File: [local] / OpenXM_contrib / gmp / doc / Attic / tasks.html (download) (as text)
Revision 1.1.1.2 (vendor branch), Mon Aug 25 16:06:11 2003 UTC (21 years, 1 month ago) by ohara
Branch: GMP
CVS Tags: VERSION_4_1_2, RELEASE_1_2_3, RELEASE_1_2_2_KNOPPIX_b, RELEASE_1_2_2_KNOPPIX Changes since 1.1.1.1: +872 -246
lines
Import gmp 4.1.2
|
<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 3.2//EN">
<html>
<head>
<title>
GMP Itemized Development Tasks
</title>
</head>
<body bgcolor=lightgreen>
<center>
<h1>
GMP Itemized Development Tasks
</h1>
</center>
<font size=-1>
Copyright 2000, 2001, 2002 Free Software Foundation, Inc. <br><br>
This file is part of the GNU MP Library. <br><br>
The GNU MP 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. <br><br>
The GNU MP 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. <br><br>
You should have received a copy of the GNU Lesser General Public License
along with the GNU MP Library; see the file COPYING.LIB. If not, write to
the Free Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
MA 02111-1307, USA.
</font>
<hr>
<!-- NB. timestamp updated automatically by emacs -->
<comment>
This file current as of 20 May 2002. An up-to-date version is available at
<a href="http://www.swox.com/gmp/tasks.html">http://www.swox.com/gmp/tasks.html</a>.
Please send comments about this page to
<a href="mailto:bug-gmp@gnu.org">bug-gmp@gnu.org</a>.
</comment>
<p> These are itemized GMP development tasks. Not all the tasks
listed here are suitable for volunteers, but many of them are.
Please see the <a href="projects.html">projects file</a> for more
sizeable projects.
<h4>Correctness and Completeness</h4>
<ul>
<li> The various reuse.c tests need to force reallocation by calling
<code>_mpz_realloc</code> with a small (1 limb) size.
<li> One reuse case is missing from mpX/tests/reuse.c:
<code>mpz_XXX(a,a,a)</code>.
<li> When printing <code>mpf_t</code> numbers with exponents >2^53 on
machines with 64-bit <code>mp_exp_t</code>, the precision of
<code>__mp_bases[base].chars_per_bit_exactly</code> is insufficient and
<code>mpf_get_str</code> aborts. Detect and compensate. Alternately,
think seriously about using some sort of fixed-point integer value.
Avoiding unnecessary floating point is probably a good thing in general,
and it might be faster on some CPUs.
<li> Make the string reading functions allow the `0x' prefix when the base is
explicitly 16. They currently only allow that prefix when the base is
unspecified (zero).
<li> <code>mpf_eq</code> is not always correct, when one operand is
1000000000... and the other operand is 0111111111..., i.e., extremely
close. There is a special case in <code>mpf_sub</code> for this
situation; put similar code in <code>mpf_eq</code>.
<li> <code>mpf_eq</code> doesn't implement what gmp.texi specifies. It should
not use just whole limbs, but partial limbs.
<li> <code>mpf_set_str</code> doesn't validate it's exponent, for instance
garbage 123.456eX789X is accepted (and an exponent 0 used), and overflow
of a <code>long</code> is not detected.
<li> <code>mpf_add</code> doesn't check for a carry from truncated portions of
the inputs, and in that respect doesn't implement the "infinite precision
followed by truncate" specified in the manual.
<li> <code>mpf_div</code> of x/x doesn't always give 1, reported by Peter
Moulder. Perhaps it suffices to put +1 on the effective divisor prec, so
that data bits rather than zeros are shifted in when normalizing. Would
prefer to switch to <code>mpn_tdiv_qr</code>, where all shifting should
disappear.
<li> Windows DLLs: tests/mpz/reuse.c and tests/mpf/reuse.c initialize global
variables with pointers to <code>mpz_add</code> etc, which doesn't work
when those routines are coming from a DLL (because they're effectively
function pointer global variables themselves). Need to rearrange perhaps
to a set of calls to a test function rather than iterating over an array.
<li> demos/pexpr.c: The local variables in <code>main</code> might be
clobbered by the <code>longjmp</code>.
</ul>
<h4>Machine Independent Optimization</h4>
<ul>
<li> <code>mpn_gcdext</code>, <code>mpz_get_d</code>,
<code>mpf_get_str</code>: Don't test <code>count_leading_zeros</code> for
zero, instead check the high bit of the operand and avoid invoking
<code>count_leading_zeros</code>. This is an optimization on all
machines, and significant on machines with slow
<code>count_leading_zeros</code>, though it's possible an already
normalized operand might not be encountered very often.
<li> Rewrite <code>umul_ppmm</code> to use floating-point for generating the
most significant limb (if <code>BITS_PER_MP_LIMB</code> <= 52 bits).
(Peter Montgomery has some ideas on this subject.)
<li> Improve the default <code>umul_ppmm</code> code in longlong.h: Add partial
products with fewer operations.
<li> Consider inlining <code>mpz_set_ui</code>. This would be both small and
fast, especially for compile-time constants, but would make application
binaries depend on having 1 limb allocated to an <code>mpz_t</code>,
preventing the "lazy" allocation scheme below.
<li> Consider inlining <code>mpz_[cft]div_ui</code> and maybe
<code>mpz_[cft]div_r_ui</code>. A <code>__gmp_divide_by_zero</code>
would be needed for the divide by zero test, unless that could be left to
<code>mpn_mod_1</code> (not sure currently whether all the risc chips
provoke the right exception there if using mul-by-inverse).
<li> Consider inlining: <code>mpz_fits_s*_p</code>. The setups for
<code>LONG_MAX</code> etc would need to go into gmp.h, and on Cray it
might, unfortunately, be necessary to forcibly include <limits.h>
since there's no apparent way to get <code>SHRT_MAX</code> with an
expression (since <code>short</code> and <code>unsigned short</code> can
be different sizes).
<li> <code>mpz_powm</code> and <code>mpz_powm_ui</code> aren't very
fast on one or two limb moduli, due to a lot of function call
overheads. These could perhaps be handled as special cases.
<li> <code>mpz_powm</code> and <code>mpz_powm_ui</code> want better
algorithm selection, and the latter should use REDC. Both could
change to use an <code>mpn_powm</code> and <code>mpn_redc</code>.
<li> <code>mpz_powm</code> REDC should do multiplications by <code>g[]</code>
using the division method when they're small, since the REDC form of a
small multiplier is normally a full size product. Probably would need a
new tuned parameter to say what size multiplier is "small", as a function
of the size of the modulus.
<li> <code>mpz_powm</code> REDC should handle even moduli if possible. Maybe
this would mean for m=n*2^k doing mod n using REDC and an auxiliary
calculation mod 2^k, then putting them together at the end.
<li> <code>mpn_gcd</code> might be able to be sped up on small to
moderate sizes by improving <code>find_a</code>, possibly just by
providing an alternate implementation for CPUs with slowish
<code>count_leading_zeros</code>.
<li> Toom3 <code>USE_MORE_MPN</code> could use a low to high cache localized
evaluate and interpolate. The necessary <code>mpn_divexact_by3c</code>
exists.
