3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. Rights for redistribution and usage in source and binary
6 # forms are granted according to the OpenSSL license.
7 # ====================================================================
9 # This is a "teaser" code, as it can be improved in several ways...
10 # First of all non-SSE2 path should be implemented (yes, for now it
11 # performs Montgomery multiplication/convolution only on SSE2-capable
12 # CPUs such as P4, others fall down to original code). Then inner loop
13 # can be unrolled and modulo-scheduled to improve ILP and possibly
14 # moved to 128-bit XMM register bank (though it would require input
15 # rearrangement and/or increase bus bandwidth utilization). Dedicated
16 # squaring procedure should give further performance improvement...
17 # Yet, for being draft, the code improves rsa512 *sign* benchmark by
18 # 110%(!), rsa1024 one - by 70% and rsa4096 - by 20%:-)
20 push(@INC,"perlasm","../../perlasm");
23 &asm_init($ARGV[0],$0);
26 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
28 &external_label("OPENSSL_ia32cap_P") if ($sse2);
30 &function_begin("bn_mul_mont",$sse2?"EXTRN\t_OPENSSL_ia32cap_P:DWORD":"");
35 $rp="edi"; $bp="edi"; # overlapping variables!!!
40 $bias=2; # amount of extra words in tp
41 # (rounded up to even value)
42 $_rp=&DWP(4*($bias+0),"esp",$num,4); # stack frame layout below tp
43 $_ap=&DWP(4*($bias+1),"esp",$num,4);
44 $_bp=&DWP(4*($bias+2),"esp",$num,4);
45 $_np=&DWP(4*($bias+3),"esp",$num,4);
46 $_n0=&DWP(4*($bias+4),"esp",$num,4);
47 $_sp=&DWP(4*($bias+5),"esp",$num,4);
49 $acc0="mm0"; # mmx register bank layout
59 &picmeup("eax","OPENSSL_ia32cap_P");
60 &bt (&DWP(0,"eax"),26);
61 &mov ("eax",0); # zero signals "we did nothing"
62 &jnc (&label("non_sse2"));
64 ################################# load argument block...
65 &mov ("eax",&wparam(0)); # BN_ULONG *rp
66 &mov ("ebx",&wparam(1)); # const BN_ULONG *ap
67 &mov ("ecx",&wparam(2)); # const BN_ULONG *bp
68 &mov ("edx",&wparam(3)); # const BN_ULONG *np
69 &mov ("esi",&wparam(4)); # BN_ULONG n0
70 &mov ($num,&wparam(5)); # int num
72 &mov ("edi","esp"); # saved stack pointer!
75 &lea ("esp",&DWP(0,"esp",$num,4)); # alloca(4*(num+$bias+6))
77 &and ("esp",-1024); # minimize TLB utilization
78 &sub ($num,$bias+6); # num is restored to its original value
79 # and will remain constant from now...
81 &mov ($_rp,"eax"); # ... save a copy of argument block
86 &mov ($_sp,"edi"); # saved stack pointer!
