Date: Tue, 4 Mar 2014 06:13:39 +0400
From: Solar Designer <solar@...nwall.com>
To: discussions@...sword-hashing.net
Subject: wider integer multiply on 32-bit x86

Hi,

Normally, on 32-bit x86 without SSE2 (thus, on Pentium 3 and older, or
when code is compiled such that SSE2 is not enabled) the widest integer
multiply available is 32x32->64, via the [I]MUL instruction.  There are
two problems with this: the instruction uses the specific EDX:EAX
registers, so we can't have more than one such multiply in progress
until we've read/replaced at least the EAX contents(*), and 32x32->64 is
not very wide.

(*) Does register renaming (on CPUs that have it) help avoid stalls
here?  Maybe.

I looked into using x86 FPU instructions to get around one or both of
these limitations.  The attached program implements 31x31->62 and
63x63->63 using the FPU.  This does force us to use memory (L1 cache)
instead of registers, but x86 is register starved anyway, so with 2+
interleaved instruction streams dealing with bigger than 32-bit values
we'd have register spills to memory anyway.  We can have 2 interleaved
instruction streams by having one use the regular [I]MUL and the other
the FPU, and we can probably have more than 2 by also using FXCH (which
is normally free starting with the original Pentium).

I think 4 instructions, including the loads and stores, for a 63x63->63
multiply is rather good.  Without this trick, it'd take 4 _multiplies_
to implement the equivalent via 32x32->32 (or perhaps 3 multiplies if we
also use the 32x32->64).  Some bigint library could use this trick,
perhaps for some nice speedup on those older CPUs/builds (does any use
it already?)

Now, if we only wanted 32x32->64 without the EDX:EAX limitation on
interleaving, things are not as nice.  We'd have to define our KDF so
that one or both source operands is at most 31-bit (not 32).  (And we
wouldn't be taking full advantage of the support for 63-bit source
operands.)  I can think of some ways to do that without an explicit mask
in the inner loop.

Why bother with interleaved instruction streams, given that one stream
is enough for the latency hardening, and the ASIC die area difference
from more complete use of just a handful of multipliers is negligible?

My thinking is that more complete use of our CPUs' multiplication
throughput is desirable for not allowing for much room for speedup in
attacks by other CPUs and similar, which _would_ interleave multiple
instruction streams for multiple candidate passwords (e.g., we do this
in JtR for many hash types).  (Of course, those other CPUs will need to
have sufficient cache and memory bandwidth to be able to benefit from
the extra parallelism.  So not all will.  But many may.)  Thus, it is
somewhat desirable to constantly feed our SIMD multipliers on SSE2+, and
when we do that it is then desirable not to incur too much of a
performance hit in pre-SSE2 builds.  Is this worth the complexity (even
if in optimized implementations only)?  I'm not sure.  Yet I thought
I'd share.

Output from widermul on a P3 (should be same on other CPUs):

00000000 * 00000000 = 0000000000000000 0000000000000000 0000000000000000
00000001 * 00000000 = 0000000000000000 0000000000000000 0000000000000000
00000000 * 00000001 = 0000000000000000 0000000000000000 0000000000000000
00000001 * 00000001 = 0000000000000001 0000000000000001 0000000000000001
00000010 * 00000010 = 0000000000000100 0000000000000100 0000000000000100
000003e8 * 000003e8 = 00000000000f4240 00000000000f4240 00000000000f4240
00001000 * 00001000 = 0000000001000000 0000000001000000 0000000001000000
00010000 * 0000ffff = 00000000ffff0000 00000000ffff0000 00000000ffff0000
00010000 * 00010000 = 0000000100000000 0000000100000000 0000000100000000
00100000 * 00100000 = 0000010000000000 0000010000000000 0000010000000000
000f4240 * 000f4240 = 000000e8d4a51000 000000e8d4a51000 000000e8d4a51000
12345678 * 78765432 = 0890f2a4d5e84370 0890f2a4d5e84370 0890f2a4d5e84370
12345678 * 7fffffff = 091a2b3bedcba988 091a2b3bedcba988 091a2b3bedcba988
7fffffff * 7fffffff = 3fffffff00000001 3fffffff00000001 3fffffff00000001
80000000 * 80000000 = 4000000000000000 4000000000000000 4000000000000000
80000000 * 80000001 = 4000000080000000 3fffffff80000000 4000000080000000
80000000 * ffffffff = 7fffffff80000000 0000000080000000 7fffffff80000000
00000001 * 80000000 = 0000000080000000 ffffffff80000000 0000000080000000
00000001 * 80000001 = 0000000080000001 ffffffff80000001 0000000080000001
12345678 * 98765432 = 0ad77d73d5e84370 f8a326fbd5e84370 0ad77d73d5e84370
7fffffff * ffffffff = 7ffffffe80000001 ffffffff80000001 7ffffffe80000001
ffffffff * ffffffff = fffffffe00000001 0000000000000001 8000000000000000

The three results are: 32x32->64 (computed without FPU) followed by
31x31->62 and 63x63->63 (but with 32-bit inputs only in this test).
As you can see, when we exceed the allowed size of inputs or output for
the latter two functions, we get incorrect results (as expected).
As long as our inputs to 31x31->62 are at most 31-bit, it works fine.
As long as the output fits in 63 bits, our 63x63->63 works fine.

Alexander

View attachment "widermul.c" of type "text/x-c" (1217 bytes)

Powered by blists - more mailing lists