2 # Copyright 2016-2018 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 # ====================================================================
17 # This module implements Poly1305 hash for x86.
21 # Numbers are cycles per processed byte with poly1305_blocks alone,
22 # measured with rdtsc at fixed clock frequency.
24 # IALU/gcc-3.4(*) SSE2(**) AVX2
28 # Core 2 4.85/+90% 1.80
29 # Westmere 4.58/+100% 1.43
30 # Sandy Bridge 3.90/+100% 1.36
31 # Haswell 3.88/+70% 1.18 0.72
32 # Skylake 3.10/+60% 1.14 0.62
33 # Silvermont 11.0/+40% 4.80
34 # Goldmont 4.10/+200% 2.10
35 # VIA Nano 6.71/+90% 2.47
36 # Sledgehammer 3.51/+180% 4.27
37 # Bulldozer 4.53/+140% 1.31
39 # (*) gcc 4.8 for some reason generated worse code;
40 # (**) besides SSE2 there are floating-point and AVX options; FP
41 # is deemed unnecessary, because pre-SSE2 processor are too
42 # old to care about, while it's not the fastest option on
43 # SSE2-capable ones; AVX is omitted, because it doesn't give
44 # a lot of improvement, 5-10% depending on processor;
46 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
47 push(@INC,"${dir}","${dir}../../perlasm");
50 $output=pop and open STDOUT,">$output";
52 &asm_init($ARGV[0],$ARGV[$#ARGV] eq "386");
55 for (@ARGV) { $sse2=1 if (/-DOPENSSL_IA32_SSE2/); }
58 &static_label("const_sse2");
59 &static_label("enter_blocks");
60 &static_label("enter_emit");
61 &external_label("OPENSSL_ia32cap_P");
63 if (`$ENV{CC} -Wa,-v -c -o /dev/null -x assembler /dev/null 2>&1`
64 =~ /GNU assembler version ([2-9]\.[0-9]+)/) {
65 $avx = ($1>=2.19) + ($1>=2.22);
68 if (!$avx && $ARGV[0] eq "win32n" &&
69 `nasm -v 2>&1` =~ /NASM version ([2-9]\.[0-9]+)/) {
70 $avx = ($1>=2.09) + ($1>=2.10);
73 if (!$avx && `$ENV{CC} -v 2>&1` =~ /((?:^clang|LLVM) version|based on LLVM) ([0-9]+\.[0-9]+)/) {
74 $avx = ($2>=3.0) + ($2>3.0);
78 ########################################################################
79 # Layout of opaque area is following.
81 # unsigned __int32 h[5]; # current hash value base 2^32
82 # unsigned __int32 pad; # is_base2_26 in vector context
83 # unsigned __int32 r[4]; # key value base 2^32
86 &function_begin("poly1305_init");
87 &mov ("edi",&wparam(0)); # context
88 &mov ("esi",&wparam(1)); # key
89 &mov ("ebp",&wparam(2)); # function table
92 &mov (&DWP(4*0,"edi"),"eax"); # zero hash value
93 &mov (&DWP(4*1,"edi"),"eax");
94 &mov (&DWP(4*2,"edi"),"eax");
95 &mov (&DWP(4*3,"edi"),"eax");
96 &mov (&DWP(4*4,"edi"),"eax");
97 &mov (&DWP(4*5,"edi"),"eax"); # is_base2_26
100 &je (&label("nokey"));
103 &call (&label("pic_point"));
104 &set_label("pic_point");
107 &lea ("eax",&DWP("poly1305_blocks-".&label("pic_point"),"ebx"));
108 &lea ("edx",&DWP("poly1305_emit-".&label("pic_point"),"ebx"));
110 &picmeup("edi","OPENSSL_ia32cap_P","ebx",&label("pic_point"));
111 &mov ("ecx",&DWP(0,"edi"));
112 &and ("ecx",1<<26|1<<24);
113 &cmp ("ecx",1<<26|1<<24); # SSE2 and XMM?
114 &jne (&label("no_sse2"));
116 &lea ("eax",&DWP("_poly1305_blocks_sse2-".&label("pic_point"),"ebx"));
117 &lea ("edx",&DWP("_poly1305_emit_sse2-".&label("pic_point"),"ebx"));
120 &mov ("ecx",&DWP(8,"edi"));
121 &test ("ecx",1<<5); # AVX2?
122 &jz (&label("no_sse2"));
124 &lea ("eax",&DWP("_poly1305_blocks_avx2-".&label("pic_point"),"ebx"));
126 &set_label("no_sse2");
127 &mov ("edi",&wparam(0)); # reload context
128 &mov (&DWP(0,"ebp"),"eax"); # fill function table
129 &mov (&DWP(4,"ebp"),"edx");
132 &mov ("eax",&DWP(4*0,"esi")); # load input key
133 &mov ("ebx",&DWP(4*1,"esi"));
134 &mov ("ecx",&DWP(4*2,"esi"));
135 &mov ("edx",&DWP(4*3,"esi"));
136 &and ("eax",0x0fffffff);
137 &and ("ebx",0x0ffffffc);
138 &and ("ecx",0x0ffffffc);
139 &and ("edx",0x0ffffffc);
140 &mov (&DWP(4*6,"edi"),"eax");
141 &mov (&DWP(4*7,"edi"),"ebx");
142 &mov (&DWP(4*8,"edi"),"ecx");
143 &mov (&DWP(4*9,"edi"),"edx");
147 &function_end("poly1305_init");
149 ($h0,$h1,$h2,$h3,$h4,
152 $s1,$s2,$s3)=map(4*$_,(0..15));
154 &function_begin("poly1305_blocks");
155 &mov ("edi",&wparam(0)); # ctx
156 &mov ("esi",&wparam(1)); # inp
157 &mov ("ecx",&wparam(2)); # len
158 &set_label("enter_blocks");
160 &jz (&label("nodata"));
163 &mov ("eax",&DWP(4*6,"edi")); # r0
164 &mov ("ebx",&DWP(4*7,"edi")); # r1
165 &lea ("ebp",&DWP(0,"esi","ecx")); # end of input
166 &mov ("ecx",&DWP(4*8,"edi")); # r2
167 &mov ("edx",&DWP(4*9,"edi")); # r3
169 &mov (&wparam(2),"ebp");
172 &mov (&DWP($r0,"esp"),"eax"); # r0
175 &mov (&DWP($r1,"esp"),"ebx"); # r1
176 &add ("eax","ebx"); # s1
179 &mov (&DWP($r2,"esp"),"ecx"); # r2
180 &add ("ebx","ecx"); # s2
183 &mov (&DWP($r3,"esp"),"edx"); # r3
184 &add ("ecx","edx"); # s3
185 &mov (&DWP($s1,"esp"),"eax"); # s1
186 &mov (&DWP($s2,"esp"),"ebx"); # s2
187 &mov (&DWP($s3,"esp"),"ecx"); # s3
189 &mov ("eax",&DWP(4*0,"edi")); # load hash value
190 &mov ("ebx",&DWP(4*1,"edi"));
191 &mov ("ecx",&DWP(4*2,"edi"));
192 &mov ("esi",&DWP(4*3,"edi"));
193 &mov ("edi",&DWP(4*4,"edi"));
194 &jmp (&label("loop"));
196 &set_label("loop",32);
197 &add ("eax",&DWP(4*0,"ebp")); # accumulate input
198 &adc ("ebx",&DWP(4*1,"ebp"));
199 &adc ("ecx",&DWP(4*2,"ebp"));
200 &adc ("esi",&DWP(4*3,"ebp"));
201 &lea ("ebp",&DWP(4*4,"ebp"));
202 &adc ("edi",&wparam(3)); # padbit
204 &mov (&DWP($h0,"esp"),"eax"); # put aside hash[+inp]
205 &mov (&DWP($h3,"esp"),"esi");
207 &mul (&DWP($r0,"esp")); # h0*r0
208 &mov (&DWP($h4,"esp"),"edi");
210 &mov ("eax","ebx"); # h1
212 &mul (&DWP($s3,"esp")); # h1*s3
214 &mov ("eax","ecx"); # h2
216 &mul (&DWP($s2,"esp")); # h2*s2
218 &mov ("eax",&DWP($h3,"esp"));
220 &mul (&DWP($s1,"esp")); # h3*s1
222 &mov ("eax",&DWP($h0,"esp"));
225 &mul (&DWP($r1,"esp")); # h0*r1
226 &mov (&DWP($d0,"esp"),"edi");
229 &mov ("eax","ebx"); # h1
231 &mul (&DWP($r0,"esp")); # h1*r0
233 &mov ("eax","ecx"); # h2
235 &mul (&DWP($s3,"esp")); # h2*s3
237 &mov ("eax",&DWP($h3,"esp"));
239 &mul (&DWP($s2,"esp")); # h3*s2
241 &mov ("eax",&DWP($h4,"esp"));
243 &imul ("eax",&DWP($s1,"esp")); # h4*s1
245 &mov ("eax",&DWP($h0,"esp"));
248 &mul (&DWP($r2,"esp")); # h0*r2
249 &mov (&DWP($d1,"esp"),"esi");
252 &mov ("eax","ebx"); # h1
254 &mul (&DWP($r1,"esp")); # h1*r1
256 &mov ("eax","ecx"); # h2
258 &mul (&DWP($r0,"esp")); # h2*r0
260 &mov ("eax",&DWP($h3,"esp"));
262 &mul (&DWP($s3,"esp")); # h3*s3
264 &mov ("eax",&DWP($h4,"esp"));
266 &imul ("eax",&DWP($s2,"esp")); # h4*s2
268 &mov ("eax",&DWP($h0,"esp"));
271 &mul (&DWP($r3,"esp")); # h0*r3
272 &mov (&DWP($d2,"esp"),"edi");
275 &mov ("eax","ebx"); # h1
277 &mul (&DWP($r2,"esp")); # h1*r2
279 &mov ("eax","ecx"); # h2
281 &mul (&DWP($r1,"esp")); # h2*r1
283 &mov ("eax",&DWP($h3,"esp"));
285 &mul (&DWP($r0,"esp")); # h3*r0
287 &mov ("ecx",&DWP($h4,"esp"));
291 &imul ("ecx",&DWP($s3,"esp")); # h4*s3
293 &mov ("eax",&DWP($d0,"esp"));
296 &imul ("edx",&DWP($r0,"esp")); # h4*r0
299 &mov ("ebx",&DWP($d1,"esp"));
300 &mov ("ecx",&DWP($d2,"esp"));
302 &mov ("edi","edx"); # last reduction step
305 &lea ("edx",&DWP(0,"edx","edx",4)); # *5
312 &cmp ("ebp",&wparam(2)); # done yet?
