#!/usr/bin/env perl
# ====================================================================
-# Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
+# Written by Andy Polyakov <appro@openssl.org> 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/.
# respectful 50%. It remains to be seen if loop unrolling and
# dedicated squaring routine can provide further improvement...
-$output=shift;
-open STDOUT,"| $^X ../perlasm/x86_64-xlate.pl $output";
+# July 2011.
+#
+# Add dedicated squaring procedure. Performance improvement varies
+# from platform to platform, but in average it's ~5%/15%/25%/33%
+# for 512-/1024-/2048-/4096-bit RSA *sign* benchmarks respectively.
+
+# August 2011.
+#
+# Unroll and modulo-schedule inner loops in such manner that they
+# are "fallen through" for input lengths of 8, which is critical for
+# 1024-bit RSA *sign*. Average performance improvement in comparison
+# to *initial* version of this module from 2005 is ~0%/30%/40%/45%
+# for 512-/1024-/2048-/4096-bit RSA *sign* benchmarks respectively.
+
+# June 2013.
+#
+# Optmize reduction in squaring procedure and improve 1024+-bit RSA
+# sign performance by 10-16% on Intel Sandy Bridge and later
+# (virtually same on non-Intel processors).
+
+$flavour = shift;
+$output = shift;
+if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
+
+$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
+
+$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
+( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
+( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
+die "can't locate x86_64-xlate.pl";
+
+open OUT,"| \"$^X\" $xlate $flavour $output";
+*STDOUT=*OUT;
# int bn_mul_mont(
$rp="%rdi"; # BN_ULONG *rp,
$num="%r9"; # int num);
$lo0="%r10";
$hi0="%r11";
-$bp="%r12"; # reassign $bp
$hi1="%r13";
$i="%r14";
$j="%r15";
.type bn_mul_mont,\@function,6
.align 16
bn_mul_mont:
+ test \$3,${num}d
+ jnz .Lmul_enter
+ cmp \$8,${num}d
+ jb .Lmul_enter
+ cmp $ap,$bp
+ jne .Lmul4x_enter
+ test \$7,${num}d
+ jz .Lsqr8x_enter
+ jmp .Lmul4x_enter
+
+.align 16
+.Lmul_enter:
push %rbx
push %rbp
push %r12
push %r14
push %r15
- lea 2($num),%rax
- mov %rsp,%rbp
- neg %rax
- lea (%rsp,%rax,8),%rsp # tp=alloca(8*(num+2))
+ mov ${num}d,${num}d
+ lea 2($num),%r10
+ mov %rsp,%r11
+ neg %r10
+ lea (%rsp,%r10,8),%rsp # tp=alloca(8*(num+2))
and \$-1024,%rsp # minimize TLB usage
- mov %rbp,8(%rsp,$num,8) # tp[num+1]=%rsp
- mov %rdx,$bp # $bp reassigned, remember?
+ mov %r11,8(%rsp,$num,8) # tp[num+1]=%rsp
+.Lmul_body:
+ mov $bp,%r12 # reassign $bp
+___
+ $bp="%r12";
+$code.=<<___;
mov ($n0),$n0 # pull n0[0] value
+ mov ($bp),$m0 # m0=bp[0]
+ mov ($ap),%rax
xor $i,$i # i=0
xor $j,$j # j=0
- mov ($bp),$m0 # m0=bp[0]
- mov ($ap),%rax
+ mov $n0,$m1
mulq $m0 # ap[0]*bp[0]
mov %rax,$lo0
- mov %rdx,$hi0
+ mov ($np),%rax
- imulq $n0,%rax # "tp[0]"*n0
- mov %rax,$m1
+ imulq $lo0,$m1 # "tp[0]"*n0
+ mov %rdx,$hi0
- mulq ($np) # np[0]*m1
- add $lo0,%rax # discarded
+ mulq $m1 # np[0]*m1
+ add %rax,$lo0 # discarded
+ mov 8($ap),%rax
adc \$0,%rdx
mov %rdx,$hi1
lea 1($j),$j # j++
+ jmp .L1st_enter
+
+.align 16
.L1st:
+ add %rax,$hi1
mov ($ap,$j,8),%rax
- mulq $m0 # ap[j]*bp[0]
- add $hi0,%rax
adc \$0,%rdx
- mov %rax,$lo0
+ add $hi0,$hi1 # np[j]*m1+ap[j]*bp[0]
+ mov $lo0,$hi0
+ adc \$0,%rdx
+ mov $hi1,-16(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$hi1
+
+.L1st_enter:
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$hi0
mov ($np,$j,8),%rax
- mov %rdx,$hi0
+ adc \$0,%rdx
+ lea 1($j),$j # j++
+ mov %rdx,$lo0
mulq $m1 # np[j]*m1
- add $hi1,%rax
- lea 1($j),$j # j++
+ cmp $num,$j
+ jne .L1st
+
+ add %rax,$hi1
+ mov ($ap),%rax # ap[0]
adc \$0,%rdx
- add $lo0,%rax # np[j]*m1+ap[j]*bp[0]
+ add $hi0,$hi1 # np[j]*m1+ap[j]*bp[0]
adc \$0,%rdx
- mov %rax,-16(%rsp,$j,8) # tp[j-1]
- cmp $num,$j
+ mov $hi1,-16(%rsp,$j,8) # tp[j-1]
mov %rdx,$hi1
- jl .L1st
+ mov $lo0,$hi0
xor %rdx,%rdx
add $hi0,$hi1
mov %rdx,(%rsp,$num,8) # store upmost overflow bit
lea 1($i),$i # i++
-.align 4
+ jmp .Louter
+.align 16
.Louter:
- xor $j,$j # j=0
-
mov ($bp,$i,8),$m0 # m0=bp[i]
- mov ($ap),%rax # ap[0]
+ xor $j,$j # j=0
+ mov $n0,$m1
+ mov (%rsp),$lo0
mulq $m0 # ap[0]*bp[i]
- add (%rsp),%rax # ap[0]*bp[i]+tp[0]
+ add %rax,$lo0 # ap[0]*bp[i]+tp[0]
+ mov ($np),%rax
adc \$0,%rdx
- mov %rax,$lo0
- mov %rdx,$hi0
- imulq $n0,%rax # tp[0]*n0
- mov %rax,$m1
+ imulq $lo0,$m1 # tp[0]*n0
+ mov %rdx,$hi0
- mulq ($np,$j,8) # np[0]*m1
- add $lo0,%rax # discarded
- mov 8(%rsp),$lo0 # tp[1]
+ mulq $m1 # np[0]*m1
+ add %rax,$lo0 # discarded
+ mov 8($ap),%rax
adc \$0,%rdx
+ mov 8(%rsp),$lo0 # tp[1]
mov %rdx,$hi1
lea 1($j),$j # j++
-.align 4
+ jmp .Linner_enter
+
+.align 16
.Linner:
+ add %rax,$hi1
mov ($ap,$j,8),%rax
- mulq $m0 # ap[j]*bp[i]
- add $hi0,%rax
adc \$0,%rdx
- add %rax,$lo0 # ap[j]*bp[i]+tp[j]
+ add $lo0,$hi1 # np[j]*m1+ap[j]*bp[i]+tp[j]
+ mov (%rsp,$j,8),$lo0
+ adc \$0,%rdx
+ mov $hi1,-16(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$hi1
+
+.Linner_enter:
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$hi0
mov ($np,$j,8),%rax
adc \$0,%rdx
+ add $hi0,$lo0 # ap[j]*bp[i]+tp[j]
mov %rdx,$hi0
+ adc \$0,$hi0
+ lea 1($j),$j # j++
mulq $m1 # np[j]*m1
- add $hi1,%rax
- lea 1($j),$j # j++
- adc \$0,%rdx
- add $lo0,%rax # np[j]*m1+ap[j]*bp[i]+tp[j]
+ cmp $num,$j
+ jne .Linner
+
+ add %rax,$hi1
+ mov ($ap),%rax # ap[0]
adc \$0,%rdx
+ add $lo0,$hi1 # np[j]*m1+ap[j]*bp[i]+tp[j]
mov (%rsp,$j,8),$lo0
- cmp $num,$j
- mov %rax,-16(%rsp,$j,8) # tp[j-1]
+ adc \$0,%rdx
+ mov $hi1,-16(%rsp,$j,8) # tp[j-1]
mov %rdx,$hi1
- jl .Linner
xor %rdx,%rdx
add $hi0,$hi1
cmp $num,$i
jl .Louter
+ xor $i,$i # i=0 and clear CF!
