3 # ====================================================================
4 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
10 # This module implements Poly1305 hash for x86_64.
14 # Numbers are cycles per processed byte with poly1305_blocks alone,
15 # measured with rdtsc at fixed clock frequency.
17 # IALU/gcc-4.8(*) AVX(**) AVX2
20 # Westmere 1.86/+120% -
21 # Sandy Bridge 1.39/+140% 1.10
22 # Haswell 1.10/+175% 1.11 0.65
23 # Skylake 1.12/+120% 0.96 0.51
24 # Silvermont 2.83/+95% -
25 # VIA Nano 1.82/+150% -
26 # Sledgehammer 1.38/+160% -
27 # Bulldozer 2.21/+130% 0.97
29 # (*) improvement coefficients relative to clang are more modest and
30 # are ~50% on most processors, in both cases we are comparing to
32 # (**) SSE2 implementation was attempted, but among non-AVX processors
33 # it was faster than integer-only code only on older Intel P4 and
34 # Core processors, 50-30%, less newer processor is, but slower on
35 # contemporary ones, for example almost 2x slower on Atom, and as
36 # former are naturally disappearing, SSE2 is deemed unnecessary;
40 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
42 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
44 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
45 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
46 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
47 die "can't locate x86_64-xlate.pl";
49 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
50 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
51 $avx = ($1>=2.19) + ($1>=2.22);
54 if (!$avx && $win64 && ($flavour =~ /nasm/ || $ENV{ASM} =~ /nasm/) &&
55 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
56 $avx = ($1>=2.09) + ($1>=2.10);
59 if (!$avx && $win64 && ($flavour =~ /masm/ || $ENV{ASM} =~ /ml64/) &&
60 `ml64 2>&1` =~ /Version ([0-9]+)\./) {
61 $avx = ($1>=10) + ($1>=12);
64 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|.*based on LLVM) ([3-9]\.[0-9]+)/) {
65 $avx = ($2>=3.0) + ($2>3.0);
68 open OUT,"| \"$^X\" $xlate $flavour $output";
71 my ($ctx,$inp,$len,$padbit)=("%rdi","%rsi","%rdx","%rcx");
72 my ($mac,$nonce)=($inp,$len); # *_emit arguments
73 my ($d1,$d2,$d3, $r0,$r1,$s1)=map("%r$_",(8..13));
74 my ($h0,$h1,$h2)=("%r14","%rbx","%rbp");
76 sub poly1305_iteration {
77 # input: copy of $r1 in %rax, $h0-$h2, $r0-$r1
78 # output: $h0-$h2 *= $r0-$r1
86 mov %rax,$h0 # future $h0
96 mov $h2,$h1 # borrow $h1
100 imulq $s1,$h1 # h2*s1
105 imulq $r0,$h2 # h2*r0
107 mov \$-4,%rax # mask value
110 and $d3,%rax # last reduction step
120 ########################################################################
121 # Layout of opaque area is following.
123 # unsigned __int64 h[3]; # current hash value base 2^64
124 # unsigned __int64 r[2]; # key value base 2^64
129 .extern OPENSSL_ia32cap_P
132 .globl poly1305_blocks
134 .type poly1305_init,\@function,3
138 mov %rax,0($ctx) # initialize hash value
145 lea poly1305_blocks(%rip),%r10
146 lea poly1305_emit(%rip),%r11
148 $code.=<<___ if ($avx);
149 mov OPENSSL_ia32cap_P+4(%rip),%r9
150 lea poly1305_blocks_avx(%rip),%rax
151 lea poly1305_emit_avx(%rip),%rcx
152 bt \$`60-32`,%r9 # AVX?
156 $code.=<<___ if ($avx>1);
157 lea poly1305_blocks_avx2(%rip),%rax
158 bt \$`5+32`,%r9 # AVX2?
162 mov \$0x0ffffffc0fffffff,%rax
163 mov \$0x0ffffffc0ffffffc,%rcx
175 .size poly1305_init,.-poly1305_init
177 .type poly1305_blocks,\@function,4
181 sub \$16,$len # too short?
192 mov $len,%r15 # reassign $len
194 mov 24($ctx),$r0 # load r
197 mov 0($ctx),$h0 # load hash value
204 add $r1,$s1 # s1 = r1 + (r1 >> 2)
209 add 0($inp),$h0 # accumulate input
214 &poly1305_iteration();
217 sub \$16,%r15 # len-=16
220 mov $h0,0($ctx) # store hash value
234 .size poly1305_blocks,.-poly1305_blocks
236 .type poly1305_emit,\@function,3
240 mov 0($ctx),%r8 # load hash value
245 add \$5,%r8 # compare to modulus
249 shr \$2,%r10 # did 130-bit value overfow?
253 add 0($nonce),%rax # accumulate nonce
255 mov %rax,0($mac) # write result
259 .size poly1305_emit,.-poly1305_emit
263 ########################################################################
264 # Layout of opaque area is following.
266 # unsigned __int32 h[5]; # current hash value base 2^26
267 # unsigned __int32 is_base2_26;
268 # unsigned __int64 r[2]; # key value base 2^64
269 # unsigned __int64 pad;
270 # struct { unsigned __int32 r^2, r^1, r^4, r^3; } r[9];
272 # where r^n are base 2^26 digits of degrees of multiplier key. There are
273 # 5 digits, but last four are interleaved with multiples of 5, totalling
274 # in 9 elements: r0, r1, 5*r1, r2, 5*r2, r3, 5*r3, r4, 5*r4.
276 my ($H0,$H1,$H2,$H3,$H4, $T0,$T1,$T2,$T3,$T4, $D0,$D1,$D2,$D3,$D4, $MASK) =
277 map("%xmm$_",(0..15));
280 .type __poly1305_block,\@abi-omnipotent
284 &poly1305_iteration();
287 .size __poly1305_block,.-__poly1305_block
289 .type __poly1305_init_avx,\@abi-omnipotent
296 lea 48+64($ctx),$ctx # size optimization
299 call __poly1305_block # r^2
301 mov \$0x3ffffff,%eax # save interleaved r^2 and r base 2^26
307 mov %eax,`16*0+0-64`($ctx)
309 mov %edx,`16*0+4-64`($ctx)
316 mov %eax,`16*1+0-64`($ctx)
317 lea (%rax,%rax,4),%eax # *5
318 mov %edx,`16*1+4-64`($ctx)
319 lea (%rdx,%rdx,4),%edx # *5
320 mov %eax,`16*2+0-64`($ctx)
322 mov %edx,`16*2+4-64`($ctx)
333 mov %eax,`16*3+0-64`($ctx)
334 lea (%rax,%rax,4),%eax # *5
335 mov %edx,`16*3+4-64`($ctx)
336 lea (%rdx,%rdx,4),%edx # *5
337 mov %eax,`16*4+0-64`($ctx)
339 mov %edx,`16*4+4-64`($ctx)
348 mov %eax,`16*5+0-64`($ctx)
349 lea (%rax,%rax,4),%eax # *5
350 mov %edx,`16*5+4-64`($ctx)
351 lea (%rdx,%rdx,4),%edx # *5
352 mov %eax,`16*6+0-64`($ctx)
354 mov %edx,`16*6+4-64`($ctx)
360 mov $d1#d,`16*7+0-64`($ctx)
361 lea ($d1,$d1,4),$d1 # *5
362 mov $d2#d,`16*7+4-64`($ctx)
363 lea ($d2,$d2,4),$d2 # *5
364 mov $d1#d,`16*8+0-64`($ctx)
365 mov $d2#d,`16*8+4-64`($ctx)
368 call __poly1305_block # r^3
370 mov \$0x3ffffff,%eax # save r^3 base 2^26
374 mov %eax,`16*0+12-64`($ctx)
378 mov %edx,`16*1+12-64`($ctx)
379 lea (%rdx,%rdx,4),%edx # *5
381 mov %edx,`16*2+12-64`($ctx)
387 mov %eax,`16*3+12-64`($ctx)
388 lea (%rax,%rax,4),%eax # *5
390 mov %eax,`16*4+12-64`($ctx)
395 mov %edx,`16*5+12-64`($ctx)
396 lea (%rdx,%rdx,4),%edx # *5
398 mov %edx,`16*6+12-64`($ctx)
403 mov $d1#d,`16*7+12-64`($ctx)
404 lea ($d1,$d1,4),$d1 # *5
405 mov $d1#d,`16*8+12-64`($ctx)
408 call __poly1305_block # r^4
410 mov \$0x3ffffff,%eax # save r^4 base 2^26
414 mov %eax,`16*0+8-64`($ctx)
418 mov %edx,`16*1+8-64`($ctx)
419 lea (%rdx,%rdx,4),%edx # *5
421 mov %edx,`16*2+8-64`($ctx)
427 mov %eax,`16*3+8-64`($ctx)
428 lea (%rax,%rax,4),%eax # *5
430 mov %eax,`16*4+8-64`($ctx)
435 mov %edx,`16*5+8-64`($ctx)
436 lea (%rdx,%rdx,4),%edx # *5
438 mov %edx,`16*6+8-64`($ctx)
443 mov $d1#d,`16*7+8-64`($ctx)
444 lea ($d1,$d1,4),$d1 # *5
445 mov $d1#d,`16*8+8-64`($ctx)
447 lea -48-64($ctx),$ctx # size [de-]optimization
449 .size __poly1305_init_avx,.-__poly1305_init_avx
451 .type poly1305_blocks_avx,\@function,4
454 mov 20($ctx),%r8d # is_base2_26
480 mov $len,%r15 # reassign $len
482 mov 0($ctx),$d1 # load hash value
486 mov 24($ctx),$r0 # load r
489 ################################# base 2^26 -> base 2^64
492 mov $d2,$r1 # borrow $r1
508 adc \$0,$h2 # can be partially reduced...