313 &jne (&label("loop"));
315 &mov ("edx",&wparam(0)); # ctx
317 &mov (&DWP(4*0,"edx"),"eax"); # store hash value
318 &mov (&DWP(4*1,"edx"),"ebx");
319 &mov (&DWP(4*2,"edx"),"ecx");
320 &mov (&DWP(4*3,"edx"),"esi");
321 &mov (&DWP(4*4,"edx"),"edi");
322 &set_label("nodata");
323 &function_end("poly1305_blocks");
325 &function_begin("poly1305_emit");
326 &mov ("ebp",&wparam(0)); # context
327 &set_label("enter_emit");
328 &mov ("edi",&wparam(1)); # output
329 &mov ("eax",&DWP(4*0,"ebp")); # load hash value
330 &mov ("ebx",&DWP(4*1,"ebp"));
331 &mov ("ecx",&DWP(4*2,"ebp"));
332 &mov ("edx",&DWP(4*3,"ebp"));
333 &mov ("esi",&DWP(4*4,"ebp"));
335 &add ("eax",5); # compare to modulus
340 &shr ("esi",2); # did it carry/borrow?
341 &neg ("esi"); # do we choose hash-modulus?
347 &mov (&DWP(4*0,"edi"),"eax");
348 &mov (&DWP(4*1,"edi"),"ebx");
349 &mov (&DWP(4*2,"edi"),"ecx");
350 &mov (&DWP(4*3,"edi"),"edx");
352 ¬ ("esi"); # or original hash value?
353 &mov ("eax",&DWP(4*0,"ebp"));
354 &mov ("ebx",&DWP(4*1,"ebp"));
355 &mov ("ecx",&DWP(4*2,"ebp"));
356 &mov ("edx",&DWP(4*3,"ebp"));
357 &mov ("ebp",&wparam(2));
362 &or ("eax",&DWP(4*0,"edi"));
363 &or ("ebx",&DWP(4*1,"edi"));
364 &or ("ecx",&DWP(4*2,"edi"));
365 &or ("edx",&DWP(4*3,"edi"));
367 &add ("eax",&DWP(4*0,"ebp")); # accumulate key
368 &adc ("ebx",&DWP(4*1,"ebp"));
369 &adc ("ecx",&DWP(4*2,"ebp"));
370 &adc ("edx",&DWP(4*3,"ebp"));
372 &mov (&DWP(4*0,"edi"),"eax");
373 &mov (&DWP(4*1,"edi"),"ebx");
374 &mov (&DWP(4*2,"edi"),"ecx");
375 &mov (&DWP(4*3,"edi"),"edx");
376 &function_end("poly1305_emit");
379 ########################################################################
380 # Layout of opaque area is following.
382 # unsigned __int32 h[5]; # current hash value base 2^26
383 # unsigned __int32 is_base2_26;
384 # unsigned __int32 r[4]; # key value base 2^32
385 # unsigned __int32 pad[2];
386 # struct { unsigned __int32 r^4, r^3, r^2, r^1; } r[9];
388 # where r^n are base 2^26 digits of degrees of multiplier key. There are
389 # 5 digits, but last four are interleaved with multiples of 5, totalling
390 # in 9 elements: r0, r1, 5*r1, r2, 5*r2, r3, 5*r3, r4, 5*r4.
392 my ($D0,$D1,$D2,$D3,$D4,$T0,$T1,$T2)=map("xmm$_",(0..7));
393 my $MASK=$T2; # borrow and keep in mind
396 &function_begin_B("_poly1305_init_sse2");
397 &movdqu ($D4,&QWP(4*6,"edi")); # key base 2^32
398 &lea ("edi",&DWP(16*3,"edi")); # size optimization
400 &sub ("esp",16*(9+5));
403 #&pand ($D4,&QWP(96,"ebx")); # magic mask
404 &movq ($MASK,&QWP(64,"ebx"));
410 &pand ($D0,$MASK); # -> base 2^26
421 &lea ("edx",&DWP(16*9,"esp")); # size optimization
423 &set_label("square");
424 &movdqa (&QWP(16*0,"esp"),$D0);
425 &movdqa (&QWP(16*1,"esp"),$D1);
426 &movdqa (&QWP(16*2,"esp"),$D2);
427 &movdqa (&QWP(16*3,"esp"),$D3);
428 &movdqa (&QWP(16*4,"esp"),$D4);
434 &paddd ($T1,$D1); # *5
435 &paddd ($T0,$D2); # *5
436 &movdqa (&QWP(16*5,"esp"),$T1);
437 &movdqa (&QWP(16*6,"esp"),$T0);
442 &paddd ($T1,$D3); # *5
443 &paddd ($T0,$D4); # *5
444 &movdqa (&QWP(16*7,"esp"),$T1);
445 &movdqa (&QWP(16*8,"esp"),$T0);
447 &pshufd ($T1,$D0,0b01000100);
449 &pshufd ($D1,$D1,0b01000100);
450 &pshufd ($D2,$D2,0b01000100);
451 &pshufd ($D3,$D3,0b01000100);
452 &pshufd ($D4,$D4,0b01000100);
453 &movdqa (&QWP(16*0,"edx"),$T1);
454 &movdqa (&QWP(16*1,"edx"),$D1);
455 &movdqa (&QWP(16*2,"edx"),$D2);
456 &movdqa (&QWP(16*3,"edx"),$D3);
457 &movdqa (&QWP(16*4,"edx"),$D4);
459 ################################################################
460 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
461 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
462 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
463 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
464 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
466 &pmuludq ($D4,$D0); # h4*r0
467 &pmuludq ($D3,$D0); # h3*r0
468 &pmuludq ($D2,$D0); # h2*r0
469 &pmuludq ($D1,$D0); # h1*r0
470 &pmuludq ($D0,$T1); # h0*r0
474 my $base = shift; $base = "esp" if (!defined($base));
476 ################################################################
477 # As for choice to "rotate" $T0-$T2 in order to move paddq
478 # past next multiplication. While it makes code harder to read
479 # and doesn't have significant effect on most processors, it
480 # makes a lot of difference on Atom, up to 30% improvement.
