1#! /usr/bin/env perl 2# Copyright 2016 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 <[email protected]> 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# June 2015 18# 19# ChaCha20 for ARMv8. 20# 21# Performance in cycles per byte out of large buffer. 22# 23# IALU/gcc-4.9 3xNEON+1xIALU 6xNEON+2xIALU 24# 25# Apple A7 5.50/+49% 3.33 1.70 26# Cortex-A53 8.40/+80% 4.72 4.72(*) 27# Cortex-A57 8.06/+43% 4.90 4.43(**) 28# Denver 4.50/+82% 2.63 2.67(*) 29# X-Gene 9.50/+46% 8.82 8.89(*) 30# Mongoose 8.00/+44% 3.64 3.25 31# Kryo 8.17/+50% 4.83 4.65 32# 33# (*) it's expected that doubling interleave factor doesn't help 34# all processors, only those with higher NEON latency and 35# higher instruction issue rate; 36# (**) expected improvement was actually higher; 37 38$flavour=shift; 39$output=shift; 40 41$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 42( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or 43( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or 44die "can't locate arm-xlate.pl"; 45 46open OUT,"| \"$^X\" \"$xlate\" $flavour \"$output\""; 47*STDOUT=*OUT; 48 49sub AUTOLOAD() # thunk [simplified] x86-style perlasm 50{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./; 51 my $arg = pop; 52 $arg = "#$arg" if ($arg*1 eq $arg); 53 $code .= "\t$opcode\t".join(',',@_,$arg)."\n"; 54} 55 56my ($out,$inp,$len,$key,$ctr) = map("x$_",(0..4)); 57 58my @x=map("x$_",(5..17,19..21)); 59my @d=map("x$_",(22..28,30)); 60 61sub ROUND { 62my ($a0,$b0,$c0,$d0)=@_; 63my ($a1,$b1,$c1,$d1)=map(($_&~3)+(($_+1)&3),($a0,$b0,$c0,$d0)); 64my ($a2,$b2,$c2,$d2)=map(($_&~3)+(($_+1)&3),($a1,$b1,$c1,$d1)); 65my ($a3,$b3,$c3,$d3)=map(($_&~3)+(($_+1)&3),($a2,$b2,$c2,$d2)); 66 67 ( 68 "&add_32 (@x[$a0],@x[$a0],@x[$b0])", 69 "&add_32 (@x[$a1],@x[$a1],@x[$b1])", 70 "&add_32 (@x[$a2],@x[$a2],@x[$b2])", 71 "&add_32 (@x[$a3],@x[$a3],@x[$b3])", 72 "&eor_32 (@x[$d0],@x[$d0],@x[$a0])", 73 "&eor_32 (@x[$d1],@x[$d1],@x[$a1])", 74 "&eor_32 (@x[$d2],@x[$d2],@x[$a2])", 75 "&eor_32 (@x[$d3],@x[$d3],@x[$a3])", 76 "&ror_32 (@x[$d0],@x[$d0],16)", 77 "&ror_32 (@x[$d1],@x[$d1],16)", 78 "&ror_32 (@x[$d2],@x[$d2],16)", 79 "&ror_32 (@x[$d3],@x[$d3],16)", 80 81 "&add_32 (@x[$c0],@x[$c0],@x[$d0])", 82 "&add_32 (@x[$c1],@x[$c1],@x[$d1])", 83 "&add_32 (@x[$c2],@x[$c2],@x[$d2])", 84 "&add_32 (@x[$c3],@x[$c3],@x[$d3])", 85 "&eor_32 (@x[$b0],@x[$b0],@x[$c0])", 86 "&eor_32 (@x[$b1],@x[$b1],@x[$c1])", 87 "&eor_32 (@x[$b2],@x[$b2],@x[$c2])", 88 "&eor_32 (@x[$b3],@x[$b3],@x[$c3])", 89 "&ror_32 (@x[$b0],@x[$b0],20)", 90 "&ror_32 (@x[$b1],@x[$b1],20)", 91 "&ror_32 (@x[$b2],@x[$b2],20)", 92 "&ror_32 (@x[$b3],@x[$b3],20)", 93 94 "&add_32 (@x[$a0],@x[$a0],@x[$b0])", 95 "&add_32 (@x[$a1],@x[$a1],@x[$b1])", 96 "&add_32 (@x[$a2],@x[$a2],@x[$b2])", 97 "&add_32 (@x[$a3],@x[$a3],@x[$b3])", 98 "&eor_32 (@x[$d0],@x[$d0],@x[$a0])", 99 "&eor_32 (@x[$d1],@x[$d1],@x[$a1])", 100 "&eor_32 (@x[$d2],@x[$d2],@x[$a2])", 101 "&eor_32 (@x[$d3],@x[$d3],@x[$a3])", 102 "&ror_32 (@x[$d0],@x[$d0],24)", 103 "&ror_32 (@x[$d1],@x[$d1],24)", 104 "&ror_32 (@x[$d2],@x[$d2],24)", 105 "&ror_32 (@x[$d3],@x[$d3],24)", 106 107 "&add_32 (@x[$c0],@x[$c0],@x[$d0])", 108 "&add_32 (@x[$c1],@x[$c1],@x[$d1])", 109 "&add_32 (@x[$c2],@x[$c2],@x[$d2])", 110 "&add_32 (@x[$c3],@x[$c3],@x[$d3])", 111 "&eor_32 (@x[$b0],@x[$b0],@x[$c0])", 112 "&eor_32 (@x[$b1],@x[$b1],@x[$c1])", 113 "&eor_32 (@x[$b2],@x[$b2],@x[$c2])", 114 "&eor_32 (@x[$b3],@x[$b3],@x[$c3])", 115 "&ror_32 (@x[$b0],@x[$b0],25)", 116 "&ror_32 (@x[$b1],@x[$b1],25)", 117 "&ror_32 (@x[$b2],@x[$b2],25)", 118 "&ror_32 (@x[$b3],@x[$b3],25)" 119 ); 120} 121 122$code.