+ mov (%rsp),%rax # tp[0]
+ lea (%rsp),$ap # borrow ap for tp
+ mov $num,$j # j=num
+ jmp .Lsub
+.align 16
+.Lsub: sbb ($np,$i,8),%rax
+ mov %rax,($rp,$i,8) # rp[i]=tp[i]-np[i]
+ mov 8($ap,$i,8),%rax # tp[i+1]
+ lea 1($i),$i # i++
+ dec $j # doesnn't affect CF!
+ jnz .Lsub
+
+ sbb \$0,%rax # handle upmost overflow bit
+ xor $i,$i
+ and %rax,$ap
+ not %rax
+ mov $rp,$np
+ and %rax,$np
+ mov $num,$j # j=num
+ or $np,$ap # ap=borrow?tp:rp
+.align 16
+.Lcopy: # copy or in-place refresh
+ mov ($ap,$i,8),%rax
+ mov $i,(%rsp,$i,8) # zap temporary vector
+ mov %rax,($rp,$i,8) # rp[i]=tp[i]
+ lea 1($i),$i
+ sub \$1,$j
+ jnz .Lcopy
+
+ mov 8(%rsp,$num,8),%rsi # restore %rsp
+ mov \$1,%rax
+ mov (%rsi),%r15
+ mov 8(%rsi),%r14
+ mov 16(%rsi),%r13
+ mov 24(%rsi),%r12
+ mov 32(%rsi),%rbp
+ mov 40(%rsi),%rbx
+ lea 48(%rsi),%rsp
+.Lmul_epilogue:
+ ret
+.size bn_mul_mont,.-bn_mul_mont
+___
+{{{
+my @A=("%r10","%r11");
+my @N=("%r13","%rdi");
+$code.=<<___;
+.type bn_mul4x_mont,\@function,6
+.align 16
+bn_mul4x_mont:
+.Lmul4x_enter:
+ push %rbx
+ push %rbp
+ push %r12
+ push %r13
+ push %r14
+ push %r15
+
+ mov ${num}d,${num}d
+ lea 4($num),%r10
+ mov %rsp,%r11
+ neg %r10
+ lea (%rsp,%r10,8),%rsp # tp=alloca(8*(num+4))
+ and \$-1024,%rsp # minimize TLB usage
+
+ mov %r11,8(%rsp,$num,8) # tp[num+1]=%rsp
+.Lmul4x_body:
+ mov $rp,16(%rsp,$num,8) # tp[num+2]=$rp
+ mov %rdx,%r12 # reassign $bp
+___
+ $bp="%r12";
+$code.=<<___;
+ mov ($n0),$n0 # pull n0[0] value
+ mov ($bp),$m0 # m0=bp[0]
+ mov ($ap),%rax
+
xor $i,$i # i=0
- lea -1($num),$j # j=num-1
- cmp \$0,%rdx # %rdx still holds upmost overflow bit
- jnz .Lsub # CF is cleared by compare with 0
- mov (%rsp,$j,8),%rax
- cmp ($np,$j,8),%rax # tp[num-1]-np[num-1]
- jae .Lsub # if taken CF was cleared by above cmp
+ xor $j,$j # j=0
+
+ mov $n0,$m1
+ mulq $m0 # ap[0]*bp[0]
+ mov %rax,$A[0]
+ mov ($np),%rax
+
+ imulq $A[0],$m1 # "tp[0]"*n0
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[0]*m1
+ add %rax,$A[0] # discarded
+ mov 8($ap),%rax
+ adc \$0,%rdx
+ mov %rdx,$N[1]
+
+ mulq $m0
+ add %rax,$A[1]
+ mov 8($np),%rax
+ adc \$0,%rdx
+ mov %rdx,$A[0]
+
+ mulq $m1
+ add %rax,$N[1]
+ mov 16($ap),%rax
+ adc \$0,%rdx
+ add $A[1],$N[1]
+ lea 4($j),$j # j++
+ adc \$0,%rdx
+ mov $N[1],(%rsp)
+ mov %rdx,$N[0]
+ jmp .L1st4x
+.align 16
+.L1st4x:
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$A[0]
+ mov -16($np,$j,8),%rax
+ adc \$0,%rdx
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[0]
+ mov -8($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0]
+ adc \$0,%rdx
+ mov $N[0],-24(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$A[1]
+ mov -8($np,$j,8),%rax
+ adc \$0,%rdx
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov ($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0]
+ adc \$0,%rdx
+ mov $N[1],-16(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[0]
+
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$A[0]
+ mov ($np,$j,8),%rax
+ adc \$0,%rdx
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[0]
+ mov 8($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0]
+ adc \$0,%rdx
+ mov $N[0],-8(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$A[1]
+ mov 8($np,$j,8),%rax
+ adc \$0,%rdx
+ lea 4($j),$j # j++
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov -16($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0]
+ adc \$0,%rdx
+ mov $N[1],-32(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[0]
+ cmp $num,$j
+ jl .L1st4x
+
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$A[0]
+ mov -16($np,$j,8),%rax
+ adc \$0,%rdx
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[0]
+ mov -8($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0]
+ adc \$0,%rdx
+ mov $N[0],-24(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[0]
+ add %rax,$A[1]
+ mov -8($np,$j,8),%rax
+ adc \$0,%rdx
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov ($ap),%rax # ap[0]
+ adc \$0,%rdx
+ add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0]
+ adc \$0,%rdx
+ mov $N[1],-16(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[0]
+
+ xor $N[1],$N[1]
+ add $A[0],$N[0]
+ adc \$0,$N[1]
+ mov $N[0],-8(%rsp,$j,8)
+ mov $N[1],(%rsp,$j,8) # store upmost overflow bit
+
+ lea 1($i),$i # i++
.align 4
-.Lcopy:
- mov (%rsp,$j,8),%rax
- mov %rax,($rp,$j,8) # rp[i]=tp[i]
- mov $i,(%rsp,$j,8) # zap temporary vector
+.Louter4x:
+ mov ($bp,$i,8),$m0 # m0=bp[i]
+ xor $j,$j # j=0
+ mov (%rsp),$A[0]
+ mov $n0,$m1
+ mulq $m0 # ap[0]*bp[i]
+ add %rax,$A[0] # ap[0]*bp[i]+tp[0]
+ mov ($np),%rax
+ adc \$0,%rdx
+
+ imulq $A[0],$m1 # tp[0]*n0
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[0]*m1
+ add %rax,$A[0] # "$N[0]", discarded
+ mov 8($ap),%rax
+ adc \$0,%rdx
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[1]
+ mov 8($np),%rax
+ adc \$0,%rdx
+ add 8(%rsp),$A[1] # +tp[1]
+ adc \$0,%rdx
