2 # Copyright 2011-2018 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
10 # ====================================================================
11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12 # project. The module is, however, dual licensed under OpenSSL and
13 # CRYPTOGAMS licenses depending on where you obtain it. For further
14 # details see http://www.openssl.org/~appro/cryptogams/.
15 # ====================================================================
19 # Companion to x86_64-mont.pl that optimizes cache-timing attack
20 # countermeasures. The subroutines are produced by replacing bp[i]
21 # references in their x86_64-mont.pl counterparts with cache-neutral
22 # references to powers table computed in BN_mod_exp_mont_consttime.
23 # In addition subroutine that scatters elements of the powers table
24 # is implemented, so that scatter-/gathering can be tuned without
25 # bn_exp.c modifications.
29 # Add MULX/AD*X code paths and additional interfaces to optimize for
30 # branch prediction unit. For input lengths that are multiples of 8
31 # the np argument is not just modulus value, but one interleaved
32 # with 0. This is to optimize post-condition...
34 # $output is the last argument if it looks like a file (it has an extension)
35 # $flavour is the first argument if it doesn't look like a file
36 $output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef;
37 $flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef;
39 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
41 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
42 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
43 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
44 die "can't locate x86_64-xlate.pl";
46 open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""
47 or die "can't call $xlate: $!";
50 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
51 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
55 if (!$addx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
56 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
60 if (!$addx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
61 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
65 if (!$addx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9])\.([0-9]+)/) {
66 my $ver = $2 + $3/100.0; # 3.1->3.01, 3.10->3.10
70 # int bn_mul_mont_gather5(
71 $rp="%rdi"; # BN_ULONG *rp,
72 $ap="%rsi"; # const BN_ULONG *ap,
73 $bp="%rdx"; # const BN_ULONG *bp,
74 $np="%rcx"; # const BN_ULONG *np,
75 $n0="%r8"; # const BN_ULONG *n0,
76 $num="%r9"; # int num,
77 # int idx); # 0 to 2^5-1, "index" in $bp holding
78 # pre-computed powers of a', interlaced
79 # in such manner that b[0] is $bp[idx],
80 # b[1] is [2^5+idx], etc.
92 .extern OPENSSL_ia32cap_P
94 .globl bn_mul_mont_gather5
95 .type bn_mul_mont_gather5,\@function,6
101 .cfi_def_cfa_register %rax
105 $code.=<<___ if ($addx);
106 mov OPENSSL_ia32cap_P+8(%rip),%r11d
113 movd `($win64?56:8)`(%rsp),%xmm5 # load 7th argument
129 lea -280(%rsp,$num,8),%r10 # future alloca(8*(num+2)+256+8)
130 neg $num # restore $num
131 and \$-1024,%r10 # minimize TLB usage
133 # An OS-agnostic version of __chkstk.
135 # Some OSes (Windows) insist on stack being "wired" to
136 # physical memory in strictly sequential manner, i.e. if stack
137 # allocation spans two pages, then reference to farmost one can
138 # be punishable by SEGV. But page walking can do good even on
139 # other OSes, because it guarantees that villain thread hits
140 # the guard page before it can make damage to innocent one...
147 jmp .Lmul_page_walk_done
154 .Lmul_page_walk_done:
157 mov %rax,8(%rsp,$num,8) # tp[num+1]=%rsp
158 .cfi_cfa_expression %rsp+8,$num,8,mul,plus,deref,+8
161 lea 128($bp),%r12 # reassign $bp (+size optimization)
164 $STRIDE=2**5*8; # 5 is "window size"
165 $N=$STRIDE/4; # should match cache line size
167 movdqa 0(%r10),%xmm0 # 00000001000000010000000000000000
168 movdqa 16(%r10),%xmm1 # 00000002000000020000000200000002
169 lea 24-112(%rsp,$num,8),%r10# place the mask after tp[num+3] (+ICache optimization)
172 pshufd \$0,%xmm5,%xmm5 # broadcast index
176 ########################################################################
177 # calculate mask by comparing 0..31 to index and save result to stack
181 pcmpeqd %xmm5,%xmm0 # compare to 1,0
185 for($k=0;$k<$STRIDE/16-4;$k+=4) {
188 pcmpeqd %xmm5,%xmm1 # compare to 3,2
189 movdqa %xmm0,`16*($k+0)+112`(%r10)
193 pcmpeqd %xmm5,%xmm2 # compare to 5,4
194 movdqa %xmm1,`16*($k+1)+112`(%r10)
198 pcmpeqd %xmm5,%xmm3 # compare to 7,6
199 movdqa %xmm2,`16*($k+2)+112`(%r10)
204 movdqa %xmm3,`16*($k+3)+112`(%r10)
208 $code.=<<___; # last iteration can be optimized
211 movdqa %xmm0,`16*($k+0)+112`(%r10)
216 movdqa %xmm1,`16*($k+1)+112`(%r10)
219 movdqa %xmm2,`16*($k+2)+112`(%r10)
220 pand `16*($k+0)-128`($bp),%xmm0 # while it's still in register
222 pand `16*($k+1)-128`($bp),%xmm1
223 pand `16*($k+2)-128`($bp),%xmm2
224 movdqa %xmm3,`16*($k+3)+112`(%r10)
225 pand `16*($k+3)-128`($bp),%xmm3
229 for($k=0;$k<$STRIDE/16-4;$k+=4) {
231 movdqa `16*($k+0)-128`($bp),%xmm4
232 movdqa `16*($k+1)-128`($bp),%xmm5
233 movdqa `16*($k+2)-128`($bp),%xmm2
234 pand `16*($k+0)+112`(%r10),%xmm4
235 movdqa `16*($k+3)-128`($bp),%xmm3
236 pand `16*($k+1)+112`(%r10),%xmm5
238 pand `16*($k+2)+112`(%r10),%xmm2
240 pand `16*($k+3)+112`(%r10),%xmm3
247 pshufd \$0x4e,%xmm0,%xmm1
250 movq %xmm0,$m0 # m0=bp[0]
252 mov ($n0),$n0 # pull n0[0] value
259 mulq $m0 # ap[0]*bp[0]
263 imulq $lo0,$m1 # "tp[0]"*n0
267 add %rax,$lo0 # discarded
280 add $hi0,$hi1 # np[j]*m1+ap[j]*bp[0]
283 mov $hi1,-16(%rsp,$j,8) # tp[j-1]
287 mulq $m0 # ap[j]*bp[0]
296 jne .L1st # note that upon exit $j==$num, so
297 # they can be used interchangeably
301 add $hi0,$hi1 # np[j]*m1+ap[j]*bp[0]
303 mov $hi1,-16(%rsp,$num,8) # tp[num-1]
310 mov $hi1,-8(%rsp,$num,8)
311 mov %rdx,(%rsp,$num,8) # store upmost overflow bit
317 lea 24+128(%rsp,$num,8),%rdx # where 256-byte mask is (+size optimization)
322 for($k=0;$k<$STRIDE/16;$k+=4) {
324 movdqa `16*($k+0)-128`($bp),%xmm0
325 movdqa `16*($k+1)-128`($bp),%xmm1
326 movdqa `16*($k+2)-128`($bp),%xmm2
327 movdqa `16*($k+3)-128`($bp),%xmm3
328 pand `16*($k+0)-128`(%rdx),%xmm0
329 pand `16*($k+1)-128`(%rdx),%xmm1
331 pand `16*($k+2)-128`(%rdx),%xmm2
333 pand `16*($k+3)-128`(%rdx),%xmm3
340 pshufd \$0x4e,%xmm4,%xmm0
344 mov ($ap),%rax # ap[0]
345 movq %xmm0,$m0 # m0=bp[i]
351 mulq $m0 # ap[0]*bp[i]
352 add %rax,$lo0 # ap[0]*bp[i]+tp[0]
356 imulq $lo0,$m1 # tp[0]*n0
360 add %rax,$lo0 # discarded
363 mov 8(%rsp),$lo0 # tp[1]
374 add $lo0,$hi1 # np[j]*m1+ap[j]*bp[i]+tp[j]
377 mov $hi1,-16(%rsp,$j,8) # tp[j-1]
381 mulq $m0 # ap[j]*bp[i]
385 add $hi0,$lo0 # ap[j]*bp[i]+tp[j]
392 jne .Linner # note that upon exit $j==$num, so
393 # they can be used interchangeably
396 add $lo0,$hi1 # np[j]*m1+ap[j]*bp[i]+tp[j]
397 mov (%rsp,$num,8),$lo0
399 mov $hi1,-16(%rsp,$num,8) # tp[num-1]
405 add $lo0,$hi1 # pull upmost overflow bit
407 mov $hi1,-8(%rsp,$num,8)
408 mov %rdx,(%rsp,$num,8) # store upmost overflow bit
414 xor $i,$i # i=0 and clear CF!
415 mov (%rsp),%rax # tp[0]
416 lea (%rsp),$ap # borrow ap for tp
420 .Lsub: sbb ($np,$i,8),%rax
421 mov %rax,($rp,$i,8) # rp[i]=tp[i]-np[i]
422 mov 8($ap,$i,8),%rax # tp[i+1]
424 dec $j # doesn't affect CF!
427 sbb \$0,%rax # handle upmost overflow bit
433 .Lcopy: # conditional copy
438 mov $i,(%rsp,$i,8) # zap temporary vector
440 mov %rdx,($rp,$i,8) # rp[i]=tp[i]
445 mov 8(%rsp,$num,8),%rsi # restore %rsp
462 .cfi_def_cfa_register %rsp
466 .size bn_mul_mont_gather5,.-bn_mul_mont_gather5
469 my @A=("%r10","%r11");
470 my @N=("%r13","%rdi");
472 .type bn_mul4x_mont_gather5,\@function,6
474 bn_mul4x_mont_gather5:
478 .cfi_def_cfa_register %rax
481 $code.=<<___ if ($addx);
483 cmp \$0x80108,%r11d # check for AD*X+BMI2+BMI1
502 shl \$3,${num}d # convert $num to bytes
503 lea ($num,$num,2),%r10 # 3*$num in bytes
506 ##############################################################
507 # Ensure that stack frame doesn't alias with $rptr+3*$num
508 # modulo 4096, which covers ret[num], am[num] and n[num]
509 # (see bn_exp.c). This is done to allow memory disambiguation
510 # logic do its magic. [Extra [num] is allocated in order
511 # to align with bn_power5's frame, which is cleansed after
512 # completing exponentiation. Extra 256 bytes is for power mask
513 # calculated from 7th argument, the index.]