<li> <code>mpn_mul_basecase</code> on NxM with big N but small M could try for
better cache locality by taking N piece by piece. The current code could
be left available for CPUs without caching. Depending how karatsuba etc
is applied to unequal size operands it might be possible to assume M is
always smallish.
<li> <code>mpn_perfect_square_p</code> on small operands might be better off
skipping the residue tests and just taking a square root.
<li> <code>mpz_perfect_power_p</code> could be improved in a number of ways.
Test for Nth power residues modulo small primes like
<code>mpn_perfect_square_p</code> does. Use p-adic arithmetic to find
possible roots. Divisibility by other primes should be tested by
grouping into a limb like <code>PP</code>.
<li> <code>mpz_perfect_power_p</code> might like to use <code>mpn_gcd_1</code>
instead of a private GCD routine. The use it's put to isn't
time-critical, and it might help be ensure correctness to use the main GCD
routine.
<li> <code>mpz_perfect_power_p</code> could use
<code>mpz_divisible_ui_p</code> instead of <code>mpz_tdiv_ui</code> for
divisibility testing, the former is faster on a number of systems. (But
all that prime test stuff is going to be rewritten some time.)
<li> Change <code>PP</code>/<code>PP_INVERTED</code> into an array of such
pairs, listing several hundred primes. Perhaps actually make the
products larger than one limb each.
<li> <code>PP</code> can have factors of 2 introduced in order to get the high
bit set and therefore a <code>PP_INVERTED</code> existing. The factors
of 2 don't affect the way the remainder r = a % ((x*y*z)*2^n) is used,
further remainders r%x, r%y, etc, are the same since x, y, etc are odd.
The advantage of this is that <code>mpn_preinv_mod_1</code> can then be
used if it's faster than plain <code>mpn_mod_1</code>. This would be a
change only for 16-bit limbs, all the rest already have <code>PP</code>
in the right form.
<li> <code>PP</code> could have extra factors of 3 or 5 or whatever introduced
if they fit, and final remainders mod 9 or 25 or whatever used, thereby
making more efficient use of the <code>mpn_mod_1</code> done. On a
16-bit limb it looks like <code>PP</code> could take an extra factor of
3.
<li> <code>mpz_probab_prime_p</code>, <code>mpn_perfect_square_p</code> and
<code>mpz_perfect_power_p</code> could use <code>mpn_mod_34lsub1</code>
to take a remainder mod 2^24-1 or 2^48-1 and quickly get remainders mod
3, 5, 7, 13 and 17 (factors of 2^24-1). This could either replace the
<code>PP</code> division currently done, or allow <code>PP</code> to do
larger primes, depending how many residue tests seem worthwhile before
launching into full root extractions or Miller-Rabin etc.
<li> <code>mpz_probab_prime_p</code> (and maybe others) could code the
divisibility tests like <code>n%7 == 0</code> in the form
<pre>
#define MP_LIMB_DIVISIBLE_7_P(n) \
((n) * MODLIMB_INVERSE_7 <= MP_LIMB_T_MAX/7)
</pre>
This would help compilers which don't know how to optimize divisions by
constants, and would help current gcc (3.0) too since gcc forms a whole
remainder rather than using a modular inverse and comparing. This
technique works for any odd modulus, and with some tweaks for even moduli
too. See Granlund and Montgomery "Division By Invariant Integers"
section 9.
<li> <code>mpz_probab_prime_p</code> and <code>mpz_nextprime</code> could
offer certainty for primes up to 2^32 by using a one limb miller-rabin
test to base 2, combined with a table of actual strong pseudoprimes in
that range (2314 of them). If that table is too big then both base 2 and
base 3 tests could be done, leaving a table of 104. The test could use
REDC and therefore be a <code>modlimb_invert</code> a remainder (maybe)
then two multiplies per bit (successively dependent). Processors with
pipelined multipliers could do base 2 and 3 in parallel. Vector systems
could do a whole bunch of bases in parallel, and perhaps offer near
certainty up to 64-bits (certainty might depend on an exhaustive search
of pseudoprimes up to that limit). Obviously 2^32 is not a big number,
but an efficient and certain calculation is attractive. It might find
other uses internally, and could even be offered as a one limb prime test
<code>mpn_probab_prime_1_p</code> or <code>gmp_probab_prime_ui_p</code>
perhaps.
<li> <code>mpz_probab_prime_p</code> doesn't need to make a copy of
<code>n</code> when the input is negative, it can setup an
<code>mpz_t</code> alias, same data pointer but a positive size. With no
need to clear before returning, the recursive function call could be
dispensed with too.
<li> <code>mpf_set_str</code> produces low zero limbs when a string has a
fraction but is exactly representable, eg. 0.5 in decimal. These could be
stripped to save work in later operations.
<li> <code>mpz_and</code>, <code>mpz_ior</code> and <code>mpz_xor</code> should
use <code>mpn_and_n</code> etc for the benefit of the small number of
targets with native versions of those routines. Need to be careful not to
pass size==0. Is some code sharing possible between the <code>mpz</code>
routines?
<li> <code>mpf_add</code>: Don't do a copy to avoid overlapping operands
unless it's really necessary (currently only sizes are tested, not
whether r really is u or v).
<li> <code>mpf_add</code>: Under the check for v having no effect on the
result, perhaps test for r==u and do nothing in that case, rather than
currently it looks like an <code>MPN_COPY_INCR</code> will be done to
reduce prec+1 limbs to prec.
<li> <code>mpn_divrem_2</code> could usefully accept unnormalized divisors and
shift the dividend on-the-fly, since this should cost nothing on
superscalar processors and avoid the need for temporary copying in
<code>mpn_tdiv_qr</code>.
<li> <code>mpf_sqrt_ui</code> calculates prec+1 limbs, whereas just prec would
satisfy the application requested precision. It should suffice to simply
reduce the rsize temporary to 2*prec-1 limbs. <code>mpf_sqrt</code>
might be similar.
<li> <code>invert_limb</code> generic C: The division could use dividend
b*(b-d)-1 which is high:low of (b-1-d):(b-1), instead of the current
(b-d):0, where b=2^<code>BITS_PER_MP_LIMB</code> and d=divisor. The
former is per the original paper and is used in the x86 code, the
advantage is that the current special case for 0x80..00 could be dropped.
The two should be equivalent, but a little check of that would be wanted.
<li> <code>mpq_cmp_ui</code> could form the <code>num1*den2</code> and
<code>num2*den1</code> products limb-by-limb from high to low and look at
each step for values differing by more than the possible carry bit from
the uncalculated portion.
<li> <code>mpq_cmp</code> could do the same high-to-low progressive multiply
and compare. The benefits of karatsuba and higher multiplication
algorithms are lost, but if it's assumed only a few high limbs will be
needed to determine an order then that's fine.
<li> <code>mpn_add_1</code>, <code>mpn_sub_1</code>, <code>mpn_add</code>,
<code>mpn_sub</code>: Internally use <code>__GMPN_ADD_1</code> etc
instead of the functions, so they get inlined on all compilers, not just
gcc and others with <code>inline</code> recognised in gmp.h.