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov 16($ap),%rax
+ adc \$0,%rdx
+ add $A[1],$N[1] # np[j]*m1+ap[j]*bp[i]+tp[j]
+ lea 4($j),$j # j+=2
+ adc \$0,%rdx
+ mov $N[1],(%rsp) # tp[j-1]
+ mov %rdx,$N[0]
+ jmp .Linner4x
+.align 16
+.Linner4x:
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[0]
+ mov -16($np,$j,8),%rax
+ adc \$0,%rdx
+ add -16(%rsp,$j,8),$A[0] # ap[j]*bp[i]+tp[j]
+ adc \$0,%rdx
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[0]
+ mov -8($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[0],$N[0]
+ adc \$0,%rdx
+ mov $N[0],-24(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[1]
+ mov -8($np,$j,8),%rax
+ adc \$0,%rdx
+ add -8(%rsp,$j,8),$A[1]
+ adc \$0,%rdx
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov ($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[1],$N[1]
+ adc \$0,%rdx
+ mov $N[1],-16(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[0]
+
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[0]
+ mov ($np,$j,8),%rax
+ adc \$0,%rdx
+ add (%rsp,$j,8),$A[0] # ap[j]*bp[i]+tp[j]
+ adc \$0,%rdx
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[0]
+ mov 8($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[0],$N[0]
+ adc \$0,%rdx
+ mov $N[0],-8(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[1]
+ mov 8($np,$j,8),%rax
+ adc \$0,%rdx
+ add 8(%rsp,$j,8),$A[1]
+ adc \$0,%rdx
+ lea 4($j),$j # j++
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov -16($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[1],$N[1]
+ adc \$0,%rdx
+ mov $N[1],-32(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[0]
+ cmp $num,$j
+ jl .Linner4x
+
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[0]
+ mov -16($np,$j,8),%rax
+ adc \$0,%rdx
+ add -16(%rsp,$j,8),$A[0] # ap[j]*bp[i]+tp[j]
+ adc \$0,%rdx
+ mov %rdx,$A[1]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[0]
+ mov -8($ap,$j,8),%rax
+ adc \$0,%rdx
+ add $A[0],$N[0]
+ adc \$0,%rdx
+ mov $N[0],-24(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[1]
+
+ mulq $m0 # ap[j]*bp[i]
+ add %rax,$A[1]
+ mov -8($np,$j,8),%rax
+ adc \$0,%rdx
+ add -8(%rsp,$j,8),$A[1]
+ adc \$0,%rdx
+ lea 1($i),$i # i++
+ mov %rdx,$A[0]
+
+ mulq $m1 # np[j]*m1
+ add %rax,$N[1]
+ mov ($ap),%rax # ap[0]
+ adc \$0,%rdx
+ add $A[1],$N[1]
+ adc \$0,%rdx
+ mov $N[1],-16(%rsp,$j,8) # tp[j-1]
+ mov %rdx,$N[0]
+
+ xor $N[1],$N[1]
+ add $A[0],$N[0]
+ adc \$0,$N[1]
+ add (%rsp,$num,8),$N[0] # pull upmost overflow bit
+ adc \$0,$N[1]
+ mov $N[0],-8(%rsp,$j,8)
+ mov $N[1],(%rsp,$j,8) # store upmost overflow bit
+
+ cmp $num,$i
+ jl .Louter4x
+___
+{
+my @ri=("%rax","%rdx",$m0,$m1);
+$code.=<<___;
+ mov 16(%rsp,$num,8),$rp # restore $rp
+ mov 0(%rsp),@ri[0] # tp[0]
+ pxor %xmm0,%xmm0
+ mov 8(%rsp),@ri[1] # tp[1]
+ shr \$2,$num # num/=4
+ lea (%rsp),$ap # borrow ap for tp
+ xor $i,$i # i=0 and clear CF!
+
+ sub 0($np),@ri[0]
+ mov 16($ap),@ri[2] # tp[2]
+ mov 24($ap),@ri[3] # tp[3]
+ sbb 8($np),@ri[1]
+ lea -1($num),$j # j=num/4-1
+ jmp .Lsub4x
+.align 16
+.Lsub4x:
+ mov @ri[0],0($rp,$i,8) # rp[i]=tp[i]-np[i]
+ mov @ri[1],8($rp,$i,8) # rp[i]=tp[i]-np[i]
+ sbb 16($np,$i,8),@ri[2]
+ mov 32($ap,$i,8),@ri[0] # tp[i+1]
+ mov 40($ap,$i,8),@ri[1]
+ sbb 24($np,$i,8),@ri[3]
+ mov @ri[2],16($rp,$i,8) # rp[i]=tp[i]-np[i]
+ mov @ri[3],24($rp,$i,8) # rp[i]=tp[i]-np[i]
+ sbb 32($np,$i,8),@ri[0]
+ mov 48($ap,$i,8),@ri[2]
+ mov 56($ap,$i,8),@ri[3]
+ sbb 40($np,$i,8),@ri[1]
+ lea 4($i),$i # i++
+ dec $j # doesnn't affect CF!
+ jnz .Lsub4x
+
+ mov @ri[0],0($rp,$i,8) # rp[i]=tp[i]-np[i]
+ mov 32($ap,$i,8),@ri[0] # load overflow bit
+ sbb 16($np,$i,8),@ri[2]
+ mov @ri[1],8($rp,$i,8) # rp[i]=tp[i]-np[i]
+ sbb 24($np,$i,8),@ri[3]
+ mov @ri[2],16($rp,$i,8) # rp[i]=tp[i]-np[i]
+
+ sbb \$0,@ri[0] # handle upmost overflow bit
+ mov @ri[3],24($rp,$i,8) # rp[i]=tp[i]-np[i]
+ xor $i,$i # i=0
+ and @ri[0],$ap
+ not @ri[0]
+ mov $rp,$np
+ and @ri[0],$np
+ lea -1($num),$j
+ or $np,$ap # ap=borrow?tp:rp
+
+ movdqu ($ap),%xmm1
+ movdqa %xmm0,(%rsp)
+ movdqu %xmm1,($rp)
+ jmp .Lcopy4x
+.align 16
+.Lcopy4x: # copy or in-place refresh
+ movdqu 16($ap,$i),%xmm2
+ movdqu 32($ap,$i),%xmm1
+ movdqa %xmm0,16(%rsp,$i)
+ movdqu %xmm2,16($rp,$i)
+ movdqa %xmm0,32(%rsp,$i)
+ movdqu %xmm1,32($rp,$i)
+ lea 32($i),$i
dec $j
- jge .Lcopy
-.align 4
-.Lexit:
- mov 8(%rsp,$num,8),%rsp # restore %rsp
+ jnz .Lcopy4x
+
+ shl \$2,$num
+ movdqu 16($ap,$i),%xmm2
+ movdqa %xmm0,16(%rsp,$i)
+ movdqu %xmm2,16($rp,$i)
+___
+}
+$code.=<<___;
+ mov 8(%rsp,$num,8),%rsi # restore %rsp
+ mov \$1,%rax
+ mov (%rsi),%r15
+ mov 8(%rsi),%r14
+ mov 16(%rsi),%r13
+ mov 24(%rsi),%r12
+ mov 32(%rsi),%rbp
+ mov 40(%rsi),%rbx
+ lea 48(%rsi),%rsp
+.Lmul4x_epilogue:
+ ret
+.size bn_mul4x_mont,.-bn_mul4x_mont
+___
+}}}
+\f{{{
+######################################################################
+# void bn_sqr8x_mont(
+my $rptr="%rdi"; # const BN_ULONG *rptr,
