2 # Copyright 2010-2018 The OpenSSL Project Authors. All Rights Reserved.
4 # Licensed under the Apache License 2.0 (the "License"). You may not use
5 # this file except in compliance with the License. You can obtain a copy
6 # in the file LICENSE in the source distribution or at
7 # https://www.openssl.org/source/license.html
10 # ====================================================================
11 # Written by Andy Polyakov <appro@openssl.org> for the OpenSSL
12 # project. The module is, however, dual licensed under OpenSSL and
13 # CRYPTOGAMS licenses depending on where you obtain it. For further
14 # details see http://www.openssl.org/~appro/cryptogams/.
15 # ====================================================================
19 # The module implements "4-bit" GCM GHASH function and underlying
20 # single multiplication operation in GF(2^128). "4-bit" means that it
21 # uses 256 bytes per-key table [+32 bytes shared table]. There is no
22 # experimental performance data available yet. The only approximation
23 # that can be made at this point is based on code size. Inner loop is
24 # 32 instructions long and on single-issue core should execute in <40
25 # cycles. Having verified that gcc 3.4 didn't unroll corresponding
26 # loop, this assembler loop body was found to be ~3x smaller than
27 # compiler-generated one...
31 # Rescheduling for dual-issue pipeline resulted in 8.5% improvement on
32 # Cortex A8 core and ~25 cycles per processed byte (which was observed
33 # to be ~3 times faster than gcc-generated code:-)
37 # Profiler-assisted and platform-specific optimization resulted in 7%
38 # improvement on Cortex A8 core and ~23.5 cycles per byte.
42 # Add NEON implementation featuring polynomial multiplication, i.e. no
43 # lookup tables involved. On Cortex A8 it was measured to process one
44 # byte in 15 cycles or 55% faster than integer-only code.
48 # Switch to multiplication algorithm suggested in paper referred
49 # below and combine it with reduction algorithm from x86 module.
50 # Performance improvement over previous version varies from 65% on
51 # Snapdragon S4 to 110% on Cortex A9. In absolute terms Cortex A8
52 # processes one byte in 8.45 cycles, A9 - in 10.2, A15 - in 7.63,
53 # Snapdragon S4 - in 9.33.
55 # Câmara, D.; Gouvêa, C. P. L.; López, J. & Dahab, R.: Fast Software
56 # Polynomial Multiplication on ARM Processors using the NEON Engine.
58 # http://conradoplg.cryptoland.net/files/2010/12/mocrysen13.pdf
60 # ====================================================================
61 # Note about "528B" variant. In ARM case it makes lesser sense to
62 # implement it for following reasons:
64 # - performance improvement won't be anywhere near 50%, because 128-
65 # bit shift operation is neatly fused with 128-bit xor here, and
66 # "538B" variant would eliminate only 4-5 instructions out of 32
67 # in the inner loop (meaning that estimated improvement is ~15%);
68 # - ARM-based systems are often embedded ones and extra memory
69 # consumption might be unappreciated (for so little improvement);
71 # Byte order [in]dependence. =========================================
73 # Caller is expected to maintain specific *dword* order in Htable,
74 # namely with *least* significant dword of 128-bit value at *lower*