<code>__GMPN_ADD_1</code> etc are meant mostly to support application
inline <code>mpn_add_1</code> etc and if they don't come out good for
internal uses then special forms can be introduced, for instance many
internal uses are in-place. Sometimes a block of code is executed based
on the carry-out, rather than using it arithmetically, and those places
might want to do their own loops entirely.
<li> <code>__gmp_extract_double</code> on 64-bit systems could use just one
bitfield for the mantissa extraction, not two, when endianness permits.
Might depend on the compiler allowing <code>long long</code> bit fields
when that's the only actual 64-bit type.
<li> <code>mpf_get_d</code> could be more like <code>mpz_get_d</code> and do
more in integers and give the float conversion as such a chance to round
in its preferred direction. Some code sharing ought to be possible. Or
if nothing else then for consistency the two ought to give identical
results on integer operands (not clear if this is so right now).
<li> <code>usqr_ppm</code> or some such could do a widening square in the
style of <code>umul_ppmm</code>. This would help 68000, and be a small
improvement for the generic C (which is used on UltraSPARC/64 for
instance). GCC recognises the generic C ul*vh and vl*uh are identical,
but does two separate additions to the rest of the result.
<li> tal-notreent.c could keep a block of memory permanently allocated.
Currently the last nested <code>TMP_FREE</code> releases all memory, so
there's an allocate and free every time a top-level function using
<code>TMP</code> is called. Would need
<code>mp_set_memory_functions</code> to tell tal-notreent.c to release
any cached memory when changing allocation functions though.
<li> <code>__gmp_tmp_alloc</code> from tal-notreent.c could be partially
inlined. If the current chunk has enough room then a couple of pointers
can be updated. Only if more space is required then a call to some sort
of <code>__gmp_tmp_increase</code> would be needed. The requirement that
<code>TMP_ALLOC</code> is an expression might make the implementation a
bit ugly and/or a bit sub-optimal.
<pre>
#define TMP_ALLOC(n)
((ROUND_UP(n) > current->end - current->point ?
__gmp_tmp_increase (ROUND_UP (n)) : 0),
current->point += ROUND_UP (n),
current->point - ROUND_UP (n))
</pre>
<li> <code>__mp_bases</code> has a lot of data for bases which are pretty much
never used. Perhaps the table should just go up to base 16, and have
code to generate data above that, if and when required. Naturally this
assumes the code would be smaller than the data saved.
<li> <code>__mp_bases</code> field <code>big_base_inverted</code> is only used
if <code>USE_PREINV_DIVREM_1</code> is true, and could be omitted
otherwise, to save space.
<li> Make <code>mpf_get_str</code> and <code>mpf_set_str</code> call the
corresponding, much faster, mpn functions.
<li> <code>mpn_mod_1</code> could pre-calculate values of R mod N, R^2 mod N,
R^3 mod N, etc, with R=2^<code>BITS_PER_MP_LIMB</code>, and use them to
process multiple limbs at each step by multiplying. Suggested by Peter
L. Montgomery.
<li> <code>mpz_get_str</code>, <code>mtox</code>: For power-of-2 bases, which
are of course fast, it seems a little silly to make a second pass over
the <code>mpn_get_str</code> output to convert to ASCII. Perhaps combine
that with the bit extractions.
<li> <code>mpz_gcdext</code>: If the caller requests only the S cofactor (of
A), and A<B, then the code ends up generating the cofactor T (of B) and
deriving S from that. Perhaps it'd be possible to arrange to get S in
the first place by calling <code>mpn_gcdext</code> with A+B,B. This
might only be an advantage if A and B are about the same size.
<li> <code>mpn_toom3_mul_n</code>, <code>mpn_toom3_sqr_n</code>: Temporaries
<code>B</code> and <code>D</code> are adjacent in memory and at the final
coefficient additions look like they could use a single
<code>mpn_add_n</code> of <code>l4</code> limbs rather than two of
<code>l2</code> limbs.
</ul>
<h4>Machine Dependent Optimization</h4>
<ul>
<li> <code>udiv_qrnnd_preinv2norm</code>, the branch-free version of
<code>udiv_qrnnd_preinv</code>, might be faster on various pipelined
chips. In particular the first <code>if (_xh != 0)</code> in
<code>udiv_qrnnd_preinv</code> might be roughly a 50/50 chance and might
branch predict poorly. (The second test is probably almost always
false.) Measuring with the tuneup program would be possible, but perhaps
a bit messy. In any case maybe the default should be the branch-free
version.
<br>
Note that the current <code>udiv_qrnnd_preinv2norm</code> implementation
assumes a right shift will sign extend, which is not guaranteed by the C
standards, and doesn't happen on Cray vector systems.
<li> Run the `tune' utility for more compiler/CPU combinations. We would like
to have gmp-mparam.h files in practically every implementation specific
mpn subdirectory, and repeat each *_THRESHOLD for gcc and the system
compiler. See the `tune' top-level directory for more information.
<pre>
#ifdef (__GNUC__)
#if __GNUC__ == 2 && __GNUC_MINOR__ == 7
...
#endif
#if __GNUC__ == 2 && __GNUC_MINOR__ == 8
...
#endif
#ifndef MUL_KARATSUBA_THRESHOLD
/* Default GNUC values */
...
#endif
#else /* system compiler */
...
#endif </pre>
<li> <code>invert_limb</code> on various processors might benefit from the
little Newton iteration done for alpha and ia64.
<li> Alpha 21264: Improve feed-in code for <code>mpn_mul_1</code>,
<code>mpn_addmul_1</code>, and <code>mpn_submul_1</code>.
<li> Alpha 21164: Rewrite <code>mpn_mul_1</code>, <code>mpn_addmul_1</code>,
and <code>mpn_submul_1</code> for the 21164. This should use both integer
multiplies and floating-point multiplies. For the floating-point
operations, the single-limb multiplier should be split into three 21-bit
chunks, or perhaps even better in four 16-bit chunks. Probably possible
to reach 9 cycles/limb.
<li> Alpha 21264 ev67: Use <code>ctlz</code> and <code>cttz</code> for
<code>count_leading_zeros</code> and<code>count_trailing_zeros</code>.
Use inline for gcc, probably want asm files for elsewhere.
<li> ARC: gcc longlong.h sets up <code>umul_ppmm</code> to call
<code>__umulsidi3</code> in libgcc. Could be copied straight across, but
perhaps ought to be tested.
<li> ARM: On v5 cpus see if the <code>clz</code> instruction can be used for
<code>count_leading_zeros</code>.
<li> Itanium: <code>mpn_divexact_by3</code> isn't particularly important, but
the generic C runs at about 27 c/l, whereas with the multiplies off the
dependent chain about 3 c/l ought to be possible.
<li> Itanium: <code>mpn_hamdist</code> could be put together based on the
current <code>mpn_popcount</code>.
<li> Itanium: <code>popc_limb</code> in gmp-impl.h could use the
<code>popcnt</code> insn.
<li> Itanium: <code>mpn_submul_1</code> is not implemented directly, only via
a combination of <code>mpn_mul_1</code> and <code>mpn_sub_n</code>.
