[Gmpcommit] /var/hg/gmp: 3 new changesets
mercurial at gmplib.org
mercurial at gmplib.org
Thu Dec 27 14:31:57 CET 2012
details: /var/hg/gmp/rev/b4d7d6d21d5b
changeset: 15216:b4d7d6d21d5b
user: Torbjorn Granlund <tege at gmplib.org>
date: Thu Dec 27 13:59:58 2012 +0100
description:
Remove an unused definition.
details: /var/hg/gmp/rev/cbaec863b98b
changeset: 15217:cbaec863b98b
user: Torbjorn Granlund <tege at gmplib.org>
date: Thu Dec 27 14:00:40 2012 +0100
description:
Suppress check_random for vax.
details: /var/hg/gmp/rev/f116a136e832
changeset: 15218:f116a136e832
user: Torbjorn Granlund <tege at gmplib.org>
date: Thu Dec 27 14:10:16 2012 +0100
description:
Complete rewrite of nonIEEE code.
diffstat:
extractdbl.c  5 +
mpn/generic/get_d.c  429 +++++++++++++++++++++++
tests/mpq/tget_d.c  2 +
3 files changed, 201 insertions(+), 235 deletions()
diffs (truncated from 569 to 300 lines):
diff r 7624a57c9d49 r f116a136e832 extractdbl.c
 a/extractdbl.c Tue Dec 25 13:45:41 2012 +0100
+++ b/extractdbl.c Thu Dec 27 14:10:16 2012 +0100
@@ 1,6 +1,7 @@
/* __gmp_extract_double  convert from double to array of mp_limb_t.
Copyright 1996, 1999, 2000, 2001, 2002, 2006 Free Software Foundation, Inc.
+Copyright 1996, 1999, 2000, 2001, 2002, 2006, 2012 Free Software Foundation,
+Inc.
This file is part of the GNU MP Library.
@@ 28,8 +29,6 @@
#define _GMP_IEEE_FLOATS 0
#endif
#define BITS_IN_MANTISSA 53

/* Extract a nonnegative double in d. */
int
diff r 7624a57c9d49 r f116a136e832 mpn/generic/get_d.c
 a/mpn/generic/get_d.c Tue Dec 25 13:45:41 2012 +0100
+++ b/mpn/generic/get_d.c Thu Dec 27 14:10:16 2012 +0100
@@ 29,33 +29,20 @@
#define _GMP_IEEE_FLOATS 0
#endif
#if ! _GMP_IEEE_FLOATS
/* dummy definition, just to let dead code compile */
union ieee_double_extract {
 struct {
 int manh, manl, sig, exp;
 } s;
 double d;
};
#endif

/* To force use of the generic C code for testing, put
"#define _GMP_IEEE_FLOATS 0" at this point. */

/* In alpha gcc prior to 3.4, signed DI comparisons involving constants are
rearranged from "x < n" to "x+(n) < 0", which is of course hopelessly
wrong if that addition overflows.
 The workaround here avoids this bug by ensuring n is not a literal
 constant. Note that this is alpha specific. The offending transformation
 is/was in alpha.c alpha_emit_conditional_branch() under "We want to use
 cmpcc/bcc".
+ The workaround here avoids this bug by ensuring n is not a literal constant.
+ Note that this is alpha specific. The offending transformation is/was in
+ alpha.c alpha_emit_conditional_branch() under "We want to use cmpcc/bcc".
 Bizarrely, it turns out this happens also with Cray cc on
 alphaev5crayunicosmk2.0.6.X, and has the same solution. Don't know why
 or how. */
+ Bizarrely, this happens also with Cray cc on alphaev5crayunicosmk2.0.6.X,
+ and has the same solution. Don't know why or how. */
#if HAVE_HOST_CPU_FAMILY_alpha \
&& ((defined (__GNUC__) && ! __GMP_GNUC_PREREQ(3,4)) \
@@ 70,55 +57,57 @@
#endif