89 &movd ($mask,"eax"); # mask 32 lower bits
91 &mov ($ap,$_ap); # load input pointers
98 &movd ($mul0,&DWP(0,$bp)); # bp[0]
99 &movd ($mul1,&DWP(0,$ap)); # ap[0]
100 &movd ($car1,&DWP(0,$np)); # np[0]
102 &pmuludq($mul1,$mul0); # ap[0]*bp[0]
104 &movq ($acc0,$mul1); # I wish movd worked for
105 &pand ($acc0,$mask); # inter-register transfers
107 &pmuludq($mul1,$_n0); # *=n0
109 &pmuludq($car1,$mul1); # "t[0]"*np[0]*n0
110 &paddq ($car1,$acc0);
117 &movd ($acc0,&DWP(0,$ap,$j,4)); # ap[j]
118 &movd ($acc1,&DWP(0,$np,$j,4)); # np[j]
119 &pmuludq($acc0,$mul0); # ap[j]*bp[0]
120 &pmuludq($acc1,$mul1); # np[j]*m1
122 &paddq ($car0,$acc0); # +=c0
126 &paddq ($car1,$acc1); # +=c1
127 &paddq ($car1,$acc0); # +=ap[j]*bp[0];
128 &movd (&DWP(-4,"esp",$j,4),$car1); # tp[j-1]=
133 &lea ($j,&DWP(1,$j));
137 &paddq ($car1,$car0);
138 &movq (&DWP(-4,"esp",$num,4),$car1);
144 &movd ($mul0,&DWP(0,$bp,$i,4)); # bp[i]
145 &movd ($mul1,&DWP(0,$ap)); # ap[0]
146 &movd ($temp,&DWP(0,"esp")); # tp[0]
147 &movd ($car1,&DWP(0,$np,$j,4)); # np[0]
148 &pmuludq($mul1,$mul0); # ap[0]*bp[i]
150 &paddq ($mul1,$temp); # +=tp[0]
155 &pmuludq($mul1,$_n0); # *=n0
157 &pmuludq($car1,$mul1);
158 &paddq ($car1,$acc0);
165 &movd ($acc0,&DWP(0,$ap,$j,4)); # ap[j]
166 &movd ($acc1,&DWP(0,$np,$j,4)); # np[j]
167 &movd ($temp,&DWP(0,"esp",$j,4)); # tp[j]
168 &pmuludq($acc0,$mul0); # ap[j]*bp[i]
169 &pmuludq($acc1,$mul1); # np[j]*m1
170 &paddq ($car0,$temp); # +=tp[j]
171 &paddq ($car0,$acc0); # +=c0
175 &paddq ($car1,$acc1); # +=c1
176 &paddq ($car1,$acc0); # +=ap[j]*bp[i]+tp[j]
177 &movd (&DWP(-4,"esp",$j,4),$car1); # tp[j-1]
182 &lea ($j,&DWP(1,$j)); # j++
184 &jl (&label("inner"));
186 &movd ($temp,&DWP(0,"esp",$num,4));
187 &paddq ($car1,$car0);
188 &paddq ($car1,$temp);
189 &movq (&DWP(-4,"esp",$num,4),$car1);
191 &lea ($i,&DWP(1,$i)); # i++
193 &jl (&label("outer"));
195 &emms (); # done with mmx bank
197 &mov ("esi",&DWP(0,"esp",$num,4)); # load upmost overflow bit
198 &mov ($rp,$_rp); # load result pointer
199 # [$ap and $bp are zapped]
201 &lea ($j,&DWP(-1,$num)); # j=num-1
202 &cmp ("esi",0); # clears CF unconditionally
203 &jnz (&label("sub"));
204 &mov ("eax",&DWP(0,"esp",$j,4));
205 &cmp ("eax",&DWP(0,$np,$j,4)); # tp[num-1]-np[num-1]?
206 &jae (&label("sub")); # if taken CF is cleared
208 &mov ("eax",&DWP(0,"esp",$j,4));
209 &mov (&DWP(0,$rp,$j,4),"eax"); # rp[i]=tp[i]
210 &mov (&DWP(0,"esp",$j,4),$j); # zap temporary vector
212 &jge (&label("copy"));
213 &jmp (&label("exit_sse2"));
216 &mov ("eax",&DWP(0,"esp",$i,4));
217 &sbb ("eax",&DWP(0,$np,$i,4));
218 &mov (&DWP(0,$rp,$i,4),"eax"); # rp[i]=tp[i]-np[i]
219 &lea ($i,&DWP(1,$i)); # i++
220 &dec ($j); # doesn't affect CF!
221 &jge (&label("sub"));
222 &lea ($j,&DWP(-1,$num)); # j=num-1
223 &sbb ("esi",0); # esi holds upmost overflow bit
224 &jc (&label("copy"));
226 &mov (&DWP(0,"esp",$j,4),$i); # zap temporary vector
228 &jge (&label("zap"));
230 &set_label("exit_sse2");
231 &mov ("esp",$_sp); # pull saved stack pointer
233 &set_label("non_sse2");
236 &function_end("bn_mul_mont");