1#! /usr/bin/env perl 2# Copyright 2014-2020 The OpenSSL Project Authors. All Rights Reserved. 3# 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 8 9# ==================================================================== 10# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL 11# project. The module is, however, dual licensed under OpenSSL and 12# CRYPTOGAMS licenses depending on where you obtain it. For further 13# details see http://www.openssl.org/~appro/cryptogams/. 14# 15# Permission to use under GPLv2 terms is granted. 16# ==================================================================== 17# 18# SHA256/512 for ARMv8. 19# 20# Performance in cycles per processed byte and improvement coefficient 21# over code generated with "default" compiler: 22# 23# SHA256-hw SHA256(*) SHA512 24# Apple A7 1.97 10.5 (+33%) 6.73 (-1%(**)) 25# Cortex-A53 2.38 15.5 (+115%) 10.0 (+150%(***)) 26# Cortex-A57 2.31 11.6 (+86%) 7.51 (+260%(***)) 27# Denver 2.01 10.5 (+26%) 6.70 (+8%) 28# X-Gene 20.0 (+100%) 12.8 (+300%(***)) 29# Mongoose 2.36 13.0 (+50%) 8.36 (+33%) 30# Kryo 1.92 17.4 (+30%) 11.2 (+8%) 31# ThunderX2 2.54 13.2 (+40%) 8.40 (+18%) 32# 33# (*) Software SHA256 results are of lesser relevance, presented 34# mostly for informational purposes. 35# (**) The result is a trade-off: it's possible to improve it by 36# 10% (or by 1 cycle per round), but at the cost of 20% loss 37# on Cortex-A53 (or by 4 cycles per round). 38# (***) Super-impressive coefficients over gcc-generated code are 39# indication of some compiler "pathology", most notably code 40# generated with -mgeneral-regs-only is significantly faster 41# and the gap is only 40-90%. 42# 43# October 2016. 44# 45# Originally it was reckoned that it makes no sense to implement NEON 46# version of SHA256 for 64-bit processors. This is because performance 47# improvement on most wide-spread Cortex-A5x processors was observed 48# to be marginal, same on Cortex-A53 and ~10% on A57. But then it was 49# observed that 32-bit NEON SHA256 performs significantly better than 50# 64-bit scalar version on *some* of the more recent processors. As 51# result 64-bit NEON version of SHA256 was added to provide best 52# all-round performance. For example it executes ~30% faster on X-Gene 53# and Mongoose. [For reference, NEON version of SHA512 is bound to 54# deliver much less improvement, likely *negative* on Cortex-A5x. 55# Which is why NEON support is limited to SHA256.] 56 57# $output is the last argument if it looks like a file (it has an extension) 58# $flavour is the first argument if it doesn't look like a file 59$output = $#ARGV >= 0 && $ARGV[$#ARGV] =~ m|\.\w+$| ? pop : undef; 60$flavour = $#ARGV >= 0 && $ARGV[0] !~ m|\.| ? shift : undef; 61 62if ($flavour && $flavour ne "void") { 63 $0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 64 ( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or 65 ( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or 66 die "can't locate arm-xlate.pl"; 67 68 open OUT,"| \"$^X\" $xlate $flavour \"$output\"" 69 or die "can't call $xlate: $!"; 70 *STDOUT=*OUT; 71} else { 72 $output and open STDOUT,">$output"; 73} 74 75if ($output =~ /512/) { 76 $BITS=512; 77 $SZ=8; 78 @Sigma0=(28,34,39); 79 @Sigma1=(14,18,41); 80 @sigma0=(1, 8, 7); 81 @sigma1=(19,61, 6); 82 $rounds=80; 83 $reg_t="x"; 84} else { 85 $BITS=256; 86 $SZ=4; 87 @Sigma0=( 2,13,22); 88 @Sigma1=( 