mirror/openssl
2
0
Fork 0
mirror of https://github.com/openssl/openssl.git synced 2024-12-09 05:51:54 +08:00
Code Issues Packages Projects Releases Wiki Activity
62aa5dd415
openssl/crypto/bn/asm/sparcv9a-mont.pl

883 lines
20 KiB
Perl
Raw Normal View History

Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
#!/usr/bin/env perl
# ====================================================================
# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
# project. The module is, however, dual licensed under OpenSSL and
# CRYPTOGAMS licenses depending on where you obtain it. For further
# details see http://www.openssl.org/~appro/cryptogams/.
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
# ====================================================================
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
# October 2005
#
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
# "Teaser" Montgomery multiplication module for UltraSPARC. Why FPU?
# Because unlike integer multiplier, which simply stalls whole CPU,
# FPU is fully pipelined and can effectively emit 48 bit partial
# product every cycle. Why not blended SPARC v9? One can argue that
# making this module dependent on UltraSPARC VIS extension limits its
Clarify binary compatibility with HAL/Fujitsu SPARC64 family.
2005-10-25 23:39:47 +08:00
# binary compatibility. Well yes, it does exclude SPARC64 prior-V(!)
# implementations from compatibility matrix. But the rest, whole Sun
# UltraSPARC family and brand new Fujitsu's SPARC64 V, all support
# VIS extension instructions used in this module. This is considered
Clarify HAL SPARC64 support situation in sparcv9a-mont.pl.
2006-11-28 19:07:36 +08:00
# good enough to not care about HAL SPARC64 users [if any] who have
# integer-only pure SPARCv9 module to "fall down" to.
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
# USI&II cores currently exhibit uniform 2x improvement [over pre-
# bn_mul_mont codebase] for all key lengths and benchmarks. On USIII
# performance improves few percents for shorter keys and worsens few
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
# percents for longer keys. This is because USIII integer multiplier
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
# is >3x faster than USI&II one, which is harder to match [but see
# TODO list below]. It should also be noted that SPARC64 V features
# out-of-order execution, which *might* mean that integer multiplier
Clarify binary compatibility with HAL/Fujitsu SPARC64 family.
2005-10-25 23:39:47 +08:00
# is pipelined, which in turn *might* be impossible to match... On
# additional note, SPARC64 V implements FP Multiply-Add instruction,
# which is perfectly usable in this context... In other words, as far
Clarify HAL SPARC64 support situation in sparcv9a-mont.pl.
2006-11-28 19:07:36 +08:00
# as Fujitsu SPARC64 V goes, talk to the author:-)
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
Unify sparcv9 assembler naming and build rules among 32- and 64-bit builds. Engage run-time switch between bn_mul_mont_fpu and bn_mul_mont_int.
2005-12-17 01:39:57 +08:00
# The implementation implies following "non-natural" limitations on
# input arguments:
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
# - num may not be less than 4;
# - num has to be even;
# Failure to meet either condition has no fatal effects, simply
# doesn't give any performance gain.
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
# TODO:
# - modulo-schedule inner loop for better performance (on in-order
# execution core such as UltraSPARC this shall result in further
# noticeable(!) improvement);
# - dedicated squaring procedure[?];
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
######################################################################
# November 2006
#
# Modulo-scheduled inner loops allow to interleave floating point and
# integer instructions and minimize Read-After-Write penalties. This
# results in *further* 20-50% perfromance improvement [depending on
# key length, more for longer keys] on USI&II cores and 30-80% - on
# USIII&IV.
Unify sparcv9 assembler naming and build rules among 32- and 64-bit builds. Engage run-time switch between bn_mul_mont_fpu and bn_mul_mont_int.
2005-12-17 01:39:57 +08:00
$fname="bn_mul_mont_fpu";
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
$bits=32;
As SPARCV9 CPU flavor is [expected to be] detected at run-time, we can afford to relax SPARCV9/8+ compiler command line and produce "unversal" binaries as we used to.
2005-12-19 17:10:06 +08:00
for (@ARGV) { $bits=64 if (/\-m64/ || /\-xarch\=v9/); }
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
if ($bits==64) {
$bias=2047;
$frame=192;
} else {
$bias=0;
$frame=128; # 96 rounded up to largest known cache-line
}
$locals=64;
# In order to provide for 32-/64-bit ABI duality, I keep integers wider
# than 32 bit in %g1-%g4 and %o0-%o5. %l0-%l7 and %i0-%i5 are used
# exclusively for pointers, indexes and other small values...
# int bn_mul_mont(
$rp="%i0"; # BN_ULONG *rp,
$ap="%i1"; # const BN_ULONG *ap,
$bp="%i2"; # const BN_ULONG *bp,
$np="%i3"; # const BN_ULONG *np,
Change bn_mul_mont declaration and BN_MONT_CTX. Update CHANGES.
