res = 0x0000000000000000ull; BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // 1 <= x < 2^64 so x can be rounded + // down to a 64-bit unsigned signed integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 fits in 127 bits + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = C1 * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; + } + } +} + +BID_RETURN (res); +} + +/***************************************************************************** + * BID128_to_uint64_xfloor + ****************************************************************************/ + +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_xfloor, x) + + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent + // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; + + // unpack x +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; + + // check for NaN or Infinity +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { + // x is special +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} + // check for non-canonical values (after the check for special values) +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // if n < 0 then n cannot be converted to uint64 with RM + if (x_sign) { // if n < 0 and q + exp = 20 + // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 0 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + // if n > 0 and q + exp = 20 + // if n >= 2^64 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*2^65 + // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0xa0000000000000000 + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] >= 0x0a) { + // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0xa0000000000000000 + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] >= 0x0a) { + // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C >= 0x10000000000000000 + if (C1.w[1] >= 0x01) { + // actually C1.w[1] == 0x01 && C1.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0xa0000000000000000 + if (C1.w[1] >= 0x0a) { + // actually C1.w[1] == 0x0a && C1.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits + C.w[1] = 0x0a; + C.w[0] = 0x0000000000000000ull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } + } + // n is not too large to be converted to int64 if 0 <= n < 2^64 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) <= 0) { // n = +0.[0...0]c(0)c(1)...c(q-1) + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + // return 0 res = 0x0000000000000000ull; BID_RETURN (res); - } else { // x is not special and is not zero + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // 1 <= x < 2^64 so x can be rounded + // down to a 64-bit unsigned signed integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 fits in 127 bits + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = C1 * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + // determine inexactness of the rounding of C* + // if (0 < f* < 10^(-x)) then + // the result is exact + // else // if (f* > T*) then + // the result is inexact + if (ind - 1 <= 2) { + if (fstar.w[1] > ten2mk128trunc[ind - 1].w[1] || + (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] && + fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 + if (fstar.w[2] || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } else { // if 22 <= ind <= 33 + if (fstar.w[3] || fstar.w[2] + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } - // if n < 0 then n cannot be converted to uint64 with RM - if (x_sign) { // if n < 0 and q + exp = 20 - // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 0 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion + } +} + +BID_RETURN (res); +} + +/***************************************************************************** + * BID128_to_uint64_ceil + ****************************************************************************/ + +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, bid128_to_uint64_ceil, + x) + + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent + // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; + + // unpack x +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; + + // check for NaN or Infinity +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { + // x is special +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} + // check for non-canonical values (after the check for special values) +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; - } - exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - // if n > 0 and q + exp = 20 - // if n >= 2^64 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*2^65 - // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n <= -1 then n cannot be converted to uint64 with RZ + // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 + // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 + if (q == 21) { + // C >= a + if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + } else { // if n > 0 and q + exp = 20 + // if n > 2^64 - 1 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) > 2^64 - 1 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 > 2^64 - 1 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 > 10 * (2^64 - 1) + // <=> C * 10^(21-q) > 0x9fffffffffffffff6, 1<=q<=34 if (q == 1) { - // C * 10^20 >= 0xa0000000000000000 - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] >= 0x0a) { - // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { + // C * 10^20 > 0x9fffffffffffffff6 + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] > 0xfffffffffffffff6ull)) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite @@ -947,10 +1523,10 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_floor, x) // else cases that can be rounded to a 64-bit int fall through // to '1 <= q + exp <= 20' } else if (q <= 19) { - // C * 10^(21-q) >= 0xa0000000000000000 - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] >= 0x0a) { - // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { + // C * 10^(21-q) > 0x9fffffffffffffff6 + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] > 0xfffffffffffffff6ull)) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite @@ -960,9 +1536,8 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_floor, x) // else cases that can be rounded to a 64-bit int fall through // to '1 <= q + exp <= 20' } else if (q == 20) { - // C >= 0x10000000000000000 - if (C1.w[1] >= 0x01) { - // actually C1.w[1] == 0x01 && C1.w[0] >= 0x0000000000000000ull) { + // C > 0xffffffffffffffff + if (C1.w[1]) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite @@ -972,9 +1547,9 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_floor, x) // else cases that can be rounded to a 64-bit int fall through // to '1 <= q + exp <= 20' } else if (q == 21) { - // C >= 0xa0000000000000000 - if (C1.w[1] >= 0x0a) { - // actually C1.w[1] == 0x0a && C1.w[0] >= 0x0000000000000000ull) { + // C > 0x9fffffffffffffff6 + if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 + && C1.w[0] > 0xfffffffffffffff6ull)) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite @@ -983,12 +1558,12 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_floor, x) } // else cases that can be rounded to a 64-bit int fall through // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits - C.w[1] = 0x0a; - C.w[0] = 0x0000000000000000ull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C > 10^(q-21) * 0x9fffffffffffffff6 max 44 bits x 68 bits + C.w[1] = 0x09; + C.w[0] = 0xfffffffffffffff6ull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] > C.w[0])) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite @@ -999,189 +1574,612 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_floor, x) // to '1 <= q + exp <= 20' } } - // n is not too large to be converted to int64 if 0 <= n < 2^64 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) <= 0) { // n = +0.