6,11,25); 89 @sigma0=( 7,18, 3); 90 @sigma1=(17,19,10); 91 $rounds=64; 92 $reg_t="w"; 93} 94 95$func="sha${BITS}_block_data_order"; 96 97($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30)); 98 99@X=map("$reg_t$_",(3..15,0..2)); 100@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27)); 101($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28)); 102 103sub BODY_00_xx { 104my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_; 105my $j=($i+1)&15; 106my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]); 107 $T0=@X[$i+3] if ($i<11); 108 109$code.=<<___ if ($i<16); 110#ifndef __AARCH64EB__ 111 rev @X[$i],@X[$i] // $i 112#endif 113___ 114$code.=<<___ if ($i<13 && ($i&1)); 115 ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ 116___ 117$code.=<<___ if ($i==13); 118 ldp @X[14],@X[15],[$inp] 119___ 120$code.=<<___ if ($i>=14); 121 ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`] 122___ 123$code.=<<___ if ($i>0 && $i<16); 124 add $a,$a,$t1 // h+=Sigma0(a) 125___ 126$code.=<<___ if ($i>=11); 127 str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`] 128___ 129# While ARMv8 specifies merged rotate-n-logical operation such as 130# 'eor x,y,z,ror#n', it was found to negatively affect performance 131# on Apple A7. The reason seems to be that it requires even 'y' to 132# be available earlier. This means that such merged instruction is 133# not necessarily best choice on critical path... On the other hand 134# Cortex-A5x handles merged instructions much better than disjoint 135# rotate and logical... See (**) footnote above. 136$code.=<<___ if ($i<15); 137 ror $t0,$e,#$Sigma1[0] 138 add $h,$h,$t2 // h+=K[i] 139 eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]` 140 and $t1,$f,$e 141 bic $t2,$g,$e 142 add $h,$h,@X[$i&15] // h+=X[i] 143 orr $t1,$t1,$t2 // Ch(e,f,g) 144 eor $t2,$a,$b // a^b, b^c in next round 145 eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e) 146 ror $T0,$a,#$Sigma0[0] 147 add $h,$h,$t1 // h+=Ch(e,f,g) 148 eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]` 149 add $h,$h,$t0 // h+=Sigma1(e) 150 and $t3,$t3,$t2 // (b^c)&=(a^b) 151 add $d,$d,$h // d+=h 152 eor $t3,$t3,$b // Maj(a,b,c) 153 eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a) 154 add $h,$h,$t3 // h+=Maj(a,b,c) 155 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 156 //add $h,$h,$t1 // h+=Sigma0(a) 157___ 158$code.=<<___ if ($i>=15); 159 ror $t0,$e,#$Sigma1[0] 160 add $h,$h,$t2 // h+=K[i] 161 ror $T1,@X[($j+1)&15],#$sigma0[0] 162 and $t1,$f,$e 163 ror $T2,@X[($j+14)&15],#$sigma1[0] 164 bic $t2,$g,$e 165 ror $T0,$a,#$Sigma0[0] 166 add $h,$h,@X[$i&15] // h+=X[i] 167 eor $t0,$t0,$e,ror#$Sigma1[1] 168 eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1] 169 orr $t1,$t1,$t2 // Ch(e,f,g) 170 eor $t2,$a,$b // a^b, b^c in next round 171 eor $t0,$t0,$e,ror#$Sigma1[2] // Sigma1(e) 172 eor $T0,$T0,$a,ror#$Sigma0[1] 173 add $h,$h,$t1 // h+=Ch(e,f,g) 174 and $t3,$t3,$t2 // (b^c)&=(a^b) 175 eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1] 176 eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1]) 177 add $h,$h,$t0 // h+=Sigma1(e) 178 eor $t3,$t3,$b // Maj(a,b,c) 179 eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a) 180 eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14]) 181 add @X[$j],@X[$j],@X[($j+9)&15] 182 add $d,$d,$h // d+=h 183 add $h,$h,$t3 // h+=Maj(a,b,c) 184 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 185 add @X[$j],@X[$j],$T1 186 add $h,$h,$t1 // h+=Sigma0(a) 187 add @X[$j],@X[$j],$T2 188___ 189 ($t2,$t3)=($t3,$t2); 190} 191 192$code.