+my $aptr="%rsi"; # const BN_ULONG *aptr,
+my $bptr="%rdx"; # not used
+my $nptr="%rcx"; # const BN_ULONG *nptr,
+my $n0 ="%r8"; # const BN_ULONG *n0);
+my $num ="%r9"; # int num, has to be divisible by 8
+
+my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
+my @A0=("%r10","%r11");
+my @A1=("%r12","%r13");
+my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
+
+$code.=<<___;
+.type bn_sqr8x_mont,\@function,6
+.align 32
+bn_sqr8x_mont:
+.Lsqr8x_enter:
+ push %rbx
+ push %rbp
+ push %r12
+ push %r13
+ push %r14
+ push %r15
+
+ shl \$3,${num}d # convert $num to bytes
+ xor %r10,%r10
+ mov %rsp,%r11 # put aside %rsp
+ sub $num,%r10 # -$num
+ mov ($n0),$n0 # *n0
+ lea -72(%rsp,%r10,2),%rsp # alloca(frame+2*$num)
+ and \$-1024,%rsp # minimize TLB usage
+ ##############################################################
+ # Stack layout
+ #
+ # +0 saved $num, used in reduction section
+ # +8 &t[2*$num], used in reduction section
+ # +32 saved $rptr
+ # +40 saved $nptr
+ # +48 saved *n0
+ # +56 saved %rsp
+ # +64 t[2*$num]
+ #
+ mov $rptr,32(%rsp) # save $rptr
+ mov $nptr,40(%rsp)
+ mov $n0, 48(%rsp)
+ mov %r11, 56(%rsp) # save original %rsp
+.Lsqr8x_body:
+ ##############################################################
+ # Squaring part:
+ #
+ # a) multiply-n-add everything but a[i]*a[i];
+ # b) shift result of a) by 1 to the left and accumulate
+ # a[i]*a[i] products;
+ #
+ ##############################################################
+ # a[1]a[0]
+ # a[2]a[0]
+ # a[3]a[0]
+ # a[2]a[1]
+ # a[4]a[0]
+ # a[3]a[1]
+ # a[5]a[0]
+ # a[4]a[1]
+ # a[3]a[2]
+ # a[6]a[0]
+ # a[5]a[1]
+ # a[4]a[2]
+ # a[7]a[0]
+ # a[6]a[1]
+ # a[5]a[2]
+ # a[4]a[3]
+ # a[7]a[1]
+ # a[6]a[2]
+ # a[5]a[3]
+ # a[7]a[2]
+ # a[6]a[3]
+ # a[5]a[4]
+ # a[7]a[3]
+ # a[6]a[4]
+ # a[7]a[4]
+ # a[6]a[5]
+ # a[7]a[5]
+ # a[7]a[6]
+ # a[1]a[0]
+ # a[2]a[0]
+ # a[3]a[0]
+ # a[4]a[0]
+ # a[5]a[0]
+ # a[6]a[0]
+ # a[7]a[0]
+ # a[2]a[1]
+ # a[3]a[1]
+ # a[4]a[1]
+ # a[5]a[1]
+ # a[6]a[1]
+ # a[7]a[1]
+ # a[3]a[2]
+ # a[4]a[2]
+ # a[5]a[2]
+ # a[6]a[2]
+ # a[7]a[2]
+ # a[4]a[3]
+ # a[5]a[3]
+ # a[6]a[3]
+ # a[7]a[3]
+ # a[5]a[4]
+ # a[6]a[4]
+ # a[7]a[4]
+ # a[6]a[5]
+ # a[7]a[5]
+ # a[7]a[6]
+ # a[0]a[0]
+ # a[1]a[1]
+ # a[2]a[2]
+ # a[3]a[3]
+ # a[4]a[4]
+ # a[5]a[5]
+ # a[6]a[6]
+ # a[7]a[7]
+
+ lea 32(%r10),$i # $i=-($num-32)
+ lea ($aptr,$num),$aptr # end of a[] buffer, ($aptr,$i)=&ap[2]
+
+ mov $num,$j # $j=$num
+
+ # comments apply to $num==8 case
+ mov -32($aptr,$i),$a0 # a[0]
+ lea 64(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num]
+ mov -24($aptr,$i),%rax # a[1]
+ lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"]
+ mov -16($aptr,$i),$ai # a[2]
+ mov %rax,$a1
+
+ mul $a0 # a[1]*a[0]
+ mov %rax,$A0[0] # a[1]*a[0]
+ mov $ai,%rax # a[2]
+ mov %rdx,$A0[1]
+ mov $A0[0],-24($tptr,$i) # t[1]
+
+ mul $a0 # a[2]*a[0]
+ add %rax,$A0[1]
+ mov $ai,%rax
+ adc \$0,%rdx
+ mov $A0[1],-16($tptr,$i) # t[2]
+ mov %rdx,$A0[0]
+
+ lea -16($i),$j # j=-16
+
+
+ mov 8($aptr,$j),$ai # a[3]
+ mul $a1 # a[2]*a[1]
+ mov %rax,$A1[0] # a[2]*a[1]+t[3]
+ mov $ai,%rax
+ mov %rdx,$A1[1]
+
+ lea 16($j),$j
+ mul $a0 # a[3]*a[0]
+ add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3]
+ mov $ai,%rax
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+ add $A1[0],$A0[0]
+ adc \$0,$A0[1]
+ mov $A0[0],-8($tptr,$j) # t[3]
+ jmp .Lsqr4x_1st
+
+.align 32
+.Lsqr4x_1st:
+ mov ($aptr,$j),$ai # a[4]
+ mul $a1 # a[3]*a[1]
+ add %rax,$A1[1] # a[3]*a[1]+t[4]
+ mov $ai,%rax
+ mov %rdx,$A1[0]
+ adc \$0,$A1[0]
+
+ mul $a0 # a[4]*a[0]
+ add %rax,$A0[1] # a[4]*a[0]+a[3]*a[1]+t[4]
+ mov $ai,%rax # a[3]
+ mov 8($aptr,$j),$ai # a[5]
+ mov %rdx,$A0[0]
+ adc \$0,$A0[0]
+ add $A1[1],$A0[1]
+ adc \$0,$A0[0]
+
+
+ mul $a1 # a[4]*a[3]
+ add %rax,$A1[0] # a[4]*a[3]+t[5]
+ mov $ai,%rax
+ mov $A0[1],($tptr,$j) # t[4]
+ mov %rdx,$A1[1]
+ adc \$0,$A1[1]
+
+ mul $a0 # a[5]*a[2]
+ add %rax,$A0[0] # a[5]*a[2]+a[4]*a[3]+t[5]
+ mov $ai,%rax
+ mov 16($aptr,$j),$ai # a[6]
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+ add $A1[0],$A0[0]
+ adc \$0,$A0[1]
+
+ mul $a1 # a[5]*a[3]
+ add %rax,$A1[1] # a[5]*a[3]+t[6]
+ mov $ai,%rax
+ mov $A0[0],8($tptr,$j) # t[5]
+ mov %rdx,$A1[0]
+ adc \$0,$A1[0]
+
+ mul $a0 # a[6]*a[2]
+ add %rax,$A0[1] # a[6]*a[2]+a[5]*a[3]+t[6]
+ mov $ai,%rax # a[3]
+ mov 24($aptr,$j),$ai # a[7]
+ mov %rdx,$A0[0]
+ adc \$0,$A0[0]
+ add $A1[1],$A0[1]
+ adc \$0,$A0[0]
+
+
+ mul $a1 # a[6]*a[5]
+ add %rax,$A1[0] # a[6]*a[5]+t[7]
+ mov $ai,%rax
+ mov $A0[1],16($tptr,$j) # t[6]
+ mov %rdx,$A1[1]
+ adc \$0,$A1[1]
+
+ mul $a0 # a[7]*a[4]
+ add %rax,$A0[0] # a[7]*a[4]+a[6]*a[5]+t[6]
+ mov $ai,%rax
+ lea 32($j),$j
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+ add $A1[0],$A0[0]
+ adc \$0,$A0[1]
+ mov $A0[0],-8($tptr,$j) # t[7]
+
+ cmp \$0,$j
+ jne .Lsqr4x_1st
+
+ mul $a1 # a[7]*a[5]
+ add %rax,$A1[1]