75 # address. This differs completely from C code and has everything to
76 # do with ldm instruction and order in which dwords are "consumed" by
77 # algorithm. *Byte* order within these dwords in turn is whatever
78 # *native* byte order on current platform. See gcm128.c for working
82 if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; }
83 else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} }
85 if ($flavour && $flavour ne "void") {
86 $0 =~ m/(.*[\/\\])[^\
/\\]+$/; $dir=$1;
87 ( $xlate="${dir}arm-xlate.pl" and -f
$xlate ) or
88 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f
$xlate) or
89 die "can't locate arm-xlate.pl";
91 open STDOUT
,"| \"$^X\" $xlate $flavour $output";
93 open STDOUT
,">$output";
96 $Xi="r0"; # argument block
101 $Zll="r4"; # variables
110 ################# r13 is stack pointer
112 ################# r15 is program counter
114 $rem_4bit=$inp; # used in gcm_gmult_4bit
120 for ($Zll,$Zlh,$Zhl,$Zhh) {
122 #if __ARM_ARCH__>=7 && defined(__ARMEL__)
125 #elif defined(__ARMEB__)
131 strb
$Tlh,[$Xi,#$i+2]
133 strb
$Thl,[$Xi,#$i+1]
137 $code.="\t".shift(@args)."\n";
143 #include "arm_arch.h"
146 #if defined(__thumb2__) || defined(__clang__)
148 #define ldrplb ldrbpl
149 #define ldrneb ldrbne
151 #if defined(__thumb2__)
157 .type rem_4bit
,%object
160 .short
0x0000,0x1C20,0x3840,0x2460
161 .short
0x7080,0x6CA0,0x48C0,0x54E0
162 .short
0xE100,0xFD20,0xD940,0xC560
163 .short
0x9180,0x8DA0,0xA9C0,0xB5E0
164 .size rem_4bit
,.-rem_4bit
166 .type rem_4bit_get
,%function
168 #if defined(__thumb2__)
169 adr
$rem_4bit,rem_4bit
171 sub $rem_4bit,pc
,#8+32 @ &rem_4bit
176 .size rem_4bit_get
,.-rem_4bit_get
178 .global gcm_ghash_4bit
179 .type gcm_ghash_4bit
,%function
182 #if defined(__thumb2__)
185 sub r12
,pc
,#8+48 @ &rem_4bit
187 add
$len,$inp,$len @
$len to point at the end
188 stmdb sp
!,{r3
-r11
,lr
} @ save
$len/end too
190 ldmia r12
,{r4
-r11
} @ copy rem_4bit
...
191 stmdb sp
!,{r4
-r11
} @
... to stack
201 add
$Zhh,$Htbl,$nlo,lsl
#4
202 ldmia
$Zhh,{$Zll-$Zhh} @ load Htbl
[nlo
]
206 and $nhi,$Zll,#0xf @ rem
207 ldmia
$Thh,{$Tll-$Thh} @ load Htbl
[nhi
]
209 eor
$Zll,$Tll,$Zll,lsr
#4
210 ldrh
$Tll,[sp
,$nhi] @ rem_4bit
[rem
]
211 eor
$Zll,$Zll,$Zlh,lsl
#28
213 eor
$Zlh,$Tlh,$Zlh,lsr
#4
214 eor
$Zlh,$Zlh,$Zhl,lsl
#28
215 eor
$Zhl,$Thl,$Zhl,lsr
#4
216 eor
$Zhl,$Zhl,$Zhh,lsl
#28
217 eor
$Zhh,$Thh,$Zhh,lsr
#4
221 eor
$Zhh,$Zhh,$Tll,lsl
#16
224 add
$Thh,$Htbl,$nlo,lsl
#4
225 and $nlo,$Zll,#0xf @ rem
228 ldmia
$Thh,{$Tll-$Thh} @ load Htbl
[nlo
]
229 eor
$Zll,$Tll,$Zll,lsr
#4
230 eor
$Zll,$Zll,$Zlh,lsl
#28
231 eor
$Zlh,$Tlh,$Zlh,lsr
#4
232 eor
$Zlh,$Zlh,$Zhl,lsl
#28
233 ldrh
$Tll,[sp
,$nlo] @ rem_4bit
[rem
]
234 eor
$Zhl,$Thl,$Zhl,lsr
#4
238 ldrplb
$nlo,[$inp,$cnt]
239 eor
$Zhl,$Zhl,$Zhh,lsl
#28
240 eor
$Zhh,$Thh,$Zhh,lsr
#4
243 and $nhi,$Zll,#0xf @ rem
244 eor
$Zhh,$Zhh,$Tll,lsl
#16 @ ^= rem_4bit[rem]
246 ldmia
$Thh,{$Tll-$Thh} @ load Htbl
[nhi
]
247 eor
$Zll,$Tll,$Zll,lsr
#4
251 ldrplb
$Tll,[$Xi,$cnt]
252 eor
$Zll,$Zll,$Zlh,lsl
#28
253 eor
$Zlh,$Tlh,$Zlh,lsr
#4
255 eor
$Zlh,$Zlh,$Zhl,lsl
#28
256 eor
$Zhl,$Thl,$Zhl,lsr
#4
257 eor
$Zhl,$Zhl,$Zhh,lsl
#28
262 eor
$Zhh,$Thh,$Zhh,lsr
#4
266 andpl
$nhi,$nlo,#0xf0
267 andpl
$nlo,$nlo,#0x0f
268 eor
$Zhh,$Zhh,$Tlh,lsl
#16 @ ^= rem_4bit[rem]
271 ldr
$len,[sp
,#32] @ re-load $len/end
275 &Zsmash
("cmp\t$inp,$len","\n".
