Debian Codegraph @

external/boringssl/crypto/modes/asm/ghash-x86_64.pl

raw
1 #!/usr/bin/env perl
2 #
3 # ====================================================================
4 # Written by Andy Polyakov <[email protected]> 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 # ====================================================================
9 #
10 # March, June 2010
11 #
12 # The module implements "4-bit" GCM GHASH function and underlying
13 # single multiplication operation in GF(2^128). "4-bit" means that
14 # it uses 256 bytes per-key table [+128 bytes shared table]. GHASH
15 # function features so called "528B" variant utilizing additional
16 # 256+16 bytes of per-key storage [+512 bytes shared table].
17 # Performance results are for this streamed GHASH subroutine and are
18 # expressed in cycles per processed byte, less is better:
19 #
20 # gcc 3.4.x(*) assembler
21 #
22 # P4 28.6 14.0 +100%
23 # Opteron 19.3 7.7 +150%
24 # Core2 17.8 8.1(**) +120%
25 # Atom 31.6 16.8 +88%
26 # VIA Nano 21.8 10.1 +115%
27 #
28 # (*) comparison is not completely fair, because C results are
29 # for vanilla "256B" implementation, while assembler results
30 # are for "528B";-)
31 # (**) it's mystery [to me] why Core2 result is not same as for
32 # Opteron;
33
34 # May 2010
35 #
36 # Add PCLMULQDQ version performing at 2.02 cycles per processed byte.
37 # See ghash-x86.pl for background information and details about coding
38 # techniques.
39 #
40 # Special thanks to David Woodhouse <[email protected]> for
41 # providing access to a Westmere-based system on behalf of Intel
42 # Open Source Technology Centre.
43
44 # December 2012
45 #
46 # Overhaul: aggregate Karatsuba post-processing, improve ILP in
47 # reduction_alg9, increase reduction aggregate factor to 4x. As for
48 # the latter. ghash-x86.pl discusses that it makes lesser sense to
49 # increase aggregate factor. Then why increase here? Critical path
50 # consists of 3 independent pclmulqdq instructions, Karatsuba post-
51 # processing and reduction. "On top" of this we lay down aggregated
52 # multiplication operations, triplets of independent pclmulqdq's. As
53 # issue rate for pclmulqdq is limited, it makes lesser sense to
54 # aggregate more multiplications than it takes to perform remaining
55 # non-multiplication operations. 2x is near-optimal coefficient for
56 # contemporary Intel CPUs (therefore modest improvement coefficient),
57 # but not for Bulldozer. Latter is because logical SIMD operations
58 # are twice as slow in comparison to Intel, so that critical path is
59 # longer. A CPU with higher pclmulqdq issue rate would also benefit
60 # from higher aggregate factor...
61 #
62 # Westmere 1.78(+13%)
63 # Sandy Bridge 1.80(+8%)
64 # Ivy Bridge 1.80(+7%)
65 # Haswell 0.55(+93%) (if system doesn't support AVX)
66 # Broadwell 0.45(+110%)(if system doesn't support AVX)
67 # Bulldozer 1.49(+27%)
68 # Silvermont 2.88(+13%)
69
70 # March 2013
71 #
72 # ... 8x aggregate factor AVX code path is using reduction algorithm
73 # suggested by Shay Gueron[1]. Even though contemporary AVX-capable
74 # CPUs such as Sandy and Ivy Bridge can execute it, the code performs
75 # sub-optimally in comparison to above mentioned version. But thanks
76 # to Ilya Albrekht and Max Locktyukhin of Intel Corp. we knew that
77 # it performs in 0.41 cycles per byte on Haswell processor, and in
78 # 0.29 on Broadwell.
79 #
80 # [1] http://rt.openssl.org/Ticket/Display.html?id=2900&user=guest&pass=guest
81
82 $flavour = shift;
83 $output = shift;
84 if ($flavour =~ /\./) { $output = $flavour; undef $flavour; }
85
86 $win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/);
87
88 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
89 ( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or
90 ( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or
91 die "can't locate x86_64-xlate.pl";
92
93 # This must be kept in sync with |$avx| in aesni-gcm-x86_64.pl; otherwise tags
94 # will be computed incorrectly.
95 #
96 # In upstream, this is controlled by shelling out to the compiler to check
97 # versions, but BoringSSL is intended to be used with pre-generated perlasm
98 # output, so this isn't useful anyway.
99 $avx = 0;
100
101 open OUT,"| \"$^X\" $xlate $flavour $output";
102 *STDOUT=*OUT;
103
104 $do4xaggr=1;
105
106 # common register layout
107 $nlo="%rax";
108 $nhi="%rbx";
109 $Zlo="%r8";
110 $Zhi="%r9";
111 $tmp="%r10";
112 $rem_4bit = "%r11";
113
114 $Xi="%rdi";
115 $Htbl="%rsi";
116
117 # per-function register layout
118 $cnt="%rcx";
119 $rem="%rdx";
120
121 sub LB() { my $r=shift; $r =~ s/%[er]([a-d])x/%\1l/ or
122 $r =~ s/%[er]([sd]i)/%\1l/ or
123 $r =~ s/%[er](bp)/%\1l/ or
124 $r =~ s/%(r[0-9]+)[d]?/%\1b/; $r; }
125
126 sub AUTOLOAD() # thunk [simplified] 32-bit style perlasm
127 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://;
128 my $arg = pop;
129 $arg = "\$$arg" if ($arg*1 eq $arg);
130 $code .= "\t$opcode\t".join(',',$arg,reverse @_)."\n";
131 }
132
133 { my $N;
134 sub loop() {
135 my $inp = shift;
136
137 $N++;
138 $code.=<<___;
139 xor $nlo,$nlo
140 xor $nhi,$nhi
141 mov `&LB("$Zlo")`,`&LB("$nlo")`
142 mov `&LB("$Zlo")`,`&LB("$nhi")`
143 shl \$4,`&LB("$nlo")`
144 mov \$14,$cnt
145 mov 8($Htbl,$nlo),$Zlo
146 mov ($Htbl,$nlo),$Zhi
147 and \$0xf0,`&LB("$nhi")`
148 mov $Zlo,$rem
149 jmp .Loop$N
150
151 .align 16
152 .Loop$N:
153 shr \$4,$Zlo
154 and \$0xf,$rem
155 mov $Zhi,$tmp
156 mov ($inp,$cnt),`&LB("$nlo")`
157 shr \$4,$Zhi
158 xor 8($Htbl,$nhi),$Zlo
159 shl \$60,$tmp
160 xor ($Htbl,$nhi),$Zhi
161 mov `&LB("$nlo")`,`&LB("$nhi")`
162 xor ($rem_4bit,$rem,8),$Zhi
163 mov $Zlo,$rem
164 shl \$4,`&LB("$nlo")`
165 xor $tmp,$Zlo
166 dec $cnt
167 js .Lbreak$N
168
169 shr \$4,$Zlo
170 and \$0xf,$rem
171 mov $Zhi,$tmp
172 shr \$4,$Zhi
173 xor 8($Htbl,$nlo),$Zlo
174 shl \$60,$tmp
175 xor ($Htbl,$nlo),$Zhi
176 and \$0xf0,`&LB("$nhi")`
177 xor ($rem_4bit,$rem,8),$Zhi
178 mov $Zlo,$rem
179 xor $tmp,$Zlo
180 jmp .Loop$N
181
182 .align 16
183 .Lbreak$N:
184 shr \$4,$Zlo
185 and \$0xf,$rem
186 mov $Zhi,$tmp
187 shr \$4,$Zhi
188 xor 8($Htbl,$nlo),$Zlo
189 shl \$60,$tmp
190 xor ($Htbl,$nlo),$Zhi
191 and \$0xf0,`&LB("$nhi")`
192 xor ($rem_4bit,$rem,8),$Zhi
193 mov $Zlo,$rem
194 xor $tmp,$Zlo
195
196 shr \$4,$Zlo
197 and \$0xf,$rem
198 mov $Zhi,$tmp
199 shr \$4,$Zhi
200 xor 8($Htbl,$nhi),$Zlo
201 shl \$60,$tmp
202 xor ($Htbl,$nhi),$Zhi
203 xor $tmp,$Zlo
204 xor ($rem_4bit,$rem,8),$Zhi
205
206 bswap $Zlo
207 bswap $Zhi
208 ___
209 }}
210
211 $code=<<___;
212 .text
213 .extern OPENSSL_ia32cap_P
214
215 .globl gcm_gmult_4bit
216 .type gcm_gmult_4bit,\@function,2
217 .align 16
218 gcm_gmult_4bit:
219 push %rbx
220 push %rbp # %rbp and %r12 are pushed exclusively in
221 push %r12 # order to reuse Win64 exception handler...
