Unobtrusive backport of 32-bit x86 Montgomery improvements from 0.9.9-dev:

you need to use "enable-montasm" to see a difference.  (Huge speed
advantage, but BN_MONT_CTX is not binary compatible, so this can't be
enabled by default in the 0.9.8 branch.)

The CHANGES entry also covers the 64-bit x86 backport in November 2007
by appro.

5 files changed, 815 insertions, 0 deletions

diff --git a/crypto/bn/.cvsignore b/crypto/bn/.cvsignore
index 57df22cf65..c2f3bc0856 100644
--- a/crypto/bn/.cvsignore
+++ b/crypto/bn/.cvsignore
@@ -4,3 +4,4 @@ Makefile.save
 semantic.cache
 co86-elf.s
 bn86-elf.s
+mo86-elf.s
diff --git a/crypto/bn/Makefile b/crypto/bn/Makefile
index 6dfd528d52..e97c751390 100644
--- a/crypto/bn/Makefile
+++ b/crypto/bn/Makefile
@@ -67,16 +67,22 @@ bn86-elf.s:	asm/bn-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) bn-586.pl elf $(CFLAGS) > ../$@)
 co86-elf.s:	asm/co-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) co-586.pl elf $(CFLAGS) > ../$@)
+mo86-elf.s:	asm/mo-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) mo-586.pl elf $(CFLAGS) > ../$@)
 # COFF
 bn86-cof.s: asm/bn-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) bn-586.pl coff $(CFLAGS) > ../$@)
 co86-cof.s: asm/co-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) co-586.pl coff $(CFLAGS) > ../$@)
+mo86-cof.s: asm/mo-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) mo-586.pl coff $(CFLAGS) > ../$@)
 # a.out
 bn86-out.s: asm/bn-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) bn-586.pl a.out $(CFLAGS) > ../$@)
 co86-out.s: asm/co-586.pl ../perlasm/x86asm.pl
 	(cd asm; $(PERL) co-586.pl a.out $(CFLAGS) > ../$@)
+mo86-out.s: asm/mo-586.pl ../perlasm/x86asm.pl
+	(cd asm; $(PERL) mo-586.pl a.out $(CFLAGS) > ../$@)
 
 sparcv8.o:	asm/sparcv8.S
 	$(CC) $(CFLAGS) -c asm/sparcv8.S
diff --git a/crypto/bn/asm/mo-586.pl b/crypto/bn/asm/mo-586.pl
new file mode 100644
index 0000000000..0982293094
--- /dev/null
+++ b/crypto/bn/asm/mo-586.pl
@@ -0,0 +1,603 @@
+#!/usr/bin/env perl
+
+# This is crypto/bn/asm/x86-mont.pl (with asciz from crypto/perlasm/x86asm.pl)
+# from OpenSSL 0.9.9-dev 
+
+sub ::asciz
+{ my @str=unpack("C*",shift);
+    push @str,0;
+    while ($#str>15) {
+	&data_byte(@str[0..15]);
+	foreach (0..15) { shift @str; }
+    }
+    &data_byte(@str) if (@str);
+}
+
+# ====================================================================
+# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# 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/.
+# ====================================================================
+
+# October 2005
+#
+# This is a "teaser" code, as it can be improved in several ways...
+# First of all non-SSE2 path should be implemented (yes, for now it
+# performs Montgomery multiplication/convolution only on SSE2-capable
+# CPUs such as P4, others fall down to original code). Then inner loop
+# can be unrolled and modulo-scheduled to improve ILP and possibly
+# moved to 128-bit XMM register bank (though it would require input
+# rearrangement and/or increase bus bandwidth utilization). Dedicated
+# squaring procedure should give further performance improvement...
+# Yet, for being draft, the code improves rsa512 *sign* benchmark by
+# 110%(!), rsa1024 one - by 70% and rsa4096 - by 20%:-)
+
+# December 2006
+#
+# Modulo-scheduling SSE2 loops results in further 15-20% improvement.
+# Integer-only code [being equipped with dedicated squaring procedure]
+# gives ~40% on rsa512 sign benchmark...
+
+push(@INC,"perlasm","../../perlasm");
+require "x86asm.pl";
+
+&asm_init($ARGV[0],$0);
+
+$sse2=0;
+for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
+
+&external_label("OPENSSL_ia32cap_P") if ($sse2);
+
+&function_begin("bn_mul_mont");
+
+$i="edx";
+$j="ecx";
+$ap="esi";	$tp="esi";		# overlapping variables!!!
+$rp="edi";	$bp="edi";		# overlapping variables!!!
+$np="ebp";
+$num="ebx";
+
+$_num=&DWP(4*0,"esp");			# stack top layout
+$_rp=&DWP(4*1,"esp");
+$_ap=&DWP(4*2,"esp");
+$_bp=&DWP(4*3,"esp");
+$_np=&DWP(4*4,"esp");
+$_n0=&DWP(4*5,"esp");	$_n0q=&QWP(4*5,"esp");
+$_sp=&DWP(4*6,"esp");
+$_bpend=&DWP(4*7,"esp");
+$frame=32;				# size of above frame rounded up to 16n
+
+	&xor	("eax","eax");
+	&mov	("edi",&wparam(5));	# int num
+	&cmp	("edi",4);
+	&jl	(&label("just_leave"));
+
+	&lea	("esi",&wparam(0));	# put aside pointer to argument block
+	&lea	("edx",&wparam(1));	# load ap
+	&mov	("ebp","esp");		# saved stack pointer!