/* Return the value {ptr,size}*2^exp, and negative if sign<0.
 Must have size>=1, and a nonzero high limb ptr[size1].
+/* Return the value {ptr,size}*2^exp, and negative if sign<0. Must have
+ size>=1, and a nonzero high limb ptr[size1].
{ptr,size} is truncated towards zero. This is consistent with other gmp
 conversions, like mpz_set_f or mpz_set_q, and is easy to implement and
 test.
+ conversions, like mpz_set_f or mpz_set_q, and is easy to implement and test.
In the past conversions had attempted (imperfectly) to let the hardware
float rounding mode take effect, but that gets tricky since multiple
roundings need to be avoided, or taken into account, and denorms mean the
effective precision of the mantissa is not constant. (For reference,
 mpz_get_d on IEEE systems was ok, except it operated on the absolute
 value. mpf_get_d and mpq_get_d suffered from multiple roundings and from
 not always using enough bits to get the rounding right.)
+ mpz_get_d on IEEE systems was ok, except it operated on the absolute value.
+ mpf_get_d and mpq_get_d suffered from multiple roundings and from not always
+ using enough bits to get the rounding right.)
It's felt that GMP is not primarily concerned with hardware floats, and
really isn't enhanced by getting involved with hardware rounding modes
 (which could even be some weird unknown style), so something unambiguous
 and straightforward is best.
+ (which could even be some weird unknown style), so something unambiguous and
+ straightforward is best.
The IEEE code below is the usual case, it knows either a 32bit or 64bit
limb and is done with shifts and masks. The 64bit case in particular
should come out nice and compact.
 The generic code works one bit at a time, which will be quite slow, but
 should support any binarybased "double" and be safe against any rounding
 mode. Note in particular it works on IEEE systems too.
+ The generic code used to work one bit at a time, which was ont only slow,
+ but implicitly relied upon denoms for intermediates, since the lowest bits'
+ weight of a perfectly valid fp number underflows in nondenorm. Therefore,
+ the generic code now works limbperlimb, initially creating a number x such
+ that 1 <= x <= BASE. (BASE is reached only as result of rounding.) Then
+ x's exponent is scaled with explicit code (not ldexp to avoid libm
+ dependency). It is a tapdance to avoid underflow or overflow, beware!
Traps:
 Hardware traps for overflow to infinity, underflow to zero, or
 unsupported denorms may or may not be taken. The IEEE code works bitwise
 and so probably won't trigger them, the generic code works by float
 operations and so probably will. This difference might be thought less
 than ideal, but again its felt straightforward code is better than trying
 to get intimate with hardware exceptions (of perhaps unknown nature).
+ Hardware traps for overflow to infinity, underflow to zero, or unsupported
+ denorms may or may not be taken. The IEEE code works bitwise and so
+ probably won't trigger them, the generic code works by float operations and
+ so probably will. This difference might be thought less than ideal, but
+ again its felt straightforward code is better than trying to get intimate
+ with hardware exceptions (of perhaps unknown nature).
Not done:
 mpz_get_d in the past handled size==1 with a cast limb>double. This
 might still be worthwhile there (for up to the mantissa many bits), but
 for mpn_get_d here, the cost of applying "exp" to the resulting exponent
 would probably use up any benefit a cast may have over bit twiddling.
 Also, if the exponent is pushed into denorm range then bit twiddling is
 the only option, to ensure the desired truncation is obtained.
+ mpz_get_d in the past handled size==1 with a cast limb>double. This might
+ still be worthwhile there (for up to the mantissa many bits), but for
+ mpn_get_d here, the cost of applying "exp" to the resulting exponent would
+ probably use up any benefit a cast may have over bit twiddling. Also, if
+ the exponent is pushed into denorm range then bit twiddling is the only
+ option, to ensure the desired truncation is obtained.
Other:
@@ 130,11 +119,11 @@
cast, neither in the IEEE or generic code. */
+
double
mpn_get_d (mp_srcptr up, mp_size_t size, mp_size_t sign, long exp)
{
int lshift, nbits;
 union ieee_double_extract u;
mp_limb_t x, mhi, mlo;
ASSERT (size >= 0);
@@ 150,8 +139,9 @@
if (UNLIKELY ((unsigned long) (GMP_NUMB_BITS * size)
> (unsigned long) (LONG_MAX  exp)))
{
 if (_GMP_IEEE_FLOATS)
 goto ieee_infinity;
+#if _GMP_IEEE_FLOATS
+ goto ieee_infinity;
+#endif
/* generic */
exp = LONG_MAX;
@@ 162,167 +152,75 @@
}
#if ! _GMP_IEEE_FLOATS
 {
 /* NonIEEE or strange limb size, do something generic. */
+ { /* NonIEEE or strange limb size, do something generic. */
+ mp_size_t i;
+ double d, weight;
+ unsigned long uexp;
 mp_size_t i;
 mp_limb_t limb, bit;
 int shift;
 double base, factor, prev_factor, d, new_d, diff;