222 .Lgmult_prologue:
223
224 movzb 15($Xi),$Zlo
225 lea .Lrem_4bit(%rip),$rem_4bit
226 ___
227 &loop ($Xi);
228 $code.=<<___;
229 mov $Zlo,8($Xi)
230 mov $Zhi,($Xi)
231
232 mov 16(%rsp),%rbx
233 lea 24(%rsp),%rsp
234 .Lgmult_epilogue:
235 ret
236 .size gcm_gmult_4bit,.-gcm_gmult_4bit
237 ___
238
239 # per-function register layout
240 $inp="%rdx";
241 $len="%rcx";
242 $rem_8bit=$rem_4bit;
243
244 $code.=<<___;
245 .globl gcm_ghash_4bit
246 .type gcm_ghash_4bit,\@function,4
247 .align 16
248 gcm_ghash_4bit:
249 push %rbx
250 push %rbp
251 push %r12
252 push %r13
253 push %r14
254 push %r15
255 sub \$280,%rsp
256 .Lghash_prologue:
257 mov $inp,%r14 # reassign couple of args
258 mov $len,%r15
259 ___
260 { my $inp="%r14";
261 my $dat="%edx";
262 my $len="%r15";
263 my @nhi=("%ebx","%ecx");
264 my @rem=("%r12","%r13");
265 my $Hshr4="%rbp";
266
267 &sub ($Htbl,-128); # size optimization
268 &lea ($Hshr4,"16+128(%rsp)");
269 { my @lo =($nlo,$nhi);
270 my @hi =($Zlo,$Zhi);
271
272 &xor ($dat,$dat);
273 for ($i=0,$j=-2;$i<18;$i++,$j++) {
274 &mov ("$j(%rsp)",&LB($dat)) if ($i>1);
275 &or ($lo[0],$tmp) if ($i>1);
276 &mov (&LB($dat),&LB($lo[1])) if ($i>0 && $i<17);
277 &shr ($lo[1],4) if ($i>0 && $i<17);
278 &mov ($tmp,$hi[1]) if ($i>0 && $i<17);
279 &shr ($hi[1],4) if ($i>0 && $i<17);
280 &mov ("8*$j($Hshr4)",$hi[0]) if ($i>1);
281 &mov ($hi[0],"16*$i+0-128($Htbl)") if ($i<16);
282 &shl (&LB($dat),4) if ($i>0 && $i<17);
283 &mov ("8*$j-128($Hshr4)",$lo[0]) if ($i>1);
284 &mov ($lo[0],"16*$i+8-128($Htbl)") if ($i<16);
285 &shl ($tmp,60) if ($i>0 && $i<17);
286
287 push (@lo,shift(@lo));
288 push (@hi,shift(@hi));
289 }
290 }
291 &add ($Htbl,-128);
292 &mov ($Zlo,"8($Xi)");
293 &mov ($Zhi,"0($Xi)");
294 &add ($len,$inp); # pointer to the end of data
295 &lea ($rem_8bit,".Lrem_8bit(%rip)");
296 &jmp (".Louter_loop");
297
298 $code.=".align 16\n.Louter_loop:\n";
299 &xor ($Zhi,"($inp)");
300 &mov ("%rdx","8($inp)");
301 &lea ($inp,"16($inp)");
302 &xor ("%rdx",$Zlo);
303 &mov ("($Xi)",$Zhi);
304 &mov ("8($Xi)","%rdx");
305 &shr ("%rdx",32);
306
307 &xor ($nlo,$nlo);
308 &rol ($dat,8);
309 &mov (&LB($nlo),&LB($dat));
310 &movz ($nhi[0],&LB($dat));
311 &shl (&LB($nlo),4);
312 &shr ($nhi[0],4);
313
314 for ($j=11,$i=0;$i<15;$i++) {
315 &rol ($dat,8);
316 &xor ($Zlo,"8($Htbl,$nlo)") if ($i>0);
317 &xor ($Zhi,"($Htbl,$nlo)") if ($i>0);
318 &mov ($Zlo,"8($Htbl,$nlo)") if ($i==0);
319 &mov ($Zhi,"($Htbl,$nlo)") if ($i==0);
320
321 &mov (&LB($nlo),&LB($dat));
322 &xor ($Zlo,$tmp) if ($i>0);
323 &movzw ($rem[1],"($rem_8bit,$rem[1],2)") if ($i>0);
324
325 &movz ($nhi[1],&LB($dat));
326 &shl (&LB($nlo),4);
327 &movzb ($rem[0],"(%rsp,$nhi[0])");
328
329 &shr ($nhi[1],4) if ($i<14);
330 &and ($nhi[1],0xf0) if ($i==14);
331 &shl ($rem[1],48) if ($i>0);
332 &xor ($rem[0],$Zlo);
333
334 &mov ($tmp,$Zhi);
335 &xor ($Zhi,$rem[1]) if ($i>0);
336 &shr ($Zlo,8);
337
338 &movz ($rem[0],&LB($rem[0]));
339 &mov ($dat,"$j($Xi)") if (--$j%4==0);
340 &shr ($Zhi,8);
341
342 &xor ($Zlo,"-128($Hshr4,$nhi[0],8)");
343 &shl ($tmp,56);
344 &xor ($Zhi,"($Hshr4,$nhi[0],8)");
345
346 unshift (@nhi,pop(@nhi)); # "rotate" registers
347 unshift (@rem,pop(@rem));
348 }
349 &movzw ($rem[1],"($rem_8bit,$rem[1],2)");
350 &xor ($Zlo,"8($Htbl,$nlo)");
351 &xor ($Zhi,"($Htbl,$nlo)");
352
353 &shl ($rem[1],48);
354 &xor ($Zlo,$tmp);
355
356 &xor ($Zhi,$rem[1]);
357 &movz ($rem[0],&LB($Zlo));
358 &shr ($Zlo,4);
359
360 &mov ($tmp,$Zhi);
361 &shl (&LB($rem[0]),4);
362 &shr ($Zhi,4);
363
364 &xor ($Zlo,"8($Htbl,$nhi[0])");
365 &movzw ($rem[0],"($rem_8bit,$rem[0],2)");
366 &shl ($tmp,60);
367
368 &xor ($Zhi,"($Htbl,$nhi[0])");
369 &xor ($Zlo,$tmp);
370 &shl ($rem[0],48);
371
372 &bswap ($Zlo);
373 &xor ($Zhi,$rem[0]);
374
375 &bswap ($Zhi);
376 &cmp ($inp,$len);
377 &jb (".Louter_loop");
378 }
379 $code.=<<___;
380 mov $Zlo,8($Xi)
381 mov $Zhi,($Xi)
382
383 lea 280(%rsp),%rsi
384 mov 0(%rsi),%r15
385 mov 8(%rsi),%r14
386 mov 16(%rsi),%r13
387 mov 24(%rsi),%r12
388 mov 32(%rsi),%rbp
389 mov 40(%rsi),%rbx
390 lea 48(%rsi),%rsp
391 .Lghash_epilogue:
392 ret
393 .size gcm_ghash_4bit,.-gcm_ghash_4bit
394 ___
395
396 ######################################################################
397 # PCLMULQDQ version.
398
399 @_4args=$win64? ("%rcx","%rdx","%r8", "%r9") : # Win64 order
400 ("%rdi","%rsi","%rdx","%rcx"); # Unix order
401
402 ($Xi,$Xhi)=("%xmm0","%xmm1"); $Hkey="%xmm2";
403 ($T1,$T2,$T3)=("%xmm3","%xmm4","%xmm5");
404
405 sub clmul64x64_T2 { # minimal register pressure
406 my ($Xhi,$Xi,$Hkey,$HK)=@_;
407
408 if (!defined($HK)) { $HK = $T2;
409 $code.=<<___;
410 movdqa $Xi,$Xhi #
411 pshufd \$0b01001110,$Xi,$T1
412 pshufd \$0b01001110,$Hkey,$T2
413 pxor $Xi,$T1 #
414 pxor $Hkey,$T2
415 ___
416 } else {
417 $code.=<<___;
418 movdqa $Xi,$Xhi #
419 pshufd \$0b01001110,$Xi,$T1
420 pxor $Xi,$T1 #
421 ___
422 }
423 $code.=<<___;
424 pclmulqdq \$0x00,$Hkey,$Xi #######
425 pclmulqdq \$0x11,$Hkey,$Xhi #######
426 pclmulqdq \$0x00,$HK,$T1 #######
427 pxor $Xi,$T1 #
428 pxor $Xhi,$T1 #
429
430 movdqa $T1,$T2 #
431 psrldq \$8,$T1
432 pslldq \$8,$T2 #
433 pxor $T1,$Xhi
434 pxor $T2,$Xi #
435 ___
436 }
437
438 sub reduction_alg9 { # 17/11 times faster than Intel version
439 my ($Xhi,$Xi) = @_;
440
441 $code.=<<___;
442 # 1st phase
443 movdqa $Xi,$T2 #
444 movdqa $Xi,$T1
445 psllq \$5,$Xi
446 pxor $Xi,$T1 #
447 psllq \$1,$Xi
448 pxor $T1,$Xi #
449 psllq \$57,$Xi #
450 movdqa $Xi,$T1 #
451 pslldq \$8,$Xi
452 psrldq \$8,$T1 #
453 pxor $T2,$Xi
454 pxor $T1,$Xhi #
455
456 # 2nd phase
457 movdqa $Xi,$T2
458 psrlq \$1,$Xi
459 pxor $T2,$Xhi #
460 pxor $Xi,$T2
461 psrlq \$5,$Xi
462 pxor $T2,$Xi #
463 psrlq \$1,$Xi #
464 pxor $Xhi,$Xi #
465 ___
466 }
467
468 { my ($Htbl,$Xip)=@_4args;
469 my $HK="%xmm6";
470
471 $code.=<<___;
472 .globl gcm_init_clmul
473 .type gcm_init_clmul,\@abi-omnipotent
474 .align 16
475 gcm_init_clmul:
476 .L_init_clmul:
477 ___
478 $code.=<<___ if ($win64);
479 .LSEH_begin_gcm_init_clmul:
480 # I can't trust assembler to use specific encoding:-(
481 .byte 0x48,0x83,0xec,0x18 #sub $0x18,%rsp
482 .byte 0x0f,0x29,0x34,0x24 #movaps %xmm6,(%rsp)
483 ___
484 $code.=<<___;
485 movdqu ($Xip),$Hkey
486 pshufd \$0b01001110,$Hkey,$Hkey # dword swap
487
488 # <<1 twist
489 pshufd \$0b11111111,$Hkey,$T2 # broadcast uppermost dword
490 movdqa $Hkey,$T1
491 psllq \$1,$Hkey
492 pxor $T3,$T3 #
493 psrlq \$63,$T1
494 pcmpgtd $T2,$T3 # broadcast carry bit
495 pslldq \$8,$T1
496 por $T1,$Hkey # H<<=1
497
498 # magic reduction
499 pand .L0x1c2_polynomial(%rip),$T3
500 pxor $T3,$Hkey # if(carry) H^=0x1c2_polynomial
501
502 # calculate H^2
503 pshufd \$0b01001110,$Hkey,$HK
504 movdqa $Hkey,$Xi
505 pxor $Hkey,$HK
506 ___
507 &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$HK);
508 &reduction_alg9 ($Xhi,$Xi);
509 $code.=<<___;
510 pshufd \$0b01001110,$Hkey,$T1
511 pshufd \$0b01001110,$Xi,$T2
512 pxor $Hkey,$T1 # Karatsuba pre-processing
513 movdqu $Hkey,0x00($Htbl) # save H
514 pxor $Xi,$T2 # Karatsuba pre-processing
515 movdqu $Xi,0x10($Htbl) # save H^2
516 palignr \$8,$T1,$T2 # low part is H.lo^H.hi...