+	&add	("edi",2);		# extra two words on top of tp
+	&neg	("edi");
+	&lea	("esp",&DWP(-$frame,"esp","edi",4));	# alloca($frame+4*(num+2))
+	&neg	("edi");
+
+	# minimize cache contention by arraning 2K window between stack
+	# pointer and ap argument [np is also position sensitive vector,
+	# but it's assumed to be near ap, as it's allocated at ~same
+	# time].
+	&mov	("eax","esp");
+	&sub	("eax","edx");
+	&and	("eax",2047);
+	&sub	("esp","eax");		# this aligns sp and ap modulo 2048
+
+	&xor	("edx","esp");
+	&and	("edx",2048);
+	&xor	("edx",2048);
+	&sub	("esp","edx");		# this splits them apart modulo 4096
+
+	&and	("esp",-64);		# align to cache line
+
+	################################# load argument block...
+	&mov	("eax",&DWP(0*4,"esi"));# BN_ULONG *rp
+	&mov	("ebx",&DWP(1*4,"esi"));# const BN_ULONG *ap
+	&mov	("ecx",&DWP(2*4,"esi"));# const BN_ULONG *bp
+	&mov	("edx",&DWP(3*4,"esi"));# const BN_ULONG *np
+	&mov	("esi",&DWP(4*4,"esi"));# const BN_ULONG *n0
+	#&mov	("edi",&DWP(5*4,"esi"));# int num
+
+	&mov	("esi",&DWP(0,"esi"));	# pull n0[0]
+	&mov	($_rp,"eax");		# ... save a copy of argument block
+	&mov	($_ap,"ebx");
+	&mov	($_bp,"ecx");
+	&mov	($_np,"edx");
+	&mov	($_n0,"esi");
+	&lea	($num,&DWP(-3,"edi"));	# num=num-1 to assist modulo-scheduling
+	#&mov	($_num,$num);		# redundant as $num is not reused
+	&mov	($_sp,"ebp");		# saved stack pointer!
+
+if($sse2) {
+$acc0="mm0";	# mmx register bank layout
+$acc1="mm1";
+$car0="mm2";
+$car1="mm3";
+$mul0="mm4";
+$mul1="mm5";
+$temp="mm6";
+$mask="mm7";
+
+	&picmeup("eax","OPENSSL_ia32cap_P");
+	&bt	(&DWP(0,"eax"),26);
+	&jnc	(&label("non_sse2"));
+
+	&mov	("eax",-1);
+	&movd	($mask,"eax");		# mask 32 lower bits
+
+	&mov	($ap,$_ap);		# load input pointers
+	&mov	($bp,$_bp);
+	&mov	($np,$_np);
+
+	&xor	($i,$i);		# i=0
+	&xor	($j,$j);		# j=0
+
+	&movd	($mul0,&DWP(0,$bp));		# bp[0]
+	&movd	($mul1,&DWP(0,$ap));		# ap[0]
+	&movd	($car1,&DWP(0,$np));		# np[0]
+
+	&pmuludq($mul1,$mul0);			# ap[0]*bp[0]
+	&movq	($car0,$mul1);
+	&movq	($acc0,$mul1);			# I wish movd worked for
+	&pand	($acc0,$mask);			# inter-register transfers
+
+	&pmuludq($mul1,$_n0q);			# *=n0
+
+	&pmuludq($car1,$mul1);			# "t[0]"*np[0]*n0
+	&paddq	($car1,$acc0);
+
+	&movd	($acc1,&DWP(4,$np));		# np[1]
+	&movd	($acc0,&DWP(4,$ap));		# ap[1]
+
+	&psrlq	($car0,32);
+	&psrlq	($car1,32);
+
+	&inc	($j);				# j++
+&set_label("1st",16);
+	&pmuludq($acc0,$mul0);			# ap[j]*bp[0]
+	&pmuludq($acc1,$mul1);			# np[j]*m1
+	&paddq	($car0,$acc0);			# +=c0
+	&paddq	($car1,$acc1);			# +=c1
+
+	&movq	($acc0,$car0);
+	&pand	($acc0,$mask);
+	&movd	($acc1,&DWP(4,$np,$j,4));	# np[j+1]
+	&paddq	($car1,$acc0);			# +=ap[j]*bp[0];
+	&movd	($acc0,&DWP(4,$ap,$j,4));	# ap[j+1]
+	&psrlq	($car0,32);
+	&movd	(&DWP($frame-4,"esp",$j,4),$car1);	# tp[j-1]=
+	&psrlq	($car1,32);
+
+	&lea	($j,&DWP(1,$j));
+	&cmp	($j,$num);
+	&jl	(&label("1st"));
+
+	&pmuludq($acc0,$mul0);			# ap[num-1]*bp[0]