=<<___; 193#include "arm_arch.h" 194#ifndef __KERNEL__ 195.extern OPENSSL_armcap_P 196.hidden OPENSSL_armcap_P 197#endif 198 199.text 200 201.globl $func 202.type $func,%function 203.align 6 204$func: 205 AARCH64_VALID_CALL_TARGET 206#ifndef __KERNEL__ 207 adrp x16,OPENSSL_armcap_P 208 ldr w16,[x16,#:lo12:OPENSSL_armcap_P] 209___ 210$code.=<<___ if ($SZ==4); 211 tst w16,#ARMV8_SHA256 212 b.ne .Lv8_entry 213 tst w16,#ARMV7_NEON 214 b.ne .Lneon_entry 215___ 216$code.=<<___ if ($SZ==8); 217 tst w16,#ARMV8_SHA512 218 b.ne .Lv8_entry 219___ 220$code.=<<___; 221#endif 222 AARCH64_SIGN_LINK_REGISTER 223 stp x29,x30,[sp,#-128]! 224 add x29,sp,#0 225 226 stp x19,x20,[sp,#16] 227 stp x21,x22,[sp,#32] 228 stp x23,x24,[sp,#48] 229 stp x25,x26,[sp,#64] 230 stp x27,x28,[sp,#80] 231 sub sp,sp,#4*$SZ 232 233 ldp $A,$B,[$ctx] // load context 234 ldp $C,$D,[$ctx,#2*$SZ] 235 ldp $E,$F,[$ctx,#4*$SZ] 236 add $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input 237 ldp $G,$H,[$ctx,#6*$SZ] 238 adrp $Ktbl,.LK$BITS 239 add $Ktbl,$Ktbl,:lo12:.LK$BITS 240 stp $ctx,$num,[x29,#96] 241 242.Loop: 243 ldp @X[0],@X[1],[$inp],#2*$SZ 244 ldr $t2,[$Ktbl],#$SZ // *K++ 245 eor $t3,$B,$C // magic seed 246 str $inp,[x29,#112] 247___ 248for ($i=0;$i<16;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 249$code.=".Loop_16_xx:\n"; 250for (;$i<32;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 251$code.=<<___; 252 cbnz $t2,.Loop_16_xx 253 254 ldp $ctx,$num,[x29,#96] 255 ldr $inp,[x29,#112] 256 sub $Ktbl,$Ktbl,#`$SZ*($rounds+1)` // rewind 257 258 ldp @X[0],@X[1],[$ctx] 259 ldp @X[2],@X[3],[$ctx,#2*$SZ] 260 add $inp,$inp,#14*$SZ // advance input pointer 261 ldp @X[4],@X[5],[$ctx,#4*$SZ] 262 add $A,$A,@X[0] 263 ldp @X[6],@X[7],[$ctx,#6*$SZ] 264 add $B,$B,@X[1] 265 add $C,$C,@X[2] 266 add $D,$D,@X[3] 267 stp $A,$B,[$ctx] 268 add $E,$E,@X[4] 269 add $F,$F,@X[5] 270 stp $C,$D,[$ctx,#2*$SZ] 271 add $G,$G,@X[6] 272 add $H,$H,@X[7] 273 cmp $inp,$num 274 stp $E,$F,[$ctx,#4*$SZ] 275 stp $G,$H,[$ctx,#6*$SZ] 276 b.ne .Loop 277 278 ldp x19,x20,[x29,#16] 279 add sp,sp,#4*$SZ 280 ldp x21,x22,[x29,#32] 281 ldp x23,x24,[x29,#48] 282 ldp x25,x26,[x29,#64] 283 ldp x27,x28,[x29,#80] 284 ldp x29,x30,[sp],#128 285 AARCH64_VALIDATE_LINK_REGISTER 286 ret 287.size $func,.-$func 288 289.rodata 290 291.align 6 292.type .LK$BITS,%object 293.LK$BITS: 294___ 295$code.=<<___ if ($SZ==8); 296 .quad 0x428a2f98d728ae22,0x7137449123ef65cd 297 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc 298 .quad 0x3956c25bf348b538,0x59f111f1b605d019 299 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118 300 .quad 0xd807aa98a3030242,0x12835b0145706fbe 301 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2 302 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1 303 .quad 0x9bdc06a725c71235,0xc19bf174cf692694 304 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3 305 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65 306 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483 307 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5 308 .quad 0x983e5152ee66dfab,0xa831c66d2db43210 309 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4 310 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725 311 .quad 