517 movdqu $T2,0x20($Htbl) # save Karatsuba "salt"
518 ___
519 if ($do4xaggr) {
520 &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$HK); # H^3
521 &reduction_alg9 ($Xhi,$Xi);
522 $code.=<<___;
523 movdqa $Xi,$T3
524 ___
525 &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$HK); # H^4
526 &reduction_alg9 ($Xhi,$Xi);
527 $code.=<<___;
528 pshufd \$0b01001110,$T3,$T1
529 pshufd \$0b01001110,$Xi,$T2
530 pxor $T3,$T1 # Karatsuba pre-processing
531 movdqu $T3,0x30($Htbl) # save H^3
532 pxor $Xi,$T2 # Karatsuba pre-processing
533 movdqu $Xi,0x40($Htbl) # save H^4
534 palignr \$8,$T1,$T2 # low part is H^3.lo^H^3.hi...
535 movdqu $T2,0x50($Htbl) # save Karatsuba "salt"
536 ___
537 }
538 $code.=<<___ if ($win64);
539 movaps (%rsp),%xmm6
540 lea 0x18(%rsp),%rsp
541 .LSEH_end_gcm_init_clmul:
542 ___
543 $code.=<<___;
544 ret
545 .size gcm_init_clmul,.-gcm_init_clmul
546 ___
547 }
548
549 { my ($Xip,$Htbl)=@_4args;
550
551 $code.=<<___;
552 .globl gcm_gmult_clmul
553 .type gcm_gmult_clmul,\@abi-omnipotent
554 .align 16
555 gcm_gmult_clmul:
556 .L_gmult_clmul:
557 movdqu ($Xip),$Xi
558 movdqa .Lbswap_mask(%rip),$T3
559 movdqu ($Htbl),$Hkey
560 movdqu 0x20($Htbl),$T2
561 pshufb $T3,$Xi
562 ___
563 &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$T2);
564 $code.=<<___ if (0 || (&reduction_alg9($Xhi,$Xi)&&0));
565 # experimental alternative. special thing about is that there
566 # no dependency between the two multiplications...
567 mov \$`0xE1<<1`,%eax
568 mov \$0xA040608020C0E000,%r10 # ((7..0)·0xE0)&0xff
569 mov \$0x07,%r11d
570 movq %rax,$T1
571 movq %r10,$T2
572 movq %r11,$T3 # borrow $T3
573 pand $Xi,$T3
574 pshufb $T3,$T2 # ($Xi&7)·0xE0
575 movq %rax,$T3
576 pclmulqdq \$0x00,$Xi,$T1 # ·(0xE1<<1)
577 pxor $Xi,$T2
578 pslldq \$15,$T2
579 paddd $T2,$T2 # <<(64+56+1)
580 pxor $T2,$Xi
581 pclmulqdq \$0x01,$T3,$Xi
582 movdqa .Lbswap_mask(%rip),$T3 # reload $T3
583 psrldq \$1,$T1
584 pxor $T1,$Xhi
585 pslldq \$7,$Xi
586 pxor $Xhi,$Xi
587 ___
588 $code.=<<___;
589 pshufb $T3,$Xi
590 movdqu $Xi,($Xip)
591 ret
592 .size gcm_gmult_clmul,.-gcm_gmult_clmul
593 ___
594 }
595
596 { my ($Xip,$Htbl,$inp,$len)=@_4args;
597 my ($Xln,$Xmn,$Xhn,$Hkey2,$HK) = map("%xmm$_",(3..7));
598 my ($T1,$T2,$T3)=map("%xmm$_",(8..10));
599
600 $code.=<<___;
601 .globl gcm_ghash_clmul
602 .type gcm_ghash_clmul,\@abi-omnipotent
603 .align 32
604 gcm_ghash_clmul:
605 .L_ghash_clmul:
606 ___
607 $code.=<<___ if ($win64);
608 lea -0x88(%rsp),%rax
609 .LSEH_begin_gcm_ghash_clmul:
610 # I can't trust assembler to use specific encoding:-(
611 .byte 0x48,0x8d,0x60,0xe0 #lea -0x20(%rax),%rsp
612 .byte 0x0f,0x29,0x70,0xe0 #movaps %xmm6,-0x20(%rax)
613 .byte 0x0f,0x29,0x78,0xf0 #movaps %xmm7,-0x10(%rax)
614 .byte 0x44,0x0f,0x29,0x00 #movaps %xmm8,0(%rax)
615 .byte 0x44,0x0f,0x29,0x48,0x10 #movaps %xmm9,0x10(%rax)
616 .byte 0x44,0x0f,0x29,0x50,0x20 #movaps %xmm10,0x20(%rax)
617 .byte 0x44,0x0f,0x29,0x58,0x30 #movaps %xmm11,0x30(%rax)
618 .byte 0x44,0x0f,0x29,0x60,0x40 #movaps %xmm12,0x40(%rax)
619 .byte 0x44,0x0f,0x29,0x68,0x50 #movaps %xmm13,0x50(%rax)
620 .byte 0x44,0x0f,0x29,0x70,0x60 #movaps %xmm14,0x60(%rax)
621 .byte 0x44,0x0f,0x29,0x78,0x70 #movaps %xmm15,0x70(%rax)
622 ___
623 $code.=<<___;
624 movdqa .Lbswap_mask(%rip),$T3
625
626 movdqu ($Xip),$Xi
627 movdqu ($Htbl),$Hkey
628 movdqu 0x20($Htbl),$HK
629 pshufb $T3,$Xi
630
631 sub \$0x10,$len
632 jz .Lodd_tail
633
634 movdqu 0x10($Htbl),$Hkey2
635 ___
636 if ($do4xaggr) {
637 my ($Xl,$Xm,$Xh,$Hkey3,$Hkey4)=map("%xmm$_",(11..15));
638
639 $code.=<<___;
640 mov OPENSSL_ia32cap_P+4(%rip),%eax
641 cmp \$0x30,$len
642 jb .Lskip4x
643
644 and \$`1<<26|1<<22`,%eax # isolate MOVBE+XSAVE
645 cmp \$`1<<22`,%eax # check for MOVBE without XSAVE
646 je .Lskip4x
647
648 sub \$0x30,$len
649 mov \$0xA040608020C0E000,%rax # ((7..0)·0xE0)&0xff
650 movdqu 0x30($Htbl),$Hkey3
651 movdqu 0x40($Htbl),$Hkey4
652
653 #######
654 # Xi+4 =[(H*Ii+3) + (H^2*Ii+2) + (H^3*Ii+1) + H^4*(Ii+Xi)] mod P
655 #
656 movdqu 0x30($inp),$Xln
657 movdqu 0x20($inp),$Xl
658 pshufb $T3,$Xln
659 pshufb $T3,$Xl
660 movdqa $Xln,$Xhn
661 pshufd \$0b01001110,$Xln,$Xmn
662 pxor $Xln,$Xmn
663 pclmulqdq \$0x00,$Hkey,$Xln
664 pclmulqdq \$0x11,$Hkey,$Xhn
665 pclmulqdq \$0x00,$HK,$Xmn
666
667 movdqa $Xl,$Xh
668 pshufd \$0b01001110,$Xl,$Xm
669 pxor $Xl,$Xm
670 pclmulqdq \$0x00,$Hkey2,$Xl
671 pclmulqdq \$0x11,$Hkey2,$Xh
672 pclmulqdq \$0x10,$HK,$Xm
673 xorps $Xl,$Xln
674 xorps $Xh,$Xhn
675 movups 0x50($Htbl),$HK
676 xorps $Xm,$Xmn
677
678 movdqu 0x10($inp),$Xl
679 movdqu 0($inp),$T1
680 pshufb $T3,$Xl
681 pshufb $T3,$T1
682 movdqa $Xl,$Xh
683 pshufd \$0b01001110,$Xl,$Xm
684 pxor $T1,$Xi
685 pxor $Xl,$Xm
686 pclmulqdq \$0x00,$Hkey3,$Xl
687 movdqa $Xi,$Xhi
688 pshufd \$0b01001110,$Xi,$T1
689 pxor $Xi,$T1
690 pclmulqdq \$0x11,$Hkey3,$Xh
691 pclmulqdq \$0x00,$HK,$Xm
692 xorps $Xl,$Xln
693 xorps $Xh,$Xhn
694
695 lea 0x40($inp),$inp
696 sub \$0x40,$len
697 jc .Ltail4x
698
699 jmp .Lmod4_loop
700 .align 32
701 .Lmod4_loop:
702 pclmulqdq \$0x00,$Hkey4,$Xi
703 xorps $Xm,$Xmn
704 movdqu 0x30($inp),$Xl
705 pshufb $T3,$Xl
706 pclmulqdq \$0x11,$Hkey4,$Xhi
707 xorps $Xln,$Xi
708 movdqu 0x20($inp),$Xln
709 movdqa $Xl,$Xh
710 pclmulqdq \$0x10,$HK,$T1
711 pshufd \$0b01001110,$Xl,$Xm
712 xorps $Xhn,$Xhi
713 pxor $Xl,$Xm
714 pshufb $T3,$Xln
715 movups 0x20($Htbl),$HK
716 xorps $Xmn,$T1
717 pclmulqdq \$0x00,$Hkey,$Xl
718 pshufd \$0b01001110,$Xln,$Xmn
719
720 pxor $Xi,$T1 # aggregated Karatsuba post-processing
721 movdqa $Xln,$Xhn
722 pxor $Xhi,$T1 #
723 pxor $Xln,$Xmn
724 movdqa $T1,$T2 #
725 pclmulqdq \$0x11,$Hkey,$Xh
726 pslldq \$8,$T1
727 psrldq \$8,$T2 #
728 pxor $T1,$Xi
729 movdqa .L7_mask(%rip),$T1
730 pxor $T2,$Xhi #
731 movq %rax,$T2
732
733 pand $Xi,$T1 # 1st phase
734 pshufb $T1,$T2 #
