1#!/usr/bin/env perl 2# 3# ==================================================================== 4# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL 5# project. The module is, however, dual licensed under OpenSSL and 6# CRYPTOGAMS licenses depending on where you obtain it. For further 7# details see http://www.openssl.org/~appro/cryptogams/. 8# ==================================================================== 9# 10# SHA256/512 for ARMv8. 11# 12# Performance in cycles per processed byte and improvement coefficient 13# over code generated with "default" compiler: 14# 15# SHA256-hw SHA256(*) SHA512 16# Apple A7 1.97 10.5 (+33%) 6.73 (-1%(**)) 17# Cortex-A53 2.38 15.5 (+115%) 10.0 (+150%(***)) 18# Cortex-A57 2.31 11.6 (+86%) 7.51 (+260%(***)) 19# Denver 2.01 10.5 (+26%) 6.70 (+8%) 20# X-Gene 20.0 (+100%) 12.8 (+300%(***)) 21# 22# (*) Software SHA256 results are of lesser relevance, presented 23# mostly for informational purposes. 24# (**) The result is a trade-off: it's possible to improve it by 25# 10% (or by 1 cycle per round), but at the cost of 20% loss 26# on Cortex-A53 (or by 4 cycles per round). 27# (***) Super-impressive coefficients over gcc-generated code are 28# indication of some compiler "pathology", most notably code 29# generated with -mgeneral-regs-only is significanty faster 30# and the gap is only 40-90%. 31 32$flavour=shift; 33# Unlike most perlasm files, sha512-armv8.pl takes an additional argument to 34# determine which hash function to emit. This differs from upstream OpenSSL so 35# that the script may continue to output to stdout. 36$variant=shift; 37$output=shift; 38 39$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 40( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or 41( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or 42die "can't locate arm-xlate.pl"; 43 44open OUT,"| \"$^X\" $xlate $flavour $output"; 45*STDOUT=*OUT; 46 47if ($variant eq "sha512") { 48 $BITS=512; 49 $SZ=8; 50 @Sigma0=(28,34,39); 51 @Sigma1=(14,18,41); 52 @sigma0=(1, 8, 7); 53 @sigma1=(19,61, 6); 54 $rounds=80; 55 $reg_t="x"; 56} elsif ($variant eq "sha256") { 57 $BITS=256; 58 $SZ=4; 59 @Sigma0=( 2,13,22); 60 @Sigma1=( 6,11,25); 61 @sigma0=( 7,18, 3); 62 @sigma1=(17,19,10); 63 $rounds=64; 64 $reg_t="w"; 65} else { 66 die "Unknown variant: $variant"; 67} 68 69$func="sha${BITS}_block_data_order"; 70 71($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30)); 72 73@X=map("$reg_t$_",(3..15,0..2)); 74@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27)); 75($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28)); 76 77sub BODY_00_xx { 78my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_; 79my $j=($i+1)&15; 80my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]); 81 $T0=@X[$i+3] if ($i<11); 82 83$code.=<<___ if ($i<16); 84#ifndef __ARMEB__ 85 rev @X[$i],@X[$i] // $i 86#endif 87___ 88$code.=<<___ if ($i<13 && ($i&1)); 89 ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ 90___ 91$code.=<<___ if ($i==13); 92 ldp @X[14],@X[15],[$inp] 93___ 94$code.=<<___ if ($i>=14); 95 ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`] 96___ 97$code.=<<___ if ($i>0 && $i<16); 98 add $a,$a,$t1 // h+=Sigma0(a) 99___ 100$code.