=<<___; 123#include <openssl/arm_arch.h> 124 125.section .rodata 126 127.align 5 128.Lsigma: 129.quad 0x3320646e61707865,0x6b20657479622d32 // endian-neutral 130.Lone: 131.long 1,0,0,0 132.asciz "ChaCha20 for ARMv8, CRYPTOGAMS by <appro\@openssl.org>" 133 134.text 135 136.globl ChaCha20_ctr32_nohw 137.type ChaCha20_ctr32_nohw,%function 138.align 5 139ChaCha20_ctr32_nohw: 140 AARCH64_SIGN_LINK_REGISTER 141 stp x29,x30,[sp,#-96]! 142 add x29,sp,#0 143 144 adrp @x[0],:pg_hi21:.Lsigma 145 add @x[0],@x[0],:lo12:.Lsigma 146 stp x19,x20,[sp,#16] 147 stp x21,x22,[sp,#32] 148 stp x23,x24,[sp,#48] 149 stp x25,x26,[sp,#64] 150 stp x27,x28,[sp,#80] 151 sub sp,sp,#64 152 153 ldp @d[0],@d[1],[@x[0]] // load sigma 154 ldp @d[2],@d[3],[$key] // load key 155 ldp @d[4],@d[5],[$key,#16] 156 ldp @d[6],@d[7],[$ctr] // load counter 157#ifdef __AARCH64EB__ 158 ror @d[2],@d[2],#32 159 ror @d[3],@d[3],#32 160 ror @d[4],@d[4],#32 161 ror @d[5],@d[5],#32 162 ror @d[6],@d[6],#32 163 ror @d[7],@d[7],#32 164#endif 165 166.Loop_outer: 167 mov.32 @x[0],@d[0] // unpack key block 168 lsr @x[1],@d[0],#32 169 mov.32 @x[2],@d[1] 170 lsr @x[3],@d[1],#32 171 mov.32 @x[4],@d[2] 172 lsr @x[5],@d[2],#32 173 mov.32 @x[6],@d[3] 174 lsr @x[7],@d[3],#32 175 mov.32 @x[8],@d[4] 176 lsr @x[9],@d[4],#32 177 mov.32 @x[10],@d[5] 178 lsr @x[11],@d[5],#32 179 mov.32 @x[12],@d[6] 180 lsr @x[13],@d[6],#32 181 mov.32 @x[14],@d[7] 182 lsr @x[15],@d[7],#32 183 184 mov $ctr,#10 185 subs $len,$len,#64 186.Loop: 187 sub $ctr,$ctr,#1 188___ 189 foreach (&ROUND(0, 4, 8,12)) { eval; } 190 foreach (&ROUND(0, 5,10,15)) { eval; } 191$code.=<<___; 192 cbnz $ctr,.Loop 193 194 add.32 @x[0],@x[0],@d[0] // accumulate key block 195 add @x[1],@x[1],@d[0],lsr#32 196 add.32 @x[2],@x[2],@d[1] 197 add @x[3],@x[3],@d[1],lsr#32 198 add.32 @x[4],@x[4],@d[2] 199 add @x[5],@x[5],@d[2],lsr#32 200 add.32 @x[6],@x[6],@d[3] 201 add @x[7],@x[7],@d[3],lsr#32 202 add.32 @x[8],@x[8],@d[4] 203 add @x[9],@x[9],@d[4],lsr#32 204 add.32 @x[10],@x[10],@d[5] 205 add @x[11],@x[11],@d[5],lsr#32 206 add.32 @x[12],@x[12],@d[6] 207 add @x[13],@x[13],@d[6],lsr#32 208 add.32 @x[14],@x[14],@d[7] 209 add @x[15],@x[15],@d[7],lsr#32 210 211 b.lo .Ltail 212 213 add @x[0],@x[0],@x[1],lsl#32 // pack 214 add @x[2],@x[2],@x[3],lsl#32 215 ldp @x[1],@x[3],[$inp,#0] // load input 216 add @x[4],@x[4],@x[5],lsl#32 217 add @x[6],@x[6],@x[7],lsl#32 218 ldp @x[5],@x[7],[$inp,#16] 219 add @x[8],@x[8],@x[9],lsl#32 220 add @x[10],@x[10],@x[11],lsl#32 221 ldp @x[9],@x[11],[$inp,#32] 222 add @x[12],@x[12],@x[13],lsl#32 223 add @x[14],@x[14],@x[15],lsl#32 224 ldp @x[13],@x[15],[$inp,#48] 225 add $inp,$inp,#64 226#ifdef __AARCH64EB__ 227 rev @x[0],@x[0] 228 rev @x[2],@x[2] 229 rev @x[4],@x[4] 230 rev @x[6],@x[6] 231 rev @x[8],@x[8] 232 rev @x[10],@x[10] 233 rev @x[12],@x[12] 234 rev @x[14],@x[14] 235#endif 236 eor @x[0],@x[0],@x[1] 237 eor @x[2],@x[2],@x[3] 238 eor @x[4],@x[4],@x[5] 239 eor @x[6],@x[6],@x[7] 240 eor @x[8],@x[8],@x[9] 241 eor @x[10],@x[10],@x[11] 242 eor @x[12],@x[12],@x[13] 243 eor @x[14],@x[14],@x[15] 244 245 stp @x[0],@x[2],[$out,#0] // store output 246 add @d[6],@d[6],#1 // increment counter 247 stp @x[4],@x[6],[$out,#16] 248 stp @x[8],@x[10],[$out,#32] 249 stp @x[12],@x[14],[$out,#48] 250 add $out,$out,#64 251 252 b.hi .Loop_outer 253 254 ldp x19,x20,[x29,#16] 255 add sp,sp,#64 256 ldp x21,x22,[x29,#32] 257 ldp x23,x24,[x29,#48] 258 ldp x25,x26,[x29,#64] 259 ldp x27,x28,[x29,#80] 260 ldp x29,x30,[sp],#96 261 AARCH64_VALIDATE_LINK_REGISTER 262 ret 263 264.align 4 265.Ltail: 266 add $len,$len,#64 267.Less_than_64: 268 sub $out,$out,#1 269 add $inp,$inp,$len 270 add $out,$out,$len 271 add $ctr,sp,$len 272 neg $len,$len 273 274 add @x[0],@x[0],@x[1],lsl#32 // pack 275 add @x[2],@x[2],@x[3],lsl#32 276 add @x[4],@x[4],@x[5],lsl#32 277 add @x[6],@x[6],@x[7],lsl#32 278 add @x[8],@x[8],@x[9],lsl#32 279 add @x[10],@x[10],@x[11],lsl#32 280 add @x[12],@x[12],@x[13],lsl#32 281 add @x[14],@x[14],@x[15],lsl#32 282#ifdef __AARCH64EB__ 283 rev @x[0],@x[0] 284 rev @x[2],@x[2] 285 rev @x[4],@x[4] 286 rev @x[6],@x[6] 287 rev @x[8],@x[8] 288 rev @x[10],@x[10] 289 rev @x[12],@x[12] 290 rev @x[14],@x[14] 291#endif 292 stp @x[0],@x[2],[sp,#0] 293 stp @x[4],@x[6],[sp,#16] 294 stp @x[8],@x[10],[sp,#32] 295 stp @x[12],@x[14],[sp,#48] 296 297.Loop_tail: 298 ldrb w10,[$inp,$len] 299 ldrb w11,[$ctr,$len] 300 add $len,$len,#1 301 eor w10,w10,w11 302 strb w10,[$out,$len] 303 cbnz $len,.Loop_tail 304 305 stp xzr,xzr,[sp,#0] 306 stp xzr,xzr,[sp,#16] 307 stp xzr,xzr,[sp,#32] 308 stp xzr,xzr,[sp,#48] 309 310 ldp x19,x20,[x29,#16] 311 add sp,sp,#64 312 ldp x21,x22,[x29,#32] 313 ldp x23,x24,[x29,#48] 314 ldp x25,x26,[x29,#64] 315 ldp x27,x28,[x29,#80] 316 ldp x29,x30,[sp],#96 317 AARCH64_VALIDATE_LINK_REGISTER 318 ret 319.size ChaCha20_ctr32_nohw,.-ChaCha20_ctr32_nohw 320___ 321 322{{{ 323my ($A0,$B0,$C0,$D0,$A1,$B1,$C1,$D1,$A2,$B2,$C2,$D2,$T0,$T1,$T2,$T3) = 324 map("v$_.4s",(0..7,16..23)); 325my (@K)=map("v$_.4s",(24..30)); 326my $ONE="v31.4s"; 327 328sub NEONROUND { 329my $odd = pop; 330my ($a,$b,$c,$d,$t)=@_; 331 332 ( 333 "&add ('$a','$a','$b')", 334 "&eor ('$d','$d','$a')", 335 "&rev32_16 ('$d','$d')", # vrot ($d,16) 336 337 "&add ('$c','$c','$d')", 338 "&eor ('$t','$b','$c')", 339 "&ushr ('$b','$t',20)", 340 "&sli ('$b','$t',12)", 341 342 "&add ('$a','$a','$b')", 343 "&eor ('$t','$d','$a')", 344 "&ushr ('$d','$t',24)", 345 "&sli ('$d','$t',8)", 346 347 "&add ('$c','$c','$d')", 348 "&eor ('$t','$b','$c')", 349 "&ushr ('$b','$t',25)", 350 "&sli ('$b','$t',7)", 351 352 "&ext ('$c','$c','$c',8)", 353 "&ext ('$d','$d','$d',$odd?4:12)", 354 "&ext ('$b','$b','$b',$odd?12:4)" 355 ); 356} 357 358$code.=<<___; 359 360.globl ChaCha20_ctr32_neon 361.type ChaCha20_ctr32_neon,%function 362.align 5 363ChaCha20_ctr32_neon: 364 AARCH64_SIGN_LINK_REGISTER 365 stp x29,x30,[sp,#-96]! 366 add x29,sp,#0 367 368 adrp @x[0],:pg_hi21:.Lsigma 369 add @x[0],@x[0],:lo12:.Lsigma 370 stp x19,x20,[sp,#16] 371 stp x21,x22,[sp,#32] 372 stp x23,x24,[sp,#48] 373 stp x25,x26,[sp,#64] 374 stp x27,x28,[sp,#80] 375 cmp $len,#512 376 b.hs .L512_or_more_neon 377 378 sub sp,sp,#64 379 380 ldp @d[0],@d[1],[@x[0]] // load sigma 381 ld1 {@K[0]},[@x[0]],#16 382 ldp @d[2],@d[3],[$key] // load key 383 ldp @d[4],@d[5],[$key,#16] 384 ld1 {@K[1],@K[2]},[$key] 385 ldp @d[6],@d[7],[$ctr] // load counter 386 ld1 {@K[3]},[$ctr] 387 ld1 {$ONE},[@x[0]] 388#ifdef __AARCH64EB__ 389 rev64 @K[0],@K[0] 390 ror @d[2],@d[2],#32 391 ror @d[3],@d[3],#32 392 ror @d[4],@d[4],#32 393 ror @d[5],@d[5],#32 394 ror @d[6],@d[6],#32 395 ror @d[7],@d[7],#32 396#endif 397 add @K[3],@K[3],$ONE // += 1 398 add @K[4],@K[3],$ONE 399 add @K[5],@K[4],$ONE 400 shl $ONE,$ONE,#2 // 1 -> 4 401 402.Loop_outer_neon: 403 mov.32 @x[0],@d[0] // unpack key block 404 lsr @x[1],@d[0],#32 405 mov $A0,@K[0] 406 mov.32 @x[2],@d[1] 407 lsr @x[3],@d[1],#32 408 mov $A1,@K[0] 409 mov.32 @x[4],@d[2] 410 lsr @x[5],@d[2],#32 411 mov $A2,@K[0] 412 mov.32 @x[6],@d[3] 413 mov $B0,@K[1] 414 lsr @x[7],@d[3],#32 415 mov $B1,@K[1] 416 mov.32 @x[8],@d[4] 417 mov $B2,@K[1] 418 lsr @x[9],@d[4],#32 419 mov $D0,@K[3] 420 mov.32 @x[10],@d[5] 421 mov $D1,@K[4] 422 lsr @x[11],@d[5],#32 423 mov $D2,@K[5] 424 mov.32 @x[12],@d[6] 425 mov $C0,@K[2] 426 lsr @x[13],@d[6],#32 427 mov $C1,@K[2] 428 mov.32 @x[14],@d[7] 429 mov $C2,@K[2] 430 lsr @x[15],@d[7],#32 431 432 mov $ctr,#10 433 subs $len,$len,#256 434.Loop_neon: 435 sub $ctr,$ctr,#1 436___ 437 my @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,0); 438 my @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,0); 439 my @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,0); 