276 "#ifdef __thumb2__\n".
279 " ldrneb $nlo,[$inp,#15]");
285 ldmia sp
!,{r4
-r11
,pc
}
287 ldmia sp
!,{r4
-r11
,lr
}
289 moveq pc
,lr @ be binary compatible with V4
, yet
290 bx lr @ interoperable with Thumb ISA
:-)
292 .size gcm_ghash_4bit
,.-gcm_ghash_4bit
294 .global gcm_gmult_4bit
295 .type gcm_gmult_4bit
,%function
297 stmdb sp
!,{r4
-r11
,lr
}
305 add
$Zhh,$Htbl,$nlo,lsl
#4
306 ldmia
$Zhh,{$Zll-$Zhh} @ load Htbl
[nlo
]
310 and $nhi,$Zll,#0xf @ rem
311 ldmia
$Thh,{$Tll-$Thh} @ load Htbl
[nhi
]
313 eor
$Zll,$Tll,$Zll,lsr
#4
314 ldrh
$Tll,[$rem_4bit,$nhi] @ rem_4bit
[rem
]
315 eor
$Zll,$Zll,$Zlh,lsl
#28
316 eor
$Zlh,$Tlh,$Zlh,lsr
#4
317 eor
$Zlh,$Zlh,$Zhl,lsl
#28
318 eor
$Zhl,$Thl,$Zhl,lsr
#4
319 eor
$Zhl,$Zhl,$Zhh,lsl
#28
320 eor
$Zhh,$Thh,$Zhh,lsr
#4
322 eor
$Zhh,$Zhh,$Tll,lsl
#16
326 add
$Thh,$Htbl,$nlo,lsl
#4
327 and $nlo,$Zll,#0xf @ rem
330 ldmia
$Thh,{$Tll-$Thh} @ load Htbl
[nlo
]
331 eor
$Zll,$Tll,$Zll,lsr
#4
332 eor
$Zll,$Zll,$Zlh,lsl
#28
333 eor
$Zlh,$Tlh,$Zlh,lsr
#4
334 eor
$Zlh,$Zlh,$Zhl,lsl
#28
335 ldrh
$Tll,[$rem_4bit,$nlo] @ rem_4bit
[rem
]
336 eor
$Zhl,$Thl,$Zhl,lsr
#4
340 ldrplb
$nlo,[$Xi,$cnt]
341 eor
$Zhl,$Zhl,$Zhh,lsl
#28
342 eor
$Zhh,$Thh,$Zhh,lsr
#4
345 and $nhi,$Zll,#0xf @ rem
346 eor
$Zhh,$Zhh,$Tll,lsl
#16 @ ^= rem_4bit[rem]
348 ldmia
$Thh,{$Tll-$Thh} @ load Htbl
[nhi
]
349 eor
$Zll,$Tll,$Zll,lsr
#4
350 eor
$Zll,$Zll,$Zlh,lsl
#28
351 eor
$Zlh,$Tlh,$Zlh,lsr
#4
352 ldrh
$Tll,[$rem_4bit,$nhi] @ rem_4bit
[rem
]
353 eor
$Zlh,$Zlh,$Zhl,lsl
#28
354 eor
$Zhl,$Thl,$Zhl,lsr
#4
355 eor
$Zhl,$Zhl,$Zhh,lsl
#28
356 eor
$Zhh,$Thh,$Zhh,lsr
#4
360 andpl
$nhi,$nlo,#0xf0
361 andpl
$nlo,$nlo,#0x0f
362 eor
$Zhh,$Zhh,$Tll,lsl
#16 @ ^= rem_4bit[rem]
368 ldmia sp
!,{r4
-r11
,pc
}
370 ldmia sp
!,{r4
-r11
,lr
}
372 moveq pc
,lr @ be binary compatible with V4
, yet
373 bx lr @ interoperable with Thumb ISA
:-)
375 .size gcm_gmult_4bit
