2 # Copyright 2007-2016 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 # sha1_block procedure for ARMv4.
21 # Size/performance trade-off
22 # ====================================================================
23 # impl size in bytes comp cycles[*] measured performance
24 # ====================================================================
26 # armv4-small 392/+29% 1958/+64% 2250/+96%
27 # armv4-compact 740/+89% 1552/+26% 1840/+22%
28 # armv4-large 1420/+92% 1307/+19% 1370/+34%[***]
29 # full unroll ~5100/+260% ~1260/+4% ~1300/+5%
30 # ====================================================================
31 # thumb = same as 'small' but in Thumb instructions[**] and
32 # with recurring code in two private functions;
33 # small = detached Xload/update, loops are folded;
34 # compact = detached Xload/update, 5x unroll;
35 # large = interleaved Xload/update, 5x unroll;
36 # full unroll = interleaved Xload/update, full unroll, estimated[!];
38 # [*] Manually counted instructions in "grand" loop body. Measured
39 # performance is affected by prologue and epilogue overhead,
40 # i-cache availability, branch penalties, etc.
41 # [**] While each Thumb instruction is twice smaller, they are not as
42 # diverse as ARM ones: e.g., there are only two arithmetic
43 # instructions with 3 arguments, no [fixed] rotate, addressing
44 # modes are limited. As result it takes more instructions to do
45 # the same job in Thumb, therefore the code is never twice as
46 # small and always slower.
47 # [***] which is also ~35% better than compiler generated code. Dual-
48 # issue Cortex A8 core was measured to process input block in
53 # Rescheduling for dual-issue pipeline resulted in 13% improvement on
54 # Cortex A8 core and in absolute terms ~870 cycles per input block
55 # [or 13.6 cycles per byte].
59 # Profiler-assisted and platform-specific optimization resulted in 10%
60 # improvement on Cortex A8 core and 12.2 cycles per byte.
64 # Add NEON implementation (see sha1-586.pl for background info). On
65 # Cortex A8 it was measured to process one byte in 6.7 cycles or >80%
66 # faster than integer-only code. Because [fully unrolled] NEON code
67 # is ~2.5x larger and there are some redundant instructions executed
68 # when processing last block, improvement is not as big for smallest
69 # blocks, only ~30%. Snapdragon S4 is a tad faster, 6.4 cycles per
70 # byte, which is also >80% faster than integer-only code. Cortex-A15
71 # is even faster spending 5.6 cycles per byte outperforming integer-
72 # only code by factor of 2.
76 # Add ARMv8 code path performing at 2.35 cpb on Apple A7.
79 if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
80 else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} }
82 if ($flavour && $flavour ne "void") {
83 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1;
84 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or
85 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or
86 die "can't locate arm-xlate.pl";
88 open STDOUT,"| \"$^X\" $xlate $flavour $output";
90 open STDOUT,">$output";
110 my ($a,$b,$c,$d,$e,$opt1,$opt2)=@_;
115 add $e,$K,$e,ror#2 @ E+=K_xx_xx
118 eor $t2,$t2,$t3 @ 1 cycle stall
119 eor $t1,$c,$d @ F_xx_xx
121 add $e,$e,$a,ror#27 @ E+=ROR(A,27)
122 eor $t0,$t0,$t2,ror#31
126 add $e,$e,$t0 @ E+=X[i]
131 my ($a,$b,$c,$d,$e)=@_;
137 add $e,$K,$e,ror#2 @ E+=K_00_19