[0...0]c(0)c(1)...c(q-1) - // return 0 + } + // n is not too large to be converted to int64 if -1 < n <= 2^64 - 1 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) + // return 0 or 1 + if (x_sign) res = 0x0000000000000000ull; + else + res = 0x0000000000000001ull; + BID_RETURN (res); + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded + // to zero to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; BID_RETURN (res); - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // 1 <= x < 2^64 so x can be rounded - // down to a 64-bit unsigned signed integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 fits in 127 bits - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = C1 * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + } + // 1 <= x <= 2^64 - 1 so x can be rounded + // to zero to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 fits in 127 bits + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = C1 * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + // if the result is positive and inexact, need to add 1 to it + + // determine inexactness of the rounding of C* + // if (0 < f* < 10^(-x)) then + // the result is exact + // else // if (f* > T*) then + // the result is inexact + if (ind - 1 <= 2) { + if (fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + if (!x_sign) { // positive and inexact + Cstar.w[0]++; + if (Cstar.w[0] == 0x0) + Cstar.w[1]++; + } + } // else the result is exact + } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 + if (fstar.w[2] || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + if (!x_sign) { // positive and inexact + Cstar.w[0]++; + if (Cstar.w[0] == 0x0) + Cstar.w[1]++; + } + } // else the result is exact + } else { // if 22 <= ind <= 33 + if (fstar.w[3] || fstar.w[2] + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + if (!x_sign) { // positive and inexact + Cstar.w[0]++; + if (Cstar.w[0] == 0x0) + Cstar.w[1]++; + } + } // else the result is exact } + + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } /***************************************************************************** - * BID128_to_uint64_xfloor + * BID128_to_uint64_xceil ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xfloor, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_xceil, x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} + // check for non-canonical values (after the check for special values) +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n <= -1 then n cannot be converted to uint64 with RZ + // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 + // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 + if (q == 21) { + // C >= a + if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); + } + } else { // if n > 0 and q + exp = 20 + // if n > 2^64 - 1 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) > 2^64 - 1 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 > 2^64 - 1 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 > 10 * (2^64 - 1) + // <=> C * 10^(21-q) > 0x9fffffffffffffff6, 1<=q<=34 + if (q == 1) { + // C * 10^20 > 0x9fffffffffffffff6 + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] > 0xfffffffffffffff6ull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) > 0x9fffffffffffffff6 + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] > 0xfffffffffffffff6ull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C > 0xffffffffffffffff + if (C1.w[1]) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C > 0x9fffffffffffffff6 + if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 + && C1.w[0] > 0xfffffffffffffff6ull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C > 10^(q-21) * 0x9fffffffffffffff6 max 44 bits x 68 bits + C.w[1] = 0x09; + C.w[0] = 0xfffffffffffffff6ull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] > C.w[0])) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } - BID_RETURN (res); } } - // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; + // n is not too large to be converted to int64 if -1 < n <= 2^64 - 1 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + // return 0 or 1 + if (x_sign) + res = 0x0000000000000000ull; + else + res = 0x0000000000000001ull; BID_RETURN (res); - } else { // x is not special and is not zero - - // if n < 0 then n cannot be converted to uint64 with RM - if (x_sign) { // if n < 0 and q + exp = 20 - // too large if c(0)c(1)...c(19).c(20)...c(q-1) > 0 + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded + // to zero to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion + // 1 <= x <= 2^64 - 1 so x can be rounded + // to zero to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 fits in 127 bits + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = C1 * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + // if the result is positive and inexact, need to add 1 to it + + // determine inexactness of the rounding of C* + // if (0 < f* < 10^(-x)) then + // the result is exact + // else // if (f* > T*) then + // the result is inexact + if (ind - 1 <= 2) { + if (fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + if (!x_sign) { // positive and inexact + Cstar.w[0]++; + if (Cstar.w[0] == 0x0) + Cstar.w[1]++; + } + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 + if (fstar.w[2] || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + if (!x_sign) { // positive and inexact + Cstar.w[0]++; + if (Cstar.w[0] == 0x0) + Cstar.w[1]++; + } + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } else { // if 22 <= ind <= 33 + if (fstar.w[3] || fstar.w[2] + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + if (!x_sign) { // positive and inexact + Cstar.w[0]++; + if (Cstar.w[0] == 0x0) + Cstar.w[1]++; + } + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } + + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; + } + } +} + +BID_RETURN (res); +} + +/***************************************************************************** + * BID128_to_uint64_int + ****************************************************************************/ + +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, bid128_to_uint64_int, + x) + + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent + // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 P256; + + // unpack x +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; + + // check for NaN or Infinity +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { + // x is special +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} + // check for non-canonical values (after the check for special values) +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; - } - exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - // if n > 0 and q + exp = 20 + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n <= -1 then n cannot be converted to uint64 with RZ + // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 + // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 + if (q == 21) { + // C >= a + if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + } else { // if n > 0 and q + exp = 20 // if n >= 2^64 then n is too large // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 @@ -1189,7 +2187,7 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xfloor, x) // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 if (q == 1) { // C * 10^20 >= 0xa0000000000000000 - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C if (C.w[1] >= 0x0a) { // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { // set invalid flag @@ -1202,7 +2200,7 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xfloor, x) // to '1 <= q + exp <= 20' } else if (q <= 19) { // C * 10^(21-q) >= 0xa0000000000000000 - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); if (C.w[1] >= 0x0a) { // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { // set invalid flag @@ -1237,12 +2235,13 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xfloor, x) } // else cases that can be rounded to a 64-bit int fall through // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits C.w[1] = 0x0a; C.w[0] = 0x0000000000000000ull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] + || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite @@ -1253,2130 +2252,1155 @@ BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xfloor, x) // to '1 <= q + exp <= 20' } } - // n is not too large to be converted to int64 if 0 <= n < 2^64 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) <= 0) { // n = +0.