510 mov \$-4,$d2 # ... so reduce
522 add $r1,$s1 # s1 = r1 + (r1 >> 2)
524 add 0($inp),$h0 # accumulate input
529 call __poly1305_block
531 test $padbit,$padbit # if $padbit is zero,
532 jz .Lstore_base2_64_avx # store hash in base 2^64 format
534 ################################# base 2^64 -> base 2^26
541 and \$0x3ffffff,%rax # h[0]
543 and \$0x3ffffff,%rdx # h[1]
547 and \$0x3ffffff,$h0 # h[2]
549 and \$0x3ffffff,$h1 # h[3]
553 jz .Lstore_base2_26_avx
563 .Lstore_base2_64_avx:
566 mov $h2,16($ctx) # note that is_base2_26 is zeroed
570 .Lstore_base2_26_avx:
571 mov %rax#d,0($ctx) # store hash value base 2^26
586 .Lblocks_avx_epilogue:
599 mov $len,%r15 # reassign $len
601 mov 24($ctx),$r0 # load r
604 mov 0($ctx),$h0 # load hash value
611 add $r1,$s1 # s1 = r1 + (r1 >> 2)
616 add 0($inp),$h0 # accumulate input
622 call __poly1305_block
625 ################################# base 2^64 -> base 2^26
632 and \$0x3ffffff,%rax # h[0]
634 and \$0x3ffffff,%rdx # h[1]
638 and \$0x3ffffff,$h0 # h[2]
640 and \$0x3ffffff,$h1 # h[3]
648 movl \$1,20($ctx) # set is_base2_26
650 call __poly1305_init_avx
663 .Lbase2_64_avx_epilogue:
668 vmovd 4*0($ctx),$H0 # load hash value
676 $code.=<<___ if (!$win64);
680 $code.=<<___ if ($win64);
683 vmovdqa %xmm6,0x50(%r11)
684 vmovdqa %xmm7,0x60(%r11)
685 vmovdqa %xmm8,0x70(%r11)
686 vmovdqa %xmm9,0x80(%r11)
687 vmovdqa %xmm10,0x90(%r11)
688 vmovdqa %xmm11,0xa0(%r11)
689 vmovdqa %xmm12,0xb0(%r11)
690 vmovdqa %xmm13,0xc0(%r11)
691 vmovdqa %xmm14,0xd0(%r11)
692 vmovdqa %xmm15,0xe0(%r11)
700 vmovdqu `16*3`($ctx),$D4 # preload r0^2
701 lea `16*3+64`($ctx),$ctx # size optimization
702 lea .Lconst(%rip),%rcx
704 ################################################################
706 vmovdqu 16*2($inp),$T0
707 vmovdqu 16*3($inp),$T1
708 vmovdqa 64(%rcx),$MASK # .Lmask26
710 vpsrldq \$6,$T0,$T2 # splat input
712 vpunpckhqdq $T1,$T0,$T4 # 4
713 vpunpcklqdq $T1,$T0,$T0 # 0:1
714 vpunpcklqdq $T3,$T2,$T3 # 2:3
716 vpsrlq \$40,$T4,$T4 # 4
718 vpand $MASK,$T0,$T0 # 0
720 vpand $MASK,$T1,$T1 # 1
722 vpand $MASK,$T2,$T2 # 2
723 vpand $MASK,$T3,$T3 # 3
724 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
728 # expand and copy pre-calculated table to stack
729 vmovdqu `16*1-64`($ctx),$D1
730 vmovdqu `16*2-64`($ctx),$D2
731 vpshufd \$0xEE,$D4,$D3 # 34xx -> 3434
732 vpshufd \$0x44,$D4,$D0 # xx12 -> 1212
733 vmovdqa $D3,-0x90(%r11)
734 vmovdqa $D0,0x00(%rsp)
735 vpshufd \$0xEE,$D1,$D4
736 vmovdqu `16*3-64`($ctx),$D0
737 vpshufd \$0x44,$D1,$D1
738 vmovdqa $D4,-0x80(%r11)
739 vmovdqa $D1,0x10(%rsp)
740 vpshufd \$0xEE,$D2,$D3
741 vmovdqu `16*4-64`($ctx),$D1
742 vpshufd \$0x44,$D2,$D2
743 vmovdqa $D3,-0x70(%r11)
744 vmovdqa $D2,0x20(%rsp)
745 vpshufd \$0xEE,$D0,$D4
746 vmovdqu `16*5-64`($ctx),$D2
747 vpshufd \$0x44,$D0,$D0
748 vmovdqa $D4,-0x60(%r11)
749 vmovdqa $D0,0x30(%rsp)
750 vpshufd \$0xEE,$D1,$D3
751 vmovdqu `16*6-64`($ctx),$D0
752 vpshufd \$0x44,$D1,$D1
753 vmovdqa $D3,-0x50(%r11)
754 vmovdqa $D1,0x40(%rsp)
755 vpshufd \$0xEE,$D2,$D4
756 vmovdqu `16*7-64`($ctx),$D1
757 vpshufd \$0x44,$D2,$D2
758 vmovdqa $D4,-0x40(%r11)
759 vmovdqa $D2,0x50(%rsp)
760 vpshufd \$0xEE,$D0,$D3
761 vmovdqu `16*8-64`($ctx),$D2
762 vpshufd \$0x44,$D0,$D0
763 vmovdqa $D3,-0x30(%r11)
764 vmovdqa $D0,0x60(%rsp)
765 vpshufd \$0xEE,$D1,$D4
766 vpshufd \$0x44,$D1,$D1
767 vmovdqa $D4,-0x20(%r11)
768 vmovdqa $D1,0x70(%rsp)
769 vpshufd \$0xEE,$D2,$D3
770 vmovdqa 0x00(%rsp),$D4 # preload r0^2
771 vpshufd \$0x44,$D2,$D2
772 vmovdqa $D3,-0x10(%r11)
773 vmovdqa $D2,0x80(%rsp)
779 ################################################################
780 # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2
781 # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^3+inp[7]*r
782 # \___________________/
783 # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2+inp[8])*r^2
784 # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^4+inp[7]*r^2+inp[9])*r
785 # \___________________/ \____________________/
787 # Note that we start with inp[2:3]*r^2. This is because it
788 # doesn't depend on reduction in previous iteration.
789 ################################################################
790 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
791 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
792 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
793 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
794 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
796 # though note that $Tx and $Hx are "reversed" in this section,
797 # and $D4 is preloaded with r0^2...
