2 # Copyright 2004-2020 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the Apache License 2.0 (the "License"). You may not use
5 # this file except in compliance with the License. You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
9 # Implemented as a Perl wrapper as we want to support several different
10 # architectures with single file. We pick up the target based on the
11 # file name we are asked to generate.
13 # It should be noted though that this perl code is nothing like
14 # <openssl>/crypto/perlasm/x86*. In this case perl is used pretty much
15 # as pre-processor to cover for platform differences in name decoration,
16 # linker tables, 32-/64-bit instruction sets...
18 # As you might know there're several PowerPC ABI in use. Most notably
19 # Linux and AIX use different 32-bit ABIs. Good news are that these ABIs
20 # are similar enough to implement leaf(!) functions, which would be ABI
21 # neutral. And that's what you find here: ABI neutral leaf functions.
22 # In case you wonder what that is...
26 # MEASUREMENTS WITH cc ON a 200 MhZ PowerPC 604e.
28 # The following is the performance of 32-bit compiler
31 # OpenSSL 0.9.6c 21 dec 2001
32 # built on: Tue Jun 11 11:06:51 EDT 2002
33 # options:bn(64,32) ...
34 #compiler: cc -DTHREADS -DAIX -DB_ENDIAN -DBN_LLONG -O3
35 # sign verify sign/s verify/s
36 #rsa 512 bits 0.0098s 0.0009s 102.0 1170.6
37 #rsa 1024 bits 0.0507s 0.0026s 19.7 387.5
38 #rsa 2048 bits 0.3036s 0.0085s 3.3 117.1
39 #rsa 4096 bits 2.0040s 0.0299s 0.5 33.4
40 #dsa 512 bits 0.0087s 0.0106s 114.3 94.5
41 #dsa 1024 bits 0.0256s 0.0313s 39.0 32.0
43 # Same benchmark with this assembler code:
45 #rsa 512 bits 0.0056s 0.0005s 178.6 2049.2
46 #rsa 1024 bits 0.0283s 0.0015s 35.3 674.1
47 #rsa 2048 bits 0.1744s 0.0050s 5.7 201.2
48 #rsa 4096 bits 1.1644s 0.0179s 0.9 55.7
49 #dsa 512 bits 0.0052s 0.0062s 191.6 162.0
50 #dsa 1024 bits 0.0149s 0.0180s 67.0 55.5
52 # Number of operations increases by at almost 75%
54 # Here are performance numbers for 64-bit compiler
57 # OpenSSL 0.9.6g [engine] 9 Aug 2002
58 # built on: Fri Apr 18 16:59:20 EDT 2003
59 # options:bn(64,64) ...
60 # compiler: cc -DTHREADS -D_REENTRANT -q64 -DB_ENDIAN -O3
61 # sign verify sign/s verify/s
62 #rsa 512 bits 0.0028s 0.0003s 357.1 3844.4
63 #rsa 1024 bits 0.0148s 0.0008s 67.5 1239.7
64 #rsa 2048 bits 0.0963s 0.0028s 10.4 353.0
65 #rsa 4096 bits 0.6538s 0.0102s 1.5 98.1
66 #dsa 512 bits 0.0026s 0.0032s 382.5 313.7
67 #dsa 1024 bits 0.0081s 0.0099s 122.8 100.6
69 # Same benchmark with this assembler code:
71 #rsa 512 bits 0.0020s 0.0002s 510.4 6273.7
72 #rsa 1024 bits 0.0088s 0.0005s 114.1 2128.3
73 #rsa 2048 bits 0.0540s 0.0016s 18.5 622.5
74 #rsa 4096 bits 0.3700s 0.0058s 2.7 171.0
75 #dsa 512 bits 0.0016s 0.0020s 610.7 507.1
76 #dsa 1024 bits 0.0047s 0.0058s 212.5 173.2
78 # Again, performance increases by at about 75%
80 # Mac OS X, Apple G5 1.8GHz (Note this is 32 bit code)
81 # OpenSSL 0.9.7c 30 Sep 2003
85 #rsa 512 bits 0.0011s 0.0001s 906.1 11012.5
86 #rsa 1024 bits 0.0060s 0.0003s 166.6 3363.1
87 #rsa 2048 bits 0.0370s 0.0010s 27.1 982.4
88 #rsa 4096 bits 0.2426s 0.0036s 4.1 280.4
89 #dsa 512 bits 0.0010s 0.0012s 1038.1 841.5
90 #dsa 1024 bits 0.0030s 0.0037s 329.6 269.7
91 #dsa 2048 bits 0.0101s 0.0127s 98.9 78.6
93 # Same benchmark with this assembler code:
95 #rsa 512 bits 0.0007s 0.0001s 1416.2 16645.9
96 #rsa 1024 bits 0.0036s 0.0002s 274.4 5380.6
97 #rsa 2048 bits 0.0222s 0.0006s 45.1 1589.5
98 #rsa 4096 bits 0.1469s 0.0022s 6.8 449.6
99 #dsa 512 bits 0.0006s 0.0007s 1664.2 1376.2
100 #dsa 1024 bits 0.0018s 0.0023s 545.0 442.2
101 #dsa 2048 bits 0.0061s 0.0075s 163.5 132.8
103 # Performance increase of ~60%
104 # Based on submission from Suresh N. Chari of IBM
106 # $output is the last argument if it looks like a file (it has an extension)
107 # $flavour is the first argument if it doesn't look like a file
108 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
109 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
111 if ($flavour =~ /32/) {
117 $LDU= "lwzu"; # load and update
119 $STU= "stwu"; # store and update
120 $UMULL= "mullw"; # unsigned multiply low
121 $UMULH= "mulhwu"; # unsigned multiply high
122 $UDIV= "divwu"; # unsigned divide
123 $UCMPI= "cmplwi"; # unsigned compare with immediate
124 $UCMP= "cmplw"; # unsigned compare
125 $CNTLZ= "cntlzw"; # count leading zeros
126 $SHL= "slw"; # shift left
127 $SHR= "srw"; # unsigned shift right
128 $SHRI= "srwi"; # unsigned shift right by immediate
129 $SHLI= "slwi"; # shift left by immediate
130 $CLRU= "clrlwi"; # clear upper bits
131 $INSR= "insrwi"; # insert right
132 $ROTL= "rotlwi"; # rotate left by immediate
133 $TR= "tw"; # conditional trap
134 } elsif ($flavour =~ /64/) {
139 # same as above, but 64-bit mnemonics...
