1#! /usr/bin/env perl 2# Copyright 2010-2018 The OpenSSL Project Authors. All Rights Reserved. 3# 4# Licensed under the OpenSSL license (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 8 9# 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# ==================================================================== 16# 17# April 2010 18# 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... 28# 29# July 2010 30# 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:-) 34# 35# February 2011 36# 37# Profiler-assisted and platform-specific optimization resulted in 7% 38# improvement on Cortex A8 core and ~23.5 cycles per byte. 39# 40# March 2011 41# 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. 45# 46# April 2014 47# 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. 54# 55# Câmara, D.; Gouvêa, C. P. L.; López, J. & Dahab, R.: Fast Software 56# Polynomial Multiplication on ARM Processors using the NEON Engine. 57# 58# http://conradoplg.cryptoland.net/files/2010/12/mocrysen13.pdf 59 60# ==================================================================== 61# Note about "528B" variant. In ARM case it makes lesser sense to 62# implement it for following reasons: 63# 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); 70# 71# Byte order [in]dependence. ========================================= 72# 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 79# example... 80 81$flavour = shift; 82if ($flavour=~/\w[\w\-]*\.\w+$/) { $output=$flavour; undef $flavour; } 83else { while (($output=shift) && ($output!~/\w[\w\-]*\.\w+$/)) {} } 84 85if ($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 $xlate) or 89 die "can't locate arm-xlate.pl"; 90 91 open STDOUT,"| \"$^X\" $xlate $flavour $output"; 92} else { 93 open STDOUT,">$output"; 94} 95 96$Xi="r0"; # argument block 97$Htbl="r1"; 98$inp="r2"; 99$len="r3"; 100 101$Zll="r4"; # variables 102$Zlh="r5"; 103$Zhl="r6"; 104$Zhh="r7"; 105$Tll="r8"; 106$Tlh="r9"; 107$Thl="r10"; 108$Thh="r11"; 109$nlo="r12"; 110################# r13 is stack pointer 111$nhi="r14"; 112################# r15 is program counter 113 114$rem_4bit=$inp; # used in gcm_gmult_4bit 115$cnt=$len; 116 117sub Zsmash() { 118 my $i=12; 119 my @args=@_; 120 for ($Zll,$Zlh,$Zhl,$Zhh) { 121 $code.=<<___; 122#if __ARM_ARCH__>=7 && defined(__ARMEL__) 123 rev $_,$_ 124 str $_,[$Xi,#$i] 125#elif defined(__ARMEB__) 126 str $_,[$Xi,#$i] 127#else 128 mov $Tlh,$_,lsr#8 129 strb $_,[$Xi,#$i+3] 130 mov $Thl,$_,lsr#16 131 strb $Tlh,[$Xi,#$i+2] 132 mov $Thh,$_,lsr#24 133 strb $Thl,[$Xi,#$i+1] 134 strb $Thh,[$Xi,#$i] 135#endif 136___ 137 $code.="\t".shift(@args)."