515 lea -320(%rsp,$num,2),%r11
521 sub %r11,%rbp # align with $rp
522 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*num*8+256)
527 lea 4096-320(,$num,2),%r10
528 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*num*8+256)
542 jmp .Lmul4x_page_walk_done
549 .Lmul4x_page_walk_done:
554 .cfi_cfa_expression %rsp+40,deref,+8
559 mov 40(%rsp),%rsi # restore %rsp
576 .cfi_def_cfa_register %rsp
580 .size bn_mul4x_mont_gather5,.-bn_mul4x_mont_gather5
582 .type mul4x_internal,\@abi-omnipotent
586 shl \$5,$num # $num was in bytes
587 movd `($win64?56:8)`(%rax),%xmm5 # load 7th argument, index
589 lea 128(%rdx,$num),%r13 # end of powers table (+size optimization)
590 shr \$5,$num # restore $num
593 $STRIDE=2**5*8; # 5 is "window size"
594 $N=$STRIDE/4; # should match cache line size
597 movdqa 0(%rax),%xmm0 # 00000001000000010000000000000000
598 movdqa 16(%rax),%xmm1 # 00000002000000020000000200000002
599 lea 88-112(%rsp,$num),%r10 # place the mask after tp[num+1] (+ICache optimization)
600 lea 128(%rdx),$bp # size optimization
602 pshufd \$0,%xmm5,%xmm5 # broadcast index
607 ########################################################################
608 # calculate mask by comparing 0..31 to index and save result to stack
612 pcmpeqd %xmm5,%xmm0 # compare to 1,0
616 for($i=0;$i<$STRIDE/16-4;$i+=4) {
619 pcmpeqd %xmm5,%xmm1 # compare to 3,2
620 movdqa %xmm0,`16*($i+0)+112`(%r10)
624 pcmpeqd %xmm5,%xmm2 # compare to 5,4
625 movdqa %xmm1,`16*($i+1)+112`(%r10)
629 pcmpeqd %xmm5,%xmm3 # compare to 7,6
630 movdqa %xmm2,`16*($i+2)+112`(%r10)
635 movdqa %xmm3,`16*($i+3)+112`(%r10)
639 $code.=<<___; # last iteration can be optimized
642 movdqa %xmm0,`16*($i+0)+112`(%r10)
647 movdqa %xmm1,`16*($i+1)+112`(%r10)
650 movdqa %xmm2,`16*($i+2)+112`(%r10)
651 pand `16*($i+0)-128`($bp),%xmm0 # while it's still in register
653 pand `16*($i+1)-128`($bp),%xmm1
654 pand `16*($i+2)-128`($bp),%xmm2
655 movdqa %xmm3,`16*($i+3)+112`(%r10)
656 pand `16*($i+3)-128`($bp),%xmm3
660 for($i=0;$i<$STRIDE/16-4;$i+=4) {
662 movdqa `16*($i+0)-128`($bp),%xmm4
663 movdqa `16*($i+1)-128`($bp),%xmm5
664 movdqa `16*($i+2)-128`($bp),%xmm2
665 pand `16*($i+0)+112`(%r10),%xmm4
666 movdqa `16*($i+3)-128`($bp),%xmm3
667 pand `16*($i+1)+112`(%r10),%xmm5
669 pand `16*($i+2)+112`(%r10),%xmm2
671 pand `16*($i+3)+112`(%r10),%xmm3
678 pshufd \$0x4e,%xmm0,%xmm1
681 movq %xmm0,$m0 # m0=bp[0]
683 mov %r13,16+8(%rsp) # save end of b[num]
684 mov $rp, 56+8(%rsp) # save $rp
686 mov ($n0),$n0 # pull n0[0] value
688 lea ($ap,$num),$ap # end of a[num]
692 mulq $m0 # ap[0]*bp[0]
696 imulq $A[0],$m1 # "tp[0]"*n0
701 add %rax,$A[0] # discarded
714 mov 16($ap,$num),%rax
717 lea 4*8($num),$j # j=4
726 mulq $m0 # ap[j]*bp[0]
737 add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0]
739 mov $N[0],-24($tp) # tp[j-1]
742 mulq $m0 # ap[j]*bp[0]
752 add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0]
754 mov $N[1],-16($tp) # tp[j-1]
757 mulq $m0 # ap[j]*bp[0]
767 add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0]
769 mov $N[0],-8($tp) # tp[j-1]
772 mulq $m0 # ap[j]*bp[0]
782 add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0]
785 mov $N[1],($tp) # tp[j-1]
791 mulq $m0 # ap[j]*bp[0]
802 add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0]
804 mov $N[0],-24($tp) # tp[j-1]
807 mulq $m0 # ap[j]*bp[0]
815 mov ($ap,$num),%rax # ap[0]
817 add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0]
819 mov $N[1],-16($tp) # tp[j-1]
822 lea ($np,$num),$np # rewind $np
833 lea 16+128($tp),%rdx # where 256-byte mask is (+size optimization)
837 for($i=0;$i<$STRIDE/16;$i+=4) {
839 movdqa `16*($i+0)-128`($bp),%xmm0
840 movdqa `16*($i+1)-128`($bp),%xmm1
841 movdqa `16*($i+2)-128`($bp),%xmm2
842 movdqa `16*($i+3)-128`($bp),%xmm3
843 pand `16*($i+0)-128`(%rdx),%xmm0
844 pand `16*($i+1)-128`(%rdx),%xmm1
846 pand `16*($i+2)-128`(%rdx),%xmm2
848 pand `16*($i+3)-128`(%rdx),%xmm3
855 pshufd \$0x4e,%xmm4,%xmm0
858 movq %xmm0,$m0 # m0=bp[i]
862 mulq $m0 # ap[0]*bp[i]
863 add %rax,$A[0] # ap[0]*bp[i]+tp[0]
867 imulq $A[0],$m1 # tp[0]*n0
869 mov $N[1],($tp) # store upmost overflow bit
871 lea ($tp,$num),$tp # rewind $tp
874 add %rax,$A[0] # "$N[0]", discarded
879 mulq $m0 # ap[j]*bp[i]
883 add 8($tp),$A[1] # +tp[1]
889 mov 16($ap,$num),%rax
891 add $A[1],$N[1] # np[j]*m1+ap[j]*bp[i]+tp[j]
892 lea 4*8($num),$j # j=4
900 mulq $m0 # ap[j]*bp[i]
904 add 16($tp),$A[0] # ap[j]*bp[i]+tp[j]
915 mov $N[1],-32($tp) # tp[j-1]
918 mulq $m0 # ap[j]*bp[i]
932 mov $N[0],-24($tp) # tp[j-1]
935 mulq $m0 # ap[j]*bp[i]
939 add ($tp),$A[0] # ap[j]*bp[i]+tp[j]
949 mov $N[1],-16($tp) # tp[j-1]
952 mulq $m0 # ap[j]*bp[i]
967 mov $N[0],-8($tp) # tp[j-1]
973 mulq $m0 # ap[j]*bp[i]
977 add 16($tp),$A[0] # ap[j]*bp[i]+tp[j]
988 mov $N[1],-32($tp) # tp[j-1]
991 mulq $m0 # ap[j]*bp[i]
1002 mov ($ap,$num),%rax # ap[0]
1006 mov $N[0],-24($tp) # tp[j-1]
1009 mov $N[1],-16($tp) # tp[j-1]
1010 lea ($np,$num),$np # rewind $np
1015 add ($tp),$N[0] # pull upmost overflow bit
1016 adc \$0,$N[1] # upmost overflow bit
1025 sub $N[0],$m1 # compare top-most words
1026 adc $j,$j # $j is zero
1028 sub $N[1],%rax # %rax=-$N[1]
1029 lea ($tp,$num),%rbx # tptr in .sqr4x_sub
1031 lea ($np),%rbp # nptr in .sqr4x_sub
1034 mov 56+8(%rsp),%rdi # rptr in .sqr4x_sub
1035 dec %r12 # so that after 'not' we get -n[0]
1040 jmp .Lsqr4x_sub_entry
1043 my @ri=("%rax",$bp,$m0,$m1);
1047 lea ($tp,$num),$tp # rewind $tp
1049 lea ($np,$N[1],8),$np
1050 mov 56+8(%rsp),$rp # restore $rp
1059 sbb 16*0($np),@ri[0]
1061 sbb 16*1($np),@ri[1]
1064 sbb 16*2($np),@ri[2]
1066 sbb 16*3($np),@ri[3]
1081 .size mul4x_internal,.-mul4x_internal
1085 ######################################################################
1087 my $rptr="%rdi"; # BN_ULONG *rptr,
1088 my $aptr="%rsi"; # const BN_ULONG *aptr,
1089 my $bptr="%rdx"; # const void *table,
1090 my $nptr="%rcx"; # const BN_ULONG *nptr,
1091 my $n0 ="%r8"; # const BN_ULONG *n0);
1092 my $num ="%r9"; # int num, has to be divisible by 8
1095 my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
1096 my @A0=("%r10","%r11");
1097 my @A1=("%r12","%r13");
1098 my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
1102 .type bn_power5,\@function,6
1107 .cfi_def_cfa_register %rax
1109 $code.=<<___ if ($addx);
1110 mov OPENSSL_ia32cap_P+8(%rip),%r11d
1112 cmp \$0x80108,%r11d # check for AD*X+BMI2+BMI1
1130 shl \$3,${num}d # convert $num to bytes
1131 lea ($num,$num,2),%r10d # 3*$num
1135 ##############################################################
1136 # Ensure that stack frame doesn't alias with $rptr+3*$num
1137 # modulo 4096, which covers ret[num], am[num] and n[num]
1138 # (see bn_exp.c). This is done to allow memory disambiguation
1139 # logic do its magic. [Extra 256 bytes is for power mask
1140 # calculated from 7th argument, the index.]