+	&pmuludq($acc1,$mul1);			# np[num-1]*m1
+	&paddq	($car0,$acc0);			# +=c0
+	&paddq	($car1,$acc1);			# +=c1
+
+	&movq	($acc0,$car0);
+	&pand	($acc0,$mask);
+	&paddq	($car1,$acc0);			# +=ap[num-1]*bp[0];
+	&movd	(&DWP($frame-4,"esp",$j,4),$car1);	# tp[num-2]=
+
+	&psrlq	($car0,32);
+	&psrlq	($car1,32);
+
+	&paddq	($car1,$car0);
+	&movq	(&QWP($frame,"esp",$num,4),$car1);	# tp[num].tp[num-1]
+
+	&inc	($i);				# i++
+&set_label("outer");
+	&xor	($j,$j);			# j=0
+
+	&movd	($mul0,&DWP(0,$bp,$i,4));	# bp[i]
+	&movd	($mul1,&DWP(0,$ap));		# ap[0]
+	&movd	($temp,&DWP($frame,"esp"));	# tp[0]
+	&movd	($car1,&DWP(0,$np));		# np[0]
+	&pmuludq($mul1,$mul0);			# ap[0]*bp[i]
+
+	&paddq	($mul1,$temp);			# +=tp[0]
+	&movq	($acc0,$mul1);
+	&movq	($car0,$mul1);
+	&pand	($acc0,$mask);
+
+	&pmuludq($mul1,$_n0q);			# *=n0
+
+	&pmuludq($car1,$mul1);
+	&paddq	($car1,$acc0);
+
+	&movd	($temp,&DWP($frame+4,"esp"));	# tp[1]
+	&movd	($acc1,&DWP(4,$np));		# np[1]
+	&movd	($acc0,&DWP(4,$ap));		# ap[1]
+
+	&psrlq	($car0,32);
+	&psrlq	($car1,32);
+	&paddq	($car0,$temp);			# +=tp[1]
+
+	&inc	($j);				# j++
+	&dec	($num);
+&set_label("inner");
+	&pmuludq($acc0,$mul0);			# ap[j]*bp[i]
+	&pmuludq($acc1,$mul1);			# np[j]*m1
+	&paddq	($car0,$acc0);			# +=c0
+	&paddq	($car1,$acc1);			# +=c1
+
+	&movq	($acc0,$car0);
+	&movd	($temp,&DWP($frame+4,"esp",$j,4));# tp[j+1]
+	&pand	($acc0,$mask);
+	&movd	($acc1,&DWP(4,$np,$j,4));	# np[j+1]
+	&paddq	($car1,$acc0);			# +=ap[j]*bp[i]+tp[j]
+	&movd	($acc0,&DWP(4,$ap,$j,4));	# ap[j+1]
+	&psrlq	($car0,32);
+	&movd	(&DWP($frame-4,"esp",$j,4),$car1);# tp[j-1]=
+	&psrlq	($car1,32);
+	&paddq	($car0,$temp);			# +=tp[j+1]
+
+	&dec	($num);
+	&lea	($j,&DWP(1,$j));		# j++
+	&jnz	(&label("inner"));
+
+	&mov	($num,$j);
+	&pmuludq($acc0,$mul0);			# ap[num-1]*bp[i]
+	&pmuludq($acc1,$mul1);			# np[num-1]*m1
+	&paddq	($car0,$acc0);			# +=c0
+	&paddq	($car1,$acc1);			# +=c1
+
+	&movq	($acc0,$car0);
+	&pand	($acc0,$mask);
+	&paddq	($car1,$acc0);			# +=ap[num-1]*bp[i]+tp[num-1]
+	&movd	(&DWP($frame-4,"esp",$j,4),$car1);	# tp[num-2]=
+	&psrlq	($car0,32);
+	&psrlq	($car1,32);
+
+	&movd	($temp,&DWP($frame+4,"esp",$num,4));	# += tp[num]
+	&paddq	($car1,$car0);
+	&paddq	($car1,$temp);
+	&movq	(&QWP($frame,"esp",$num,4),$car1);	# tp[num].tp[num-1]
+
+	&lea	($i,&DWP(1,$i));		# i++
+	&cmp	($i,$num);
+	&jle	(&label("outer"));
+
+	&emms	();				# done with mmx bank
+	&jmp	(&label("common_tail"));
+
+&set_label("non_sse2",16);
+}
+
+if (0) {
+	&mov	("esp",$_sp);
+	&xor	("eax","eax");	# signal "not fast enough [yet]"
+	&jmp	(&label("just_leave"));
+	# While the below code provides competitive performance for
+	# all key lengthes on modern Intel cores, it's still more
+	# than 10% slower for 4096-bit key elsewhere:-( "Competitive"
+	# means compared to the original integer-only assembler.
+	# 512-bit RSA sign is better by ~40%, but that's about all
+	# one can say about all CPUs...