\n"; 138 $i-=4; 139 } 140} 141 142$code=<<___; 143#include <openssl/arm_arch.h> 144 145@ Silence ARMv8 deprecated IT instruction warnings. This file is used by both 146@ ARMv7 and ARMv8 processors and does not use ARMv8 instructions. (ARMv8 PMULL 147@ instructions are in aesv8-armx.pl.) 148.arch armv7-a 149 150.text 151#if defined(__thumb2__) || defined(__clang__) 152.syntax unified 153#define ldrplb ldrbpl 154#define ldrneb ldrbne 155#endif 156#if defined(__thumb2__) 157.thumb 158#else 159.code 32 160#endif 161 162.type rem_4bit,%object 163.align 5 164rem_4bit: 165.short 0x0000,0x1C20,0x3840,0x2460 166.short 0x7080,0x6CA0,0x48C0,0x54E0 167.short 0xE100,0xFD20,0xD940,0xC560 168.short 0x9180,0x8DA0,0xA9C0,0xB5E0 169.size rem_4bit,.-rem_4bit 170 171.type rem_4bit_get,%function 172rem_4bit_get: 173#if defined(__thumb2__) 174 adr $rem_4bit,rem_4bit 175#else 176 sub $rem_4bit,pc,#8+32 @ &rem_4bit 177#endif 178 b .Lrem_4bit_got 179 nop 180 nop 181.size rem_4bit_get,.-rem_4bit_get 182 183.global gcm_ghash_4bit 184.type gcm_ghash_4bit,%function 185.align 4 186gcm_ghash_4bit: 187#if defined(__thumb2__) 188 adr r12,rem_4bit 189#else 190 sub r12,pc,#8+48 @ &rem_4bit 191#endif 192 add $len,$inp,$len @ $len to point at the end 193 stmdb sp!,{r3-r11,lr} @ save $len/end too 194 195 ldmia r12,{r4-r11} @ copy rem_4bit ... 196 stmdb sp!,{r4-r11} @ ... to stack 197 198 ldrb $nlo,[$inp,#15] 199 ldrb $nhi,[$Xi,#15] 200.Louter: 201 eor $nlo,$nlo,$nhi 202 and $nhi,$nlo,#0xf0 203 and $nlo,$nlo,#0x0f 204 mov $cnt,#14 205 206 add $Zhh,$Htbl,$nlo,lsl#4 207 ldmia $Zhh,{$Zll-$Zhh} @ load Htbl[nlo] 208 add $Thh,$Htbl,$nhi 209 ldrb $nlo,[$inp,#14] 210 211 and $nhi,$Zll,#0xf @ rem 212 ldmia $Thh,{$Tll-$Thh} @ load Htbl[nhi] 213 add $nhi,$nhi,$nhi 214 eor $Zll,$Tll,$Zll,lsr#4 215 ldrh $Tll,[sp,$nhi] @ rem_4bit[rem] 216 eor $Zll,$Zll,$Zlh,lsl#28 217 ldrb $nhi,[$Xi,#14] 218 eor $Zlh,$Tlh,$Zlh,lsr#4 219 eor $Zlh,$Zlh,$Zhl,lsl#28 220 eor $Zhl,$Thl,$Zhl,lsr#4 221 eor $Zhl,$Zhl,$Zhh,lsl#28 222 eor $Zhh,$Thh,$Zhh,lsr#4 223 eor $nlo,$nlo,$nhi 224 and $nhi,$nlo,#0xf0 225 and $nlo,$nlo,#0x0f 226 eor $Zhh,$Zhh,$Tll,lsl#16 227 228.Linner: 229 add $Thh,$Htbl,$nlo,lsl#4 230 and $nlo,$Zll,#0xf @ rem 231 subs $cnt,$cnt,#1 232 add $nlo,$nlo,$nlo 233 ldmia $Thh,{$Tll-$Thh} @ load Htbl[nlo] 234 eor $Zll,$Tll,$Zll,lsr#4 235 eor $Zll,$Zll,$Zlh,lsl#28 236 eor $Zlh,$Tlh,$Zlh,lsr#4 237 eor $Zlh,$Zlh,$Zhl,lsl#28 238 ldrh $Tll,[sp,$nlo] @ rem_4bit[rem] 239 eor $Zhl,$Thl,$Zhl,lsr#4 240#ifdef __thumb2__ 241 it pl 242#endif 243 ldrplb $nlo,[$inp,$cnt] 244 eor $Zhl,$Zhl,$Zhh,lsl#28 245 eor $Zhh,$Thh,$Zhh,lsr#4 246 247 add $Thh,$Htbl,$nhi 248 and $nhi,$Zll,#0xf @ rem 249 eor $Zhh,$Zhh,$Tll,lsl#16 @ ^= rem_4bit[rem] 250 add $nhi,$nhi,$nhi 