1142 lea -320(%rsp,$num,2),%r11
1148 sub %r11,%rbp # align with $aptr
1149 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*num*8+256)
1154 lea 4096-320(,$num,2),%r10
1155 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*num*8+256)
1165 lea (%rbp,%r11),%rsp
1169 jmp .Lpwr_page_walk_done
1172 lea -4096(%rsp),%rsp
1176 .Lpwr_page_walk_done:
1181 ##############################################################
1184 # +0 saved $num, used in reduction section
1185 # +8 &t[2*$num], used in reduction section
1191 mov %rax, 40(%rsp) # save original %rsp
1192 .cfi_cfa_expression %rsp+40,deref,+8
1194 movq $rptr,%xmm1 # save $rptr, used in sqr8x
1195 movq $nptr,%xmm2 # save $nptr
1196 movq %r10, %xmm3 # -$num, used in sqr8x
1199 call __bn_sqr8x_internal
1200 call __bn_post4x_internal
1201 call __bn_sqr8x_internal
1202 call __bn_post4x_internal
1203 call __bn_sqr8x_internal
1204 call __bn_post4x_internal
1205 call __bn_sqr8x_internal
1206 call __bn_post4x_internal
1207 call __bn_sqr8x_internal
1208 call __bn_post4x_internal
1218 mov 40(%rsp),%rsi # restore %rsp
1234 .cfi_def_cfa_register %rsp
1238 .size bn_power5,.-bn_power5
1240 .globl bn_sqr8x_internal
1241 .hidden bn_sqr8x_internal
1242 .type bn_sqr8x_internal,\@abi-omnipotent
1245 __bn_sqr8x_internal:
1247 ##############################################################
1250 # a) multiply-n-add everything but a[i]*a[i];
1251 # b) shift result of a) by 1 to the left and accumulate
1252 # a[i]*a[i] products;
1254 ##############################################################
1320 lea 32(%r10),$i # $i=-($num-32)
1321 lea ($aptr,$num),$aptr # end of a[] buffer, ($aptr,$i)=&ap[2]
1323 mov $num,$j # $j=$num
1325 # comments apply to $num==8 case
1326 mov -32($aptr,$i),$a0 # a[0]
1327 lea 48+8(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num]
1328 mov -24($aptr,$i),%rax # a[1]
1329 lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"]
1330 mov -16($aptr,$i),$ai # a[2]
1334 mov %rax,$A0[0] # a[1]*a[0]
1337 mov $A0[0],-24($tptr,$i) # t[1]
1343 mov $A0[1],-16($tptr,$i) # t[2]
1347 mov -8($aptr,$i),$ai # a[3]
1349 mov %rax,$A1[0] # a[2]*a[1]+t[3]
1355 add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3]
1361 mov $A0[0],-8($tptr,$j) # t[3]
1366 mov ($aptr,$j),$ai # a[4]
1368 add %rax,$A1[1] # a[3]*a[1]+t[4]
1374 add %rax,$A0[1] # a[4]*a[0]+a[3]*a[1]+t[4]
1376 mov 8($aptr,$j),$ai # a[5]
1384 add %rax,$A1[0] # a[4]*a[3]+t[5]
1386 mov $A0[1],($tptr,$j) # t[4]
1391 add %rax,$A0[0] # a[5]*a[2]+a[4]*a[3]+t[5]
1393 mov 16($aptr,$j),$ai # a[6]
1400 add %rax,$A1[1] # a[5]*a[3]+t[6]
1402 mov $A0[0],8($tptr,$j) # t[5]
1407 add %rax,$A0[1] # a[6]*a[2]+a[5]*a[3]+t[6]
1409 mov 24($aptr,$j),$ai # a[7]
1417 add %rax,$A1[0] # a[6]*a[5]+t[7]
1419 mov $A0[1],16($tptr,$j) # t[6]
1425 add %rax,$A0[0] # a[7]*a[4]+a[6]*a[5]+t[6]
1431 mov $A0[0],-8($tptr,$j) # t[7]
1443 mov $A1[1],($tptr) # t[8]
1445 mov %rdx,8($tptr) # t[9]
1449 .Lsqr4x_outer: # comments apply to $num==6 case
1450 mov -32($aptr,$i),$a0 # a[0]
1451 lea 48+8(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num]
1452 mov -24($aptr,$i),%rax # a[1]
1453 lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"]
1454 mov -16($aptr,$i),$ai # a[2]
1458 mov -24($tptr,$i),$A0[0] # t[1]
1459 add %rax,$A0[0] # a[1]*a[0]+t[1]
1462 mov $A0[0],-24($tptr,$i) # t[1]
1469 add -16($tptr,$i),$A0[1] # a[2]*a[0]+t[2]
1472 mov $A0[1],-16($tptr,$i) # t[2]
1476 mov -8($aptr,$i),$ai # a[3]
1478 add %rax,$A1[0] # a[2]*a[1]+t[3]
1481 add -8($tptr,$i),$A1[0]
1486 add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3]
1492 mov $A0[0],-8($tptr,$i) # t[3]
1499 mov ($aptr,$j),$ai # a[4]
1501 add %rax,$A1[1] # a[3]*a[1]+t[4]
1505 add ($tptr,$j),$A1[1]
1510 add %rax,$A0[1] # a[4]*a[0]+a[3]*a[1]+t[4]
1512 mov 8($aptr,$j),$ai # a[5]
1519 add %rax,$A1[0] # a[4]*a[3]+t[5]
1520 mov $A0[1],($tptr,$j) # t[4]
1524 add 8($tptr,$j),$A1[0]
1529 add %rax,$A0[0] # a[5]*a[2]+a[4]*a[3]+t[5]
1535 mov $A0[0],-8($tptr,$j) # t[5], "preloaded t[1]" below
1547 mov $A1[1],($tptr) # t[6], "preloaded t[2]" below
1549 mov %rdx,8($tptr) # t[7], "preloaded t[3]" below
1554 # comments apply to $num==4 case
1555 mov -32($aptr),$a0 # a[0]
1556 lea 48+8(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num]
1557 mov -24($aptr),%rax # a[1]
1558 lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"]
1559 mov -16($aptr),$ai # a[2]
1563 add %rax,$A0[0] # a[1]*a[0]+t[1], preloaded t[1]
1571 mov $A0[0],-24($tptr) # t[1]
1574 add $A1[1],$A0[1] # a[2]*a[0]+t[2], preloaded t[2]
1575 mov -8($aptr),$ai # a[3]
1579 add %rax,$A1[0] # a[2]*a[1]+t[3], preloaded t[3]
1581 mov $A0[1],-16($tptr) # t[2]
1586 add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3]
1592 mov $A0[0],-8($tptr) # t[3]
1596 mov -16($aptr),%rax # a[2]
1601 mov $A1[1],($tptr) # t[4]
1603 mov %rdx,8($tptr) # t[5]
1608 my ($shift,$carry)=($a0,$a1);
1609 my @S=(@A1,$ai,$n0);
1613 sub $num,$i # $i=16-$num
1616 add $A1[0],%rax # t[5]
1618 mov %rax,8($tptr) # t[5]
1619 mov %rdx,16($tptr) # t[6]
1620 mov $carry,24($tptr) # t[7]
1622 mov -16($aptr,$i),%rax # a[0]
1623 lea 48+8(%rsp),$tptr
1624 xor $A0[0],$A0[0] # t[0]
1625 mov 8($tptr),$A0[1] # t[1]
1627 lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
1629 lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
1631 or $A0[0],$S[1] # | t[2*i]>>63
1632 mov 16($tptr),$A0[0] # t[2*i+2] # prefetch
1633 mov $A0[1],$shift # shift=t[2*i+1]>>63
1634 mul %rax # a[i]*a[i]
1635 neg $carry # mov $carry,cf
1636 mov 24($tptr),$A0[1] # t[2*i+2+1] # prefetch
1638 mov -8($aptr,$i),%rax # a[i+1] # prefetch
1642 lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift
1644 sbb $carry,$carry # mov cf,$carry
1646 lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
1648 or $A0[0],$S[3] # | t[2*i]>>63
1649 mov 32($tptr),$A0[0] # t[2*i+2] # prefetch
1650 mov $A0[1],$shift # shift=t[2*i+1]>>63
1651 mul %rax # a[i]*a[i]
1652 neg $carry # mov $carry,cf
1653 mov 40($tptr),$A0[1] # t[2*i+2+1] # prefetch
1655 mov 0($aptr,$i),%rax # a[i+1] # prefetch
1660 sbb $carry,$carry # mov cf,$carry
1662 jmp .Lsqr4x_shift_n_add
1665 .Lsqr4x_shift_n_add:
1666 lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
1668 lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
1670 or $A0[0],$S[1] # | t[2*i]>>63
1671 mov -16($tptr),$A0[0] # t[2*i+2] # prefetch
1672 mov $A0[1],$shift # shift=t[2*i+1]>>63
1673 mul %rax # a[i]*a[i]
1674 neg $carry # mov $carry,cf
1675 mov -8($tptr),$A0[1] # t[2*i+2+1] # prefetch
1677 mov -8($aptr,$i),%rax # a[i+1] # prefetch
1678 mov $S[0],-32($tptr)
1681 lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift
1682 mov $S[1],-24($tptr)
1683 sbb $carry,$carry # mov cf,$carry
1685 lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
1687 or $A0[0],$S[3] # | t[2*i]>>63
1688 mov 0($tptr),$A0[0] # t[2*i+2] # prefetch
1689 mov $A0[1],$shift # shift=t[2*i+1]>>63
1690 mul %rax # a[i]*a[i]
1691 neg $carry # mov $carry,cf
1692 mov 8($tptr),$A0[1] # t[2*i+2+1] # prefetch
1694 mov 0($aptr,$i),%rax # a[i+1] # prefetch
1695 mov $S[2],-16($tptr)
1698 lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
1700 sbb $carry,$carry # mov cf,$carry
1702 lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
1704 or $A0[0],$S[1] # | t[2*i]>>63
1705 mov 16($tptr),$A0[0] # t[2*i+2] # prefetch
1706 mov $A0[1],$shift # shift=t[2*i+1]>>63
1707 mul %rax # a[i]*a[i]
1708 neg $carry # mov $carry,cf
1709 mov 24($tptr),$A0[1] # t[2*i+2+1] # prefetch
1711 mov 8($aptr,$i),%rax # a[i+1] # prefetch
1715 lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift
1717 sbb $carry,$carry # mov cf,$carry
1719 lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
1721 or $A0[0],$S[3] # | t[2*i]>>63
1722 mov 32($tptr),$A0[0] # t[2*i+2] # prefetch
1723 mov $A0[1],$shift # shift=t[2*i+1]>>63
1724 mul %rax # a[i]*a[i]
1725 neg $carry # mov $carry,cf
1726 mov 40($tptr),$A0[1] # t[2*i+2+1] # prefetch
1728 mov 16($aptr,$i),%rax # a[i+1] # prefetch
1732 sbb $carry,$carry # mov cf,$carry
1735 jnz .Lsqr4x_shift_n_add
1737 lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift
1740 lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 |
1742 or $A0[0],$S[1] # | t[2*i]>>63
1743 mov -16($tptr),$A0[0] # t[2*i+2] # prefetch
1744 mov $A0[1],$shift # shift=t[2*i+1]>>63
1745 mul %rax # a[i]*a[i]
1746 neg $carry # mov $carry,cf
1747 mov -8($tptr),$A0[1] # t[2*i+2+1] # prefetch
1749 mov -8($aptr),%rax # a[i+1] # prefetch
1750 mov $S[0],-32($tptr)
1753 lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1|shift
1754 mov $S[1],-24($tptr)
1755 sbb $carry,$carry # mov cf,$carry
1757 lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 |
1759 or $A0[0],$S[3] # | t[2*i]>>63
1760 mul %rax # a[i]*a[i]
1761 neg $carry # mov $carry,cf
1764 mov $S[2],-16($tptr)
1768 ######################################################################
1769 # Montgomery reduction part, "word-by-word" algorithm.