0x06ca6351e003826f,0x142929670a0e6e70 312 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926 313 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df 314 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8 315 .quad 0x81c2c92e47edaee6,0x92722c851482353b 316 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001 317 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30 318 .quad 0xd192e819d6ef5218,0xd69906245565a910 319 .quad 0xf40e35855771202a,0x106aa07032bbd1b8 320 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53 321 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8 322 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb 323 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3 324 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60 325 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec 326 .quad 0x90befffa23631e28,0xa4506cebde82bde9 327 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b 328 .quad 0xca273eceea26619c,0xd186b8c721c0c207 329 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178 330 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6 331 .quad 0x113f9804bef90dae,0x1b710b35131c471b 332 .quad 0x28db77f523047d84,0x32caab7b40c72493 333 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c 334 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a 335 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817 336 .quad 0 // terminator 337___ 338$code.=<<___ if ($SZ==4); 339 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 340 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 341 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 342 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 343 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 344 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 345 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 346 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 347 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 348 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 349 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 350 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 351 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 352 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 353 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 354 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 355 .long 0 //terminator 356___ 357$code.=<<___; 358.size .LK$BITS,.-.LK$BITS 359.asciz "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>" 360.align 2 361 362.text 363___ 364 365if ($SZ==4) { 366my $Ktbl="x3"; 367 368my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2)); 369my @MSG=map("v$_.16b",(4..7)); 370my ($W0,$W1)=("v16.4s","v17.4s"); 371my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b"); 372 373$code.=<<___; 374#ifndef __KERNEL__ 375.type sha256_block_armv8,%function 376.align 6 377sha256_block_armv8: 378.Lv8_entry: 379 // Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later. 380 stp x29,x30,[sp,#-16]! 381 add x29,sp,#0 382 383 ld1.32 {$ABCD,$EFGH},[$ctx] 384 adrp $Ktbl,.LK256 385 add $Ktbl,$Ktbl,:lo12:.LK256 386 387.Loop_hw: 388 ld1 {@MSG[0]-@MSG[3]},[$inp],#64 389 sub $num,$num,#1 390 ld1.32 {$W0},[$Ktbl],#16 391 rev32 @MSG[0],@MSG[0] 392 rev32 @MSG[1],@MSG[1] 393 rev32 @MSG[2],@MSG[2] 394 rev32 @MSG[3],@MSG[3] 395 orr $ABCD_SAVE,$ABCD,$ABCD // offload 396 orr $EFGH_SAVE,$EFGH,$EFGH 397___ 398for($i=0;$i<12;$i++) { 399$code.