glibc/sysdeps/ieee754/ldbl-128ibm/mpn2ldbl.c
Alan Modra 4cf69995e2 Fix for [BZ #15680] IBM long double inaccuracy
http://sourceware.org/ml/libc-alpha/2013-06/msg00919.html

I discovered a number of places where denormals and other corner cases
were being handled wrongly.

- printf_fphex.c: Testing for the low double exponent being zero is
unnecessary.  If the difference in exponents is less than 53 then the
high double exponent must be nearing the low end of its range, and the
low double exponent hit rock bottom.

- ldbl2mpn.c: A denormal (ie. exponent of zero) value is treated as
if the exponent was one, so shift mantissa left by one.  Code handling
normalisation of the low double mantissa lacked a test for shift count
greater than bits in type being shifted, and lacked anything to handle
the case where the difference in exponents is less than 53 as in
printf_fphex.c.

- math_ldbl.h (ldbl_extract_mantissa): Same as above, but worse, with
code testing for exponent > 1 for some reason, probably a typo for >= 1.

- math_ldbl.h (ldbl_insert_mantissa): Round the high double as per
mpn2ldbl.c (hi is odd or explicit mantissas non-zero) so that the
number we return won't change when applying ldbl_canonicalize().
Add missing overflow checks and normalisation of high mantissa.
Correct misleading comment: "The hidden bit of the lo mantissa is
zero" is not always true as can be seen from the code rounding the hi
mantissa.  Also by inspection, lzcount can never be less than zero so
remove that test.  Lastly, masking bitfields to their widths can be
left to the compiler.

- mpn2ldbl.c: The overflow checks here on rounding of high double were
just plain wrong.  Incrementing the exponent must be accompanied by a
shift right of the mantissa to keep the value unchanged.  Above notes
for ldbl_insert_mantissa are also relevant.

	[BZ #15680]
	* sysdeps/ieee754/ldbl-128ibm/e_rem_pio2l.c: Comment fix.
	* sysdeps/ieee754/ldbl-128ibm/printf_fphex.c
	(PRINT_FPHEX_LONG_DOUBLE): Tidy code by moving -53 into ediff
	calculation.  Remove unnecessary test for denormal exponent.
	* sysdeps/ieee754/ldbl-128ibm/ldbl2mpn.c (__mpn_extract_long_double):
	Correct handling of denormals.  Avoid undefined shift behaviour.
	Correct normalisation of low mantissa when low double is denormal.
	* sysdeps/ieee754/ldbl-128ibm/math_ldbl.h
	(ldbl_extract_mantissa): Likewise.  Comment.  Use uint64_t* for hi64.
	(ldbl_insert_mantissa): Make both hi64 and lo64 parms uint64_t.
	Correct normalisation of low mantissa.  Test for overflow of high
	mantissa and normalise.
	(ldbl_nearbyint): Use more readable constant for two52.
	* sysdeps/ieee754/ldbl-128ibm/mpn2ldbl.c
	(__mpn_construct_long_double): Fix test for overflow of high
	mantissa and correct normalisation.  Avoid undefined shift.
2013-10-04 10:30:56 +09:30