 /* "limb" is "up[i]" the limb being examined, "bit" is a mask for the
 bit being examined, initially the highest nonzero bit. */
 i = size1;
 limb = up[i];
 count_leading_zeros (shift, limb);
 bit = GMP_LIMB_HIGHBIT >> shift;

 /* relative to just under high nonzero bit */
 exp = (shift  GMP_NAIL_BITS) + 1;

 /* Power up "factor" to 2^exp, being the value of the "bit" in "limb"
 being examined. */
 base = (exp >= 0 ? 2.0 : 0.5);
 exp = ABS (exp);
 factor = 1.0;
 for (;;)
+ /* First generate an fp number disregarding exp, instead keeping things
+ within the numb base factor from 1, which should prevent overflow and
+ underflow even for the most exponent limited fp formats. The
+ termination criteria should be refined, since we now include too many
+ limbs. */
+ weight = 1/MP_BASE_AS_DOUBLE;
+ d = up[size  1];
+ for (i = size  2; i >= 0; i)
{
 if (exp & 1)
 {
 prev_factor = factor;
 factor *= base;
 FORCE_DOUBLE (factor);
 if (factor == 0.0)
 return 0.0; /* underflow */
 if (factor == prev_factor)
 {
 d = factor; /* overflow, apparent infinity */
 goto generic_done;
 }
 }
 exp >>= 1;
 if (exp == 0)
+ d += up[i] * weight;
+ weight /= MP_BASE_AS_DOUBLE;
+ if (weight == 0)
break;
 base *= base;
}
 /* Add a "factor" for each nonzero bit, working from high to low.
 Stop if any rounding occurs, hence implementing a truncation.

 Note no attention is paid to DBL_MANT_DIG, since the effective
 number of bits in the mantissa isn't constant when in denorm range.
 We also encountered an ARM system with apparently somewhat doubtful
 software floats where DBL_MANT_DIG claimed 53 bits but only 32
 actually worked. */

 d = factor; /* high bit */
 for (;;)
+ /* Now apply exp. */
+ exp = GMP_NUMB_BITS;
+ if (exp > 0)
{
 factor *= 0.5; /* next bit */
 bit >>= 1;
 if (bit == 0)
 {
 /* next limb, if any */
 i;
 if (i < 0)
 break;
 limb = up[i];
 bit = GMP_NUMB_HIGHBIT;
 }

 if (bit & limb)
 {
 new_d = d + factor;
 FORCE_DOUBLE (new_d);
 diff = new_d  d;
 if (diff != factor)
 break; /* rounding occured, stop now */
 d = new_d;
 }
 }

 generic_done:
 return (sign >= 0 ? d : d);
 }
#endif

 up += size;

#if GMP_LIMB_BITS == 64
 mlo = up[1];
 count_leading_zeros (lshift, mlo);

 exp = (lshift  GMP_NAIL_BITS) + 1;
 mlo <<= lshift;

 nbits = GMP_LIMB_BITS  lshift;

 if (nbits < 53 && size > 1)
 {
 x = up[2];
 x <<= GMP_NAIL_BITS;
 x >>= nbits;
 mlo = x;
 nbits += GMP_NUMB_BITS;

 if (LIMBS_PER_DOUBLE >= 3 && nbits < 53 && size > 2)
 {
 x = up[3];
 x <<= GMP_NAIL_BITS;
More information about the gmpcommit
mailing list