<li> UltraSPARC/64: Optimize <code>mpn_mul_1</code>, <code>mpn_addmul_1</code>,
for s2 < 2^32 (or perhaps for any zero 16-bit s2 chunk). Not sure how
much this can improve the speed, though, since the symmetry that we rely
on is lost. Perhaps we can just gain cycles when s2 < 2^16, or more
accurately, when two 16-bit s2 chunks which are 16 bits apart are zero.
<li> UltraSPARC/64: Write native <code>mpn_submul_1</code>, analogous to
<code>mpn_addmul_1</code>.
<li> UltraSPARC/64: Write <code>umul_ppmm</code>. Using four
"<code>mulx</code>"s either with an asm block or via the generic C code is
about 90 cycles. Try using fp operations, and also try using karatsuba
for just three "<code>mulx</code>"s.
<li> UltraSPARC/64: <code>mpn_divrem_1</code>, <code>mpn_mod_1</code>,
<code>mpn_divexact_1</code> and <code>mpn_modexact_1_odd</code> could
process 32 bits at a time when the divisor fits 32-bits. This will need
only 4 <code>mulx</code>'s per limb instead of 8 in the general case.
<li> UltraSPARC/32: Rewrite <code>mpn_lshift</code>, <code>mpn_rshift</code>.
Will give 2 cycles/limb. Trivial modifications of mpn/sparc64 should do.
<li> UltraSPARC/32: Write special mpn_Xmul_1 loops for s2 < 2^16.
<li> UltraSPARC/32: Use <code>mulx</code> for <code>umul_ppmm</code> if
possible (see commented out code in longlong.h). This is unlikely to
save more than a couple of cycles, so perhaps isn't worth bothering with.
<li> UltraSPARC/32: On Solaris gcc doesn't give us <code>__sparc_v9__</code>
or anything to indicate V9 support when -mcpu=v9 is selected. See
gcc/config/sol2-sld-64.h. Will need to pass something through from
./configure to select the right code in longlong.h. (Currently nothing
is lost because <code>mulx</code> for multiplying is commented out.)
<li> UltraSPARC: <code>modlimb_invert</code> might save a few cycles from
masking down to just the useful bits at each point in the calculation,
since <code>mulx</code> speed depends on the highest bit set. Either
explicit masks or small types like <code>short</code> and
<code>int</code> ought to work.
<li> Sparc64 HAL R1: <code>mpn_popcount</code> and <code>mpn_hamdist</code>
could use <code>popc</code> currently commented out in gmp-impl.h. This
chip reputedly implements <code>popc</code> properly (see gcc sparc.md),
would need to recognise the chip as <code>sparchalr1</code> or something
in configure / config.sub / config.guess.
<li> PA64: Improve <code>mpn_addmul_1</code>, <code>mpn_submul_1</code>, and
<code>mpn_mul_1</code>. The current code runs at 11 cycles/limb. It
should be possible to saturate the cache, which will happen at 8
cycles/limb (7.5 for mpn_mul_1). Write special loops for s2 < 2^32;
it should be possible to make them run at about 5 cycles/limb.
<li> PPC630: Rewrite <code>mpn_addmul_1</code>, <code>mpn_submul_1</code>, and
<code>mpn_mul_1</code>. Use both integer and floating-point operations,
possibly two floating-point and one integer limb per loop. Split operands
into four 16-bit chunks for fast fp operations. Should easily reach 9
cycles/limb (using one int + one fp), but perhaps even 7 cycles/limb
(using one int + two fp).
<li> PPC630: <code>mpn_rshift</code> could do the same sort of unrolled loop
as <code>mpn_lshift</code>. Some judicious use of m4 might let the two
share source code, or with a register to control the loop direction
perhaps even share object code.
<li> PowerPC-32: <code>mpn_rshift</code> should do the same sort of unrolled
loop as <code>mpn_lshift</code>.
<li> Implement <code>mpn_mul_basecase</code> and <code>mpn_sqr_basecase</code>
for important machines. Helping the generic sqr_basecase.c with an
<code>mpn_sqr_diagonal</code> might be enough for some of the RISCs.
<li> POWER2/POWER2SC: Schedule <code>mpn_lshift</code>/<code>mpn_rshift</code>.
Will bring time from 1.75 to 1.25 cycles/limb.
<li> X86: Optimize non-MMX <code>mpn_lshift</code> for shifts by 1. (See
Pentium code.)
<li> X86: Good authority has it that in the past an inline <code>rep
movs</code> would upset GCC register allocation for the whole function.
Is this still true in GCC 3? It uses <code>rep movs</code> itself for
<code>__builtin_memcpy</code>. Examine the code for some simple and
complex functions to find out. Inlining <code>rep movs</code> would be
desirable, it'd be both smaller and faster.
<li> Pentium P54: <code>mpn_lshift</code> and <code>mpn_rshift</code> can come
down from 6.0 c/l to 5.5 or 5.375 by paying attention to pairing after
<code>shrdl</code> and <code>shldl</code>, see mpn/x86/pentium/README.
<li> Pentium P55 MMX: <code>mpn_lshift</code> and <code>mpn_rshift</code>
might benefit from some destination prefetching.
<li> PentiumPro: <code>mpn_divrem_1</code> might be able to use a
mul-by-inverse, hoping for maybe 30 c/l.
<li> P6: <code>mpn_add_n</code> and <code>mpn_sub_n</code> should be able to go
faster than the generic x86 code at 3.5 c/l. The athlon code for instance
runs at about 2.7.
<li> K7: <code>mpn_lshift</code> and <code>mpn_rshift</code> might be able to
do something branch-free for unaligned startups, and shaving one insn
from the loop with alternative indexing might save a cycle.
<li> PPC32: Try using fewer registers in the current <code>mpn_lshift</code>.
The pipeline is now extremely deep, perhaps unnecessarily deep.
<li> PPC32: Write <code>mpn_rshift</code> based on new <code>mpn_lshift</code>.
<li> PPC32: Rewrite <code>mpn_add_n</code> and <code>mpn_sub_n</code>. Should
run at just 3.25 cycles/limb.
<li> Fujitsu VPP: Vectorize main functions, perhaps in assembly language.
<li> Fujitsu VPP: Write <code>mpn_mul_basecase</code> and
<code>mpn_sqr_basecase</code>. This should use a "vertical multiplication
method", to avoid carry propagation. splitting one of the operands in
11-bit chunks.
<li> 68k, Pentium: <code>mpn_lshift</code> by 31 should use the special rshift
by 1 code, and vice versa <code>mpn_rshift</code> by 31 should use the
special lshift by 1. This would be best as a jump across to the other
routine, could let both live in lshift.asm and omit rshift.asm on finding
<code>mpn_rshift</code> already provided.
<li> Cray T3E: Experiment with optimization options. In particular,
-hpipeline3 seems promising. We should at least up -O to -O2 or -O3.
<li> Cray: <code>mpn_com_n</code> and <code>mpn_and_n</code> etc very probably
wants a pragma like <code>MPN_COPY_INCR</code>.
<li> Cray vector systems: <code>mpn_lshift</code>, <code>mpn_rshift</code>,
<code>mpn_popcount</code> and <code>mpn_hamdist</code> are nice and small
and could be inlined to avoid function calls.