=<<___; 400 ld1.32 {$W1},[$Ktbl],#16 401 add.i32 $W0,$W0,@MSG[0] 402 sha256su0 @MSG[0],@MSG[1] 403 orr $abcd,$ABCD,$ABCD 404 sha256h $ABCD,$EFGH,$W0 405 sha256h2 $EFGH,$abcd,$W0 406 sha256su1 @MSG[0],@MSG[2],@MSG[3] 407___ 408 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG)); 409} 410$code.=<<___; 411 ld1.32 {$W1},[$Ktbl],#16 412 add.i32 $W0,$W0,@MSG[0] 413 orr $abcd,$ABCD,$ABCD 414 sha256h $ABCD,$EFGH,$W0 415 sha256h2 $EFGH,$abcd,$W0 416 417 ld1.32 {$W0},[$Ktbl],#16 418 add.i32 $W1,$W1,@MSG[1] 419 orr $abcd,$ABCD,$ABCD 420 sha256h $ABCD,$EFGH,$W1 421 sha256h2 $EFGH,$abcd,$W1 422 423 ld1.32 {$W1},[$Ktbl] 424 add.i32 $W0,$W0,@MSG[2] 425 sub $Ktbl,$Ktbl,#$rounds*$SZ-16 // rewind 426 orr $abcd,$ABCD,$ABCD 427 sha256h $ABCD,$EFGH,$W0 428 sha256h2 $EFGH,$abcd,$W0 429 430 add.i32 $W1,$W1,@MSG[3] 431 orr $abcd,$ABCD,$ABCD 432 sha256h $ABCD,$EFGH,$W1 433 sha256h2 $EFGH,$abcd,$W1 434 435 add.i32 $ABCD,$ABCD,$ABCD_SAVE 436 add.i32 $EFGH,$EFGH,$EFGH_SAVE 437 438 cbnz $num,.Loop_hw 439 440 st1.32 {$ABCD,$EFGH},[$ctx] 441 442 ldr x29,[sp],#16 443 ret 444.size sha256_block_armv8,.-sha256_block_armv8 445#endif 446___ 447} 448 449if ($SZ==4) { ######################################### NEON stuff # 450# You'll surely note a lot of similarities with sha256-armv4 module, 451# and of course it's not a coincidence. sha256-armv4 was used as 452# initial template, but was adapted for ARMv8 instruction set and 453# extensively re-tuned for all-round performance. 454 455my @V = ($A,$B,$C,$D,$E,$F,$G,$H) = map("w$_",(3..10)); 456my ($t0,$t1,$t2,$t3,$t4) = map("w$_",(11..15)); 457my $Ktbl="x16"; 458my $Xfer="x17"; 459my @X = map("q$_",(0..3)); 460my ($T0,$T1,$T2,$T3,$T4,$T5,$T6,$T7) = map("q$_",(4..7,16..19)); 461my $j=0; 462 463sub AUTOLOAD() # thunk [simplified] x86-style perlasm 464{ my $opcode = $AUTOLOAD; $opcode =~ s/.*:://; $opcode =~ s/_/\./; 465 my $arg = pop; 466 $arg = "#$arg" if ($arg*1 eq $arg); 467 $code .= "\t$opcode\t".join(',',@_,$arg)."\n"; 468} 469 470sub Dscalar { shift =~ m|[qv]([0-9]+)|?"d$1":""; } 471sub Dlo { shift =~ m|[qv]([0-9]+)|?"v$1.d[0]":""; } 472sub Dhi { shift =~ m|[qv]([0-9]+)|?"v$1.d[1]":""; } 473 474sub Xupdate() 475{ use integer; 476 my $body = shift; 477 my @insns = (&$body,&$body,&$body,&$body); 478 my ($a,$b,$c,$d,$e,$f,$g,$h); 479 480 &ext_8 ($T0,@X[0],@X[1],4); # X[1..4] 481 eval(shift(@insns)); 482 eval(shift(@insns)); 483 eval(shift(@insns)); 484 &ext_8 ($T3,@X[2],@X[3],4); # X[9..12] 485 eval(shift(@insns)); 486 eval(shift(@insns)); 487 &mov (&Dscalar($T7),&Dhi(@X[3])); # X[14..15] 488 eval(shift(@insns)); 489 eval(shift(@insns)); 490 &ushr_32 ($T2,$T0,$sigma0[0]); 491 eval(shift(@insns)); 492 &ushr_32 ($T1,$T0,$sigma0[2]); 493 eval(shift(@insns)); 494 &add_32 (@X[0],@X[0],$T3); # X[0..3] += X[9..12] 495 eval(shift(@insns)); 496 &sli_32 ($T2,$T0,32-$sigma0[0]); 497 eval(shift(@insns)); 498 eval(shift(@insns)); 499 &ushr_32 ($T3,$T0,$sigma0[1]); 500 eval(shift(@insns)); 501 eval(shift(@insns)); 502 &eor_8 ($T1,$T1,$T2); 503 eval(shift(@insns)); 504 eval(shift(@insns)); 505 &sli_32 ($T3,$T0,32-$sigma0[1]); 506 eval(shift(@insns)); 507 eval(shift(@insns)); 508 &ushr_32 ($T4,$T7,$sigma1[0]); 509 eval(shift(@insns)); 510 eval(shift(@insns)); 511 &eor_8 ($T1,$T1,$T3); # sigma0(X[1..4]) 512 eval(shift(@insns)); 513 eval(shift(@insns)); 514 &sli_32 ($T4,$T7,32-$sigma1[0]); 515 