799 vpmuludq $T0,$D4,$D0 # d0 = h0*r0
800 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
801 vmovdqa $H2,0x20(%r11) # offload hash
802 vpmuludq $T2,$D4,$D2 # d3 = h2*r0
803 vmovdqa 0x10(%rsp),$H2 # r1^2
804 vpmuludq $T3,$D4,$D3 # d3 = h3*r0
805 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
807 vmovdqa $H0,0x00(%r11) #
808 vpmuludq 0x20(%rsp),$T4,$H0 # h4*s1
809 vmovdqa $H1,0x10(%r11) #
810 vpmuludq $T3,$H2,$H1 # h3*r1
811 vpaddq $H0,$D0,$D0 # d0 += h4*s1
812 vpaddq $H1,$D4,$D4 # d4 += h3*r1
813 vmovdqa $H3,0x30(%r11) #
814 vpmuludq $T2,$H2,$H0 # h2*r1
815 vpmuludq $T1,$H2,$H1 # h1*r1
816 vpaddq $H0,$D3,$D3 # d3 += h2*r1
817 vmovdqa 0x30(%rsp),$H3 # r2^2
818 vpaddq $H1,$D2,$D2 # d2 += h1*r1
819 vmovdqa $H4,0x40(%r11) #
820 vpmuludq $T0,$H2,$H2 # h0*r1
821 vpmuludq $T2,$H3,$H0 # h2*r2
822 vpaddq $H2,$D1,$D1 # d1 += h0*r1
824 vmovdqa 0x40(%rsp),$H4 # s2^2
825 vpaddq $H0,$D4,$D4 # d4 += h2*r2
826 vpmuludq $T1,$H3,$H1 # h1*r2
827 vpmuludq $T0,$H3,$H3 # h0*r2
828 vpaddq $H1,$D3,$D3 # d3 += h1*r2
829 vmovdqa 0x50(%rsp),$H2 # r3^2
830 vpaddq $H3,$D2,$D2 # d2 += h0*r2
831 vpmuludq $T4,$H4,$H0 # h4*s2
832 vpmuludq $T3,$H4,$H4 # h3*s2
833 vpaddq $H0,$D1,$D1 # d1 += h4*s2
834 vmovdqa 0x60(%rsp),$H3 # s3^2
835 vpaddq $H4,$D0,$D0 # d0 += h3*s2
837 vmovdqa 0x80(%rsp),$H4 # s4^2
838 vpmuludq $T1,$H2,$H1 # h1*r3
839 vpmuludq $T0,$H2,$H2 # h0*r3
840 vpaddq $H1,$D4,$D4 # d4 += h1*r3
841 vpaddq $H2,$D3,$D3 # d3 += h0*r3
842 vpmuludq $T4,$H3,$H0 # h4*s3
843 vpmuludq $T3,$H3,$H1 # h3*s3
844 vpaddq $H0,$D2,$D2 # d2 += h4*s3
845 vmovdqu 16*0($inp),$H0 # load input
846 vpaddq $H1,$D1,$D1 # d1 += h3*s3
847 vpmuludq $T2,$H3,$H3 # h2*s3
848 vpmuludq $T2,$H4,$T2 # h2*s4
849 vpaddq $H3,$D0,$D0 # d0 += h2*s3
851 vmovdqu 16*1($inp),$H1 #
852 vpaddq $T2,$D1,$D1 # d1 += h2*s4
853 vpmuludq $T3,$H4,$T3 # h3*s4
854 vpmuludq $T4,$H4,$T4 # h4*s4
855 vpsrldq \$6,$H0,$H2 # splat input
856 vpaddq $T3,$D2,$D2 # d2 += h3*s4
857 vpaddq $T4,$D3,$D3 # d3 += h4*s4
858 vpsrldq \$6,$H1,$H3 #
859 vpmuludq 0x70(%rsp),$T0,$T4 # h0*r4
860 vpmuludq $T1,$H4,$T0 # h1*s4
861 vpunpckhqdq $H1,$H0,$H4 # 4
862 vpaddq $T4,$D4,$D4 # d4 += h0*r4
863 vmovdqa -0x90(%r11),$T4 # r0^4
864 vpaddq $T0,$D0,$D0 # d0 += h1*s4
866 vpunpcklqdq $H1,$H0,$H0 # 0:1
867 vpunpcklqdq $H3,$H2,$H3 # 2:3
869 #vpsrlq \$40,$H4,$H4 # 4
870 vpsrldq \$`40/8`,$H4,$H4 # 4
872 vpand $MASK,$H0,$H0 # 0
874 vpand $MASK,$H1,$H1 # 1
875 vpand 0(%rcx),$H4,$H4 # .Lmask24
877 vpand $MASK,$H2,$H2 # 2
878 vpand $MASK,$H3,$H3 # 3
879 vpor 32(%rcx),$H4,$H4 # padbit, yes, always
881 vpaddq 0x00(%r11),$H0,$H0 # add hash value
882 vpaddq 0x10(%r11),$H1,$H1
883 vpaddq 0x20(%r11),$H2,$H2
884 vpaddq 0x30(%r11),$H3,$H3
885 vpaddq 0x40(%r11),$H4,$H4
892 ################################################################
893 # Now we accumulate (inp[0:1]+hash)*r^4
894 ################################################################
895 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
896 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
897 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
898 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
899 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
901 vpmuludq $H0,$T4,$T0 # h0*r0
902 vpmuludq $H1,$T4,$T1 # h1*r0
905 vmovdqa -0x80(%r11),$T2 # r1^4
906 vpmuludq $H2,$T4,$T0 # h2*r0
907 vpmuludq $H3,$T4,$T1 # h3*r0
910 vpmuludq $H4,$T4,$T4 # h4*r0
911 vpmuludq -0x70(%r11),$H4,$T0 # h4*s1
914 vpaddq $T0,$D0,$D0 # d0 += h4*s1
915 vpmuludq $H2,$T2,$T1 # h2*r1
916 vpmuludq $H3,$T2,$T0 # h3*r1
917 vpaddq $T1,$D3,$D3 # d3 += h2*r1
918 vmovdqa -0x60(%r11),$T3 # r2^4
919 vpaddq $T0,$D4,$D4 # d4 += h3*r1
920 vpmuludq $H1,$T2,$T1 # h1*r1
921 vpmuludq $H0,$T2,$T2 # h0*r1
922 vpaddq $T1,$D2,$D2 # d2 += h1*r1
923 vpaddq $T2,$D1,$D1 # d1 += h0*r1
925 vmovdqa -0x50(%r11),$T4 # s2^4
926 vpmuludq $H2,$T3,$T0 # h2*r2
927 vpmuludq $H1,$T3,$T1 # h1*r2
928 vpaddq $T0,$D4,$D4 # d4 += h2*r2
929 vpaddq $T1,$D3,$D3 # d3 += h1*r2
930 vmovdqa -0x40(%r11),$T2 # r3^4
931 vpmuludq $H0,$T3,$T3 # h0*r2
932 vpmuludq $H4,$T4,$T0 # h4*s2
933 vpaddq $T3,$D2,$D2 # d2 += h0*r2
934 vpaddq $T0,$D1,$D1 # d1 += h4*s2
935 vmovdqa -0x30(%r11),$T3 # s3^4
936 vpmuludq $H3,$T4,$T4 # h3*s2
937 vpmuludq $H1,$T2,$T1 # h1*r3
938 vpaddq $T4,$D0,$D0 # d0 += h3*s2
940 vmovdqa -0x10(%r11),$T4 # s4^4
941 vpaddq $T1,$D4,$D4 # d4 += h1*r3
942 vpmuludq $H0,$T2,$T2 # h0*r3
943 vpmuludq $H4,$T3,$T0 # h4*s3
944 vpaddq $T2,$D3,$D3 # d3 += h0*r3
945 vpaddq $T0,$D2,$D2 # d2 += h4*s3