141 $LDU= "ldu"; # load and update
143 $STU= "stdu"; # store and update
144 $UMULL= "mulld"; # unsigned multiply low
145 $UMULH= "mulhdu"; # unsigned multiply high
146 $UDIV= "divdu"; # unsigned divide
147 $UCMPI= "cmpldi"; # unsigned compare with immediate
148 $UCMP= "cmpld"; # unsigned compare
149 $CNTLZ= "cntlzd"; # count leading zeros
150 $SHL= "sld"; # shift left
151 $SHR= "srd"; # unsigned shift right
152 $SHRI= "srdi"; # unsigned shift right by immediate
153 $SHLI= "sldi"; # shift left by immediate
154 $CLRU= "clrldi"; # clear upper bits
155 $INSR= "insrdi"; # insert right
156 $ROTL= "rotldi"; # rotate left by immediate
157 $TR= "td"; # conditional trap
158 } else { die "nonsense $flavour"; }
160 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
161 ( $xlate="${dir}ppc-xlate.pl" and -f $xlate ) or
162 ( $xlate="${dir}../../perlasm/ppc-xlate.pl" and -f $xlate) or
163 die "can't locate ppc-xlate.pl";
165 open STDOUT,"| $^X $xlate $flavour \"$output\""
166 or die "can't call $xlate: $!";
169 #--------------------------------------------------------------------
176 # Created by: Suresh Chari
177 # IBM Thomas J. Watson Research Library
181 # Description: Optimized assembly routines for OpenSSL crypto
182 # on the 32 bitPowerPC platform.
187 # 2. Fixed bn_add,bn_sub and bn_div_words, added comments,
188 # cleaned up code. Also made a single version which can
189 # be used for both the AIX and Linux compilers. See NOTE
191 # 12/05/03 Suresh Chari
192 # (with lots of help from) Andy Polyakov
194 # 1. Initial version 10/20/02 Suresh Chari
197 # The following file works for the xlc,cc
200 # NOTE: To get the file to link correctly with the gcc compiler
201 # you have to change the names of the routines and remove
202 # the first .(dot) character. This should automatically
203 # be done in the build process.
205 # Hand optimized assembly code for the following routines
218 # NOTE: It is possible to optimize this code more for
219 # specific PowerPC or Power architectures. On the Northstar
220 # architecture the optimizations in this file do
221 # NOT provide much improvement.
223 # If you have comments or suggestions to improve code send
224 # me a note at schari\@us.ibm.com
226 #--------------------------------------------------------------------------
228 # Defines to be used in the assembly code.
230 #.set r0,0 # we use it as storage for value of 0
231 #.set SP,1 # preserved
232 #.set RTOC,2 # preserved
233 #.set r3,3 # 1st argument/return value
234 #.set r4,4 # 2nd argument/volatile register
235 #.set r5,5 # 3rd argument/volatile register
243 #.set r13,13 # not used, nor any other "below" it...
245 # Declare function names to be global
246 # NOTE: For gcc these names MUST be changed to remove
247 # the first . i.e. for example change ".bn_sqr_comba4"
248 # to "bn_sqr_comba4". This should be automatically done
251 .globl .bn_sqr_comba4
252 .globl .bn_sqr_comba8
253 .globl .bn_mul_comba4
254 .globl .bn_mul_comba8
260 .globl .bn_mul_add_words
268 # NOTE: The following label name should be changed to
269 # "bn_sqr_comba4" i.e. remove the first dot
270 # for the gcc compiler. This should be automatically
277 # Optimized version of bn_sqr_comba4.
279 # void bn_sqr_comba4(BN_ULONG *r, BN_ULONG *a)
283 # Freely use registers r5,r6,r7,r8,r9,r10,r11 as follows:
285 # r5,r6 are the two BN_ULONGs being multiplied.
286 # r7,r8 are the results of the 32x32 giving 64 bit multiply.
287 # r9,r10, r11 are the equivalents of c1,c2, c3.
288 # Here's the assembly
291 xor r0,r0,r0 # set r0 = 0. Used in the addze
294 #sqr_add_c(a,0,c1,c2,c3)
297 $UMULH r10,r5,r5 #in first iteration. No need
298 #to add since c1=c2=c3=0.
299 # Note c3(r11) is NOT set to 0
302 $ST r9,`0*$BNSZ`(r3) # r[0]=c1;
303 # sqr_add_c2(a,1,0,c2,c3,c1);
308 addc r7,r7,r7 # compute (r7,r8)=2*(r7,r8)
310 addze r9,r0 # catch carry if any.
311 # r9= r0(=0) and carry
313 addc r10,r7,r10 # now add to temp result.