,.-gcm_gmult_4bit
378 my ($Xl,$Xm,$Xh,$IN)=map("q$_",(0..3));
379 my ($t0,$t1,$t2,$t3)=map("q$_",(8..12));
380 my ($Hlo,$Hhi,$Hhl,$k48,$k32,$k16)=map("d$_",(26..31));
385 vext
.8 $t0#lo, $a, $a, #1 @ A1
386 vmull
.p8
$t0, $t0#lo, $b @ F = A1*B
387 vext
.8 $r#lo, $b, $b, #1 @ B1
388 vmull
.p8
$r, $a, $r#lo @ E = A*B1
389 vext
.8 $t1#lo, $a, $a, #2 @ A2
390 vmull
.p8
$t1, $t1#lo, $b @ H = A2*B
391 vext
.8 $t3#lo, $b, $b, #2 @ B2
392 vmull
.p8
$t3, $a, $t3#lo @ G = A*B2
393 vext
.8 $t2#lo, $a, $a, #3 @ A3
394 veor
$t0, $t0, $r @ L
= E
+ F
395 vmull
.p8
$t2, $t2#lo, $b @ J = A3*B
396 vext
.8 $r#lo, $b, $b, #3 @ B3
397 veor
$t1, $t1, $t3 @ M
= G
+ H
398 vmull
.p8
$r, $a, $r#lo @ I = A*B3
399 veor
$t0#lo, $t0#lo, $t0#hi @ t0 = (L) (P0 + P1) << 8
400 vand
$t0#hi, $t0#hi, $k48
401 vext
.8 $t3#lo, $b, $b, #4 @ B4
402 veor
$t1#lo, $t1#lo, $t1#hi @ t1 = (M) (P2 + P3) << 16
403 vand
$t1#hi, $t1#hi, $k32
404 vmull
.p8
$t3, $a, $t3#lo @ K = A*B4
405 veor
$t2, $t2, $r @ N
= I
+ J
406 veor
$t0#lo, $t0#lo, $t0#hi
407 veor
$t1#lo, $t1#lo, $t1#hi
408 veor
$t2#lo, $t2#lo, $t2#hi @ t2 = (N) (P4 + P5) << 24
409 vand
$t2#hi, $t2#hi, $k16
410 vext
.8 $t0, $t0, $t0, #15
411 veor
$t3#lo, $t3#lo, $t3#hi @ t3 = (K) (P6 + P7) << 32
413 vext
.8 $t1, $t1, $t1, #14
414 veor
$t2#lo, $t2#lo, $t2#hi
415 vmull
.p8
$r, $a, $b @ D
= A
*B
416 vext
.8 $t3, $t3, $t3, #12
417 vext
.8 $t2, $t2, $t2, #13
426 #if __ARM_MAX_ARCH__>=7
430 .global gcm_init_neon
431 .type gcm_init_neon
,%function
434 vld1
.64
$IN#hi,[r1]! @ load H
438 vshr
.u64
$t0#lo,#63 @ t0=0xc2....01
440 vshr
.u64
$Hlo,$IN#lo,#63
441 vshr
.s8
$t1,#7 @ broadcast carry bit
444 vorr
$IN#hi,$Hlo @ H<<<=1
445 veor
$IN,$IN,$t0 @ twisted H
449 .size gcm_init_neon
,.-gcm_init_neon
451 .global gcm_gmult_neon
452 .type gcm_gmult_neon
,%function
455 vld1
.64
$IN#hi,[$Xi]! @ load Xi
456 vld1
.64
$IN#lo,[$Xi]!