2005-10-23 01:57:18 +08:00
$n0="%i4"; # const BN_ULONG *n0,
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
$num="%i5"; # int num);
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
$tp="%l0"; # t[num]
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
$ap_l="%l1"; # a[num],n[num] are smashed to 32-bit words and saved
$ap_h="%l2"; # to these four vectors as double-precision FP values.
$np_l="%l3"; # This way a bunch of fxtods are eliminated in second
$np_h="%l4"; # loop and L1-cache aliasing is minimized...
$i="%l5";
$j="%l6";
$mask="%l7"; # 16-bit mask, 0xffff
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
$n0="%g4"; # reassigned(!) to "64-bit" register
$carry="%i4"; # %i4 reused(!) for a carry bit
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
# FP register naming chart
#
# ..HILO
# dcba
# --------
# LOa
# LOb
# LOc
# LOd
# HIa
# HIb
# HIc
# HId
# ..a
# ..b
$ba="%f0"; $bb="%f2"; $bc="%f4"; $bd="%f6";
$na="%f8"; $nb="%f10"; $nc="%f12"; $nd="%f14";
$alo="%f16"; $alo_="%f17"; $ahi="%f18"; $ahi_="%f19";
$nlo="%f20"; $nlo_="%f21"; $nhi="%f22"; $nhi_="%f23";
$dota="%f24"; $dotb="%f26";
$aloa="%f32"; $alob="%f34"; $aloc="%f36"; $alod="%f38";
$ahia="%f40"; $ahib="%f42"; $ahic="%f44"; $ahid="%f46";
$nloa="%f48"; $nlob="%f50"; $nloc="%f52"; $nlod="%f54";
$nhia="%f56"; $nhib="%f58"; $nhic="%f60"; $nhid="%f62";
$ASI_FL16_P=0xD2; # magic ASI value to engage 16-bit FP load
$code=<<___;
.section ".text",#alloc,#execinstr
.global $fname
.align 32
$fname:
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
save %sp,-$frame-$locals,%sp
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
cmp $num,4
bl,a,pn %icc,.Lret
clr %i0
andcc $num,1,%g0 ! $num has to be even...
bnz,a,pn %icc,.Lret
clr %i0 ! signal "unsupported input value"
sparcv9a-mont was modified to handle 32-bit aligned input, but check for 64-bit alignment was not removed.
2007-03-20 16:54:51 +08:00
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
srl $num,1,$num
sparcv9a-mont was modified to handle 32-bit aligned input, but check for 64-bit alignment was not removed.
2007-03-20 16:54:51 +08:00
sethi %hi(0xffff),$mask
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ld [%i4+0],$n0 ! $n0 reassigned, remember?
sparcv9a-mont was modified to handle 32-bit aligned input, but check for 64-bit alignment was not removed.
2007-03-20 16:54:51 +08:00
or $mask,%lo(0xffff),$mask
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ld [%i4+4],%o0
sllx %o0,32,%o0
or %o0,$n0,$n0 ! $n0=n0[1].n0[0]
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
sll $num,3,$num ! num*=8
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %sp,$bias,%o0 ! real top of stack
sll $num,2,%o1
add %o1,$num,%o1 ! %o1=num*5
sub %o0,%o1,%o0
and %o0,-2048,%o0 ! optimize TLB utilization
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
sub %o0,$bias,%sp ! alloca(5*num*8)
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
rd %asi,%o7 ! save %asi
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %sp,$bias+$frame+$locals,$tp
add $tp,$num,$ap_l
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
add $ap_l,$num,$ap_l ! [an]p_[lh] point at the vectors' ends !