440 my @thread3=&ROUND(0,4,8,12); 441 442 foreach (@thread0) { 443 eval; eval(shift(@thread3)); 444 eval(shift(@thread1)); eval(shift(@thread3)); 445 eval(shift(@thread2)); eval(shift(@thread3)); 446 } 447 448 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,1); 449 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,1); 450 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,1); 451 @thread3=&ROUND(0,5,10,15); 452 453 foreach (@thread0) { 454 eval; eval(shift(@thread3)); 455 eval(shift(@thread1)); eval(shift(@thread3)); 456 eval(shift(@thread2)); eval(shift(@thread3)); 457 } 458$code.=<<___; 459 cbnz $ctr,.Loop_neon 460 461 add.32 @x[0],@x[0],@d[0] // accumulate key block 462 add $A0,$A0,@K[0] 463 add @x[1],@x[1],@d[0],lsr#32 464 add $A1,$A1,@K[0] 465 add.32 @x[2],@x[2],@d[1] 466 add $A2,$A2,@K[0] 467 add @x[3],@x[3],@d[1],lsr#32 468 add $C0,$C0,@K[2] 469 add.32 @x[4],@x[4],@d[2] 470 add $C1,$C1,@K[2] 471 add @x[5],@x[5],@d[2],lsr#32 472 add $C2,$C2,@K[2] 473 add.32 @x[6],@x[6],@d[3] 474 add $D0,$D0,@K[3] 475 add @x[7],@x[7],@d[3],lsr#32 476 add.32 @x[8],@x[8],@d[4] 477 add $D1,$D1,@K[4] 478 add @x[9],@x[9],@d[4],lsr#32 479 add.32 @x[10],@x[10],@d[5] 480 add $D2,$D2,@K[5] 481 add @x[11],@x[11],@d[5],lsr#32 482 add.32 @x[12],@x[12],@d[6] 483 add $B0,$B0,@K[1] 484 add @x[13],@x[13],@d[6],lsr#32 485 add.32 @x[14],@x[14],@d[7] 486 add $B1,$B1,@K[1] 487 add @x[15],@x[15],@d[7],lsr#32 488 add $B2,$B2,@K[1] 489 490 b.lo .Ltail_neon 491 492 add @x[0],@x[0],@x[1],lsl#32 // pack 493 add @x[2],@x[2],@x[3],lsl#32 494 ldp @x[1],@x[3],[$inp,#0] // load input 495 add @x[4],@x[4],@x[5],lsl#32 496 add @x[6],@x[6],@x[7],lsl#32 497 ldp @x[5],@x[7],[$inp,#16] 498 add @x[8],@x[8],@x[9],lsl#32 499 add @x[10],@x[10],@x[11],lsl#32 500 ldp @x[9],@x[11],[$inp,#32] 501 add @x[12],@x[12],@x[13],lsl#32 502 add @x[14],@x[14],@x[15],lsl#32 503 ldp @x[13],@x[15],[$inp,#48] 504 add $inp,$inp,#64 505#ifdef __AARCH64EB__ 506 rev @x[0],@x[0] 507 rev @x[2],@x[2] 508 rev @x[4],@x[4] 509 rev @x[6],@x[6] 510 rev @x[8],@x[8] 511 rev @x[10],@x[10] 512 rev @x[12],@x[12] 513 rev @x[14],@x[14] 514#endif 515 ld1.8 {$T0-$T3},[$inp],#64 516 eor @x[0],@x[0],@x[1] 517 eor @x[2],@x[2],@x[3] 518 eor @x[4],@x[4],@x[5] 519 eor @x[6],@x[6],@x[7] 520 eor @x[8],@x[8],@x[9] 521 eor $A0,$A0,$T0 522 eor @x[10],@x[10],@x[11] 523 eor $B0,$B0,$T1 524 eor @x[12],@x[12],@x[13] 525 eor $C0,$C0,$T2 526 eor @x[14],@x[14],@x[15] 527 eor $D0,$D0,$T3 528 ld1.8 {$T0-$T3},[$inp],#64 529 530 stp @x[0],@x[2],[$out,#0] // store output 531 add @d[6],@d[6],#4 // increment counter 532 stp @x[4],@x[6],[$out,#16] 533 add @K[3],@K[3],$ONE // += 4 534 stp @x[8],@x[10],[$out,#32] 535 add @K[4],@K[4],$ONE 536 stp @x[12],@x[14],[$out,#48] 537 add @K[5],@K[5],$ONE 538 add $out,$out,#64 539 540 st1.8 {$A0-$D0},[$out],#64 541 ld1.8 {$A0-$D0},[$inp],#64 542 543 eor $A1,$A1,$T0 544 eor $B1,$B1,$T1 545 eor $C1,$C1,$T2 546 eor $D1,$D1,$T3 547 st1.8 {$A1-$D1},[$out],#64 548 549 eor $A2,$A2,$A0 550 eor $B2,$B2,$B0 551 eor $C2,$C2,$C0 552 eor $D2,$D2,$D0 553 st1.8 {$A2-$D2},[$out],#64 554 555 b.hi .Loop_outer_neon 556 557 ldp x19,x20,[x29,#16] 558 add sp,sp,#64 559 ldp x21,x22,[x29,#32] 560 ldp x23,x24,[x29,#48] 561 ldp x25,x26,[x29,#64] 562 ldp x27,x28,[x29,#80] 563 ldp x29,x30,[sp],#96 564 AARCH64_VALIDATE_LINK_REGISTER 565 ret 566 567.Ltail_neon: 568 add $len,$len,#256 569 cmp $len,#64 570 b.lo .Less_than_64 571 572 add @x[0],@x[0],@x[1],lsl#32 // pack 573 add @x[2],@x[2],@x[3],lsl#32 574 ldp @x[1],@x[3],[$inp,#0] // load input 575 add @x[4],@x[4],@x[5],lsl#32 576 add @x[6],@x[6],@x[7],lsl#32 577 ldp @x[5],@x[7],[$inp,#16] 578 add @x[8],@x[8],@x[9],lsl#32 579 add @x[10],@x[10],@x[11],lsl#32 580 ldp @x[9],@x[11],[$inp,#32] 