+} else {
+$inp="esi";	# integer path uses these registers differently
+$word="edi";
+$carry="ebp";
+
+	&mov	($inp,$_ap);
+	&lea	($carry,&DWP(1,$num));
+	&mov	($word,$_bp);
+	&xor	($j,$j);				# j=0
+	&mov	("edx",$inp);
+	&and	($carry,1);				# see if num is even
+	&sub	("edx",$word);				# see if ap==bp
+	&lea	("eax",&DWP(4,$word,$num,4));		# &bp[num]
+	&or	($carry,"edx");
+	&mov	($word,&DWP(0,$word));			# bp[0]
+	&jz	(&label("bn_sqr_mont"));
+	&mov	($_bpend,"eax");
+	&mov	("eax",&DWP(0,$inp));
+	&xor	("edx","edx");
+
+&set_label("mull",16);
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[j]*bp[0]
+	&add	($carry,"eax");
+	&lea	($j,&DWP(1,$j));
+	&adc	("edx",0);
+	&mov	("eax",&DWP(0,$inp,$j,4));		# ap[j+1]
+	&cmp	($j,$num);
+	&mov	(&DWP($frame-4,"esp",$j,4),$carry);	# tp[j]=
+	&jl	(&label("mull"));
+
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[num-1]*bp[0]
+	 &mov	($word,$_n0);
+	&add	("eax",$carry);
+	 &mov	($inp,$_np);
+	&adc	("edx",0);
+	 &imul	($word,&DWP($frame,"esp"));		# n0*tp[0]
+
+	&mov	(&DWP($frame,"esp",$num,4),"eax");	# tp[num-1]=
+	&xor	($j,$j);
+	&mov	(&DWP($frame+4,"esp",$num,4),"edx");	# tp[num]=
+	&mov	(&DWP($frame+8,"esp",$num,4),$j);	# tp[num+1]=
+
+	&mov	("eax",&DWP(0,$inp));			# np[0]
+	&mul	($word);				# np[0]*m
+	&add	("eax",&DWP($frame,"esp"));		# +=tp[0]
+	&mov	("eax",&DWP(4,$inp));			# np[1]
+	&adc	("edx",0);
+	&inc	($j);
+
+	&jmp	(&label("2ndmadd"));
+
+&set_label("1stmadd",16);
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[j]*bp[i]
+	&add	($carry,&DWP($frame,"esp",$j,4));	# +=tp[j]
+	&lea	($j,&DWP(1,$j));
+	&adc	("edx",0);
+	&add	($carry,"eax");
+	&mov	("eax",&DWP(0,$inp,$j,4));		# ap[j+1]
+	&adc	("edx",0);
+	&cmp	($j,$num);
+	&mov	(&DWP($frame-4,"esp",$j,4),$carry);	# tp[j]=
+	&jl	(&label("1stmadd"));
+
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[num-1]*bp[i]
+	&add	("eax",&DWP($frame,"esp",$num,4));	# +=tp[num-1]
+	 &mov	($word,$_n0);
+	&adc	("edx",0);
+	 &mov	($inp,$_np);
+	&add	($carry,"eax");
+	&adc	("edx",0);
+	 &imul	($word,&DWP($frame,"esp"));		# n0*tp[0]
+
+	&xor	($j,$j);
+	&add	("edx",&DWP($frame+4,"esp",$num,4));	# carry+=tp[num]
+	&mov	(&DWP($frame,"esp",$num,4),$carry);	# tp[num-1]=
+	&adc	($j,0);
+	 &mov	("eax",&DWP(0,$inp));			# np[0]
+	&mov	(&DWP($frame+4,"esp",$num,4),"edx");	# tp[num]=
+	&mov	(&DWP($frame+8,"esp",$num,4),$j);	# tp[num+1]=
+
+	&mul	($word);				# np[0]*m
+	&add	("eax",&DWP($frame,"esp"));		# +=tp[0]
+	&mov	("eax",&DWP(4,$inp));			# np[1]
+	&adc	("edx",0);
+	&mov	($j,1);
+
+&set_label("2ndmadd",16);
+	&mov	($carry,"edx");
+	&mul	($word);				# np[j]*m
+	&add	($carry,&DWP($frame,"esp",$j,4));	# +=tp[j]
+	&lea	($j,&DWP(1,$j));
+	&adc	("edx",0);
+	&add	($carry,"eax");
+	&mov	("eax",&DWP(0,$inp,$j,4));		# np[j+1]
+	&adc	("edx",0);
+	&cmp	($j,$num);
+	&mov	(&DWP($frame-8,"esp",$j,4),$carry);	# tp[j-1]=
+	&jl	(&label("2ndmadd"));
+
+	&mov	($carry,"edx");
+	&mul	($word);				# np[j]*m
+	&add	($carry,&DWP($frame,"esp",$num,4));	# +=tp[num-1]
+	&adc	("edx",0);
+	&add	($carry,"eax");
+	&adc	("edx",0);
+	&mov	(&DWP($frame-4,"esp",$num,4),$carry);	# tp[num-2]=
+
+	&xor	("eax","eax");
+	 &mov	($j,$_bp);				# &bp[i]
+	&add	("edx",&DWP($frame+4,"esp",$num,4));	# carry+=tp[num]
+	&adc	("eax",&DWP($frame+8,"esp",$num,4));	# +=tp[num+1]
+	 &lea	($j,&DWP(4,$j));
+	&mov	(&DWP($frame,"esp",$num,4),"edx");	# tp[num-1]=
+	 &cmp	($j,$_bpend);
+	&mov	(&DWP($frame+4,"esp",$num,4),"eax");	# tp[num]=
+	&je	(&label("common_tail"));
+