139 orr $t0,$t0,$t1,lsl#8
140 eor $t1,$c,$d @ F_xx_xx
141 orr $t0,$t0,$t2,lsl#16
142 add $e,$e,$a,ror#27 @ E+=ROR(A,27)
143 orr $t0,$t0,$t3,lsl#24
145 ldr $t0,[$inp],#4 @ handles unaligned
146 add $e,$K,$e,ror#2 @ E+=K_00_19
147 eor $t1,$c,$d @ F_xx_xx
148 add $e,$e,$a,ror#27 @ E+=ROR(A,27)
150 rev $t0,$t0 @ byte swap
154 add $e,$e,$t0 @ E+=X[i]
155 eor $t1,$t1,$d,ror#2 @ F_00_19(B,C,D)
157 add $e,$e,$t1 @ E+=F_00_19(B,C,D)
162 my ($a,$b,$c,$d,$e)=@_;
163 &Xupdate(@_,"and $t1,$b,$t1,ror#2");
165 eor $t1,$t1,$d,ror#2 @ F_00_19(B,C,D)
166 add $e,$e,$t1 @ E+=F_00_19(B,C,D)
171 my ($a,$b,$c,$d,$e)=@_;
172 &Xupdate(@_,"eor $t1,$b,$t1,ror#2");
174 add $e,$e,$t1 @ E+=F_20_39(B,C,D)
179 my ($a,$b,$c,$d,$e)=@_;
180 &Xupdate(@_,"and $t1,$b,$t1,ror#2","and $t2,$c,$d");
182 add $e,$e,$t1 @ E+=F_40_59(B,C,D)
188 #include "arm_arch.h"
190 #if defined(__thumb2__)
199 .global sha1_block_data_order
200 .type sha1_block_data_order,%function
203 sha1_block_data_order:
204 #if __ARM_MAX_ARCH__>=7
206 ldr r12,.LOPENSSL_armcap
207 # if !defined(_WIN32)
209 ldr r12,[r3,r12] @ OPENSSL_armcap_P
211 # if defined(__APPLE__) || defined(_WIN32)
219 stmdb sp!,{r4-r12,lr}
220 add $len,$inp,$len,lsl#6 @ $len to point at the end of $inp
221 ldmia $ctx,{$a,$b,$c,$d,$e}
228 mov $e,$e,ror#30 @ [6]
231 for($i=0;$i<5;$i++) {
232 &BODY_00_15(@V); unshift(@V,pop(@V));
235 #if defined(__thumb2__)
241 bne .L_00_15 @ [((11+4)*5+2)*3]
244 &BODY_00_15(@V); unshift(@V,pop(@V));
245 &BODY_16_19(@V); unshift(@V,pop(@V));
246 &BODY_16_19(@V); unshift(@V,pop(@V));
247 &BODY_16_19(@V); unshift(@V,pop(@V));
248 &BODY_16_19(@V); unshift(@V,pop(@V));
251 ldr $K,.LK_20_39 @ [+15+16*4]
252 cmn sp,#0 @ [+3], clear carry to denote 20_39
255 for($i=0;$i<5;$i++) {
256 &BODY_20_39(@V); unshift(@V,pop(@V));
259 #if defined(__thumb2__)
263 teq $Xi,sp @ preserve carry
265 bne .L_20_39_or_60_79 @ [+((12+3)*5+2)*4]
266 bcs .L_done @ [+((12+3)*5+2)*4], spare 300 bytes
269 sub sp,sp,#20*4 @ [+2]
272 for($i=0;$i<5;$i++) {
273 &BODY_40_59(@V); unshift(@V,pop(@V));
276 #if defined(__thumb2__)
282 bne .L_40_59 @ [+((12+5)*5+2)*4]
286 cmp sp,#0 @ set carry to denote 60_79
287 b .L_20_39_or_60_79 @ [+4], spare 300 bytes
289 add sp,sp,#80*4 @ "deallocate" stack frame
290 ldmia $ctx,{$K,$t0,$t1,$t2,$t3}
296 stmia $ctx,{$a,$b,$c,$d,$e}
298 bne .Lloop @ [+18], total 1307
301 ldmia sp!,{r4-r12,pc}
303 ldmia sp!,{r4-r12,lr}
305 moveq pc,lr @ be binary compatible with V4, yet
306 bx lr @ interoperable with Thumb ISA:-)
308 .size sha1_block_data_order,.-sha1_block_data_order
311 .LK_00_19: .word 0x5a827999
312 .LK_20_39: .word 0x6ed9eba1
313 .LK_40_59: .word 0x8f1bbcdc
314 .LK_60_79: .word 0xca62c1d6
315 #if __ARM_MAX_ARCH__>=7
318 .word OPENSSL_armcap_P
320 .word OPENSSL_armcap_P-.Lsha1_block
323 .asciz "SHA1 block transform for ARMv4/NEON/ARMv8, CRYPTOGAMS by <appro\@openssl.org>"
326 #####################################################################
330 my @V=($a,$b,$c,$d,$e);
331 my ($K_XX_XX,$Ki,$t0,$t1,$Xfer,$saved_sp)=map("r$_",(8..12,14));
333 my @X=map("q$_",(8..11,0..3));
334 my @Tx=("q12","q13");
335 my ($K,$zero)=("q14","q15");
338 sub AUTOLOAD() # thunk [simplified] x86-style perlasm
339 { my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./;
341 $arg = "#$arg" if ($arg*1 eq $arg);
342 $code .= "\t$opcode\t".join(',',@_,$arg)."\n";
347 '($a,$b,$c,$d,$e)=@V;'. # '$code.="@ $j\n";'.