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add $ap_l,$num,$ap_h
add $ap_h,$num,$np_l
add $np_l,$num,$np_h
wr %g0,$ASI_FL16_P,%asi ! setup %asi for 16-bit FP loads
add $rp,$num,$rp ! readjust input pointers to point
add $ap,$num,$ap ! at the ends too...
add $bp,$num,$bp
add $np,$num,$np
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
stx %o7,[%sp+$bias+$frame+48] ! save %asi
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
sub %g0,$num,$i ! i=-num
sub %g0,$num,$j ! j=-num
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add $ap,$j,%o3
add $bp,$i,%o4
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
ld [%o3+4],%g1 ! bp[0]
ld [%o3+0],%o0
ld [%o4+4],%g5 ! ap[0]
sllx %g1,32,%g1
ld [%o4+0],%o1
sllx %g5,32,%g5
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
or %g1,%o0,%o0
or %g5,%o1,%o1
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
add $np,$j,%o5
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
mulx %o1,%o0,%o0 ! ap[0]*bp[0]
mulx $n0,%o0,%o0 ! ap[0]*bp[0]*n0
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
stx %o0,[%sp+$bias+$frame+0]
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ld [%o3+0],$alo_ ! load a[j] as pair of 32-bit words
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $alo
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ld [%o3+4],$ahi_
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $ahi
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ld [%o5+0],$nlo_ ! load n[j] as pair of 32-bit words
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $nlo
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ld [%o5+4],$nhi_
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $nhi
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
! transfer b[i] to FPU as 4x16-bit values
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ldda [%o4+2]%asi,$ba
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $alo,$alo
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ldda [%o4+0]%asi,$bb
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $ahi,$ahi
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ldda [%o4+6]%asi,$bc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $nlo,$nlo
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ldda [%o4+4]%asi,$bd
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $nhi,$nhi
! transfer ap[0]*b[0]*n0 to FPU as 4x16-bit values
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+6]%asi,$na
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $ba,$ba
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+4]%asi,$nb
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $bb,$bb
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+2]%asi,$nc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $bc,$bc
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+0]%asi,$nd
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $bd,$bd
std $alo,[$ap_l+$j] ! save smashed ap[j] in double format
fxtod $na,$na
std $ahi,[$ap_h+$j]
fxtod $nb,$nb
std $nlo,[$np_l+$j] ! save smashed np[j] in double format
fxtod $nc,$nc
std $nhi,[$np_h+$j]
fxtod $nd,$nd
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$ba,$aloa
fmuld $nlo,$na,$nloa
fmuld $alo,$bb,$alob
fmuld $nlo,$nb,$nlob
fmuld $alo,$bc,$aloc
faddd $aloa,$nloa,$nloa
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nlo,$nc,$nloc
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$bd,$alod
faddd $alob,$nlob,$nlob
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nlo,$nd,$nlod
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$ba,$ahia
faddd $aloc,$nloc,$nloc
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$na,$nhia
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bb,$ahib
faddd $alod,$nlod,$nlod
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nb,$nhib
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bc,$ahic
faddd $ahia,$nhia,$nhia
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nc,$nhic
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bd,$ahid
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
faddd $ahib,$nhib,$nhib
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $nhi,$nd,$nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
faddd $ahic,$nhic,$dota ! $nhic
faddd $ahid,$nhid,$dotb ! $nhid
faddd $nloc,$nhia,$nloc
faddd $nlod,$nhib,$nlod
fdtox $nloa,$nloa
fdtox $nlob,$nlob
fdtox $nloc,$nloc
fdtox $nlod,$nlod
std $nloa,[%sp+$bias+$frame+0]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add $j,8,$j
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nlob,[%sp+$bias+$frame+8]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add $ap,$j,%o4
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nloc,[%sp+$bias+$frame+16]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add $np,$j,%o5
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nlod,[%sp+$bias+$frame+24]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
ld [%o4+0],$alo_ ! load a[j] as pair of 32-bit words
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $alo
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
ld [%o4+4],$ahi_
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $ahi
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
ld [%o5+0],$nlo_ ! load n[j] as pair of 32-bit words
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $nlo
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
ld [%o5+4],$nhi_
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fzeros $nhi
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $alo,$alo
fxtod $ahi,$ahi
fxtod $nlo,$nlo
fxtod $nhi,$nhi
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+0],%o0
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$ba,$aloa
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+8],%o1
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $nlo,$na,$nloa
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+16],%o2
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$bb,$alob