314 addze r11,r8 # r8 added to r11 which is 0
317 $ST r10,`1*$BNSZ`(r3) #r[1]=c2;
318 #sqr_add_c(a,1,c3,c1,c2)
324 #sqr_add_c2(a,2,0,c3,c1,c2)
336 $ST r11,`2*$BNSZ`(r3) #r[2]=c3
337 #sqr_add_c2(a,3,0,c1,c2,c3);
348 #sqr_add_c2(a,2,1,c1,c2,c3);
360 $ST r9,`3*$BNSZ`(r3) #r[3]=c1
361 #sqr_add_c(a,2,c2,c3,c1);
367 #sqr_add_c2(a,3,1,c2,c3,c1);
378 $ST r10,`4*$BNSZ`(r3) #r[4]=c2
379 #sqr_add_c2(a,3,2,c3,c1,c2);
390 $ST r11,`5*$BNSZ`(r3) #r[5] = c3
391 #sqr_add_c(a,3,c1,c2,c3);
397 $ST r9,`6*$BNSZ`(r3) #r[6]=c1
398 $ST r10,`7*$BNSZ`(r3) #r[7]=c2
401 .byte 0,12,0x14,0,0,0,2,0
403 .size .bn_sqr_comba4,.-.bn_sqr_comba4
406 # NOTE: The following label name should be changed to
407 # "bn_sqr_comba8" i.e. remove the first dot
408 # for the gcc compiler. This should be automatically
415 # This is an optimized version of the bn_sqr_comba8 routine.
416 # Tightly uses the adde instruction
419 # void bn_sqr_comba8(BN_ULONG *r, BN_ULONG *a)
423 # Freely use registers r5,r6,r7,r8,r9,r10,r11 as follows:
425 # r5,r6 are the two BN_ULONGs being multiplied.
426 # r7,r8 are the results of the 32x32 giving 64 bit multiply.
427 # r9,r10, r11 are the equivalents of c1,c2, c3.
429 # Possible optimization of loading all 8 longs of a into registers
430 # doesn't provide any speedup
433 xor r0,r0,r0 #set r0 = 0.Used in addze
436 #sqr_add_c(a,0,c1,c2,c3);
438 $UMULL r9,r5,r5 #1st iteration: no carries.
440 $ST r9,`0*$BNSZ`(r3) # r[0]=c1;
441 #sqr_add_c2(a,1,0,c2,c3,c1);
446 addc r10,r7,r10 #add the two register number
447 adde r11,r8,r0 # (r8,r7) to the three register
448 addze r9,r0 # number (r9,r11,r10).NOTE:r0=0
450 addc r10,r7,r10 #add the two register number
451 adde r11,r8,r11 # (r8,r7) to the three register
452 addze r9,r9 # number (r9,r11,r10).
454 $ST r10,`1*$BNSZ`(r3) # r[1]=c2
456 #sqr_add_c(a,1,c3,c1,c2);
462 #sqr_add_c2(a,2,0,c3,c1,c2);
475 $ST r11,`2*$BNSZ`(r3) #r[2]=c3
476 #sqr_add_c2(a,3,0,c1,c2,c3);
477 $LD r6,`3*$BNSZ`(r4) #r6 = a[3]. r5 is already a[0].
488 #sqr_add_c2(a,2,1,c1,c2,c3);
502 $ST r9,`3*$BNSZ`(r3) #r[3]=c1;
503 #sqr_add_c(a,2,c2,c3,c1);
510 #sqr_add_c2(a,3,1,c2,c3,c1);
522 #sqr_add_c2(a,4,0,c2,c3,c1);
535 $ST r10,`4*$BNSZ`(r3) #r[4]=c2;
536 #sqr_add_c2(a,5,0,c3,c1,c2);
548 #sqr_add_c2(a,4,1,c3,c1,c2);
561 #sqr_add_c2(a,3,2,c3,c1,c2);
574 $ST r11,`5*$BNSZ`(r3) #r[5]=c3;
575 #sqr_add_c(a,3,c1,c2,c3);
581 #sqr_add_c2(a,4,2,c1,c2,c3);
593 #sqr_add_c2(a,5,1,c1,c2,c3);
606 #sqr_add_c2(a,6,0,c1,c2,c3);
617 $ST r9,`6*$BNSZ`(r3) #r[6]=c1;
618 #sqr_add_c2(a,7,0,c2,c3,c1);
629 #sqr_add_c2(a,6,1,c2,c3,c1);
641 #sqr_add_c2(a,5,2,c2,c3,c1);
652 #sqr_add_c2(a,4,3,c2,c3,c1);
664 $ST r10,`7*$BNSZ`(r3) #r[7]=c2;
665 #sqr_add_c(a,4,c3,c1,c2);
671 #sqr_add_c2(a,5,3,c3,c1,c2);
681 #sqr_add_c2(a,6,2,c3,c1,c2);
693 #sqr_add_c2(a,7,1,c3,c1,c2);
704 $ST r11,`8*$BNSZ`(r3) #r[8]=c3;
705 #sqr_add_c2(a,7,2,c1,c2,c3);
716 #sqr_add_c2(a,6,3,c1,c2,c3);
727 #sqr_add_c2(a,5,4,c1,c2,c3);
738 $ST r9,`9*$BNSZ`(r3) #r[9]=c1;
739 #sqr_add_c(a,5,c2,c3,c1);
745 #sqr_add_c2(a,6,4,c2,c3,c1);
755 #sqr_add_c2(a,7,3,c2,c3,c1);
766 $ST r10,`10*$BNSZ`(r3) #r[10]=c2;
767 #sqr_add_c2(a,7,4,c3,c1,c2);
777 #sqr_add_c2(a,6,5,c3,c1,c2);
788 $ST r11,`11*$BNSZ`(r3) #r[11]=c3;
789 #sqr_add_c(a,6,c1,c2,c3);
795 #sqr_add_c2(a,7,5,c1,c2,c3)
805 $ST r9,`12*$BNSZ`(r3) #r[12]=c1;
807 #sqr_add_c2(a,7,6,c2,c3,c1)
817 $ST r10,`13*$BNSZ`(r3) #r[13]=c2;
818 #sqr_add_c(a,7,c3,c1,c2);
823 $ST r11,`14*$BNSZ`(r3) #r[14]=c3;
824 $ST r9, `15*$BNSZ`(r3) #r[15]=c1;
829 .byte 0,12,0x14,0,0,0,2,0
831 .size .bn_sqr_comba8,.-.bn_sqr_comba8
834 # NOTE: The following label name should be changed to
835 # "bn_mul_comba4" i.e. remove the first dot
836 # for the gcc compiler. This should be automatically
843 # This is an optimized version of the bn_mul_comba4 routine.