=<<___ if ($i>=11); 101 str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`] 102___ 103# While ARMv8 specifies merged rotate-n-logical operation such as 104# 'eor x,y,z,ror#n', it was found to negatively affect performance 105# on Apple A7. The reason seems to be that it requires even 'y' to 106# be available earlier. This means that such merged instruction is 107# not necessarily best choice on critical path... On the other hand 108# Cortex-A5x handles merged instructions much better than disjoint 109# rotate and logical... See (**) footnote above. 110$code.=<<___ if ($i<15); 111 ror $t0,$e,#$Sigma1[0] 112 add $h,$h,$t2 // h+=K[i] 113 eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]` 114 and $t1,$f,$e 115 bic $t2,$g,$e 116 add $h,$h,@X[$i&15] // h+=X[i] 117 orr $t1,$t1,$t2 // Ch(e,f,g) 118 eor $t2,$a,$b // a^b, b^c in next round 119 eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e) 120 ror $T0,$a,#$Sigma0[0] 121 add $h,$h,$t1 // h+=Ch(e,f,g) 122 eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]` 123 add $h,$h,$t0 // h+=Sigma1(e) 124 and $t3,$t3,$t2 // (b^c)&=(a^b) 125 add $d,$d,$h // d+=h 126 eor $t3,$t3,$b // Maj(a,b,c) 127 eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a) 128 add $h,$h,$t3 // h+=Maj(a,b,c) 129 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 130 //add $h,$h,$t1 // h+=Sigma0(a) 131___ 132$code.=<<___ if ($i>=15); 133 ror $t0,$e,#$Sigma1[0] 134 add $h,$h,$t2 // h+=K[i] 135 ror $T1,@X[($j+1)&15],#$sigma0[0] 136 and $t1,$f,$e 137 ror $T2,@X[($j+14)&15],#$sigma1[0] 138 bic $t2,$g,$e 139 ror $T0,$a,#$Sigma0[0] 140 add $h,$h,@X[$i&15] // h+=X[i] 141 eor $t0,$t0,$e,ror#$Sigma1[1] 142 eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1] 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,$e,ror#$Sigma1[2] // Sigma1(e) 146 eor $T0,$T0,$a,ror#$Sigma0[1] 147 add $h,$h,$t1 // h+=Ch(e,f,g) 148 and $t3,$t3,$t2 // (b^c)&=(a^b) 149 eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1] 150 eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1]) 151 add $h,$h,$t0 // h+=Sigma1(e) 152 eor $t3,$t3,$b // Maj(a,b,c) 153 eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a) 154 eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14]) 155 add @X[$j],@X[$j],@X[($j+9)&15] 156 add $d,$d,$h // d+=h 157 add $h,$h,$t3 // h+=Maj(a,b,c) 158 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 159 add @X[$j],@X[$j],$T1 160 add $h,$h,$t1 // h+=Sigma0(a) 161 add @X[$j],@X[$j],$T2 162___ 163 ($t2,$t3)=($t3,$t2); 164} 165 166$code.=<<___; 167#include "arm_arch.h" 168 169.text 170 171.extern OPENSSL_armcap_P 172.globl $func 173.type $func,%function 174.align 6 175$func: 176___ 177$code.=<<___ if ($SZ==4); 178 ldr x16,.LOPENSSL_armcap_P 179 adr x17,.LOPENSSL_armcap_P 180 add x16,x16,x17 181 ldr w16,[x16] 182 tst w16,#ARMV8_SHA256 183 b.ne .Lv8_entry 184___ 185$code.=<<___; 186 stp x29,x30,[sp,#-128]! 187 add x29,sp,#0 188 189 stp x19,x20,[sp,#16] 190 stp x21,x22,[sp,#32] 191 stp x23,x24,[sp,#48] 192 stp x25,x26,[sp,#64] 193 stp x27,x28,[sp,#80] 194 sub sp,sp,#4*$SZ 195 196 ldp $A,$B,[$ctx] // load context 197 ldp $C,$D,[$ctx,#2*$SZ] 198 ldp $E,$F,[$ctx,#4*$SZ] 199 add $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input 200 ldp $G,$H,[$ctx,#6*$SZ] 201 adr $Ktbl,.LK$BITS 202 stp $ctx,$num,[x29,#96] 203 204.Loop: 205 ldp @X[0],@X[1],[$inp],#2*$SZ 206 ldr $t2,[$Ktbl],#$SZ // *K++ 207 eor $t3,$B,$C // magic seed 208 str $inp,[x29,#112] 209___ 210for ($i=0;$i<16;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 211$code.=".Loop_16_xx:\n"; 212for (;$i<32;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 213$code.