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+24],%o3
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $nlo,$nb,$nlob
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o0,16,%o7
std $alo,[$ap_l+$j] ! save smashed ap[j] in double format
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$bc,$aloc
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add %o7,%o1,%o1
std $ahi,[$ap_h+$j]
faddd $aloa,$nloa,$nloa
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nlo,$nc,$nloc
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o1,16,%o7
std $nlo,[$np_l+$j] ! save smashed np[j] in double format
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$bd,$alod
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add %o7,%o2,%o2
std $nhi,[$np_h+$j]
faddd $alob,$nlob,$nlob
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nlo,$nd,$nlod
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o2,16,%o7
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$ba,$ahia
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15]
faddd $aloc,$nloc,$nloc
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$na,$nhia
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
!and %o0,$mask,%o0
!and %o1,$mask,%o1
!and %o2,$mask,%o2
!sllx %o1,16,%o1
!sllx %o2,32,%o2
!sllx %o3,48,%o7
!or %o1,%o0,%o0
!or %o2,%o0,%o0
!or %o7,%o0,%o0 ! 64-bit result
srlx %o3,16,%g1 ! 34-bit carry
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bb,$ahib
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
faddd $alod,$nlod,$nlod
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nb,$nhib
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bc,$ahic
faddd $ahia,$nhia,$nhia
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nc,$nhic
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bd,$ahid
faddd $ahib,$nhib,$nhib
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nd,$nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
faddd $dota,$nloa,$nloa
faddd $dotb,$nlob,$nlob
faddd $ahic,$nhic,$dota ! $nhic
faddd $ahid,$nhid,$dotb ! $nhid
faddd $nloc,$nhia,$nloc
faddd $nlod,$nhib,$nlod
fdtox $nloa,$nloa
fdtox $nlob,$nlob
fdtox $nloc,$nloc
fdtox $nlod,$nlod
std $nloa,[%sp+$bias+$frame+0]
std $nlob,[%sp+$bias+$frame+8]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
addcc $j,8,$j
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nloc,[%sp+$bias+$frame+16]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
bz,pn %icc,.L1stskip
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nlod,[%sp+$bias+$frame+24]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
SPARC Solaris and Linux assemblers treat .align directive differently. PR: 1547
2007-06-20 20:24:22 +08:00
.align 32 ! incidentally already aligned !
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
.L1st:
add $ap,$j,%o4
add $np,$j,%o5
ld [%o4+0],$alo_ ! load a[j] as pair of 32-bit words
fzeros $alo
ld [%o4+4],$ahi_
fzeros $ahi
ld [%o5+0],$nlo_ ! load n[j] as pair of 32-bit words
fzeros $nlo
ld [%o5+4],$nhi_
fzeros $nhi
fxtod $alo,$alo
fxtod $ahi,$ahi
fxtod $nlo,$nlo
fxtod $nhi,$nhi
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+0],%o0
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $alo,$ba,$aloa
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+8],%o1
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nlo,$na,$nloa
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+16],%o2
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $alo,$bb,$alob
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+24],%o3
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nlo,$nb,$nlob
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o0,16,%o7
std $alo,[$ap_l+$j] ! save smashed ap[j] in double format
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $alo,$bc,$aloc
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add %o7,%o1,%o1
std $ahi,[$ap_h+$j]
faddd $aloa,$nloa,$nloa
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nlo,$nc,$nloc
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o1,16,%o7
std $nlo,[$np_l+$j] ! save smashed np[j] in double format
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $alo,$bd,$alod
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add %o7,%o2,%o2
std $nhi,[$np_h+$j]
faddd $alob,$nlob,$nlob
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nlo,$nd,$nlod
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o2,16,%o7
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $ahi,$ba,$ahia
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15]
and %o0,$mask,%o0
faddd $aloc,$nloc,$nloc
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nhi,$na,$nhia
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
and %o1,$mask,%o1
and %o2,$mask,%o2
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $ahi,$bb,$ahib
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
sllx %o1,16,%o1
faddd $alod,$nlod,$nlod
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nhi,$nb,$nhib
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
sllx %o2,32,%o2
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $ahi,$bc,$ahic
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
sllx %o3,48,%o7
or %o1,%o0,%o0
faddd $ahia,$nhia,$nhia
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nhi,$nc,$nhic
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
or %o2,%o0,%o0
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $ahi,$bd,$ahid
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
or %o7,%o0,%o0 ! 64-bit result
faddd $ahib,$nhib,$nhib
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
fmuld $nhi,$nd,$nhid
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
addcc %g1,%o0,%o0
faddd $dota,$nloa,$nloa
srlx %o3,16,%g1 ! 34-bit carry
faddd $dotb,$nlob,$nlob
bcs,a %xcc,.+8
add %g1,1,%g1
stx %o0,[$tp] ! tp[j-1]=
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
faddd $ahic,$nhic,$dota ! $nhic
faddd $ahid,$nhid,$dotb ! $nhid
