This commit fulfills the promise of the previous one: now one of the
branches of sshbn.h's big ifdef _doesn't_ define a BignumDblInt, and
instead provides implementations of the primitive arithmetic macros in
terms of Visual Studio's x86-64 compiler intrinsics. So now, when this
codebase is compiled with 64-bit VS, it can use a 64-bit BignumInt and
everything still seems to work.
As I mentioned in the previous commit, I'm going to want PuTTY to be
able to run sensibly when compiled with 64-bit Visual Studio,
including handling bignums in 64-bit chunks for speed. Unfortunately,
64-bit VS does not provide any type we can use as BignumDblInt in that
situation (unlike 64-bit gcc and clang, which give us __uint128_t).
The only facilities it provides are compiler intrinsics to access an
add-with-carry operation and a 64x64->128 multiplication (the latter
delivering its product in two separate 64-bit output chunks).
Hence, here's a substantial rework of the bignum code to make it
implement everything in terms of _those_ primitives, rather than
depending throughout on having BignumDblInt available to use ad-hoc.
BignumDblInt does still exist, for the moment, but now it's an
internal implementation detail of sshbn.h, only declared inside a new
set of macros implementing arithmetic primitives, and not accessible
to any code outside sshbn.h (which confirms that I really did catch
all uses of it and remove them).
The resulting code is surprisingly nice-looking, actually. You'd
expect more hassle and roundabout circumlocutions when you drop down
to using a more basic set of primitive operations, but actually, in
many cases it's turned out shorter to write things in terms of the new
BignumADC and BignumMUL macros - because almost all my uses of
BignumDblInt were implementing those operations anyway, taking several
lines at a time, and now they can do each thing in just one line.
The biggest headache was Poly1305: I wasn't able to find any sensible
way to adapt the existing Python script that generates the various
per-int-size implementations of arithmetic mod 2^130-5, and so I had
to rewrite it from scratch instead, with nothing in common with the
old version beyond a handful of comments. But even that seems to have
worked out nicely: the new version has much more legible descriptions
of the high-level algorithms, by virtue of having a 'Multiprecision'
type which wraps up the division into words, and yet Multiprecision's
range analysis allows it to automatically drop out special cases such
as multiplication by 5 being much easier than multiplication by
another multi-word integer.
DIVMOD_WORD is a portability hazard, because implementing it requires
either a way to get direct access to the x86 DIV instruction or
equivalent (be it inline assembler or a compiler intrinsic), or else
an integer type we can use as BignumDblInt. But I'm starting to think
about porting to 64-bit Visual Studio with a 64-bit BignumInt, and in
that situation neither of those options will be available.
I could write a piece of _out_-of-line x86-64 assembler in a separate
source file and put a function call in DIVMOD_WORD, but instead I've
decided to solve the problem in a more futureproof way: remove
DIVMOD_WORD totally and write a division function that doesn't need it
at all, solving not only today's porting headache but all future ones
in this area.
The new implementation works by precomputing (a good enough
approximation to) the leading word of the reciprocal of the modulus,
and then getting each word of quotient by multiplying by that
reciprocal, where we previously used DIVMOD_WORD to divide by the
leading word of the actual modulus. The reciprocal itself is computed
outside internal_mod() and passed in as a parameter, allowing me to
save time by only computing it once when I'm about to do a modpow.
To some extent this complicates the implementation: the advantage of
DIVMOD_WORD was that it yielded a full word q of quotient every time
it was used, so the subtraction of q*m from the input could be done in
a nicely word-aligned way. But the reciprocal multiply approach yields
_almost_ a full word of quotient, because you have to make the
reciprocal a bit short to avoid overflow at multiplication time. For a
start, this means we have to do fractionally more iterations of the
main loop; but more painfully, we can no longer depend on the
subtraction of q*m at every step being word-aligned, and instead we
have to be prepared to do it at any bit shift.
But the flip side is that once we've implemented that, the rest of the
algorithm becomes a lot less full of horrible special cases: in
particular, we can now completely throw away the horribleness at all
the call sites where we shift the modulus up by a fractional word to
set its top bit, and then have to do a little dance to get the last
few bits of quotient involving a second call to internal_mod.
So there are points both for and against the new implementation in
simplicity terms; but I think on balance it's more comprehensible than
the old one, and a quick timing test suggests it also ends up a touch
faster overall - the new testbn gets through the output of
testdata/bignum.py in 4.034s where the old one took 4.392s.
gcc and clang both provide a type called __uint128_t when compiling
for 64-bit targets, code-generated more or less similarly to the way
64-bit long longs are handled on 32-bit targets (spanning two
registers, using ADD/ADC, that sort of thing). Where this is available
(and they also provide a handy macro to make it easy to detect), we
should obviously use it, so that we can handle bignums a larger chunk
at a time and make use of the full width of the hardware's multiplier.
Preliminary benchmarking using 'testbn' suggests a factor of about 2.5
improvement.
I've added the new possibility to the ifdefs in sshbn.h, and also
re-run contrib/make1305.py to generate a set of variants of the
poly1305 arithmetic for the new size of BignumInt.
This allows files other than sshbn.c to work with the primitives
necessary to build multi-word arithmetic functions satisfying all of
PuTTY's portability constraints.