946 vmovdqu 16*2($inp),$T0 # load input
947 vpmuludq $H3,$T3,$T2 # h3*s3
948 vpmuludq $H2,$T3,$T3 # h2*s3
949 vpaddq $T2,$D1,$D1 # d1 += h3*s3
950 vmovdqu 16*3($inp),$T1 #
951 vpaddq $T3,$D0,$D0 # d0 += h2*s3
953 vpmuludq $H2,$T4,$H2 # h2*s4
954 vpmuludq $H3,$T4,$H3 # h3*s4
955 vpsrldq \$6,$T0,$T2 # splat input
956 vpaddq $H2,$D1,$D1 # d1 += h2*s4
957 vpmuludq $H4,$T4,$H4 # h4*s4
958 vpsrldq \$6,$T1,$T3 #
959 vpaddq $H3,$D2,$H2 # h2 = d2 + h3*s4
960 vpaddq $H4,$D3,$H3 # h3 = d3 + h4*s4
961 vpmuludq -0x20(%r11),$H0,$H4 # h0*r4
963 vpunpckhqdq $T1,$T0,$T4 # 4
964 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
965 vpaddq $H0,$D0,$H0 # h0 = d0 + h1*s4
967 vpunpcklqdq $T1,$T0,$T0 # 0:1
968 vpunpcklqdq $T3,$T2,$T3 # 2:3
970 #vpsrlq \$40,$T4,$T4 # 4
971 vpsrldq \$`40/8`,$T4,$T4 # 4
973 vmovdqa 0x00(%rsp),$D4 # preload r0^2
974 vpand $MASK,$T0,$T0 # 0
976 vpand $MASK,$T1,$T1 # 1
977 vpand 0(%rcx),$T4,$T4 # .Lmask24
979 vpand $MASK,$T2,$T2 # 2
980 vpand $MASK,$T3,$T3 # 3
981 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
983 ################################################################
984 # lazy reduction as discussed in "NEON crypto" by D.J. Bernstein
989 vpaddq $D3,$H4,$H4 # h3 -> h4
993 vpaddq $D0,$D1,$H1 # h0 -> h1
1000 vpaddq $D1,$H2,$H2 # h1 -> h2
1004 vpaddq $D0,$H0,$H0 # h4 -> h0
1008 vpaddq $D2,$H3,$H3 # h2 -> h3
1012 vpaddq $D0,$H1,$H1 # h0 -> h1
1016 vpaddq $D3,$H4,$H4 # h3 -> h4
1021 ################################################################
1022 # multiply (inp[0:1]+hash) or inp[2:3] by r^2:r^1
1024 vpshufd \$0x10,$D4,$D4 # r0^n, xx12 -> x1x2
1035 vmovdqa $H2,0x20(%r11)
1036 vmovdqa $H0,0x00(%r11)
1037 vmovdqa $H1,0x10(%r11)
1038 vmovdqa $H3,0x30(%r11)
1039 vmovdqa $H4,0x40(%r11)
1041 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
1042 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
1043 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1044 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
1045 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1047 vpmuludq $T2,$D4,$D2 # d2 = h2*r0
1048 vpmuludq $T0,$D4,$D0 # d0 = h0*r0
1049 vpshufd \$0x10,`16*1-64`($ctx),$H2 # r1^n
1050 vpmuludq $T1,$D4,$D1 # d1 = h1*r0
1051 vpmuludq $T3,$D4,$D3 # d3 = h3*r0
1052 vpmuludq $T4,$D4,$D4 # d4 = h4*r0
1054 vpmuludq $T3,$H2,$H0 # h3*r1
1055 vpaddq $H0,$D4,$D4 # d4 += h3*r1
1056 vpshufd \$0x10,`16*2-64`($ctx),$H3 # s1^n
1057 vpmuludq $T2,$H2,$H1 # h2*r1
1058 vpaddq $H1,$D3,$D3 # d3 += h2*r1
1059 vpshufd \$0x10,`16*3-64`($ctx),$H4 # r2^n
1060 vpmuludq $T1,$H2,$H0 # h1*r1
1061 vpaddq $H0,$D2,$D2 # d2 += h1*r1
1062 vpmuludq $T0,$H2,$H2 # h0*r1
1063 vpaddq $H2,$D1,$D1 # d1 += h0*r1
1064 vpmuludq $T4,$H3,$H3 # h4*s1
1065 vpaddq $H3,$D0,$D0 # d0 += h4*s1
1067 vpshufd \$0x10,`16*4-64`($ctx),$H2 # s2^n
1068 vpmuludq $T2,$H4,$H1 # h2*r2
1069 vpaddq $H1,$D4,$D4 # d4 += h2*r2
1070 vpmuludq $T1,$H4,$H0 # h1*r2
1071 vpaddq $H0,$D3,$D3 # d3 += h1*r2
1072 vpshufd \$0x10,`16*5-64`($ctx),$H3 # r3^n
1073 vpmuludq $T0,$H4,$H4 # h0*r2
1074 vpaddq $H4,$D2,$D2 # d2 += h0*r2
1075 vpmuludq $T4,$H2,$H1 # h4*s2
1076 vpaddq $H1,$D1,$D1 # d1 += h4*s2
1077 vpshufd \$0x10,`16*6-64`($ctx),$H4 # s3^n
1078 vpmuludq $T3,$H2,$H2 # h3*s2
1079 vpaddq $H2,$D0,$D0 # d0 += h3*s2
1081 vpmuludq $T1,$H3,$H0 # h1*r3
1082 vpaddq $H0,$D4,$D4 # d4 += h1*r3
1083 vpmuludq $T0,$H3,$H3 # h0*r3
1084 vpaddq $H3,$D3,$D3 # d3 += h0*r3
1085 vpshufd \$0x10,`16*7-64`($ctx),$H2 # r4^n
1086 vpmuludq $T4,$H4,$H1 # h4*s3
1087 vpaddq $H1,$D2,$D2 # d2 += h4*s3
1088 vpshufd \$0x10,`16*8-64`($ctx),$H3 # s4^n
1089 vpmuludq $T3,$H4,$H0 # h3*s3
1090 vpaddq $H0,$D1,$D1 # d1 += h3*s3
1091 vpmuludq $T2,$H4,$H4 # h2*s3
1092 vpaddq $H4,$D0,$D0 # d0 += h2*s3
1094 vpmuludq $T0,$H2,$H2 # h0*r4
1095 vpaddq $H2,$D4,$D4 # h4 = d4 + h0*r4
1096 vpmuludq $T4,$H3,$H1 # h4*s4
1097 vpaddq $H1,$D3,$D3 # h3 = d3 + h4*s4
1098 vpmuludq $T3,$H3,$H0 # h3*s4
1099 vpaddq $H0,$D2,$D2 # h2 = d2 + h3*s4
1100 vpmuludq $T2,$H3,$H1 # h2*s4
1101 vpaddq $H1,$D1,$D1 # h1 = d1 + h2*s4
1102 vpmuludq $T1,$H3,$H3 # h1*s4
1103 vpaddq $H3,$D0,$D0 # h0 = d0 + h1*s4
1107 vmovdqu 16*0($inp),$H0 # load input
1108 vmovdqu 16*1($inp),$H1
1110 vpsrldq \$6,$H0,$H2 # splat input
1112 vpunpckhqdq $H1,$H0,$H4 # 4
1113 vpunpcklqdq $H1,$H0,$H0 # 0:1
1114 vpunpcklqdq $H3,$H2,$H3 # 2:3
1116 vpsrlq \$40,$H4,$H4 # 4
1118 vpand $MASK,$H0,$H0 # 0
1120 vpand $MASK,$H1,$H1 # 1
1122 vpand $MASK,$H2,$H2 # 2
1123 vpand $MASK,$H3,$H3 # 3
1124 vpor 32(%rcx),$H4,$H4 # padbit, yes, always
