#
# Modulo-scheduling SSE2 loops results in further 15-20% improvement.
# Integer-only code [being equipped with dedicated squaring procedure]
-# gives >=30% on rsa512 sign benchmark...
+# gives ~40% on rsa512 sign benchmark...
$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
push(@INC,"${dir}","${dir}../../perlasm");
$i="edx";
$j="ecx";
-$ap="esi";
+$ap="esi"; $tp="esi"; # overlapping variables!!!
$rp="edi"; $bp="edi"; # overlapping variables!!!
$np="ebp";
$num="ebx";
-$_rp=&DWP(4*0,"esp"); # stack top layout
-$_ap=&DWP(4*1,"esp");
-$_bp=&DWP(4*2,"esp");
-$_np=&DWP(4*3,"esp");
-$_n0=&DWP(4*4,"esp");
-$_num=&DWP(4*5,"esp");
+$_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");
$frame=32; # size of above frame rounded up to 16n
&cmp ("edi",4);
&jl (&label("just_leave"));
- ################################# load argument block...
- &mov ("eax",&wparam(0)); # BN_ULONG *rp
- &mov ("ebx",&wparam(1)); # const BN_ULONG *ap
- &mov ("ecx",&wparam(2)); # const BN_ULONG *bp
- &mov ("edx",&wparam(3)); # const BN_ULONG *np
- &mov ("esi",&wparam(4)); # const BN_ULONG *n0
- #&mov ("edi",&wparam(5)); # int num
-
+ &lea ("esi",&wparam(0)); # put aside pointer to argument block
+ &lea ("edx",&wparam(1)); # load ap
&mov ("ebp","esp"); # saved stack pointer!
&add ("edi",2); # extra two words on top of tp
&neg ("edi");
&lea ("esp",&DWP(-$frame,"esp","edi",4)); # alloca($frame+4*(num+2))
&neg ("edi");
- &and ("esp",-4096); # minimize TLB utilization
+
+ # minimize cache contention by arraning 2K window between stack
+ # pointer and ap argument [np is also position sensitive vector,
+ # but it's assumed to be near ap, as it's allocated at ~same
+ # time].
+ &mov ("eax","esp");
+ &sub ("eax","edx");
+ &and ("eax",2047);
+ &sub ("esp","eax"); # this aligns sp and ap modulo 2048
+
+ &xor ("edx","esp");
+ &and ("edx",2048);
+ &xor ("edx",2048);
+ &sub ("esp","edx"); # this splits them apart modulo 4096
+
+ &and ("esp",-64); # align to cache line
+
+ ################################# load argument block...
+ &mov ("eax",&DWP(0*4,"esi"));# BN_ULONG *rp
+ &mov ("ebx",&DWP(1*4,"esi"));# const BN_ULONG *ap
+ &mov ("ecx",&DWP(2*4,"esi"));# const BN_ULONG *bp
+ &mov ("edx",&DWP(3*4,"esi"));# const BN_ULONG *np
+ &mov ("esi",&DWP(4*4,"esi"));# const BN_ULONG *n0
+ #&mov ("edi",&DWP(5*4,"esi"));# int num
&mov ("esi",&DWP(0,"esi")); # pull n0[0]
&mov ($_rp,"eax"); # ... save a copy of argument block
&movq ($acc0,$mul1); # I wish movd worked for
&pand ($acc0,$mask); # inter-register transfers
- &pmuludq($mul1,$_n0); # *=n0
+ &pmuludq($mul1,$_n0q); # *=n0
&pmuludq($car1,$mul1); # "t[0]"*np[0]*n0
&paddq ($car1,$acc0);
&psrlq ($car1,32);
&inc ($j); # j++
-&set_label("1st");
+&set_label("1st",16);
&pmuludq($acc0,$mul0); # ap[j]*bp[0]
&pmuludq($acc1,$mul1); # np[j]*m1
&paddq ($car0,$acc0); # +=c0
&psrlq ($car1,32);
&paddq ($car1,$car0);
- &movq (&DWP($frame,"esp",$num,4),$car1); # tp[num].tp[num-1]
+ &movq (&QWP($frame,"esp",$num,4),$car1); # tp[num].tp[num-1]
\f
&inc ($i); # i++
&set_label("outer");
&movq ($car0,$mul1);
&pand ($acc0,$mask);
- &pmuludq($mul1,$_n0); # *=n0
+ &pmuludq($mul1,$_n0q); # *=n0
&pmuludq($car1,$mul1);
&paddq ($car1,$acc0);
&movd ($temp,&DWP($frame+4,"esp",$num,4)); # += tp[num]
&paddq ($car1,$car0);
&paddq ($car1,$temp);
- &movq (&DWP($frame,"esp",$num,4),$car1); # tp[num].tp[num-1]
+ &movq (&QWP($frame,"esp",$num,4),$car1); # tp[num].tp[num-1]
&lea ($i,&DWP(1,$i)); # i++
&cmp ($i,$num);
&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 cores, it's still a tad slower
- # for >=2048-bits keys on *elder* CPUs:-( "Competitive" means
- # compared to the original integer-only assembler. 512-bit
- # RSA sign is better by >=30%, but that's about all one can
- # say about all CPUs...