845 # void bn_mul_comba4(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
849 # r6, r7 are the 2 BN_ULONGs being multiplied.
850 # r8, r9 are the results of the 32x32 giving 64 multiply.
851 # r10, r11, r12 are the equivalents of c1, c2, and c3.
853 xor r0,r0,r0 #r0=0. Used in addze below.
854 #mul_add_c(a[0],b[0],c1,c2,c3);
859 $ST r10,`0*$BNSZ`(r3) #r[0]=c1
860 #mul_add_c(a[0],b[1],c2,c3,c1);
867 #mul_add_c(a[1],b[0],c2,c3,c1);
868 $LD r6, `1*$BNSZ`(r4)
869 $LD r7, `0*$BNSZ`(r5)
875 $ST r11,`1*$BNSZ`(r3) #r[1]=c2
876 #mul_add_c(a[2],b[0],c3,c1,c2);
883 #mul_add_c(a[1],b[1],c3,c1,c2);
891 #mul_add_c(a[0],b[2],c3,c1,c2);
899 $ST r12,`2*$BNSZ`(r3) #r[2]=c3
900 #mul_add_c(a[0],b[3],c1,c2,c3);
907 #mul_add_c(a[1],b[2],c1,c2,c3);
915 #mul_add_c(a[2],b[1],c1,c2,c3);
923 #mul_add_c(a[3],b[0],c1,c2,c3);
931 $ST r10,`3*$BNSZ`(r3) #r[3]=c1
932 #mul_add_c(a[3],b[1],c2,c3,c1);
939 #mul_add_c(a[2],b[2],c2,c3,c1);
947 #mul_add_c(a[1],b[3],c2,c3,c1);
955 $ST r11,`4*$BNSZ`(r3) #r[4]=c2
956 #mul_add_c(a[2],b[3],c3,c1,c2);
963 #mul_add_c(a[3],b[2],c3,c1,c2);
971 $ST r12,`5*$BNSZ`(r3) #r[5]=c3
972 #mul_add_c(a[3],b[3],c1,c2,c3);
979 $ST r10,`6*$BNSZ`(r3) #r[6]=c1
980 $ST r11,`7*$BNSZ`(r3) #r[7]=c2
983 .byte 0,12,0x14,0,0,0,3,0
985 .size .bn_mul_comba4,.-.bn_mul_comba4
988 # NOTE: The following label name should be changed to
989 # "bn_mul_comba8" i.e. remove the first dot
990 # for the gcc compiler. This should be automatically
997 # Optimized version of the bn_mul_comba8 routine.
999 # void bn_mul_comba8(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b)
1003 # r6, r7 are the 2 BN_ULONGs being multiplied.
1004 # r8, r9 are the results of the 32x32 giving 64 multiply.
1005 # r10, r11, r12 are the equivalents of c1, c2, and c3.
1007 xor r0,r0,r0 #r0=0. Used in addze below.
1009 #mul_add_c(a[0],b[0],c1,c2,c3);
1010 $LD r6,`0*$BNSZ`(r4) #a[0]
1011 $LD r7,`0*$BNSZ`(r5) #b[0]
1014 $ST r10,`0*$BNSZ`(r3) #r[0]=c1;
1015 #mul_add_c(a[0],b[1],c2,c3,c1);
1016 $LD r7,`1*$BNSZ`(r5)
1020 addze r12,r9 # since we didn't set r12 to zero before.
1022 #mul_add_c(a[1],b[0],c2,c3,c1);
1023 $LD r6,`1*$BNSZ`(r4)
1024 $LD r7,`0*$BNSZ`(r5)
1030 $ST r11,`1*$BNSZ`(r3) #r[1]=c2;
1031 #mul_add_c(a[2],b[0],c3,c1,c2);
1032 $LD r6,`2*$BNSZ`(r4)
1038 #mul_add_c(a[1],b[1],c3,c1,c2);
1039 $LD r6,`1*$BNSZ`(r4)
1040 $LD r7,`1*$BNSZ`(r5)
1046 #mul_add_c(a[0],b[2],c3,c1,c2);
1047 $LD r6,`0*$BNSZ`(r4)
1048 $LD r7,`2*$BNSZ`(r5)
1054 $ST r12,`2*$BNSZ`(r3) #r[2]=c3;
1055 #mul_add_c(a[0],b[3],c1,c2,c3);
1056 $LD r7,`3*$BNSZ`(r5)
1062 #mul_add_c(a[1],b[2],c1,c2,c3);
1063 $LD r6,`1*$BNSZ`(r4)
1064 $LD r7,`2*$BNSZ`(r5)
1071 #mul_add_c(a[2],b[1],c1,c2,c3);
1072 $LD r6,`2*$BNSZ`(r4)
1073 $LD r7,`1*$BNSZ`(r5)
1079 #mul_add_c(a[3],b[0],c1,c2,c3);
1080 $LD r6,`3*$BNSZ`(r4)
1081 $LD r7,`0*$BNSZ`(r5)
1087 $ST r10,`3*$BNSZ`(r3) #r[3]=c1;
1088 #mul_add_c(a[4],b[0],c2,c3,c1);
1089 $LD r6,`4*$BNSZ`(r4)
1095 #mul_add_c(a[3],b[1],c2,c3,c1);