581 add @x[12],@x[12],@x[13],lsl#32 582 add @x[14],@x[14],@x[15],lsl#32 583 ldp @x[13],@x[15],[$inp,#48] 584 add $inp,$inp,#64 585#ifdef __AARCH64EB__ 586 rev @x[0],@x[0] 587 rev @x[2],@x[2] 588 rev @x[4],@x[4] 589 rev @x[6],@x[6] 590 rev @x[8],@x[8] 591 rev @x[10],@x[10] 592 rev @x[12],@x[12] 593 rev @x[14],@x[14] 594#endif 595 eor @x[0],@x[0],@x[1] 596 eor @x[2],@x[2],@x[3] 597 eor @x[4],@x[4],@x[5] 598 eor @x[6],@x[6],@x[7] 599 eor @x[8],@x[8],@x[9] 600 eor @x[10],@x[10],@x[11] 601 eor @x[12],@x[12],@x[13] 602 eor @x[14],@x[14],@x[15] 603 604 stp @x[0],@x[2],[$out,#0] // store output 605 add @d[6],@d[6],#4 // increment counter 606 stp @x[4],@x[6],[$out,#16] 607 stp @x[8],@x[10],[$out,#32] 608 stp @x[12],@x[14],[$out,#48] 609 add $out,$out,#64 610 b.eq .Ldone_neon 611 sub $len,$len,#64 612 cmp $len,#64 613 b.lo .Less_than_128 614 615 ld1.8 {$T0-$T3},[$inp],#64 616 eor $A0,$A0,$T0 617 eor $B0,$B0,$T1 618 eor $C0,$C0,$T2 619 eor $D0,$D0,$T3 620 st1.8 {$A0-$D0},[$out],#64 621 b.eq .Ldone_neon 622 sub $len,$len,#64 623 cmp $len,#64 624 b.lo .Less_than_192 625 626 ld1.8 {$T0-$T3},[$inp],#64 627 eor $A1,$A1,$T0 628 eor $B1,$B1,$T1 629 eor $C1,$C1,$T2 630 eor $D1,$D1,$T3 631 st1.8 {$A1-$D1},[$out],#64 632 b.eq .Ldone_neon 633 sub $len,$len,#64 634 635 st1.8 {$A2-$D2},[sp] 636 b .Last_neon 637 638.Less_than_128: 639 st1.8 {$A0-$D0},[sp] 640 b .Last_neon 641.Less_than_192: 642 st1.8 {$A1-$D1},[sp] 643 b .Last_neon 644 645.align 4 646.Last_neon: 647 sub $out,$out,#1 648 add $inp,$inp,$len 649 add $out,$out,$len 650 add $ctr,sp,$len 651 neg $len,$len 652 653.Loop_tail_neon: 654 ldrb w10,[$inp,$len] 655 ldrb w11,[$ctr,$len] 656 add $len,$len,#1 657 eor w10,w10,w11 658 strb w10,[$out,$len] 659 cbnz $len,.Loop_tail_neon 660 661 stp xzr,xzr,[sp,#0] 662 stp xzr,xzr,[sp,#16] 663 stp xzr,xzr,[sp,#32] 664 stp xzr,xzr,[sp,#48] 665 666.Ldone_neon: 667 ldp x19,x20,[x29,#16] 668 add sp,sp,#64 669 ldp x21,x22,[x29,#32] 670 ldp x23,x24,[x29,#48] 671 ldp x25,x26,[x29,#64] 672 ldp x27,x28,[x29,#80] 673 ldp x29,x30,[sp],#96 674 AARCH64_VALIDATE_LINK_REGISTER 675 ret 676.size ChaCha20_ctr32_neon,.-ChaCha20_ctr32_neon 677___ 678{ 679my ($T0,$T1,$T2,$T3,$T4,$T5)=@K; 680my ($A0,$B0,$C0,$D0,$A1,$B1,$C1,$D1,$A2,$B2,$C2,$D2, 681 $A3,$B3,$C3,$D3,$A4,$B4,$C4,$D4,$A5,$B5,$C5,$D5) = map("v$_.4s",(0..23)); 682 683$code.=<<___; 684.type ChaCha20_512_neon,%function 685.align 5 686ChaCha20_512_neon: 687 AARCH64_SIGN_LINK_REGISTER 688 stp x29,x30,[sp,#-96]! 689 add x29,sp,#0 690 691 adrp @x[0],:pg_hi21:.Lsigma 692 add @x[0],@x[0],:lo12:.Lsigma 693 stp x19,x20,[sp,#16] 694 stp x21,x22,[sp,#32] 695 stp x23,x24,[sp,#48] 696 stp x25,x26,[sp,#64] 697 stp x27,x28,[sp,#80] 698 699.L512_or_more_neon: 700 sub sp,sp,#128+64 701 702 ldp @d[0],@d[1],[@x[0]] // load sigma 703 ld1 {@K[0]},[@x[0]],#16 704 ldp @d[2],@d[3],[$key] // load key 705 ldp @d[4],@d[5],[$key,#16] 706 ld1 {@K[1],@K[2]},[$key] 707 ldp @d[6],@d[7],[$ctr] // load counter 708 ld1 {@K[3]},[$ctr] 709 ld1 {$ONE},[@x[0]] 710#ifdef __AARCH64EB__ 711 rev64 @K[0],@K[0] 712 ror @d[2],@d[2],#32 713 ror @d[3],@d[3],#32 714 ror @d[4],@d[4],#32 715 ror @d[5],@d[5],#32 716 ror @d[6],@d[6],#32 717 ror @d[7],@d[7],#32 718#endif 719 add @K[3],@K[3],$ONE // += 1 720 stp @K[0],@K[1],[sp,#0] // off-load key block, invariant part 721 add @K[3],@K[3],$ONE // not typo 722 str @K[2],[sp,#32] 723 add @K[4],@K[3],$ONE 724 add @K[5],@K[4],$ONE 725 add @K[6],@K[5],$ONE 726 shl $ONE,$ONE,#2 // 1 -> 4 727 728 stp d8,d9,[sp,#128+0] // meet ABI requirements 729 stp d10,d11,[sp,#128+16] 730 stp d12,d13,[sp,#128+32] 731 stp d14,d15,[sp,#128+48] 732 733 sub $len,$len,#512 // not typo 734 735.Loop_outer_512_neon: 736 mov $A0,@K[0] 737 mov $A1,@K[0] 738 mov $A2,@K[0] 739 mov $A3,@K[0] 