+ lea 16($i),$i
+ adc \$0,%rdx
+ add $A0[1],$A1[1]
+ adc \$0,%rdx
+
+ mov $A1[1],($tptr) # t[8]
+ mov %rdx,$A1[0]
+ mov %rdx,8($tptr) # t[9]
+ jmp .Lsqr4x_outer
+
+.align 32
+.Lsqr4x_outer: # comments apply to $num==6 case
+ mov -32($aptr,$i),$a0 # a[0]
+ lea 64(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num]
+ mov -24($aptr,$i),%rax # a[1]
+ lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"]
+ mov -16($aptr,$i),$ai # a[2]
+ mov %rax,$a1
+
+ mov -24($tptr,$i),$A0[0] # t[1]
+ mul $a0 # a[1]*a[0]
+ add %rax,$A0[0] # a[1]*a[0]+t[1]
+ mov $ai,%rax # a[2]
+ adc \$0,%rdx
+ mov $A0[0],-24($tptr,$i) # t[1]
+ mov %rdx,$A0[1]
+
+ mul $a0 # a[2]*a[0]
+ add %rax,$A0[1]
+ mov $ai,%rax
+ adc \$0,%rdx
+ add -16($tptr,$i),$A0[1] # a[2]*a[0]+t[2]
+ mov %rdx,$A0[0]
+ adc \$0,$A0[0]
+ mov $A0[1],-16($tptr,$i) # t[2]
+
+ lea -16($i),$j # j=-16
+ xor $A1[0],$A1[0]
+
+
+ mov 8($aptr,$j),$ai # a[3]
+ mul $a1 # a[2]*a[1]
+ add %rax,$A1[0] # a[2]*a[1]+t[3]
+ mov $ai,%rax
+ adc \$0,%rdx
+ add 8($tptr,$j),$A1[0]
+ mov %rdx,$A1[1]
+ adc \$0,$A1[1]
+
+ mul $a0 # a[3]*a[0]
+ add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3]
+ mov $ai,%rax
+ adc \$0,%rdx
+ add $A1[0],$A0[0]
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+ mov $A0[0],8($tptr,$j) # t[3]
+
+ lea 16($j),$j
+ jmp .Lsqr4x_inner
+
+.align 32
+.Lsqr4x_inner:
+ mov ($aptr,$j),$ai # a[4]
+ mul $a1 # a[3]*a[1]
+ add %rax,$A1[1] # a[3]*a[1]+t[4]
+ mov $ai,%rax
+ mov %rdx,$A1[0]
+ adc \$0,$A1[0]
+ add ($tptr,$j),$A1[1]
+ adc \$0,$A1[0]
+
+ mul $a0 # a[4]*a[0]
+ add %rax,$A0[1] # a[4]*a[0]+a[3]*a[1]+t[4]
+ mov $ai,%rax # a[3]
+ mov 8($aptr,$j),$ai # a[5]
+ mov %rdx,$A0[0]
+ adc \$0,$A0[0]
+ add $A1[1],$A0[1]
+ adc \$0,$A0[0]
+
+ mul $a1 # a[4]*a[3]
+ add %rax,$A1[0] # a[4]*a[3]+t[5]
+ mov $A0[1],($tptr,$j) # t[4]
+ mov $ai,%rax
+ mov %rdx,$A1[1]
+ adc \$0,$A1[1]
+ add 8($tptr,$j),$A1[0]
+ lea 16($j),$j # j++
+ adc \$0,$A1[1]
+
+ mul $a0 # a[5]*a[2]
+ add %rax,$A0[0] # a[5]*a[2]+a[4]*a[3]+t[5]
+ mov $ai,%rax
+ adc \$0,%rdx
+ add $A1[0],$A0[0]
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+ mov $A0[0],-8($tptr,$j) # t[5], "preloaded t[1]" below
+
+ cmp \$0,$j
+ jne .Lsqr4x_inner
+
+ mul $a1 # a[5]*a[3]
+ add %rax,$A1[1]
+ adc \$0,%rdx
+ add $A0[1],$A1[1]
+ adc \$0,%rdx
+
+ mov $A1[1],($tptr) # t[6], "preloaded t[2]" below
+ mov %rdx,$A1[0]
+ mov %rdx,8($tptr) # t[7], "preloaded t[3]" below
+
+ add \$16,$i
+ jnz .Lsqr4x_outer
+
+ # comments apply to $num==4 case
+ mov -32($aptr),$a0 # a[0]
+ lea 64(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num]
+ mov -24($aptr),%rax # a[1]
+ lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"]
+ mov -16($aptr),$ai # a[2]
+ mov %rax,$a1
+
+ mul $a0 # a[1]*a[0]
+ add %rax,$A0[0] # a[1]*a[0]+t[1], preloaded t[1]
+ mov $ai,%rax # a[2]
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+
+ mul $a0 # a[2]*a[0]
+ add %rax,$A0[1]
+ mov $ai,%rax
+ mov $A0[0],-24($tptr) # t[1]
+ mov %rdx,$A0[0]
+ adc \$0,$A0[0]
+ add $A1[1],$A0[1] # a[2]*a[0]+t[2], preloaded t[2]
+ mov -8($aptr),$ai # a[3]
+ adc \$0,$A0[0]
+
+ mul $a1 # a[2]*a[1]
+ add %rax,$A1[0] # a[2]*a[1]+t[3], preloaded t[3]
+ mov $ai,%rax
+ mov $A0[1],-16($tptr) # t[2]
+ mov %rdx,$A1[1]
+ adc \$0,$A1[1]
+
+ mul $a0 # a[3]*a[0]
+ add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3]
+ mov $ai,%rax
+ mov %rdx,$A0[1]
+ adc \$0,$A0[1]
+ add $A1[0],$A0[0]
+ adc \$0,$A0[1]
+ mov $A0[0],-8($tptr) # t[3]
+
+ mul $a1 # a[3]*a[1]
+ add %rax,$A1[1]
+ mov -16($aptr),%rax # a[2]
+ adc \$0,%rdx
+ add $A0[1],$A1[1]
+ adc \$0,%rdx
+
+ mov $A1[1],($tptr) # t[4]
+ mov %rdx,$A1[0]
+ mov %rdx,8($tptr) # t[5]
+
+ mul $ai # a[2]*a[3]
+___
+{
+my ($shift,$carry)=($a0,$a1);
+my @S=(@A1,$ai,$n0);
+$code.=<<___;
+ add \$16,$i
+ xor $shift,$shift
+ sub $num,$i # $i=16-$num
+ xor $carry,$carry
+
+ add $A1[0],%rax # t[5]
+ adc \$0,%rdx
+ mov %rax,8($tptr) # t[5]
+ mov %rdx,16($tptr) # t[6]
+ mov $carry,24($tptr) # t[7]
+
+ mov -16($aptr,$i),%rax # a[0]
+ lea 64(%rsp,$num,2),$tptr
+ xor $A0[0],$A0[0] # t[0]
+ mov -24($tptr,$i,2),$A0[1] # t[1]
+
+ lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[1] # | t[2*i]>>63
+ mov -16($tptr,$i,2),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov -8($tptr,$i,2),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[0]
+ mov -8($aptr,$i),%rax # a[i+1] # prefetch
+ mov $S[0],-32($tptr,$i,2)
+ adc %rdx,$S[1]
+
+ lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift
+ mov $S[1],-24($tptr,$i,2)
+ sbb $carry,$carry # mov cf,$carry
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[3] # | t[2*i]>>63
+ mov 0($tptr,$i,2),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov 8($tptr,$i,2),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[2]
+ mov 0($aptr,$i),%rax # a[i+1] # prefetch
+ mov $S[2],-16($tptr,$i,2)
+ adc %rdx,$S[3]
+ lea 16($i),$i
+ mov $S[3],-40($tptr,$i,2)