[0...0]c(0)c(1)...c(q-1) - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - // return 0 - res = 0x0000000000000000ull; + } + // n is not too large to be converted to int64 if -1 < n < 2^64 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) + // return 0 + res = 0x0000000000000000ull; + BID_RETURN (res); + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded + // to zero to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; BID_RETURN (res); - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // 1 <= x < 2^64 so x can be rounded - // down to a 64-bit unsigned signed integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 fits in 127 bits - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = C1 * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - // determine inexactness of the rounding of C* - // if (0 < f* < 10^(-x)) then - // the result is exact - // else // if (f* > T*) then - // the result is inexact - if (ind - 1 <= 2) { - if (fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 - if (fstar.w[2] || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // if 22 <= ind <= 33 - if (fstar.w[3] || fstar.w[2] - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } - - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + } + // 1 <= x < 2^64 so x can be rounded + // to zero to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 fits in 127 bits + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = C1 * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 } + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } /***************************************************************************** - * BID128_to_uint64_ceil + * BID128_to_uint64_xint ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_ceil, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, bid128_to_uint64_xint, + x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; - } - exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n <= -1 then n cannot be converted to uint64 with RZ - // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 - // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 - if (q == 21) { - // C >= a - if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n <= -1 then n cannot be converted to uint64 with RZ + // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 + // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 + if (q == 21) { + // C >= a + if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // if n > 0 and q + exp = 20 - // if n > 2^64 - 1 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) > 2^64 - 1 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 > 2^64 - 1 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 > 10 * (2^64 - 1) - // <=> C * 10^(21-q) > 0x9fffffffffffffff6, 1<=q<=34 - if (q == 1) { - // C * 10^20 > 0x9fffffffffffffff6 - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] > 0xfffffffffffffff6ull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) > 0x9fffffffffffffff6 - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] > 0xfffffffffffffff6ull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C > 0xffffffffffffffff - if (C1.w[1] > 0x0 || (C1.w[1] == 0x0 - && C1.w[0] > 0xffffffffffffffffull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C > 0x9fffffffffffffff6 - if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 - && C1.w[0] > 0xfffffffffffffff6ull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C > 10^(q-21) * 0x9fffffffffffffff6 max 44 bits x 68 bits - C.w[1] = 0x09; - C.w[0] = 0xfffffffffffffff6ull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] > C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } - } - } - // n is not too large to be converted to int64 if -1 < n <= 2^64 - 1 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) - // return 0 or 1 - if (x_sign) - res = 0x0000000000000000ull; - else - res = 0x0000000000000001ull; - BID_RETURN (res); - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded - // to zero to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - // 1 <= x <= 2^64 - 1 so x can be rounded - // to zero to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 fits in 127 bits - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = C1 * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; + } else { // if n > 0 and q + exp = 20 + // if n >= 2^64 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*2^65 + // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0xa0000000000000000 + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] >= 0x0a) { + // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0xa0000000000000000 + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] >= 0x0a) { + // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - // if the result is positive and inexact, need to add 1 to it - - // determine inexactness of the rounding of C* - // if (0 < f* < 10^(-x)) then - // the result is exact - // else // if (f* > T*) then - // the result is inexact - if (ind - 1 <= 2) { - if (fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - if (!x_sign) { // positive and inexact - Cstar.w[0]++; - if (Cstar.w[0] == 0x0) - Cstar.w[1]++; - } - } // else the result is exact - } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 - if (fstar.w[2] || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - if (!x_sign) { // positive and inexact - Cstar.w[0]++; - if (Cstar.w[0] == 0x0) - Cstar.w[1]++; - } - } // else the result is exact - } else { // if 22 <= ind <= 33 - if (fstar.w[3] || fstar.w[2] - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - if (!x_sign) { // positive and inexact - Cstar.w[0]++; - if (Cstar.w[0] == 0x0) - Cstar.w[1]++; - } - } // else the result is exact + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C >= 0x10000000000000000 + if (C1.w[1] >= 0x01) { + // actually C1.w[1] == 0x01 && C1.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; - } - } - } - BID_RETURN (res); -} - -/***************************************************************************** - * BID128_to_uint64_xceil - ****************************************************************************/ - -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xceil, x) - - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent - // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; - - // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; - - // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { - // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0xa0000000000000000 + if (C1.w[1] >= 0x0a) { + // actually C1.w[1] == 0x0a && C1.w[0] >= 0x0000000000000000ull) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits + C.w[1] = 0x0a; + C.w[0] = 0x0000000000000000ull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] + || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } - BID_RETURN (res); } } - // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + // n is not too large to be converted to int64 if -1 < n < 2^64 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + // return 0 res = 0x0000000000000000ull; BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion - x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; - } - exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded + // to zero to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n <= -1 then n cannot be converted to uint64 with RZ - // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 - // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 - if (q == 21) { - // C >= a - if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - } else { // if n > 0 and q + exp = 20 - // if n > 2^64 - 1 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) > 2^64 - 1 