<li> Cray: Variable length arrays seem to be faster than the tal-notreent.c
scheme. Not sure why, maybe they merely give the compiler more
information about aliasing (or the lack thereof). Would like to modify
<code>TMP_ALLOC</code> to use them, or introduce a new scheme. Memory
blocks wanted unconditionally are easy enough, those wanted only
sometimes are a problem. Perhaps a special size calculation to ask for a
dummy length 1 when unwanted, or perhaps an inlined subroutine
duplicating code under each conditional. Don't really want to turn
everything into a dog's dinner just because Cray don't offer an
<code>alloca</code>.
<li> Cray: <code>mpn_get_str</code> on power-of-2 bases ought to vectorize.
Does it? <code>bits_per_digit</code> and the inner loop over bits in a
limb might prevent it. Perhaps special cases for binary, octal and hex
would be worthwhile (very possibly for all processors too).
<li> Cray: <code>popc_limb</code> could use the Cray <code>_popc</code>
intrinsic. That would help <code>mpz_hamdist</code> and might make the
generic C versions of <code>mpn_popcount</code> and
<code>mpn_hamdist</code> suffice for Cray (if it vectorizes, or can be
given a hint to do so).
<li> 68000: <code>mpn_mul_1</code>, <code>mpn_addmul_1</code>,
<code>mpn_submul_1</code>: Check for a 16-bit multiplier and use two
multiplies per limb, not four.
<li> 68000: <code>mpn_lshift</code> and <code>mpn_rshift</code> could use a
<code>roll</code> and mask instead of <code>lsrl</code> and
<code>lsll</code>. This promises to be a speedup, effectively trading a
6+2*n shift for one or two 4 cycle masks. Suggested by Jean-Charles
Meyrignac.
<li> Improve <code>count_leading_zeros</code> for 64-bit machines:
<pre>
if ((x >> 32) == 0) { x <<= 32; cnt += 32; }
if ((x >> 48) == 0) { x <<= 16; cnt += 16; }
... </pre>
<li> IRIX 6 MIPSpro compiler has an <code>__inline</code> which could perhaps
be used in <code>__GMP_EXTERN_INLINE</code>. What would be the right way
to identify suitable versions of that compiler?
<li> VAX D and G format <code>double</code> floats are straightforward and
could perhaps be handled directly in <code>__gmp_extract_double</code>
and maybe in <code>mpz_get_d</code>, rather than falling back on the
generic code. (Both formats are detected by <code>configure</code>.)
<li> <code>mpn_get_str</code> final divisions by the base with
<code>udiv_qrnd_unnorm</code> could use some sort of multiply-by-inverse
on suitable machines. This ends up happening for decimal by presenting
the compiler with a run-time constant, but the same for other bases would
be good. Perhaps use could be made of the fact base<256.
<li> <code>mpn_umul_ppmm</code>, <code>mpn_udiv_qrnnd</code>: Return a
structure like <code>div_t</code> to avoid going through memory, in
particular helping RISCs that don't do store-to-load forwarding. Clearly
this is only possible if the ABI returns a structure of two
<code>mp_limb_t</code>s in registers.
</ul>
<h4>New Functionality</h4>
<ul>
<li> Add in-memory versions of <code>mp?_out_raw</code> and
<code>mp?_inp_raw</code>.
<li> <code>mpz_get_nth_ui</code>. Return the nth word (not necessarily the
nth limb).
<li> Maybe add <code>mpz_crr</code> (Chinese Remainder Reconstruction).
<li> Let `0b' and `0B' mean binary input everywhere.
<li> <code>mpz_init</code> and <code>mpq_init</code> could do lazy allocation.
Set <code>ALLOC(var)</code> to 0 to indicate nothing allocated, and let
<code>_mpz_realloc</code> do the initial alloc. Set
<code>z->_mp_d</code> to a dummy that <code>mpz_get_ui</code> and
similar can unconditionally fetch from. Niels Möller has had a go at
this.
<br>
The advantages of the lazy scheme would be:
<ul>
<li> Initial allocate would be the size required for the first value
stored, rather than getting 1 limb in <code>mpz_init</code> and then
more or less immediately reallocating.
<li> <code>mpz_init</code> would only store magic values in the
<code>mpz_t</code> fields, and could be inlined.
<li> A fixed initializer could even be used by applications, like
<code>mpz_t z = MPZ_INITIALIZER;</code>, which might be convenient
for globals.
</ul>
The advantages of the current scheme are:
<ul>
<li> <code>mpz_set_ui</code> and other similar routines needn't check the
size allocated and can just store unconditionally.
<li> <code>mpz_set_ui</code> and perhaps others like
<code>mpz_tdiv_r_ui</code> and a prospective
<code>mpz_set_ull</code> could be inlined.
</ul>
<li> Add <code>mpf_out_raw</code> and <code>mpf_inp_raw</code>. Make sure
format is portable between 32-bit and 64-bit machines, and between
little-endian and big-endian machines.
<li> <code>mpn_and_n</code> ... <code>mpn_copyd</code>: Perhaps make the mpn
logops and copys available in gmp.h, either as library functions or
inlines, with the availability of library functions instantiated in the
generated gmp.h at build time.
<li> <code>mpz_set_str</code> etc variants taking string lengths rather than
null-terminators.
<li> Consider changing the thresholds to apply the simpler algorithm when
"<code><=</code>" rather than "<code><</code>", so a threshold can
be set to <code>MP_SIZE_T_MAX</code> to get only the simpler code (the
compiler will know <code>size <= MP_SIZE_T_MAX</code> is always true).
Alternately it looks like the <code>ABOVE_THRESHOLD</code> and
<code>BELOW_THRESHOLD</code> macros can do this adequately, and also pick
up cases where a threshold of zero should mean only the second algorithm.
<li> <code>mpz_nthprime</code>.
<li> Perhaps <code>mpz_init2</code>, initializing and making initial room for
N bits. The actual size would be rounded up to a limb, and perhaps an
extra limb added since so many <code>mpz</code> routines need that on
their destination.
<li> <code>mpz_andn</code>, <code>mpz_iorn</code>, <code>mpz_nand</code>,
<code>mpz_nior</code>, <code>mpz_xnor</code> might be useful additions,
if they could share code with the current such functions (which should be
possible).
<li> <code>mpz_and_ui</code> etc might be of use sometimes. Suggested by
Niels Möller.
<li> <code>mpf_set_str</code> and <code>mpf_inp_str</code> could usefully
accept 0x, 0b etc when base==0. Perhaps the exponent could default to
decimal in this case, with a further 0x, 0b etc allowed there.
Eg. 0xFFAA@0x5A. A leading "0" for octal would match the integers, but
probably something like "0.123" ought not mean octal.
<li> <code>GMP_LONG_LONG_LIMB</code> or some such could become a documented
feature of gmp.h, so applications could know whether to
<code>printf</code> a limb using <code>%lu</code> or <code>%Lu</code>.
<li> <code>PRIdMP_LIMB</code> and similar defines following C99
<inttypes.h> might be of use to applications printing limbs.