149 lines
4.3 KiB
C

/* Copyright (C) 1995-2013 Free Software Foundation, Inc.
This file is part of the GNU C Library.
The GNU C Library is free software; you can redistribute it and/or
modify it under the terms of the GNU Lesser General Public
License as published by the Free Software Foundation; either
version 2.1 of the License, or (at your option) any later version.
The GNU C Library is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
Lesser General Public License for more details.
You should have received a copy of the GNU Lesser General Public
License along with the GNU C Library; if not, see
<http://www.gnu.org/licenses/>. */
#include "gmp.h"
#include "gmp-impl.h"
#include <ieee754.h>
#include <float.h>
#include <math.h>
/* Convert a multi-precision integer of the needed number of bits (106
for long double) and an integral power of two to a `long double' in
IBM extended format. */
long double
__mpn_construct_long_double (mp_srcptr frac_ptr, int expt, int sign)
{
union ibm_extended_long_double u;
unsigned long lzcount;
unsigned long long hi, lo;
int exponent2;
u.d[0].ieee.negative = sign;
u.d[1].ieee.negative = sign;
u.d[0].ieee.exponent = expt + IEEE754_DOUBLE_BIAS;
u.d[1].ieee.exponent = 0;
exponent2 = expt - 53 + IEEE754_DOUBLE_BIAS;
#if BITS_PER_MP_LIMB == 32
/* The low order 53 bits (52 + hidden) go into the lower double */
lo = frac_ptr[0];
lo |= (frac_ptr[1] & ((1LL << (53 - 32)) - 1)) << 32;
/* The high order 53 bits (52 + hidden) go into the upper double */
hi = (frac_ptr[1] >> (53 - 32)) & ((1 << 11) - 1);
hi |= ((unsigned long long) frac_ptr[2]) << 11;
hi |= ((unsigned long long) frac_ptr[3]) << (32 + 11);
#elif BITS_PER_MP_LIMB == 64
/* The low order 53 bits (52 + hidden) go into the lower double */
lo = frac_ptr[0] & (((mp_limb_t) 1 << 53) - 1);
/* The high order 53 bits (52 + hidden) go into the upper double */
hi = (frac_ptr[0] >> 53) & (((mp_limb_t) 1 << 11) - 1);
hi |= (frac_ptr[1] << 11);
#else
#error "mp_limb size " BITS_PER_MP_LIMB "not accounted for"
#endif
if ((hi & (1LL << 52)) == 0 && (hi | lo) != 0)
{
/* denormal number */
unsigned long long val = hi ? hi : lo;
if (sizeof (val) == sizeof (long))
lzcount = __builtin_clzl (val);
else if ((val >> 32) != 0)
lzcount = __builtin_clzl ((long) (val >> 32));
else
lzcount = __builtin_clzl ((long) val) + 32;
if (hi)
lzcount = lzcount - (64 - 53);
else
lzcount = lzcount + 53 - (64 - 53);
if (lzcount > u.d[0].ieee.exponent)
{
lzcount = u.d[0].ieee.exponent;
u.d[0].ieee.exponent = 0;
exponent2 -= lzcount;
}
else
{
u.d[0].ieee.exponent -= (lzcount - 1);
exponent2 -= (lzcount - 1);
}
if (lzcount <= 53)
{
hi = (hi << lzcount) | (lo >> (53 - lzcount));
lo = (lo << lzcount) & ((1LL << 53) - 1);
}
else
{
hi = lo << (lzcount - 53);
lo = 0;
}
}
if (lo != 0)
{
/* hidden bit of low double controls rounding of the high double.
If hidden is '1' and either the explicit mantissa is non-zero
or hi is odd, then round up hi and adjust lo (2nd mantissa)
plus change the sign of the low double to compensate. */
if ((lo & (1LL << 52)) != 0
&& ((hi & 1) != 0 || (lo & ((1LL << 52) - 1)) != 0))
{
hi++;
if ((hi & (1LL << 53)) != 0)
{
hi >>= 1;
u.d[0].ieee.exponent++;
}
u.d[1].ieee.negative = !sign;
lo = (1LL << 53) - lo;
}
/* Normalize the low double. Shift the mantissa left until
the hidden bit is '1' and adjust the exponent accordingly. */
if (sizeof (lo) == sizeof (long))
lzcount = __builtin_clzl (lo);
else if ((lo >> 32) != 0)
lzcount = __builtin_clzl ((long) (lo >> 32));
else
lzcount = __builtin_clzl ((long) lo) + 32;
lzcount = lzcount - (64 - 53);
lo <<= lzcount;
exponent2 -= lzcount;
if (exponent2 > 0)
u.d[1].ieee.exponent = exponent2;
else if (exponent2 > -53)
lo >>= 1 - exponent2;
else
lo = 0;
}
else
u.d[1].ieee.negative = 0;
u.d[1].ieee.mantissa1 = lo;
u.d[1].ieee.mantissa0 = lo >> 32;
u.d[0].ieee.mantissa1 = hi;
u.d[0].ieee.mantissa0 = hi >> 32;
return u.ld;
}