1771 # This new path is inspired by multiple submissions from Intel, by
1772 # Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
1775 my ($nptr,$tptr,$carry,$m0)=("%rbp","%rdi","%rsi","%rbx");
1779 __bn_sqr8x_reduction:
1781 lea ($nptr,$num),%rcx # end of n[]
1782 lea 48+8(%rsp,$num,2),%rdx # end of t[] buffer
1784 lea 48+8(%rsp,$num),$tptr # end of initial t[] window
1787 jmp .L8x_reduction_loop
1790 .L8x_reduction_loop:
1791 lea ($tptr,$num),$tptr # start of current t[] window
1801 mov %rax,(%rdx) # store top-most carry bit
1802 lea 8*8($tptr),$tptr
1806 imulq 32+8(%rsp),$m0 # n0*a[0]
1807 mov 8*0($nptr),%rax # n[0]
1814 mov 8*1($nptr),%rax # n[1]
1824 mov $m0,48-8+8(%rsp,%rcx,8) # put aside n0*a[i]
1833 mov 32+8(%rsp),$carry # pull n0, borrow $carry
1841 imulq %r8,$carry # modulo-scheduled
1871 mov $carry,$m0 # n0*a[i]
1873 mov 8*0($nptr),%rax # n[0]
1882 lea 8*8($nptr),$nptr
1884 mov 8+8(%rsp),%rdx # pull end of t[]
1885 cmp 0+8(%rsp),$nptr # end of n[]?
1897 sbb $carry,$carry # top carry
1899 mov 48+56+8(%rsp),$m0 # pull n0*a[0]
1909 mov %r8,($tptr) # save result
1918 lea 8($tptr),$tptr # $tptr++
1963 mov 48-16+8(%rsp,%rcx,8),$m0# pull n0*a[i]
1967 mov 8*0($nptr),%rax # pull n[0]
1974 lea 8*8($nptr),$nptr
1975 mov 8+8(%rsp),%rdx # pull end of t[]
1976 cmp 0+8(%rsp),$nptr # end of n[]?
1977 jae .L8x_tail_done # break out of loop
1979 mov 48+56+8(%rsp),$m0 # pull n0*a[0]
1981 mov 8*0($nptr),%rax # pull n[0]
1990 sbb $carry,$carry # top carry
1998 add (%rdx),%r8 # can this overflow?
2018 adc \$0,%rax # top-most carry
2019 mov -8($nptr),%rcx # np[num-1]
2022 movq %xmm2,$nptr # restore $nptr
2024 mov %r8,8*0($tptr) # store top 512 bits
2026 movq %xmm3,$num # $num is %r9, can't be moved upwards
2033 lea 8*8($tptr),$tptr
2035 cmp %rdx,$tptr # end of t[]?
2036 jb .L8x_reduction_loop
2039 .size bn_sqr8x_internal,.-bn_sqr8x_internal
2042 ##############################################################
2043 # Post-condition, 4x unrolled
2046 my ($tptr,$nptr)=("%rbx","%rbp");
2048 .type __bn_post4x_internal,\@abi-omnipotent
2050 __bn_post4x_internal:
2053 lea (%rdi,$num),$tptr # %rdi was $tptr above
2055 movq %xmm1,$rptr # restore $rptr
2057 movq %xmm1,$aptr # prepare for back-to-back call
2059 dec %r12 # so that after 'not' we get -n[0]
2064 jmp .Lsqr4x_sub_entry
2073 lea 8*4($nptr),$nptr
2083 neg %r10 # mov %r10,%cf
2089 lea 8*4($tptr),$tptr
2091 sbb %r10,%r10 # mov %cf,%r10
2094 lea 8*4($rptr),$rptr
2099 mov $num,%r10 # prepare for back-to-back call
2100 neg $num # restore $num
2103 .size __bn_post4x_internal,.-__bn_post4x_internal
2108 .globl bn_from_montgomery
2109 .type bn_from_montgomery,\@abi-omnipotent
2113 testl \$7,`($win64?"48(%rsp)":"%r9d")`
2118 .size bn_from_montgomery,.-bn_from_montgomery
2120 .type bn_from_mont8x,\@function,6
2126 .cfi_def_cfa_register %rax
2141 shl \$3,${num}d # convert $num to bytes
2142 lea ($num,$num,2),%r10 # 3*$num in bytes
2146 ##############################################################
2147 # Ensure that stack frame doesn't alias with $rptr+3*$num
2148 # modulo 4096, which covers ret[num], am[num] and n[num]
2149 # (see bn_exp.c). The stack is allocated to aligned with
2150 # bn_power5's frame, and as bn_from_montgomery happens to be
2151 # last operation, we use the opportunity to cleanse it.
2153 lea -320(%rsp,$num,2),%r11
2159 sub %r11,%rbp # align with $aptr
2160 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*$num*8+256)
2165 lea 4096-320(,$num,2),%r10
2166 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*$num*8+256)
2176 lea (%rbp,%r11),%rsp
2180 jmp .Lfrom_page_walk_done
2183 lea -4096(%rsp),%rsp
2187 .Lfrom_page_walk_done:
2192 ##############################################################
2195 # +0 saved $num, used in reduction section
2196 # +8 &t[2*$num], used in reduction section
2202 mov %rax, 40(%rsp) # save original %rsp
2203 .cfi_cfa_expression %rsp+40,deref,+8
2212 movdqu ($aptr),%xmm1
2213 movdqu 16($aptr),%xmm2
2214 movdqu 32($aptr),%xmm3
2215 movdqa %xmm0,(%rax,$num)
2216 movdqu 48($aptr),%xmm4
2217 movdqa %xmm0,16(%rax,$num)
2218 .byte 0x48,0x8d,0xb6,0x40,0x00,0x00,0x00 # lea 64($aptr),$aptr
2220 movdqa %xmm0,32(%rax,$num)
2221 movdqa %xmm2,16(%rax)
2222 movdqa %xmm0,48(%rax,$num)
2223 movdqa %xmm3,32(%rax)
2224 movdqa %xmm4,48(%rax)
2233 movq %r10, %xmm3 # -num
2235 $code.=<<___ if ($addx);
2236 mov OPENSSL_ia32cap_P+8(%rip),%r11d
2238 cmp \$0x80108,%r11d # check for AD*X+BMI2+BMI1
2241 lea (%rax,$num),$rptr
2242 call __bn_sqrx8x_reduction
2243 call __bn_postx4x_internal
2247 jmp .Lfrom_mont_zero
2253 call __bn_sqr8x_reduction
2254 call __bn_post4x_internal
2258 jmp .Lfrom_mont_zero
2262 mov 40(%rsp),%rsi # restore %rsp
2264 movdqa %xmm0,16*0(%rax)
2265 movdqa %xmm0,16*1(%rax)
2266 movdqa %xmm0,16*2(%rax)
2267 movdqa %xmm0,16*3(%rax)
2270 jnz .Lfrom_mont_zero
2286 .cfi_def_cfa_register %rsp
2290 .size bn_from_mont8x,.-bn_from_mont8x
2296 my $bp="%rdx"; # restore original value
2299 .type bn_mulx4x_mont_gather5,\@function,6
2301 bn_mulx4x_mont_gather5:
2304 .cfi_def_cfa_register %rax
2320 shl \$3,${num}d # convert $num to bytes
2321 lea ($num,$num,2),%r10 # 3*$num in bytes
2325 ##############################################################
2326 # Ensure that stack frame doesn't alias with $rptr+3*$num
2327 # modulo 4096, which covers ret[num], am[num] and n[num]
2328 # (see bn_exp.c). This is done to allow memory disambiguation
2329 # logic do its magic. [Extra [num] is allocated in order
2330 # to align with bn_power5's frame, which is cleansed after
2331 # completing exponentiation. Extra 256 bytes is for power mask
2332 # calculated from 7th argument, the index.]