+	&mov	($word,&DWP(0,$j));			# bp[i+1]
+	&mov	($inp,$_ap);
+	&mov	($_bp,$j);				# &bp[++i]
+	&xor	($j,$j);
+	&xor	("edx","edx");
+	&mov	("eax",&DWP(0,$inp));
+	&jmp	(&label("1stmadd"));
+
+&set_label("bn_sqr_mont",16);
+$sbit=$num;
+	&mov	($_num,$num);
+	&mov	($_bp,$j);				# i=0
+
+	&mov	("eax",$word);				# ap[0]
+	&mul	($word);				# ap[0]*ap[0]
+	&mov	(&DWP($frame,"esp"),"eax");		# tp[0]=
+	&mov	($sbit,"edx");
+	&shr	("edx",1);
+	&and	($sbit,1);
+	&inc	($j);
+&set_label("sqr",16);
+	&mov	("eax",&DWP(0,$inp,$j,4));		# ap[j]
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[j]*ap[0]
+	&add	("eax",$carry);
+	&lea	($j,&DWP(1,$j));
+	&adc	("edx",0);
+	&lea	($carry,&DWP(0,$sbit,"eax",2));
+	&shr	("eax",31);
+	&cmp	($j,$_num);
+	&mov	($sbit,"eax");
+	&mov	(&DWP($frame-4,"esp",$j,4),$carry);	# tp[j]=
+	&jl	(&label("sqr"));
+
+	&mov	("eax",&DWP(0,$inp,$j,4));		# ap[num-1]
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[num-1]*ap[0]
+	&add	("eax",$carry);
+	 &mov	($word,$_n0);
+	&adc	("edx",0);
+	 &mov	($inp,$_np);
+	&lea	($carry,&DWP(0,$sbit,"eax",2));
+	 &imul	($word,&DWP($frame,"esp"));		# n0*tp[0]
+	&shr	("eax",31);
+	&mov	(&DWP($frame,"esp",$j,4),$carry);	# tp[num-1]=
+
+	&lea	($carry,&DWP(0,"eax","edx",2));
+	 &mov	("eax",&DWP(0,$inp));			# np[0]
+	&shr	("edx",31);
+	&mov	(&DWP($frame+4,"esp",$j,4),$carry);	# tp[num]=
+	&mov	(&DWP($frame+8,"esp",$j,4),"edx");	# tp[num+1]=
+
+	&mul	($word);				# np[0]*m
+	&add	("eax",&DWP($frame,"esp"));		# +=tp[0]
+	&mov	($num,$j);
+	&adc	("edx",0);
+	&mov	("eax",&DWP(4,$inp));			# np[1]
+	&mov	($j,1);
+
+&set_label("3rdmadd",16);
+	&mov	($carry,"edx");
+	&mul	($word);				# np[j]*m
+	&add	($carry,&DWP($frame,"esp",$j,4));	# +=tp[j]
+	&adc	("edx",0);
+	&add	($carry,"eax");
+	&mov	("eax",&DWP(4,$inp,$j,4));		# np[j+1]
+	&adc	("edx",0);
+	&mov	(&DWP($frame-4,"esp",$j,4),$carry);	# tp[j-1]=
+
+	&mov	($carry,"edx");
+	&mul	($word);				# np[j+1]*m
+	&add	($carry,&DWP($frame+4,"esp",$j,4));	# +=tp[j+1]
+	&lea	($j,&DWP(2,$j));
+	&adc	("edx",0);
+	&add	($carry,"eax");
+	&mov	("eax",&DWP(0,$inp,$j,4));		# np[j+2]
+	&adc	("edx",0);
+	&cmp	($j,$num);
+	&mov	(&DWP($frame-8,"esp",$j,4),$carry);	# tp[j]=
+	&jl	(&label("3rdmadd"));
+
+	&mov	($carry,"edx");
+	&mul	($word);				# np[j]*m
+	&add	($carry,&DWP($frame,"esp",$num,4));	# +=tp[num-1]
+	&adc	("edx",0);
+	&add	($carry,"eax");
+	&adc	("edx",0);
+	&mov	(&DWP($frame-4,"esp",$num,4),$carry);	# tp[num-2]=
+
+	&mov	($j,$_bp);				# i
+	&xor	("eax","eax");
+	&mov	($inp,$_ap);
+	&add	("edx",&DWP($frame+4,"esp",$num,4));	# carry+=tp[num]
+	&adc	("eax",&DWP($frame+8,"esp",$num,4));	# +=tp[num+1]
+	&mov	(&DWP($frame,"esp",$num,4),"edx");	# tp[num-1]=
+	&cmp	($j,$num);
+	&mov	(&DWP($frame+4,"esp",$num,4),"eax");	# tp[num]=
+	&je	(&label("common_tail"));
+
+	&mov	($word,&DWP(4,$inp,$j,4));		# ap[i]
+	&lea	($j,&DWP(1,$j));
+	&mov	("eax",$word);
+	&mov	($_bp,$j);				# ++i
+	&mul	($word);				# ap[i]*ap[i]
+	&add	("eax",&DWP($frame,"esp",$j,4));	# +=tp[i]
+	&adc	("edx",0);
+	&mov	(&DWP($frame,"esp",$j,4),"eax");	# tp[i]=
+	&xor	($carry,$carry);
+	&cmp	($j,$num);
+	&lea	($j,&DWP(1,$j));
+	&je	(&label("sqrlast"));
+
+	&mov	($sbit,"edx");				# zaps $num
+	&shr	("edx",1);
+	&and	($sbit,1);
+&set_label("sqradd",16);
+	&mov	("eax",&DWP(0,$inp,$j,4));		# ap[j]
+	&mov	($carry,"edx");
+	&mul	($word);				# ap[j]*ap[i]
+	&add	("eax",$carry);
+	&lea	($carry,&DWP(0,"eax","eax"));
+	&adc	("edx",0);
+	&shr	("eax",31);
+	&add	($carry,&DWP($frame,"esp",$j,4));	# +=tp[j]