457 vmov
.i64
$k48,#0x0000ffffffffffff
458 vldmia
$Htbl,{$Hlo-$Hhi} @ load twisted H
459 vmov
.i64
$k32,#0x00000000ffffffff
463 vmov
.i64
$k16,#0x000000000000ffff
464 veor
$Hhl,$Hlo,$Hhi @ Karatsuba pre
-processing
467 .size gcm_gmult_neon
,.-gcm_gmult_neon
469 .global gcm_ghash_neon
470 .type gcm_ghash_neon
,%function
473 vld1
.64
$Xl#hi,[$Xi]! @ load Xi
474 vld1
.64
$Xl#lo,[$Xi]!
475 vmov
.i64
$k48,#0x0000ffffffffffff
476 vldmia
$Htbl,{$Hlo-$Hhi} @ load twisted H
477 vmov
.i64
$k32,#0x00000000ffffffff
481 vmov
.i64
$k16,#0x000000000000ffff
482 veor
$Hhl,$Hlo,$Hhi @ Karatsuba pre
-processing
485 vld1
.64
$IN#hi,[$inp]! @ load inp
486 vld1
.64
$IN#lo,[$inp]!
490 veor
$IN,$Xl @ inp
^=Xi
493 &clmul64x64
($Xl,$Hlo,"$IN#lo"); # H.lo·Xi.lo
495 veor
$IN#lo,$IN#lo,$IN#hi @ Karatsuba pre-processing
497 &clmul64x64
($Xm,$Hhl,"$IN#lo"); # (H.lo+H.hi)·(Xi.lo+Xi.hi)
498 &clmul64x64
($Xh,$Hhi,"$IN#hi"); # H.hi·Xi.hi
500 veor
$Xm,$Xm,$Xl @ Karatsuba post
-processing
502 veor
$Xl#hi,$Xl#hi,$Xm#lo
503 veor
$Xh#lo,$Xh#lo,$Xm#hi @ Xh|Xl - 256-bit result
505 @ equivalent of reduction_avx from ghash
-x86_64
.pl
506 vshl
.i64
$t1,$Xl,#57 @ 1st phase
511 veor
$Xl#hi,$Xl#hi,$t2#lo @
512 veor
$Xh#lo,$Xh#lo,$t2#hi
514 vshr
.u64
$t2,$Xl,#1 @ 2nd phase
518 vshr
.u64
$Xl,$Xl,#1 @
529 vst1
.64
$Xl#hi,[$Xi]! @ write out Xi
533 .size gcm_ghash_neon
,.-gcm_ghash_neon
538 .asciz
"GHASH for ARMv4/NEON, CRYPTOGAMS by <appro\@openssl.org>"
542 foreach (split("\n",$code)) {
543 s/\`([^\`]*)\`/eval $1/geo;
545 s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo or
546 s/\bret\b/bx lr/go or
547 s/\bbx\s+lr\b/.word\t0xe12fff1e/go; # make it possible to compile with -march=armv4
551 close STDOUT
; # enforce flush