251 ldmia $Thh,{$Tll-$Thh} @ load Htbl[nhi] 252 eor $Zll,$Tll,$Zll,lsr#4 253#ifdef __thumb2__ 254 it pl 255#endif 256 ldrplb $Tll,[$Xi,$cnt] 257 eor $Zll,$Zll,$Zlh,lsl#28 258 eor $Zlh,$Tlh,$Zlh,lsr#4 259 ldrh $Tlh,[sp,$nhi] 260 eor $Zlh,$Zlh,$Zhl,lsl#28 261 eor $Zhl,$Thl,$Zhl,lsr#4 262 eor $Zhl,$Zhl,$Zhh,lsl#28 263#ifdef __thumb2__ 264 it pl 265#endif 266 eorpl $nlo,$nlo,$Tll 267 eor $Zhh,$Thh,$Zhh,lsr#4 268#ifdef __thumb2__ 269 itt pl 270#endif 271 andpl $nhi,$nlo,#0xf0 272 andpl $nlo,$nlo,#0x0f 273 eor $Zhh,$Zhh,$Tlh,lsl#16 @ ^= rem_4bit[rem] 274 bpl .Linner 275 276 ldr $len,[sp,#32] @ re-load $len/end 277 add $inp,$inp,#16 278 mov $nhi,$Zll 279___ 280 &Zsmash("cmp\t$inp,$len","\n". 281 "#ifdef __thumb2__\n". 282 " it ne\n". 283 "#endif\n". 284 " ldrneb $nlo,[$inp,#15]"); 285$code.=<<___; 286 bne .Louter 287 288 add sp,sp,#36 289#if __ARM_ARCH__>=5 290 ldmia sp!,{r4-r11,pc} 291#else 292 ldmia sp!,{r4-r11,lr} 293 tst lr,#1 294 moveq pc,lr @ be binary compatible with V4, yet 295 bx lr @ interoperable with Thumb ISA:-) 296#endif 297.size gcm_ghash_4bit,.-gcm_ghash_4bit 298 299.global gcm_gmult_4bit 300.type gcm_gmult_4bit,%function 301gcm_gmult_4bit: 302 stmdb sp!,{r4-r11,lr} 303 ldrb $nlo,[$Xi,#15] 304 b rem_4bit_get 305.Lrem_4bit_got: 306 and $nhi,$nlo,#0xf0 307 and $nlo,$nlo,#0x0f 308 mov $cnt,#14 309 310 add $Zhh,$Htbl,$nlo,lsl#4 311 ldmia $Zhh,{$Zll-$Zhh} @ load Htbl[nlo] 312 ldrb $nlo,[$Xi,#14] 313 314 add $Thh,$Htbl,$nhi 315 and $nhi,$Zll,#0xf @ rem 316 ldmia $Thh,{$Tll-$Thh} @ load Htbl[nhi] 317 add $nhi,$nhi,$nhi 318 eor $Zll,$Tll,$Zll,lsr#4 319 ldrh $Tll,[$rem_4bit,$nhi] @ rem_4bit[rem] 320 eor $Zll,$Zll,$Zlh,lsl#28 321 eor $Zlh,$Tlh,$Zlh,lsr#4 322 eor $Zlh,$Zlh,$Zhl,lsl#28 323 eor $Zhl,$Thl,$Zhl,lsr#4 324 eor $Zhl,$Zhl,$Zhh,lsl#28 325 eor $Zhh,$Thh,$Zhh,lsr#4 326 and $nhi,$nlo,#0xf0 327 eor $Zhh,$Zhh,$Tll,lsl#16 328 and $nlo,$nlo,#0x0f 329 330.Loop: 331 add $Thh,$Htbl,$nlo,lsl#4 332 and $nlo,$Zll,#0xf @ rem 333 subs $cnt,$cnt,#1 334 add $nlo,$nlo,$nlo 335 ldmia $Thh,{$Tll-$Thh} @ load Htbl[nlo] 336 eor $Zll,$Tll,$Zll,lsr#4 337 eor $Zll,$Zll,$Zlh,lsl#28 338 eor $Zlh,$Tlh,$Zlh,lsr#4 339 eor $Zlh,$Zlh,$Zhl,lsl#28 340 ldrh $Tll,[$rem_4bit,$nlo] @ rem_4bit[rem] 341 eor $Zhl,$Thl,$Zhl,lsr#4 342#ifdef __thumb2__ 343 it pl 344#endif 345 ldrplb $nlo,[$Xi,$cnt] 346 eor $Zhl,$Zhl,$Zhh,lsl#28 347 eor $Zhh,$Thh,$Zhh,lsr#4 348 349 add $Thh,$Htbl,$nhi 350 and $nhi,$Zll,#0xf @ rem 351 eor $Zhh,$Zhh,$Tll,lsl#16 @ ^= rem_4bit[rem] 352 add $nhi,$nhi,$nhi 353 ldmia $Thh,{$Tll-$Thh} @ load Htbl[nhi] 354 eor $Zll,$Tll,$Zll,lsr#4 355 eor $Zll,$Zll,$Zlh,lsl#28 356 eor $Zlh,$Tlh,$Zlh,lsr#4 357 ldrh $Tll,[$rem_4bit,$nhi] @ rem_4bit[rem] 358 eor $Zlh,$Zlh,$Zhl,lsl#28 359 eor $Zhl,$Thl,$Zhl,lsr#4 360 eor $Zhl,$Zhl,$Zhh,lsl#28 361 eor $Zhh,$Thh,$Zhh,lsr#4 362#ifdef __thumb2__ 363 itt pl 364#endif 365 andpl $nhi,$nlo,#0xf0 366 andpl $nlo,$nlo,#0x0f 367 eor $Zhh,$Zhh,$Tll,lsl#16 @ ^= rem_4bit[rem] 368 bpl .Loop 369___ 370 &Zsmash(); 371$code.