eval(shift(@insns)); 516 eval(shift(@insns)); 517 &ushr_32 ($T5,$T7,$sigma1[2]); 518 eval(shift(@insns)); 519 eval(shift(@insns)); 520 &ushr_32 ($T3,$T7,$sigma1[1]); 521 eval(shift(@insns)); 522 eval(shift(@insns)); 523 &add_32 (@X[0],@X[0],$T1); # X[0..3] += sigma0(X[1..4]) 524 eval(shift(@insns)); 525 eval(shift(@insns)); 526 &sli_u32 ($T3,$T7,32-$sigma1[1]); 527 eval(shift(@insns)); 528 eval(shift(@insns)); 529 &eor_8 ($T5,$T5,$T4); 530 eval(shift(@insns)); 531 eval(shift(@insns)); 532 eval(shift(@insns)); 533 &eor_8 ($T5,$T5,$T3); # sigma1(X[14..15]) 534 eval(shift(@insns)); 535 eval(shift(@insns)); 536 eval(shift(@insns)); 537 &add_32 (@X[0],@X[0],$T5); # X[0..1] += sigma1(X[14..15]) 538 eval(shift(@insns)); 539 eval(shift(@insns)); 540 eval(shift(@insns)); 541 &ushr_32 ($T6,@X[0],$sigma1[0]); 542 eval(shift(@insns)); 543 &ushr_32 ($T7,@X[0],$sigma1[2]); 544 eval(shift(@insns)); 545 eval(shift(@insns)); 546 &sli_32 ($T6,@X[0],32-$sigma1[0]); 547 eval(shift(@insns)); 548 &ushr_32 ($T5,@X[0],$sigma1[1]); 549 eval(shift(@insns)); 550 eval(shift(@insns)); 551 &eor_8 ($T7,$T7,$T6); 552 eval(shift(@insns)); 553 eval(shift(@insns)); 554 &sli_32 ($T5,@X[0],32-$sigma1[1]); 555 eval(shift(@insns)); 556 eval(shift(@insns)); 557 &ld1_32 ("{$T0}","[$Ktbl], #16"); 558 eval(shift(@insns)); 559 &eor_8 ($T7,$T7,$T5); # sigma1(X[16..17]) 560 eval(shift(@insns)); 561 eval(shift(@insns)); 562 &eor_8 ($T5,$T5,$T5); 563 eval(shift(@insns)); 564 eval(shift(@insns)); 565 &mov (&Dhi($T5), &Dlo($T7)); 566 eval(shift(@insns)); 567 eval(shift(@insns)); 568 eval(shift(@insns)); 569 &add_32 (@X[0],@X[0],$T5); # X[2..3] += sigma1(X[16..17]) 570 eval(shift(@insns)); 571 eval(shift(@insns)); 572 eval(shift(@insns)); 573 &add_32 ($T0,$T0,@X[0]); 574 while($#insns>=1) { eval(shift(@insns)); } 575 &st1_32 ("{$T0}","[$Xfer], #16"); 576 eval(shift(@insns)); 577 578 push(@X,shift(@X)); # "rotate" X[] 579} 580 581sub Xpreload() 582{ use integer; 583 my $body = shift; 584 my @insns = (&$body,&$body,&$body,&$body); 585 my ($a,$b,$c,$d,$e,$f,$g,$h); 586 587 eval(shift(@insns)); 588 eval(shift(@insns)); 589 &ld1_8 ("{@X[0]}","[$inp],#16"); 590 eval(shift(@insns)); 591 eval(shift(@insns)); 592 &ld1_32 ("{$T0}","[$Ktbl],#16"); 593 eval(shift(@insns)); 594 eval(shift(@insns)); 595 eval(shift(@insns)); 596 eval(shift(@insns)); 597 &rev32 (@X[0],@X[0]); 598 eval(shift(@insns)); 599 eval(shift(@insns)); 600 eval(shift(@insns)); 601 eval(shift(@insns)); 602 &add_32 ($T0,$T0,@X[0]); 603 foreach (@insns) { eval; } # remaining instructions 604 &st1_32 ("{$T0}","[$Xfer], #16"); 605 606 push(@X,shift(@X)); # "rotate" X[] 607} 608 609sub body_00_15 () { 610 ( 611 '($a,$b,$c,$d,$e,$f,$g,$h)=@V;'. 612 '&add ($h,$h,$t1)', # h+=X[i]+K[i] 613 '&add ($a,$a,$t4);'. # h+=Sigma0(a) from the past 614 '&and ($t1,$f,$e)', 615 '&bic ($t4,$g,$e)', 616 '&eor ($t0,$e,$e,"ror#".($Sigma1[1]-$Sigma1[0]))', 617 '&add ($a,$a,$t2)', # h+=Maj(a,b,c) from the past 618 '&orr ($t1,$t1,$t4)', # Ch(e,f,g) 619 '&eor ($t0,$t0,$e,"ror#".($Sigma1[2]-$Sigma1[0]))', # Sigma1(e) 620 '&eor ($t4,$a,$a,"ror#".($Sigma0[1]-$Sigma0[0]))', 621 '&add ($h,$h,$t1)', # h+=Ch(e,f,g) 622 '&ror ($t0,$t0,"#$Sigma1[0]")', 623 '&eor ($t2,$a,$b)', # a^b, b^c in next round 624 '&eor ($t4,$t4,$a,"ror#".($Sigma0[2]-$Sigma0[0]))', # Sigma0(a) 625 '&add ($h,$h,$t0)', # h+=Sigma1(e) 626 '&ldr ($t1,sprintf "[sp,#%d]",4*(($j+1)&15)) if (($j&15)!=15);'. 