735 pxor $Xi,$T2 #
736 pclmulqdq \$0x00,$HK,$Xm
737 psllq \$57,$T2 #
738 movdqa $T2,$T1 #
739 pslldq \$8,$T2
740 pclmulqdq \$0x00,$Hkey2,$Xln
741 psrldq \$8,$T1 #
742 pxor $T2,$Xi
743 pxor $T1,$Xhi #
744 movdqu 0($inp),$T1
745
746 movdqa $Xi,$T2 # 2nd phase
747 psrlq \$1,$Xi
748 pclmulqdq \$0x11,$Hkey2,$Xhn
749 xorps $Xl,$Xln
750 movdqu 0x10($inp),$Xl
751 pshufb $T3,$Xl
752 pclmulqdq \$0x10,$HK,$Xmn
753 xorps $Xh,$Xhn
754 movups 0x50($Htbl),$HK
755 pshufb $T3,$T1
756 pxor $T2,$Xhi #
757 pxor $Xi,$T2
758 psrlq \$5,$Xi
759
760 movdqa $Xl,$Xh
761 pxor $Xm,$Xmn
762 pshufd \$0b01001110,$Xl,$Xm
763 pxor $T2,$Xi #
764 pxor $T1,$Xhi
765 pxor $Xl,$Xm
766 pclmulqdq \$0x00,$Hkey3,$Xl
767 psrlq \$1,$Xi #
768 pxor $Xhi,$Xi #
769 movdqa $Xi,$Xhi
770 pclmulqdq \$0x11,$Hkey3,$Xh
771 xorps $Xl,$Xln
772 pshufd \$0b01001110,$Xi,$T1
773 pxor $Xi,$T1
774
775 pclmulqdq \$0x00,$HK,$Xm
776 xorps $Xh,$Xhn
777
778 lea 0x40($inp),$inp
779 sub \$0x40,$len
780 jnc .Lmod4_loop
781
782 .Ltail4x:
783 pclmulqdq \$0x00,$Hkey4,$Xi
784 pclmulqdq \$0x11,$Hkey4,$Xhi
785 pclmulqdq \$0x10,$HK,$T1
786 xorps $Xm,$Xmn
787 xorps $Xln,$Xi
788 xorps $Xhn,$Xhi
789 pxor $Xi,$Xhi # aggregated Karatsuba post-processing
790 pxor $Xmn,$T1
791
792 pxor $Xhi,$T1 #
793 pxor $Xi,$Xhi
794
795 movdqa $T1,$T2 #
796 psrldq \$8,$T1
797 pslldq \$8,$T2 #
798 pxor $T1,$Xhi
799 pxor $T2,$Xi #
800 ___
801 &reduction_alg9($Xhi,$Xi);
802 $code.=<<___;
803 add \$0x40,$len
804 jz .Ldone
805 movdqu 0x20($Htbl),$HK
806 sub \$0x10,$len
807 jz .Lodd_tail
808 .Lskip4x:
809 ___
810 }
811 $code.=<<___;
812 #######
813 # Xi+2 =[H*(Ii+1 + Xi+1)] mod P =
814 # [(H*Ii+1) + (H*Xi+1)] mod P =
815 # [(H*Ii+1) + H^2*(Ii+Xi)] mod P
816 #
817 movdqu ($inp),$T1 # Ii
818 movdqu 16($inp),$Xln # Ii+1
819 pshufb $T3,$T1
820 pshufb $T3,$Xln
821 pxor $T1,$Xi # Ii+Xi
822
823 movdqa $Xln,$Xhn
824 pshufd \$0b01001110,$Xln,$Xmn
825 pxor $Xln,$Xmn
826 pclmulqdq \$0x00,$Hkey,$Xln
827 pclmulqdq \$0x11,$Hkey,$Xhn
828 pclmulqdq \$0x00,$HK,$Xmn
829
830 lea 32($inp),$inp # i+=2
831 nop
832 sub \$0x20,$len
833 jbe .Leven_tail
834 nop
835 jmp .Lmod_loop
836
837 .align 32
838 .Lmod_loop:
839 movdqa $Xi,$Xhi
840 movdqa $Xmn,$T1
841 pshufd \$0b01001110,$Xi,$Xmn #
842 pxor $Xi,$Xmn #
843
844 pclmulqdq \$0x00,$Hkey2,$Xi
845 pclmulqdq \$0x11,$Hkey2,$Xhi
846 pclmulqdq \$0x10,$HK,$Xmn
847
848 pxor $Xln,$Xi # (H*Ii+1) + H^2*(Ii+Xi)
849 pxor $Xhn,$Xhi
850 movdqu ($inp),$T2 # Ii
851 pxor $Xi,$T1 # aggregated Karatsuba post-processing
852 pshufb $T3,$T2
853 movdqu 16($inp),$Xln # Ii+1
854
855 pxor $Xhi,$T1
856 pxor $T2,$Xhi # "Ii+Xi", consume early
857 pxor $T1,$Xmn
858 pshufb $T3,$Xln
859 movdqa $Xmn,$T1 #
860 psrldq \$8,$T1
861 pslldq \$8,$Xmn #
862 pxor $T1,$Xhi
863 pxor $Xmn,$Xi #
864
865 movdqa $Xln,$Xhn #
866
867 movdqa $Xi,$T2 # 1st phase
868 movdqa $Xi,$T1
869 psllq \$5,$Xi
870 pxor $Xi,$T1 #
871 pclmulqdq \$0x00,$Hkey,$Xln #######
872 psllq \$1,$Xi
873 pxor $T1,$Xi #
874 psllq \$57,$Xi #
875 movdqa $Xi,$T1 #
876 pslldq \$8,$Xi
877 psrldq \$8,$T1 #
878 pxor $T2,$Xi
879 pshufd \$0b01001110,$Xhn,$Xmn
880 pxor $T1,$Xhi #
881 pxor $Xhn,$Xmn #
882
883 movdqa $Xi,$T2 # 2nd phase
884 psrlq \$1,$Xi
885 pclmulqdq \$0x11,$Hkey,$Xhn #######
886 pxor $T2,$Xhi #
887 pxor $Xi,$T2
888 psrlq \$5,$Xi
889 pxor $T2,$Xi #
890 lea 32($inp),$inp
891 psrlq \$1,$Xi #
892 pclmulqdq \$0x00,$HK,$Xmn #######
893 pxor $Xhi,$Xi #
894
895 sub \$0x20,$len
896 ja .Lmod_loop
897
898 .Leven_tail:
899 movdqa $Xi,$Xhi
900 movdqa $Xmn,$T1
901 pshufd \$0b01001110,$Xi,$Xmn #
902 pxor $Xi,$Xmn #
903
904 pclmulqdq \$0x00,$Hkey2,$Xi
905 pclmulqdq \$0x11,$Hkey2,$Xhi
906 pclmulqdq \$0x10,$HK,$Xmn
907
908 pxor $Xln,$Xi # (H*Ii+1) + H^2*(Ii+Xi)
909 pxor $Xhn,$Xhi
910 pxor $Xi,$T1
911 pxor $Xhi,$T1
912 pxor $T1,$Xmn
913 movdqa $Xmn,$T1 #
914 psrldq \$8,$T1
915 pslldq \$8,$Xmn #
916 pxor $T1,$Xhi
917 pxor $Xmn,$Xi #
918 ___
919 &reduction_alg9 ($Xhi,$Xi);
920 $code.=<<___;
921 test $len,$len
922 jnz .Ldone
923
924 .Lodd_tail:
925 movdqu ($inp),$T1 # Ii
926 pshufb $T3,$T1
927 pxor $T1,$Xi # Ii+Xi
928 ___
929 &clmul64x64_T2 ($Xhi,$Xi,$Hkey,$HK); # H*(Ii+Xi)
930 &reduction_alg9 ($Xhi,$Xi);
931 $code.=<<___;
932 .Ldone:
933 pshufb $T3,$Xi
934 movdqu $Xi,($Xip)
935 ___
936 $code.=<<___ if ($win64);
937 movaps (%rsp),%xmm6
938 movaps 0x10(%rsp),%xmm7
939 movaps 0x20(%rsp),%xmm8
940 movaps 0x30(%rsp),%xmm9
941 movaps 0x40(%rsp),%xmm10
942 movaps 0x50(%rsp),%xmm11
943 movaps 0x60(%rsp),%xmm12
944 movaps 0x70(%rsp),%xmm13
945 movaps 0x80(%rsp),%xmm14
946 movaps 0x90(%rsp),%xmm15
947 lea 0xa8(%rsp),%rsp
948 .LSEH_end_gcm_ghash_clmul:
949 ___
950 $code.=<<___;
951 ret
952 .size gcm_ghash_clmul,.-gcm_ghash_clmul
953 ___
954 }
955
956 $code.=<<___;
957 .globl gcm_init_avx
958 .type gcm_init_avx,\@abi-omnipotent
959 .align 32
960 gcm_init_avx:
961 ___
962 if ($avx) {
963 my ($Htbl,$Xip)=@_4args;
964 my $HK="%xmm6";
965
966 $code.=<<___ if ($win64);
967 .LSEH_begin_gcm_init_avx:
968 # I can't trust assembler to use specific encoding:-(
969 .byte 0x48,0x83,0xec,0x18 #sub $0x18,%rsp
970 .byte 0x0f,0x29,0x34,0x24 #movaps %xmm6,(%rsp)
971 ___
972 $code.=<<___;
973 vzeroupper
974
975 vmovdqu ($Xip),$Hkey
976 vpshufd \$0b01001110,$Hkey,$Hkey # dword swap
977
978 # <<1 twist
979 vpshufd \$0b11111111,$Hkey,$T2 # broadcast uppermost dword
980 vpsrlq \$63,$Hkey,$T1
981 vpsllq \$1,$Hkey,$Hkey
982 vpxor $T3,$T3,$T3 #
983 vpcmpgtd $T2,$T3,$T3 # broadcast carry bit
984 vpslldq \$8,$T1,$T1
985 vpor $T1,$Hkey,$Hkey # H<<=1
986
987 # magic reduction
988 vpand .L0x1c2_polynomial(%rip),$T3,$T3
989 vpxor $T3,$Hkey,$Hkey # if(carry) H^=0x1c2_polynomial
990
991 vpunpckhqdq $Hkey,$Hkey,$HK
992 vmovdqa $Hkey,$Xi
993 vpxor $Hkey,$HK,$HK
994 mov \$4,%r10 # up to H^8
995 jmp .Linit_start_avx
996 ___
997
998 sub clmul64x64_avx {