2334 lea -320(%rsp,$num,2),%r11
2340 sub %r11,%rbp # align with $aptr
2341 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*$num*8+256)
2345 lea 4096-320(,$num,2),%r10
2346 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*$num*8+256)
2352 and \$-64,%rbp # ensure alignment
2356 lea (%rbp,%r11),%rsp
2359 ja .Lmulx4x_page_walk
2360 jmp .Lmulx4x_page_walk_done
2363 lea -4096(%rsp),%rsp
2366 ja .Lmulx4x_page_walk
2367 .Lmulx4x_page_walk_done:
2369 ##############################################################
2372 # +8 off-loaded &b[i]
2381 mov $n0, 32(%rsp) # save *n0
2382 mov %rax,40(%rsp) # save original %rsp
2383 .cfi_cfa_expression %rsp+40,deref,+8
2385 call mulx4x_internal
2387 mov 40(%rsp),%rsi # restore %rsp
2404 .cfi_def_cfa_register %rsp
2408 .size bn_mulx4x_mont_gather5,.-bn_mulx4x_mont_gather5
2410 .type mulx4x_internal,\@abi-omnipotent
2414 mov $num,8(%rsp) # save -$num (it was in bytes)
2416 neg $num # restore $num
2418 neg %r10 # restore $num
2419 lea 128($bp,$num),%r13 # end of powers table (+size optimization)
2421 movd `($win64?56:8)`(%rax),%xmm5 # load 7th argument
2423 lea .Linc(%rip),%rax
2424 mov %r13,16+8(%rsp) # end of b[num]
2425 mov $num,24+8(%rsp) # inner counter
2426 mov $rp, 56+8(%rsp) # save $rp
2428 my ($aptr, $bptr, $nptr, $tptr, $mi, $bi, $zero, $num)=
2429 ("%rsi","%rdi","%rcx","%rbx","%r8","%r9","%rbp","%rax");
2431 my $STRIDE=2**5*8; # 5 is "window size"
2432 my $N=$STRIDE/4; # should match cache line size
2434 movdqa 0(%rax),%xmm0 # 00000001000000010000000000000000
2435 movdqa 16(%rax),%xmm1 # 00000002000000020000000200000002
2436 lea 88-112(%rsp,%r10),%r10 # place the mask after tp[num+1] (+ICache optimization)
2437 lea 128($bp),$bptr # size optimization
2439 pshufd \$0,%xmm5,%xmm5 # broadcast index
2444 ########################################################################
2445 # calculate mask by comparing 0..31 to index and save result to stack
2450 pcmpeqd %xmm5,%xmm0 # compare to 1,0
2453 for($i=0;$i<$STRIDE/16-4;$i+=4) {
2456 pcmpeqd %xmm5,%xmm1 # compare to 3,2
2457 movdqa %xmm0,`16*($i+0)+112`(%r10)
2461 pcmpeqd %xmm5,%xmm2 # compare to 5,4
2462 movdqa %xmm1,`16*($i+1)+112`(%r10)
2466 pcmpeqd %xmm5,%xmm3 # compare to 7,6
2467 movdqa %xmm2,`16*($i+2)+112`(%r10)
2472 movdqa %xmm3,`16*($i+3)+112`(%r10)
2476 $code.=<<___; # last iteration can be optimized
2480 movdqa %xmm0,`16*($i+0)+112`(%r10)
2484 movdqa %xmm1,`16*($i+1)+112`(%r10)
2487 movdqa %xmm2,`16*($i+2)+112`(%r10)
2489 pand `16*($i+0)-128`($bptr),%xmm0 # while it's still in register
2490 pand `16*($i+1)-128`($bptr),%xmm1
2491 pand `16*($i+2)-128`($bptr),%xmm2
2492 movdqa %xmm3,`16*($i+3)+112`(%r10)
2493 pand `16*($i+3)-128`($bptr),%xmm3
2497 for($i=0;$i<$STRIDE/16-4;$i+=4) {
2499 movdqa `16*($i+0)-128`($bptr),%xmm4
2500 movdqa `16*($i+1)-128`($bptr),%xmm5
2501 movdqa `16*($i+2)-128`($bptr),%xmm2
2502 pand `16*($i+0)+112`(%r10),%xmm4
2503 movdqa `16*($i+3)-128`($bptr),%xmm3
2504 pand `16*($i+1)+112`(%r10),%xmm5
2506 pand `16*($i+2)+112`(%r10),%xmm2
2508 pand `16*($i+3)+112`(%r10),%xmm3
2515 pshufd \$0x4e,%xmm0,%xmm1
2517 lea $STRIDE($bptr),$bptr
2518 movq %xmm0,%rdx # bp[0]
2519 lea 64+8*4+8(%rsp),$tptr
2522 mulx 0*8($aptr),$mi,%rax # a[0]*b[0]
2523 mulx 1*8($aptr),%r11,%r12 # a[1]*b[0]
2525 mulx 2*8($aptr),%rax,%r13 # ...
2528 mulx 3*8($aptr),%rax,%r14
2531 imulq 32+8(%rsp),$mi # "t[0]"*n0
2532 xor $zero,$zero # cf=0, of=0
2535 mov $bptr,8+8(%rsp) # off-load &b[i]
2537 lea 4*8($aptr),$aptr
2539 adcx $zero,%r14 # cf=0
2541 mulx 0*8($nptr),%rax,%r10
2542 adcx %rax,%r15 # discarded
2544 mulx 1*8($nptr),%rax,%r11
2547 mulx 2*8($nptr),%rax,%r12
2548 mov 24+8(%rsp),$bptr # counter value
2549 mov %r10,-8*4($tptr)
2552 mulx 3*8($nptr),%rax,%r15
2554 mov %r11,-8*3($tptr)
2556 adox $zero,%r15 # of=0
2557 lea 4*8($nptr),$nptr
2558 mov %r12,-8*2($tptr)
2563 adcx $zero,%r15 # cf=0, modulo-scheduled
2564 mulx 0*8($aptr),%r10,%rax # a[4]*b[0]
2566 mulx 1*8($aptr),%r11,%r14 # a[5]*b[0]
2568 mulx 2*8($aptr),%r12,%rax # ...
2570 mulx 3*8($aptr),%r13,%r14
2574 adcx $zero,%r14 # cf=0
2575 lea 4*8($aptr),$aptr
2576 lea 4*8($tptr),$tptr
2579 mulx 0*8($nptr),%rax,%r15
2582 mulx 1*8($nptr),%rax,%r15
2585 mulx 2*8($nptr),%rax,%r15
2586 mov %r10,-5*8($tptr)
2588 mov %r11,-4*8($tptr)
2590 mulx 3*8($nptr),%rax,%r15
2592 mov %r12,-3*8($tptr)
2595 lea 4*8($nptr),$nptr
2596 mov %r13,-2*8($tptr)
2598 dec $bptr # of=0, pass cf
2601 mov 8(%rsp),$num # load -num
2602 adc $zero,%r15 # modulo-scheduled
2603 lea ($aptr,$num),$aptr # rewind $aptr
2605 mov 8+8(%rsp),$bptr # re-load &b[i]
2606 adc $zero,$zero # top-most carry
2607 mov %r14,-1*8($tptr)
2612 lea 16-256($tptr),%r10 # where 256-byte mask is (+density control)
2617 for($i=0;$i<$STRIDE/16;$i+=4) {
2619 movdqa `16*($i+0)-128`($bptr),%xmm0
2620 movdqa `16*($i+1)-128`($bptr),%xmm1
2621 movdqa `16*($i+2)-128`($bptr),%xmm2
2622 pand `16*($i+0)+256`(%r10),%xmm0
2623 movdqa `16*($i+3)-128`($bptr),%xmm3
2624 pand `16*($i+1)+256`(%r10),%xmm1
2626 pand `16*($i+2)+256`(%r10),%xmm2
2628 pand `16*($i+3)+256`(%r10),%xmm3
2635 pshufd \$0x4e,%xmm4,%xmm0
2637 lea $STRIDE($bptr),$bptr
2638 movq %xmm0,%rdx # m0=bp[i]
2640 mov $zero,($tptr) # save top-most carry
2641 lea 4*8($tptr,$num),$tptr # rewind $tptr
2642 mulx 0*8($aptr),$mi,%r11 # a[0]*b[i]
2643 xor $zero,$zero # cf=0, of=0
2645 mulx 1*8($aptr),%r14,%r12 # a[1]*b[i]
2646 adox -4*8($tptr),$mi # +t[0]
2648 mulx 2*8($aptr),%r15,%r13 # ...
2649 adox -3*8($tptr),%r11
2651 mulx 3*8($aptr),%rdx,%r14
2652 adox -2*8($tptr),%r12
2654 lea ($nptr,$num),$nptr # rewind $nptr
2655 lea 4*8($aptr),$aptr
2656 adox -1*8($tptr),%r13
2661 imulq 32+8(%rsp),$mi # "t[0]"*n0
2664 xor $zero,$zero # cf=0, of=0
2665 mov $bptr,8+8(%rsp) # off-load &b[i]
2667 mulx 0*8($nptr),%rax,%r10
2668 adcx %rax,%r15 # discarded
2670 mulx 1*8($nptr),%rax,%r11
2673 mulx 2*8($nptr),%rax,%r12
2676 mulx 3*8($nptr),%rax,%r15
2678 mov 24+8(%rsp),$bptr # counter value
2679 mov %r10,-8*4($tptr)
2681 mov %r11,-8*3($tptr)
2682 adox $zero,%r15 # of=0
2683 mov %r12,-8*2($tptr)
2684 lea 4*8($nptr),$nptr
2689 mulx 0*8($aptr),%r10,%rax # a[4]*b[i]
2690 adcx $zero,%r15 # cf=0, modulo-scheduled
2692 mulx 1*8($aptr),%r11,%r14 # a[5]*b[i]
2693 adcx 0*8($tptr),%r10
2695 mulx 2*8($aptr),%r12,%rax # ...