faddd $nloc,$nhia,$nloc
faddd $nlod,$nhib,$nlod
fdtox $nloa,$nloa
fdtox $nlob,$nlob
fdtox $nloc,$nloc
fdtox $nlod,$nlod
std $nloa,[%sp+$bias+$frame+0]
std $nlob,[%sp+$bias+$frame+8]
std $nloc,[%sp+$bias+$frame+16]
std $nlod,[%sp+$bias+$frame+24]
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
addcc $j,8,$j
bnz,pt %icc,.L1st
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add $tp,8,$tp
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
.L1stskip:
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
fdtox $dota,$dota
fdtox $dotb,$dotb
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
ldx [%sp+$bias+$frame+0],%o0
ldx [%sp+$bias+$frame+8],%o1
ldx [%sp+$bias+$frame+16],%o2
ldx [%sp+$bias+$frame+24],%o3
srlx %o0,16,%o7
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
std $dota,[%sp+$bias+$frame+32]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
add %o7,%o1,%o1
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
std $dotb,[%sp+$bias+$frame+40]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
srlx %o1,16,%o7
add %o7,%o2,%o2
srlx %o2,16,%o7
add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15]
and %o0,$mask,%o0
and %o1,$mask,%o1
and %o2,$mask,%o2
sllx %o1,16,%o1
sllx %o2,32,%o2
sllx %o3,48,%o7
or %o1,%o0,%o0
or %o2,%o0,%o0
or %o7,%o0,%o0 ! 64-bit result
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
ldx [%sp+$bias+$frame+32],%o4
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
addcc %g1,%o0,%o0
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
ldx [%sp+$bias+$frame+40],%o5
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
srlx %o3,16,%g1 ! 34-bit carry
bcs,a %xcc,.+8
add %g1,1,%g1
stx %o0,[$tp] ! tp[j-1]=
add $tp,8,$tp
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
srlx %o4,16,%o7
add %o7,%o5,%o5
and %o4,$mask,%o4
sllx %o5,16,%o7
or %o7,%o4,%o4
addcc %g1,%o4,%o4
srlx %o5,48,%g1
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
bcs,a %xcc,.+8
add %g1,1,%g1
mov %g1,$carry
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
stx %o4,[$tp] ! tp[num-1]=
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ba .Louter
add $i,8,$i
.align 32
.Louter:
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
sub %g0,$num,$j ! j=-num
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %sp,$bias+$frame+$locals,$tp
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
add $ap,$j,%o3
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add $bp,$i,%o4
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
ld [%o3+4],%g1 ! bp[i]
ld [%o3+0],%o0
ld [%o4+4],%g5 ! ap[0]
sllx %g1,32,%g1
ld [%o4+0],%o1
sllx %g5,32,%g5
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
or %g1,%o0,%o0
or %g5,%o1,%o1
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [$tp],%o2 ! tp[0]
mulx %o1,%o0,%o0
addcc %o2,%o0,%o0
mulx $n0,%o0,%o0 ! (ap[0]*bp[i]+t[0])*n0
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
stx %o0,[%sp+$bias+$frame+0]
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
! transfer b[i] to FPU as 4x16-bit values
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
ldda [%o4+2]%asi,$ba
ldda [%o4+0]%asi,$bb
ldda [%o4+6]%asi,$bc
ldda [%o4+4]%asi,$bd
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
! transfer (ap[0]*b[i]+t[0])*n0 to FPU as 4x16-bit values
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+6]%asi,$na
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $ba,$ba
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+4]%asi,$nb
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $bb,$bb
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+2]%asi,$nc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $bc,$bc
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldda [%sp+$bias+$frame+0]%asi,$nd
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
fxtod $bd,$bd
ldd [$ap_l+$j],$alo ! load a[j] in double format
fxtod $na,$na
ldd [$ap_h+$j],$ahi
fxtod $nb,$nb
ldd [$np_l+$j],$nlo ! load n[j] in double format
fxtod $nc,$nc
ldd [$np_h+$j],$nhi
fxtod $nd,$nd
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$ba,$aloa
fmuld $nlo,$na,$nloa
fmuld $alo,$bb,$alob
fmuld $nlo,$nb,$nlob
fmuld $alo,$bc,$aloc
faddd $aloa,$nloa,$nloa
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nlo,$nc,$nloc
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $alo,$bd,$alod
faddd $alob,$nlob,$nlob
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nlo,$nd,$nlod
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$ba,$ahia
faddd $aloc,$nloc,$nloc
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$na,$nhia
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bb,$ahib
faddd $alod,$nlod,$nlod
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nb,$nhib
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bc,$ahic
faddd $ahia,$nhia,$nhia
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
fmuld $nhi,$nc,$nhic
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $ahi,$bd,$ahid
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
faddd $ahib,$nhib,$nhib
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
fmuld $nhi,$nd,$nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
faddd $ahic,$nhic,$dota ! $nhic
faddd $ahid,$nhid,$dotb ! $nhid
faddd $nloc,$nhia,$nloc
faddd $nlod,$nhib,$nlod
fdtox $nloa,$nloa
fdtox $nlob,$nlob
fdtox $nloc,$nloc
fdtox $nlod,$nlod
std $nloa,[%sp+$bias+$frame+0]
std $nlob,[%sp+$bias+$frame+8]
std $nloc,[%sp+$bias+$frame+16]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
add $j,8,$j
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nlod,[%sp+$bias+$frame+24]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldd [$ap_l+$j],$alo ! load a[j] in double format
ldd [$ap_h+$j],$ahi
ldd [$np_l+$j],$nlo ! load n[j] in double format
ldd [$np_h+$j],$nhi
fmuld $alo,$ba,$aloa
fmuld $nlo,$na,$nloa
fmuld $alo,$bb,$alob
fmuld $nlo,$nb,$nlob
fmuld $alo,$bc,$aloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+0],%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $aloa,$nloa,$nloa