999 my ($Xhi,$Xi,$Hkey,$HK)=@_;
1000
1001 if (!defined($HK)) { $HK = $T2;
1002 $code.=<<___;
1003 vpunpckhqdq $Xi,$Xi,$T1
1004 vpunpckhqdq $Hkey,$Hkey,$T2
1005 vpxor $Xi,$T1,$T1 #
1006 vpxor $Hkey,$T2,$T2
1007 ___
1008 } else {
1009 $code.=<<___;
1010 vpunpckhqdq $Xi,$Xi,$T1
1011 vpxor $Xi,$T1,$T1 #
1012 ___
1013 }
1014 $code.=<<___;
1015 vpclmulqdq \$0x11,$Hkey,$Xi,$Xhi #######
1016 vpclmulqdq \$0x00,$Hkey,$Xi,$Xi #######
1017 vpclmulqdq \$0x00,$HK,$T1,$T1 #######
1018 vpxor $Xi,$Xhi,$T2 #
1019 vpxor $T2,$T1,$T1 #
1020
1021 vpslldq \$8,$T1,$T2 #
1022 vpsrldq \$8,$T1,$T1
1023 vpxor $T2,$Xi,$Xi #
1024 vpxor $T1,$Xhi,$Xhi
1025 ___
1026 }
1027
1028 sub reduction_avx {
1029 my ($Xhi,$Xi) = @_;
1030
1031 $code.=<<___;
1032 vpsllq \$57,$Xi,$T1 # 1st phase
1033 vpsllq \$62,$Xi,$T2
1034 vpxor $T1,$T2,$T2 #
1035 vpsllq \$63,$Xi,$T1
1036 vpxor $T1,$T2,$T2 #
1037 vpslldq \$8,$T2,$T1 #
1038 vpsrldq \$8,$T2,$T2
1039 vpxor $T1,$Xi,$Xi #
1040 vpxor $T2,$Xhi,$Xhi
1041
1042 vpsrlq \$1,$Xi,$T2 # 2nd phase
1043 vpxor $Xi,$Xhi,$Xhi
1044 vpxor $T2,$Xi,$Xi #
1045 vpsrlq \$5,$T2,$T2
1046 vpxor $T2,$Xi,$Xi #
1047 vpsrlq \$1,$Xi,$Xi #
1048 vpxor $Xhi,$Xi,$Xi #
1049 ___
1050 }
1051
1052 $code.=<<___;
1053 .align 32
1054 .Linit_loop_avx:
1055 vpalignr \$8,$T1,$T2,$T3 # low part is H.lo^H.hi...
1056 vmovdqu $T3,-0x10($Htbl) # save Karatsuba "salt"
1057 ___
1058 &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK); # calculate H^3,5,7
1059 &reduction_avx ($Xhi,$Xi);
1060 $code.=<<___;
1061 .Linit_start_avx:
1062 vmovdqa $Xi,$T3
1063 ___
1064 &clmul64x64_avx ($Xhi,$Xi,$Hkey,$HK); # calculate H^2,4,6,8
1065 &reduction_avx ($Xhi,$Xi);
1066 $code.=<<___;
1067 vpshufd \$0b01001110,$T3,$T1
1068 vpshufd \$0b01001110,$Xi,$T2
1069 vpxor $T3,$T1,$T1 # Karatsuba pre-processing
1070 vmovdqu $T3,0x00($Htbl) # save H^1,3,5,7
1071 vpxor $Xi,$T2,$T2 # Karatsuba pre-processing
1072 vmovdqu $Xi,0x10($Htbl) # save H^2,4,6,8
1073 lea 0x30($Htbl),$Htbl
1074 sub \$1,%r10
1075 jnz .Linit_loop_avx
1076
1077 vpalignr \$8,$T2,$T1,$T3 # last "salt" is flipped
1078 vmovdqu $T3,-0x10($Htbl)
1079
1080 vzeroupper
1081 ___
1082 $code.=<<___ if ($win64);
1083 movaps (%rsp),%xmm6
1084 lea 0x18(%rsp),%rsp
1085 .LSEH_end_gcm_init_avx:
1086 ___
1087 $code.=<<___;
1088 ret
1089 .size gcm_init_avx,.-gcm_init_avx
1090 ___
1091 } else {
1092 $code.=<<___;
1093 jmp .L_init_clmul
1094 .size gcm_init_avx,.-gcm_init_avx
1095 ___
1096 }
1097
1098 $code.=<<___;
1099 .globl gcm_gmult_avx
1100 .type gcm_gmult_avx,\@abi-omnipotent
1101 .align 32
1102 gcm_gmult_avx:
1103 jmp .L_gmult_clmul
1104 .size gcm_gmult_avx,.-gcm_gmult_avx
1105 ___
1106
1107 $code.=<<___;
1108 .globl gcm_ghash_avx
1109 .type gcm_ghash_avx,\@abi-omnipotent
1110 .align 32
1111 gcm_ghash_avx:
1112 ___
1113 if ($avx) {
1114 my ($Xip,$Htbl,$inp,$len)=@_4args;
1115 my ($Xlo,$Xhi,$Xmi,
1116 $Zlo,$Zhi,$Zmi,
1117 $Hkey,$HK,$T1,$T2,
1118 $Xi,$Xo,$Tred,$bswap,$Ii,$Ij) = map("%xmm$_",(0..15));
1119
1120 $code.=<<___ if ($win64);
1121 lea -0x88(%rsp),%rax
1122 .LSEH_begin_gcm_ghash_avx:
1123 # I can't trust assembler to use specific encoding:-(
1124 .byte 0x48,0x8d,0x60,0xe0 #lea -0x20(%rax),%rsp
1125 .byte 0x0f,0x29,0x70,0xe0 #movaps %xmm6,-0x20(%rax)
1126 .byte 0x0f,0x29,0x78,0xf0 #movaps %xmm7,-0x10(%rax)
1127 .byte 0x44,0x0f,0x29,0x00 #movaps %xmm8,0(%rax)
1128 .byte 0x44,0x0f,0x29,0x48,0x10 #movaps %xmm9,0x10(%rax)
1129 .byte 0x44,0x0f,0x29,0x50,0x20 #movaps %xmm10,0x20(%rax)
1130 .byte 0x44,0x0f,0x29,0x58,0x30 #movaps %xmm11,0x30(%rax)
1131 .byte 0x44,0x0f,0x29,0x60,0x40 #movaps %xmm12,0x40(%rax)
1132 .byte 0x44,0x0f,0x29,0x68,0x50 #movaps %xmm13,0x50(%rax)
1133 .byte 0x44,0x0f,0x29,0x70,0x60 #movaps %xmm14,0x60(%rax)
1134 .byte 0x44,0x0f,0x29,0x78,0x70 #movaps %xmm15,0x70(%rax)
1135 ___
1136 $code.=<<___;
1137 vzeroupper
1138
1139 vmovdqu ($Xip),$Xi # load $Xi
1140 lea .L0x1c2_polynomial(%rip),%r10
1141 lea 0x40($Htbl),$Htbl # size optimization
1142 vmovdqu .Lbswap_mask(%rip),$bswap
1143 vpshufb $bswap,$Xi,$Xi
1144 cmp \$0x80,$len
1145 jb .Lshort_avx
1146 sub \$0x80,$len
1147
1148 vmovdqu 0x70($inp),$Ii # I[7]
1149 vmovdqu 0x00-0x40($Htbl),$Hkey # $Hkey^1
1150 vpshufb $bswap,$Ii,$Ii
1151 vmovdqu 0x20-0x40($Htbl),$HK
1152
1153 vpunpckhqdq $Ii,$Ii,$T2
1154 vmovdqu 0x60($inp),$Ij # I[6]
1155 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1156 vpxor $Ii,$T2,$T2
1157 vpshufb $bswap,$Ij,$Ij
1158 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1159 vmovdqu 0x10-0x40($Htbl),$Hkey # $Hkey^2
1160 vpunpckhqdq $Ij,$Ij,$T1
1161 vmovdqu 0x50($inp),$Ii # I[5]
1162 vpclmulqdq \$0x00,$HK,$T2,$Xmi
1163 vpxor $Ij,$T1,$T1
1164
1165 vpshufb $bswap,$Ii,$Ii
1166 vpclmulqdq \$0x00,$Hkey,$Ij,$Zlo
1167 vpunpckhqdq $Ii,$Ii,$T2
1168 vpclmulqdq \$0x11,$Hkey,$Ij,$Zhi
1169 vmovdqu 0x30-0x40($Htbl),$Hkey # $Hkey^3
1170 vpxor $Ii,$T2,$T2
1171 vmovdqu 0x40($inp),$Ij # I[4]
1172 vpclmulqdq \$0x10,$HK,$T1,$Zmi
1173 vmovdqu 0x50-0x40($Htbl),$HK
1174
1175 vpshufb $bswap,$Ij,$Ij
1176 vpxor $Xlo,$Zlo,$Zlo
1177 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1178 vpxor $Xhi,$Zhi,$Zhi
1179 vpunpckhqdq $Ij,$Ij,$T1
1180 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1181 vmovdqu 0x40-0x40($Htbl),$Hkey # $Hkey^4
1182 vpxor $Xmi,$Zmi,$Zmi
1183 vpclmulqdq \$0x00,$HK,$T2,$Xmi
1184 vpxor $Ij,$T1,$T1
1185
1186 vmovdqu 0x30($inp),$Ii # I[3]
1187 vpxor $Zlo,$Xlo,$Xlo
1188 vpclmulqdq \$0x00,$Hkey,$Ij,$Zlo
1189 vpxor $Zhi,$Xhi,$Xhi
1190 vpshufb $bswap,$Ii,$Ii
1191 vpclmulqdq \$0x11,$Hkey,$Ij,$Zhi
1192 vmovdqu 0x60-0x40($Htbl),$Hkey # $Hkey^5
1193 vpxor $Zmi,$Xmi,$Xmi
1194 vpunpckhqdq $Ii,$Ii,$T2
1195 vpclmulqdq \$0x10,$HK,$T1,$Zmi
1196 vmovdqu 0x80-0x40($Htbl),$HK
1197 vpxor $Ii,$T2,$T2
1198
1199 vmovdqu 0x20($inp),$Ij # I[2]
1200 vpxor $Xlo,$Zlo,$Zlo