483 &pmuludq ($T0,&QWP(16*3,$base)); # r1*h3
485 &pmuludq ($T1,&QWP(16*2,$base)); # r1*h2
488 &pmuludq ($T2,&QWP(16*1,$base)); # r1*h1
491 &pmuludq ($T0,&QWP(16*0,$base)); # r1*h0
493 &pmuludq ($T1,&QWP(16*4,$base)); # s1*h4
494 &$load ($T2,2); # r2^n
498 &pmuludq ($T2,&QWP(16*2,$base)); # r2*h2
501 &pmuludq ($T0,&QWP(16*1,$base)); # r2*h1
503 &$load ($T2,6); # s2^n
504 &pmuludq ($T1,&QWP(16*0,$base)); # r2*h0
507 &pmuludq ($T2,&QWP(16*4,$base)); # s2*h4
509 &pmuludq ($T0,&QWP(16*3,$base)); # s2*h3
510 &$load ($T1,3); # r3^n
514 &pmuludq ($T1,&QWP(16*1,$base)); # r3*h1
516 &$load ($T0,7); # s3^n
517 &pmuludq ($T2,&QWP(16*0,$base)); # r3*h0
520 &pmuludq ($T0,&QWP(16*4,$base)); # s3*h4
523 &pmuludq ($T1,&QWP(16*3,$base)); # s3*h3
525 &pmuludq ($T2,&QWP(16*2,$base)); # s3*h2
526 &$load ($T0,4); # r4^n
529 &$load ($T1,8); # s4^n
530 &pmuludq ($T0,&QWP(16*0,$base)); # r4*h0
533 &pmuludq ($T1,&QWP(16*4,$base)); # s4*h4
536 &pmuludq ($T2,&QWP(16*1,$base)); # s4*h1
539 &pmuludq ($T0,&QWP(16*2,$base)); # s4*h2
541 &pmuludq ($T1,&QWP(16*3,$base)); # s4*h3
542 &movdqa ($MASK,&QWP(64,"ebx"));
546 &pmuladd (sub { my ($reg,$i)=@_;
547 &movdqa ($reg,&QWP(16*$i,"esp"));
553 ################################################################
554 # lazy reduction as discussed in "NEON crypto" by D.J. Bernstein
557 # [(*) see discussion in poly1305-armv4 module]
562 &$extra () if (defined($extra));
563 &paddq ($T0,$D4); # h3 -> h4
568 &paddq ($T1,$D1); # h0 -> h1
573 &paddd ($D0,$T0); # favour paddd when
578 &paddq ($T1,$D2); # h1 -> h2
579 &paddq ($T0,$D0); # h4 -> h0 (*)
584 &paddd ($T1,$D3); # h2 -> h3
590 &paddd ($D1,$T0); # h0 -> h1
592 &paddd ($D4,$T1); # h3 -> h4
597 &jz (&label("square_break"));
599 &punpcklqdq ($D0,&QWP(16*0,"esp")); # 0:r^1:0:r^2
600 &punpcklqdq ($D1,&QWP(16*1,"esp"));
601 &punpcklqdq ($D2,&QWP(16*2,"esp"));
602 &punpcklqdq ($D3,&QWP(16*3,"esp"));
603 &punpcklqdq ($D4,&QWP(16*4,"esp"));
604 &jmp (&label("square"));
606 &set_label("square_break");
607 &psllq ($D0,32); # -> r^3:0:r^4:0
612 &por ($D0,&QWP(16*0,"esp")); # r^3:r^1:r^4:r^2
613 &por ($D1,&QWP(16*1,"esp"));
614 &por ($D2,&QWP(16*2,"esp"));
615 &por ($D3,&QWP(16*3,"esp"));
616 &por ($D4,&QWP(16*4,"esp"));
618 &pshufd ($D0,$D0,0b10001101); # -> r^1:r^2:r^3:r^4
619 &pshufd ($D1,$D1,0b10001101);
620 &pshufd ($D2,$D2,0b10001101);
621 &pshufd ($D3,$D3,0b10001101);
622 &pshufd ($D4,$D4,0b10001101);
624 &movdqu (&QWP(16*0,"edi"),$D0); # save the table
625 &movdqu (&QWP(16*1,"edi"),$D1);
626 &movdqu (&QWP(16*2,"edi"),$D2);
627 &movdqu (&QWP(16*3,"edi"),$D3);
628 &movdqu (&QWP(16*4,"edi"),$D4);
634 &paddd ($T1,$D1); # *5
635 &paddd ($T0,$D2); # *5
636 &movdqu (&QWP(16*5,"edi"),$T1);
637 &movdqu (&QWP(16*6,"edi"),$T0);
642 &paddd ($T1,$D3); # *5
643 &paddd ($T0,$D4); # *5
644 &movdqu (&QWP(16*7,"edi"),$T1);
645 &movdqu (&QWP(16*8,"edi"),$T0);
648 &lea ("edi",&DWP(-16*3,"edi")); # size de-optimization
650 &function_end_B("_poly1305_init_sse2");
653 &function_begin("_poly1305_blocks_sse2");
654 &mov ("edi",&wparam(0)); # ctx
655 &mov ("esi",&wparam(1)); # inp
656 &mov ("ecx",&wparam(2)); # len
658 &mov ("eax",&DWP(4*5,"edi")); # is_base2_26
660 &jz (&label("nodata"));
662 &jae (&label("enter_sse2"));
663 &test ("eax","eax"); # is_base2_26?
664 &jz (&label("enter_blocks"));
666 &set_label("enter_sse2",16);
667 &call (&label("pic_point"));
668 &set_label("pic_point");
670 &lea ("ebx",&DWP(&label("const_sse2")."-".&label("pic_point"),"ebx"));
672 &test ("eax","eax"); # is_base2_26?
673 &jnz (&label("base2_26"));
675 &call ("_poly1305_init_sse2");
677 ################################################# base 2^32 -> base 2^26
678 &mov ("eax",&DWP(0,"edi"));
679 &mov ("ecx",&DWP(3,"edi"));
680 &mov ("edx",&DWP(6,"edi"));
681 &mov ("esi",&DWP(9,"edi"));
682 &mov ("ebp",&DWP(13,"edi"));
683 &mov (&DWP(4*5,"edi"),1); # is_base2_26
686 &and ("eax",0x3ffffff);
688 &and ("ecx",0x3ffffff);
690 &and ("edx",0x3ffffff);
698 &mov ("esi",&wparam(1)); # [reload] inp
699 &mov ("ecx",&wparam(2)); # [reload] len
700 &jmp (&label("base2_32"));
702 &set_label("base2_26",16);
703 &movd ($D0,&DWP(4*0,"edi")); # load hash value
704 &movd ($D1,&DWP(4*1,"edi"));
705 &movd ($D2,&DWP(4*2,"edi"));
706 &movd ($D3,&DWP(4*3,"edi"));
707 &movd ($D4,&DWP(4*4,"edi"));
708 &movdqa ($MASK,&QWP(64,"ebx"));
710 &set_label("base2_32");
711 &mov ("eax",&wparam(3)); # padbit
714 &sub ("esp",16*(5+5+5+9+9));
717 &lea ("edi",&DWP(16*3,"edi")); # size optimization
718 &shl ("eax",24); # padbit
721 &jz (&label("even"));
723 ################################################################
724 # process single block, with SSE2, because it's still faster
725 # even though half of result is discarded
727 &movdqu ($T1,&QWP(0,"esi")); # input
728 &lea ("esi",&DWP(16,"esi"));
730 &movdqa ($T0,$T1); # -> base 2^26 ...
732 &paddd ($D0,$T1); # ... and accumulate
751 &movd ($T0,"eax"); # padbit
753 &movd ($T1,&DWP(16*0+12,"edi")); # r0
756 &movdqa (&QWP(16*0,"esp"),$D0);
757 &movdqa (&QWP(16*1,"esp"),$D1);
758 &movdqa (&QWP(16*2,"esp"),$D2);
759 &movdqa (&QWP(16*3,"esp"),$D3);
760 &movdqa (&QWP(16*4,"esp"),$D4);
762 ################################################################
763 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
764 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
765 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
766 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
767 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
769 &pmuludq ($D0,$T1); # h4*r0
770 &pmuludq ($D1,$T1); # h3*r0
771 &pmuludq ($D2,$T1); # h2*r0
772 &movd ($T0,&DWP(16*1+12,"edi")); # r1
773 &pmuludq ($D3,$T1); # h1*r0
774 &pmuludq ($D4,$T1); # h0*r0
776 &pmuladd (sub { my ($reg,$i)=@_;
777 &movd ($reg,&DWP(16*$i+12,"edi"));
783 &jz (&label("done"));
786 &lea ("edx",&DWP(16*(5+5+5+9),"esp"));# size optimization
787 &lea ("eax",&DWP(-16*2,"esi"));
790 ################################################################
791 # expand and copy pre-calculated table to stack