+ sbb $carry,$carry # mov cf,$carry
+ jmp .Lsqr4x_shift_n_add
+
+.align 32
+.Lsqr4x_shift_n_add:
+ lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[1] # | t[2*i]>>63
+ mov -16($tptr,$i,2),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov -8($tptr,$i,2),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[0]
+ mov -8($aptr,$i),%rax # a[i+1] # prefetch
+ mov $S[0],-32($tptr,$i,2)
+ adc %rdx,$S[1]
+
+ lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift
+ mov $S[1],-24($tptr,$i,2)
+ sbb $carry,$carry # mov cf,$carry
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[3] # | t[2*i]>>63
+ mov 0($tptr,$i,2),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov 8($tptr,$i,2),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[2]
+ mov 0($aptr,$i),%rax # a[i+1] # prefetch
+ mov $S[2],-16($tptr,$i,2)
+ adc %rdx,$S[3]
+
+ lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
+ mov $S[3],-8($tptr,$i,2)
+ sbb $carry,$carry # mov cf,$carry
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[1] # | t[2*i]>>63
+ mov 16($tptr,$i,2),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov 24($tptr,$i,2),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[0]
+ mov 8($aptr,$i),%rax # a[i+1] # prefetch
+ mov $S[0],0($tptr,$i,2)
+ adc %rdx,$S[1]
+
+ lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift
+ mov $S[1],8($tptr,$i,2)
+ sbb $carry,$carry # mov cf,$carry
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[3] # | t[2*i]>>63
+ mov 32($tptr,$i,2),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov 40($tptr,$i,2),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[2]
+ mov 16($aptr,$i),%rax # a[i+1] # prefetch
+ mov $S[2],16($tptr,$i,2)
+ adc %rdx,$S[3]
+ mov $S[3],24($tptr,$i,2)
+ sbb $carry,$carry # mov cf,$carry
+ add \$32,$i
+ jnz .Lsqr4x_shift_n_add
+
+ lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[1] # | t[2*i]>>63
+ mov -16($tptr),$A0[0] # t[2*i+2] # prefetch
+ mov $A0[1],$shift # shift=t[2*i+1]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ mov -8($tptr),$A0[1] # t[2*i+2+1] # prefetch
+ adc %rax,$S[0]
+ mov -8($aptr),%rax # a[i+1] # prefetch
+ mov $S[0],-32($tptr)
+ adc %rdx,$S[1]
+
+ lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1|shift
+ mov $S[1],-24($tptr)
+ sbb $carry,$carry # mov cf,$carry
+ shr \$63,$A0[0]
+ lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
+ shr \$63,$A0[1]
+ or $A0[0],$S[3] # | t[2*i]>>63
+ mul %rax # a[i]*a[i]
+ neg $carry # mov $carry,cf
+ adc %rax,$S[2]
+ adc %rdx,$S[3]
+ mov $S[2],-16($tptr)
+ mov $S[3],-8($tptr)
+___
+}\f
+######################################################################
+# Montgomery reduction part, "word-by-word" algorithm.
+#
+# This new path is inspired by multiple submissions from Intel, by
+# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
+# Vinodh Gopal...
+{
+my ($nptr,$tptr,$carry,$m0)=("%rbp","%rdi","%rsi","%rbx");
+
+$code.=<<___;
+ mov 40(%rsp),$nptr # pull $nptr
+ xor %rax,%rax
+ lea ($nptr,$num),%rdx # end of n[]
+ lea 64(%rsp,$num,2),$tptr # end of t[] buffer
+ mov %rdx,0(%rsp)
+ mov $tptr,8(%rsp)
+ mov %rax,($tptr) # clear top-most carry bit
+ lea 64(%rsp,$num),$tptr # end of initial t[] window
+ neg $num
+ jmp .L8x_reduction_loop
+
+.align 32
+.L8x_reduction_loop:
+ lea ($tptr,$num),$tptr # start of current t[] window
+ mov 8*0($tptr),$m0
+ mov 8*1($tptr),%r9
+ mov 8*2($tptr),%r10
+ mov 8*3($tptr),%r11
+ mov 8*4($tptr),%r12
+ mov 8*5($tptr),%r13
+ mov 8*6($tptr),%r14
+ mov 8*7($tptr),%r15
+ lea 8*8($tptr),$tptr
+
+ mov $m0,%r8
+ imulq 48(%rsp),$m0 # n0*a[0]
+ mov 8*0($nptr),%rax # n[0]
+ mov \$8,%ecx
+ jmp .L8x_reduce
+
+.align 32
+.L8x_reduce:
+ mulq $m0
+ mov 8*1($nptr),%rax # n[1]
+ neg %r8
+ mov %rdx,%r8
+ adc \$0,%r8
+
+ mulq $m0
+ add %rax,%r9
+ mov 8*2($nptr),%rax
+ adc \$0,%rdx
+ add %r9,%r8
+ mov $m0,64-8(%rsp,%rcx,8) # put aside n0*a[i]
+ mov %rdx,%r9
+ adc \$0,%r9
+
+ mulq $m0
+ add %rax,%r10
+ mov 8*3($nptr),%rax
+ adc \$0,%rdx
+ add %r10,%r9
+ mov 48(%rsp),$carry # pull n0, borrow $carry
+ mov %rdx,%r10
+ adc \$0,%r10
+
+ mulq $m0
+ add %rax,%r11
+ mov 8*4($nptr),%rax
+ adc \$0,%rdx
+ imulq %r8,$carry # modulo-scheduled
+ add %r11,%r10
+ mov %rdx,%r11
+ adc \$0,%r11
+
+ mulq $m0
+ add %rax,%r12
+ mov 8*5($nptr),%rax
+ adc \$0,%rdx
+ add %r12,%r11
+ mov %rdx,%r12
+ adc \$0,%r12
+
+ mulq $m0
+ add %rax,%r13
+ mov 8*6($nptr),%rax
+ adc \$0,%rdx
+ add %r13,%r12
+ mov %rdx,%r13
+ adc \$0,%r13
+
+ mulq $m0
+ add %rax,%r14
+ mov 8*7($nptr),%rax
+ adc \$0,%rdx
+ add %r14,%r13
+ mov %rdx,%r14
+ adc \$0,%r14
+
+ mulq $m0
+ mov $carry,$m0 # n0*a[i]
+ add %rax,%r15
+ mov 8*0($nptr),%rax # n[0]
+ adc \$0,%rdx
+ add %r15,%r14
+ mov %rdx,%r15
+ adc \$0,%r15
+
+ dec %ecx
+ jnz .L8x_reduce
+
+ lea 8*8($nptr),$nptr
+ xor %rax,%rax
+ mov 8(%rsp),%rdx # pull end of t[]
+ cmp 0(%rsp),$nptr # end of n[]?
