3 # ====================================================================
4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
10 # This module doesn't present direct interest for OpenSSL, because it
11 # doesn't provide better performance for longer keys, at least not on
12 # in-order-execution cores. While 512-bit RSA sign operations can be
13 # 65% faster in 64-bit mode, 1024-bit ones are only 15% faster, and
14 # 4096-bit ones are up to 15% slower. In 32-bit mode it varies from
15 # 16% improvement for 512-bit RSA sign to -33% for 4096-bit RSA
16 # verify:-( All comparisons are against bn_mul_mont-free assembler.
17 # The module might be of interest to embedded system developers, as
18 # the code is smaller than 1KB, yet offers >3x improvement on MIPS64
19 # and 75-30% [less for longer keys] on MIPS32 over compiler-generated
22 ######################################################################
23 # There is a number of MIPS ABI in use, O32 and N32/64 are most
24 # widely used. Then there is a new contender: NUBI. It appears that if
25 # one picks the latter, it's possible to arrange code in ABI neutral
26 # manner. Therefore let's stick to NUBI register layout:
28 ($zero,$at,$t0,$t1,$t2)=map("\$$_",(0..2,24,25));
29 ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
30 ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7,$s8,$s9,$s10,$s11)=map("\$$_",(12..23));
31 ($gp,$tp,$sp,$fp,$ra)=map("\$$_",(3,28..31));
33 # The return value is placed in $a0. Following coding rules facilitate
36 # - never ever touch $tp, "thread pointer", former $gp;
37 # - copy return value to $t0, former $v0 [or to $a0 if you're adapting
39 # - on O32 populate $a4-$a7 with 'lw $aN,4*N($sp)' if necessary;
41 # For reference here is register layout for N32/64 MIPS ABIs:
43 # ($zero,$at,$v0,$v1)=map("\$$_",(0..3));
44 # ($a0,$a1,$a2,$a3,$a4,$a5,$a6,$a7)=map("\$$_",(4..11));
45 # ($t0,$t1,$t2,$t3,$t8,$t9)=map("\$$_",(12..15,24,25));
46 # ($s0,$s1,$s2,$s3,$s4,$s5,$s6,$s7)=map("\$$_",(16..23));
47 # ($gp,$sp,$fp,$ra)=map("\$$_",(28..31));
49 $flavour = shift || "o32"; # supported flavours are o32,n32,64,nubi32,nubi64
51 if ($flavour =~ /64|n32/i) {
52 $PTR_ADD="dadd"; # incidentally works even on n32
53 $PTR_SUB="dsub"; # incidentally works even on n32
64 $SAVED_REGS_MASK = ($flavour =~ /nubi/i) ? 0x00fff000 : 0x00ff0000;
68 ######################################################################
70 while (($output=shift) && ($output!~/^\w[\w\-]*\.\w+$/)) {}
71 open STDOUT,">$output";
73 if ($flavour =~ /64|n32/i) {
90 $rp=$a0; # BN_ULONG *rp,
91 $ap=$a1; # const BN_ULONG *ap,
92 $bp=$a2; # const BN_ULONG *bp,
93 $np=$a3; # const BN_ULONG *np,
94 $n0=$a4; # const BN_ULONG *n0,
127 $code.=<<___ if ($flavour =~ /o32/i);
135 slt $at,$num,17 # on in-order CPU
136 bnez $at,bn_mul_mont_internal
143 .ent bn_mul_mont_internal
144 bn_mul_mont_internal:
145 .frame $fp,$FRAMESIZE*$SZREG,$ra
146 .mask 0x40000000|$SAVED_REGS_MASK,-$SZREG
147 $PTR_SUB $sp,$FRAMESIZE*$SZREG
148 $REG_S $fp,($FRAMESIZE-1)*$SZREG($sp)
149 $REG_S $s11,($FRAMESIZE-2)*$SZREG($sp)
150 $REG_S $s10,($FRAMESIZE-3)*$SZREG($sp)
151 $REG_S $s9,($FRAMESIZE-4)*$SZREG($sp)
152 $REG_S $s8,($FRAMESIZE-5)*$SZREG($sp)
153 $REG_S $s7,($FRAMESIZE-6)*$SZREG($sp)
154 $REG_S $s6,($FRAMESIZE-7)*$SZREG($sp)
155 $REG_S $s5,($FRAMESIZE-8)*$SZREG($sp)
156 $REG_S $s4,($FRAMESIZE-9)*$SZREG($sp)
158 $code.=<<___ if ($flavour =~ /nubi/i);
159 $REG_S $s3,($FRAMESIZE-10)*$SZREG($sp)
160 $REG_S $s2,($FRAMESIZE-11)*$SZREG($sp)
161 $REG_S $s1,($FRAMESIZE-12)*$SZREG($sp)
162 $REG_S $s0,($FRAMESIZE-13)*$SZREG($sp)
169 $LD $bi,0($bp) # bp[0]
170 $LD $aj,0($ap) # ap[0]
171 $LD $nj,0($np) # np[0]
173 $PTR_SUB $sp,2*$BNSZ # place for two extra words
174 sll $num,`log($BNSZ)/log(2)`
348 $ST $hi1,2*$BNSZ($tp)
355 $PTR_ADD $tj,$sp,$num # &tp[num]
358 li $hi0,0 # clear borrow bit
361 .Lsub: $LD $lo0,($tp)
365 $SUBU $lo1,$lo0,$lo1 # tp[i]-np[i]
375 $SUBU $hi0,$hi1,$hi0 # handle upmost overflow bit
377 $PTR_SUB $rp,$num # restore rp
382 or $ap,$ap,$bp # ap=borrow?tp:rp
385 .Lcopy: $LD $aj,($ap)
399 $REG_L $fp,($FRAMESIZE-1)*$SZREG($sp)
400 $REG_L $s11,($FRAMESIZE-2)*$SZREG($sp)
401 $REG_L $s10,($FRAMESIZE-3)*$SZREG($sp)
402 $REG_L $s9,($FRAMESIZE-4)*$SZREG($sp)
403 $REG_L $s8,($FRAMESIZE-5)*$SZREG($sp)
404 $REG_L $s7,($FRAMESIZE-6)*$SZREG($sp)
405 $REG_L $s6,($FRAMESIZE-7)*$SZREG($sp)
406 $REG_L $s5,($FRAMESIZE-8)*$SZREG($sp)
407 $REG_L $s4,($FRAMESIZE-9)*$SZREG($sp)
409 $code.=<<___ if ($flavour =~ /nubi/i);
410 $REG_L $s3,($FRAMESIZE-10)*$SZREG($sp)
411 $REG_L $s2,($FRAMESIZE-11)*$SZREG($sp)
412 $REG_L $s1,($FRAMESIZE-12)*$SZREG($sp)
413 $REG_L $s0,($FRAMESIZE-13)*$SZREG($sp)
417 $PTR_ADD $sp,$FRAMESIZE*$SZREG
418 .end bn_mul_mont_internal
420 .asciiz "Montgomery Multiplication for MIPS, CRYPTOGAMS by <appro\@openssl.org>"
423 $code =~ s/\`([^\`]*)\`/eval $1/gem;