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 > 2^64 - 1 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 > 10 * (2^64 - 1) - // <=> C * 10^(21-q) > 0x9fffffffffffffff6, 1<=q<=34 - if (q == 1) { - // C * 10^20 > 0x9fffffffffffffff6 - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] > 0xfffffffffffffff6ull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) > 0x9fffffffffffffff6 - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] > 0xfffffffffffffff6ull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C > 0xffffffffffffffff - if (C1.w[1] > 0x0 || (C1.w[1] == 0x0 - && C1.w[0] > 0xffffffffffffffffull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C > 0x9fffffffffffffff6 - if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 - && C1.w[0] > 0xfffffffffffffff6ull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C > 10^(q-21) * 0x9fffffffffffffff6 max 44 bits x 68 bits - C.w[1] = 0x09; - C.w[0] = 0xfffffffffffffff6ull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] > C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } - } } - // n is not too large to be converted to int64 if -1 < n <= 2^64 - 1 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - // return 0 or 1 - if (x_sign) - res = 0x0000000000000000ull; - else - res = 0x0000000000000001ull; - BID_RETURN (res); - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded - // to zero to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); + // 1 <= x < 2^64 so x can be rounded + // to zero to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 fits in 127 bits + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = C1 * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; } - // 1 <= x <= 2^64 - 1 so x can be rounded - // to zero to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 fits in 127 bits - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = C1 * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - // if the result is positive and inexact, need to add 1 to it - - // determine inexactness of the rounding of C* - // if (0 < f* < 10^(-x)) then - // the result is exact - // else // if (f* > T*) then - // the result is inexact - if (ind - 1 <= 2) { - if (fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - if (!x_sign) { // positive and inexact - Cstar.w[0]++; - if (Cstar.w[0] == 0x0) - Cstar.w[1]++; - } - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 - if (fstar.w[2] || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - if (!x_sign) { // positive and inexact - Cstar.w[0]++; - if (Cstar.w[0] == 0x0) - Cstar.w[1]++; - } - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // if 22 <= ind <= 33 - if (fstar.w[3] || fstar.w[2] - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - if (!x_sign) { // positive and inexact - Cstar.w[0]++; - if (Cstar.w[0] == 0x0) - Cstar.w[1]++; - } - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } - - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 } + // determine inexactness of the rounding of C* + // if (0 < f* < 10^(-x)) then + // the result is exact + // else // if (f* > T*) then + // the result is inexact + if (ind - 1 <= 2) { + if (fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 + if (fstar.w[2] || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } else { // if 22 <= ind <= 33 + if (fstar.w[3] || fstar.w[2] + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } // else the result is exact + } + + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } /***************************************************************************** - * BID128_to_uint64_int + * BID128_to_uint64_rninta ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_int, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_rninta, x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 P256; + UINT64 tmp64; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - exp = (x_exp >> 49) - 6176; - - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n <= -1 then n cannot be converted to uint64 with RZ - // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 - // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 - if (q == 21) { - // C >= a - if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n <= -1/2 then n cannot be converted to uint64 with RN + // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1/2 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x05, 1<=q<=34 + // <=> C * 10^(21-q) >= 0x05, 1<=q<=34 + if (q == 21) { + // C >= 5 + if (C1.w[1] != 0 || C1.w[0] >= 0x05ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // if n > 0 and q + exp = 20 - // if n >= 2^64 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*2^65 - // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 - if (q == 1) { - // C * 10^20 >= 0xa0000000000000000 - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] >= 0x0a) { - // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) >= 0xa0000000000000000 - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] >= 0x0a) { - // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C >= 0x10000000000000000 - if (C1.w[1] >= 0x01) { - // actually C1.w[1] == 0x01 && C1.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C >= 0xa0000000000000000 - if (C1.w[1] >= 0x0a) { - // actually C1.w[1] == 0x0a && C1.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits - C.w[1] = 0x0a; - C.w[0] = 0x0000000000000000ull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] - && C1.w[0] >= C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } - } - } - // n is not too large to be converted to int64 if -1 < n < 2^64 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) - // return 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded - // to zero to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) >= 5 is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - // 1 <= x < 2^64 so x can be rounded - // to zero to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 fits in 127 bits - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = C1 * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; + } else { // if n > 0 and q + exp = 20 + // if n >= 2^64 - 1/2 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) + // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0x9fffffffffffffffb + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0x9fffffffffffffffb + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff + C.w[0] = C1.w[0] + C1.w[0]; + C.w[1] = C1.w[1] + C1.w[1]; + if (C.w[0] < C1.w[0]) + C.w[1]++; + if (C.w[1] > 0x01 || (C.w[1] == 0x01 + && C.w[0] >= 0xffffffffffffffffull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0x9fffffffffffffffb + if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 + && C1.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits + C.w[1] = 0x09; + C.w[0] = 0xfffffffffffffffbull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] + || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } } } - BID_RETURN (res); -} - -/***************************************************************************** - * BID128_to_uint64_xint - ****************************************************************************/ - -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xint, x) - - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent - // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; - - // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; - - // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { - // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN + // n is not too large to be converted to int64 if -1/2 < n < 2^64 - 1/2 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) + // return 0 + res = 0x0000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) + // if 0.c(0)c(1)...c(q-1) < 0.5 <=> c(0)c(1)...c(q-1) < 5 * 10^(q-1) + // res = 0 + // else if x > 0 + // res = +1 + // else // if x < 0 + // invalid exc + ind = q - 1; + if (ind <= 18) { // 0 <= ind <= 18 + if ((C1.w[1] == 0) && (C1.w[0] < midpoint64[ind])) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf + } else { // 19 <= ind <= 33 + if ((C1.w[1] < midpoint128[ind - 19].w[1]) + || ((C1.w[1] == midpoint128[ind - 19].w[1]) + && (C1.w[0] < midpoint128[ind - 19].w[0]))) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); } - BID_RETURN (res); - } - } - // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion - x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; } - exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded + // to nearest to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n <= -1 then n cannot be converted to uint64 with RZ - // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x0a, 1<=q<=34 - // <=> C * 10^(21-q) >= 0x0a, 1<=q<=34 - if (q == 21) { - // C >= a - if (C1.w[1] != 0 || C1.w[0] >= 0x0aull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) >= a is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= a * 10^(q-21) is true because C > 2^64 and a*10^(q-21) < 2^64 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - } else { // if n > 0 and q + exp = 20 - // if n >= 2^64 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*2^65 - // <=> C * 10^(21-q) >= 0xa0000000000000000, 1<=q<=34 - if (q == 1) { - // C * 10^20 >= 0xa0000000000000000 - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] >= 0x0a) { - // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) >= 0xa0000000000000000 - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] >= 0x0a) { - // actually C.w[1] == 0x0a && C.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C >= 0x10000000000000000 - if (C1.w[1] >= 0x01) { - // actually C1.w[1] == 0x01 && C1.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C >= 0xa0000000000000000 - if (C1.w[1] >= 0x0a) { - // actually C1.w[1] == 0x0a && C1.w[0] >= 0x0000000000000000ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0xa0000000000000000 max 44 bits x 68 bits - C.w[1] = 0x0a; - C.w[0] = 0x0000000000000000ull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } - } } - // n is not too large to be converted to int64 if -1 < n < 2^64 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) <= 0) { // n = +/-0.[0...0]c(0)c(1)...c(q-1) - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - // return 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64 so if positive x can be rounded - // to zero to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // 1 <= x < 2^64 so x can be rounded - // to zero to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 fits in 127 bits - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = C1 * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - // determine inexactness of the rounding of C* - // if (0 < f* < 10^(-x)) then - // the result is exact - // else // if (f* > T*) then - // the result is inexact - if (ind - 1 <= 2) { - if (fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 - if (fstar.w[2] || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // if 22 <= ind <= 33 - if (fstar.w[3] || fstar.w[2] - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } - - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; + // 1 <= x < 2^64-1/2 so x can be rounded + // to nearest to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits + tmp64 = C1.w[0]; + if (ind <= 19) { + C1.w[0] = C1.w[0] + midpoint64[ind - 1]; } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + C1.w[0] = C1.w[0] + midpoint128[ind - 20].w[0]; + C1.w[1] = C1.w[1] + midpoint128[ind - 20].w[1]; + } + if (C1.w[0] < tmp64) + C1.w[1]++; + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = (C1 + 1/2 * 10^x) * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; } + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // if (0 < f* < 10^(-x)) then the result is a midpoint + // if floor(C*) is even then C* = floor(C*) - logical right + // shift; C* has p decimal digits, correct by Prop. 1) + // else if floor(C*) is odd C* = floor(C*)-1 (logical right + // shift; C* has p decimal digits, correct by Pr. 1) + // else + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + + // if the result was a midpoint it was rounded away from zero + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } /***************************************************************************** - * BID128_to_uint64_rninta + * BID128_to_uint64_xrninta ****************************************************************************/ -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_rninta, x) +BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE (UINT64, + bid128_to_uint64_xrninta, x) - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent + UINT64 res; + UINT64 x_sign; + UINT64 x_exp; + int exp; // unbiased exponent // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - UINT64 tmp64; - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 P256; + UINT64 tmp64, tmp64A; + BID_UI64DOUBLE tmp1; + unsigned int x_nr_bits; + int q, ind, shift; + UINT128 C1, C; + UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits + UINT256 fstar; + UINT256 P256; // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; +x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative +x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions +C1.w[1] = x.w[1] & MASK_COEFF; +C1.w[0] = x.w[0]; // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { +if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is QNaN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - } - BID_RETURN (res); - } +if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN + if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is QNaN + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; } + BID_RETURN (res); +} else { // x is not a NaN, so it must be infinity + if (!x_sign) { // x is +inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } else { // x is -inf + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + } + BID_RETURN (res); +} +} // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion +if ((C1.w[1] > 0x0001ed09bead87c0ull) + || (C1.w[1] == 0x0001ed09bead87c0ull + && (C1.w[0] > 0x378d8e63ffffffffull)) + || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { + res = 0x0000000000000000ull; + BID_RETURN (res); +} else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { + // x is 0 + res = 0x0000000000000000ull; + BID_RETURN (res); +} else { // x is not special and is not zero + + // q = nr. of decimal digits in x + // determine first the nr. of bits in x + if (C1.w[1] == 0) { + if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 + // split the 64-bit value in two 32-bit halves to avoid rounding errors + if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 + tmp1.d = (double) (C1.w[0] >> 32); // exact conversion + x_nr_bits = + 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } else { // x < 2^32 + tmp1.d = (double) (C1.w[0]); // exact conversion x_nr_bits = 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion - x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; + } else { // if x < 2^53 + tmp1.d = (double) C1.w[0]; // exact conversion + x_nr_bits = + 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); } - exp = (x_exp >> 49) - 6176; - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n <= -1/2 then n cannot be converted to uint64 with RN - // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1/2 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x05, 1<=q<=34 - // <=> C * 10^(21-q) >= 0x05, 1<=q<=34 - if (q == 21) { - // C >= 5 - if (C1.w[1] != 0 || C1.w[0] >= 0x05ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) >= 5 is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 + } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) + tmp1.d = (double) C1.w[1]; // exact conversion + x_nr_bits = + 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); + } + q = nr_digits[x_nr_bits - 1].digits; + if (q == 0) { + q = nr_digits[x_nr_bits - 1].digits1; + if (C1.w[1] > nr_digits[x_nr_bits - 1].threshold_hi + || (C1.w[1] == nr_digits[x_nr_bits - 