349 '&add ($e,$e,$Ki)', # e+=X[i]+K
351 '&ldr ($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))',
352 '&add ($e,$e,$a,"ror#27")', # e+=ROR(A,27)
353 '&eor ($t1,$t1,$t0)', # F_00_19
354 '&mov ($b,$b,"ror#2")', # b=ROR(b,2)
355 '&add ($e,$e,$t1);'. # e+=F_00_19
356 '$j++; unshift(@V,pop(@V));'
361 '($a,$b,$c,$d,$e)=@V;'. # '$code.="@ $j\n";'.
363 '&add ($e,$e,$Ki)', # e+=X[i]+K
364 '&ldr ($Ki,sprintf "[sp,#%d]",4*(($j+1)&15)) if ($j<79)',
365 '&eor ($t1,$t0,$c)', # F_20_39
366 '&add ($e,$e,$a,"ror#27")', # e+=ROR(A,27)
367 '&mov ($b,$b,"ror#2")', # b=ROR(b,2)
368 '&add ($e,$e,$t1);'. # e+=F_20_39
369 '$j++; unshift(@V,pop(@V));'
374 '($a,$b,$c,$d,$e)=@V;'. # '$code.="@ $j\n";'.
375 '&add ($e,$e,$Ki)', # e+=X[i]+K
377 '&ldr ($Ki,sprintf "[sp,#%d]",4*(($j+1)&15))',
378 '&add ($e,$e,$a,"ror#27")', # e+=ROR(A,27)
382 '&mov ($b,$b,"ror#2")', # b=ROR(b,2)
383 '&add ($e,$e,$t1);'. # e+=F_40_59
384 '$j++; unshift(@V,pop(@V));'
391 my @insns = (&$body,&$body,&$body,&$body);
394 &vext_8 (@X[0],@X[-4&7],@X[-3&7],8); # compose "X[-14]" in "X[0]"
398 &vadd_i32 (@Tx[1],@X[-1&7],$K);
400 &vld1_32 ("{$K\[]}","[$K_XX_XX,:32]!") if ($Xi%5==0);
402 &vext_8 (@Tx[0],@X[-1&7],$zero,4); # "X[-3]", 3 words
406 &veor (@X[0],@X[0],@X[-4&7]); # "X[0]"^="X[-16]"
409 &veor (@Tx[0],@Tx[0],@X[-2&7]); # "X[-3]"^"X[-8]"
412 &veor (@Tx[0],@Tx[0],@X[0]); # "X[0]"^="X[-3]"^"X[-8]
415 &vst1_32 ("{@Tx[1]}","[$Xfer,:128]!"); # X[]+K xfer
416 &sub ($Xfer,$Xfer,64) if ($Xi%4==0);
419 &vext_8 (@Tx[1],$zero,@Tx[0],4); # "X[0]"<<96, extract one dword
422 &vadd_i32 (@X[0],@Tx[0],@Tx[0]);
425 &vsri_32 (@X[0],@Tx[0],31); # "X[0]"<<<=1
429 &vshr_u32 (@Tx[0],@Tx[1],30);
432 &vshl_u32 (@Tx[1],@Tx[1],2);
435 &veor (@X[0],@X[0],@Tx[0]);
438 &veor (@X[0],@X[0],@Tx[1]); # "X[0]"^=("X[0]">>96)<<<2
440 foreach (@insns) { eval; } # remaining instructions [if any]
442 $Xi++; push(@X,shift(@X)); # "rotate" X[]
448 my @insns = (&$body,&$body,&$body,&$body);
451 &vext_8 (@Tx[0],@X[-2&7],@X[-1&7],8); # compose "X[-6]"
455 &veor (@X[0],@X[0],@X[-4&7]); # "X[0]"="X[-32]"^"X[-16]"
458 &veor (@X[0],@X[0],@X[-7&7]); # "X[0]"^="X[-28]"
461 &vadd_i32 (@Tx[1],@X[-1&7],$K);
463 &vld1_32 ("{$K\[]}","[$K_XX_XX,:32]!") if ($Xi%5==0);
465 &veor (@Tx[0],@Tx[0],@X[0]); # "X[-6]"^="X[0]"
468 &vshr_u32 (@X[0],@Tx[0],30);
471 &vst1_32 ("{@Tx[1]}","[$Xfer,:128]!"); # X[]+K xfer
472 &sub ($Xfer,$Xfer,64) if ($Xi%4==0);
475 &vsli_32 (@X[0],@Tx[0],2); # "X[0]"="X[-6]"<<<2
477 foreach (@insns) { eval; } # remaining instructions [if any]
479 $Xi++; push(@X,shift(@X)); # "rotate" X[]
485 my @insns = (&$body,&$body,&$body,&$body);
488 &vadd_i32 (@Tx[1],@X[-1&7],$K);
491 &vst1_32 ("{@Tx[1]}","[$Xfer,:128]!");
492 &sub ($Xfer,$Xfer,64);
495 &sub ($K_XX_XX,$K_XX_XX,16); # rewind $K_XX_XX
497 &subeq ($inp,$inp,64); # reload last block to avoid SEGV
498 &vld1_8 ("{@X[-4&7]-@X[-3&7]}","[$inp]!");