1201 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1202 vpxor $Xhi,$Zhi,$Zhi
1203 vpshufb $bswap,$Ij,$Ij
1204 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1205 vmovdqu 0x70-0x40($Htbl),$Hkey # $Hkey^6
1206 vpxor $Xmi,$Zmi,$Zmi
1207 vpunpckhqdq $Ij,$Ij,$T1
1208 vpclmulqdq \$0x00,$HK,$T2,$Xmi
1209 vpxor $Ij,$T1,$T1
1210
1211 vmovdqu 0x10($inp),$Ii # I[1]
1212 vpxor $Zlo,$Xlo,$Xlo
1213 vpclmulqdq \$0x00,$Hkey,$Ij,$Zlo
1214 vpxor $Zhi,$Xhi,$Xhi
1215 vpshufb $bswap,$Ii,$Ii
1216 vpclmulqdq \$0x11,$Hkey,$Ij,$Zhi
1217 vmovdqu 0x90-0x40($Htbl),$Hkey # $Hkey^7
1218 vpxor $Zmi,$Xmi,$Xmi
1219 vpunpckhqdq $Ii,$Ii,$T2
1220 vpclmulqdq \$0x10,$HK,$T1,$Zmi
1221 vmovdqu 0xb0-0x40($Htbl),$HK
1222 vpxor $Ii,$T2,$T2
1223
1224 vmovdqu ($inp),$Ij # I[0]
1225 vpxor $Xlo,$Zlo,$Zlo
1226 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1227 vpxor $Xhi,$Zhi,$Zhi
1228 vpshufb $bswap,$Ij,$Ij
1229 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1230 vmovdqu 0xa0-0x40($Htbl),$Hkey # $Hkey^8
1231 vpxor $Xmi,$Zmi,$Zmi
1232 vpclmulqdq \$0x10,$HK,$T2,$Xmi
1233
1234 lea 0x80($inp),$inp
1235 cmp \$0x80,$len
1236 jb .Ltail_avx
1237
1238 vpxor $Xi,$Ij,$Ij # accumulate $Xi
1239 sub \$0x80,$len
1240 jmp .Loop8x_avx
1241
1242 .align 32
1243 .Loop8x_avx:
1244 vpunpckhqdq $Ij,$Ij,$T1
1245 vmovdqu 0x70($inp),$Ii # I[7]
1246 vpxor $Xlo,$Zlo,$Zlo
1247 vpxor $Ij,$T1,$T1
1248 vpclmulqdq \$0x00,$Hkey,$Ij,$Xi
1249 vpshufb $bswap,$Ii,$Ii
1250 vpxor $Xhi,$Zhi,$Zhi
1251 vpclmulqdq \$0x11,$Hkey,$Ij,$Xo
1252 vmovdqu 0x00-0x40($Htbl),$Hkey # $Hkey^1
1253 vpunpckhqdq $Ii,$Ii,$T2
1254 vpxor $Xmi,$Zmi,$Zmi
1255 vpclmulqdq \$0x00,$HK,$T1,$Tred
1256 vmovdqu 0x20-0x40($Htbl),$HK
1257 vpxor $Ii,$T2,$T2
1258
1259 vmovdqu 0x60($inp),$Ij # I[6]
1260 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1261 vpxor $Zlo,$Xi,$Xi # collect result
1262 vpshufb $bswap,$Ij,$Ij
1263 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1264 vxorps $Zhi,$Xo,$Xo
1265 vmovdqu 0x10-0x40($Htbl),$Hkey # $Hkey^2
1266 vpunpckhqdq $Ij,$Ij,$T1
1267 vpclmulqdq \$0x00,$HK, $T2,$Xmi
1268 vpxor $Zmi,$Tred,$Tred
1269 vxorps $Ij,$T1,$T1
1270
1271 vmovdqu 0x50($inp),$Ii # I[5]
1272 vpxor $Xi,$Tred,$Tred # aggregated Karatsuba post-processing
1273 vpclmulqdq \$0x00,$Hkey,$Ij,$Zlo
1274 vpxor $Xo,$Tred,$Tred
1275 vpslldq \$8,$Tred,$T2
1276 vpxor $Xlo,$Zlo,$Zlo
1277 vpclmulqdq \$0x11,$Hkey,$Ij,$Zhi
1278 vpsrldq \$8,$Tred,$Tred
1279 vpxor $T2, $Xi, $Xi
1280 vmovdqu 0x30-0x40($Htbl),$Hkey # $Hkey^3
1281 vpshufb $bswap,$Ii,$Ii
1282 vxorps $Tred,$Xo, $Xo
1283 vpxor $Xhi,$Zhi,$Zhi
1284 vpunpckhqdq $Ii,$Ii,$T2
1285 vpclmulqdq \$0x10,$HK, $T1,$Zmi
1286 vmovdqu 0x50-0x40($Htbl),$HK
1287 vpxor $Ii,$T2,$T2
1288 vpxor $Xmi,$Zmi,$Zmi
1289
1290 vmovdqu 0x40($inp),$Ij # I[4]
1291 vpalignr \$8,$Xi,$Xi,$Tred # 1st phase
1292 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1293 vpshufb $bswap,$Ij,$Ij
1294 vpxor $Zlo,$Xlo,$Xlo
1295 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1296 vmovdqu 0x40-0x40($Htbl),$Hkey # $Hkey^4
1297 vpunpckhqdq $Ij,$Ij,$T1
1298 vpxor $Zhi,$Xhi,$Xhi
1299 vpclmulqdq \$0x00,$HK, $T2,$Xmi
1300 vxorps $Ij,$T1,$T1
1301 vpxor $Zmi,$Xmi,$Xmi
1302
1303 vmovdqu 0x30($inp),$Ii # I[3]
1304 vpclmulqdq \$0x10,(%r10),$Xi,$Xi
1305 vpclmulqdq \$0x00,$Hkey,$Ij,$Zlo
1306 vpshufb $bswap,$Ii,$Ii
1307 vpxor $Xlo,$Zlo,$Zlo
1308 vpclmulqdq \$0x11,$Hkey,$Ij,$Zhi
1309 vmovdqu 0x60-0x40($Htbl),$Hkey # $Hkey^5
1310 vpunpckhqdq $Ii,$Ii,$T2
1311 vpxor $Xhi,$Zhi,$Zhi
1312 vpclmulqdq \$0x10,$HK, $T1,$Zmi
1313 vmovdqu 0x80-0x40($Htbl),$HK
1314 vpxor $Ii,$T2,$T2
1315 vpxor $Xmi,$Zmi,$Zmi
1316
1317 vmovdqu 0x20($inp),$Ij # I[2]
1318 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1319 vpshufb $bswap,$Ij,$Ij
1320 vpxor $Zlo,$Xlo,$Xlo
1321 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1322 vmovdqu 0x70-0x40($Htbl),$Hkey # $Hkey^6
1323 vpunpckhqdq $Ij,$Ij,$T1
1324 vpxor $Zhi,$Xhi,$Xhi
1325 vpclmulqdq \$0x00,$HK, $T2,$Xmi
1326 vpxor $Ij,$T1,$T1
1327 vpxor $Zmi,$Xmi,$Xmi
1328 vxorps $Tred,$Xi,$Xi
1329
1330 vmovdqu 0x10($inp),$Ii # I[1]
1331 vpalignr \$8,$Xi,$Xi,$Tred # 2nd phase
1332 vpclmulqdq \$0x00,$Hkey,$Ij,$Zlo
1333 vpshufb $bswap,$Ii,$Ii
1334 vpxor $Xlo,$Zlo,$Zlo
1335 vpclmulqdq \$0x11,$Hkey,$Ij,$Zhi
1336 vmovdqu 0x90-0x40($Htbl),$Hkey # $Hkey^7
1337 vpclmulqdq \$0x10,(%r10),$Xi,$Xi
1338 vxorps $Xo,$Tred,$Tred
1339 vpunpckhqdq $Ii,$Ii,$T2
1340 vpxor $Xhi,$Zhi,$Zhi
1341 vpclmulqdq \$0x10,$HK, $T1,$Zmi
1342 vmovdqu 0xb0-0x40($Htbl),$HK
1343 vpxor $Ii,$T2,$T2
1344 vpxor $Xmi,$Zmi,$Zmi
1345
1346 vmovdqu ($inp),$Ij # I[0]
1347 vpclmulqdq \$0x00,$Hkey,$Ii,$Xlo
1348 vpshufb $bswap,$Ij,$Ij
1349 vpclmulqdq \$0x11,$Hkey,$Ii,$Xhi
1350 vmovdqu 0xa0-0x40($Htbl),$Hkey # $Hkey^8
1351 vpxor $Tred,$Ij,$Ij
1352 vpclmulqdq \$0x10,$HK, $T2,$Xmi
1353 vpxor $Xi,$Ij,$Ij # accumulate $Xi
1354
1355 lea 0x80($inp),$inp
1356 sub \$0x80,$len
1357 jnc .Loop8x_avx
1358
1359 add \$0x80,$len
1360 jmp .Ltail_no_xor_avx
1361
1362 .align 32
1363 .Lshort_avx:
1364 vmovdqu -0x10($inp,$len),$Ii # very last word
1365 lea ($inp,$len),$inp
1366 vmovdqu 0x00-0x40($Htbl),$Hkey # $Hkey^1
1367 vmovdqu 0x20-0x40($Htbl),$HK
1368 vpshufb $bswap,$Ii,$Ij
1369
1370 vmovdqa $Xlo,$Zlo # subtle way to zero $Zlo,
1371 vmovdqa $Xhi,$Zhi # $Zhi and
1372 vmovdqa $Xmi,$Zmi # $Zmi
1373 sub \$0x10,$len
1374 jz .Ltail_avx
1375
1376 vpunpckhqdq $Ij,$Ij,$T1
1377 vpxor $Xlo,$Zlo,$Zlo
1378 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1379 vpxor $Ij,$T1,$T1
1380 vmovdqu -0x20($inp),$Ii
1381 vpxor $Xhi,$Zhi,$Zhi
1382 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1383 vmovdqu 0x10-0x40($Htbl),$Hkey # $Hkey^2
1384 vpshufb $bswap,$Ii,$Ij
1385 vpxor $Xmi,$Zmi,$Zmi
1386 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1387 vpsrldq \$8,$HK,$HK
1388 sub \$0x10,$len