=<<___; 214 cbnz $t2,.Loop_16_xx 215 216 ldp $ctx,$num,[x29,#96] 217 ldr $inp,[x29,#112] 218 sub $Ktbl,$Ktbl,#`$SZ*($rounds+1)` // rewind 219 220 ldp @X[0],@X[1],[$ctx] 221 ldp @X[2],@X[3],[$ctx,#2*$SZ] 222 add $inp,$inp,#14*$SZ // advance input pointer 223 ldp @X[4],@X[5],[$ctx,#4*$SZ] 224 add $A,$A,@X[0] 225 ldp @X[6],@X[7],[$ctx,#6*$SZ] 226 add $B,$B,@X[1] 227 add $C,$C,@X[2] 228 add $D,$D,@X[3] 229 stp $A,$B,[$ctx] 230 add $E,$E,@X[4] 231 add $F,$F,@X[5] 232 stp $C,$D,[$ctx,#2*$SZ] 233 add $G,$G,@X[6] 234 add $H,$H,@X[7] 235 cmp $inp,$num 236 stp $E,$F,[$ctx,#4*$SZ] 237 stp $G,$H,[$ctx,#6*$SZ] 238 b.ne .Loop 239 240 ldp x19,x20,[x29,#16] 241 add sp,sp,#4*$SZ 242 ldp x21,x22,[x29,#32] 243 ldp x23,x24,[x29,#48] 244 ldp x25,x26,[x29,#64] 245 ldp x27,x28,[x29,#80] 246 ldp x29,x30,[sp],#128 247 ret 248.size $func,.-$func 249 250.align 6 251.type .LK$BITS,%object 252.LK$BITS: 253___ 254$code.=<<___ if ($SZ==8); 255 .quad 0x428a2f98d728ae22,0x7137449123ef65cd 256 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc 257 .quad 0x3956c25bf348b538,0x59f111f1b605d019 258 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118 259 .quad 0xd807aa98a3030242,0x12835b0145706fbe 260 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2 261 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1 262 .quad 0x9bdc06a725c71235,0xc19bf174cf692694 263 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3 264 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65 265 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483 266 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5 267 .quad 0x983e5152ee66dfab,0xa831c66d2db43210 268 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4 269 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725 270 .quad 0x06ca6351e003826f,0x142929670a0e6e70 271 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926 272 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df 273 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8 274 .quad 0x81c2c92e47edaee6,0x92722c851482353b 275 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001 276 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30 277 .quad 0xd192e819d6ef5218,0xd69906245565a910 278 .quad 0xf40e35855771202a,0x106aa07032bbd1b8 279 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53 280 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8 281 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb 282 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3 283 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60 284 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec 285 .quad 0x90befffa23631e28,0xa4506cebde82bde9 286 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b 287 .quad 0xca273eceea26619c,0xd186b8c721c0c207 288 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178 289 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6 290 .quad 0x113f9804bef90dae,0x1b710b35131c471b 291 .quad 0x28db77f523047d84,0x32caab7b40c72493 292 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c 293 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a 294 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817 295 .quad 0 // terminator 296___ 297$code.