=<<___; 372#if __ARM_ARCH__>=5 373 ldmia sp!,{r4-r11,pc} 374#else 375 ldmia sp!,{r4-r11,lr} 376 tst lr,#1 377 moveq pc,lr @ be binary compatible with V4, yet 378 bx lr @ interoperable with Thumb ISA:-) 379#endif 380.size gcm_gmult_4bit,.-gcm_gmult_4bit 381___ 382{ 383my ($Xl,$Xm,$Xh,$IN)=map("q$_",(0..3)); 384my ($t0,$t1,$t2,$t3)=map("q$_",(8..12)); 385my ($Hlo,$Hhi,$Hhl,$k48,$k32,$k16)=map("d$_",(26..31)); 386 387sub clmul64x64 { 388my ($r,$a,$b)=@_; 389$code.=<<___; 390 vext.8 $t0#lo, $a, $a, #1 @ A1 391 vmull.p8 $t0, $t0#lo, $b @ F = A1*B 392 vext.8 $r#lo, $b, $b, #1 @ B1 393 vmull.p8 $r, $a, $r#lo @ E = A*B1 394 vext.8 $t1#lo, $a, $a, #2 @ A2 395 vmull.p8 $t1, $t1#lo, $b @ H = A2*B 396 vext.8 $t3#lo, $b, $b, #2 @ B2 397 vmull.p8 $t3, $a, $t3#lo @ G = A*B2 398 vext.8 $t2#lo, $a, $a, #3 @ A3 399 veor $t0, $t0, $r @ L = E + F 400 vmull.p8 $t2, $t2#lo, $b @ J = A3*B 401 vext.8 $r#lo, $b, $b, #3 @ B3 402 veor $t1, $t1, $t3 @ M = G + H 403 vmull.p8 $r, $a, $r#lo @ I = A*B3 404 veor $t0#lo, $t0#lo, $t0#hi @ t0 = (L) (P0 + P1) << 8 405 vand $t0#hi, $t0#hi, $k48 406 vext.8 $t3#lo, $b, $b, #4 @ B4 407 veor $t1#lo, $t1#lo, $t1#hi @ t1 = (M) (P2 + P3) << 16 408 vand $t1#hi, $t1#hi, $k32 409 vmull.p8 $t3, $a, $t3#lo @ K = A*B4 410 veor $t2, $t2, $r @ N = I + J 411 veor $t0#lo, $t0#lo, $t0#hi 412 veor $t1#lo, $t1#lo, $t1#hi 413 veor $t2#lo, $t2#lo, $t2#hi @ t2 = (N) (P4 + P5) << 24 414 vand $t2#hi, $t2#hi, $k16 415 vext.8 $t0, $t0, $t0, #15 416 veor $t3#lo, $t3#lo, $t3#hi @ t3 = (K) (P6 + P7) << 32 417 vmov.i64 $t3#hi, #0 418 vext.8 $t1, $t1, $t1, #14 419 veor $t2#lo, $t2#lo, $t2#hi 420 vmull.p8 $r, $a, $b @ D = A*B 421 vext.8 $t3, $t3, $t3, #12 422 vext.8 $t2, $t2, $t2, #13 423 veor $t0, $t0, $t1 424 veor $t2, $t2, $t3 425 veor $r, $r, $t0 426 veor $r, $r, $t2 427___ 428} 429 430$code.=<<___; 431#if __ARM_MAX_ARCH__>=7 432.arch armv7-a 433.fpu neon 434 435.global gcm_init_neon 436.type gcm_init_neon,%function 437.align 4 438gcm_init_neon: 439 vld1.64 $IN#hi,[r1]! @ load H 440 vmov.i8 $t0,#0xe1 441 vld1.64 $IN#lo,[r1] 442 vshl.i64 $t0#hi,#57 443 vshr.u64 $t0#lo,#63 @ t0=0xc2....01 444 vdup.8 $t1,$IN#hi[7] 445 vshr.u64 $Hlo,$IN#lo,#63 446 vshr.s8 $t1,#7 @ broadcast carry bit 447 vshl.i64 $IN,$IN,#1 448 vand $t0,$t0,$t1 449 vorr $IN#hi,$Hlo @ H<<<=1 450 veor $IN,$IN,$t0 @ twisted H 451 vstmia r0,{$IN} 452 453 ret @ bx lr 454.size gcm_init_neon,.