+ # 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
$word="edi";
&mov ($carry,"edx");
&mul ($word); # np[j]*m
- &add ($carry,&DWP($frame,"esp",$j,4)); # +=tp[j]
+ &add ($carry,&DWP($frame,"esp",$num,4)); # +=tp[num-1]
&adc ("edx",0);
&add ($carry,"eax");
&adc ("edx",0);
&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("x86done"));
+ &je (&label("common_tail"));
- &mov ($word,&DWP(0,$j)); # bp[i]
+ &mov ($word,&DWP(0,$j)); # bp[i+1]
&mov ($inp,$_ap);
&mov ($_bp,$j); # &bp[++i]
&xor ($j,$j);
&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("x86done"));
+ &je (&label("common_tail"));
\f
&mov ($word,&DWP(4,$inp,$j,4)); # ap[i]
&lea ($j,&DWP(1,$j));
&mov ($carry,"edx");
&mul ($word); # ap[j]*ap[i]
&add ("eax",$carry);
- &lea ($j,&DWP(1,$j));
+ &lea ($carry,&DWP(0,"eax","eax"));
&adc ("edx",0);
- &lea ($carry,&DWP(0,$sbit,"eax",2));
&shr ("eax",31);
- &add ($carry,&DWP($frame-4,"esp",$j,4)); # +=tp[j]
+ &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 ("eax",&DWP(4,$inp)); # np[1]
&jmp (&label("3rdmadd"));
-\f
-&set_label("x86done",4);
- &mov ($np,$_np); # make adjustments for tail processing
}
-
+\f
&set_label("common_tail",16);
- &mov ("esi",&DWP($frame+4,"esp",$num,4));# load upmost overflow bit
+ &mov ($np,$_np); # load modulus pointer
&mov ($rp,$_rp); # load result pointer
- # [$ap and $bp are zapped]
- &xor ($i,$i); # i=0
+ &lea ($tp,&DWP($frame,"esp")); # [$ap and $bp are zapped]
+
+ &mov ("eax",&DWP(0,$tp)); # tp[0]
&mov ($j,$num); # j=num-1
- &cmp ("esi",0); # clears CF unconditionally
- &jnz (&label("sub"));
- &mov ("eax",&DWP($frame,"esp",$j,4));
- &cmp ("eax",&DWP(0,$np,$j,4)); # tp[num-1]-np[num-1]?
- &jae (&label("sub")); # if taken CF is cleared
-&set_label("copy",16);
- &mov ("eax",&DWP($frame,"esp",$j,4));
- &mov (&DWP(0,$rp,$j,4),"eax"); # rp[i]=tp[i]
- &mov (&DWP($frame,"esp",$j,4),$j); # zap temporary vector
- &dec ($j);
- &jge (&label("copy"));
- &jmp (&label("exit"));
+ &xor ($i,$i); # i=0 and clear CF!
&set_label("sub",16);
- &mov ("eax",&DWP($frame,"esp",$i,4));
&sbb ("eax",&DWP(0,$np,$i,4));
&mov (&DWP(0,$rp,$i,4),"eax"); # rp[i]=tp[i]-np[i]
- &lea ($i,&DWP(1,$i)); # i++
&dec ($j); # doesn't affect CF!
+ &mov ("eax",&DWP(4,$tp,$i,4)); # tp[i+1]
+ &lea ($i,&DWP(1,$i)); # i++
&jge (&label("sub"));
- &mov ($j,$num); # j=num-1
- &sbb ("esi",0); # esi holds upmost overflow bit
- &jc (&label("copy"));
-&set_label("zap",16);
- &mov (&DWP($frame,"esp",$j,4),$i); # zap temporary vector
- &dec ($j);
- &jge (&label("zap"));
-
-&set_label("exit",4);
+
+ &sbb ("eax",0); # handle upmost overflow bit
+ &and ($tp,"eax");
+ ¬ ("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");