2696 adcx 1*8($tptr),%r11
2698 mulx 3*8($aptr),%r13,%r14
2700 adcx 2*8($tptr),%r12
2702 adcx 3*8($tptr),%r13
2703 adox $zero,%r14 # of=0
2704 lea 4*8($aptr),$aptr
2705 lea 4*8($tptr),$tptr
2706 adcx $zero,%r14 # cf=0
2709 mulx 0*8($nptr),%rax,%r15
2712 mulx 1*8($nptr),%rax,%r15
2715 mulx 2*8($nptr),%rax,%r15
2716 mov %r10,-5*8($tptr)
2719 mov %r11,-4*8($tptr)
2720 mulx 3*8($nptr),%rax,%r15
2722 lea 4*8($nptr),$nptr
2723 mov %r12,-3*8($tptr)
2726 mov %r13,-2*8($tptr)
2728 dec $bptr # of=0, pass cf
2731 mov 0+8(%rsp),$num # load -num
2732 adc $zero,%r15 # modulo-scheduled
2733 sub 0*8($tptr),$bptr # pull top-most carry to %cf
2734 mov 8+8(%rsp),$bptr # re-load &b[i]
2737 lea ($aptr,$num),$aptr # rewind $aptr
2738 adc $zero,$zero # top-most carry
2739 mov %r14,-1*8($tptr)
2746 mov ($nptr,$num),%r12
2747 lea ($nptr,$num),%rbp # rewind $nptr
2749 lea ($tptr,$num),%rdi # rewind $tptr
2752 sub %r14,%r10 # compare top-most words
2756 sub %r8,%rax # %rax=-%r8
2757 mov 56+8(%rsp),%rdx # restore rp
2758 dec %r12 # so that after 'not' we get -n[0]
2763 jmp .Lsqrx4x_sub_entry # common post-condition
2765 .size mulx4x_internal,.-mulx4x_internal
2768 ######################################################################
2770 my $rptr="%rdi"; # BN_ULONG *rptr,
2771 my $aptr="%rsi"; # const BN_ULONG *aptr,
2772 my $bptr="%rdx"; # const void *table,
2773 my $nptr="%rcx"; # const BN_ULONG *nptr,
2774 my $n0 ="%r8"; # const BN_ULONG *n0);
2775 my $num ="%r9"; # int num, has to be divisible by 8
2778 my ($i,$j,$tptr)=("%rbp","%rcx",$rptr);
2779 my @A0=("%r10","%r11");
2780 my @A1=("%r12","%r13");
2781 my ($a0,$a1,$ai)=("%r14","%r15","%rbx");
2784 .type bn_powerx5,\@function,6
2789 .cfi_def_cfa_register %rax
2805 shl \$3,${num}d # convert $num to bytes
2806 lea ($num,$num,2),%r10 # 3*$num in bytes
2810 ##############################################################
2811 # Ensure that stack frame doesn't alias with $rptr+3*$num
2812 # modulo 4096, which covers ret[num], am[num] and n[num]
2813 # (see bn_exp.c). This is done to allow memory disambiguation
2814 # logic do its magic. [Extra 256 bytes is for power mask
2815 # calculated from 7th argument, the index.]
2817 lea -320(%rsp,$num,2),%r11
2823 sub %r11,%rbp # align with $aptr
2824 lea -320(%rbp,$num,2),%rbp # future alloca(frame+2*$num*8+256)
2829 lea 4096-320(,$num,2),%r10
2830 lea -320(%rbp,$num,2),%rbp # alloca(frame+2*$num*8+256)
2840 lea (%rbp,%r11),%rsp
2844 jmp .Lpwrx_page_walk_done
2847 lea -4096(%rsp),%rsp
2851 .Lpwrx_page_walk_done:
2856 ##############################################################
2859 # +0 saved $num, used in reduction section
2860 # +8 &t[2*$num], used in reduction section
2861 # +16 intermediate carry bit
2862 # +24 top-most carry bit, used in reduction section
2868 movq $rptr,%xmm1 # save $rptr
2869 movq $nptr,%xmm2 # save $nptr
2870 movq %r10, %xmm3 # -$num
2873 mov %rax, 40(%rsp) # save original %rsp
2874 .cfi_cfa_expression %rsp+40,deref,+8
2877 call __bn_sqrx8x_internal
2878 call __bn_postx4x_internal
2879 call __bn_sqrx8x_internal
2880 call __bn_postx4x_internal
2881 call __bn_sqrx8x_internal
2882 call __bn_postx4x_internal
2883 call __bn_sqrx8x_internal
2884 call __bn_postx4x_internal
2885 call __bn_sqrx8x_internal
2886 call __bn_postx4x_internal
2888 mov %r10,$num # -num
2894 call mulx4x_internal
2896 mov 40(%rsp),%rsi # restore %rsp
2913 .cfi_def_cfa_register %rsp
2917 .size bn_powerx5,.-bn_powerx5
2919 .globl bn_sqrx8x_internal
2920 .hidden bn_sqrx8x_internal
2921 .type bn_sqrx8x_internal,\@abi-omnipotent
2924 __bn_sqrx8x_internal:
2926 ##################################################################
2929 # a) multiply-n-add everything but a[i]*a[i];
2930 # b) shift result of a) by 1 to the left and accumulate
2931 # a[i]*a[i] products;
2933 ##################################################################
2934 # a[7]a[7]a[6]a[6]a[5]a[5]a[4]a[4]a[3]a[3]a[2]a[2]a[1]a[1]a[0]a[0]
2965 # a[7]a[7]a[6]a[6]a[5]a[5]a[4]a[4]a[3]a[3]a[2]a[2]a[1]a[1]a[0]a[0]
2968 my ($zero,$carry)=("%rbp","%rcx");
2971 lea 48+8(%rsp),$tptr
2972 lea ($aptr,$num),$aaptr
2973 mov $num,0+8(%rsp) # save $num
2974 mov $aaptr,8+8(%rsp) # save end of $aptr
2975 jmp .Lsqr8x_zero_start
2978 .byte 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00
2981 movdqa %xmm0,0*8($tptr)
2982 movdqa %xmm0,2*8($tptr)
2983 movdqa %xmm0,4*8($tptr)
2984 movdqa %xmm0,6*8($tptr)
2985 .Lsqr8x_zero_start: # aligned at 32
2986 movdqa %xmm0,8*8($tptr)
2987 movdqa %xmm0,10*8($tptr)
2988 movdqa %xmm0,12*8($tptr)
2989 movdqa %xmm0,14*8($tptr)
2990 lea 16*8($tptr),$tptr
2994 mov 0*8($aptr),%rdx # a[0], modulo-scheduled
2995 #xor %r9,%r9 # t[1], ex-$num, zero already
3002 lea 48+8(%rsp),$tptr
3003 xor $zero,$zero # cf=0, cf=0
3004 jmp .Lsqrx8x_outer_loop
3007 .Lsqrx8x_outer_loop:
3008 mulx 1*8($aptr),%r8,%rax # a[1]*a[0]
3009 adcx %r9,%r8 # a[1]*a[0]+=t[1]
3011 mulx 2*8($aptr),%r9,%rax # a[2]*a[0]
3014 .byte 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00 # mulx 3*8($aptr),%r10,%rax # ...
3017 .byte 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00 # mulx 4*8($aptr),%r11,%rax
3020 mulx 5*8($aptr),%r12,%rax
3023 mulx 6*8($aptr),%r13,%rax
3026 mulx 7*8($aptr),%r14,%r15
3027 mov 1*8($aptr),%rdx # a[1]
3031 mov %r8,1*8($tptr) # t[1]
3032 mov %r9,2*8($tptr) # t[2]
3033 sbb $carry,$carry # mov %cf,$carry
3034 xor $zero,$zero # cf=0, of=0
3037 mulx 2*8($aptr),%r8,%rbx # a[2]*a[1]
3038 mulx 3*8($aptr),%r9,%rax # a[3]*a[1]
3041 mulx 4*8($aptr),%r10,%rbx # ...
3044 .byte 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00 # mulx 5*8($aptr),%r11,%rax
3047 .byte 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00 # mulx 6*8($aptr),%r12,%rbx
3050 .byte 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00 # mulx 7*8($aptr),%r13,%r14
3051 mov 2*8($aptr),%rdx # a[2]
3055 adox $zero,%r14 # of=0
3056 adcx $zero,%r14 # cf=0
3058 mov %r8,3*8($tptr) # t[3]
3059 mov %r9,4*8($tptr) # t[4]
3061 mulx 3*8($aptr),%r8,%rbx # a[3]*a[2]
3062 mulx 4*8($aptr),%r9,%rax # a[4]*a[2]
3065 mulx 5*8($aptr),%r10,%rbx # ...
3068 .byte 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00 # mulx 6*8($aptr),%r11,%rax
3071 .byte 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00 # mulx 7*8($aptr),%r12,%r13
3073 mov 3*8($aptr),%rdx # a[3]
3077 mov %r8,5*8($tptr) # t[5]
3078 mov %r9,6*8($tptr) # t[6]
3079 mulx 4*8($aptr),%r8,%rax # a[4]*a[3]
3080 adox $zero,%r13 # of=0
3081 adcx $zero,%r13 # cf=0
3083 mulx 5*8($aptr),%r9,%rbx # a[5]*a[3]
3086 mulx 6*8($aptr),%r10,%rax # ...
3089 mulx 7*8($aptr),%r11,%r12
3090 mov 4*8($aptr),%rdx # a[4]
3091 mov 5*8($aptr),%r14 # a[5]
3094 mov 6*8($aptr),%r15 # a[6]
3096 adox $zero,%r12 # of=0
3097 adcx $zero,%r12 # cf=0
3099 mov %r8,7*8($tptr) # t[7]
3100 mov %r9,8*8($tptr) # t[8]
3102 mulx %r14,%r9,%rax # a[5]*a[4]
3103 mov 7*8($aptr),%r8 # a[7]
3105 mulx %r15,%r10,%rbx # a[6]*a[4]
3108 mulx %r8,%r11,%rax # a[7]*a[4]
3109 mov %r14,%rdx # a[5]
3112 #adox $zero,%rax # of=0
3113 adcx $zero,%rax # cf=0
3115 mulx %r15,%r14,%rbx # a[6]*a[5]
3116 mulx %r8,%r12,%r13 # a[7]*a[5]
3117 mov %r15,%rdx # a[6]
3118 lea 8*8($aptr),$aptr
3125 mulx %r8,%r8,%r14 # a[7]*a[6]
3130 je .Lsqrx8x_outer_break
3132 neg $carry # mov $carry,%cf
3136 adcx 9*8($tptr),%r9 # +=t[9]
3137 adcx 10*8($tptr),%r10 # ...