1096 $LD r6,`3*$BNSZ`(r4)
1097 $LD r7,`1*$BNSZ`(r5)
1103 #mul_add_c(a[2],b[2],c2,c3,c1);
1104 $LD r6,`2*$BNSZ`(r4)
1105 $LD r7,`2*$BNSZ`(r5)
1111 #mul_add_c(a[1],b[3],c2,c3,c1);
1112 $LD r6,`1*$BNSZ`(r4)
1113 $LD r7,`3*$BNSZ`(r5)
1119 #mul_add_c(a[0],b[4],c2,c3,c1);
1120 $LD r6,`0*$BNSZ`(r4)
1121 $LD r7,`4*$BNSZ`(r5)
1127 $ST r11,`4*$BNSZ`(r3) #r[4]=c2;
1128 #mul_add_c(a[0],b[5],c3,c1,c2);
1129 $LD r7,`5*$BNSZ`(r5)
1135 #mul_add_c(a[1],b[4],c3,c1,c2);
1136 $LD r6,`1*$BNSZ`(r4)
1137 $LD r7,`4*$BNSZ`(r5)
1143 #mul_add_c(a[2],b[3],c3,c1,c2);
1144 $LD r6,`2*$BNSZ`(r4)
1145 $LD r7,`3*$BNSZ`(r5)
1151 #mul_add_c(a[3],b[2],c3,c1,c2);
1152 $LD r6,`3*$BNSZ`(r4)
1153 $LD r7,`2*$BNSZ`(r5)
1159 #mul_add_c(a[4],b[1],c3,c1,c2);
1160 $LD r6,`4*$BNSZ`(r4)
1161 $LD r7,`1*$BNSZ`(r5)
1167 #mul_add_c(a[5],b[0],c3,c1,c2);
1168 $LD r6,`5*$BNSZ`(r4)
1169 $LD r7,`0*$BNSZ`(r5)
1175 $ST r12,`5*$BNSZ`(r3) #r[5]=c3;
1176 #mul_add_c(a[6],b[0],c1,c2,c3);
1177 $LD r6,`6*$BNSZ`(r4)
1183 #mul_add_c(a[5],b[1],c1,c2,c3);
1184 $LD r6,`5*$BNSZ`(r4)
1185 $LD r7,`1*$BNSZ`(r5)
1191 #mul_add_c(a[4],b[2],c1,c2,c3);
1192 $LD r6,`4*$BNSZ`(r4)
1193 $LD r7,`2*$BNSZ`(r5)
1199 #mul_add_c(a[3],b[3],c1,c2,c3);
1200 $LD r6,`3*$BNSZ`(r4)
1201 $LD r7,`3*$BNSZ`(r5)
1207 #mul_add_c(a[2],b[4],c1,c2,c3);
1208 $LD r6,`2*$BNSZ`(r4)
1209 $LD r7,`4*$BNSZ`(r5)
1215 #mul_add_c(a[1],b[5],c1,c2,c3);
1216 $LD r6,`1*$BNSZ`(r4)
1217 $LD r7,`5*$BNSZ`(r5)
1223 #mul_add_c(a[0],b[6],c1,c2,c3);
1224 $LD r6,`0*$BNSZ`(r4)
1225 $LD r7,`6*$BNSZ`(r5)
1231 $ST r10,`6*$BNSZ`(r3) #r[6]=c1;
1232 #mul_add_c(a[0],b[7],c2,c3,c1);
1233 $LD r7,`7*$BNSZ`(r5)
1239 #mul_add_c(a[1],b[6],c2,c3,c1);
1240 $LD r6,`1*$BNSZ`(r4)
1241 $LD r7,`6*$BNSZ`(r5)
1247 #mul_add_c(a[2],b[5],c2,c3,c1);
1248 $LD r6,`2*$BNSZ`(r4)
1249 $LD r7,`5*$BNSZ`(r5)
1255 #mul_add_c(a[3],b[4],c2,c3,c1);
1256 $LD r6,`3*$BNSZ`(r4)
1257 $LD r7,`4*$BNSZ`(r5)
1263 #mul_add_c(a[4],b[3],c2,c3,c1);
1264 $LD r6,`4*$BNSZ`(r4)
1265 $LD r7,`3*$BNSZ`(r5)
1271 #mul_add_c(a[5],b[2],c2,c3,c1);
1272 $LD r6,`5*$BNSZ`(r4)
1273 $LD r7,`2*$BNSZ`(r5)
1279 #mul_add_c(a[6],b[1],c2,c3,c1);
1280 $LD r6,`6*$BNSZ`(r4)
1281 $LD r7,`1*$BNSZ`(r5)
1287 #mul_add_c(a[7],b[0],c2,c3,c1);
1288 $LD r6,`7*$BNSZ`(r4)
1289 $LD r7,`0*$BNSZ`(r5)
1295 $ST r11,`7*$BNSZ`(r3) #r[7]=c2;
1296 #mul_add_c(a[7],b[1],c3,c1,c2);
1297 $LD r7,`1*$BNSZ`(r5)
1303 #mul_add_c(a[6],b[2],c3,c1,c2);
1304 $LD r6,`6*$BNSZ`(r4)
1305 $LD r7,`2*$BNSZ`(r5)
1311 #mul_add_c(a[5],b[3],c3,c1,c2);
1312 $LD r6,`5*$BNSZ`(r4)
1313 $LD r7,`3*$BNSZ`(r5)
1319 #mul_add_c(a[4],b[4],c3,c1,c2);
1320 $LD r6,`4*$BNSZ`(r4)
1321 $LD r7,`4*$BNSZ`(r5)
1327 #mul_add_c(a[3],b[5],c3,c1,c2);
1328 $LD r6,`3*$BNSZ`(r4)
1329 $LD r7,`5*$BNSZ`(r5)
1335 #mul_add_c(a[2],b[6],c3,c1,c2);
1336 $LD r6,`2*$BNSZ`(r4)
1337 $LD r7,`6*$BNSZ`(r5)
1343 #mul_add_c(a[1],b[7],c3,c1,c2);
1344 $LD r6,`1*$BNSZ`(r4)
1345 $LD r7,`7*$BNSZ`(r5)
1351 $ST r12,`8*$BNSZ`(r3) #r[8]=c3;