627 '&ldr ($t1,"[$Ktbl]") if ($j==15);'. 628 '&and ($t3,$t3,$t2)', # (b^c)&=(a^b) 629 '&ror ($t4,$t4,"#$Sigma0[0]")', 630 '&add ($d,$d,$h)', # d+=h 631 '&eor ($t3,$t3,$b)', # Maj(a,b,c) 632 '$j++; unshift(@V,pop(@V)); ($t2,$t3)=($t3,$t2);' 633 ) 634} 635 636$code.=<<___; 637#ifdef __KERNEL__ 638.globl sha256_block_neon 639#endif 640.type sha256_block_neon,%function 641.align 4 642sha256_block_neon: 643 AARCH64_VALID_CALL_TARGET 644.Lneon_entry: 645 // Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later 646 stp x29, x30, [sp, #-16]! 647 mov x29, sp 648 sub sp,sp,#16*4 649 650 adrp $Ktbl,.LK256 651 add $Ktbl,$Ktbl,:lo12:.LK256 652 add $num,$inp,$num,lsl#6 // len to point at the end of inp 653 654 ld1.8 {@X[0]},[$inp], #16 655 ld1.8 {@X[1]},[$inp], #16 656 ld1.8 {@X[2]},[$inp], #16 657 ld1.8 {@X[3]},[$inp], #16 658 ld1.32 {$T0},[$Ktbl], #16 659 ld1.32 {$T1},[$Ktbl], #16 660 ld1.32 {$T2},[$Ktbl], #16 661 ld1.32 {$T3},[$Ktbl], #16 662 rev32 @X[0],@X[0] // yes, even on 663 rev32 @X[1],@X[1] // big-endian 664 rev32 @X[2],@X[2] 665 rev32 @X[3],@X[3] 666 mov $Xfer,sp 667 add.32 $T0,$T0,@X[0] 668 add.32 $T1,$T1,@X[1] 669 add.32 $T2,$T2,@X[2] 670 st1.32 {$T0-$T1},[$Xfer], #32 671 add.32 $T3,$T3,@X[3] 672 st1.32 {$T2-$T3},[$Xfer] 673 sub $Xfer,$Xfer,#32 674 675 ldp $A,$B,[$ctx] 676 ldp $C,$D,[$ctx,#8] 677 ldp $E,$F,[$ctx,#16] 678 ldp $G,$H,[$ctx,#24] 679 ldr $t1,[sp,#0] 680 mov $t2,wzr 681 eor $t3,$B,$C 682 mov $t4,wzr 683 b .L_00_48 684 685.align 4 686.L_00_48: 687___ 688 &Xupdate(\&body_00_15); 689 &Xupdate(\&body_00_15); 690 &Xupdate(\&body_00_15); 691 &Xupdate(\&body_00_15); 692$code.=<<___; 693 cmp $t1,#0 // check for K256 terminator 694 ldr $t1,[sp,#0] 695 sub $Xfer,$Xfer,#64 696 bne .L_00_48 697 698 sub $Ktbl,$Ktbl,#256 // rewind $Ktbl 699 cmp $inp,$num 700 mov $Xfer, #64 701 csel $Xfer, $Xfer, xzr, eq 702 sub $inp,$inp,$Xfer // avoid SEGV 703 mov $Xfer,sp 704___ 705 &Xpreload(\&body_00_15); 706 &Xpreload(\&body_00_15); 707 &Xpreload(\&body_00_15); 708 &Xpreload(\&body_00_15); 709$code.=<<___; 710 add $A,$A,$t4 // h+=Sigma0(a) from the past 711 ldp $t0,$t1,[$ctx,#0] 712 add $A,$A,$t2 // h+=Maj(a,b,c) from the past 713 ldp $t2,$t3,[$ctx,#8] 714 add $A,$A,$t0 // accumulate 715 add $B,$B,$t1 716 ldp $t0,$t1,[$ctx,#16] 717 add $C,$C,$t2 718 add $D,$D,$t3 719 ldp $t2,$t3,[$ctx,#24] 720 add $E,$E,$t0 721 add $F,$F,$t1 722 ldr $t1,[sp,#0] 723 stp $A,$B,[$ctx,#0] 724 add $G,$G,$t2 725 mov $t2,wzr 726 stp $C,$D,[$ctx,#8] 727 add $H,$H,$t3 728 stp $E,$F,[$ctx,#16] 729 eor $t3,$B,$C 730 stp $G,$H,[$ctx,#24] 731 mov $t4,wzr 732 mov $Xfer,sp 733 b.ne .L_00_48 734 735 ldr x29,[x29] 736 add sp,sp,#16*4+16 737 ret 738.size sha256_block_neon,.-sha256_block_neon 739___ 740} 741 742if ($SZ==8) { 743my $Ktbl="x3"; 744 745my @H = map("v$_.16b",(0..4)); 746my ($fg,$de,$m9_10)=map("v$_.16b",(5..7)); 747my @MSG=map("v$_.16b",(16..23)); 748my ($W0,$W1)=("v24.2d","v25.2d"); 749my ($AB,$CD,$EF,$GH)=map("v$_.16b",(26..29)); 750 751$code.=<<___; 752#ifndef __KERNEL__ 753.type sha512_block_armv8,%function 754.align 6 755sha512_block_armv8: 756.Lv8_entry: 757 // Armv8.3-A PAuth: even though x30 is pushed to stack it is not popped later 758 stp x29,x30,[sp,#-16]! 