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 # ====================================================================
11 # This is a "teaser" code, as it can be improved in several ways...
12 # First of all non-SSE2 path should be implemented (yes, for now it
13 # performs Montgomery multiplication/convolution only on SSE2-capable
14 # CPUs such as P4, others fall down to original code). Then inner loop
15 # can be unrolled and modulo-scheduled to improve ILP and possibly
16 # moved to 128-bit XMM register bank (though it would require input
17 # rearrangement and/or increase bus bandwidth utilization). Dedicated
18 # squaring procedure should give further performance improvement...
19 # Yet, for being draft, the code improves rsa512 *sign* benchmark by
20 # 110%(!), rsa1024 one - by 70% and rsa4096 - by 20%:-)
22 push(@INC,"perlasm","../../perlasm");
25 &asm_init($ARGV[0],$0);
28 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
30 &external_label("OPENSSL_ia32cap_P") if ($sse2);
32 &function_begin("bn_mul_mont",$sse2?"EXTRN\t_OPENSSL_ia32cap_P:DWORD":"");
37 $rp="edi"; $bp="edi"; # overlapping variables!!!
41 $_rp=&DWP(4*0,"esp"); # stack top layout
46 $_num=&DWP(4*5,"esp");
48 $frame=32; # size of above frame rounded up to 16n
50 $acc0="mm0"; # mmx register bank layout
60 &picmeup("eax","OPENSSL_ia32cap_P");
61 &bt (&DWP(0,"eax"),26);
62 &jnc (&label("non_sse2"));
64 ################################# load argument block...
65 &mov ("eax",&wparam(0)); # BN_ULONG *rp
66 &mov ("ebx",&wparam(1)); # const BN_ULONG *ap
67 &mov ("ecx",&wparam(2)); # const BN_ULONG *bp
68 &mov ("edx",&wparam(3)); # const BN_ULONG *np
69 &mov ("esi",&wparam(4)); # BN_ULONG n0
70 &mov ($num,&wparam(5)); # int num
72 &mov ("edi","esp"); # saved stack pointer!
73 &add ($num,1); # extra word on top of tp
75 &lea ("esp",&DWP(-$frame,"esp",$num,4)); # alloca($frame+8*($num+1))
77 &and ("esp",-1024); # minimize TLB utilization
78 &sub ($num,1); # num is restored to its original value
79 # and will remain constant from now...
81 &mov ($_rp,"eax"); # ... save a copy of argument block
86 #&mov ($_num,$num); # redundant in sse2 context
87 &mov ($_sp,"edi"); # saved stack pointer!
90 &movd ($mask,"eax"); # mask 32 lower bits
92 &mov ($ap,$_ap); # load input pointers
99 &movd ($mul0,&DWP(0,$bp)); # bp[0]
100 &movd ($mul1,&DWP(0,$ap)); # ap[0]
101 &movd ($car1,&DWP(0,$np)); # np[0]
103 &pmuludq($mul1,$mul0); # ap[0]*bp[0]
105 &movq ($acc0,$mul1); # I wish movd worked for
106 &pand ($acc0,$mask); # inter-register transfers
108 &pmuludq($mul1,$_n0); # *=n0
110 &pmuludq($car1,$mul1); # "t[0]"*np[0]*n0
111 &paddq ($car1,$acc0);
118 &movd ($acc0,&DWP(0,$ap,$j,4)); # ap[j]