+ jae .L8x_no_tail
+
+ add 8*0($tptr),%r8
+ adc 8*1($tptr),%r9
+ adc 8*2($tptr),%r10
+ adc 8*3($tptr),%r11
+ adc 8*4($tptr),%r12
+ adc 8*5($tptr),%r13
+ adc 8*6($tptr),%r14
+ adc 8*7($tptr),%r15
+ sbb $carry,$carry # top carry
+
+ mov 64+56(%rsp),$m0 # pull n0*a[0]
+ mov \$8,%ecx
+ mov 8*0($nptr),%rax
+ jmp .L8x_tail
+
+.align 32
+.L8x_tail:
+ mulq $m0
+ add %rax,%r8
+ mov 8*1($nptr),%rax
+ mov %r8,($tptr) # save result
+ mov %rdx,%r8
+ adc \$0,%r8
+
+ mulq $m0
+ add %rax,%r9
+ mov 8*2($nptr),%rax
+ adc \$0,%rdx
+ add %r9,%r8
+ lea 8($tptr),$tptr # $tptr++
+ mov %rdx,%r9
+ adc \$0,%r9
+
+ mulq $m0
+ add %rax,%r10
+ mov 8*3($nptr),%rax
+ adc \$0,%rdx
+ add %r10,%r9
+ mov %rdx,%r10
+ adc \$0,%r10
+
+ mulq $m0
+ add %rax,%r11
+ mov 8*4($nptr),%rax
+ adc \$0,%rdx
+ add %r11,%r10
+ mov %rdx,%r11
+ adc \$0,%r11
+
+ mulq $m0
+ add %rax,%r12
+ mov 8*5($nptr),%rax
+ adc \$0,%rdx
+ add %r12,%r11
+ mov %rdx,%r12
+ adc \$0,%r12
+
+ mulq $m0
+ add %rax,%r13
+ mov 8*6($nptr),%rax
+ adc \$0,%rdx
+ add %r13,%r12
+ mov %rdx,%r13
+ adc \$0,%r13
+
+ mulq $m0
+ add %rax,%r14
+ mov 8*7($nptr),%rax
+ adc \$0,%rdx
+ add %r14,%r13
+ mov %rdx,%r14
+ adc \$0,%r14
+
+ mulq $m0
+ mov 64-16(%rsp,%rcx,8),$m0 # pull n0*a[i]
+ add %rax,%r15
+ adc \$0,%rdx
+ add %r15,%r14
+ mov 8*0($nptr),%rax # pull n[0]
+ mov %rdx,%r15
+ adc \$0,%r15
+
+ dec %ecx
+ jnz .L8x_tail
+
+ lea 8*8($nptr),$nptr
+ mov 8(%rsp),%rdx # pull end of t[]
+ cmp 0(%rsp),$nptr # end of n[]?
+ jae .L8x_tail_done # break out of loop
+
+ mov 64+56(%rsp),$m0 # pull n0*a[0]
+ neg $carry
+ mov 8*0($nptr),%rax # pull n[0]
+ adc 8*0($tptr),%r8
+ adc 8*1($tptr),%r9
+ adc 8*2($tptr),%r10
+ adc 8*3($tptr),%r11
+ adc 8*4($tptr),%r12
+ adc 8*5($tptr),%r13
+ adc 8*6($tptr),%r14
+ adc 8*7($tptr),%r15
+ sbb $carry,$carry # top carry
+
+ mov \$8,%ecx
+ jmp .L8x_tail
+
+.align 32
+.L8x_tail_done:
+ add (%rdx),%r8 # can this overflow?
+ xor %rax,%rax
+
+ neg $carry
+.L8x_no_tail:
+ adc 8*0($tptr),%r8
+ adc 8*1($tptr),%r9
+ adc 8*2($tptr),%r10
+ adc 8*3($tptr),%r11
+ adc 8*4($tptr),%r12
+ adc 8*5($tptr),%r13
+ adc 8*6($tptr),%r14
+ adc 8*7($tptr),%r15
+ adc \$0,%rax # top-most carry
+
+ mov 40(%rsp),$nptr # restore $nptr
+
+ mov %r8,8*0($tptr) # store top 512 bits
+ mov %r9,8*1($tptr)
+ mov $nptr,$num # $num is %r9, can't be moved upwards
+ mov %r10,8*2($tptr)
+ sub 0(%rsp),$num # -$num
+ mov %r11,8*3($tptr)
+ mov %r12,8*4($tptr)
+ mov %r13,8*5($tptr)
+ mov %r14,8*6($tptr)
+ mov %r15,8*7($tptr)
+ lea 8*8($tptr),$tptr
+ mov %rax,(%rdx) # store top-most carry
+
+ cmp %rdx,$tptr # end of t[]?
+ jb .L8x_reduction_loop
+
+ neg $num # restore $num
+___
+}\f
+##############################################################
+# Post-condition, 4x unrolled copy from bn_mul_mont
+#
+{
+my ($tptr,$nptr)=("%rbx",$aptr);
+my @ri=("%rax","%rdx","%r10","%r11");
+$code.=<<___;
+ mov 64(%rsp,$num),@ri[0] # tp[0]
+ lea 64(%rsp,$num),$tptr # upper half of t[2*$num] holds result
+ mov 40(%rsp),$nptr # restore $nptr
+ shr \$5,$num # num/4
+ mov 8($tptr),@ri[1] # t[1]
+ xor $i,$i # i=0 and clear CF!
+
+ mov 32(%rsp),$rptr # restore $rptr
+ sub 0($nptr),@ri[0]
+ mov 16($tptr),@ri[2] # t[2]
+ mov 24($tptr),@ri[3] # t[3]
+ sbb 8($nptr),@ri[1]
+ lea -1($num),$j # j=num/4-1
+ jmp .Lsqr4x_sub
+.align 32
+.Lsqr4x_sub:
+ mov @ri[0],0($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ mov @ri[1],8($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ sbb 16($nptr,$i,8),@ri[2]
+ mov 32($tptr,$i,8),@ri[0] # tp[i+1]
+ mov 40($tptr,$i,8),@ri[1]
+ sbb 24($nptr,$i,8),@ri[3]
+ mov @ri[2],16($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ mov @ri[3],24($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ sbb 32($nptr,$i,8),@ri[0]
+ mov 48($tptr,$i,8),@ri[2]
+ mov 56($tptr,$i,8),@ri[3]
+ sbb 40($nptr,$i,8),@ri[1]
+ lea 4($i),$i # i++
+ dec $j # doesn't affect CF!