1352 #mul_add_c(a[2],b[7],c1,c2,c3);
1353 $LD r6,`2*$BNSZ`(r4)
1359 #mul_add_c(a[3],b[6],c1,c2,c3);
1360 $LD r6,`3*$BNSZ`(r4)
1361 $LD r7,`6*$BNSZ`(r5)
1367 #mul_add_c(a[4],b[5],c1,c2,c3);
1368 $LD r6,`4*$BNSZ`(r4)
1369 $LD r7,`5*$BNSZ`(r5)
1375 #mul_add_c(a[5],b[4],c1,c2,c3);
1376 $LD r6,`5*$BNSZ`(r4)
1377 $LD r7,`4*$BNSZ`(r5)
1383 #mul_add_c(a[6],b[3],c1,c2,c3);
1384 $LD r6,`6*$BNSZ`(r4)
1385 $LD r7,`3*$BNSZ`(r5)
1391 #mul_add_c(a[7],b[2],c1,c2,c3);
1392 $LD r6,`7*$BNSZ`(r4)
1393 $LD r7,`2*$BNSZ`(r5)
1399 $ST r10,`9*$BNSZ`(r3) #r[9]=c1;
1400 #mul_add_c(a[7],b[3],c2,c3,c1);
1401 $LD r7,`3*$BNSZ`(r5)
1407 #mul_add_c(a[6],b[4],c2,c3,c1);
1408 $LD r6,`6*$BNSZ`(r4)
1409 $LD r7,`4*$BNSZ`(r5)
1415 #mul_add_c(a[5],b[5],c2,c3,c1);
1416 $LD r6,`5*$BNSZ`(r4)
1417 $LD r7,`5*$BNSZ`(r5)
1423 #mul_add_c(a[4],b[6],c2,c3,c1);
1424 $LD r6,`4*$BNSZ`(r4)
1425 $LD r7,`6*$BNSZ`(r5)
1431 #mul_add_c(a[3],b[7],c2,c3,c1);
1432 $LD r6,`3*$BNSZ`(r4)
1433 $LD r7,`7*$BNSZ`(r5)
1439 $ST r11,`10*$BNSZ`(r3) #r[10]=c2;
1440 #mul_add_c(a[4],b[7],c3,c1,c2);
1441 $LD r6,`4*$BNSZ`(r4)
1447 #mul_add_c(a[5],b[6],c3,c1,c2);
1448 $LD r6,`5*$BNSZ`(r4)
1449 $LD r7,`6*$BNSZ`(r5)
1455 #mul_add_c(a[6],b[5],c3,c1,c2);
1456 $LD r6,`6*$BNSZ`(r4)
1457 $LD r7,`5*$BNSZ`(r5)
1463 #mul_add_c(a[7],b[4],c3,c1,c2);
1464 $LD r6,`7*$BNSZ`(r4)
1465 $LD r7,`4*$BNSZ`(r5)
1471 $ST r12,`11*$BNSZ`(r3) #r[11]=c3;
1472 #mul_add_c(a[7],b[5],c1,c2,c3);
1473 $LD r7,`5*$BNSZ`(r5)
1479 #mul_add_c(a[6],b[6],c1,c2,c3);
1480 $LD r6,`6*$BNSZ`(r4)
1481 $LD r7,`6*$BNSZ`(r5)
1487 #mul_add_c(a[5],b[7],c1,c2,c3);
1488 $LD r6,`5*$BNSZ`(r4)
1489 $LD r7,`7*$BNSZ`(r5)
1495 $ST r10,`12*$BNSZ`(r3) #r[12]=c1;
1496 #mul_add_c(a[6],b[7],c2,c3,c1);
1497 $LD r6,`6*$BNSZ`(r4)
1503 #mul_add_c(a[7],b[6],c2,c3,c1);
1504 $LD r6,`7*$BNSZ`(r4)
1505 $LD r7,`6*$BNSZ`(r5)
1511 $ST r11,`13*$BNSZ`(r3) #r[13]=c2;
1512 #mul_add_c(a[7],b[7],c3,c1,c2);
1513 $LD r7,`7*$BNSZ`(r5)
1518 $ST r12,`14*$BNSZ`(r3) #r[14]=c3;
1519 $ST r10,`15*$BNSZ`(r3) #r[15]=c1;
1522 .byte 0,12,0x14,0,0,0,3,0
1524 .size .bn_mul_comba8,.-.bn_mul_comba8
1527 # NOTE: The following label name should be changed to
1528 # "bn_sub_words" i.e. remove the first dot
1529 # for the gcc compiler. This should be automatically
1536 # Handcoded version of bn_sub_words
1538 #BN_ULONG bn_sub_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
1545 # Note: No loop unrolling done since this is not a performance
1548 xor r0,r0,r0 #set r0 = 0
1550 # check for r6 = 0 AND set carry bit.
1552 subfc. r7,r0,r6 # If r6 is 0 then result is 0.
1553 # if r6 > 0 then result !=0
1554 # In either case carry bit is set.
1555 beq Lppcasm_sub_adios
1560 Lppcasm_sub_mainloop:
1563 subfe r6,r8,r7 # r6 = r7+carry bit + onescomplement(r8)
1564 # if carry = 1 this is r7-r8. Else it
1565 # is r7-r8 -1 as we need.
1567 bdnz Lppcasm_sub_mainloop
1569 subfze r3,r0 # if carry bit is set then r3 = 0 else -1
1570 andi. r3,r3,1 # keep only last bit.