793 &movdqu ($T0,&QWP(16*0,"edi")); # r^1:r^2:r^3:r^4
794 &pshufd ($T1,$T0,0b01000100); # duplicate r^3:r^4
795 &cmovb ("esi","eax");
796 &pshufd ($T0,$T0,0b11101110); # duplicate r^1:r^2
797 &movdqa (&QWP(16*0,"edx"),$T1);
798 &lea ("eax",&DWP(16*10,"esp"));
799 &movdqu ($T1,&QWP(16*1,"edi"));
800 &movdqa (&QWP(16*(0-9),"edx"),$T0);
801 &pshufd ($T0,$T1,0b01000100);
802 &pshufd ($T1,$T1,0b11101110);
803 &movdqa (&QWP(16*1,"edx"),$T0);
804 &movdqu ($T0,&QWP(16*2,"edi"));
805 &movdqa (&QWP(16*(1-9),"edx"),$T1);
806 &pshufd ($T1,$T0,0b01000100);
807 &pshufd ($T0,$T0,0b11101110);
808 &movdqa (&QWP(16*2,"edx"),$T1);
809 &movdqu ($T1,&QWP(16*3,"edi"));
810 &movdqa (&QWP(16*(2-9),"edx"),$T0);
811 &pshufd ($T0,$T1,0b01000100);
812 &pshufd ($T1,$T1,0b11101110);
813 &movdqa (&QWP(16*3,"edx"),$T0);
814 &movdqu ($T0,&QWP(16*4,"edi"));
815 &movdqa (&QWP(16*(3-9),"edx"),$T1);
816 &pshufd ($T1,$T0,0b01000100);
817 &pshufd ($T0,$T0,0b11101110);
818 &movdqa (&QWP(16*4,"edx"),$T1);
819 &movdqu ($T1,&QWP(16*5,"edi"));
820 &movdqa (&QWP(16*(4-9),"edx"),$T0);
821 &pshufd ($T0,$T1,0b01000100);
822 &pshufd ($T1,$T1,0b11101110);
823 &movdqa (&QWP(16*5,"edx"),$T0);
824 &movdqu ($T0,&QWP(16*6,"edi"));
825 &movdqa (&QWP(16*(5-9),"edx"),$T1);
826 &pshufd ($T1,$T0,0b01000100);
827 &pshufd ($T0,$T0,0b11101110);
828 &movdqa (&QWP(16*6,"edx"),$T1);
829 &movdqu ($T1,&QWP(16*7,"edi"));
830 &movdqa (&QWP(16*(6-9),"edx"),$T0);
831 &pshufd ($T0,$T1,0b01000100);
832 &pshufd ($T1,$T1,0b11101110);
833 &movdqa (&QWP(16*7,"edx"),$T0);
834 &movdqu ($T0,&QWP(16*8,"edi"));
835 &movdqa (&QWP(16*(7-9),"edx"),$T1);
836 &pshufd ($T1,$T0,0b01000100);
837 &pshufd ($T0,$T0,0b11101110);
838 &movdqa (&QWP(16*8,"edx"),$T1);
839 &movdqa (&QWP(16*(8-9),"edx"),$T0);
842 my ($inpbase,$offbase)=@_;
844 &movdqu ($T0,&QWP($inpbase+0,"esi")); # load input
845 &movdqu ($T1,&QWP($inpbase+16,"esi"));
846 &lea ("esi",&DWP(16*2,"esi"));
848 &movdqa (&QWP($offbase+16*2,"esp"),$D2);
849 &movdqa (&QWP($offbase+16*3,"esp"),$D3);
850 &movdqa (&QWP($offbase+16*4,"esp"),$D4);
852 &movdqa ($D2,$T0); # splat input
857 &punpcklqdq ($D2,$D3); # 2:3
858 &punpckhqdq ($D4,$T1); # 4
859 &punpcklqdq ($T0,$T1); # 0:1
867 &pand ($T0,$MASK); # 0
868 &pand ($T1,$MASK); # 1
869 &pand ($D2,$MASK); # 2
870 &pand ($D3,$MASK); # 3
871 &por ($D4,&QWP(0,"ebx")); # padbit, yes, always
873 &movdqa (&QWP($offbase+16*0,"esp"),$D0) if ($offbase);
874 &movdqa (&QWP($offbase+16*1,"esp"),$D1) if ($offbase);
876 &load_input (16*2,16*5);
878 &jbe (&label("skip_loop"));
879 &jmp (&label("loop"));
881 &set_label("loop",32);
882 ################################################################
883 # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2
884 # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^3+inp[7]*r
885 # \___________________/
886 # ((inp[0]*r^4+inp[2]*r^2+inp[4])*r^4+inp[6]*r^2+inp[8])*r^2
887 # ((inp[1]*r^4+inp[3]*r^2+inp[5])*r^4+inp[7]*r^2+inp[9])*r
888 # \___________________/ \____________________/
889 ################################################################
891 &movdqa ($T2,&QWP(16*(0-9),"edx")); # r0^2
892 &movdqa (&QWP(16*1,"eax"),$T1);
893 &movdqa (&QWP(16*2,"eax"),$D2);
894 &movdqa (&QWP(16*3,"eax"),$D3);
895 &movdqa (&QWP(16*4,"eax"),$D4);
897 ################################################################
898 # d4 = h4*r0 + h0*r4 + h1*r3 + h2*r2 + h3*r1
899 # d3 = h3*r0 + h0*r3 + h1*r2 + h2*r1 + h4*5*r4
900 # d2 = h2*r0 + h0*r2 + h1*r1 + h3*5*r4 + h4*5*r3
901 # d1 = h1*r0 + h0*r1 + h2*5*r4 + h3*5*r3 + h4*5*r2
902 # d0 = h0*r0 + h1*5*r4 + h2*5*r3 + h3*5*r2 + h4*5*r1
905 &pmuludq ($T0,$T2); # h0*r0
907 &pmuludq ($T1,$T2); # h1*r0
908 &pmuludq ($D2,$T2); # h2*r0
909 &pmuludq ($D3,$T2); # h3*r0
910 &pmuludq ($D4,$T2); # h4*r0
915 &pmuludq ($D0,&$addr(8)); # h1*s4
917 &pmuludq ($D1,&$addr(1)); # h0*r1
920 &pmuludq ($T2,&$addr(2)); # h0*r2
923 &pmuludq ($T0,&$addr(3)); # h0*r3
925 &movdqa ($T2,&QWP(16*1,"eax")); # pull h1
926 &pmuludq ($T1,&$addr(4)); # h0*r4
930 &pmuludq ($T2,&$addr(1)); # h1*r1
933 &pmuludq ($T0,&$addr(2)); # h1*r2
935 &movdqa ($T2,&QWP(16*2,"eax")); # pull h2
936 &pmuludq ($T1,&$addr(3)); # h1*r3
939 &pmuludq ($T2,&$addr(7)); # h2*s3
942 &pmuludq ($T0,&$addr(8)); # h2*s4
946 &pmuludq ($T1,&$addr(1)); # h2*r1
948 &movdqa ($T0,&QWP(16*3,"eax")); # pull h3
949 &pmuludq ($T2,&$addr(2)); # h2*r2
952 &pmuludq ($T0,&$addr(6)); # h3*s2
955 &pmuludq ($T1,&$addr(7)); # h3*s3
958 &pmuludq ($T2,&$addr(8)); # h3*s4
961 &movdqa ($T1,&QWP(16*4,"eax")); # pull h4
962 &pmuludq ($T0,&$addr(1)); # h3*r1
965 &pmuludq ($T1,&$addr(8)); # h4*s4
968 &pmuludq ($T2,&$addr(5)); # h4*s1
971 &pmuludq ($T0,&$addr(6)); # h4*s2
973 &movdqa ($MASK,&QWP(64,"ebx"));
974 &pmuludq ($T1,&$addr(7)); # h4*s3
978 &pmuladd_alt (sub { my $i=shift; &QWP(16*($i-9),"edx"); });
980 &load_input (-16*2,0);
981 &lea ("eax",&DWP(-16*2,"esi"));
984 &paddd ($T0,&QWP(16*(5+0),"esp")); # add hash value
985 &paddd ($T1,&QWP(16*(5+1),"esp"));
986 &paddd ($D2,&QWP(16*(5+2),"esp"));
987 &paddd ($D3,&QWP(16*(5+3),"esp"));
988 &paddd ($D4,&QWP(16*(5+4),"esp"));
990 &cmovb ("esi","eax");
991 &lea ("eax",&DWP(16*10,"esp"));
993 &movdqa ($T2,&QWP(16*0,"edx")); # r0^4
994 &movdqa (&QWP(16*1,"esp"),$D1);
995 &movdqa (&QWP(16*1,"eax"),$T1);
996 &movdqa (&QWP(16*2,"eax"),$D2);
997 &movdqa (&QWP(16*3,"eax"),$D3);
998 &movdqa (&QWP(16*4,"eax"),$D4);
1000 ################################################################
1001 # d4 += h4*r0 + h0*r4 + h1*r3 + h2*r2 + h3*r1
1002 # d3 += h3*r0 + h0*r3 + h1*r2 + h2*r1 + h4*5*r4
1003 # d2 += h2*r0 + h0*r2 + h1*r1 + h3*5*r4 + h4*5*r3
1004 # d1 += h1*r0 + h0*r1 + h2*5*r4 + h3*5*r3 + h4*5*r2
1005 # d0 += h0*r0 + h1*5*r4 + h2*5*r3 + h3*5*r2 + h4*5*r1
1008 &pmuludq ($T0,$T2); # h0*r0
1011 &pmuludq ($T1,$T2); # h1*r0
1012 &pmuludq ($D2,$T2); # h2*r0
1013 &pmuludq ($D3,$T2); # h3*r0
1014 &pmuludq ($D4,$T2); # h4*r0
1016 &paddq ($T1,&QWP(16*1,"esp"));
1017 &paddq ($D2,&QWP(16*2,"esp"));
1018 &paddq ($D3,&QWP(16*3,"esp"));