fmuld $nlo,$nc,$nloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+8],%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $alo,$bd,$alod
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+16],%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $alob,$nlob,$nlob
fmuld $nlo,$nd,$nlod
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+24],%o3
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$ba,$ahia
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o0,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $aloc,$nloc,$nloc
fmuld $nhi,$na,$nhia
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %o7,%o1,%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$bb,$ahib
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o1,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $alod,$nlod,$nlod
fmuld $nhi,$nb,$nhib
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %o7,%o2,%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$bc,$ahic
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o2,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahia,$nhia,$nhia
fmuld $nhi,$nc,$nhic
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15]
! why?
and %o0,$mask,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$bd,$ahid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
and %o1,$mask,%o1
and %o2,$mask,%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahib,$nhib,$nhib
fmuld $nhi,$nd,$nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
sllx %o1,16,%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $dota,$nloa,$nloa
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
sllx %o2,32,%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $dotb,$nlob,$nlob
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
sllx %o3,48,%o7
or %o1,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahic,$nhic,$dota ! $nhic
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
or %o2,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahid,$nhid,$dotb ! $nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
or %o7,%o0,%o0 ! 64-bit result
ldx [$tp],%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $nloc,$nhia,$nloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
addcc %o7,%o0,%o0
! end-of-why?
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $nlod,$nhib,$nlod
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o3,16,%g1 ! 34-bit carry
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fdtox $nloa,$nloa
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
bcs,a %xcc,.+8
add %g1,1,%g1
fdtox $nlob,$nlob
fdtox $nloc,$nloc
fdtox $nlod,$nlod
std $nloa,[%sp+$bias+$frame+0]
std $nlob,[%sp+$bias+$frame+8]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
addcc $j,8,$j
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nloc,[%sp+$bias+$frame+16]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
bz,pn %icc,.Linnerskip
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
std $nlod,[%sp+$bias+$frame+24]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
ba .Linner
nop
.align 32
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
.Linner:
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldd [$ap_l+$j],$alo ! load a[j] in double format
ldd [$ap_h+$j],$ahi
ldd [$np_l+$j],$nlo ! load n[j] in double format
ldd [$np_h+$j],$nhi
fmuld $alo,$ba,$aloa
fmuld $nlo,$na,$nloa
fmuld $alo,$bb,$alob
fmuld $nlo,$nb,$nlob
fmuld $alo,$bc,$aloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+0],%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $aloa,$nloa,$nloa
fmuld $nlo,$nc,$nloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+8],%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $alo,$bd,$alod
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+16],%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $alob,$nlob,$nlob
fmuld $nlo,$nd,$nlod
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ldx [%sp+$bias+$frame+24],%o3
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$ba,$ahia
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o0,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $aloc,$nloc,$nloc
fmuld $nhi,$na,$nhia
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %o7,%o1,%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$bb,$ahib
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o1,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $alod,$nlod,$nlod
fmuld $nhi,$nb,$nhib
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %o7,%o2,%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$bc,$ahic
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o2,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahia,$nhia,$nhia
fmuld $nhi,$nc,$nhic
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15]
and %o0,$mask,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fmuld $ahi,$bd,$ahid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
and %o1,$mask,%o1
and %o2,$mask,%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahib,$nhib,$nhib
fmuld $nhi,$nd,$nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
sllx %o1,16,%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $dota,$nloa,$nloa
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
sllx %o2,32,%o2
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $dotb,$nlob,$nlob
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
sllx %o3,48,%o7
or %o1,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahic,$nhic,$dota ! $nhic
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
or %o2,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $ahid,$nhid,$dotb ! $nhid
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
or %o7,%o0,%o0 ! 64-bit result
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $nloc,$nhia,$nloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
addcc %g1,%o0,%o0
Minor optimizations based on intruction level profiler feedback.