+ jnz .Lsqr4x_sub
+
+ mov @ri[0],0($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ mov 32($tptr,$i,8),@ri[0] # load overflow bit
+ sbb 16($nptr,$i,8),@ri[2]
+ mov @ri[1],8($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ sbb 24($nptr,$i,8),@ri[3]
+ mov @ri[2],16($rptr,$i,8) # rp[i]=tp[i]-np[i]
+
+ sbb \$0,@ri[0] # handle upmost overflow bit
+ mov @ri[3],24($rptr,$i,8) # rp[i]=tp[i]-np[i]
+ xor $i,$i # i=0
+ and @ri[0],$tptr
+ not @ri[0]
+ mov $rptr,$nptr
+ and @ri[0],$nptr
+ lea -1($num),$j
+ or $nptr,$tptr # tp=borrow?tp:rp
+
+ pxor %xmm0,%xmm0
+ lea 64(%rsp,$num,8),$nptr
+ movdqu ($tptr),%xmm1
+ lea ($nptr,$num,8),$nptr
+ movdqa %xmm0,64(%rsp) # zap lower half of temporary vector
+ movdqa %xmm0,($nptr) # zap upper half of temporary vector
+ movdqu %xmm1,($rptr)
+ jmp .Lsqr4x_copy
+.align 32
+.Lsqr4x_copy: # copy or in-place refresh
+ movdqu 16($tptr,$i),%xmm2
+ movdqu 32($tptr,$i),%xmm1
+ movdqa %xmm0,80(%rsp,$i) # zap lower half of temporary vector
+ movdqa %xmm0,96(%rsp,$i) # zap lower half of temporary vector
+ movdqa %xmm0,16($nptr,$i) # zap upper half of temporary vector
+ movdqa %xmm0,32($nptr,$i) # zap upper half of temporary vector
+ movdqu %xmm2,16($rptr,$i)
+ movdqu %xmm1,32($rptr,$i)
+ lea 32($i),$i
+ dec $j
+ jnz .Lsqr4x_copy
+
+ movdqu 16($tptr,$i),%xmm2
+ movdqa %xmm0,80(%rsp,$i) # zap lower half of temporary vector
+ movdqa %xmm0,16($nptr,$i) # zap upper half of temporary vector
+ movdqu %xmm2,16($rptr,$i)
+___
+}
+$code.=<<___;
+ mov 56(%rsp),%rsi # restore %rsp
mov \$1,%rax
+ mov 0(%rsi),%r15
+ mov 8(%rsi),%r14
+ mov 16(%rsi),%r13
+ mov 24(%rsi),%r12
+ mov 32(%rsi),%rbp
+ mov 40(%rsi),%rbx
+ lea 48(%rsi),%rsp
+.Lsqr8x_epilogue:
+ ret
+.size bn_sqr8x_mont,.-bn_sqr8x_mont
+___
+}}}
+$code.=<<___;
+.asciz "Montgomery Multiplication for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+.align 16
+___
+
+# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
+# CONTEXT *context,DISPATCHER_CONTEXT *disp)
+if ($win64) {
+$rec="%rcx";
+$frame="%rdx";
+$context="%r8";
+$disp="%r9";
+
+$code.=<<___;
+.extern __imp_RtlVirtualUnwind
+.type mul_handler,\@abi-omnipotent
+.align 16
+mul_handler:
+ push %rsi
+ push %rdi
+ push %rbx
+ push %rbp
+ push %r12
+ push %r13
+ push %r14
+ push %r15
+ pushfq
+ sub \$64,%rsp
+
+ mov 120($context),%rax # pull context->Rax
+ mov 248($context),%rbx # pull context->Rip
+
+ mov 8($disp),%rsi # disp->ImageBase
+ mov 56($disp),%r11 # disp->HandlerData
+
+ mov 0(%r11),%r10d # HandlerData[0]
+ lea (%rsi,%r10),%r10 # end of prologue label
+ cmp %r10,%rbx # context->Rip<end of prologue label
+ jb .Lcommon_seh_tail
+
+ mov 152($context),%rax # pull context->Rsp
+
+ mov 4(%r11),%r10d # HandlerData[1]
+ lea (%rsi,%r10),%r10 # epilogue label
+ cmp %r10,%rbx # context->Rip>=epilogue label
+ jae .Lcommon_seh_tail
+
+ mov 192($context),%r10 # pull $num
+ mov 8(%rax,%r10,8),%rax # pull saved stack pointer
+ lea 48(%rax),%rax
+
+ mov -8(%rax),%rbx
+ mov -16(%rax),%rbp
+ mov -24(%rax),%r12
+ mov -32(%rax),%r13
+ mov -40(%rax),%r14
+ mov -48(%rax),%r15
+ mov %rbx,144($context) # restore context->Rbx
+ mov %rbp,160($context) # restore context->Rbp
+ mov %r12,216($context) # restore context->R12
+ mov %r13,224($context) # restore context->R13
+ mov %r14,232($context) # restore context->R14
+ mov %r15,240($context) # restore context->R15
+
+ jmp .Lcommon_seh_tail
+.size mul_handler,.-mul_handler
+
+.type sqr_handler,\@abi-omnipotent
+.align 16
+sqr_handler:
+ push %rsi
+ push %rdi
+ push %rbx
+ push %rbp
+ push %r12
+ push %r13
+ push %r14
+ push %r15
+ pushfq
+ sub \$64,%rsp
+
+ mov 120($context),%rax # pull context->Rax
+ mov 248($context),%rbx # pull context->Rip
+
+ lea .Lsqr8x_body(%rip),%r10
+ cmp %r10,%rbx # context->Rip<.Lsqr_body
+ jb .Lcommon_seh_tail
+
+ mov 152($context),%rax # pull context->Rsp
+
+ lea .Lsqr8x_epilogue(%rip),%r10
+ cmp %r10,%rbx # context->Rip>=.Lsqr_epilogue
+ jae .Lcommon_seh_tail
+
+ mov 56(%rax),%rax # pull saved stack pointer
+ lea 48(%rax),%rax
+
+ mov -8(%rax),%rbx
+ mov -16(%rax),%rbp
+ mov -24(%rax),%r12
+ mov -32(%rax),%r13
+ mov -40(%rax),%r14
+ mov -48(%rax),%r15
+ mov %rbx,144($context) # restore context->Rbx
+ mov %rbp,160($context) # restore context->Rbp
+ mov %r12,216($context) # restore context->R12
+ mov %r13,224($context) # restore context->R13
+ mov %r14,232($context) # restore context->R14
+ mov %r15,240($context) # restore context->R15
+
+.Lcommon_seh_tail:
+ mov 8(%rax),%rdi
+ mov 16(%rax),%rsi
+ mov %rax,152($context) # restore context->Rsp
+ mov %rsi,168($context) # restore context->Rsi
+ mov %rdi,176($context) # restore context->Rdi
+
+ mov 40($disp),%rdi # disp->ContextRecord
+ mov $context,%rsi # context
+ mov \$154,%ecx # sizeof(CONTEXT)
+ .long 0xa548f3fc # cld; rep movsq
+
+ mov $disp,%rsi
+ xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
+ mov 8(%rsi),%rdx # arg2, disp->ImageBase
+ mov 0(%rsi),%r8 # arg3, disp->ControlPc
+ mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
+ mov 40(%rsi),%r10 # disp->ContextRecord
+ lea 56(%rsi),%r11 # &disp->HandlerData
+ lea 24(%rsi),%r12 # &disp->EstablisherFrame
+ mov %r10,32(%rsp) # arg5
+ mov %r11,40(%rsp) # arg6
+ mov %r12,48(%rsp) # arg7
+ mov %rcx,56(%rsp) # arg8, (NULL)
+ call *__imp_RtlVirtualUnwind(%rip)
+
+ mov \$1,%eax # ExceptionContinueSearch
+ add \$64,%rsp
+ popfq
pop %r15
pop %r14
pop %r13
pop %r12
pop %rbp
pop %rbx
+ pop %rdi
+ pop %rsi
ret
+.size sqr_handler,.-sqr_handler
-.align 16
-.Lsub: mov (%rsp,$i,8),%rax
- sbb ($np,$i,8),%rax
- mov %rax,($rp,$i,8) # rp[i]=tp[i]-np[j]
- lea 1($i),$i # i++
- dec $j # doesn't affect CF!
- jge .Lsub
- lea -1($num),$j # j=num-1
- sbb \$0,%rdx
- jc .Lcopy # tp was less than np
+.section .pdata
.align 4
-.Lzap: mov $i,(%rsp,$j,8) # zap temporary vector
- dec $j
- jge .Lzap
- jmp .Lexit
-.size bn_mul_mont,.-bn_mul_mont
-.asciz "Montgomery Multiplication for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
+ .rva .LSEH_begin_bn_mul_mont
+ .rva .LSEH_end_bn_mul_mont
+ .rva .LSEH_info_bn_mul_mont
+
+ .rva .LSEH_begin_bn_mul4x_mont
+ .rva .LSEH_end_bn_mul4x_mont
+ .rva .LSEH_info_bn_mul4x_mont
+
+ .rva .LSEH_begin_bn_sqr8x_mont
+ .rva .LSEH_end_bn_sqr8x_mont
+ .rva .LSEH_info_bn_sqr8x_mont
+
+.section .xdata
+.align 8
+.LSEH_info_bn_mul_mont:
+ .byte 9,0,0,0
+ .rva mul_handler
+ .rva .Lmul_body,.Lmul_epilogue # HandlerData[]
+.LSEH_info_bn_mul4x_mont:
+ .byte 9,0,0,0
+ .rva mul_handler
+ .rva .Lmul4x_body,.Lmul4x_epilogue # HandlerData[]
+.LSEH_info_bn_sqr8x_mont:
+ .byte 9,0,0,0
+ .rva sqr_handler
___
+}
print $code;
close STDOUT;
projects / openssl.git / blobdiff