=<<___ if ($SZ==4); 298 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 299 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 300 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 301 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 302 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 303 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 304 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 305 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 306 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 307 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 308 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 309 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 310 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 311 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 312 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 313 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 314 .long 0 //terminator 315___ 316$code.=<<___; 317.size .LK$BITS,.-.LK$BITS 318.align 3 319.LOPENSSL_armcap_P: 320 .quad OPENSSL_armcap_P-. 321.asciz "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>" 322.align 2 323___ 324 325if ($SZ==4) { 326my $Ktbl="x3"; 327 328my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2)); 329my @MSG=map("v$_.16b",(4..7)); 330my ($W0,$W1)=("v16.4s","v17.4s"); 331my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b"); 332 333$code.=<<___; 334.type sha256_block_armv8,%function 335.align 6 336sha256_block_armv8: 337.Lv8_entry: 338 stp x29,x30,[sp,#-16]! 339 add x29,sp,#0 340 341 ld1.32 {$ABCD,$EFGH},[$ctx] 342 adr $Ktbl,.LK256 343 344.Loop_hw: 345 ld1 {@MSG[0]-@MSG[3]},[$inp],#64 346 sub $num,$num,#1 347 ld1.32 {$W0},[$Ktbl],#16 348 rev32 @MSG[0],@MSG[0] 349 rev32 @MSG[1],@MSG[1] 350 rev32 @MSG[2],@MSG[2] 351 rev32 @MSG[3],@MSG[3] 352 orr $ABCD_SAVE,$ABCD,$ABCD // offload 353 orr $EFGH_SAVE,$EFGH,$EFGH 354___ 355for($i=0;$i<12;$i++) { 356$code.=<<___; 357 ld1.32 {$W1},[$Ktbl],#16 358 add.i32 $W0,$W0,@MSG[0] 359 sha256su0 @MSG[0],@MSG[1] 360 orr $abcd,$ABCD,$ABCD 361 sha256h $ABCD,$EFGH,$W0 362 sha256h2 $EFGH,$abcd,$W0 363 sha256su1 @MSG[0],@MSG[2],@MSG[3] 364___ 365 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG)); 366} 367$code.=<<___; 368 ld1.32 {$W1},[$Ktbl],#16 369 add.i32 $W0,$W0,@MSG[0] 370 orr $abcd,$ABCD,$ABCD 371 sha256h $ABCD,$EFGH,$W0 372 sha256h2 $EFGH,$abcd,$W0 373 374 ld1.32 {$W0},[$Ktbl],#16 375 add.i32 $W1,$W1,@MSG[1] 376 orr $abcd,$ABCD,$ABCD 377 sha256h $ABCD,$EFGH,$W1 378 sha256h2 $EFGH,$abcd,$W1 379 380 ld1.32 {$W1},[$Ktbl] 381 add.i32 $W0,$W0,@MSG[2] 382 sub $Ktbl,$Ktbl,#$rounds*$SZ-16 // rewind 383 orr $abcd,$ABCD,$ABCD 384 sha256h $ABCD,$EFGH,$W0 385 sha256h2 $EFGH,$abcd,$W0 386 387 add.i32 $W1,$W1,@MSG[3] 388 orr $abcd,$ABCD,$ABCD 389 sha256h $ABCD,$EFGH,$W1 390 sha256h2 $EFGH,$abcd,$W1 391 392 add.i32 $ABCD,$ABCD,$ABCD_SAVE 393 add.i32 $EFGH,$EFGH,$EFGH_SAVE 394 395 cbnz $num,.Loop_hw 396 397 st1.32 {$ABCD,$EFGH},[$ctx] 398 399 ldr x29,[sp],#16 400 ret 401.size sha256_block_armv8,.-sha256_block_armv8 402___ 403} 404 405$code.=<<___; 406.comm OPENSSL_armcap_P,4,4 407___ 408 409{ my %opcode = ( 410 "sha256h" => 0x5e004000, "sha256h2" => 0x5e005000, 411 "sha256su0" => 0x5e282800, "sha256su1" => 0x5e006000 ); 412 413 sub unsha256 { 414 my ($mnemonic,$arg)=@_; 415 416 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o 417 && 418 sprintf ".inst\t0x%08x\t//%s %s", 419 $opcode{$mnemonic}|$1|($2<<5)|($3<<16), 420 $mnemonic,$arg; 421 } 422} 423 424foreach(split("\n",$code)) { 425 426 s/\`([^\`]*)\`/eval($1)/geo; 427 428 s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo; 429 430 s/\.\w?32\b//o and s/\.16b/\.4s/go; 431 m/(ld|st)1[^\[]+\[0\]/o and s/\.4s/\.s/go; 432 433 print $_,"\n"; 434} 435 436close STDOUT; 437