740 mov $A4,@K[0] 741 mov $A5,@K[0] 742 mov $B0,@K[1] 743 mov.32 @x[0],@d[0] // unpack key block 744 mov $B1,@K[1] 745 lsr @x[1],@d[0],#32 746 mov $B2,@K[1] 747 mov.32 @x[2],@d[1] 748 mov $B3,@K[1] 749 lsr @x[3],@d[1],#32 750 mov $B4,@K[1] 751 mov.32 @x[4],@d[2] 752 mov $B5,@K[1] 753 lsr @x[5],@d[2],#32 754 mov $D0,@K[3] 755 mov.32 @x[6],@d[3] 756 mov $D1,@K[4] 757 lsr @x[7],@d[3],#32 758 mov $D2,@K[5] 759 mov.32 @x[8],@d[4] 760 mov $D3,@K[6] 761 lsr @x[9],@d[4],#32 762 mov $C0,@K[2] 763 mov.32 @x[10],@d[5] 764 mov $C1,@K[2] 765 lsr @x[11],@d[5],#32 766 add $D4,$D0,$ONE // +4 767 mov.32 @x[12],@d[6] 768 add $D5,$D1,$ONE // +4 769 lsr @x[13],@d[6],#32 770 mov $C2,@K[2] 771 mov.32 @x[14],@d[7] 772 mov $C3,@K[2] 773 lsr @x[15],@d[7],#32 774 mov $C4,@K[2] 775 stp @K[3],@K[4],[sp,#48] // off-load key block, variable part 776 mov $C5,@K[2] 777 str @K[5],[sp,#80] 778 779 mov $ctr,#5 780 subs $len,$len,#512 781.Loop_upper_neon: 782 sub $ctr,$ctr,#1 783___ 784 my @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,0); 785 my @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,0); 786 my @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,0); 787 my @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,0); 788 my @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,0); 789 my @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,0); 790 my @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15)); 791 my $diff = ($#thread0+1)*6 - $#thread67 - 1; 792 my $i = 0; 793 794 foreach (@thread0) { 795 eval; eval(shift(@thread67)); 796 eval(shift(@thread1)); eval(shift(@thread67)); 797 eval(shift(@thread2)); eval(shift(@thread67)); 798 eval(shift(@thread3)); eval(shift(@thread67)); 799 eval(shift(@thread4)); eval(shift(@thread67)); 800 eval(shift(@thread5)); eval(shift(@thread67)); 801 } 802 803 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,1); 804 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,1); 805 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,1); 806 @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,1); 807 @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,1); 808 @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,1); 809 @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15)); 810 811 foreach (@thread0) { 812 eval; eval(shift(@thread67)); 813 eval(shift(@thread1)); eval(shift(@thread67)); 814 eval(shift(@thread2)); eval(shift(@thread67)); 815 eval(shift(@thread3)); eval(shift(@thread67)); 816 eval(shift(@thread4)); eval(shift(@thread67)); 817 eval(shift(@thread5)); eval(shift(@thread67)); 818 } 819$code.=<<___; 820 cbnz $ctr,.Loop_upper_neon 821 822 add.32 @x[0],@x[0],@d[0] // accumulate key block 823 add @x[1],@x[1],@d[0],lsr#32 824 add.32 @x[2],@x[2],@d[1] 825 add @x[3],@x[3],@d[1],lsr#32 826 add.32 @x[4],@x[4],@d[2] 827 add @x[5],@x[5],@d[2],lsr#32 828 add.32 @x[6],@x[6],@d[3] 829 add @x[7],@x[7],@d[3],lsr#32 830 add.32 @x[8],@x[8],@d[4] 831 add @x[9],@x[9],@d[4],lsr#32 832 add.32 @x[10],@x[10],@d[5] 833 add @x[11],@x[11],@d[5],lsr#32 834 add.32 @x[12],@x[12],@d[6] 835 add @x[13],@x[13],@d[6],lsr#32 836 add.32 @x[14],@x[14],@d[7] 837 add @x[15],@x[15],@d[7],lsr#32 838 839 add @x[0],@x[0],@x[1],lsl#32 // pack 840 add @x[2],@x[2],@x[3],lsl#32 841 ldp @x[1],@x[3],[$inp,#0] // load input 842 add @x[4],@x[4],@x[5],lsl#32 843 add @x[6],@x[6],@x[7],lsl#32 844 ldp @x[5],@x[7],[$inp,#16] 845 add @x[8],@x[8],@x[9],lsl#32 846 add @x[10],@x[10],@x[11],lsl#32 847 ldp @x[9],@x[11],[$inp,#32] 848 add @x[12],@x[12],@x[13],lsl#32 849 add @x[14],@x[14],@x[15],lsl#32 850 ldp @x[13],@x[15],[$inp,#48] 851 add $inp,$inp,#64 852#ifdef __AARCH64EB__ 853 rev @x[0],@x[0] 854 rev @x[2],@x[2] 855 rev @x[4],@x[4] 856 rev @x[6],@x[6] 857 rev @x[8],@x[8] 