1019 &paddq ($D4,&QWP(16*4,"esp"));
1021 &pmuladd_alt (sub { my $i=shift; &QWP(16*$i,"edx"); });
1025 &load_input (16*2,16*5);
1027 &ja (&label("loop"));
1029 &set_label("skip_loop");
1030 ################################################################
1031 # multiply (inp[0:1]+hash) or inp[2:3] by r^2:r^1
1033 &pshufd ($T2,&QWP(16*(0-9),"edx"),0x10);# r0^n
1035 &jnz (&label("long_tail"));
1037 &paddd ($T0,$D0); # add hash value
1039 &paddd ($D2,&QWP(16*7,"esp"));
1040 &paddd ($D3,&QWP(16*8,"esp"));
1041 &paddd ($D4,&QWP(16*9,"esp"));
1043 &set_label("long_tail");
1045 &movdqa (&QWP(16*0,"eax"),$T0);
1046 &movdqa (&QWP(16*1,"eax"),$T1);
1047 &movdqa (&QWP(16*2,"eax"),$D2);
1048 &movdqa (&QWP(16*3,"eax"),$D3);
1049 &movdqa (&QWP(16*4,"eax"),$D4);
1051 ################################################################
1052 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
1053 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
1054 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1055 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
1056 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1058 &pmuludq ($T0,$T2); # h0*r0
1059 &pmuludq ($T1,$T2); # h1*r0
1060 &pmuludq ($D2,$T2); # h2*r0
1062 &pshufd ($T0,&QWP(16*(1-9),"edx"),0x10);# r1^n
1063 &pmuludq ($D3,$T2); # h3*r0
1065 &pmuludq ($D4,$T2); # h4*r0
1067 &pmuladd (sub { my ($reg,$i)=@_;
1068 &pshufd ($reg,&QWP(16*($i-9),"edx"),0x10);
1071 &jz (&label("short_tail"));
1073 &load_input (-16*2,0);
1075 &pshufd ($T2,&QWP(16*0,"edx"),0x10); # r0^n
1076 &paddd ($T0,&QWP(16*5,"esp")); # add hash value
1077 &paddd ($T1,&QWP(16*6,"esp"));
1078 &paddd ($D2,&QWP(16*7,"esp"));
1079 &paddd ($D3,&QWP(16*8,"esp"));
1080 &paddd ($D4,&QWP(16*9,"esp"));
1082 ################################################################
1083 # multiply inp[0:1] by r^4:r^3 and accumulate
1085 &movdqa (&QWP(16*0,"esp"),$T0);
1086 &pmuludq ($T0,$T2); # h0*r0
1087 &movdqa (&QWP(16*1,"esp"),$T1);
1088 &pmuludq ($T1,$T2); # h1*r0
1091 &pmuludq ($D2,$T2); # h2*r0
1094 &pmuludq ($D3,$T2); # h3*r0
1095 &paddq ($D2,&QWP(16*2,"esp"));
1096 &movdqa (&QWP(16*2,"esp"),$T0);
1097 &pshufd ($T0,&QWP(16*1,"edx"),0x10); # r1^n
1098 &paddq ($D3,&QWP(16*3,"esp"));
1099 &movdqa (&QWP(16*3,"esp"),$T1);
1101 &pmuludq ($D4,$T2); # h4*r0
1102 &paddq ($D4,&QWP(16*4,"esp"));
1103 &movdqa (&QWP(16*4,"esp"),$T1);
1105 &pmuladd (sub { my ($reg,$i)=@_;
1106 &pshufd ($reg,&QWP(16*$i,"edx"),0x10);
1109 &set_label("short_tail");
1111 ################################################################
1112 # horizontal addition
1114 &pshufd ($T1,$D4,0b01001110);
1115 &pshufd ($T0,$D3,0b01001110);
1118 &pshufd ($T1,$D0,0b01001110);
1119 &pshufd ($T0,$D1,0b01001110);
1122 &pshufd ($T1,$D2,0b01001110);
1125 &lazy_reduction (sub { &paddq ($D2,$T1) });
1128 &movd (&DWP(-16*3+4*0,"edi"),$D0); # store hash value
1129 &movd (&DWP(-16*3+4*1,"edi"),$D1);
1130 &movd (&DWP(-16*3+4*2,"edi"),$D2);
1131 &movd (&DWP(-16*3+4*3,"edi"),$D3);
1132 &movd (&DWP(-16*3+4*4,"edi"),$D4);
1134 &set_label("nodata");
1135 &function_end("_poly1305_blocks_sse2");
1138 &function_begin("_poly1305_emit_sse2");
1139 &mov ("ebp",&wparam(0)); # context
1141 &cmp (&DWP(4*5,"ebp"),0); # is_base2_26?
1142 &je (&label("enter_emit"));
1144 &mov ("eax",&DWP(4*0,"ebp")); # load hash value
1145 &mov ("edi",&DWP(4*1,"ebp"));
1146 &mov ("ecx",&DWP(4*2,"ebp"));
1147 &mov ("edx",&DWP(4*3,"ebp"));
1148 &mov ("esi",&DWP(4*4,"ebp"));
1150 &mov ("ebx","edi"); # base 2^26 -> base 2^32
1172 &adc ("esi",0); # can be partially reduced
1174 &mov ("edi","esi"); # final reduction
1177 &lea ("ebp",&DWP(0,"edi","edi",4)); # *5
1178 &mov ("edi",&wparam(1)); # output
1180 &mov ("ebp",&wparam(2)); # key
1186 &movd ($D0,"eax"); # offload original hash value
1187 &add ("eax",5); # compare to modulus
1195 &shr ("esi",2); # did it carry/borrow?
1197 &neg ("esi"); # do we choose (hash-modulus) ...
1202 &mov (&DWP(4*0,"edi"),"eax");
1204 &mov (&DWP(4*1,"edi"),"ebx");
1206 &mov (&DWP(4*2,"edi"),"ecx");
1208 &mov (&DWP(4*3,"edi"),"edx");
1211 ¬ ("esi"); # ... or original hash value?
1214 &or ("eax",&DWP(4*0,"edi"));
1216 &or ("ebx",&DWP(4*1,"edi"));
1218 &or ("ecx",&DWP(4*2,"edi"));
1219 &or ("edx",&DWP(4*3,"edi"));
1221 &add ("eax",&DWP(4*0,"ebp")); # accumulate key
1222 &adc ("ebx",&DWP(4*1,"ebp"));
1223 &mov (&DWP(4*0,"edi"),"eax");
1224 &adc ("ecx",&DWP(4*2,"ebp"));
1225 &mov (&DWP(4*1,"edi"),"ebx");
1226 &adc ("edx",&DWP(4*3,"ebp"));
1227 &mov (&DWP(4*2,"edi"),"ecx");
1228 &mov (&DWP(4*3,"edi"),"edx");
1229 &function_end("_poly1305_emit_sse2");
1232 ########################################################################
1233 # Note that poly1305_init_avx2 operates on %xmm, I could have used
1234 # poly1305_init_sse2...
1237 &function_begin_B("_poly1305_init_avx2");
1238 &vmovdqu ($D4,&QWP(4*6,"edi")); # key base 2^32
1239 &lea ("edi",&DWP(16*3,"edi")); # size optimization
1241 &sub ("esp",16*(9+5));
1244 #&vpand ($D4,$D4,&QWP(96,"ebx")); # magic mask
1245 &vmovdqa ($MASK,&QWP(64,"ebx"));
1247 &vpand ($D0,$D4,$MASK); # -> base 2^26
1248 &vpsrlq ($D1,$D4,26);
1249 &vpsrldq ($D3,$D4,6);
1250 &vpand ($D1,$D1,$MASK);
1252 &vpsrlq ($D3,$D3,30);
1253 &vpand ($D2,$D2,$MASK);
1254 &vpand ($D3,$D3,$MASK);
1255 &vpsrldq ($D4,$D4,13);
1257 &lea ("edx",&DWP(16*9,"esp")); # size optimization
1259 &set_label("square");
1260 &vmovdqa (&QWP(16*0,"esp"),$D0);
1261 &vmovdqa (&QWP(16*1,"esp"),$D1);
1262 &vmovdqa (&QWP(16*2,"esp"),$D2);
1263 &vmovdqa (&QWP(16*3,"esp"),$D3);
1264 &vmovdqa (&QWP(16*4,"esp"),$D4);
1266 &vpslld ($T1,$D1,2);
1267 &vpslld ($T0,$D2,2);
1268 &vpaddd ($T1,$T1,$D1); # *5
1269 &vpaddd ($T0,$T0,$D2); # *5
1270 &vmovdqa (&QWP(16*5,"esp"),$T1);
1271 &vmovdqa (&QWP(16*6,"esp"),$T0);
1272 &vpslld ($T1,$D3,2);
1273 &vpslld ($T0,$D4,2);
1274 &vpaddd ($T1,$T1,$D3); # *5
1275 &vpaddd ($T0,$T0,$D4); # *5
1276 &vmovdqa (&QWP(16*7,"esp"),$T1);
1277 &vmovdqa (&QWP(16*8,"esp"),$T0);
1279 &vpshufd ($T0,$D0,0b01000100);