+	&lea	($j,&DWP(1,$j));
+	&adc	("eax",0);
+	&add	($carry,$sbit);
+	&adc	("eax",0);
+	&cmp	($j,$_num);
+	&mov	(&DWP($frame-4,"esp",$j,4),$carry);	# tp[j]=
+	&mov	($sbit,"eax");
+	&jle	(&label("sqradd"));
+
+	&mov	($carry,"edx");
+	&lea	("edx",&DWP(0,$sbit,"edx",2));
+	&shr	($carry,31);
+&set_label("sqrlast");
+	&mov	($word,$_n0);
+	&mov	($inp,$_np);
+	&imul	($word,&DWP($frame,"esp"));		# n0*tp[0]
+
+	&add	("edx",&DWP($frame,"esp",$j,4));	# +=tp[num]
+	&mov	("eax",&DWP(0,$inp));			# np[0]
+	&adc	($carry,0);
+	&mov	(&DWP($frame,"esp",$j,4),"edx");	# tp[num]=
+	&mov	(&DWP($frame+4,"esp",$j,4),$carry);	# tp[num+1]=
+
+	&mul	($word);				# np[0]*m
+	&add	("eax",&DWP($frame,"esp"));		# +=tp[0]
+	&lea	($num,&DWP(-1,$j));
+	&adc	("edx",0);
+	&mov	($j,1);
+	&mov	("eax",&DWP(4,$inp));			# np[1]
+
+	&jmp	(&label("3rdmadd"));
+}
+
+&set_label("common_tail",16);
+	&mov	($np,$_np);			# load modulus pointer
+	&mov	($rp,$_rp);			# load result pointer
+	&lea	($tp,&DWP($frame,"esp"));	# [$ap and $bp are zapped]
+
+	&mov	("eax",&DWP(0,$tp));		# tp[0]
+	&mov	($j,$num);			# j=num-1
+	&xor	($i,$i);			# i=0 and clear CF!
+
+&set_label("sub",16);
+	&sbb	("eax",&DWP(0,$np,$i,4));
+	&mov	(&DWP(0,$rp,$i,4),"eax");	# rp[i]=tp[i]-np[i]
+	&dec	($j);				# doesn't affect CF!
+	&mov	("eax",&DWP(4,$tp,$i,4));	# tp[i+1]
+	&lea	($i,&DWP(1,$i));		# i++
+	&jge	(&label("sub"));
+
+	&sbb	("eax",0);			# handle upmost overflow bit
+	&and	($tp,"eax");
+	&not	("eax");
+	&mov	($np,$rp);
+	&and	($np,"eax");
+	&or	($tp,$np);			# tp=carry?tp:rp
+
+&set_label("copy",16);				# copy or in-place refresh
+	&mov	("eax",&DWP(0,$tp,$num,4));
+	&mov	(&DWP(0,$rp,$num,4),"eax");	# rp[i]=tp[i]
+	&mov	(&DWP($frame,"esp",$num,4),$j);	# zap temporary vector
+	&dec	($num);
+	&jge	(&label("copy"));
+
+	&mov	("esp",$_sp);		# pull saved stack pointer
+	&mov	("eax",1);
+&set_label("just_leave");
+&function_end("bn_mul_mont");
+
+&asciz("Montgomery Multiplication for x86, CRYPTOGAMS by <appro\@openssl.org>");
+
+&asm_finish();
diff --git a/crypto/bn/bn.h b/crypto/bn/bn.h
index df6eea29a7..de39a72074 100644
--- a/crypto/bn/bn.h
+++ b/crypto/bn/bn.h
@@ -303,7 +303,14 @@ struct bn_mont_ctx_st
 	BIGNUM N;      /* The modulus */
 	BIGNUM Ni;     /* R*(1/R mod N) - N*Ni = 1
 	                * (Ni is only stored for bignum algorithm) */
+#if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
+	/* Non-default compile (can only happen with "enable-montasm"),
+	 * uses the new type from 0.9.9 to accomodate two words: */
+	BN_ULONG n0[2];/* least significant word(s) of Ni */
+#else
+	/* By default, use old type: */
 	BN_ULONG n0;   /* least significant word of Ni */
+#endif
 	int flags;
 	};
 
diff --git a/crypto/bn/bn_mont.c b/crypto/bn/bn_mont.c
index 23e4ba5140..e17c697e39 100644
--- a/crypto/bn/bn_mont.c
+++ b/crypto/bn/bn_mont.c
@@ -122,6 +122,12 @@
 
 #define MONT_WORD /* use the faster word-based algorithm */
 
+#if defined(MONT_WORD) && defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
+static int BN_from_montgomery_word(BIGNUM *ret, BIGNUM *r, BN_MONT_CTX *mont);
+#endif
+
+
+
 int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
 			  BN_MONT_CTX *mont, BN_CTX *ctx)
 	{
@@ -133,7 +139,11 @@ int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
 	if (num>1 && a->top==num && b->top==num)
 		{
 		if (bn_wexpand(r,num) == NULL) return(0);
+#if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32) /* non-default compile */
+		if (bn_mul_mont(r->d,a->d,b->d,mont->N.d,mont->n0,num))