-gcm_init_neon 455 456.global gcm_gmult_neon 457.type gcm_gmult_neon,%function 458.align 4 459gcm_gmult_neon: 460 vld1.64 $IN#hi,[$Xi]! @ load Xi 461 vld1.64 $IN#lo,[$Xi]! 462 vmov.i64 $k48,#0x0000ffffffffffff 463 vldmia $Htbl,{$Hlo-$Hhi} @ load twisted H 464 vmov.i64 $k32,#0x00000000ffffffff 465#ifdef __ARMEL__ 466 vrev64.8 $IN,$IN 467#endif 468 vmov.i64 $k16,#0x000000000000ffff 469 veor $Hhl,$Hlo,$Hhi @ Karatsuba pre-processing 470 mov $len,#16 471 b .Lgmult_neon 472.size gcm_gmult_neon,.-gcm_gmult_neon 473 474.global gcm_ghash_neon 475.type gcm_ghash_neon,%function 476.align 4 477gcm_ghash_neon: 478 vld1.64 $Xl#hi,[$Xi]! @ load Xi 479 vld1.64 $Xl#lo,[$Xi]! 480 vmov.i64 $k48,#0x0000ffffffffffff 481 vldmia $Htbl,{$Hlo-$Hhi} @ load twisted H 482 vmov.i64 $k32,#0x00000000ffffffff 483#ifdef __ARMEL__ 484 vrev64.8 $Xl,$Xl 485#endif 486 vmov.i64 $k16,#0x000000000000ffff 487 veor $Hhl,$Hlo,$Hhi @ Karatsuba pre-processing 488 489.Loop_neon: 490 vld1.64 $IN#hi,[$inp]! @ load inp 491 vld1.64 $IN#lo,[$inp]! 492#ifdef __ARMEL__ 493 vrev64.8 $IN,$IN 494#endif 495 veor $IN,$Xl @ inp^=Xi 496.Lgmult_neon: 497___ 498 &clmul64x64 ($Xl,$Hlo,"$IN#lo"); # H.lo·Xi.lo 499$code.=<<___; 500 veor $IN#lo,$IN#lo,$IN#hi @ Karatsuba pre-processing 501___ 502 &clmul64x64 ($Xm,$Hhl,"$IN#lo"); # (H.lo+H.hi)·(Xi.lo+Xi.hi) 503 &clmul64x64 ($Xh,$Hhi,"$IN#hi"); # H.hi·Xi.hi 504$code.=<<___; 505 veor $Xm,$Xm,$Xl @ Karatsuba post-processing 506 veor $Xm,$Xm,$Xh 507 veor $Xl#hi,$Xl#hi,$Xm#lo 508 veor $Xh#lo,$Xh#lo,$Xm#hi @ Xh|Xl - 256-bit result 509 510 @ equivalent of reduction_avx from ghash-x86_64.pl 511 vshl.i64 $t1,$Xl,#57 @ 1st phase 512 vshl.i64 $t2,$Xl,#62 513 veor $t2,$t2,$t1 @ 514 vshl.i64 $t1,$Xl,#63 515 veor $t2, $t2, $t1 @ 516 veor $Xl#hi,$Xl#hi,$t2#lo @ 517 veor $Xh#lo,$Xh#lo,$t2#hi 518 519 vshr.u64 $t2,$Xl,#1 @ 2nd phase 520 veor $Xh,$Xh,$Xl 521 veor $Xl,$Xl,$t2 @ 522 vshr.u64 $t2,$t2,#6 523 vshr.u64 $Xl,$Xl,#1 @ 524 veor $Xl,$Xl,$Xh @ 525 veor $Xl,$Xl,$t2 @ 526 527 subs $len,#16 528 bne .Loop_neon 529 530#ifdef __ARMEL__ 531 vrev64.8 $Xl,$Xl 532#endif 533 sub $Xi,#16 534 vst1.64 $Xl#hi,[$Xi]! @ write out Xi 535 vst1.64 $Xl#lo,[$Xi] 536 537 ret @ bx lr 538.size gcm_ghash_neon,.-gcm_ghash_neon 539#endif 540___ 541} 542$code.=<<___; 543.asciz "GHASH for ARMv4/NEON, CRYPTOGAMS by <appro\@openssl.org>" 544.align 2 545___ 546 547foreach (split("\n",$code)) { 548 s/\`([^\`]*)\`/eval $1/geo; 549 550 s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo or 551 s/\bret\b/bx lr/go or 552 s/\bbx\s+lr\b/.word\t0xe12fff1e/go; # make it possible to compile with -march=armv4 553 554 print $_,"\n"; 555} 556close STDOUT; # enforce flush 557