759 add x29,sp,#0 760 761 ld1 {@MSG[0]-@MSG[3]},[$inp],#64 // load input 762 ld1 {@MSG[4]-@MSG[7]},[$inp],#64 763 764 ld1.64 {@H[0]-@H[3]},[$ctx] // load context 765 adrp $Ktbl,.LK512 766 add $Ktbl,$Ktbl,:lo12:.LK512 767 768 rev64 @MSG[0],@MSG[0] 769 rev64 @MSG[1],@MSG[1] 770 rev64 @MSG[2],@MSG[2] 771 rev64 @MSG[3],@MSG[3] 772 rev64 @MSG[4],@MSG[4] 773 rev64 @MSG[5],@MSG[5] 774 rev64 @MSG[6],@MSG[6] 775 rev64 @MSG[7],@MSG[7] 776 b .Loop_hw 777 778.align 4 779.Loop_hw: 780 ld1.64 {$W0},[$Ktbl],#16 781 subs $num,$num,#1 782 sub x4,$inp,#128 783 orr $AB,@H[0],@H[0] // offload 784 orr $CD,@H[1],@H[1] 785 orr $EF,@H[2],@H[2] 786 orr $GH,@H[3],@H[3] 787 csel $inp,$inp,x4,ne // conditional rewind 788___ 789for($i=0;$i<32;$i++) { 790$code.=<<___; 791 add.i64 $W0,$W0,@MSG[0] 792 ld1.64 {$W1},[$Ktbl],#16 793 ext $W0,$W0,$W0,#8 794 ext $fg,@H[2],@H[3],#8 795 ext $de,@H[1],@H[2],#8 796 add.i64 @H[3],@H[3],$W0 // "T1 + H + K512[i]" 797 sha512su0 @MSG[0],@MSG[1] 798 ext $m9_10,@MSG[4],@MSG[5],#8 799 sha512h @H[3],$fg,$de 800 sha512su1 @MSG[0],@MSG[7],$m9_10 801 add.i64 @H[4],@H[1],@H[3] // "D + T1" 802 sha512h2 @H[3],$H[1],@H[0] 803___ 804 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG)); 805 @H = (@H[3],@H[0],@H[4],@H[2],@H[1]); 806} 807for(;$i<40;$i++) { 808$code.=<<___ if ($i<39); 809 ld1.64 {$W1},[$Ktbl],#16 810___ 811$code.=<<___ if ($i==39); 812 sub $Ktbl,$Ktbl,#$rounds*$SZ // rewind 813___ 814$code.=<<___; 815 add.i64 $W0,$W0,@MSG[0] 816 ld1 {@MSG[0]},[$inp],#16 // load next input 817 ext $W0,$W0,$W0,#8 818 ext $fg,@H[2],@H[3],#8 819 ext $de,@H[1],@H[2],#8 820 add.i64 @H[3],@H[3],$W0 // "T1 + H + K512[i]" 821 sha512h @H[3],$fg,$de 822 rev64 @MSG[0],@MSG[0] 823 add.i64 @H[4],@H[1],@H[3] // "D + T1" 824 sha512h2 @H[3],$H[1],@H[0] 825___ 826 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG)); 827 @H = (@H[3],@H[0],@H[4],@H[2],@H[1]); 828} 829$code.=<<___; 830 add.i64 @H[0],@H[0],$AB // accumulate 831 add.i64 @H[1],@H[1],$CD 832 add.i64 @H[2],@H[2],$EF 833 add.i64 @H[3],@H[3],$GH 834 835 cbnz $num,.Loop_hw 836 837 st1.64 {@H[0]-@H[3]},[$ctx] // store context 838 839 ldr x29,[sp],#16 840 ret 841.size sha512_block_armv8,.-sha512_block_armv8 842#endif 843___ 844} 845 846{ my %opcode = ( 847 "sha256h" => 0x5e004000, "sha256h2" => 0x5e005000, 848 "sha256su0" => 0x5e282800, "sha256su1" => 0x5e006000 ); 849 850 sub unsha256 { 851 my ($mnemonic,$arg)=@_; 852 853 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o 854 && 855 sprintf ".inst\t0x%08x\t//%s %s", 856 $opcode{$mnemonic}|$1|($2<<5)|($3<<16), 857 $mnemonic,$arg; 858 } 859} 860 861{ my %opcode = ( 862 "sha512h" => 0xce608000, "sha512h2" => 0xce608400, 863 "sha512su0" => 0xcec08000, "sha512su1" => 0xce608800 ); 864 865 sub unsha512 { 866 my ($mnemonic,$arg)=@_; 867 868 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o 869 && 870 sprintf ".inst\t0x%08x\t//%s %s", 871 $opcode{$mnemonic}|$1|($2<<5)|($3<<16), 872 $mnemonic,$arg; 873 } 874} 875 876open SELF,$0; 877while(<SELF>) { 878 next if (/^#!/); 879 last if (!s/^#/\/\// and !/^$/); 880 print; 881} 882close SELF; 883 884foreach(split("\n",$code)) { 885 886 s/\`([^\`]*)\`/eval($1)/ge; 887 888 s/\b(sha512\w+)\s+([qv].*)/unsha512($1,$2)/ge or 889 s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/ge; 890 891 s/\bq([0-9]+)\b/v$1.16b/g; # old->new registers 892 893 s/\.[ui]?8(\s)/$1/; 894 s/\.\w?64\b// and s/\.16b/\.2d/g or 895 s/\.\w?32\b// and s/\.16b/\.4s/g; 896 m/\bext\b/ and s/\.2d/\.16b/g or 897 m/(ld|st)1[^\[]+\[0\]/ and s/\.4s/\.s/g; 898 899 print $_,"\n"; 900} 901 902close STDOUT or die "error closing STDOUT: $!"; 903