1126 vpshufd \$0x32,`16*0-64`($ctx),$T4 # r0^n, 34xx -> x3x4
1127 vpaddq 0x00(%r11),$H0,$H0
1128 vpaddq 0x10(%r11),$H1,$H1
1129 vpaddq 0x20(%r11),$H2,$H2
1130 vpaddq 0x30(%r11),$H3,$H3
1131 vpaddq 0x40(%r11),$H4,$H4
1133 ################################################################
1134 # multiply (inp[0:1]+hash) by r^4:r^3 and accumulate
1136 vpmuludq $H0,$T4,$T0 # h0*r0
1137 vpaddq $T0,$D0,$D0 # d0 += h0*r0
1138 vpmuludq $H1,$T4,$T1 # h1*r0
1139 vpaddq $T1,$D1,$D1 # d1 += h1*r0
1140 vpmuludq $H2,$T4,$T0 # h2*r0
1141 vpaddq $T0,$D2,$D2 # d2 += h2*r0
1142 vpshufd \$0x32,`16*1-64`($ctx),$T2 # r1^n
1143 vpmuludq $H3,$T4,$T1 # h3*r0
1144 vpaddq $T1,$D3,$D3 # d3 += h3*r0
1145 vpmuludq $H4,$T4,$T4 # h4*r0
1146 vpaddq $T4,$D4,$D4 # d4 += h4*r0
1148 vpmuludq $H3,$T2,$T0 # h3*r1
1149 vpaddq $T0,$D4,$D4 # d4 += h3*r1
1150 vpshufd \$0x32,`16*2-64`($ctx),$T3 # s1
1151 vpmuludq $H2,$T2,$T1 # h2*r1
1152 vpaddq $T1,$D3,$D3 # d3 += h2*r1
1153 vpshufd \$0x32,`16*3-64`($ctx),$T4 # r2
1154 vpmuludq $H1,$T2,$T0 # h1*r1
1155 vpaddq $T0,$D2,$D2 # d2 += h1*r1
1156 vpmuludq $H0,$T2,$T2 # h0*r1
1157 vpaddq $T2,$D1,$D1 # d1 += h0*r1
1158 vpmuludq $H4,$T3,$T3 # h4*s1
1159 vpaddq $T3,$D0,$D0 # d0 += h4*s1
1161 vpshufd \$0x32,`16*4-64`($ctx),$T2 # s2
1162 vpmuludq $H2,$T4,$T1 # h2*r2
1163 vpaddq $T1,$D4,$D4 # d4 += h2*r2
1164 vpmuludq $H1,$T4,$T0 # h1*r2
1165 vpaddq $T0,$D3,$D3 # d3 += h1*r2
1166 vpshufd \$0x32,`16*5-64`($ctx),$T3 # r3
1167 vpmuludq $H0,$T4,$T4 # h0*r2
1168 vpaddq $T4,$D2,$D2 # d2 += h0*r2
1169 vpmuludq $H4,$T2,$T1 # h4*s2
1170 vpaddq $T1,$D1,$D1 # d1 += h4*s2
1171 vpshufd \$0x32,`16*6-64`($ctx),$T4 # s3
1172 vpmuludq $H3,$T2,$T2 # h3*s2
1173 vpaddq $T2,$D0,$D0 # d0 += h3*s2
1175 vpmuludq $H1,$T3,$T0 # h1*r3
1176 vpaddq $T0,$D4,$D4 # d4 += h1*r3
1177 vpmuludq $H0,$T3,$T3 # h0*r3
1178 vpaddq $T3,$D3,$D3 # d3 += h0*r3
1179 vpshufd \$0x32,`16*7-64`($ctx),$T2 # r4
1180 vpmuludq $H4,$T4,$T1 # h4*s3
1181 vpaddq $T1,$D2,$D2 # d2 += h4*s3
1182 vpshufd \$0x32,`16*8-64`($ctx),$T3 # s4
1183 vpmuludq $H3,$T4,$T0 # h3*s3
1184 vpaddq $T0,$D1,$D1 # d1 += h3*s3
1185 vpmuludq $H2,$T4,$T4 # h2*s3
1186 vpaddq $T4,$D0,$D0 # d0 += h2*s3
1188 vpmuludq $H0,$T2,$T2 # h0*r4
1189 vpaddq $T2,$D4,$D4 # d4 += h0*r4
1190 vpmuludq $H4,$T3,$T1 # h4*s4
1191 vpaddq $T1,$D3,$D3 # d3 += h4*s4
1192 vpmuludq $H3,$T3,$T0 # h3*s4
1193 vpaddq $T0,$D2,$D2 # d2 += h3*s4
1194 vpmuludq $H2,$T3,$T1 # h2*s4
1195 vpaddq $T1,$D1,$D1 # d1 += h2*s4
1196 vpmuludq $H1,$T3,$T3 # h1*s4
1197 vpaddq $T3,$D0,$D0 # d0 += h1*s4
1200 ################################################################
1205 vpaddq $H3,$D4,$D4 # h3 -> h4
1209 vpaddq $H0,$D1,$D1 # h0 -> h1
1216 vpaddq $H1,$D2,$D2 # h1 -> h2
1220 vpaddq $H4,$D0,$D0 # h4 -> h0
1224 vpaddq $H2,$D3,$D3 # h2 -> h3
1228 vpaddq $H0,$D1,$D1 # h0 -> h1
1232 vpaddq $H3,$D4,$D4 # h3 -> h4
1234 ################################################################
1235 # horizontal addition
1248 vmovd $H0,`4*0-48-64`($ctx) # save partially reduced
1249 vmovd $H1,`4*1-48-64`($ctx)
1250 vmovd $H2,`4*2-48-64`($ctx)
1251 vmovd $H3,`4*3-48-64`($ctx)
1252 vmovd $H4,`4*4-48-64`($ctx)
1254 $code.=<<___ if ($win64);
1255 vmovdqa 0x50(%r11),%xmm6
1256 vmovdqa 0x60(%r11),%xmm7
1257 vmovdqa 0x70(%r11),%xmm8
1258 vmovdqa 0x80(%r11),%xmm9
1259 vmovdqa 0x90(%r11),%xmm10
1260 vmovdqa 0xa0(%r11),%xmm11
1261 vmovdqa 0xb0(%r11),%xmm12
1262 vmovdqa 0xc0(%r11),%xmm13
1263 vmovdqa 0xd0(%r11),%xmm14
1264 vmovdqa 0xe0(%r11),%xmm15
1268 $code.=<<___ if (!$win64);
1274 .size poly1305_blocks_avx,.-poly1305_blocks_avx
1276 .type poly1305_emit_avx,\@function,3
1279 cmpl \$0,20($ctx) # is_base2_26?
1282 mov 0($ctx),%eax # load hash value base 2^26
1288 shl \$26,%rcx # base 2^26 -> base 2^64
1304 mov %r10,%rax # could be partially reduced, so reduce
1314 add \$5,%r8 # compare to modulus
1318 shr \$2,%r10 # did 130-bit value overfow?
1322 add 0($nonce),%rax # accumulate nonce
1324 mov %rax,0($mac) # write result
1328 .size poly1305_emit_avx,.-poly1305_emit_avx
1332 my ($H0,$H1,$H2,$H3,$H4, $MASK, $T4,$T0,$T1,$T2,$T3, $D0,$D1,$D2,$D3,$D4) =
1333 map("%ymm$_",(0..15));
1337 .type poly1305_blocks_avx2,\@function,4
1339 poly1305_blocks_avx2:
1340 mov 20($ctx),%r8d # is_base2_26
1366 mov $len,%r15 # reassign $len
1368 mov 0($ctx),$d1 # load hash value
1372 mov 24($ctx),$r0 # load r
1375 ################################# base 2^26 -> base 2^64
1378 mov $d2,$r1 # borrow $r1
1394 adc \$0,$h2 # can be partially reduced...