3138 adcx 11*8($tptr),%r11
3139 adc 12*8($tptr),%r12
3140 adc 13*8($tptr),%r13
3141 adc 14*8($tptr),%r14
3142 adc 15*8($tptr),%r15
3144 lea 2*64($tptr),$tptr
3145 sbb %rax,%rax # mov %cf,$carry
3147 mov -64($aptr),%rdx # a[0]
3148 mov %rax,16+8(%rsp) # offload $carry
3149 mov $tptr,24+8(%rsp)
3151 #lea 8*8($tptr),$tptr # see 2*8*8($tptr) above
3152 xor %eax,%eax # cf=0, of=0
3158 mulx 0*8($aaptr),%rax,%r8 # a[8]*a[i]
3159 adcx %rax,%rbx # +=t[8]
3162 mulx 1*8($aaptr),%rax,%r9 # ...
3166 mulx 2*8($aaptr),%rax,%r10
3170 mulx 3*8($aaptr),%rax,%r11
3174 .byte 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 # mulx 4*8($aaptr),%rax,%r12
3178 mulx 5*8($aaptr),%rax,%r13
3182 mulx 6*8($aaptr),%rax,%r14
3183 mov %rbx,($tptr,%rcx,8) # store t[8+i]
3188 .byte 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00 # mulx 7*8($aaptr),%rax,%r15
3189 mov 8($aptr,%rcx,8),%rdx # a[i]
3191 adox %rbx,%r15 # %rbx is 0, of=0
3192 adcx %rbx,%r15 # cf=0
3198 lea 8*8($aaptr),$aaptr
3200 cmp 8+8(%rsp),$aaptr # done?
3203 sub 16+8(%rsp),%rbx # mov 16(%rsp),%cf
3214 lea 8*8($tptr),$tptr
3216 sbb %rax,%rax # mov %cf,%rax
3217 xor %ebx,%ebx # cf=0, of=0
3218 mov %rax,16+8(%rsp) # offload carry
3224 sub 16+8(%rsp),%rbx # mov 16(%rsp),%cf
3226 mov 24+8(%rsp),$carry # initial $tptr, borrow $carry
3228 mov 0*8($aptr),%rdx # a[8], modulo-scheduled
3236 cmp $carry,$tptr # cf=0, of=0
3237 je .Lsqrx8x_outer_loop
3242 mov 2*8($carry),%r10
3244 mov 3*8($carry),%r11
3246 mov 4*8($carry),%r12
3248 mov 5*8($carry),%r13
3250 mov 6*8($carry),%r14
3252 mov 7*8($carry),%r15
3254 jmp .Lsqrx8x_outer_loop
3257 .Lsqrx8x_outer_break:
3258 mov %r9,9*8($tptr) # t[9]
3259 movq %xmm3,%rcx # -$num
3260 mov %r10,10*8($tptr) # ...
3261 mov %r11,11*8($tptr)
3262 mov %r12,12*8($tptr)
3263 mov %r13,13*8($tptr)
3264 mov %r14,14*8($tptr)
3269 lea 48+8(%rsp),$tptr
3270 mov ($aptr,$i),%rdx # a[0]
3272 mov 8($tptr),$A0[1] # t[1]
3273 xor $A0[0],$A0[0] # t[0], of=0, cf=0
3274 mov 0+8(%rsp),$num # restore $num
3276 mov 16($tptr),$A1[0] # t[2] # prefetch
3277 mov 24($tptr),$A1[1] # t[3] # prefetch
3278 #jmp .Lsqrx4x_shift_n_add # happens to be aligned
3281 .Lsqrx4x_shift_n_add:
3285 .byte 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00 # mov 8($aptr,$i),%rdx # a[i+1] # prefetch
3286 .byte 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00 # mov 32($tptr),$A0[0] # t[2*i+4] # prefetch
3289 mov 40($tptr),$A0[1] # t[2*i+4+1] # prefetch
3296 mov 16($aptr,$i),%rdx # a[i+2] # prefetch
3297 mov 48($tptr),$A1[0] # t[2*i+6] # prefetch
3300 mov 56($tptr),$A1[1] # t[2*i+6+1] # prefetch
3307 mov 24($aptr,$i),%rdx # a[i+3] # prefetch
3309 mov 64($tptr),$A0[0] # t[2*i+8] # prefetch
3312 mov 72($tptr),$A0[1] # t[2*i+8+1] # prefetch
3319 jrcxz .Lsqrx4x_shift_n_add_break
3320 .byte 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00 # mov 0($aptr,$i),%rdx # a[i+4] # prefetch
3323 mov 80($tptr),$A1[0] # t[2*i+10] # prefetch
3324 mov 88($tptr),$A1[1] # t[2*i+10+1] # prefetch
3329 jmp .Lsqrx4x_shift_n_add
3332 .Lsqrx4x_shift_n_add_break:
3336 lea 64($tptr),$tptr # end of t[] buffer
3339 ######################################################################
3340 # Montgomery reduction part, "word-by-word" algorithm.
3342 # This new path is inspired by multiple submissions from Intel, by
3343 # Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford,
3346 my ($nptr,$carry,$m0)=("%rbp","%rsi","%rdx");
3350 __bn_sqrx8x_reduction:
3351 xor %eax,%eax # initial top-most carry bit
3352 mov 32+8(%rsp),%rbx # n0
3353 mov 48+8(%rsp),%rdx # "%r8", 8*0($tptr)
3354 lea -8*8($nptr,$num),%rcx # end of n[]
3355 #lea 48+8(%rsp,$num,2),$tptr # end of t[] buffer
3356 mov %rcx, 0+8(%rsp) # save end of n[]
3357 mov $tptr,8+8(%rsp) # save end of t[]
3359 lea 48+8(%rsp),$tptr # initial t[] window
3360 jmp .Lsqrx8x_reduction_loop
3363 .Lsqrx8x_reduction_loop:
3369 imulq %rbx,%rdx # n0*a[i]
3373 mov %rax,24+8(%rsp) # store top-most carry bit
3375 lea 8*8($tptr),$tptr
3376 xor $carry,$carry # cf=0,of=0
3383 mulx 8*0($nptr),%rax,%r8 # n[0]
3384 adcx %rbx,%rax # discarded
3387 mulx 8*1($nptr),%rbx,%r9 # n[1]
3391 mulx 8*2($nptr),%rbx,%r10
3395 mulx 8*3($nptr),%rbx,%r11
3399 .byte 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00 # mulx 8*4($nptr),%rbx,%r12
3405 mulx 32+8(%rsp),%rbx,%rdx # %rdx discarded
3407 mov %rax,64+48+8(%rsp,%rcx,8) # put aside n0*a[i]
3409 mulx 8*5($nptr),%rax,%r13
3413 mulx 8*6($nptr),%rax,%r14
3417 mulx 8*7($nptr),%rax,%r15
3420 adox $carry,%r15 # $carry is 0
3421 adcx $carry,%r15 # cf=0
3423 .byte 0x67,0x67,0x67
3427 mov $carry,%rax # xor %rax,%rax
3428 cmp 0+8(%rsp),$nptr # end of n[]?
3429 jae .Lsqrx8x_no_tail
3431 mov 48+8(%rsp),%rdx # pull n0*a[0]
3433 lea 8*8($nptr),$nptr
3436 adcx 8*2($tptr),%r10
3442 lea 8*8($tptr),$tptr
3443 sbb %rax,%rax # top carry
3445 xor $carry,$carry # of=0, cf=0
3452 mulx 8*0($nptr),%rax,%r8
3456 mulx 8*1($nptr),%rax,%r9
3460 mulx 8*2($nptr),%rax,%r10
3464 mulx 8*3($nptr),%rax,%r11
3468 .byte 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 # mulx 8*4($nptr),%rax,%r12
3472 mulx 8*5($nptr),%rax,%r13
3476 mulx 8*6($nptr),%rax,%r14
3480 mulx 8*7($nptr),%rax,%r15
3481 mov 72+48+8(%rsp,%rcx,8),%rdx # pull n0*a[i]
3484 mov %rbx,($tptr,%rcx,8) # save result
3486 adcx $carry,%r15 # cf=0
3491 cmp 0+8(%rsp),$nptr # end of n[]?
3492 jae .Lsqrx8x_tail_done # break out of loop
3494 sub 16+8(%rsp),$carry # mov 16(%rsp),%cf
3495 mov 48+8(%rsp),%rdx # pull n0*a[0]
3496 lea 8*8($nptr),$nptr
3505 lea 8*8($tptr),$tptr
3507 sub \$8,%rcx # mov \$-8,%rcx
3509 xor $carry,$carry # of=0, cf=0
3516 add 24+8(%rsp),%r8 # can this overflow?
3526 sub 16+8(%rsp),$carry # mov 16(%rsp),%cf
3527 .Lsqrx8x_no_tail: # %cf is 0 if jumped here
3531 mov 8*7($nptr),$carry
3532 movq %xmm2,$nptr # restore $nptr
3539 adc \$0,%rax # top-most carry
3541 mov 32+8(%rsp),%rbx # n0
3542 mov 8*8($tptr,%rcx),%rdx # modulo-scheduled "%r8"
3544 mov %r8,8*0($tptr) # store top 512 bits
3545 lea 8*8($tptr),%r8 # borrow %r8
3554 lea 8*8($tptr,%rcx),$tptr # start of current t[] window
3555 cmp 8+8(%rsp),%r8 # end of t[]?