Perhaps they should be defined only if specifically requested, the way
<inttypes.h> does. But if <code>GMP_LONG_LONG_LIMB</code> or
whatever is added then perhaps this can easily enough be left to
applications.
<li> <code>mpf_get_ld</code> and <code>mpf_set_ld</code> converting
<code>mpf_t</code> to and from <code>long double</code>. Other
<code>long double</code> routines would be desirable too, but these would
be a start. Often <code>long double</code> is the same as
<code>double</code>, which is easy but pretty pointless. Should
recognise the Intel 80-bit format on i386, and IEEE 128-bit quad on
sparc, hppa and power. Might like an ABI sub-option or something when
it's a compiler option for 64-bit or 128-bit <code>long double</code>.
<li> <code>gmp_printf</code> could accept <code>%b</code> for binary output.
It'd be nice if it worked for plain <code>int</code> etc too, not just
<code>mpz_t</code> etc.
<li> <code>gmp_printf</code> in fact could usefully accept an arbitrary base,
for both integer and float conversions. A base either in the format
string or as a parameter with <code>*</code> should be allowed. Maybe
<code>&13b</code> (b for base) or something like that.
<li> <code>gmp_printf</code> could perhaps have a type code for an
<code>mp_limb_t</code>. That would save an application from having to
worry whether it's a <code>long</code> or a <code>long long</code>.
<li> <code>gmp_printf</code> could perhaps accept <code>mpq_t</code> for float
conversions, eg. <code>"%.4Qf"</code>. This would be merely for
convenience, but still might be useful. Rounding would be the same as
for an <code>mpf_t</code> (ie. currently round-to-nearest, but not
actually documented). Alternately, perhaps a separate
<code>mpq_get_str_point</code> or some such might be more use. Suggested
by Pedro Gimeno.
<li> <code>gmp_printf</code> could usefully accept a flag to control the
rounding of float conversions. The wouldn't do much for
<code>mpf_t</code>, but would be good if <code>mpfr_t</code> was
supported in the future, or perhaps for <code>mpq_t</code>. Something
like <code>&*r</code> (r for rounding, and mpfr style
<code>GMP_RND</code> parameter).
<li> <code>mpz_combit</code> to toggle a bit would be a good companion for
<code>mpz_setbit</code> and <code>mpz_clrbit</code>. Suggested by Niels
Möller (and has done some work towards it).
<li> <code>mpz_scan0_reverse</code> or <code>mpz_scan0low</code> or some such
searching towards the low end of an integer might match
<code>mpz_scan0</code> nicely. Likewise for <code>scan1</code>.
Suggested by Roberto Bagnara.
<li> <code>mpz_bit_subset</code> or some such to test whether one integer is a
bitwise subset of another might be of use. Some sort of return value
indicating whether it's a proper or non-proper subset would be good and
wouldn't cost anything in the implementation. Suggested by Roberto
Bagnara.
<li> <code>gmp_randinit_r</code> and maybe <code>gmp_randstate_set</code> to
init-and-copy or to just copy a <code>gmp_randstate_t</code>. Suggested
by Pedro Gimeno.
<li> <code>mpf_get_ld</code>, <code>mpf_set_ld</code>: Conversions between
<code>mpf_t</code> and <code>long double</code>, suggested by Dan
Christensen. There'd be some work to be done by <code>configure</code>
to recognise the format in use. xlc on aix for instance apparently has
an option for either plain double 64-bit or quad 128-bit precision. This
might mean library contents vary with the compiler used to build, which
is undesirable. It might be possible to detect the mode the application
is compiling with, and try to avoid mismatch problems.
<li> <code>mpz_sqrt_if_perfect_square</code>: When
<code>mpz_perfect_square_p</code> does its tests it calculates a square
root and then discards it. For some applications it might be useful to
return that root. Suggested by Jason Moxham.
<li> <code>mpz_get_ull</code>, <code>mpz_set_ull</code>,
<code>mpz_get_sll</code>, <code>mpz_get_sll</code>: Conversions for
<code>long long</code>. These would aid interoperability, though a
mixture of GMP and <code>long long</code> would probably not be too
common. Disadvantages of using <code>long long</code> in libgmp.a would
be
<ul>
<li> Library contents vary according to the build compiler.
<li> gmp.h would need an ugly <code>#ifdef</code> block to decide if the
application compiler could take the <code>long long</code>
prototypes.
<li> Some sort of <code>LIBGMP_HAS_LONGLONG</code> would be wanted to
indicate whether the functions are available. (Applications using
autoconf could probe the library too.)
</ul>
It'd be possible to defer the need for <code>long long</code> to
application compile time, by having something like
<code>mpz_set_2ui</code> called with two halves of a <code>long
long</code>. Disadvantages of this would be,
<ul>
<li> Bigger code in the application, though perhaps not if a <code>long
long</code> is normally passed as two halves anyway.
<li> <code>mpz_get_ull</code> would be a rather big inline, or would have
to be two function calls.
<li> <code>mpz_get_sll</code> would be a worse inline, and would put the
treatment of <code>-0x10..00</code> into applications (see
<code>mpz_get_si</code> correctness above).
<li> Although having libgmp.a independent of the build compiler is nice,
it sort of sacrifices the capabilities of a good compiler to
uniformity with inferior ones.
</ul>
Plain use of <code>long long</code> is probably the lesser evil, if only
because it makes best use of gcc.
<li> <code>mpz_strtoz</code> parsing the same as <code>strtol</code>.
Suggested by Alexander Kruppa.
</ul>
<h4>Configuration</h4>
<ul>
<li> Floating-point format: <code>GMP_C_DOUBLE_FORMAT</code> seems to work
well. Get rid of the <code>#ifdef</code> mess in gmp-impl.h and use the
results of the test instead.
<li> a29k: umul.s and udiv.s exist but don't get used.
<li> ARM: <code>umul_ppmm</code> in longlong.h always uses <code>umull</code>,
but is that available only for M series chips or some such? Perhaps it
should be configured in some way.
<li> HPPA: config.guess should recognize 7000, 7100, 7200, and 8x00.
<li> HPPA 2.0w: gcc is rumoured to support 2.0w as of version 3, though
perhaps just as a build-time choice. In any case, figure out how to
identify a suitable gcc or put it in the right mode, for the GMP compiler
choices.
<li> IA64: Latest libtool has some nonsense to detect ELF-32 or ELF-64 on
<code>ia64-*-hpux*</code>. Does GMP need to know anything about that?
<li> Mips: config.guess should say mipsr3000, mipsr4000, mipsr10000, etc.
"hinv -c processor" gives lots of information on Irix. Standard
config.guess appends "el" to indicate endianness, but
<code>AC_C_BIGENDIAN</code> seems the best way to handle that for GMP.
<li> PowerPC: The function descriptor nonsense for AIX is currently driven by
<code>*-*-aix*</code>. It might be more reliable to do some sort of
feature test, examining the compiler output perhaps. It might also be
nice to merge the aix.m4 files into powerpc-defs.m4.
<li> Sparc: <code>config.guess</code> recognises various exact sparcs, make
use of that information in <code>configure</code> (work on this is in
progress).