1396 mov \$-4,$d2 # ... so reduce
1408 add $r1,$s1 # s1 = r1 + (r1 >> 2)
1410 .Lbase2_26_pre_avx2:
1411 add 0($inp),$h0 # accumulate input
1417 call __poly1305_block
1421 jnz .Lbase2_26_pre_avx2
1423 test $padbit,$padbit # if $padbit is zero,
1424 jz .Lstore_base2_64_avx2 # store hash in base 2^64 format
1426 ################################# base 2^64 -> base 2^26
1433 and \$0x3ffffff,%rax # h[0]
1435 and \$0x3ffffff,%rdx # h[1]
1439 and \$0x3ffffff,$h0 # h[2]
1441 and \$0x3ffffff,$h1 # h[3]
1445 jz .Lstore_base2_26_avx2
1455 .Lstore_base2_64_avx2:
1458 mov $h2,16($ctx) # note that is_base2_26 is zeroed
1462 .Lstore_base2_26_avx2:
1463 mov %rax#d,0($ctx) # store hash value base 2^26
1478 .Lblocks_avx2_epilogue:
1489 .Lbase2_64_avx2_body:
1491 mov $len,%r15 # reassign $len
1493 mov 24($ctx),$r0 # load r
1496 mov 0($ctx),$h0 # load hash value
1503 add $r1,$s1 # s1 = r1 + (r1 >> 2)
1508 .Lbase2_64_pre_avx2:
1509 add 0($inp),$h0 # accumulate input
1515 call __poly1305_block
1519 jnz .Lbase2_64_pre_avx2
1522 ################################# base 2^64 -> base 2^26
1529 and \$0x3ffffff,%rax # h[0]
1531 and \$0x3ffffff,%rdx # h[1]
1535 and \$0x3ffffff,$h0 # h[2]
1537 and \$0x3ffffff,$h1 # h[3]
1545 movl \$1,20($ctx) # set is_base2_26
1547 call __poly1305_init_avx
1560 .Lbase2_64_avx2_epilogue:
1565 vmovd 4*0($ctx),%x#$H0 # load hash value base 2^26
1566 vmovd 4*1($ctx),%x#$H1
1567 vmovd 4*2($ctx),%x#$H2
1568 vmovd 4*3($ctx),%x#$H3
1569 vmovd 4*4($ctx),%x#$H4
1573 $code.=<<___ if (!$win64);
1577 $code.=<<___ if ($win64);
1578 lea -0xf8(%rsp),%r11
1580 vmovdqa %xmm6,0x50(%r11)
1581 vmovdqa %xmm7,0x60(%r11)
1582 vmovdqa %xmm8,0x70(%r11)
1583 vmovdqa %xmm9,0x80(%r11)
1584 vmovdqa %xmm10,0x90(%r11)
1585 vmovdqa %xmm11,0xa0(%r11)
1586 vmovdqa %xmm12,0xb0(%r11)
1587 vmovdqa %xmm13,0xc0(%r11)
1588 vmovdqa %xmm14,0xd0(%r11)
1589 vmovdqa %xmm15,0xe0(%r11)
1593 lea 48+64($ctx),$ctx # size optimization
1594 lea .Lconst(%rip),%rcx
1596 # expand and copy pre-calculated table to stack
1597 vmovdqu `16*0-64`($ctx),%x#$T2
1599 vmovdqu `16*1-64`($ctx),%x#$T3
1600 vmovdqu `16*2-64`($ctx),%x#$T4
1601 vmovdqu `16*3-64`($ctx),%x#$D0
1602 vmovdqu `16*4-64`($ctx),%x#$D1
1603 vmovdqu `16*5-64`($ctx),%x#$D2
1604 vmovdqu `16*6-64`($ctx),%x#$D3
1605 vpermq \$0x15,$T2,$T2 # 00003412 -> 12343434
1606 vmovdqu `16*7-64`($ctx),%x#$D4
1607 vpermq \$0x15,$T3,$T3
1608 vpshufd \$0xc8,$T2,$T2 # 12343434 -> 14243444
1609 vmovdqu `16*8-64`($ctx),%x#$MASK
1610 vpermq \$0x15,$T4,$T4
1611 vpshufd \$0xc8,$T3,$T3
1612 vmovdqa $T2,0x00(%rsp)
1613 vpermq \$0x15,$D0,$D0
1614 vpshufd \$0xc8,$T4,$T4
1615 vmovdqa $T3,0x20(%rsp)
1616 vpermq \$0x15,$D1,$D1
1617 vpshufd \$0xc8,$D0,$D0
1618 vmovdqa $T4,0x40(%rsp)
1619 vpermq \$0x15,$D2,$D2
1620 vpshufd \$0xc8,$D1,$D1
1621 vmovdqa $D0,0x60(%rsp)
1622 vpermq \$0x15,$D3,$D3
1623 vpshufd \$0xc8,$D2,$D2
1624 vmovdqa $D1,0x80(%rsp)
1625 vpermq \$0x15,$D4,$D4
1626 vpshufd \$0xc8,$D3,$D3
1627 vmovdqa $D2,0xa0(%rsp)
1628 vpermq \$0x15,$MASK,$MASK
1629 vpshufd \$0xc8,$D4,$D4
1630 vmovdqa $D3,0xc0(%rsp)
1631 vpshufd \$0xc8,$MASK,$MASK
1632 vmovdqa $D4,0xe0(%rsp)
1633 vmovdqa $MASK,0x100(%rsp)
1634 vmovdqa 64(%rcx),$MASK # .Lmask26
1636 ################################################################
1638 vmovdqu 16*0($inp),%x#$T0
1639 vmovdqu 16*1($inp),%x#$T1
1640 vinserti128 \$1,16*2($inp),$T0,$T0
1641 vinserti128 \$1,16*3($inp),$T1,$T1
1644 vpsrldq \$6,$T0,$T2 # splat input
1646 vpunpckhqdq $T1,$T0,$T4 # 4
1647 vpunpcklqdq $T3,$T2,$T2 # 2:3
1648 vpunpcklqdq $T1,$T0,$T0 # 0:1
1653 vpsrlq \$40,$T4,$T4 # 4
1654 vpand $MASK,$T2,$T2 # 2
1655 vpand $MASK,$T0,$T0 # 0
1656 vpand $MASK,$T1,$T1 # 1
1657 vpand $MASK,$T3,$T3 # 3
1658 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
1660 lea 0x90(%rsp),%rax # size optimization
1661 vpaddq $H2,$T2,$H2 # accumulate input
1668 ################################################################
1669 # ((inp[0]*r^4+r[4])*r^4+r[8])*r^4
1670 # ((inp[1]*r^4+r[5])*r^4+r[9])*r^3
1671 # ((inp[2]*r^4+r[6])*r^4+r[10])*r^2
1672 # ((inp[3]*r^4+r[7])*r^4+r[11])*r^1
1673 # \________/\________/
1674 ################################################################
1675 #vpaddq $H2,$T2,$H2 # accumulate input
1677 vmovdqa `32*0`(%rsp),$T0 # r0^4
1679 vmovdqa `32*1`(%rsp),$T1 # r1^4
1681 vmovdqa `32*3`(%rsp),$T2 # r2^4
1683 vmovdqa `32*6-0x90`(%rax),$T3 # s3^4
1684 vmovdqa `32*8-0x90`(%rax),$S4 # s4^4
1686 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
1687 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
1688 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1689 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
1690 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1692 # however, as h2 is "chronologically" first one available pull
1693 # corresponding operations up, so it's
1695 # d4 = h2*r2 + h4*r0 + h3*r1 + h1*r3 + h0*r4
1696 # d3 = h2*r1 + h3*r0 + h1*r2 + h0*r3 + h4*5*r4
1697 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1698 # d1 = h2*5*r4 + h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3
1699 # d0 = h2*5*r3 + h0*r0 + h4*5*r1 + h3*5*r2 + h1*5*r4
1701 vpmuludq $H2,$T0,$D2 # d2 = h2*r0
1702 vpmuludq $H2,$T1,$D3 # d3 = h2*r1
1703 vpmuludq $H2,$T2,$D4 # d4 = h2*r2
1704 vpmuludq $H2,$T3,$D0 # d0 = h2*s3
1705 vpmuludq $H2,$S4,$D1 # d1 = h2*s4
1707 vpmuludq $H0,$T1,$T4 # h0*r1
1708 vpmuludq $H1,$T1,$H2 # h1*r1, borrow $H2 as temp
1709 vpaddq $T4,$D1,$D1 # d1 += h0*r1
1710 vpaddq $H2,$D2,$D2 # d2 += h1*r1
1711 vpmuludq $H3,$T1,$T4 # h3*r1
1712 vpmuludq `32*2`(%rsp),$H4,$H2 # h4*s1