1281 &vpshufd ($D1,$D1,0b01000100);
1282 &vpshufd ($D2,$D2,0b01000100);
1283 &vpshufd ($D3,$D3,0b01000100);
1284 &vpshufd ($D4,$D4,0b01000100);
1285 &vmovdqa (&QWP(16*0,"edx"),$T0);
1286 &vmovdqa (&QWP(16*1,"edx"),$D1);
1287 &vmovdqa (&QWP(16*2,"edx"),$D2);
1288 &vmovdqa (&QWP(16*3,"edx"),$D3);
1289 &vmovdqa (&QWP(16*4,"edx"),$D4);
1291 ################################################################
1292 # d4 = h4*r0 + h3*r1 + h2*r2 + h1*r3 + h0*r4
1293 # d3 = h3*r0 + h2*r1 + h1*r2 + h0*r3 + h4*5*r4
1294 # d2 = h2*r0 + h1*r1 + h0*r2 + h4*5*r3 + h3*5*r4
1295 # d1 = h1*r0 + h0*r1 + h4*5*r2 + h3*5*r3 + h2*5*r4
1296 # d0 = h0*r0 + h4*5*r1 + h3*5*r2 + h2*5*r3 + h1*5*r4
1298 &vpmuludq ($D4,$D4,$D0); # h4*r0
1299 &vpmuludq ($D3,$D3,$D0); # h3*r0
1300 &vpmuludq ($D2,$D2,$D0); # h2*r0
1301 &vpmuludq ($D1,$D1,$D0); # h1*r0
1302 &vpmuludq ($D0,$T0,$D0); # h0*r0
1304 &vpmuludq ($T0,$T1,&QWP(16*3,"edx")); # r1*h3
1305 &vpaddq ($D4,$D4,$T0);
1306 &vpmuludq ($T2,$T1,&QWP(16*2,"edx")); # r1*h2
1307 &vpaddq ($D3,$D3,$T2);
1308 &vpmuludq ($T0,$T1,&QWP(16*1,"edx")); # r1*h1
1309 &vpaddq ($D2,$D2,$T0);
1310 &vmovdqa ($T2,&QWP(16*5,"esp")); # s1
1311 &vpmuludq ($T1,$T1,&QWP(16*0,"edx")); # r1*h0
1312 &vpaddq ($D1,$D1,$T1);
1313 &vmovdqa ($T0,&QWP(16*2,"esp")); # r2
1314 &vpmuludq ($T2,$T2,&QWP(16*4,"edx")); # s1*h4
1315 &vpaddq ($D0,$D0,$T2);
1317 &vpmuludq ($T1,$T0,&QWP(16*2,"edx")); # r2*h2
1318 &vpaddq ($D4,$D4,$T1);
1319 &vpmuludq ($T2,$T0,&QWP(16*1,"edx")); # r2*h1
1320 &vpaddq ($D3,$D3,$T2);
1321 &vmovdqa ($T1,&QWP(16*6,"esp")); # s2
1322 &vpmuludq ($T0,$T0,&QWP(16*0,"edx")); # r2*h0
1323 &vpaddq ($D2,$D2,$T0);
1324 &vpmuludq ($T2,$T1,&QWP(16*4,"edx")); # s2*h4
1325 &vpaddq ($D1,$D1,$T2);
1326 &vmovdqa ($T0,&QWP(16*3,"esp")); # r3
1327 &vpmuludq ($T1,$T1,&QWP(16*3,"edx")); # s2*h3
1328 &vpaddq ($D0,$D0,$T1);
1330 &vpmuludq ($T2,$T0,&QWP(16*1,"edx")); # r3*h1
1331 &vpaddq ($D4,$D4,$T2);
1332 &vmovdqa ($T1,&QWP(16*7,"esp")); # s3
1333 &vpmuludq ($T0,$T0,&QWP(16*0,"edx")); # r3*h0
1334 &vpaddq ($D3,$D3,$T0);
1335 &vpmuludq ($T2,$T1,&QWP(16*4,"edx")); # s3*h4
1336 &vpaddq ($D2,$D2,$T2);
1337 &vpmuludq ($T0,$T1,&QWP(16*3,"edx")); # s3*h3
1338 &vpaddq ($D1,$D1,$T0);
1339 &vmovdqa ($T2,&QWP(16*4,"esp")); # r4
1340 &vpmuludq ($T1,$T1,&QWP(16*2,"edx")); # s3*h2
1341 &vpaddq ($D0,$D0,$T1);
1343 &vmovdqa ($T0,&QWP(16*8,"esp")); # s4
1344 &vpmuludq ($T2,$T2,&QWP(16*0,"edx")); # r4*h0
1345 &vpaddq ($D4,$D4,$T2);
1346 &vpmuludq ($T1,$T0,&QWP(16*4,"edx")); # s4*h4
1347 &vpaddq ($D3,$D3,$T1);
1348 &vpmuludq ($T2,$T0,&QWP(16*1,"edx")); # s4*h1
1349 &vpaddq ($D0,$D0,$T2);
1350 &vpmuludq ($T1,$T0,&QWP(16*2,"edx")); # s4*h2
1351 &vpaddq ($D1,$D1,$T1);
1352 &vmovdqa ($MASK,&QWP(64,"ebx"));
1353 &vpmuludq ($T0,$T0,&QWP(16*3,"edx")); # s4*h3
1354 &vpaddq ($D2,$D2,$T0);
1356 ################################################################
1358 &vpsrlq ($T0,$D3,26);
1359 &vpand ($D3,$D3,$MASK);
1360 &vpsrlq ($T1,$D0,26);
1361 &vpand ($D0,$D0,$MASK);
1362 &vpaddq ($D4,$D4,$T0); # h3 -> h4
1363 &vpaddq ($D1,$D1,$T1); # h0 -> h1
1364 &vpsrlq ($T0,$D4,26);
1365 &vpand ($D4,$D4,$MASK);
1366 &vpsrlq ($T1,$D1,26);
1367 &vpand ($D1,$D1,$MASK);
1368 &vpaddq ($D2,$D2,$T1); # h1 -> h2
1369 &vpaddd ($D0,$D0,$T0);
1370 &vpsllq ($T0,$T0,2);
1371 &vpsrlq ($T1,$D2,26);
1372 &vpand ($D2,$D2,$MASK);
1373 &vpaddd ($D0,$D0,$T0); # h4 -> h0
1374 &vpaddd ($D3,$D3,$T1); # h2 -> h3
1375 &vpsrlq ($T1,$D3,26);
1376 &vpsrlq ($T0,$D0,26);
1377 &vpand ($D0,$D0,$MASK);
1378 &vpand ($D3,$D3,$MASK);
1379 &vpaddd ($D1,$D1,$T0); # h0 -> h1
1380 &vpaddd ($D4,$D4,$T1); # h3 -> h4
1383 &jz (&label("square_break"));
1385 &vpunpcklqdq ($D0,$D0,&QWP(16*0,"esp")); # 0:r^1:0:r^2
1386 &vpunpcklqdq ($D1,$D1,&QWP(16*1,"esp"));
1387 &vpunpcklqdq ($D2,$D2,&QWP(16*2,"esp"));
1388 &vpunpcklqdq ($D3,$D3,&QWP(16*3,"esp"));
1389 &vpunpcklqdq ($D4,$D4,&QWP(16*4,"esp"));
1390 &jmp (&label("square"));
1392 &set_label("square_break");
1393 &vpsllq ($D0,$D0,32); # -> r^3:0:r^4:0
1394 &vpsllq ($D1,$D1,32);
1395 &vpsllq ($D2,$D2,32);
1396 &vpsllq ($D3,$D3,32);
1397 &vpsllq ($D4,$D4,32);
1398 &vpor ($D0,$D0,&QWP(16*0,"esp")); # r^3:r^1:r^4:r^2
1399 &vpor ($D1,$D1,&QWP(16*1,"esp"));
1400 &vpor ($D2,$D2,&QWP(16*2,"esp"));
1401 &vpor ($D3,$D3,&QWP(16*3,"esp"));
1402 &vpor ($D4,$D4,&QWP(16*4,"esp"));
1404 &vpshufd ($D0,$D0,0b10001101); # -> r^1:r^2:r^3:r^4
1405 &vpshufd ($D1,$D1,0b10001101);
1406 &vpshufd ($D2,$D2,0b10001101);
1407 &vpshufd ($D3,$D3,0b10001101);
1408 &vpshufd ($D4,$D4,0b10001101);
1410 &vmovdqu (&QWP(16*0,"edi"),$D0); # save the table
1411 &vmovdqu (&QWP(16*1,"edi"),$D1);
1412 &vmovdqu (&QWP(16*2,"edi"),$D2);
1413 &vmovdqu (&QWP(16*3,"edi"),$D3);
1414 &vmovdqu (&QWP(16*4,"edi"),$D4);
1416 &vpslld ($T1,$D1,2);
1417 &vpslld ($T0,$D2,2);
1418 &vpaddd ($T1,$T1,$D1); # *5
1419 &vpaddd ($T0,$T0,$D2); # *5
1420 &vmovdqu (&QWP(16*5,"edi"),$T1);
1421 &vmovdqu (&QWP(16*6,"edi"),$T0);
1422 &vpslld ($T1,$D3,2);
1423 &vpslld ($T0,$D4,2);
1424 &vpaddd ($T1,$T1,$D3); # *5
1425 &vpaddd ($T0,$T0,$D4); # *5
1426 &vmovdqu (&QWP(16*7,"edi"),$T1);
1427 &vmovdqu (&QWP(16*8,"edi"),$T0);
1430 &lea ("edi",&DWP(-16*3,"edi")); # size de-optimization
1432 &function_end_B("_poly1305_init_avx2");
1434 ########################################################################
1435 # now it's time to switch to %ymm
1437 my ($D0,$D1,$D2,$D3,$D4,$T0,$T1,$T2)=map("ymm$_",(0..7));
1440 sub X { my $reg=shift; $reg=~s/^ymm/xmm/; $reg; }
1443 &function_begin("_poly1305_blocks_avx2");
1444 &mov ("edi",&wparam(0)); # ctx
1445 &mov ("esi",&wparam(1)); # inp
1446 &mov ("ecx",&wparam(2)); # len
1448 &mov ("eax",&DWP(4*5,"edi")); # is_base2_26
1450 &jz (&label("nodata"));
1452 &jae (&label("enter_avx2"));
1453 &test ("eax","eax"); # is_base2_26?
1454 &jz (&label("enter_blocks"));
1456 &set_label("enter_avx2");
1459 &call (&label("pic_point"));
1460 &set_label("pic_point");
1462 &lea ("ebx",&DWP(&label("const_sse2")."-".&label("pic_point"),"ebx"));
1464 &test ("eax","eax"); # is_base2_26?