1].threshold_hi + && C1.w[0] >= nr_digits[x_nr_bits - 1].threshold_lo)) + q++; + } + exp = (x_exp >> 49) - 6176; + + if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) + // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... + // so x rounded to an integer may or may not fit in an unsigned 64-bit int + // the cases that do not fit are identified here; the ones that fit + // fall through and will be handled with other cases further, + // under '1 <= q + exp <= 20' + if (x_sign) { // if n < 0 and q + exp = 20 + // if n <= -1/2 then n cannot be converted to uint64 with RN + // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1/2 + // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x05, 1<=q<=34 + // <=> C * 10^(21-q) >= 0x05, 1<=q<=34 + if (q == 21) { + // C >= 5 + if (C1.w[1] != 0 || C1.w[0] >= 0x05ull) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // if n > 0 and q + exp = 20 - // if n >= 2^64 - 1/2 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) - // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 - if (q == 1) { - // C * 10^20 >= 0x9fffffffffffffffb - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) >= 0x9fffffffffffffffb - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff - C.w[0] = C1.w[0] + C1.w[0]; - C.w[1] = C1.w[1] + C1.w[1]; - if (C.w[0] < C1.w[0]) - C.w[1]++; - if (C.w[1] > 0x01 || (C.w[1] == 0x01 - && C.w[0] >= 0xffffffffffffffffull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C >= 0x9fffffffffffffffb - if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 - && C1.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits - C.w[1] = 0x09; - C.w[0] = 0xfffffffffffffffbull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } + // else cases that can be rounded to 64-bit unsigned int fall through + // to '1 <= q + exp <= 20' + } else { + // if 1 <= q <= 20 + // C * 10^(21-q) >= 5 is true because C >= 1 and 10^(21-q) >= 10 + // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - } - // n is not too large to be converted to int64 if -1/2 < n < 2^64 - 1/2 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) - // return 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) - // if 0.c(0)c(1)...c(q-1) < 0.5 <=> c(0)c(1)...c(q-1) < 5 * 10^(q-1) - // res = 0 - // else if x > 0 - // res = +1 - // else // if x < 0 - // invalid exc - ind = q - 1; - if (ind <= 18) { // 0 <= ind <= 18 - if ((C1.w[1] == 0) && (C1.w[0] < __bid_midpoint64[ind])) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { + } else { // if n > 0 and q + exp = 20 + // if n >= 2^64 - 1/2 then n is too large + // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 + // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) + // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 + if (q == 1) { + // C * 10^20 >= 0x9fffffffffffffffb + __mul_128x64_to_128 (C, C1.w[0], ten2k128[0]); // 10^20 * C + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } else { // 19 <= ind <= 33 - if ((C1.w[1] < __bid_midpoint128[ind - 19].w[1]) - || ((C1.w[1] == __bid_midpoint128[ind - 19].w[1]) - && (C1.w[0] < __bid_midpoint128[ind - 19].w[0]))) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q <= 19) { + // C * 10^(21-q) >= 0x9fffffffffffffffb + __mul_64x64_to_128MACH (C, C1.w[0], ten2k64[21 - q]); + if (C.w[1] > 0x09 || (C.w[1] == 0x09 + && C.w[0] >= 0xfffffffffffffffbull)) { // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); } - } - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded - // to nearest to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // 1 <= x < 2^64-1/2 so x can be rounded - // to nearest to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits - tmp64 = C1.w[0]; - if (ind <= 19) { - C1.w[0] = C1.w[0] + __bid_midpoint64[ind - 1]; - } else { - C1.w[0] = C1.w[0] + __bid_midpoint128[ind - 20].w[0]; - C1.w[1] = C1.w[1] + __bid_midpoint128[ind - 20].w[1]; + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 20) { + // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff + C.w[0] = C1.w[0] + C1.w[0]; + C.w[1] = C1.w[1] + C1.w[1]; + if (C.w[0] < C1.w[0]) + C.w[1]++; + if (C.w[1] > 0x01 || (C.w[1] == 0x01 + && C.w[0] >= 0xffffffffffffffffull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - if (C1.w[0] < tmp64) - C1.w[1]++; - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = (C1 + 1/2 * 10^x) * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else if (q == 21) { + // C >= 0x9fffffffffffffffb + if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 + && C1.w[0] >= 0xfffffffffffffffbull)) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // if (0 < f* < 10^(-x)) then the result is a midpoint - // if floor(C*) is even then C* = floor(C*) - logical right - // shift; C* has p decimal digits, correct by Prop. 1) - // else if floor(C*) is odd C* = floor(C*)-1 (logical right - // shift; C* has p decimal digits, correct by Pr. 1) - // else - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' + } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 + // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits + C.w[1] = 0x09; + C.w[0] = 0xfffffffffffffffbull; + __mul_128x64_to_128 (C, ten2k64[q - 21], C); + if (C1.w[1] > C.w[1] + || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { + // set invalid flag + *pfpsf |= INVALID_EXCEPTION; + // return Integer Indefinite + res = 0x8000000000000000ull; + BID_RETURN (res); } - - // if the result was a midpoint it was rounded away from zero - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; + // else cases that can be rounded to a 64-bit int fall through + // to '1 <= q + exp <= 20' } } } - BID_RETURN (res); -} - -/***************************************************************************** - * BID128_to_uint64_xrninta - ****************************************************************************/ - -BID128_FUNCTION_ARG1_NORND_CUSTOMRESTYPE(UINT64, __bid128_to_uint64_xrninta, x) - - UINT64 res; - UINT64 x_sign; - UINT64 x_exp; - int exp; // unbiased exponent - // Note: C1.w[1], C1.w[0] represent x_signif_hi, x_signif_lo (all are UINT64) - UINT64 tmp64, tmp64A; - BID_UI64DOUBLE tmp1; - unsigned int x_nr_bits; - int q, ind, shift; - UINT128 C1, C; - UINT128 Cstar; // C* represents up to 34 decimal digits ~ 113 bits - UINT256 fstar; - UINT256 P256; - - // unpack x - x_sign = x.w[1] & MASK_SIGN; // 0 for positive, MASK_SIGN for negative - x_exp = x.w[1] & MASK_EXP; // biased and shifted left 49 bit positions - C1.w[1] = x.w[1] & MASK_COEFF; - C1.w[0] = x.w[0]; - - // check for NaN or Infinity - if ((x.w[1] & MASK_SPECIAL) == MASK_SPECIAL) { - // x is special - if ((x.w[1] & MASK_NAN) == MASK_NAN) { // x is NAN - if ((x.w[1] & MASK_SNAN) == MASK_SNAN) { // x is SNAN - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite + // n is not too large to be converted to int64 if -1/2 < n < 2^64 - 1/2 + // Note: some of the cases tested for above fall through to this point + if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + // return 0 + res = 0x0000000000000000ull; + BID_RETURN (res); + } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) + // if 0.c(0)c(1)...c(q-1) < 0.5 <=> c(0)c(1)...c(q-1) < 5 * 10^(q-1) + // res = 0 + // else if x > 0 + // res = +1 + // else // if x < 0 + // invalid exc + ind = q - 1; + if (ind <= 18) { // 0 <= ind <= 18 + if ((C1.w[1] == 0) && (C1.w[0] < midpoint64[ind])) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { res = 0x8000000000000000ull; - } else { // x is QNaN // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); } - BID_RETURN (res); - } else { // x is not a NaN, so it must be infinity - if (!x_sign) { // x is +inf - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite + } else { // 19 <= ind <= 33 + if ((C1.w[1] < midpoint128[ind - 19].w[1]) + || ((C1.w[1] == midpoint128[ind - 19].w[1]) + && (C1.w[0] < midpoint128[ind - 19].w[0]))) { + res = 0x0000000000000000ull; // return 0 + } else if (!x_sign) { // n > 0 + res = 0x00000001; // return +1 + } else { res = 0x8000000000000000ull; - } else { // x is -inf - // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; + BID_RETURN (res); } - BID_RETURN (res); - } - } - // check for non-canonical values (after the check for special values) - if ((C1.w[1] > 0x0001ed09bead87c0ull) - || (C1.w[1] == 0x0001ed09bead87c0ull - && (C1.w[0] > 0x378d8e63ffffffffull)) - || ((x.w[1] & 0x6000000000000000ull) == 0x6000000000000000ull)) { - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((C1.w[1] == 0x0ull) && (C1.w[0] == 0x0ull)) { - // x is 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else { // x is not special and is not zero - - // q = nr. of decimal digits in x - // determine first the nr. of bits in x - if (C1.w[1] == 0) { - if (C1.w[0] >= 0x0020000000000000ull) { // x >= 2^53 - // split the 64-bit value in two 32-bit halves to avoid rounding errors - if (C1.w[0] >= 0x0000000100000000ull) { // x >= 2^32 - tmp1.d = (double) (C1.w[0] >> 32); // exact conversion - x_nr_bits = - 33 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } else { // x < 2^32 - tmp1.d = (double) (C1.w[0]); // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // if x < 2^53 - tmp1.d = (double) C1.w[0]; // exact conversion - x_nr_bits = - 1 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - } else { // C1.w[1] != 0 => nr. bits = 64 + nr_bits (C1.w[1]) - tmp1.d = (double) C1.w[1]; // exact conversion - x_nr_bits = - 65 + ((((unsigned int) (tmp1.ui64 >> 52)) & 0x7ff) - 0x3ff); - } - q = __bid_nr_digits[x_nr_bits - 1].digits; - if (q == 0) { - q = __bid_nr_digits[x_nr_bits - 1].digits1; - if (C1.w[1] > __bid_nr_digits[x_nr_bits - 1].threshold_hi - || (C1.w[1] == __bid_nr_digits[x_nr_bits - 1].threshold_hi - && C1.w[0] >= __bid_nr_digits[x_nr_bits - 1].threshold_lo)) - q++; } - exp = (x_exp >> 49) - 6176; - - if ((q + exp) > 20) { // x >= 10^20 ~= 2^66.45... (cannot fit in 64 bits) + // set inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) + // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded + // to nearest to a 64-bit unsigned signed integer + if (x_sign) { // x <= -1 // set invalid flag *pfpsf |= INVALID_EXCEPTION; // return Integer Indefinite res = 0x8000000000000000ull; BID_RETURN (res); - } else if ((q + exp) == 20) { // x = c(0)c(1)...c(19).c(20)...c(q-1) - // in this case 2^63.11... ~= 10^19 <= x < 10^20 ~= 2^66.43... - // so x rounded to an integer may or may not fit in an unsigned 64-bit int - // the cases that do not fit are identified here; the ones that fit - // fall through and will be handled with other cases further, - // under '1 <= q + exp <= 20' - if (x_sign) { // if n < 0 and q + exp = 20 - // if n <= -1/2 then n cannot be converted to uint64 with RN - // too large if c(0)c(1)...c(19).c(20)...c(q-1) >= 1/2 - // <=> 0.c(0)c(1)...c(q-1) * 10^21 >= 0x05, 1<=q<=34 - // <=> C * 10^(21-q) >= 0x05, 1<=q<=34 - if (q == 21) { - // C >= 5 - if (C1.w[1] != 0 || C1.w[0] >= 0x05ull) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to 64-bit unsigned int fall through - // to '1 <= q + exp <= 20' - } else { - // if 1 <= q <= 20 - // C * 10^(21-q) >= 5 is true because C >= 1 and 10^(21-q) >= 10 - // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 5 * 10^(q-21) is true because C > 2^64 and 5*10^(q-21) < 2^64 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - } else { // if n > 0 and q + exp = 20 - // if n >= 2^64 - 1/2 then n is too large - // <=> c(0)c(1)...c(19).c(20)...c(q-1) >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^20 >= 2^64-1/2 - // <=> 0.c(0)c(1)...c(19)c(20)...c(q-1) * 10^21 >= 5*(2^65-1) - // <=> C * 10^(21-q) >= 0x9fffffffffffffffb, 1<=q<=34 - if (q == 1) { - // C * 10^20 >= 0x9fffffffffffffffb - __mul_128x64_to_128 (C, C1.w[0], __bid_ten2k128[0]); // 10^20 * C - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q <= 19) { - // C * 10^(21-q) >= 0x9fffffffffffffffb - __mul_64x64_to_128MACH (C, C1.w[0], __bid_ten2k64[21 - q]); - if (C.w[1] > 0x09 || (C.w[1] == 0x09 - && C.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 20) { - // C * 10 >= 0x9fffffffffffffffb <=> C * 2 > 1ffffffffffffffff - C.w[0] = C1.w[0] + C1.w[0]; - C.w[1] = C1.w[1] + C1.w[1]; - if (C.w[0] < C1.w[0]) - C.w[1]++; - if (C.w[1] > 0x01 || (C.w[1] == 0x01 - && C.w[0] >= 0xffffffffffffffffull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else if (q == 21) { - // C >= 0x9fffffffffffffffb - if (C1.w[1] > 0x09 || (C1.w[1] == 0x09 - && C1.w[0] >= 0xfffffffffffffffbull)) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } else { // if 22 <= q <= 34 => 1 <= q - 21 <= 13 - // C >= 10^(q-21) * 0x9fffffffffffffffb max 44 bits x 68 bits - C.w[1] = 0x09; - C.w[0] = 0xfffffffffffffffbull; - __mul_128x64_to_128 (C, __bid_ten2k64[q - 21], C); - if (C1.w[1] > C.w[1] || (C1.w[1] == C.w[1] && C1.w[0] >= C.w[0])) { - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - // else cases that can be rounded to a 64-bit int fall through - // to '1 <= q + exp <= 20' - } - } } - // n is not too large to be converted to int64 if -1/2 < n < 2^64 - 1/2 - // Note: some of the cases tested for above fall through to this point - if ((q + exp) < 0) { // n = +/-0.0...c(0)c(1)...c(q-1) - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - // return 0 - res = 0x0000000000000000ull; - BID_RETURN (res); - } else if ((q + exp) == 0) { // n = +/-0.c(0)c(1)...c(q-1) - // if 0.c(0)c(1)...c(q-1) < 0.5 <=> c(0)c(1)...c(q-1) < 5 * 10^(q-1) - // res = 0 - // else if x > 0 - // res = +1 - // else // if x < 0 - // invalid exc - ind = q - 1; - if (ind <= 18) { // 0 <= ind <= 18 - if ((C1.w[1] == 0) && (C1.w[0] < __bid_midpoint64[ind])) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { - res = 0x8000000000000000ull; - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } - } else { // 19 <= ind <= 33 - if ((C1.w[1] < __bid_midpoint128[ind - 19].w[1]) - || ((C1.w[1] == __bid_midpoint128[ind - 19].w[1]) - && (C1.w[0] < __bid_midpoint128[ind - 19].w[0]))) { - res = 0x0000000000000000ull; // return 0 - } else if (!x_sign) { // n > 0 - res = 0x00000001; // return +1 - } else { - res = 0x8000000000000000ull; - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); - } + // 1 <= x < 2^64-1/2 so x can be rounded + // to nearest to a 64-bit unsigned integer + if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 + ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' + // chop off ind digits from the lower part of C1 + // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits + tmp64 = C1.w[0]; + if (ind <= 19) { + C1.w[0] = C1.w[0] + midpoint64[ind - 1]; + } else { + C1.w[0] = C1.w[0] + midpoint128[ind - 20].w[0]; + C1.w[1] = C1.w[1] + midpoint128[ind - 20].w[1]; } - // set inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } else { // if (1 <= q + exp <= 20, 1 <= q <= 34, -33 <= exp <= 19) - // x <= -1 or 1 <= x < 2^64-1/2 so if positive x can be rounded - // to nearest to a 64-bit unsigned signed integer - if (x_sign) { // x <= -1 - // set invalid flag - *pfpsf |= INVALID_EXCEPTION; - // return Integer Indefinite - res = 0x8000000000000000ull; - BID_RETURN (res); + if (C1.w[0] < tmp64) + C1.w[1]++; + // calculate C* and f* + // C* is actually floor(C*) in this case + // C* and f* need shifting and masking, as shown by + // shiftright128[] and maskhigh128[] + // 1 <= x <= 33 + // kx = 10^(-x) = ten2mk128[ind - 1] + // C* = (C1 + 1/2 * 10^x) * 10^(-x) + // the approximation of 10^(-x) was rounded up to 118 bits + __mul_128x128_to_256 (P256, C1, ten2mk128[ind - 1]); + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[1] = P256.w[3]; + Cstar.w[0] = P256.w[2]; + fstar.w[3] = 0; + fstar.w[2] = P256.w[2] & maskhigh128[ind - 1]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; + } else { // 22 <= ind - 1 <= 33 + Cstar.w[1] = 0; + Cstar.w[0] = P256.w[3]; + fstar.w[3] = P256.w[3] & maskhigh128[ind - 1]; + fstar.w[2] = P256.w[2]; + fstar.w[1] = P256.w[1]; + fstar.w[0] = P256.w[0]; } - // 1 <= x < 2^64-1/2 so x can be rounded - // to nearest to a 64-bit unsigned integer - if (exp < 0) { // 2 <= q <= 