1713 vpaddq $T4,$D4,$D4 # d4 += h3*r1
1714 vpaddq $H2,$D0,$D0 # d0 += h4*s1
1715 vmovdqa `32*4-0x90`(%rax),$T1 # s2
1717 vpmuludq $H0,$T0,$T4 # h0*r0
1718 vpmuludq $H1,$T0,$H2 # h1*r0
1719 vpaddq $T4,$D0,$D0 # d0 += h0*r0
1720 vpaddq $H2,$D1,$D1 # d1 += h1*r0
1721 vpmuludq $H3,$T0,$T4 # h3*r0
1722 vpmuludq $H4,$T0,$H2 # h4*r0
1723 vmovdqu 16*0($inp),%x#$T0 # load input
1724 vpaddq $T4,$D3,$D3 # d3 += h3*r0
1725 vpaddq $H2,$D4,$D4 # d4 += h4*r0
1726 vinserti128 \$1,16*2($inp),$T0,$T0
1728 vpmuludq $H3,$T1,$T4 # h3*s2
1729 vpmuludq $H4,$T1,$H2 # h4*s2
1730 vmovdqu 16*1($inp),%x#$T1
1731 vpaddq $T4,$D0,$D0 # d0 += h3*s2
1732 vpaddq $H2,$D1,$D1 # d1 += h4*s2
1733 vmovdqa `32*5-0x90`(%rax),$H2 # r3
1734 vpmuludq $H1,$T2,$T4 # h1*r2
1735 vpmuludq $H0,$T2,$T2 # h0*r2
1736 vpaddq $T4,$D3,$D3 # d3 += h1*r2
1737 vpaddq $T2,$D2,$D2 # d2 += h0*r2
1738 vinserti128 \$1,16*3($inp),$T1,$T1
1741 vpmuludq $H1,$H2,$T4 # h1*r3
1742 vpmuludq $H0,$H2,$H2 # h0*r3
1743 vpsrldq \$6,$T0,$T2 # splat input
1744 vpaddq $T4,$D4,$D4 # d4 += h1*r3
1745 vpaddq $H2,$D3,$D3 # d3 += h0*r3
1746 vpmuludq $H3,$T3,$T4 # h3*s3
1747 vpmuludq $H4,$T3,$H2 # h4*s3
1749 vpaddq $T4,$D1,$D1 # d1 += h3*s3
1750 vpaddq $H2,$D2,$D2 # d2 += h4*s3
1751 vpunpckhqdq $T1,$T0,$T4 # 4
1753 vpmuludq $H3,$S4,$H3 # h3*s4
1754 vpmuludq $H4,$S4,$H4 # h4*s4
1755 vpunpcklqdq $T1,$T0,$T0 # 0:1
1756 vpaddq $H3,$D2,$H2 # h2 = d2 + h3*r4
1757 vpaddq $H4,$D3,$H3 # h3 = d3 + h4*r4
1758 vpunpcklqdq $T3,$T2,$T3 # 2:3
1759 vpmuludq `32*7-0x90`(%rax),$H0,$H4 # h0*r4
1760 vpmuludq $H1,$S4,$H0 # h1*s4
1761 vmovdqa 64(%rcx),$MASK # .Lmask26
1762 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
1763 vpaddq $H0,$D0,$H0 # h0 = d0 + h1*s4
1765 ################################################################
1766 # lazy reduction (interleaved with tail of input splat)
1770 vpaddq $D3,$H4,$H4 # h3 -> h4
1774 vpaddq $D0,$D1,$H1 # h0 -> h1
1783 vpaddq $D1,$H2,$H2 # h1 -> h2
1787 vpaddq $D4,$H0,$H0 # h4 -> h0
1789 vpand $MASK,$T2,$T2 # 2
1794 vpaddq $D2,$H3,$H3 # h2 -> h3
1796 vpaddq $T2,$H2,$H2 # modulo-scheduled
1801 vpaddq $D0,$H1,$H1 # h0 -> h1
1803 vpsrlq \$40,$T4,$T4 # 4
1807 vpaddq $D3,$H4,$H4 # h3 -> h4
1809 vpand $MASK,$T0,$T0 # 0
1810 vpand $MASK,$T1,$T1 # 1
1811 vpand $MASK,$T3,$T3 # 3
1812 vpor 32(%rcx),$T4,$T4 # padbit, yes, always
1819 ################################################################
1820 # while above multiplications were by r^4 in all lanes, in last
1821 # iteration we multiply least significant lane by r^4 and most
1822 # significant one by r, so copy of above except that references
1823 # to the precomputed table are displaced by 4...
1825 #vpaddq $H2,$T2,$H2 # accumulate input
1827 vmovdqu `32*0+4`(%rsp),$T0 # r0^4
1829 vmovdqu `32*1+4`(%rsp),$T1 # r1^4
1831 vmovdqu `32*3+4`(%rsp),$T2 # r2^4
1833 vmovdqu `32*6+4-0x90`(%rax),$T3 # s3^4
1834 vmovdqu `32*8+4-0x90`(%rax),$S4 # s4^4
1836 vpmuludq $H2,$T0,$D2 # d2 = h2*r0
1837 vpmuludq $H2,$T1,$D3 # d3 = h2*r1
1838 vpmuludq $H2,$T2,$D4 # d4 = h2*r2
1839 vpmuludq $H2,$T3,$D0 # d0 = h2*s3
1840 vpmuludq $H2,$S4,$D1 # d1 = h2*s4
1842 vpmuludq $H0,$T1,$T4 # h0*r1
1843 vpmuludq $H1,$T1,$H2 # h1*r1
1844 vpaddq $T4,$D1,$D1 # d1 += h0*r1
1845 vpaddq $H2,$D2,$D2 # d2 += h1*r1
1846 vpmuludq $H3,$T1,$T4 # h3*r1
1847 vpmuludq `32*2+4`(%rsp),$H4,$H2 # h4*s1
1848 vpaddq $T4,$D4,$D4 # d4 += h3*r1
1849 vpaddq $H2,$D0,$D0 # d0 += h4*s1
1851 vpmuludq $H0,$T0,$T4 # h0*r0
1852 vpmuludq $H1,$T0,$H2 # h1*r0
1853 vpaddq $T4,$D0,$D0 # d0 += h0*r0
1854 vmovdqu `32*4+4-0x90`(%rax),$T1 # s2
1855 vpaddq $H2,$D1,$D1 # d1 += h1*r0
1856 vpmuludq $H3,$T0,$T4 # h3*r0
1857 vpmuludq $H4,$T0,$H2 # h4*r0
1858 vpaddq $T4,$D3,$D3 # d3 += h3*r0
1859 vpaddq $H2,$D4,$D4 # d4 += h4*r0
1861 vpmuludq $H3,$T1,$T4 # h3*s2
1862 vpmuludq $H4,$T1,$H2 # h4*s2
1863 vpaddq $T4,$D0,$D0 # d0 += h3*s2
1864 vpaddq $H2,$D1,$D1 # d1 += h4*s2
1865 vmovdqu `32*5+4-0x90`(%rax),$H2 # r3
1866 vpmuludq $H1,$T2,$T4 # h1*r2
1867 vpmuludq $H0,$T2,$T2 # h0*r2
1868 vpaddq $T4,$D3,$D3 # d3 += h1*r2
1869 vpaddq $T2,$D2,$D2 # d2 += h0*r2
1871 vpmuludq $H1,$H2,$T4 # h1*r3
1872 vpmuludq $H0,$H2,$H2 # h0*r3
1873 vpaddq $T4,$D4,$D4 # d4 += h1*r3
1874 vpaddq $H2,$D3,$D3 # d3 += h0*r3
1875 vpmuludq $H3,$T3,$T4 # h3*s3
1876 vpmuludq $H4,$T3,$H2 # h4*s3
1877 vpaddq $T4,$D1,$D1 # d1 += h3*s3
1878 vpaddq $H2,$D2,$D2 # d2 += h4*s3
1880 vpmuludq $H3,$S4,$H3 # h3*s4
1881 vpmuludq $H4,$S4,$H4 # h4*s4
1882 vpaddq $H3,$D2,$H2 # h2 = d2 + h3*r4
1883 vpaddq $H4,$D3,$H3 # h3 = d3 + h4*r4
1884 vpmuludq `32*7+4-0x90`(%rax),$H0,$H4 # h0*r4
1885 vpmuludq $H1,$S4,$H0 # h1*s4
1886 vmovdqa 64(%rcx),$MASK # .Lmask26
1887 vpaddq $H4,$D4,$H4 # h4 = d4 + h0*r4
1888 vpaddq $H0,$D0,$H0 # h0 = d0 + h1*s4
1890 ################################################################
1895 vpaddq $D3,$H4,$H4 # h3 -> h4
1899 vpaddq $D0,$D1,$H1 # h0 -> h1
1906 vpaddq $D1,$H2,$H2 # h1 -> h2
1910 vpaddq $D4,$H0,$H0 # h4 -> h0
1914 vpaddq $D2,$H3,$H3 # h2 -> h3
1918 vpaddq $D0,$H1,$H1 # h0 -> h1
1922 vpaddq $D3,$H4,$H4 # h3 -> h4
1924 ################################################################
1925 # horizontal addition
1938 vpermq \$0x2,$H2,$T2
1939 vpermq \$0x2,$H0,$T0
1940 vpermq \$0x2,$H1,$T1
1941 vpermq \$0x2,$H3,$T3
1942 vpermq \$0x2,$H4,$T4
1949 vmovd %x#$H0,`4*0-48-64`($ctx)# save partially reduced
1950 vmovd %x#$H1,`4*1-48-64`($ctx)
1951 vmovd %x#$H2,`4*2-48-64`($ctx)
1952 vmovd %x#$H3,`4*3-48-64`($ctx)
1953 vmovd %x#$H4,`4*4-48-64`($ctx)
1955 $code.=<<___ if ($win64);
1956 vmovdqa 0x50(%r11),%xmm6