+#else
 		if (bn_mul_mont(r->d,a->d,b->d,mont->N.d,&mont->n0,num))
+#endif
 			{
 			r->neg = a->neg^b->neg;
 			r->top = num;
@@ -157,7 +167,11 @@ int BN_mod_mul_montgomery(BIGNUM *r, const BIGNUM *a, const BIGNUM *b,
 		if (!BN_mul(tmp,a,b,ctx)) goto err;
 		}
 	/* reduce from aRR to aR */
+#if defined(MONT_WORD) && defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
+	if (!BN_from_montgomery_word(r,tmp,mont)) goto err;
+#else
 	if (!BN_from_montgomery(r,tmp,mont,ctx)) goto err;
+#endif
 	bn_check_top(r);
 	ret=1;
 err:
@@ -165,6 +179,145 @@ err:
 	return(ret);
 	}
 
+#if defined(MONT_WORD) && defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
+static int BN_from_montgomery_word(BIGNUM *ret, BIGNUM *r, BN_MONT_CTX *mont)
+	{
+	BIGNUM *n;
+	BN_ULONG *ap,*np,*rp,n0,v,*nrp;
+	int al,nl,max,i,x,ri;
+
+	n= &(mont->N);
+	/* mont->ri is the size of mont->N in bits (rounded up
+	   to the word size) */
+	al=ri=mont->ri/BN_BITS2;
+
+	nl=n->top;
+	if ((al == 0) || (nl == 0)) { ret->top=0; return(1); }
+
+	max=(nl+al+1); /* allow for overflow (no?) XXX */
+	if (bn_wexpand(r,max) == NULL) return(0);
+
+	r->neg^=n->neg;
+	np=n->d;
+	rp=r->d;
+	nrp= &(r->d[nl]);
+
+	/* clear the top words of T */
+	for (i=r->top; i<max; i++) /* memset? XXX */
+		r->d[i]=0;
+
+	r->top=max;
+	n0=mont->n0[0];
+
+#ifdef BN_COUNT
+	fprintf(stderr,"word BN_from_montgomery_word %d * %d\n",nl,nl);
+#endif
+	for (i=0; i<nl; i++)
+		{
+#ifdef __TANDEM
+                {
+                   long long t1;
+                   long long t2;
+                   long long t3;
+                   t1 = rp[0] * (n0 & 0177777);
+                   t2 = 037777600000l;
+                   t2 = n0 & t2;
+                   t3 = rp[0] & 0177777;
+                   t2 = (t3 * t2) & BN_MASK2;
+                   t1 = t1 + t2;
+                   v=bn_mul_add_words(rp,np,nl,(BN_ULONG) t1);
+                }
+#else
+		v=bn_mul_add_words(rp,np,nl,(rp[0]*n0)&BN_MASK2);
+#endif
+		nrp++;
+		rp++;
+		if (((nrp[-1]+=v)&BN_MASK2) >= v)
+			continue;
+		else
+			{
+			if (((++nrp[0])&BN_MASK2) != 0) continue;
+			if (((++nrp[1])&BN_MASK2) != 0) continue;
+			for (x=2; (((++nrp[x])&BN_MASK2) == 0); x++) ;
+			}
+		}
+	bn_correct_top(r);
+
+	/* mont->ri will be a multiple of the word size and below code
+	 * is kind of BN_rshift(ret,r,mont->ri) equivalent */
+	if (r->top <= ri)
+		{
+		ret->top=0;
+		return(1);
+		}
+	al=r->top-ri;
+
+	if (bn_wexpand(ret,ri) == NULL) return(0);
+	x=0-(((al-ri)>>(sizeof(al)*8-1))&1);
+	ret->top=x=(ri&~x)|(al&x);	/* min(ri,al) */
+	ret->neg=r->neg;
+
+	rp=ret->d;
+	ap=&(r->d[ri]);
+
+	{
+	size_t m1,m2;
+
+	v=bn_sub_words(rp,ap,np,ri);
+	/* this ----------------^^ works even in al<ri case
+	 * thanks to zealous zeroing of top of the vector in the
+	 * beginning. */
+
+	/* if (al==ri && !v) || al>ri) nrp=rp; else nrp=ap; */
+	/* in other words if subtraction result is real, then
+	 * trick unconditional memcpy below to perform in-place
+	 * "refresh" instead of actual copy. */
+	m1=0-(size_t)(((al-ri)>>(sizeof(al)*8-1))&1);	/* al<ri */
+	m2=0-(size_t)(((ri-al)>>(sizeof(al)*8-1))&1);	/* al>ri */
+	m1|=m2;			/* (al!=ri) */
+	m1|=(0-(size_t)v);	/* (al!=ri || v) */
+	m1&=~m2;		/* (al!=ri || v) && !al>ri */
+	nrp=(BN_ULONG *)(((size_t)rp&~m1)|((size_t)ap&m1));
+	}
+
+	/* 'i<ri' is chosen to eliminate dependency on input data, even
+	 * though it results in redundant copy in al<ri case. */
+	for (i=0,ri-=4; i<ri; i+=4)