119 &movd ($acc1,&DWP(0,$np,$j,4)); # np[j]
120 &pmuludq($acc0,$mul0); # ap[j]*bp[0]
121 &pmuludq($acc1,$mul1); # np[j]*m1
123 &paddq ($car0,$acc0); # +=c0
127 &paddq ($car1,$acc1); # +=c1
128 &paddq ($car1,$acc0); # +=ap[j]*bp[0];
129 &movd (&DWP($frame-4,"esp",$j,4),$car1); # tp[j-1]=
134 &lea ($j,&DWP(1,$j));
138 &paddq ($car1,$car0);
139 &movq (&DWP($frame-4,"esp",$num,4),$car1);
145 &movd ($mul0,&DWP(0,$bp,$i,4)); # bp[i]
146 &movd ($mul1,&DWP(0,$ap)); # ap[0]
147 &movd ($temp,&DWP($frame,"esp")); # tp[0]
148 &movd ($car1,&DWP(0,$np)); # np[0]
149 &pmuludq($mul1,$mul0); # ap[0]*bp[i]
151 &paddq ($mul1,$temp); # +=tp[0]
156 &pmuludq($mul1,$_n0); # *=n0
158 &pmuludq($car1,$mul1);
159 &paddq ($car1,$acc0);
166 &movd ($acc0,&DWP(0,$ap,$j,4)); # ap[j]
167 &movd ($acc1,&DWP(0,$np,$j,4)); # np[j]
168 &movd ($temp,&DWP($frame,"esp",$j,4));# tp[j]
169 &pmuludq($acc0,$mul0); # ap[j]*bp[i]
170 &pmuludq($acc1,$mul1); # np[j]*m1
171 &paddq ($car0,$temp); # +=tp[j]
172 &paddq ($car0,$acc0); # +=c0
176 &paddq ($car1,$acc1); # +=c1
177 &paddq ($car1,$acc0); # +=ap[j]*bp[i]+tp[j]
178 &movd (&DWP($frame-4,"esp",$j,4),$car1); # tp[j-1]=
183 &lea ($j,&DWP(1,$j)); # j++
185 &jl (&label("inner"));
187 &movd ($temp,&DWP($frame,"esp",$num,4));
188 &paddq ($car1,$car0);
189 &paddq ($car1,$temp);
190 &movq (&DWP($frame-4,"esp",$num,4),$car1);
192 &lea ($i,&DWP(1,$i)); # i++
194 &jl (&label("outer"));
196 &emms (); # done with mmx bank
198 &mov ("esi",&DWP($frame,"esp",$num,4));# load upmost overflow bit
199 &mov ($rp,$_rp); # load result pointer
200 # [$ap and $bp are zapped]
202 &lea ($j,&DWP(-1,$num)); # j=num-1
203 &cmp ("esi",0); # clears CF unconditionally
204 &jnz (&label("sub"));
205 &mov ("eax",&DWP($frame,"esp",$j,4));
206 &cmp ("eax",&DWP(0,$np,$j,4)); # tp[num-1]-np[num-1]?
207 &jae (&label("sub")); # if taken CF is cleared
209 &mov ("eax",&DWP($frame,"esp",$j,4));
210 &mov (&DWP(0,$rp,$j,4),"eax"); # rp[i]=tp[i]
211 &mov (&DWP($frame,"esp",$j,4),$j); # zap temporary vector
213 &jge (&label("copy"));
214 &jmp (&label("exit_sse2"));
217 &mov ("eax",&DWP($frame,"esp",$i,4));
218 &sbb ("eax",&DWP(0,$np,$i,4));
219 &mov (&DWP(0,$rp,$i,4),"eax"); # rp[i]=tp[i]-np[i]
220 &lea ($i,&DWP(1,$i)); # i++
221 &dec ($j); # doesn't affect CF!
222 &jge (&label("sub"));
223 &lea ($j,&DWP(-1,$num)); # j=num-1
224 &sbb ("esi",0); # esi holds upmost overflow bit
225 &jc (&label("copy"));
227 &mov (&DWP($frame,"esp",$j,4),$i); # zap temporary vector
229 &jge (&label("zap"));
231 &set_label("exit_sse2");
232 &mov ("esp",$_sp); # pull saved stack pointer
234 &jmp (&label("leave"));
235 &set_label("non_sse2");
238 &xor ("eax","eax"); # zero signals "not implemented [yet]"
241 &function_end("bn_mul_mont");