1957 vmovdqa 0x60(%r11),%xmm7
1958 vmovdqa 0x70(%r11),%xmm8
1959 vmovdqa 0x80(%r11),%xmm9
1960 vmovdqa 0x90(%r11),%xmm10
1961 vmovdqa 0xa0(%r11),%xmm11
1962 vmovdqa 0xb0(%r11),%xmm12
1963 vmovdqa 0xc0(%r11),%xmm13
1964 vmovdqa 0xd0(%r11),%xmm14
1965 vmovdqa 0xe0(%r11),%xmm15
1969 $code.=<<___ if (!$win64);
1975 .size poly1305_blocks_avx2,.-poly1305_blocks_avx2
1982 .long 0x0ffffff,0,0x0ffffff,0,0x0ffffff,0,0x0ffffff,0
1984 .long 1<<24,0,1<<24,0,1<<24,0,1<<24,0
1986 .long 0x3ffffff,0,0x3ffffff,0,0x3ffffff,0,0x3ffffff,0
1988 .long 5,0,5,0,5,0,5,0
1993 .asciz "Poly1305 for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1997 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1998 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
2006 .extern __imp_RtlVirtualUnwind
2007 .type se_handler,\@abi-omnipotent
2021 mov 120($context),%rax # pull context->Rax
2022 mov 248($context),%rbx # pull context->Rip
2024 mov 8($disp),%rsi # disp->ImageBase
2025 mov 56($disp),%r11 # disp->HandlerData
2027 mov 0(%r11),%r10d # HandlerData[0]
2028 lea (%rsi,%r10),%r10 # prologue label
2029 cmp %r10,%rbx # context->Rip<.Lprologue
2030 jb .Lcommon_seh_tail
2032 mov 152($context),%rax # pull context->Rsp
2034 mov 4(%r11),%r10d # HandlerData[1]
2035 lea (%rsi,%r10),%r10 # epilogue label
2036 cmp %r10,%rbx # context->Rip>=.Lepilogue
2037 jae .Lcommon_seh_tail
2047 mov %rbx,144($context) # restore context->Rbx
2048 mov %rbp,160($context) # restore context->Rbp
2049 mov %r12,216($context) # restore context->R12
2050 mov %r13,224($context) # restore context->R13
2051 mov %r14,232($context) # restore context->R14
2052 mov %r15,240($context) # restore context->R14
2054 jmp .Lcommon_seh_tail
2055 .size se_handler,.-se_handler
2057 .type avx_handler,\@abi-omnipotent
2071 mov 120($context),%rax # pull context->Rax
2072 mov 248($context),%rbx # pull context->Rip
2074 mov 8($disp),%rsi # disp->ImageBase
2075 mov 56($disp),%r11 # disp->HandlerData
2077 mov 0(%r11),%r10d # HandlerData[0]
2078 lea (%rsi,%r10),%r10 # prologue label
2079 cmp %r10,%rbx # context->Rip<prologue label
2080 jb .Lcommon_seh_tail
2082 mov 152($context),%rax # pull context->Rsp
2084 mov 4(%r11),%r10d # HandlerData[1]
2085 lea (%rsi,%r10),%r10 # epilogue label
2086 cmp %r10,%rbx # context->Rip>=epilogue label
2087 jae .Lcommon_seh_tail
2089 mov 208($context),%rax # pull context->R11
2093 lea 512($context),%rdi # &context.Xmm6
2095 .long 0xa548f3fc # cld; rep movsq
2100 mov %rax,152($context) # restore context->Rsp
2101 mov %rsi,168($context) # restore context->Rsi
2102 mov %rdi,176($context) # restore context->Rdi
2104 mov 40($disp),%rdi # disp->ContextRecord
2105 mov $context,%rsi # context
2106 mov \$154,%ecx # sizeof(CONTEXT)
2107 .long 0xa548f3fc # cld; rep movsq
2110 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
2111 mov 8(%rsi),%rdx # arg2, disp->ImageBase
2112 mov 0(%rsi),%r8 # arg3, disp->ControlPc
2113 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
2114 mov 40(%rsi),%r10 # disp->ContextRecord
2115 lea 56(%rsi),%r11 # &disp->HandlerData
2116 lea 24(%rsi),%r12 # &disp->EstablisherFrame
2117 mov %r10,32(%rsp) # arg5
2118 mov %r11,40(%rsp) # arg6
2119 mov %r12,48(%rsp) # arg7
2120 mov %rcx,56(%rsp) # arg8, (NULL)
2121 call *__imp_RtlVirtualUnwind(%rip)
2123 mov \$1,%eax # ExceptionContinueSearch
2135 .size avx_handler,.-avx_handler
2139 .rva .LSEH_begin_poly1305_init
2140 .rva .LSEH_end_poly1305_init
2141 .rva .LSEH_info_poly1305_init
2143 .rva .LSEH_begin_poly1305_blocks
2144 .rva .LSEH_end_poly1305_blocks
2145 .rva .LSEH_info_poly1305_blocks
2147 .rva .LSEH_begin_poly1305_emit
2148 .rva .LSEH_end_poly1305_emit
2149 .rva .LSEH_info_poly1305_emit
2151 $code.=<<___ if ($avx);
2152 .rva .LSEH_begin_poly1305_blocks_avx
2154 .rva .LSEH_info_poly1305_blocks_avx_1
2158 .rva .LSEH_info_poly1305_blocks_avx_2
2161 .rva .LSEH_end_poly1305_blocks_avx
2162 .rva .LSEH_info_poly1305_blocks_avx_3
2164 .rva .LSEH_begin_poly1305_emit_avx
2165 .rva .LSEH_end_poly1305_emit_avx
2166 .rva .LSEH_info_poly1305_emit_avx
2168 $code.=<<___ if ($avx>1);
2169 .rva .LSEH_begin_poly1305_blocks_avx2
2170 .rva .Lbase2_64_avx2
2171 .rva .LSEH_info_poly1305_blocks_avx2_1
2173 .rva .Lbase2_64_avx2
2175 .rva .LSEH_info_poly1305_blocks_avx2_2
2178 .rva .LSEH_end_poly1305_blocks_avx2
2179 .rva .LSEH_info_poly1305_blocks_avx2_3
2184 .LSEH_info_poly1305_init:
2187 .rva .LSEH_begin_poly1305_init,.LSEH_begin_poly1305_init
2189 .LSEH_info_poly1305_blocks:
2192 .rva .Lblocks_body,.Lblocks_epilogue
2194 .LSEH_info_poly1305_emit:
2197 .rva .LSEH_begin_poly1305_emit,.LSEH_begin_poly1305_emit
2199 $code.=<<___ if ($avx);
2200 .LSEH_info_poly1305_blocks_avx_1:
2203 .rva .Lblocks_avx_body,.Lblocks_avx_epilogue # HandlerData[]
2205 .LSEH_info_poly1305_blocks_avx_2:
2208 .rva .Lbase2_64_avx_body,.Lbase2_64_avx_epilogue # HandlerData[]
2210 .LSEH_info_poly1305_blocks_avx_3:
2213 .rva .Ldo_avx_body,.Ldo_avx_epilogue # HandlerData[]
2215 .LSEH_info_poly1305_emit_avx:
2218 .rva .LSEH_begin_poly1305_emit_avx,.LSEH_begin_poly1305_emit_avx
2220 $code.=<<___ if ($avx>1);
2221 .LSEH_info_poly1305_blocks_avx2_1:
2224 .rva .Lblocks_avx2_body,.Lblocks_avx2_epilogue # HandlerData[]
2226 .LSEH_info_poly1305_blocks_avx2_2:
2229 .rva .Lbase2_64_avx2_body,.Lbase2_64_avx2_epilogue # HandlerData[]
2231 .LSEH_info_poly1305_blocks_avx2_3:
2234 .rva .Ldo_avx2_body,.Ldo_avx2_epilogue # HandlerData[]
2238 foreach (split('\n',$code)) {
2239 s/\`([^\`]*)\`/eval($1)/ge;
2240 s/%r([a-z]+)#d/%e$1/g;
2241 s/%r([0-9]+)#d/%r$1d/g;