1573 .byte 0,12,0x14,0,0,0,4,0
1575 .size .bn_sub_words,.-.bn_sub_words
1578 # NOTE: The following label name should be changed to
1579 # "bn_add_words" i.e. remove the first dot
1580 # for the gcc compiler. This should be automatically
1587 # Handcoded version of bn_add_words
1589 #BN_ULONG bn_add_words(BN_ULONG *r, BN_ULONG *a, BN_ULONG *b, int n)
1596 # Note: No loop unrolling done since this is not a performance
1601 # check for r6 = 0. Is this needed?
1603 addic. r6,r6,0 #test r6 and clear carry bit.
1604 beq Lppcasm_add_adios
1609 Lppcasm_add_mainloop:
1614 bdnz Lppcasm_add_mainloop
1616 addze r3,r0 #return carry bit.
1619 .byte 0,12,0x14,0,0,0,4,0
1621 .size .bn_add_words,.-.bn_add_words
1624 # NOTE: The following label name should be changed to
1625 # "bn_div_words" i.e. remove the first dot
1626 # for the gcc compiler. This should be automatically
1633 # This is a cleaned up version of code generated by
1634 # the AIX compiler. The only optimization is to use
1635 # the PPC instruction to count leading zeros instead
1636 # of call to num_bits_word. Since this was compiled
1637 # only at level -O2 we can possibly squeeze it more?
1643 $UCMPI 0,r5,0 # compare r5 and 0
1644 bne Lppcasm_div1 # proceed if d!=0
1645 li r3,-1 # d=0 return -1
1650 $CNTLZ. r7,r5 #r7 = num leading 0s in d.
1651 beq Lppcasm_div2 #proceed if no leading zeros
1652 subf r8,r7,r8 #r8 = BN_num_bits_word(d)
1653 $SHR. r9,r3,r8 #are there any bits above r8'th?
1654 $TR 16,r9,r0 #if there're, signal to dump core...
1656 $UCMP 0,r3,r5 #h>=d?
1657 blt Lppcasm_div3 #goto Lppcasm_div3 if not
1658 subf r3,r5,r3 #h-=d ;
1659 Lppcasm_div3: #r7 = BN_BITS2-i. so r7=i
1660 cmpi 0,0,r7,0 # is (i == 0)?
1662 $SHL r3,r3,r7 # h = (h<< i)
1663 $SHR r8,r4,r8 # r8 = (l >> BN_BITS2 -i)
1664 $SHL r5,r5,r7 # d<<=i
1665 or r3,r3,r8 # h = (h<<i)|(l>>(BN_BITS2-i))
1666 $SHL r4,r4,r7 # l <<=i
1668 $SHRI r9,r5,`$BITS/2` # r9 = dh
1669 # dl will be computed when needed
1670 # as it saves registers.
1672 mtctr r6 #counter will be in count.
1673 Lppcasm_divouterloop:
1674 $SHRI r8,r3,`$BITS/2` #r8 = (h>>BN_BITS4)
1675 $SHRI r11,r4,`$BITS/2` #r11= (l&BN_MASK2h)>>BN_BITS4
1676 # compute here for innerloop.
1677 $UCMP 0,r8,r9 # is (h>>BN_BITS4)==dh
1678 bne Lppcasm_div5 # goto Lppcasm_div5 if not
1681 $CLRU r8,r8,`$BITS/2` #q = BN_MASK2l
1684 $UDIV r8,r3,r9 #q = h/dh
1686 $UMULL r12,r9,r8 #th = q*dh
1687 $CLRU r10,r5,`$BITS/2` #r10=dl
1688 $UMULL r6,r8,r10 #tl = q*dl
1690 Lppcasm_divinnerloop:
1691 subf r10,r12,r3 #t = h -th
1692 $SHRI r7,r10,`$BITS/2` #r7= (t &BN_MASK2H), sort of...
1693 addic. r7,r7,0 #test if r7 == 0. used below.
1694 # now want to compute
1695 # r7 = (t<<BN_BITS4)|((l&BN_MASK2h)>>BN_BITS4)
1696 # the following 2 instructions do that
1697 $SHLI r7,r10,`$BITS/2` # r7 = (t<<BN_BITS4)
1698 or r7,r7,r11 # r7|=((l&BN_MASK2h)>>BN_BITS4)
1699 $UCMP cr1,r6,r7 # compare (tl <= r7)
1700 bne Lppcasm_divinnerexit
1701 ble cr1,Lppcasm_divinnerexit
1703 subf r12,r9,r12 #th -=dh
1704 $CLRU r10,r5,`$BITS/2` #r10=dl. t is no longer needed in loop.
1705 subf r6,r10,r6 #tl -=dl
1706 b Lppcasm_divinnerloop
1707 Lppcasm_divinnerexit:
1708 $SHRI r10,r6,`$BITS/2` #t=(tl>>BN_BITS4)
1709 $SHLI r11,r6,`$BITS/2` #tl=(tl<<BN_BITS4)&BN_MASK2h;
1710 $UCMP cr1,r4,r11 # compare l and tl
1711 add r12,r12,r10 # th+=t
1712 bge cr1,Lppcasm_div7 # if (l>=tl) goto Lppcasm_div7
1713 addi r12,r12,1 # th++
1715 subf r11,r11,r4 #r11=l-tl
1716 $UCMP cr1,r3,r12 #compare h and th
1717 bge cr1,Lppcasm_div8 #if (h>=th) goto Lppcasm_div8
1721 subf r12,r12,r3 #r12 = h-th
1722 $SHLI r4,r11,`$BITS/2` #l=(l&BN_MASK2l)<<BN_BITS4
1724 # h = ((h<<BN_BITS4)|(l>>BN_BITS4))&BN_MASK2
1725 # the following 2 instructions will do this.
1726 $INSR r11,r12,`$BITS/2`,`$BITS/2` # r11 is the value we want rotated $BITS/2.