1389 jz .Ltail_avx
1390
1391 vpunpckhqdq $Ij,$Ij,$T1
1392 vpxor $Xlo,$Zlo,$Zlo
1393 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1394 vpxor $Ij,$T1,$T1
1395 vmovdqu -0x30($inp),$Ii
1396 vpxor $Xhi,$Zhi,$Zhi
1397 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1398 vmovdqu 0x30-0x40($Htbl),$Hkey # $Hkey^3
1399 vpshufb $bswap,$Ii,$Ij
1400 vpxor $Xmi,$Zmi,$Zmi
1401 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1402 vmovdqu 0x50-0x40($Htbl),$HK
1403 sub \$0x10,$len
1404 jz .Ltail_avx
1405
1406 vpunpckhqdq $Ij,$Ij,$T1
1407 vpxor $Xlo,$Zlo,$Zlo
1408 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1409 vpxor $Ij,$T1,$T1
1410 vmovdqu -0x40($inp),$Ii
1411 vpxor $Xhi,$Zhi,$Zhi
1412 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1413 vmovdqu 0x40-0x40($Htbl),$Hkey # $Hkey^4
1414 vpshufb $bswap,$Ii,$Ij
1415 vpxor $Xmi,$Zmi,$Zmi
1416 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1417 vpsrldq \$8,$HK,$HK
1418 sub \$0x10,$len
1419 jz .Ltail_avx
1420
1421 vpunpckhqdq $Ij,$Ij,$T1
1422 vpxor $Xlo,$Zlo,$Zlo
1423 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1424 vpxor $Ij,$T1,$T1
1425 vmovdqu -0x50($inp),$Ii
1426 vpxor $Xhi,$Zhi,$Zhi
1427 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1428 vmovdqu 0x60-0x40($Htbl),$Hkey # $Hkey^5
1429 vpshufb $bswap,$Ii,$Ij
1430 vpxor $Xmi,$Zmi,$Zmi
1431 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1432 vmovdqu 0x80-0x40($Htbl),$HK
1433 sub \$0x10,$len
1434 jz .Ltail_avx
1435
1436 vpunpckhqdq $Ij,$Ij,$T1
1437 vpxor $Xlo,$Zlo,$Zlo
1438 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1439 vpxor $Ij,$T1,$T1
1440 vmovdqu -0x60($inp),$Ii
1441 vpxor $Xhi,$Zhi,$Zhi
1442 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1443 vmovdqu 0x70-0x40($Htbl),$Hkey # $Hkey^6
1444 vpshufb $bswap,$Ii,$Ij
1445 vpxor $Xmi,$Zmi,$Zmi
1446 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1447 vpsrldq \$8,$HK,$HK
1448 sub \$0x10,$len
1449 jz .Ltail_avx
1450
1451 vpunpckhqdq $Ij,$Ij,$T1
1452 vpxor $Xlo,$Zlo,$Zlo
1453 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1454 vpxor $Ij,$T1,$T1
1455 vmovdqu -0x70($inp),$Ii
1456 vpxor $Xhi,$Zhi,$Zhi
1457 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1458 vmovdqu 0x90-0x40($Htbl),$Hkey # $Hkey^7
1459 vpshufb $bswap,$Ii,$Ij
1460 vpxor $Xmi,$Zmi,$Zmi
1461 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1462 vmovq 0xb8-0x40($Htbl),$HK
1463 sub \$0x10,$len
1464 jmp .Ltail_avx
1465
1466 .align 32
1467 .Ltail_avx:
1468 vpxor $Xi,$Ij,$Ij # accumulate $Xi
1469 .Ltail_no_xor_avx:
1470 vpunpckhqdq $Ij,$Ij,$T1
1471 vpxor $Xlo,$Zlo,$Zlo
1472 vpclmulqdq \$0x00,$Hkey,$Ij,$Xlo
1473 vpxor $Ij,$T1,$T1
1474 vpxor $Xhi,$Zhi,$Zhi
1475 vpclmulqdq \$0x11,$Hkey,$Ij,$Xhi
1476 vpxor $Xmi,$Zmi,$Zmi
1477 vpclmulqdq \$0x00,$HK,$T1,$Xmi
1478
1479 vmovdqu (%r10),$Tred
1480
1481 vpxor $Xlo,$Zlo,$Xi
1482 vpxor $Xhi,$Zhi,$Xo
1483 vpxor $Xmi,$Zmi,$Zmi
1484
1485 vpxor $Xi, $Zmi,$Zmi # aggregated Karatsuba post-processing
1486 vpxor $Xo, $Zmi,$Zmi
1487 vpslldq \$8, $Zmi,$T2
1488 vpsrldq \$8, $Zmi,$Zmi
1489 vpxor $T2, $Xi, $Xi
1490 vpxor $Zmi,$Xo, $Xo
1491
1492 vpclmulqdq \$0x10,$Tred,$Xi,$T2 # 1st phase
1493 vpalignr \$8,$Xi,$Xi,$Xi
1494 vpxor $T2,$Xi,$Xi
1495
1496 vpclmulqdq \$0x10,$Tred,$Xi,$T2 # 2nd phase
1497 vpalignr \$8,$Xi,$Xi,$Xi
1498 vpxor $Xo,$Xi,$Xi
1499 vpxor $T2,$Xi,$Xi
1500
1501 cmp \$0,$len
1502 jne .Lshort_avx
1503
1504 vpshufb $bswap,$Xi,$Xi
1505 vmovdqu $Xi,($Xip)
1506 vzeroupper
1507 ___
1508 $code.=<<___ if ($win64);
1509 movaps (%rsp),%xmm6
1510 movaps 0x10(%rsp),%xmm7
1511 movaps 0x20(%rsp),%xmm8
1512 movaps 0x30(%rsp),%xmm9
1513 movaps 0x40(%rsp),%xmm10
1514 movaps 0x50(%rsp),%xmm11
1515 movaps 0x60(%rsp),%xmm12
1516 movaps 0x70(%rsp),%xmm13
1517 movaps 0x80(%rsp),%xmm14
1518 movaps 0x90(%rsp),%xmm15
1519 lea 0xa8(%rsp),%rsp
1520 .LSEH_end_gcm_ghash_avx:
1521 ___
1522 $code.=<<___;
1523 ret
1524 .size gcm_ghash_avx,.-gcm_ghash_avx
1525 ___
1526 } else {
1527 $code.=<<___;
1528 jmp .L_ghash_clmul
1529 .size gcm_ghash_avx,.-gcm_ghash_avx
1530 ___
1531 }
1532
1533 $code.=<<___;
1534 .align 64
1535 .Lbswap_mask:
1536 .byte 15,14,13,12,11,10,9,8,7,6,5,4,3,2,1,0
1537 .L0x1c2_polynomial:
1538 .byte 1,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0xc2
1539 .L7_mask:
1540 .long 7,0,7,0
1541 .L7_mask_poly:
1542 .long 7,0,`0xE1<<1`,0
1543 .align 64
1544 .type .Lrem_4bit,\@object
1545 .Lrem_4bit:
1546 .long 0,`0x0000<<16`,0,`0x1C20<<16`,0,`0x3840<<16`,0,`0x2460<<16`
1547 .long 0,`0x7080<<16`,0,`0x6CA0<<16`,0,`0x48C0<<16`,0,`0x54E0<<16`
1548 .long 0,`0xE100<<16`,0,`0xFD20<<16`,0,`0xD940<<16`,0,`0xC560<<16`
1549 .long 0,`0x9180<<16`,0,`0x8DA0<<16`,0,`0xA9C0<<16`,0,`0xB5E0<<16`
1550 .type .Lrem_8bit,\@object
1551 .Lrem_8bit:
1552 .value 0x0000,0x01C2,0x0384,0x0246,0x0708,0x06CA,0x048C,0x054E
1553 .value 0x0E10,0x0FD2,0x0D94,0x0C56,0x0918,0x08DA,0x0A9C,0x0B5E
1554 .value 0x1C20,0x1DE2,0x1FA4,0x1E66,0x1B28,0x1AEA,0x18AC,0x196E
1555 .value 0x1230,0x13F2,0x11B4,0x1076,0x1538,0x14FA,0x16BC,0x177E
1556 .value 0x3840,0x3982,0x3BC4,0x3A06,0x3F48,0x3E8A,0x3CCC,0x3D0E
1557 .value 0x3650,0x3792,0x35D4,0x3416,0x3158,0x309A,0x32DC,0x331E
1558 .value 0x2460,0x25A2,0x27E4,0x2626,0x2368,0x22AA,0x20EC,0x212E
1559 .value 0x2A70,0x2BB2,0x29F4,0x2836,0x2D78,0x2CBA,0x2EFC,0x2F3E
1560 .value 0x7080,0x7142,0x7304,0x72C6,0x7788,0x764A,0x740C,0x75CE
1561 .value 0x7E90,0x7F52,0x7D14,0x7CD6,0x7998,0x785A,0x7A1C,0x7BDE
1562 .value 0x6CA0,0x6D62,0x6F24,0x6EE6,0x6BA8,0x6A6A,0x682C,0x69EE
1563 .value 0x62B0,0x6372,0x6134,0x60F6,0x65B8,0x647A,0x663C,0x67FE
1564 .value 0x48C0,0x4902,0x4B44,0x4A86,0x4FC8,0x4E0A,0x4C4C,0x4D8E
1565 .value 0x46D0,0x4712,0x4554,0x4496,0x41D8,0x401A,0x425C,0x439E
1566 .value 0x54E0,0x5522,0x5764,0x56A6,0x53E8,0x522A,0x506C,0x51AE