1465 &jnz (&label("base2_26"));
1467 &call ("_poly1305_init_avx2");
1469 ################################################# base 2^32 -> base 2^26
1470 &mov ("eax",&DWP(0,"edi"));
1471 &mov ("ecx",&DWP(3,"edi"));
1472 &mov ("edx",&DWP(6,"edi"));
1473 &mov ("esi",&DWP(9,"edi"));
1474 &mov ("ebp",&DWP(13,"edi"));
1477 &and ("eax",0x3ffffff);
1479 &and ("ecx",0x3ffffff);
1481 &and ("edx",0x3ffffff);
1483 &mov (&DWP(4*0,"edi"),"eax");
1484 &mov (&DWP(4*1,"edi"),"ecx");
1485 &mov (&DWP(4*2,"edi"),"edx");
1486 &mov (&DWP(4*3,"edi"),"esi");
1487 &mov (&DWP(4*4,"edi"),"ebp");
1488 &mov (&DWP(4*5,"edi"),1); # is_base2_26
1490 &mov ("esi",&wparam(1)); # [reload] inp
1491 &mov ("ecx",&wparam(2)); # [reload] len
1493 &set_label("base2_26");
1494 &mov ("eax",&wparam(3)); # padbit
1497 &sub ("esp",32*(5+9));
1498 &and ("esp",-512); # ensure that frame
1499 # doesn't cross page
1500 # boundary, which is
1502 # misaligned 32-byte
1505 ################################################################
1506 # expand and copy pre-calculated table to stack
1508 &vmovdqu (&X($D0),&QWP(16*(3+0),"edi"));
1509 &lea ("edx",&DWP(32*5+128,"esp")); # +128 size optimization
1510 &vmovdqu (&X($D1),&QWP(16*(3+1),"edi"));
1511 &vmovdqu (&X($D2),&QWP(16*(3+2),"edi"));
1512 &vmovdqu (&X($D3),&QWP(16*(3+3),"edi"));
1513 &vmovdqu (&X($D4),&QWP(16*(3+4),"edi"));
1514 &lea ("edi",&DWP(16*3,"edi")); # size optimization
1515 &vpermq ($D0,$D0,0b01000000); # 00001234 -> 12343434
1516 &vpermq ($D1,$D1,0b01000000);
1517 &vpermq ($D2,$D2,0b01000000);
1518 &vpermq ($D3,$D3,0b01000000);
1519 &vpermq ($D4,$D4,0b01000000);
1520 &vpshufd ($D0,$D0,0b11001000); # 12343434 -> 14243444
1521 &vpshufd ($D1,$D1,0b11001000);
1522 &vpshufd ($D2,$D2,0b11001000);
1523 &vpshufd ($D3,$D3,0b11001000);
1524 &vpshufd ($D4,$D4,0b11001000);
1525 &vmovdqa (&QWP(32*0-128,"edx"),$D0);
1526 &vmovdqu (&X($D0),&QWP(16*5,"edi"));
1527 &vmovdqa (&QWP(32*1-128,"edx"),$D1);
1528 &vmovdqu (&X($D1),&QWP(16*6,"edi"));
1529 &vmovdqa (&QWP(32*2-128,"edx"),$D2);
1530 &vmovdqu (&X($D2),&QWP(16*7,"edi"));
1531 &vmovdqa (&QWP(32*3-128,"edx"),$D3);
1532 &vmovdqu (&X($D3),&QWP(16*8,"edi"));
1533 &vmovdqa (&QWP(32*4-128,"edx"),$D4);
1534 &vpermq ($D0,$D0,0b01000000);
1535 &vpermq ($D1,$D1,0b01000000);
1536 &vpermq ($D2,$D2,0b01000000);
1537 &vpermq ($D3,$D3,0b01000000);
1538 &vpshufd ($D0,$D0,0b11001000);
1539 &vpshufd ($D1,$D1,0b11001000);
1540 &vpshufd ($D2,$D2,0b11001000);
1541 &vpshufd ($D3,$D3,0b11001000);
1542 &vmovdqa (&QWP(32*5-128,"edx"),$D0);
1543 &vmovd (&X($D0),&DWP(-16*3+4*0,"edi"));# load hash value
1544 &vmovdqa (&QWP(32*6-128,"edx"),$D1);
1545 &vmovd (&X($D1),&DWP(-16*3+4*1,"edi"));
1546 &vmovdqa (&QWP(32*7-128,"edx"),$D2);
1547 &vmovd (&X($D2),&DWP(-16*3+4*2,"edi"));
1548 &vmovdqa (&QWP(32*8-128,"edx"),$D3);
1549 &vmovd (&X($D3),&DWP(-16*3+4*3,"edi"));
1550 &vmovd (&X($D4),&DWP(-16*3+4*4,"edi"));
1551 &vmovdqa ($MASK,&QWP(64,"ebx"));
1552 &neg ("eax"); # padbit
1555 &jz (&label("even"));
1561 &vmovdqu (&X($T0),&QWP(16*0,"esi"));
1563 &jb (&label("one"));
1565 &vmovdqu (&X($T1),&QWP(16*1,"esi"));
1566 &je (&label("two"));
1568 &vinserti128 ($T0,$T0,&QWP(16*2,"esi"),1);
1569 &lea ("esi",&DWP(16*3,"esi"));
1570 &lea ("ebx",&DWP(8,"ebx")); # three padbits
1571 &lea ("edx",&DWP(32*5+128+8,"esp")); # --:r^1:r^2:r^3 (*)
1572 &jmp (&label("tail"));
1575 &lea ("esi",&DWP(16*2,"esi"));
1576 &lea ("ebx",&DWP(16,"ebx")); # two padbits
1577 &lea ("edx",&DWP(32*5+128+16,"esp"));# --:--:r^1:r^2 (*)
1578 &jmp (&label("tail"));
1581 &lea ("esi",&DWP(16*1,"esi"));
1582 &vpxor ($T1,$T1,$T1);
1583 &lea ("ebx",&DWP(32,"ebx","eax",8)); # one or no padbits
1584 &lea ("edx",&DWP(32*5+128+24,"esp"));# --:--:--:r^1 (*)
1585 &jmp (&label("tail"));
1587 # (*) spots marked with '--' are data from next table entry, but they
1588 # are multiplied by 0 and therefore rendered insignificant
1590 &set_label("even",32);
1591 &vmovdqu (&X($T0),&QWP(16*0,"esi")); # load input
1592 &vmovdqu (&X($T1),&QWP(16*1,"esi"));
1593 &vinserti128 ($T0,$T0,&QWP(16*2,"esi"),1);
1594 &vinserti128 ($T1,$T1,&QWP(16*3,"esi"),1);
1595 &lea ("esi",&DWP(16*4,"esi"));
1597 &jz (&label("tail"));
1600 ################################################################
1601 # ((inp[0]*r^4+r[4])*r^4+r[8])*r^4
1602 # ((inp[1]*r^4+r[5])*r^4+r[9])*r^3
1603 # ((inp[2]*r^4+r[6])*r^4+r[10])*r^2
1604 # ((inp[3]*r^4+r[7])*r^4+r[11])*r^1
1605 # \________/ \_______/
1606 ################################################################
1609 &vmovdqa (&QWP(32*2,"esp"),$D2);
1610 &vpsrldq ($D2,$T0,6); # splat input
1611 &vmovdqa (&QWP(32*0,"esp"),$D0);
1612 &vpsrldq ($D0,$T1,6);
1613 &vmovdqa (&QWP(32*1,"esp"),$D1);
1614 &vpunpckhqdq ($D1,$T0,$T1); # 4
1615 &vpunpcklqdq ($T0,$T0,$T1); # 0:1
1616 &vpunpcklqdq ($D2,$D2,$D0); # 2:3
1618 &vpsrlq ($D0,$D2,30);
1619 &vpsrlq ($D2,$D2,4);
1620 &vpsrlq ($T1,$T0,26);
1621 &vpsrlq ($D1,$D1,40); # 4
1622 &vpand ($D2,$D2,$MASK); # 2
1623 &vpand ($T0,$T0,$MASK); # 0
1624 &vpand ($T1,$T1,$MASK); # 1
1625 &vpand ($D0,$D0,$MASK); # 3 (*)
1626 &vpor ($D1,$D1,&QWP(0,"ebx")); # padbit, yes, always
1628 # (*) note that output is counterintuitive, inp[3:4] is
1629 # returned in $D1-2, while $D3-4 are preserved;
1636 &vpaddq ($D2,$D2,&QWP(32*2,"esp")); # add hash value
1637 &vpaddq ($T0,$T0,&QWP(32*0,"esp"));
1638 &vpaddq ($T1,$T1,&QWP(32*1,"esp"));
1639 &vpaddq ($D0,$D0,$D3);
1640 &vpaddq ($D1,$D1,$D4);
1642 ################################################################
1643 # d3 = h2*r1 + h0*r3 + h1*r2 + h3*r0 + h4*5*r4