<li> Sparc32: floating point or integer <code>udiv</code> should be selected
according to the CPU target. Currently floating point ends up being
used on all sparcs, which is probably not right for generic V7 and V8.
<li> Sparc: The use of <code>-xtarget=native</code> with <code>cc</code> is
incorrect when cross-compiling, the target should be set according to the
configured <code>$host</code> CPU.
<li> m68k: config.guess can detect 68000, 68010, CPU32 and 68020, but relies
on system information for 030, 040 and 060. Can they be identified by
running some code?
<li> m68k: gas 2.11.90.0.1 pads with zero bytes in text segments, which is not
valid code. Probably need <code>.balignw <n>,0x4e7f</code> to get
nops, if <code>ALIGN</code> is going to be used for anything that's
executed across.
<li> Some CPUs have <code>umul</code> and <code>udiv</code> code not being
used. Check all such for bit rot and then put umul and udiv in
<code>$gmp_mpn_functions_optional</code> as "standard optional" objects.
<br> In particular Sparc and SparcV8 on non-gcc should benefit from
umul.asm enabled; the generic umul is suspected to be making sqr_basecase
slower than mul_basecase.
<li> HPPA <code>mpn_umul_ppmm</code> and <code>mpn_udiv_qrnnd</code> have a
different parameter order than those functions on other CPUs. It might
avoid confusion to have them under different names, maybe
<code>mpn_umul_ppmm_r</code> or some such. Prototypes then wouldn't
be conditionalized, and the appropriate form could be selected with the
<code>HAVE_NATIVE</code> scheme if/when the code switches to use a
<code>PROLOGUE</code> style.
<li> <code>DItype</code>: The setup in gmp-impl.h for non-GCC could use an
autoconf test to determine whether <code>long long</code> is available.
<li> m88k: Make the assembler code work on non-underscore systems. Conversion
to .asm would be desirable. Ought to be easy, but would want to be
tested.
<li> z8k: The use of a 32-bit limb in mpn/z8000x as opposed to 16-bits in
mpn/z8000 could be an ABI choice. But this chip is obsolete and nothing
is likely to be done unless someone is actively using it.
<li> config.m4 is generated only by the configure script, it won't be
regenerated by config.status. Creating it as an <code>AC_OUTPUT</code>
would work, but it might upset "make" to have things like <code>L$</code>
get into the Makefiles through <code>AC_SUBST</code>.
<code>AC_CONFIG_COMMANDS</code> would be the alternative. With some
careful m4 quoting the <code>changequote</code> calls might not be
needed, which might free up the order in which things had to be output.
<li> <code>make distclean</code>: Only the mpn directory links which were
created are removed, but perhaps all possible links should be removed, in
case someone runs configure a second time without a
<code>distclean</code> in between. The only tricky part would be making
sure all possible <code>extra_functions</code> are covered.
<li> MinGW: Apparently a Cygwin version of gcc can be used by passing
<code>-mno-cygwin</code>. For <code>--host=*-*-mingw32*</code> it might
be convenient to automatically use that option, if it works. Needs
someone with a dual cygwin/mingw setup to test.
<li> Automake: Latest automake has a <code>CCAS</code>, <code>CCASFLAGS</code>
scheme. Though we probably wouldn't be using its assembler support we
could try to use those variables in compatible ways.
</ul>
<h4>Random Numbers</h4>
<ul>
<li> <code>_gmp_rand</code> is not particularly fast on the linear
congruential algorithm and could stand various improvements.
<ul>
<li> Make a second seed area within <code>gmp_randstate_t</code> (or
<code>_mp_algdata</code> rather) to save some copying.
<li> Make a special case for a single limb <code>2exp</code> modulus, to
avoid <code>mpn_mul</code> calls. Perhaps the same for two limbs.
<li> Inline the <code>lc</code> code, to avoid a function call and
<code>TMP_ALLOC</code> for every chunk.
<li> The special case for <code>seedn==0</code> will be very rarely used,
and on that basis seems unnecessary.
<li> Perhaps the <code>2exp</code> and general LC cases should be split,
for clarity (if the general case is retained).
</ul>
<li> <code>gmp_randinit_mers</code> for a Mersenne Twister generator. It's
likely to be more random and about the same speed as Knuth's 55-element
Fibonacci generator, and can probably become the default. Pedro Gimeno
has started on this.
<li> <code>gmp_randinit_lc</code>: Finish or remove. Doing a division for
every every step won't be very fast, so check whether the usefulness of
this algorithm can be justified. (Consensus is that it's not useful and
can be removed.)
<li> Blum-Blum-Shub: Finish or remove. A separate
<code>gmp_randinit_bbs</code> would be wanted, not the currently
commented out case in <code>gmp_randinit</code>.
<li> <code>_gmp_rand</code> could be done as a function pointer within
<code>gmp_randstate_t</code> (or rather in the <code>_mp_algdata</code>
part), instead of switching on a <code>gmp_randalg_t</code>. Likewise
<code>gmp_randclear</code>, and perhaps <code>gmp_randseed</code> if it
became algorithm-specific. This would be more modular, and would ensure
only code for the desired algorithms is dragged into the link.
<li> <code>mpz_urandomm</code> should do something for n<=0, but what?
<li> <code>mpz_urandomm</code> implementation looks like it could be improved.
Perhaps it's enough to calculate <code>nbits</code> as ceil(log2(n)) and
call <code>_gmp_rand</code> until a value <code><n</code> is obtained.
<li> <code>gmp_randstate_t</code> used for parameters perhaps should become
<code>gmp_randstate_ptr</code> the same as other types.
<li> Some of the empirical randomness tests could be included in a "make
check". They ought to work everywhere, for a given seed at least.
</ul>
<h4>Miscellaneous</h4>
<ul>
<li> Make <code>mpz_div</code> and <code>mpz_divmod</code> use rounding
analogous to <code>mpz_mod</code>. Document, and list as an
incompatibility.
<li> <code>mpz_gcdext</code> and <code>mpn_gcdext</code> ought to document
what range of values the generated cofactors can take, and preferably
ensure the definition uniquely specifies the cofactors for given inputs.
A basic extended Euclidean algorithm or multi-step variant leads to
|x|<|b| and |y|<|a| or something like that, but there's probably
two solutions under just those restrictions.
<li> <code>mpz_invert</code> should call <code>mpn_gcdext</code> directly.
<li> demos/factorize.c: use <code>mpz_divisible_ui_p</code> rather than
<code>mpz_tdiv_qr_ui</code>. (Of course dividing multiple primes at a
time would be better still.)
<li> The various test programs use quite a bit of the main
<code>libgmp</code>. This establishes good cross-checks, but it might be
better to use simple reference routines where possible. Where it's not
possible some attention could be paid to the order of the tests, so a
<code>libgmp</code> routine is only used for tests once it seems to be
good.
<li> <code>mpf_set_q</code> is very similar to <code>mpf_div</code>, it'd be
good for the two to share code. Perhaps <code>mpf_set_q</code> should
make some <code>mpf_t</code> aliases for its numerator and denominator
and just call <code>mpf_div</code>. Both would be simplified a good deal
by switching to <code>mpn_tdiv_qr</code> perhaps making them small enough
not to bother with sharing (especially since <code>mpf_set_q</code>
wouldn't need to watch out for overlaps).