1567 .value 0x5AF0,0x5B32,0x5974,0x58B6,0x5DF8,0x5C3A,0x5E7C,0x5FBE
1568 .value 0xE100,0xE0C2,0xE284,0xE346,0xE608,0xE7CA,0xE58C,0xE44E
1569 .value 0xEF10,0xEED2,0xEC94,0xED56,0xE818,0xE9DA,0xEB9C,0xEA5E
1570 .value 0xFD20,0xFCE2,0xFEA4,0xFF66,0xFA28,0xFBEA,0xF9AC,0xF86E
1571 .value 0xF330,0xF2F2,0xF0B4,0xF176,0xF438,0xF5FA,0xF7BC,0xF67E
1572 .value 0xD940,0xD882,0xDAC4,0xDB06,0xDE48,0xDF8A,0xDDCC,0xDC0E
1573 .value 0xD750,0xD692,0xD4D4,0xD516,0xD058,0xD19A,0xD3DC,0xD21E
1574 .value 0xC560,0xC4A2,0xC6E4,0xC726,0xC268,0xC3AA,0xC1EC,0xC02E
1575 .value 0xCB70,0xCAB2,0xC8F4,0xC936,0xCC78,0xCDBA,0xCFFC,0xCE3E
1576 .value 0x9180,0x9042,0x9204,0x93C6,0x9688,0x974A,0x950C,0x94CE
1577 .value 0x9F90,0x9E52,0x9C14,0x9DD6,0x9898,0x995A,0x9B1C,0x9ADE
1578 .value 0x8DA0,0x8C62,0x8E24,0x8FE6,0x8AA8,0x8B6A,0x892C,0x88EE
1579 .value 0x83B0,0x8272,0x8034,0x81F6,0x84B8,0x857A,0x873C,0x86FE
1580 .value 0xA9C0,0xA802,0xAA44,0xAB86,0xAEC8,0xAF0A,0xAD4C,0xAC8E
1581 .value 0xA7D0,0xA612,0xA454,0xA596,0xA0D8,0xA11A,0xA35C,0xA29E
1582 .value 0xB5E0,0xB422,0xB664,0xB7A6,0xB2E8,0xB32A,0xB16C,0xB0AE
1583 .value 0xBBF0,0xBA32,0xB874,0xB9B6,0xBCF8,0xBD3A,0xBF7C,0xBEBE
1584
1585 .asciz "GHASH for x86_64, CRYPTOGAMS by <appro\@openssl.org>"
1586 .align 64
1587 ___
1588
1589 # EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame,
1590 # CONTEXT *context,DISPATCHER_CONTEXT *disp)
1591 if ($win64) {
1592 $rec="%rcx";
1593 $frame="%rdx";
1594 $context="%r8";
1595 $disp="%r9";
1596
1597 $code.=<<___;
1598 .extern __imp_RtlVirtualUnwind
1599 .type se_handler,\@abi-omnipotent
1600 .align 16
1601 se_handler:
1602 push %rsi
1603 push %rdi
1604 push %rbx
1605 push %rbp
1606 push %r12
1607 push %r13
1608 push %r14
1609 push %r15
1610 pushfq
1611 sub \$64,%rsp
1612
1613 mov 120($context),%rax # pull context->Rax
1614 mov 248($context),%rbx # pull context->Rip
1615
1616 mov 8($disp),%rsi # disp->ImageBase
1617 mov 56($disp),%r11 # disp->HandlerData
1618
1619 mov 0(%r11),%r10d # HandlerData[0]
1620 lea (%rsi,%r10),%r10 # prologue label
1621 cmp %r10,%rbx # context->Rip<prologue label
1622 jb .Lin_prologue
1623
1624 mov 152($context),%rax # pull context->Rsp
1625
1626 mov 4(%r11),%r10d # HandlerData[1]
1627 lea (%rsi,%r10),%r10 # epilogue label
1628 cmp %r10,%rbx # context->Rip>=epilogue label
1629 jae .Lin_prologue
1630
1631 lea 24(%rax),%rax # adjust "rsp"
1632
1633 mov -8(%rax),%rbx
1634 mov -16(%rax),%rbp
1635 mov -24(%rax),%r12
1636 mov %rbx,144($context) # restore context->Rbx
1637 mov %rbp,160($context) # restore context->Rbp
1638 mov %r12,216($context) # restore context->R12
1639
1640 .Lin_prologue:
1641 mov 8(%rax),%rdi
1642 mov 16(%rax),%rsi
1643 mov %rax,152($context) # restore context->Rsp
1644 mov %rsi,168($context) # restore context->Rsi
1645 mov %rdi,176($context) # restore context->Rdi
1646
1647 mov 40($disp),%rdi # disp->ContextRecord
1648 mov $context,%rsi # context
1649 mov \$`1232/8`,%ecx # sizeof(CONTEXT)
1650 .long 0xa548f3fc # cld; rep movsq
1651
1652 mov $disp,%rsi
1653 xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER
1654 mov 8(%rsi),%rdx # arg2, disp->ImageBase
1655 mov 0(%rsi),%r8 # arg3, disp->ControlPc
1656 mov 16(%rsi),%r9 # arg4, disp->FunctionEntry
1657 mov 40(%rsi),%r10 # disp->ContextRecord
1658 lea 56(%rsi),%r11 # &disp->HandlerData
1659 lea 24(%rsi),%r12 # &disp->EstablisherFrame
1660 mov %r10,32(%rsp) # arg5
1661 mov %r11,40(%rsp) # arg6
1662 mov %r12,48(%rsp) # arg7
1663 mov %rcx,56(%rsp) # arg8, (NULL)
1664 call *__imp_RtlVirtualUnwind(%rip)
1665
1666 mov \$1,%eax # ExceptionContinueSearch
1667 add \$64,%rsp
1668 popfq
1669 pop %r15
1670 pop %r14
1671 pop %r13
1672 pop %r12
1673 pop %rbp
1674 pop %rbx
1675 pop %rdi
1676 pop %rsi
1677 ret
1678 .size se_handler,.-se_handler
1679
1680 .section .pdata
1681 .align 4
1682 .rva .LSEH_begin_gcm_gmult_4bit
1683 .rva .LSEH_end_gcm_gmult_4bit
1684 .rva .LSEH_info_gcm_gmult_4bit
1685
1686 .rva .LSEH_begin_gcm_ghash_4bit
1687 .rva .LSEH_end_gcm_ghash_4bit
1688 .rva .LSEH_info_gcm_ghash_4bit
1689
1690 .rva .LSEH_begin_gcm_init_clmul
1691 .rva .LSEH_end_gcm_init_clmul
1692 .rva .LSEH_info_gcm_init_clmul
1693
1694 .rva .LSEH_begin_gcm_ghash_clmul
1695 .rva .LSEH_end_gcm_ghash_clmul
1696 .rva .LSEH_info_gcm_ghash_clmul
1697 ___
1698 $code.=<<___ if ($avx);
1699 .rva .LSEH_begin_gcm_init_avx
1700 .rva .LSEH_end_gcm_init_avx
1701 .rva .LSEH_info_gcm_init_clmul
1702
1703 .rva .LSEH_begin_gcm_ghash_avx
1704 .rva .LSEH_end_gcm_ghash_avx
1705 .rva .LSEH_info_gcm_ghash_clmul
1706 ___
1707 $code.=<<___;
1708 .section .xdata
1709 .align 8
1710 .LSEH_info_gcm_gmult_4bit:
1711 .byte 9,0,0,0
1712 .rva se_handler
1713 .rva .Lgmult_prologue,.Lgmult_epilogue # HandlerData
1714 .LSEH_info_gcm_ghash_4bit:
1715 .byte 9,0,0,0
1716 .rva se_handler
1717 .rva .Lghash_prologue,.Lghash_epilogue # HandlerData
1718 .LSEH_info_gcm_init_clmul:
1719 .byte 0x01,0x08,0x03,0x00
1720 .byte 0x08,0x68,0x00,0x00 #movaps 0x00(rsp),xmm6
1721 .byte 0x04,0x22,0x00,0x00 #sub rsp,0x18
1722 .LSEH_info_gcm_ghash_clmul:
1723 .byte 0x01,0x33,0x16,0x00
1724 .byte 0x33,0xf8,0x09,0x00 #movaps 0x90(rsp),xmm15
1725 .byte 0x2e,0xe8,0x08,0x00 #movaps 0x80(rsp),xmm14
1726 .byte 0x29,0xd8,0x07,0x00 #movaps 0x70(rsp),xmm13
1727 .byte 0x24,0xc8,0x06,0x00 #movaps 0x60(rsp),xmm12
1728 .byte 0x1f,0xb8,0x05,0x00 #movaps 0x50(rsp),xmm11
1729 .byte 0x1a,0xa8,0x04,0x00 #movaps 0x40(rsp),xmm10
1730 .byte 0x15,0x98,0x03,0x00 #movaps 0x30(rsp),xmm9
1731 .byte 0x10,0x88,0x02,0x00 #movaps 0x20(rsp),xmm8
1732 .byte 0x0c,0x78,0x01,0x00 #movaps 0x10(rsp),xmm7
1733 .byte 0x08,0x68,0x00,0x00 #movaps 0x00(rsp),xmm6
1734 .byte 0x04,0x01,0x15,0x00 #sub rsp,0xa8
1735 ___
1736 }
1737
1738 $code =~ s/\`([^\`]*)\`/eval($1)/gem;
1739
1740 print $code;
1741
1742 close STDOUT;