3556 jb .Lsqrx8x_reduction_loop
3559 .size bn_sqrx8x_internal,.-bn_sqrx8x_internal
3562 ##############################################################
3563 # Post-condition, 4x unrolled
3566 my ($rptr,$nptr)=("%rdx","%rbp");
3569 __bn_postx4x_internal:
3572 mov %rcx,%r10 # -$num
3573 mov %rcx,%r9 # -$num
3576 #lea 48+8(%rsp,%r9),$tptr
3577 movq %xmm1,$rptr # restore $rptr
3578 movq %xmm1,$aptr # prepare for back-to-back call
3579 dec %r12 # so that after 'not' we get -n[0]
3584 jmp .Lsqrx4x_sub_entry
3594 lea 8*4($nptr),$nptr
3599 neg %r8 # mov %r8,%cf
3605 lea 8*4($tptr),$tptr
3607 sbb %r8,%r8 # mov %cf,%r8
3610 lea 8*4($rptr),$rptr
3615 neg %r9 # restore $num
3619 .size __bn_postx4x_internal,.-__bn_postx4x_internal
3624 my ($inp,$num,$tbl,$idx)=$win64?("%rcx","%edx","%r8", "%r9d") : # Win64 order
3625 ("%rdi","%esi","%rdx","%ecx"); # Unix order
3632 .type bn_get_bits5,\@abi-omnipotent
3645 movzw (%r10,$num,2),%eax
3650 .size bn_get_bits5,.-bn_get_bits5
3653 .type bn_scatter5,\@abi-omnipotent
3658 jz .Lscatter_epilogue
3659 lea ($tbl,$idx,8),$tbl
3670 .size bn_scatter5,.-bn_scatter5
3673 .type bn_gather5,\@abi-omnipotent
3676 .LSEH_begin_bn_gather5: # Win64 thing, but harmless in other cases
3678 # I can't trust assembler to use specific encoding:-(
3679 .byte 0x4c,0x8d,0x14,0x24 #lea (%rsp),%r10
3680 .byte 0x48,0x81,0xec,0x08,0x01,0x00,0x00 #sub $0x108,%rsp
3681 lea .Linc(%rip),%rax
3682 and \$-16,%rsp # shouldn't be formally required
3685 movdqa 0(%rax),%xmm0 # 00000001000000010000000000000000
3686 movdqa 16(%rax),%xmm1 # 00000002000000020000000200000002
3687 lea 128($tbl),%r11 # size optimization
3688 lea 128(%rsp),%rax # size optimization
3690 pshufd \$0,%xmm5,%xmm5 # broadcast $idx
3694 ########################################################################
3695 # calculate mask by comparing 0..31 to $idx and save result to stack
3697 for($i=0;$i<$STRIDE/16;$i+=4) {
3700 pcmpeqd %xmm5,%xmm0 # compare to 1,0
3702 $code.=<<___ if ($i);
3703 movdqa %xmm3,`16*($i-1)-128`(%rax)
3709 pcmpeqd %xmm5,%xmm1 # compare to 3,2
3710 movdqa %xmm0,`16*($i+0)-128`(%rax)
3714 pcmpeqd %xmm5,%xmm2 # compare to 5,4
3715 movdqa %xmm1,`16*($i+1)-128`(%rax)
3719 pcmpeqd %xmm5,%xmm3 # compare to 7,6
3720 movdqa %xmm2,`16*($i+2)-128`(%rax)
3725 movdqa %xmm3,`16*($i-1)-128`(%rax)
3733 for($i=0;$i<$STRIDE/16;$i+=4) {
3735 movdqa `16*($i+0)-128`(%r11),%xmm0
3736 movdqa `16*($i+1)-128`(%r11),%xmm1
3737 movdqa `16*($i+2)-128`(%r11),%xmm2
3738 pand `16*($i+0)-128`(%rax),%xmm0
3739 movdqa `16*($i+3)-128`(%r11),%xmm3
3740 pand `16*($i+1)-128`(%rax),%xmm1
3742 pand `16*($i+2)-128`(%rax),%xmm2
3744 pand `16*($i+3)-128`(%rax),%xmm3
3751 lea $STRIDE(%r11),%r11
3752 pshufd \$0x4e,%xmm4,%xmm0
3754 movq %xmm0,($out) # m0=bp[0]
3761 .LSEH_end_bn_gather5:
3763 .size bn_gather5,.-bn_gather5
3771 .asciz "Montgomery Multiplication with scatter/gather for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
3774 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
3775 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
3783 .extern __imp_RtlVirtualUnwind
3784 .type mul_handler,\@abi-omnipotent
3798 mov 120($context),%rax # pull context->Rax
3799 mov 248($context),%rbx # pull context->Rip
3801 mov 8($disp),%rsi # disp->ImageBase
3802 mov 56($disp),%r11 # disp->HandlerData
3804 mov 0(%r11),%r10d # HandlerData[0]
3805 lea (%rsi,%r10),%r10 # end of prologue label
3806 cmp %r10,%rbx # context->Rip<end of prologue label
3807 jb .Lcommon_seh_tail
3809 mov 4(%r11),%r10d # HandlerData[1]
3810 lea (%rsi,%r10),%r10 # beginning of body label
3811 cmp %r10,%rbx # context->Rip<body label
3812 jb .Lcommon_pop_regs
3814 mov 152($context),%rax # pull context->Rsp
3816 mov 8(%r11),%r10d # HandlerData[2]
3817 lea (%rsi,%r10),%r10 # epilogue label
3818 cmp %r10,%rbx # context->Rip>=epilogue label
3819 jae .Lcommon_seh_tail
3821 lea .Lmul_epilogue(%rip),%r10
3825 mov 192($context),%r10 # pull $num
3826 mov 8(%rax,%r10,8),%rax # pull saved stack pointer
3828 jmp .Lcommon_pop_regs
3831 mov 40(%rax),%rax # pull saved stack pointer
3839 mov %rbx,144($context) # restore context->Rbx
3840 mov %rbp,160($context) # restore context->Rbp
3841 mov %r12,216($context) # restore context->R12
3842 mov %r13,224($context) # restore context->R13
3843 mov %r14,232($context) # restore context->R14
3844 mov %r15,240($context) # restore context->R15
3849 mov %rax,152($context) # restore context->Rsp
3850 mov %rsi,168($context) # restore context->Rsi
3851 mov %rdi,176($context) # restore context->Rdi
3853 mov 40($disp),%rdi # disp->ContextRecord
3854 mov $context,%rsi # context
3855 mov \$154,%ecx # sizeof(CONTEXT)
3856 .long 0xa548f3fc # cld; rep movsq
3859 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
3860 mov 8(%rsi),%rdx # arg2, disp->ImageBase
3861 mov 0(%rsi),%r8 # arg3, disp->ControlPc
3862 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
3863 mov 40(%rsi),%r10 # disp->ContextRecord
3864 lea 56(%rsi),%r11 # &disp->HandlerData
3865 lea 24(%rsi),%r12 # &disp->EstablisherFrame
3866 mov %r10,32(%rsp) # arg5
3867 mov %r11,40(%rsp) # arg6
3868 mov %r12,48(%rsp) # arg7
3869 mov %rcx,56(%rsp) # arg8, (NULL)
3870 call *__imp_RtlVirtualUnwind(%rip)
3872 mov \$1,%eax # ExceptionContinueSearch
3884 .size mul_handler,.-mul_handler
3888 .rva .LSEH_begin_bn_mul_mont_gather5
3889 .rva .LSEH_end_bn_mul_mont_gather5
3890 .rva .LSEH_info_bn_mul_mont_gather5
3892 .rva .LSEH_begin_bn_mul4x_mont_gather5
3893 .rva .LSEH_end_bn_mul4x_mont_gather5
3894 .rva .LSEH_info_bn_mul4x_mont_gather5
3896 .rva .LSEH_begin_bn_power5
3897 .rva .LSEH_end_bn_power5
3898 .rva .LSEH_info_bn_power5
3900 .rva .LSEH_begin_bn_from_mont8x
3901 .rva .LSEH_end_bn_from_mont8x
3902 .rva .LSEH_info_bn_from_mont8x
3904 $code.=<<___ if ($addx);
3905 .rva .LSEH_begin_bn_mulx4x_mont_gather5
3906 .rva .LSEH_end_bn_mulx4x_mont_gather5
3907 .rva .LSEH_info_bn_mulx4x_mont_gather5
3909 .rva .LSEH_begin_bn_powerx5
3910 .rva .LSEH_end_bn_powerx5
3911 .rva .LSEH_info_bn_powerx5
3914 .rva .LSEH_begin_bn_gather5
3915 .rva .LSEH_end_bn_gather5
3916 .rva .LSEH_info_bn_gather5
3920 .LSEH_info_bn_mul_mont_gather5:
3923 .rva .Lmul_body,.Lmul_body,.Lmul_epilogue # HandlerData[]
3925 .LSEH_info_bn_mul4x_mont_gather5:
3928 .rva .Lmul4x_prologue,.Lmul4x_body,.Lmul4x_epilogue # HandlerData[]
3930 .LSEH_info_bn_power5:
3933 .rva .Lpower5_prologue,.Lpower5_body,.Lpower5_epilogue # HandlerData[]
3935 .LSEH_info_bn_from_mont8x:
3938 .rva .Lfrom_prologue,.Lfrom_body,.Lfrom_epilogue # HandlerData[]
3940 $code.=<<___ if ($addx);
3942 .LSEH_info_bn_mulx4x_mont_gather5:
3945 .rva .Lmulx4x_prologue,.Lmulx4x_body,.Lmulx4x_epilogue # HandlerData[]
3947 .LSEH_info_bn_powerx5:
3950 .rva .Lpowerx5_prologue,.Lpowerx5_body,.Lpowerx5_epilogue # HandlerData[]
3954 .LSEH_info_bn_gather5:
3955 .byte 0x01,0x0b,0x03,0x0a
3956 .byte 0x0b,0x01,0x21,0x00 # sub rsp,0x108
3957 .byte 0x04,0xa3,0x00,0x00 # lea r10,(rsp)
3962 $code =~ s/\`([^\`]*)\`/eval($1)/gem;