501 &vld1_8 ("{@X[-2&7]-@X[-1&7]}","[$inp]!");
504 &vld1_32 ("{$K\[]}","[$K_XX_XX,:32]!"); # load K_00_19
507 &vrev32_8 (@X[-4&7],@X[-4&7]);
509 foreach (@insns) { eval; } # remaining instructions
517 my @insns = (&$body,&$body,&$body,&$body);
520 &vrev32_8 (@X[($Xi-3)&7],@X[($Xi-3)&7]);
523 &vadd_i32 (@X[$Xi&7],@X[($Xi-4)&7],$K);
526 &vst1_32 ("{@X[$Xi&7]}","[$Xfer,:128]!");# X[]+K xfer to IALU
528 foreach (@insns) { eval; }
534 #if __ARM_MAX_ARCH__>=7
538 .type sha1_block_data_order_neon,%function
540 sha1_block_data_order_neon:
542 stmdb sp!,{r4-r12,lr}
543 add $len,$inp,$len,lsl#6 @ $len to point at the end of $inp
544 @ dmb @ errata #451034 on early Cortex A8
545 @ vstmdb sp!,{d8-d15} @ ABI specification says so
548 adr $K_XX_XX,.LK_00_19
549 bic $Xfer,$Xfer,#15 @ align for 128-bit stores
551 ldmia $ctx,{$a,$b,$c,$d,$e} @ load context
552 mov sp,$Xfer @ alloca
554 vld1.8 {@X[-4&7]-@X[-3&7]},[$inp]! @ handles unaligned
555 veor $zero,$zero,$zero
556 vld1.8 {@X[-2&7]-@X[-1&7]},[$inp]!
557 vld1.32 {${K}\[]},[$K_XX_XX,:32]! @ load K_00_19
558 vrev32.8 @X[-4&7],@X[-4&7] @ yes, even on
559 vrev32.8 @X[-3&7],@X[-3&7] @ big-endian...
560 vrev32.8 @X[-2&7],@X[-2&7]
561 vadd.i32 @X[0],@X[-4&7],$K
562 vrev32.8 @X[-1&7],@X[-1&7]
563 vadd.i32 @X[1],@X[-3&7],$K
564 vst1.32 {@X[0]},[$Xfer,:128]!
565 vadd.i32 @X[2],@X[-2&7],$K
566 vst1.32 {@X[1]},[$Xfer,:128]!
567 vst1.32 {@X[2]},[$Xfer,:128]!
568 ldr $Ki,[sp] @ big RAW stall
572 &Xupdate_16_31(\&body_00_19);
573 &Xupdate_16_31(\&body_00_19);
574 &Xupdate_16_31(\&body_00_19);
575 &Xupdate_16_31(\&body_00_19);
576 &Xupdate_32_79(\&body_00_19);
577 &Xupdate_32_79(\&body_20_39);
578 &Xupdate_32_79(\&body_20_39);
579 &Xupdate_32_79(\&body_20_39);
580 &Xupdate_32_79(\&body_20_39);
581 &Xupdate_32_79(\&body_20_39);
582 &Xupdate_32_79(\&body_40_59);
583 &Xupdate_32_79(\&body_40_59);
584 &Xupdate_32_79(\&body_40_59);
585 &Xupdate_32_79(\&body_40_59);
586 &Xupdate_32_79(\&body_40_59);
587 &Xupdate_32_79(\&body_20_39);
588 &Xuplast_80(\&body_20_39);
589 &Xloop(\&body_20_39);
590 &Xloop(\&body_20_39);
591 &Xloop(\&body_20_39);
593 ldmia $ctx,{$Ki,$t0,$t1,$Xfer} @ accumulate context
604 stmia $ctx,{$a,$b,$c,$d,$e}
609 @ vldmia sp!,{d8-d15}
610 ldmia sp!,{r4-r12,pc}
611 .size sha1_block_data_order_neon,.-sha1_block_data_order_neon
615 #####################################################################
619 my ($ABCD,$E,$E0,$E1)=map("q$_",(0..3));
620 my @MSG=map("q$_",(4..7));
621 my @Kxx=map("q$_",(8..11));
622 my ($W0,$W1,$ABCD_SAVE)=map("q$_",(12..14));
623 my $_byte = ($flavour =~ /win/ ? "DCB" : ".byte");
626 #if __ARM_MAX_ARCH__>=7
628 # if defined(__thumb2__)
629 # define INST(a,b,c,d) $_byte c,d|0xf,a,b
631 # define INST(a,b,c,d) $_byte a,b,c,d|0x10
634 .type sha1_block_data_order_armv8,%function
636 sha1_block_data_order_armv8:
638 vstmdb sp!,{d8-d15} @ ABI specification says so
642 vld1.32 {$ABCD},[$ctx]!