+		{
+		BN_ULONG t1,t2,t3,t4;
+		
+		t1=nrp[i+0];
+		t2=nrp[i+1];
+		t3=nrp[i+2];	ap[i+0]=0;
+		t4=nrp[i+3];	ap[i+1]=0;
+		rp[i+0]=t1;	ap[i+2]=0;
+		rp[i+1]=t2;	ap[i+3]=0;
+		rp[i+2]=t3;
+		rp[i+3]=t4;
+		}
+	for (ri+=4; i<ri; i++)
+		rp[i]=nrp[i], ap[i]=0;
+	bn_correct_top(r);
+	bn_correct_top(ret);
+	bn_check_top(ret);
+
+	return(1);
+	}
+
+int BN_from_montgomery(BIGNUM *ret, const BIGNUM *a, BN_MONT_CTX *mont,
+	     BN_CTX *ctx)
+	{
+	int retn=0;
+	BIGNUM *t;
+
+	BN_CTX_start(ctx);
+	if ((t = BN_CTX_get(ctx)) && BN_copy(t,a))
+		retn = BN_from_montgomery_word(ret,t,mont);
+	BN_CTX_end(ctx);
+	return retn;
+	}
+#else
+
 int BN_from_montgomery(BIGNUM *ret, const BIGNUM *a, BN_MONT_CTX *mont,
 	     BN_CTX *ctx)
 	{
@@ -357,6 +510,7 @@ int BN_from_montgomery(BIGNUM *ret, const BIGNUM *a, BN_MONT_CTX *mont,
 	BN_CTX_end(ctx);
 	return(retn);
 	}
+#endif /* defined(MONT_WORD) && defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32) */
 
 BN_MONT_CTX *BN_MONT_CTX_new(void)
 	{
@@ -376,6 +530,11 @@ void BN_MONT_CTX_init(BN_MONT_CTX *ctx)
 	BN_init(&(ctx->RR));
 	BN_init(&(ctx->N));
 	BN_init(&(ctx->Ni));
+#if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32) /* non-default compile */
+	ctx->n0[0] = ctx->n0[1] = 0;
+#else
+	ctx->n0 = 0;
+#endif
 	ctx->flags=0;
 	}
 
@@ -409,7 +568,11 @@ int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx)
 
 		mont->ri=(BN_num_bits(mod)+(BN_BITS2-1))/BN_BITS2*BN_BITS2;
 		BN_zero(R);
+#if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32) /* non-default compile */
+		if (!(BN_set_bit(R,2*BN_BITS2))) goto err;	/* R */
+#else
 		if (!(BN_set_bit(R,BN_BITS2))) goto err;	/* R */
+#endif
 
 		buf[0]=mod->d[0]; /* tmod = N mod word size */
 		buf[1]=0;
@@ -419,6 +582,35 @@ int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx)
 		tmod.top = buf[0] != 0 ? 1 : 0;
 		tmod.dmax=2;
 		tmod.neg=0;
+
+#if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32)
+								tmod.top=0;
+		if ((buf[0] = mod->d[0]))			tmod.top=1;
+		if ((buf[1] = mod->top>1 ? mod->d[1] : 0))	tmod.top=2;
+
+		if ((BN_mod_inverse(Ri,R,&tmod,ctx)) == NULL)
+			goto err;
+		if (!BN_lshift(Ri,Ri,2*BN_BITS2)) goto err; /* R*Ri */
+		if (!BN_is_zero(Ri))
+			{
+			if (!BN_sub_word(Ri,1)) goto err;
+			}
+		else /* if N mod word size == 1 */
+			{
+			if (bn_expand(Ri,(int)sizeof(BN_ULONG)*2) == NULL)
+				goto err;
+			/* Ri-- (mod double word size) */
+			Ri->neg=0;
+			Ri->d[0]=BN_MASK2;
+			Ri->d[1]=BN_MASK2;
+			Ri->top=2;
+			}
+		if (!BN_div(Ri,NULL,Ri,&tmod,ctx)) goto err;
+		/* Ni = (R*Ri-1)/N,
+		 * keep only couple of least significant words: */
+		mont->n0[0] = (Ri->top > 0) ? Ri->d[0] : 0;
+		mont->n0[1] = (Ri->top > 1) ? Ri->d[1] : 0;
+#else
 							/* Ri = R^-1 mod N*/
 		if ((BN_mod_inverse(Ri,R,&tmod,ctx)) == NULL)
 			goto err;
@@ -435,6 +627,7 @@ int BN_MONT_CTX_set(BN_MONT_CTX *mont, const BIGNUM *mod, BN_CTX *ctx)
 		/* Ni = (R*Ri-1)/N,
 		 * keep only least significant word: */
 		mont->n0 = (Ri->top > 0) ? Ri->d[0] : 0;
+#endif
 		}
 #else /* !MONT_WORD */
 		{ /* bignum version */
@@ -470,7 +663,12 @@ BN_MONT_CTX *BN_MONT_CTX_copy(BN_MONT_CTX *to, BN_MONT_CTX *from)
 	if (!BN_copy(&(to->N),&(from->N))) return NULL;
 	if (!BN_copy(&(to->Ni),&(from->Ni))) return NULL;
 	to->ri=from->ri;
+#if defined(OPENSSL_BN_ASM_MONT) && (BN_BITS2<=32) /* non-default compile */
+	to->n0[0]=from->n0[0];
+	to->n0[1]=from->n0[1];
+#else
 	to->n0=from->n0;
+#endif
 	return(to);
 	}