Mercurial > dropbear
diff tomsfastmath/changes.txt @ 643:a362b62d38b2 dropbear-tfm
Add tomsfastmath from git rev bfa4582842bc3bab42e4be4aed5703437049502a
with Makefile.in renamed
author | Matt Johnston <matt@ucc.asn.au> |
---|---|
date | Wed, 23 Nov 2011 18:10:20 +0700 |
parents | |
children |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/tomsfastmath/changes.txt Wed Nov 23 18:10:20 2011 +0700 @@ -0,0 +1,95 @@ +March 14th, 2007 +0.12 -- Christophe Devine contributed MIPS asm w00t + ++ quick release to get the MIPS code out there + +March 10th, 2007 +0.11 -- re-org'ed the source tree, it's cooler now + -- cleaned up mul/sqr so there is only one file per unit, should help building with older compilers and/or on slower boxes + -- [CRI] optimized fp_read_unsigned_bin + -- fixed -0 ... again ... I now have less hair on my head. + -- [CRI] renamed bn_reverse() -> fp_reverse() + +November 1st, 2006 +0.10 -- Unrolled Montgomery for 1..16 digits with TFM_SMALL_MONT_SET between 10% and 25% speedup depending on size. + -- fixed fp_sqr_comba.c so it builds in ISO C mode [Andreas Lange] + -- [email protected] pointed out fp_radix_size() had a few typos that affected correctness. Fixed. + -- Added support for ECC performance, e.g. define "-DTFM_ALREADY_SET -DTFM_ECC192" and it will disable + all of the unrolled code EXCEPT what is required for ECC P-192. It autodetects 32/64-bit platforms too. It's super neato. + Support for 192, 224, 256, 384 and 521 bit curves through the defines [see tfm.h] + -- AVR32 support added, define TFM_AVR32 to enable + +April 4th, 2006 +0.09 -- Bruce Guenter suggested I use --tag=CC for libtool builds where the compiler may think it's C++. + -- Added support for k=1 in exptmod for RSA exponents. Makes it more competitive with other libraries + -- added cutoffs to the comba32 sqr/mul code to better handle 640, 786 and 896 bit operands (e.g. for RSA-1280, RSA-1536 and RSA-1792), really no effect + for 64-bit boxes as these represent 1280, 1536 and 1792 bit operands (not likely to be invoked). + -- Removed karatsuba from the mul/sqr since they're not useful and slow. + -- added 20, 24 and 28 digit multipliers for oddsized RSA support. You can easily disable them by uncommenting the TFM_SQRXX and TFM_MULXX lines in + tfm.h to save space. Now GMP and TFM are roughly the same speed on the Pentium-M for RSA with LTC. + -- unrolled SSE2 code and optimize for platforms with load/store pipes (e.g. can store and load in a cycle). Got 4% or so boost on my Dothan laptop (marginal improvement on a P4 Prescott) + +Jan 26th, 2006 +0.08 -- Fixed a bug in the generic mult/sqr where we overflowed by one digit + +November 18th, 2005 +0.07 -- Fixes to fp_mul and fp_sqr to clean up the handling of the defines, fix to tfm.h to also clear up the prototypes. + -- Updates to build and run on a IBM PPC 405 [using GCC 3.4.4] + -- Made the "make" command renamable in the build system + +October 31st, 2005 +0.06 -- fixed fp_mul() and fp_sqr() to trim digits when overflows would occur. Produces numerically inprecise results + (e.g. the lower FP_SIZE digits) but shouldn't segfault at least ;-) + -- Updated the combas so you can turn on and off specific unrolled loops at build time + -- Michael Heyman reported a bug in s_fp_sub() that was pretty substantial and a bug in fp_montgomery_calc_normalization(). Fixed. + +August 1st, 2005 +0.05 -- Quick fix to the fp_invmod.c code to let it handle even moduli [required for LTC] + -- Added makefile.shared to make shared objects [required for LTC] + -- Improved makefiles to make them way more configurable + -- Added timing resistant fp_exptmod() enabled with TFM_TIMING_RESISTANT + +July 23rd, 2005 +0.04 -- Fixed bugs in the SSE2 squaring code + -- Rewrote the multipliers to be optimized for small inputs + -- Nelson Bolyard of the NSS crew submitted [among other things] new faster Montgomery reduction + code. It brings the performance for small numbers on the AMD64 and all numbers on the P4 + to a new level. Thanks! + -- Added missing ARM support for fp_montgomery_reduce.c that the NSS folk left off, Officially + the ARM code is for v4 and above WITH the "M" multiplier support (e.g. umlal instruction) + -- Added PPC32 support, define TFM_PPC32 to enable it, I used the "PowerPC 6xx" instruction + databook for reference. Does not require altivec. Should be fairly portable to the other + 32-bit PPCs provided they have mullw and mulhwu instructions. + [Note: porting the macros to PPC64 should be trivial, anyone with a shell to lend... email me!] + -- Rewrote the config a bit in tfm.h so you can better choose which set of "oh my god that's huge" code to + enable for your task. "generic" functions are ALWAYS included which are smaller but will cover the + gaps in the coverage for ya. + -- The PPC32 code has been verified to function on a Darwin box running GCC 2.95.2 + [Thanks to the folk at PeerSec for lending me a shell to use] + -- Fixed a bug in fp_exptmod() where if the exponent was negative AND the destination the output + would have the sign set to FP_NEG. + +March 1st, 2005 +0.03 -- Optimized squaring + -- Applied new license header to all files (still PD) + +September 18th, 2004 +0.02 -- Added TFM_LARGE to turn on/off 16x combas to save even more space. + This also helps prevent killing the cache on smaller cpus. + -- Cast memset to void in fp_init() to catch people who misuse the function (e.g. expect return) + Thanks to Johan Lindh + -- Cleaned up x86-64 support [faster montgomery reductions] + -- Autodetects x86-32 and x86-64 and enables it's asm now + -- Made test demo build cleaner in multilib platforms [e.g. mixed 32/64 bits] + -- Fix to fp_mod to ensure that remainder is of the same sign as the modulus. + -- Fixed bug in fp_montgomery_calc_normalization for single digit moduli + -- cleaned up ISO C macros in comba/mont to avoid branches [works best with GCC 3.4.x branch] + -- Added more testing to tfm.h to help detect misconfigured builds + -- Added TFM_NO_ASM which forces ASM off [even if it was autodetected]. + -- Added fp_radix_size() to API + -- Cleaned up demo/test.c to build with far fewer warnings (mostly %d => %lu fixes) + -- fp_exptmod() now supports negative exponent and base>modulus cases + -- Added fp_ident() which gives a string showing how TFM was configured. Useful for debuging... + -- fix gen.pl script so it includes the whole source tree now + +August 25th, 2004 +0.01 -- Initial Release