643 vld1.32 {$E\[0]},[$ctx]
645 vld1.32 {@Kxx[0]\[]},[r3,:32]!
646 vld1.32 {@Kxx[1]\[]},[r3,:32]!
647 vld1.32 {@Kxx[2]\[]},[r3,:32]!
648 vld1.32 {@Kxx[3]\[]},[r3,:32]
651 vld1.8 {@MSG[0]-@MSG[1]},[$inp]!
652 vld1.8 {@MSG[2]-@MSG[3]},[$inp]!
653 vrev32.8 @MSG[0],@MSG[0]
654 vrev32.8 @MSG[1],@MSG[1]
656 vadd.i32 $W0,@Kxx[0],@MSG[0]
657 vrev32.8 @MSG[2],@MSG[2]
658 vmov $ABCD_SAVE,$ABCD @ offload
661 vadd.i32 $W1,@Kxx[0],@MSG[1]
662 vrev32.8 @MSG[3],@MSG[3]
665 vadd.i32 $W0,@Kxx[$j],@MSG[2]
666 sha1su0 @MSG[0],@MSG[1],@MSG[2]
668 for ($j=0,$i=1;$i<20-3;$i++) {
669 my $f=("c","p","m","p")[$i/5];
673 vadd.i32 $W1,@Kxx[$j],@MSG[3]
674 sha1su1 @MSG[0],@MSG[3]
676 $code.=<<___ if ($i<20-4);
677 sha1su0 @MSG[1],@MSG[2],@MSG[3]
679 ($E0,$E1)=($E1,$E0); ($W0,$W1)=($W1,$W0);
680 push(@MSG,shift(@MSG)); $j++ if ((($i+3)%5)==0);
685 vadd.i32 $W1,@Kxx[$j],@MSG[3]
694 vadd.i32 $ABCD,$ABCD,$ABCD_SAVE
697 vst1.32 {$ABCD},[$ctx]!
698 vst1.32 {$E\[0]},[$ctx]
702 .size sha1_block_data_order_armv8,.-sha1_block_data_order_armv8
707 #if __ARM_MAX_ARCH__>=7
708 .comm OPENSSL_armcap_P,4,4
713 "sha1c" => 0xf2000c40, "sha1p" => 0xf2100c40,
714 "sha1m" => 0xf2200c40, "sha1su0" => 0xf2300c40,
715 "sha1h" => 0xf3b902c0, "sha1su1" => 0xf3ba0380 );
718 my ($mnemonic,$arg)=@_;
720 if ($arg =~ m/q([0-9]+)(?:,\s*q([0-9]+))?,\s*q([0-9]+)/o) {
721 my $word = $opcode{$mnemonic}|(($1&7)<<13)|(($1&8)<<19)
722 |(($2&7)<<17)|(($2&8)<<4)
723 |(($3&7)<<1) |(($3&8)<<2);
724 # since ARMv7 instructions are always encoded little-endian.
725 # correct solution is to use .inst directive, but older
726 # assemblers don't implement it:-(
728 # this fix-up provides Thumb encoding in conjunction with INST
729 $word &= ~0x10000000 if (($word & 0x0f000000) == 0x02000000);
730 sprintf "INST(0x%02x,0x%02x,0x%02x,0x%02x)\t@ %s %s",
731 $word&0xff,($word>>8)&0xff,
732 ($word>>16)&0xff,($word>>24)&0xff,
738 foreach (split($/,$code)) {
739 s/{q([0-9]+)\[\]}/sprintf "{d%d[],d%d[]}",2*$1,2*$1+1/eo or
740 s/{q([0-9]+)\[0\]}/sprintf "{d%d[0]}",2*$1/eo;
742 s/\b(sha1\w+)\s+(q.*)/unsha1($1,$2)/geo;
745 s/\bbx\s+lr\b/.word\t0xe12fff1e/o; # make it possible to compile with -march=armv4
750 close STDOUT; # enforce flush