2006-11-28 18:34:51 +08:00
ldx [$tp+8],%o7 ! tp[j]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
faddd $nlod,$nhib,$nlod
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
srlx %o3,16,%g1 ! 34-bit carry
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fdtox $nloa,$nloa
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
bcs,a %xcc,.+8
add %g1,1,%g1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fdtox $nlob,$nlob
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
addcc %o7,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fdtox $nloc,$nloc
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
bcs,a %xcc,.+8
add %g1,1,%g1
stx %o0,[$tp] ! tp[j-1]
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fdtox $nlod,$nlod
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
std $nloa,[%sp+$bias+$frame+0]
std $nlob,[%sp+$bias+$frame+8]
std $nloc,[%sp+$bias+$frame+16]
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
addcc $j,8,$j
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
std $nlod,[%sp+$bias+$frame+24]
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
bnz,pt %icc,.Linner
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
add $tp,8,$tp
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
.Linnerskip:
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
fdtox $dota,$dota
fdtox $dotb,$dotb
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
ldx [%sp+$bias+$frame+0],%o0
ldx [%sp+$bias+$frame+8],%o1
ldx [%sp+$bias+$frame+16],%o2
ldx [%sp+$bias+$frame+24],%o3
srlx %o0,16,%o7
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
std $dota,[%sp+$bias+$frame+32]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
add %o7,%o1,%o1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
std $dotb,[%sp+$bias+$frame+40]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
srlx %o1,16,%o7
add %o7,%o2,%o2
srlx %o2,16,%o7
add %o7,%o3,%o3 ! %o3.%o2[0..15].%o1[0..15].%o0[0..15]
and %o0,$mask,%o0
and %o1,$mask,%o1
and %o2,$mask,%o2
sllx %o1,16,%o1
sllx %o2,32,%o2
sllx %o3,48,%o7
or %o1,%o0,%o0
or %o2,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+32],%o4
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
or %o7,%o0,%o0 ! 64-bit result
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [%sp+$bias+$frame+40],%o5
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
addcc %g1,%o0,%o0
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
ldx [$tp+8],%o7 ! tp[j]
This is "informational" commit. Its mere purpose is to expose "modulo factor" in inner loops.
2006-11-28 15:20:36 +08:00
srlx %o3,16,%g1 ! 34-bit carry
bcs,a %xcc,.+8
add %g1,1,%g1
addcc %o7,%o0,%o0
bcs,a %xcc,.+8
add %g1,1,%g1
stx %o0,[$tp] ! tp[j-1]
add $tp,8,$tp
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
srlx %o4,16,%o7
add %o7,%o5,%o5
and %o4,$mask,%o4
sllx %o5,16,%o7
or %o7,%o4,%o4
addcc %g1,%o4,%o4
srlx %o5,48,%g1
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
bcs,a %xcc,.+8
add %g1,1,%g1
Modulo-schedule loops in sparcv9a-mont.pl. Overall improvement factor over 0.9.8 is up to 3x on USI&II cores and up to 80% - on USIII&IV.
2006-11-28 15:24:26 +08:00
addcc $carry,%o4,%o4
stx %o4,[$tp] ! tp[num-1]
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
mov %g1,$carry
bcs,a %xcc,.+8
add $carry,1,$carry
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
addcc $i,8,$i
bnz %icc,.Louter
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
nop
Eliminate conditional final subtraction in Montgomery assembler modules.
2007-06-18 01:10:03 +08:00
add $tp,8,$tp ! adjust tp to point at the end
orn %g0,%g0,%g4
sub %g0,$num,%o7 ! n=-num
SPARC Solaris and Linux assemblers treat .align directive differently. PR: 1547
2007-06-20 20:24:22 +08:00
ba .Lsub
Latest bn_mont.c modification broke ECDSA test. I've got math wrong, which is fixed now.