<li> PowerPC: The cpu time base registers (per <code>mftb</code> and
<code>mftbu</code>) could be used for the speed and tune programs. Would
need to know its frequency of course. Usually it's 1/4 of bus speed
(eg. 25 MHz) but some chips drive it from an external input. Probably
have to measure to be sure.
<li> <code>MUL_FFT_THRESHOLD</code> etc: the FFT thresholds should allow a
return to a previous k at certain sizes. This arises basically due to
the step effect caused by size multiples effectively used for each k.
Looking at a graph makes it fairly clear.
<li> <code>__gmp_doprnt_mpf</code> does a rather unattractive round-to-nearest
on the string returned by <code>mpf_get_str</code>. Perhaps some variant
of <code>mpf_get_str</code> could be made which would better suit.
</ul>
<h4>Aids to Development</h4>
<ul>
<li> Add <code>ASSERT</code>s at the start of each user-visible mpz/mpq/mpf
function to check the validity of each <code>mp?_t</code> parameter, in
particular to check they've been <code>mp?_init</code>ed. This might
catch elementary mistakes in user programs. Care would need to be taken
over <code>MPZ_TMP_INIT</code>ed variables used internally. If nothing
else then consistency checks like size<=alloc, ptr not
<code>NULL</code> and ptr+size not wrapping around the address space,
would be possible. A more sophisticated scheme could track
<code>_mp_d</code> pointers and ensure only a valid one is used. Such a
scheme probably wouldn't be reentrant, not without some help from the
system.
<li> tune/time.c could try to determine at runtime whether
<code>getrusage</code> and <code>gettimeofday</code> are reliable.
Currently we pretend in configure that the dodgy m68k netbsd 1.4.1
<code>getrusage</code> doesn't exist. If a test might take a long time
to run then perhaps cache the result in a file somewhere.
</ul>
<h4>Documentation</h4>
<ul>
<li> Document conventions, like that <code> unsigned long int</code> is used for
bit counts/ranges, and that <code>mp_size_t</code> is used for limb counts.
<li> <code>mpz_inp_str</code> (etc) doesn't say when it stops reading digits.
</ul>
<h4>Bright Ideas</h4>
The following may or may not be feasible, and aren't likely to get done in the
near future, but are at least worth thinking about.
<ul>
<li> Reorganize longlong.h so that we can inline the operations even for the
system compiler. When there is no such compiler feature, make calls to
stub functions. Write such stub functions for as many machines as
possible.
<li> longlong.h could declare when it's using, or would like to use,
<code>mpn_umul_ppmm</code>, and the corresponding umul.asm file could be
included in libgmp only in that case, the same as is effectively done for
<code>__clz_tab</code>. Likewise udiv.asm and perhaps cntlz.asm. This
would only be a very small space saving, so perhaps not worth the
complexity.
<li> longlong.h could be built at configure time by concatenating or
#including fragments from each directory in the mpn path. This would
select CPU specific macros the same way as CPU specific assembler code.
Code used would no longer depend on cpp predefines, and the current
nested conditionals could be flattened out.
<li> <code>mpz_get_si</code> returns 0x80000000 for -0x100000000, whereas it's
sort of supposed to return the low 31 (or 63) bits. But this is
undocumented, and perhaps not too important.
<li> <code>mpz_*_ui</code> division routines currently return abs(a%b).
Perhaps make them return the real remainder instead? Return type would
be <code>signed long int</code>. But this would be an incompatible
change, so it might have to be under newly named functions.
<li> <code>mpz_init_set*</code> and <code>mpz_realloc</code> could allocate
say an extra 16 limbs over what's needed, so as to reduce the chance of
having to do a reallocate if the <code>mpz_t</code> grows a bit more.
This could only be an option, since it'd badly bloat memory usage in
applications using many small values.
<li> <code>mpq</code> functions could perhaps check for numerator or
denominator equal to 1, on the assumption that integers or
denominator-only values might be expected to occur reasonably often.
<li> <code>count_trailing_zeros</code> is used on more or less uniformly
distributed numbers in a couple of places. For some CPUs
<code>count_trailing_zeros</code> is slow and it's probably worth handling
the frequently occurring 0 to 2 trailing zeros cases specially.
<li> <code>mpf_t</code> might like to let the exponent be undefined when
size==0, instead of requiring it 0 as now. It should be possible to do
size==0 tests before paying attention to the exponent. The advantage is
not needing to set exp in the various places a zero result can arise,
which avoids some tedium but is otherwise perhaps not too important.
Currently <code>mpz_set_f</code> and <code>mpf_cmp_ui</code> depend on
exp==0, maybe elsewhere too.
<li> <code>__gmp_allocate_func</code>: Could use GCC <code>__attribute__
((malloc))</code> on this, though don't know if it'd do much. GCC 3.0
allows that attribute on functions, but not function pointers (see info
node "Attribute Syntax"), so would need a new autoconf test. This can
wait until there's a GCC that supports it.
<li> <code>mpz_add_ui</code> contains two <code>__GMPN_COPY</code>s, one from
<code>mpn_add_1</code> and one from <code>mpn_sub_1</code>. If those two
routines were opened up a bit maybe that code could be shared. When a
copy needs to be done there's no carry to append for the add, and if the
copy is non-empty no high zero for the sub. <br> An alternative would be
to do a copy at the start and then an in-place add or sub. Obviously
that duplicates the fetches and stores for carry propagation, but that's
normally only one or two limbs. The same applies to <code>mpz_add</code>
when one operand is longer than the other, and to <code>mpz_com</code>
since it's just -(x+1).
<li> <code>restrict</code>'ed pointers: Does the C99 definition of restrict
(one writer many readers, or whatever it is) suit the GMP style "same or
separate" function parameters? If so, judicious use might improve the
code generated a bit. Do any compilers have their own flavour of
restrict as "completely unaliased", and is that still usable?
<li> 68000: A 16-bit limb might suit 68000 better than 32-bits, since the
native multiply is only 16x16. Could have this as an <code>ABI</code>
option, selecting <code>_SHORT_LIMB</code> in gmp.h. Naturally a new set
of asm subroutines would be necessary. Would need new
<code>mpz_set_ui</code> etc since the current code assumes limb>=long,
but 2-limb operand forms would find a use for <code>long long</code> on
other processors too.
<li> Nx1 remainders can be taken at multiplier throughput speed by
pre-calculating an array "p[i] = 2^(i*<code>BITS_PER_MP_LIMB</code>) mod
m", then for the input limbs x calculating an inner product "sum
p[i]*x[i]", and a final 3x1 limb remainder mod m. If those powers take
roughly N divide steps to calculate then there'd be an advantage any time
the same m is used three or more times. Suggested by Victor Shoup in
connection with chinese-remainder style decompositions, but perhaps with
other uses.
</ul>
<hr>
</body>
</html>
<!--
Local variables:
eval: (add-hook 'write-file-hooks 'time-stamp)
time-stamp-start: "This file current as of "
time-stamp-format: "%:d %3b %:y"
time-stamp-end: "\\."
time-stamp-line-limit: 50
End:
-->