1727 $ROTL r3,r11,`$BITS/2` # rotate by $BITS/2 and store in r3
1728 bdz Lppcasm_div9 #if (count==0) break ;
1729 $SHLI r0,r8,`$BITS/2` #ret =q<<BN_BITS4
1730 b Lppcasm_divouterloop
1735 .byte 0,12,0x14,0,0,0,3,0
1737 .size .bn_div_words,.-.bn_div_words
1740 # NOTE: The following label name should be changed to
1741 # "bn_sqr_words" i.e. remove the first dot
1742 # for the gcc compiler. This should be automatically
1748 # Optimized version of bn_sqr_words
1750 # void bn_sqr_words(BN_ULONG *r, BN_ULONG *a, int n)
1759 # No unrolling done here. Not performance critical.
1761 addic. r5,r5,0 #test r5.
1762 beq Lppcasm_sqr_adios
1766 Lppcasm_sqr_mainloop:
1767 #sqr(r[0],r[1],a[0]);
1773 bdnz Lppcasm_sqr_mainloop
1777 .byte 0,12,0x14,0,0,0,3,0
1779 .size .bn_sqr_words,.-.bn_sqr_words
1782 # NOTE: The following label name should be changed to
1783 # "bn_mul_words" i.e. remove the first dot
1784 # for the gcc compiler. This should be automatically
1791 # BN_ULONG bn_mul_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
1798 xor r12,r12,r12 # used for carry
1799 rlwinm. r7,r5,30,2,31 # num >> 2
1803 #mul(rp[0],ap[0],w,c1);
1804 $LD r8,`0*$BNSZ`(r4)
1808 #addze r10,r10 #carry is NOT ignored.
1809 #will be taken care of
1810 #in second spin below
1812 $ST r9,`0*$BNSZ`(r3)
1813 #mul(rp[1],ap[1],w,c1);
1814 $LD r8,`1*$BNSZ`(r4)
1819 $ST r11,`1*$BNSZ`(r3)
1820 #mul(rp[2],ap[2],w,c1);
1821 $LD r8,`2*$BNSZ`(r4)
1826 $ST r9,`2*$BNSZ`(r3)
1827 #mul_add(rp[3],ap[3],w,c1);
1828 $LD r8,`3*$BNSZ`(r4)
1832 addze r12,r12 #this spin we collect carry into
1834 $ST r11,`3*$BNSZ`(r3)
1836 addi r3,r3,`4*$BNSZ`
1837 addi r4,r4,`4*$BNSZ`
1838 bdnz Lppcasm_mw_LOOP
1843 #mul(rp[0],ap[0],w,c1);
1844 $LD r8,`0*$BNSZ`(r4)
1849 $ST r9,`0*$BNSZ`(r3)
1857 #mul(rp[1],ap[1],w,c1);
1858 $LD r8,`1*$BNSZ`(r4)
1863 $ST r9,`1*$BNSZ`(r3)
1870 #mul_add(rp[2],ap[2],w,c1);
1871 $LD r8,`2*$BNSZ`(r4)
1876 $ST r9,`2*$BNSZ`(r3)
1883 .byte 0,12,0x14,0,0,0,4,0
1885 .size .bn_mul_words,.-.bn_mul_words
1888 # NOTE: The following label name should be changed to
1889 # "bn_mul_add_words" i.e. remove the first dot
1890 # for the gcc compiler. This should be automatically
1897 # BN_ULONG bn_mul_add_words(BN_ULONG *rp, BN_ULONG *ap, int num, BN_ULONG w)
1904 # empirical evidence suggests that unrolled version performs best!!
1906 xor r0,r0,r0 #r0 = 0
1907 xor r12,r12,r12 #r12 = 0 . used for carry
1908 rlwinm. r7,r5,30,2,31 # num >> 2
1909 beq Lppcasm_maw_leftover # if (num < 4) go LPPCASM_maw_leftover
1911 Lppcasm_maw_mainloop:
1912 #mul_add(rp[0],ap[0],w,c1);
1913 $LD r8,`0*$BNSZ`(r4)
1914 $LD r11,`0*$BNSZ`(r3)
1917 addc r9,r9,r12 #r12 is carry.
1921 #the above instruction addze
1922 #is NOT needed. Carry will NOT
1923 #be ignored. It's not affected
1924 #by multiply and will be collected
1926 $ST r9,`0*$BNSZ`(r3)
1928 #mul_add(rp[1],ap[1],w,c1);
1929 $LD r8,`1*$BNSZ`(r4)
1930 $LD r9,`1*$BNSZ`(r3)
1933 adde r11,r11,r10 #r10 is carry.
1937 $ST r11,`1*$BNSZ`(r3)
1939 #mul_add(rp[2],ap[2],w,c1);
1940 $LD r8,`2*$BNSZ`(r4)
1942 $LD r11,`2*$BNSZ`(r3)
1948 $ST r9,`2*$BNSZ`(r3)
1950 #mul_add(rp[3],ap[3],w,c1);
1951 $LD r8,`3*$BNSZ`(r4)
1953 $LD r9,`3*$BNSZ`(r3)
1959 $ST r11,`3*$BNSZ`(r3)
1960 addi r3,r3,`4*$BNSZ`
1961 addi r4,r4,`4*$BNSZ`
1962 bdnz Lppcasm_maw_mainloop
1964 Lppcasm_maw_leftover:
1966 beq Lppcasm_maw_adios
1969 #mul_add(rp[0],ap[0],w,c1);
1981 bdz Lppcasm_maw_adios
1982 #mul_add(rp[1],ap[1],w,c1);
1993 bdz Lppcasm_maw_adios
1994 #mul_add(rp[2],ap[2],w,c1);
2009 .byte 0,12,0x14,0,0,0,4,0
2011 .size .bn_mul_add_words,.-.bn_mul_add_words
2014 $data =~ s/\`([^\`]*)\`/eval $1/gem;
2016 close STDOUT or die "error closing STDOUT: $!";