34, -33 <= exp <= -1, 1 <= q + exp <= 20 - ind = -exp; // 1 <= ind <= 33; ind is a synonym for 'x' - // chop off ind digits from the lower part of C1 - // C1 = C1 + 1/2 * 10^ind where the result C1 fits in 127 bits - tmp64 = C1.w[0]; - if (ind <= 19) { - C1.w[0] = C1.w[0] + __bid_midpoint64[ind - 1]; - } else { - C1.w[0] = C1.w[0] + __bid_midpoint128[ind - 20].w[0]; - C1.w[1] = C1.w[1] + __bid_midpoint128[ind - 20].w[1]; - } - if (C1.w[0] < tmp64) - C1.w[1]++; - // calculate C* and f* - // C* is actually floor(C*) in this case - // C* and f* need shifting and masking, as shown by - // __bid_shiftright128[] and __bid_maskhigh128[] - // 1 <= x <= 33 - // kx = 10^(-x) = __bid_ten2mk128[ind - 1] - // C* = (C1 + 1/2 * 10^x) * 10^(-x) - // the approximation of 10^(-x) was rounded up to 118 bits - __mul_128x128_to_256 (P256, C1, __bid_ten2mk128[ind - 1]); - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[1] = P256.w[3]; - Cstar.w[0] = P256.w[2]; - fstar.w[3] = 0; - fstar.w[2] = P256.w[2] & __bid_maskhigh128[ind - 1]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } else { // 22 <= ind - 1 <= 33 - Cstar.w[1] = 0; - Cstar.w[0] = P256.w[3]; - fstar.w[3] = P256.w[3] & __bid_maskhigh128[ind - 1]; - fstar.w[2] = P256.w[2]; - fstar.w[1] = P256.w[1]; - fstar.w[0] = P256.w[0]; - } - // the top Ex bits of 10^(-x) are T* = __bid_ten2mk128trunc[ind], e.g. - // if x=1, T*=__bid_ten2mk128trunc[0]=0x19999999999999999999999999999999 - // if (0 < f* < 10^(-x)) then the result is a midpoint - // if floor(C*) is even then C* = floor(C*) - logical right - // shift; C* has p decimal digits, correct by Prop. 1) - // else if floor(C*) is odd C* = floor(C*)-1 (logical right - // shift; C* has p decimal digits, correct by Pr. 1) - // else - // C* = floor(C*) (logical right shift; C has p decimal digits, - // correct by Property 1) - // n = C* * 10^(e+x) - - // shift right C* by Ex-128 = __bid_shiftright128[ind] - shift = __bid_shiftright128[ind - 1]; // 0 <= shift <= 102 - if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 - Cstar.w[0] = - (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); - // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); - } else { // 22 <= ind - 1 <= 33 - Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 - } - // determine inexactness of the rounding of C* - // if (0 < f* - 1/2 < 10^(-x)) then - // the result is exact - // else // if (f* - 1/2 > T*) then - // the result is inexact - if (ind - 1 <= 2) { - if (fstar.w[1] > 0x8000000000000000ull || - (fstar.w[1] == 0x8000000000000000ull && fstar.w[0] > 0x0ull)) { - // f* > 1/2 and the result may be exact - tmp64 = fstar.w[1] - 0x8000000000000000ull; // f* - 1/2 - if ((tmp64 > __bid_ten2mk128trunc[ind - 1].w[1] - || (tmp64 == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] >= __bid_ten2mk128trunc[ind - 1].w[0]))) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // the result is inexact; f2* <= 1/2 + // the top Ex bits of 10^(-x) are T* = ten2mk128trunc[ind], e.g. + // if x=1, T*=ten2mk128trunc[0]=0x19999999999999999999999999999999 + // if (0 < f* < 10^(-x)) then the result is a midpoint + // if floor(C*) is even then C* = floor(C*) - logical right + // shift; C* has p decimal digits, correct by Prop. 1) + // else if floor(C*) is odd C* = floor(C*)-1 (logical right + // shift; C* has p decimal digits, correct by Pr. 1) + // else + // C* = floor(C*) (logical right shift; C has p decimal digits, + // correct by Property 1) + // n = C* * 10^(e+x) + + // shift right C* by Ex-128 = shiftright128[ind] + shift = shiftright128[ind - 1]; // 0 <= shift <= 102 + if (ind - 1 <= 21) { // 0 <= ind - 1 <= 21 + Cstar.w[0] = + (Cstar.w[0] >> shift) | (Cstar.w[1] << (64 - shift)); + // redundant, it will be 0! Cstar.w[1] = (Cstar.w[1] >> shift); + } else { // 22 <= ind - 1 <= 33 + Cstar.w[0] = (Cstar.w[0] >> (shift - 64)); // 2 <= shift - 64 <= 38 + } + // determine inexactness of the rounding of C* + // if (0 < f* - 1/2 < 10^(-x)) then + // the result is exact + // else // if (f* - 1/2 > T*) then + // the result is inexact + if (ind - 1 <= 2) { + if (fstar.w[1] > 0x8000000000000000ull || + (fstar.w[1] == 0x8000000000000000ull + && fstar.w[0] > 0x0ull)) { + // f* > 1/2 and the result may be exact + tmp64 = fstar.w[1] - 0x8000000000000000ull; // f* - 1/2 + if (tmp64 > ten2mk128trunc[ind - 1].w[1] + || (tmp64 == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] >= ten2mk128trunc[ind - 1].w[0])) { // set the inexact flag *pfpsf |= INEXACT_EXCEPTION; - } - } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 - if (fstar.w[3] > 0x0 || (fstar.w[3] == 0x0 - && fstar.w[2] > __bid_one_half128[ind - 1]) || (fstar.w[3] == 0x0 - && fstar.w[2] == __bid_one_half128[ind - 1] && (fstar.w[1] - || fstar.w[0]))) { - // f2* > 1/2 and the result may be exact - // Calculate f2* - 1/2 - tmp64 = fstar.w[2] - __bid_one_half128[ind - 1]; - tmp64A = fstar.w[3]; - if (tmp64 > fstar.w[2]) - tmp64A--; - if (tmp64A || tmp64 - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // the result is inexact; f2* <= 1/2 + } // else the result is exact + } else { // the result is inexact; f2* <= 1/2 + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } + } else if (ind - 1 <= 21) { // if 3 <= ind <= 21 + if (fstar.w[3] > 0x0 || + (fstar.w[3] == 0x0 && fstar.w[2] > onehalf128[ind - 1]) || + (fstar.w[3] == 0x0 && fstar.w[2] == onehalf128[ind - 1] && + (fstar.w[1] || fstar.w[0]))) { + // f2* > 1/2 and the result may be exact + // Calculate f2* - 1/2 + tmp64 = fstar.w[2] - onehalf128[ind - 1]; + tmp64A = fstar.w[3]; + if (tmp64 > fstar.w[2]) + tmp64A--; + if (tmp64A || tmp64 + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { // set the inexact flag *pfpsf |= INEXACT_EXCEPTION; - } - } else { // if 22 <= ind <= 33 - if (fstar.w[3] > __bid_one_half128[ind - 1] || - (fstar.w[3] == __bid_one_half128[ind - 1] && (fstar.w[2] || - fstar.w[1] || fstar.w[0]))) { - // f2* > 1/2 and the result may be exact - // Calculate f2* - 1/2 - tmp64 = fstar.w[3] - __bid_one_half128[ind - 1]; - if (tmp64 || fstar.w[2] - || fstar.w[1] > __bid_ten2mk128trunc[ind - 1].w[1] - || (fstar.w[1] == __bid_ten2mk128trunc[ind - 1].w[1] - && fstar.w[0] > __bid_ten2mk128trunc[ind - 1].w[0])) { - // set the inexact flag - *pfpsf |= INEXACT_EXCEPTION; - } // else the result is exact - } else { // the result is inexact; f2* <= 1/2 + } // else the result is exact + } else { // the result is inexact; f2* <= 1/2 + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; + } + } else { // if 22 <= ind <= 33 + if (fstar.w[3] > onehalf128[ind - 1] || + (fstar.w[3] == onehalf128[ind - 1] && + (fstar.w[2] || fstar.w[1] || fstar.w[0]))) { + // f2* > 1/2 and the result may be exact + // Calculate f2* - 1/2 + tmp64 = fstar.w[3] - onehalf128[ind - 1]; + if (tmp64 || fstar.w[2] + || fstar.w[1] > ten2mk128trunc[ind - 1].w[1] + || (fstar.w[1] == ten2mk128trunc[ind - 1].w[1] + && fstar.w[0] > ten2mk128trunc[ind - 1].w[0])) { // set the inexact flag *pfpsf |= INEXACT_EXCEPTION; - } + } // else the result is exact + } else { // the result is inexact; f2* <= 1/2 + // set the inexact flag + *pfpsf |= INEXACT_EXCEPTION; } - - // if the result was a midpoint it was rounded away from zero - res = Cstar.w[0]; // the result is positive - } else if (exp == 0) { - // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 - // res = C (exact) - res = C1.w[0]; - } else { - // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 - // res = C * 10^exp (exact) - must fit in 64 bits - res = C1.w[0] * __bid_ten2k64[exp]; } + + // if the result was a midpoint it was rounded away from zero + res = Cstar.w[0]; // the result is positive + } else if (exp == 0) { + // 1 <= q <= 20, but x < 2^64 - 1/2 so in this case C1.w[1] has to be 0 + // res = C (exact) + res = C1.w[0]; + } else { + // if (exp > 0) => 1 <= exp <= 19, 1 <= q < 19, 2 <= q + exp <= 20 + // res = C * 10^exp (exact) - must fit in 64 bits + res = C1.w[0] * ten2k64[exp]; } } - BID_RETURN (res); +} + +BID_RETURN (res); } |