2007-06-29 21:10:19 +08:00
subcc %g0,%g0,%g0 ! clear %icc.c
SPARC Solaris and Linux assemblers treat .align directive differently. PR: 1547
2007-06-20 20:24:22 +08:00
.align 32
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
.Lsub:
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
ldx [$tp+%o7],%o0
add $np,%o7,%g1
ld [%g1+0],%o2
ld [%g1+4],%o3
srlx %o0,32,%o1
subccc %o0,%o2,%o2
add $rp,%o7,%g1
subccc %o1,%o3,%o3
st %o2,[%g1+0]
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
add %o7,8,%o7
brnz,pt %o7,.Lsub
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
st %o3,[%g1+4]
Eliminate conditional final subtraction in Montgomery assembler modules.
2007-06-18 01:10:03 +08:00
subc $carry,0,%g4
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
sub %g0,$num,%o7 ! n=-num
SPARC Solaris and Linux assemblers treat .align directive differently. PR: 1547
2007-06-20 20:24:22 +08:00
ba .Lcopy
nop
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
SPARC Solaris and Linux assemblers treat .align directive differently. PR: 1547
2007-06-20 20:24:22 +08:00
.align 32
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
.Lcopy:
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
ldx [$tp+%o7],%o0
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
add $rp,%o7,%g1
Eliminate conditional final subtraction in Montgomery assembler modules.
2007-06-18 01:10:03 +08:00
ld [%g1+0],%o2
ld [%g1+4],%o3
stx %g0,[$tp+%o7]
and %o0,%g4,%o0
srlx %o0,32,%o1
andn %o2,%g4,%o2
andn %o3,%g4,%o3
or %o2,%o0,%o0
or %o3,%o1,%o1
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
st %o0,[%g1+0]
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
add %o7,8,%o7
brnz,pt %o7,.Lcopy
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
st %o1,[%g1+4]
Switch 64-bit sparcv9 platforms from bn(64,64) to bn(64,32). This doesn't have impact on performance, because amount of multiplications does not increase with this switch, not on sparcv9 that is. On the contrary, it actually improves performance, because it spares a load of instructions used to chase carries. Not to mention that BN assembler modules can be shared more freely between 32- and 64-bit builts.
2005-12-16 06:40:58 +08:00
sub %g0,$num,%o7 ! n=-num
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
.Lzap:
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
stx %g0,[$ap_l+%o7]
stx %g0,[$ap_h+%o7]
stx %g0,[$np_l+%o7]
stx %g0,[$np_h+%o7]
add %o7,8,%o7
brnz,pt %o7,.Lzap
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
nop
ldx [%sp+$bias+$frame+48],%o7
wr %g0,%o7,%asi ! restore %asi
mov 1,%i0
Add support for 32-bit ABI to sparcv9a-mont.pl module.
2005-10-23 02:16:09 +08:00
.Lret:
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
ret
restore
.type $fname,#function
.size $fname,(.-$fname)
Eliminate 64-bit alignment limitation in sparcv9a-mont.
2006-12-08 23:18:41 +08:00
.asciz "Montgomery Multipltication for UltraSPARC, CRYPTOGAMS by <appro\@openssl.org>"
SPARC Solaris and Linux assemblers treat .align directive differently. PR: 1547
2007-06-20 20:24:22 +08:00
.align 32
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
___
$code =~ s/\`([^\`]*)\`/eval($1)/gem;
As SPARCV9 CPU flavor is [expected to be] detected at run-time, we can afford to relax SPARCV9/8+ compiler command line and produce "unversal" binaries as we used to.
2005-12-19 17:10:06 +08:00
# Below substitution makes it possible to compile without demanding
# VIS extentions on command line, e.g. -xarch=v9 vs. -xarch=v9a. I
# dare to do this, because VIS capability is detected at run-time now
# and this routine is not called on CPU not capable to execute it. Do
# note that fzeros is not the only VIS dependency! Another dependency
# is implicit and is just _a_ numerical value loaded to %asi register,
# which assembler can't recognize as VIS specific...
$code =~ s/fzeros\s+%f([0-9]+)/
sprintf(".word\t0x%x\t! fzeros %%f%d",0x81b00c20|($1<<25),$1)
/gem;
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
print $code;
As SPARCV9 CPU flavor is [expected to be] detected at run-time, we can afford to relax SPARCV9/8+ compiler command line and produce "unversal" binaries as we used to.
2005-12-19 17:10:06 +08:00
# flush
Yet another "teaser" Montgomery multiply module, for UltraSPARC. It's not integrated yet, but it's tested and benchmarked [see commentary section for further details].
2005-10-19 15:12:06 +08:00
close STDOUT;
Reference in New Issue Copy Permalink