1644 # d4 = h2*r2 + h0*r4 + h1*r3 + h3*r1 + h4*r0
1645 # d0 = h2*5*r3 + h0*r0 + h1*5*r4 + h3*5*r2 + h4*5*r1
1646 # d1 = h2*5*r4 + h0*r1 + h1*r0 + h3*5*r3 + h4*5*r2
1647 # d2 = h2*r0 + h0*r2 + h1*r1 + h3*5*r4 + h4*5*r3
1649 &vpmuludq ($D3,$D2,&$addr(1)); # d3 = h2*r1
1650 &vmovdqa (QWP(32*1,"esp"),$T1);
1651 &vpmuludq ($D4,$D2,&$addr(2)); # d4 = h2*r2
1652 &vmovdqa (QWP(32*3,"esp"),$D0);
1653 &vpmuludq ($D0,$D2,&$addr(7)); # d0 = h2*s3
1654 &vmovdqa (QWP(32*4,"esp"),$D1);
1655 &vpmuludq ($D1,$D2,&$addr(8)); # d1 = h2*s4
1656 &vpmuludq ($D2,$D2,&$addr(0)); # d2 = h2*r0
1658 &vpmuludq ($T2,$T0,&$addr(3)); # h0*r3
1659 &vpaddq ($D3,$D3,$T2); # d3 += h0*r3
1660 &vpmuludq ($T1,$T0,&$addr(4)); # h0*r4
1661 &vpaddq ($D4,$D4,$T1); # d4 + h0*r4
1662 &vpmuludq ($T2,$T0,&$addr(0)); # h0*r0
1663 &vpaddq ($D0,$D0,$T2); # d0 + h0*r0
1664 &vmovdqa ($T2,&QWP(32*1,"esp")); # h1
1665 &vpmuludq ($T1,$T0,&$addr(1)); # h0*r1
1666 &vpaddq ($D1,$D1,$T1); # d1 += h0*r1
1667 &vpmuludq ($T0,$T0,&$addr(2)); # h0*r2
1668 &vpaddq ($D2,$D2,$T0); # d2 += h0*r2
1670 &vpmuludq ($T1,$T2,&$addr(2)); # h1*r2
1671 &vpaddq ($D3,$D3,$T1); # d3 += h1*r2
1672 &vpmuludq ($T0,$T2,&$addr(3)); # h1*r3
1673 &vpaddq ($D4,$D4,$T0); # d4 += h1*r3
1674 &vpmuludq ($T1,$T2,&$addr(8)); # h1*s4
1675 &vpaddq ($D0,$D0,$T1); # d0 += h1*s4
1676 &vmovdqa ($T1,&QWP(32*3,"esp")); # h3
1677 &vpmuludq ($T0,$T2,&$addr(0)); # h1*r0
1678 &vpaddq ($D1,$D1,$T0); # d1 += h1*r0
1679 &vpmuludq ($T2,$T2,&$addr(1)); # h1*r1
1680 &vpaddq ($D2,$D2,$T2); # d2 += h1*r1
1682 &vpmuludq ($T0,$T1,&$addr(0)); # h3*r0
1683 &vpaddq ($D3,$D3,$T0); # d3 += h3*r0
1684 &vpmuludq ($T2,$T1,&$addr(1)); # h3*r1
1685 &vpaddq ($D4,$D4,$T2); # d4 += h3*r1
1686 &vpmuludq ($T0,$T1,&$addr(6)); # h3*s2
1687 &vpaddq ($D0,$D0,$T0); # d0 += h3*s2
1688 &vmovdqa ($T0,&QWP(32*4,"esp")); # h4
1689 &vpmuludq ($T2,$T1,&$addr(7)); # h3*s3
1690 &vpaddq ($D1,$D1,$T2); # d1+= h3*s3
1691 &vpmuludq ($T1,$T1,&$addr(8)); # h3*s4
1692 &vpaddq ($D2,$D2,$T1); # d2 += h3*s4
1694 &vpmuludq ($T2,$T0,&$addr(8)); # h4*s4
1695 &vpaddq ($D3,$D3,$T2); # d3 += h4*s4
1696 &vpmuludq ($T1,$T0,&$addr(5)); # h4*s1
1697 &vpaddq ($D0,$D0,$T1); # d0 += h4*s1
1698 &vpmuludq ($T2,$T0,&$addr(0)); # h4*r0
1699 &vpaddq ($D4,$D4,$T2); # d4 += h4*r0
1700 &vmovdqa ($MASK,&QWP(64,"ebx"));
1701 &vpmuludq ($T1,$T0,&$addr(6)); # h4*s2
1702 &vpaddq ($D1,$D1,$T1); # d1 += h4*s2
1703 &vpmuludq ($T0,$T0,&$addr(7)); # h4*s3
1704 &vpaddq ($D2,$D2,$T0); # d2 += h4*s3
1706 &vpmuladd (sub { my $i=shift; &QWP(32*$i-128,"edx"); });
1708 sub vlazy_reduction {
1709 ################################################################
1712 &vpsrlq ($T0,$D3,26);
1713 &vpand ($D3,$D3,$MASK);
1714 &vpsrlq ($T1,$D0,26);
1715 &vpand ($D0,$D0,$MASK);
1716 &vpaddq ($D4,$D4,$T0); # h3 -> h4
1717 &vpaddq ($D1,$D1,$T1); # h0 -> h1
1718 &vpsrlq ($T0,$D4,26);
1719 &vpand ($D4,$D4,$MASK);
1720 &vpsrlq ($T1,$D1,26);
1721 &vpand ($D1,$D1,$MASK);
1722 &vpaddq ($D2,$D2,$T1); # h1 -> h2
1723 &vpaddq ($D0,$D0,$T0);
1724 &vpsllq ($T0,$T0,2);
1725 &vpsrlq ($T1,$D2,26);
1726 &vpand ($D2,$D2,$MASK);
1727 &vpaddq ($D0,$D0,$T0); # h4 -> h0
1728 &vpaddq ($D3,$D3,$T1); # h2 -> h3
1729 &vpsrlq ($T1,$D3,26);
1730 &vpsrlq ($T0,$D0,26);
1731 &vpand ($D0,$D0,$MASK);
1732 &vpand ($D3,$D3,$MASK);
1733 &vpaddq ($D1,$D1,$T0); # h0 -> h1
1734 &vpaddq ($D4,$D4,$T1); # h3 -> h4
1738 &vmovdqu (&X($T0),&QWP(16*0,"esi")); # load input
1739 &vmovdqu (&X($T1),&QWP(16*1,"esi"));
1740 &vinserti128 ($T0,$T0,&QWP(16*2,"esi"),1);
1741 &vinserti128 ($T1,$T1,&QWP(16*3,"esi"),1);
1742 &lea ("esi",&DWP(16*4,"esi"));
1744 &jnz (&label("loop"));
1748 &and ("ebx",-64); # restore pointer
1750 &vpmuladd (sub { my $i=shift; &QWP(4+32*$i-128,"edx"); });
1752 ################################################################
1753 # horizontal addition
1755 &vpsrldq ($T0,$D4,8);
1756 &vpsrldq ($T1,$D3,8);
1757 &vpaddq ($D4,$D4,$T0);
1758 &vpsrldq ($T0,$D0,8);
1759 &vpaddq ($D3,$D3,$T1);
1760 &vpsrldq ($T1,$D1,8);
1761 &vpaddq ($D0,$D0,$T0);
1762 &vpsrldq ($T0,$D2,8);
1763 &vpaddq ($D1,$D1,$T1);
1764 &vpermq ($T1,$D4,2); # keep folding
1765 &vpaddq ($D2,$D2,$T0);
1766 &vpermq ($T0,$D3,2);
1767 &vpaddq ($D4,$D4,$T1);
1768 &vpermq ($T1,$D0,2);
1769 &vpaddq ($D3,$D3,$T0);
1770 &vpermq ($T0,$D1,2);
1771 &vpaddq ($D0,$D0,$T1);
1772 &vpermq ($T1,$D2,2);
1773 &vpaddq ($D1,$D1,$T0);
1774 &vpaddq ($D2,$D2,$T1);
1779 &je (&label("done"));
1781 ################################################################
1782 # clear all but single word
1784 &vpshufd (&X($D0),&X($D0),0b11111100);
1785 &lea ("edx",&DWP(32*5+128,"esp")); # restore pointer
1786 &vpshufd (&X($D1),&X($D1),0b11111100);
1787 &vpshufd (&X($D2),&X($D2),0b11111100);
1788 &vpshufd (&X($D3),&X($D3),0b11111100);
1789 &vpshufd (&X($D4),&X($D4),0b11111100);
1790 &jmp (&label("even"));
1792 &set_label("done",16);
1793 &vmovd (&DWP(-16*3+4*0,"edi"),&X($D0));# store hash value
1794 &vmovd (&DWP(-16*3+4*1,"edi"),&X($D1));
1795 &vmovd (&DWP(-16*3+4*2,"edi"),&X($D2));
1796 &vmovd (&DWP(-16*3+4*3,"edi"),&X($D3));
1797 &vmovd (&DWP(-16*3+4*4,"edi"),&X($D4));
1800 &set_label("nodata");
1801 &function_end("_poly1305_blocks_avx2");
1803 &set_label("const_sse2",64);
1804 &data_word(1<<24,0, 1<<24,0, 1<<24,0, 1<<24,0);
1805 &data_word(0,0, 0,0, 0,0, 0,0);
1806 &data_word(0x03ffffff,0,0x03ffffff,0, 0x03ffffff,0, 0x03ffffff,0);
1807 &data_word(0x0fffffff,0x0ffffffc,0x0ffffffc,0x0ffffffc);
1809 &asciz ("Poly1305 for x86, CRYPTOGAMS by <appro\@openssl.org>");
1814 close STDOUT or die "error closing STDOUT";