858 rev @x[10],@x[10] 859 rev @x[12],@x[12] 860 rev @x[14],@x[14] 861#endif 862 eor @x[0],@x[0],@x[1] 863 eor @x[2],@x[2],@x[3] 864 eor @x[4],@x[4],@x[5] 865 eor @x[6],@x[6],@x[7] 866 eor @x[8],@x[8],@x[9] 867 eor @x[10],@x[10],@x[11] 868 eor @x[12],@x[12],@x[13] 869 eor @x[14],@x[14],@x[15] 870 871 stp @x[0],@x[2],[$out,#0] // store output 872 add @d[6],@d[6],#1 // increment counter 873 mov.32 @x[0],@d[0] // unpack key block 874 lsr @x[1],@d[0],#32 875 stp @x[4],@x[6],[$out,#16] 876 mov.32 @x[2],@d[1] 877 lsr @x[3],@d[1],#32 878 stp @x[8],@x[10],[$out,#32] 879 mov.32 @x[4],@d[2] 880 lsr @x[5],@d[2],#32 881 stp @x[12],@x[14],[$out,#48] 882 add $out,$out,#64 883 mov.32 @x[6],@d[3] 884 lsr @x[7],@d[3],#32 885 mov.32 @x[8],@d[4] 886 lsr @x[9],@d[4],#32 887 mov.32 @x[10],@d[5] 888 lsr @x[11],@d[5],#32 889 mov.32 @x[12],@d[6] 890 lsr @x[13],@d[6],#32 891 mov.32 @x[14],@d[7] 892 lsr @x[15],@d[7],#32 893 894 mov $ctr,#5 895.Loop_lower_neon: 896 sub $ctr,$ctr,#1 897___ 898 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,0); 899 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,0); 900 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,0); 901 @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,0); 902 @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,0); 903 @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,0); 904 @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15)); 905 906 foreach (@thread0) { 907 eval; eval(shift(@thread67)); 908 eval(shift(@thread1)); eval(shift(@thread67)); 909 eval(shift(@thread2)); eval(shift(@thread67)); 910 eval(shift(@thread3)); eval(shift(@thread67)); 911 eval(shift(@thread4)); eval(shift(@thread67)); 912 eval(shift(@thread5)); eval(shift(@thread67)); 913 } 914 915 @thread0=&NEONROUND($A0,$B0,$C0,$D0,$T0,1); 916 @thread1=&NEONROUND($A1,$B1,$C1,$D1,$T1,1); 917 @thread2=&NEONROUND($A2,$B2,$C2,$D2,$T2,1); 918 @thread3=&NEONROUND($A3,$B3,$C3,$D3,$T3,1); 919 @thread4=&NEONROUND($A4,$B4,$C4,$D4,$T4,1); 920 @thread5=&NEONROUND($A5,$B5,$C5,$D5,$T5,1); 921 @thread67=(&ROUND(0,4,8,12),&ROUND(0,5,10,15)); 922 923 foreach (@thread0) { 924 eval; eval(shift(@thread67)); 925 eval(shift(@thread1)); eval(shift(@thread67)); 926 eval(shift(@thread2)); eval(shift(@thread67)); 927 eval(shift(@thread3)); eval(shift(@thread67)); 928 eval(shift(@thread4)); eval(shift(@thread67)); 929 eval(shift(@thread5)); eval(shift(@thread67)); 930 } 931$code.=<<___; 932 cbnz $ctr,.Loop_lower_neon 933 934 add.32 @x[0],@x[0],@d[0] // accumulate key block 935 ldp @K[0],@K[1],[sp,#0] 936 add @x[1],@x[1],@d[0],lsr#32 937 ldp @K[2],@K[3],[sp,#32] 938 add.32 @x[2],@x[2],@d[1] 939 ldp @K[4],@K[5],[sp,#64] 940 add @x[3],@x[3],@d[1],lsr#32 941 add $A0,$A0,@K[0] 942 add.32 @x[4],@x[4],@d[2] 943 add $A1,$A1,@K[0] 944 add @x[5],@x[5],@d[2],lsr#32 945 add $A2,$A2,@K[0] 946 add.32 @x[6],@x[6],@d[3] 947 add $A3,$A3,@K[0] 948 add @x[7],@x[7],@d[3],lsr#32 949 add $A4,$A4,@K[0] 950 add.32 @x[8],@x[8],@d[4] 951 add $A5,$A5,@K[0] 952 add @x[9],@x[9],@d[4],lsr#32 953 add $C0,$C0,@K[2] 954 add.32 @x[10],@x[10],@d[5] 955 add $C1,$C1,@K[2] 956 add @x[11],@x[11],@d[5],lsr#32 957 add $C2,$C2,@K[2] 958 add.32 @x[12],@x[12],@d[6] 959 add $C3,$C3,@K[2] 960 add @x[13],@x[13],@d[6],lsr#32 961 add $C4,$C4,@K[2] 962 add.32 @x[14],@x[14],@d[7] 963 add $C5,$C5,@K[2] 964 add @x[15],@x[15],@d[7],lsr#32 965 add $D4,$D4,$ONE // +4 966 add @x[0],@x[0],@x[1],lsl#32 // pack 967 add $D5,$D5,$ONE // +4 968 add @x[2],@x[2],@x[3],lsl#32 969 add $D0,$D0,@K[3] 970 ldp @x[1],@x[3],[$inp,#0] // load input 971 add $D1,$D1,@K[4] 972 add @x[4],@x[4],@x[5],lsl#32 973 add $D2,$D2,@K[5] 974 add @x[6],@x[6],@x[7],lsl#32 975 add $D3,$D3,@K[6] 976 ldp @x[5],@x[7],[$inp,#16] 977 add $D4,$D4,@K[3] 978 add @x[8],@x[8],@x[9],lsl#32 979 add $D5,$D5,@K[4] 980 add @x[10],@x[10],@x[11],lsl#32 981 add $B0,$B0,@K[1] 982 ldp @x[9],@x[11],[$inp,#32] 983 add $B1,$B1,@K[1] 984 add @x[12],@x[12],@x[13],lsl#32 985 add $B2,$B2,@K[1] 986 add @x[14],@x[14],@x[15],lsl#32 987 add $B3,$B3,@K[1] 988 ldp @x[13],@x[15],[$inp,#48] 989 add $B4,$B4,@K[1] 990 add $inp,$inp,#64 991 add $B5,$B5,@K[1] 992 993#ifdef __AARCH64EB__ 994 rev @x[0],@x[0] 995 rev @x[2],@x[2] 996 rev @x[4],@x[4] 997 rev @x[6],@x[6] 998 rev @x[8],@x[8] 999 rev @x[10],@x[10] 1000 rev @x[12],@x[12] 1001 rev @x[14],@x[14] 1002#endif 1003 ld1.8 {$T0-$T3},[$inp],#64 1004 eor @x[0],@x[0],@x[1] 1005 eor @x[2],@x[2],@x[3] 1006 eor @x[4],@x[4],@x[5] 1007 eor @x[6],@x[6],@x[7] 1008 eor @x[8],@x[8],@x[9] 1009 eor $A0,$A0,$T0 1010 eor @x[10],@x[10],@x[11] 1011 eor $B0,$B0,$T1 1012 eor @x[12],@x[12],@x[13] 1013 eor $C0,$C0,$T2 1014 eor @x[14],@x[14],@x[15] 1015 eor $D0,$D0,$T3 1016 ld1.8 {$T0-$T3},[$inp],#64 1017 1018 stp @x[0],@x[2],[$out,#0] // store output 1019 add @d[6],@d[6],#7 // increment counter 1020 stp @x[4],@x[6],[$out,#16] 1021 stp @x[8],@x[10],[$out,#32] 1022 stp @x[12],@x[14],[$out,#48] 1023 add $out,$out,#64 1024 st1.8 {$A0-$D0},[$out],#64 1025 1026 ld1.8 {$A0-$D0},[$inp],#64 1027 eor $A1,$A1,$T0 1028 eor $B1,$B1,$T1 1029 eor $C1,$C1,$T2 1030 eor $D1,$D1,$T3 1031 st1.8 {$A1-$D1},[$out],#64 1032 1033 ld1.8 {$A1-$D1},[$inp],#64 1034 eor $A2,$A2,$A0 1035 ldp @K[0],@K[1],[sp,#0] 1036 eor $B2,$B2,$B0 1037 ldp @K[2],@K[3],[sp,#32] 1038 eor $C2,$C2,$C0 1039 eor $D2,$D2,$D0 1040 st1.8 {$A2-$D2},[$out],#64 1041 1042 ld1.8 {$A2-$D2},[$inp],#64 1043 eor $A3,$A3,$A1 1044 eor $B3,$B3,$B1 1045 eor $C3,$C3,$C1 1046 eor $D3,$D3,$D1 1047 st1.8 {$A3-$D3},[$out],#64 1048 1049 ld1.8 {$A3-$D3},[$inp],#64 1050 eor $A4,$A4,$A2 1051 eor $B4,$B4,$B2 1052 eor $C4,$C4,$C2 1053 eor $D4,$D4,$D2 1054 st1.8 {$A4-$D4},[$out],#64 1055 1056 shl $A0,$ONE,#1 // 4 -> 8 1057 eor $A5,$A5,$A3 1058 eor $B5,$B5,$B3 1059 eor $C5,$C5,$C3 1060 eor $D5,$D5,$D3 1061 st1.8 {$A5-$D5},[$out],#64 1062 1063 add @K[3],@K[3],$A0 // += 8 1064 add @K[4],@K[4],$A0 1065 add @K[5],@K[5],$A0 1066 add @K[6],@K[6],$A0 1067 1068 b.hs .Loop_outer_512_neon 1069 1070 adds $len,$len,#512 1071 ushr $A0,$ONE,#2 // 4 -> 1 1072 1073 ldp d8,d9,[sp,#128+0] // meet ABI requirements 1074 ldp d10,d11,[sp,#128+16] 1075 ldp d12,d13,[sp,#128+32] 1076 ldp d14,d15,[sp,#128+48] 1077 1078 stp @K[0],$ONE,[sp,#0] // wipe off-load area 1079 stp @K[0],$ONE,[sp,#32] 1080 stp @K[0],$ONE,[sp,#64] 1081 1082 b.eq .Ldone_512_neon 1083 1084 cmp $len,#192 1085 sub @K[3],@K[3],$A0 // -= 1 1086 sub @K[4],@K[4],$A0 1087 sub @K[5],@K[5],$A0 1088 add sp,sp,#128 1089 b.hs .Loop_outer_neon 1090 1091 eor @K[1],@K[1],@K[1] 1092 eor @K[2],@K[2],@K[2] 1093 eor @K[3],@K[3],@K[3] 1094 eor @K[4],@K[4],@K[4] 1095 eor @K[5],@K[5],@K[5] 1096 eor @K[6],@K[6],@K[6] 1097 b .Loop_outer 1098 1099.Ldone_512_neon: 1100 ldp x19,x20,[x29,#16] 1101 add sp,sp,#128+64 1102 ldp x21,x22,[x29,#32] 1103 ldp x23,x24,[x29,#48] 1104 ldp x25,x26,[x29,#64] 1105 ldp x27,x28,[x29,#80] 1106 ldp x29,x30,[sp],#96 1107 AARCH64_VALIDATE_LINK_REGISTER 1108 ret 1109.size ChaCha20_512_neon,.-ChaCha20_512_neon 1110___ 1111} 1112}}} 1113 1114foreach (split("\n",$code)) { 1115 s/\`([^\`]*)\`/eval $1/geo; 1116 1117 (s/\b([a-z]+)\.32\b/$1/ and (s/x([0-9]+)/w$1/g or 1)) or 1118 (m/\b(eor|ext|mov)\b/ and (s/\.4s/\.16b/g or 1)) or 1119 (s/\b((?:ld|st)1)\.8\b/$1/ and (s/\.4s/\.16b/g or 1)) or 1120 (m/\b(ld|st)[rp]\b/ and (s/v([0-9]+)\.4s/q$1/g or 1)) or 1121 (s/\brev32\.16\b/rev32/ and (s/\.4s/\.8h/g or 1)); 1122 1123 #s/\bq([0-9]+)#(lo|hi)/sprintf "d%d",2*$1+($2 eq "hi")/geo; 1124 1125 print $_,"\n"; 1126} 1127close STDOUT or die "error closing STDOUT: $!"; # flush 1128