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$flavour=shift; 33$output=shift; 34open STDOUT,">$output"; 35 36if ($output =~ /512/) { 37 $BITS=512; 38 $SZ=8; 39 @Sigma0=(28,34,39); 40 @Sigma1=(14,18,41); 41 @sigma0=(1, 8, 7); 42 @sigma1=(19,61, 6); 43 $rounds=80; 44 $reg_t="x"; 45} else { 46 $BITS=256; 47 $SZ=4; 48 @Sigma0=( 2,13,22); 49 @Sigma1=( 6,11,25); 50 @sigma0=( 7,18, 3); 51 @sigma1=(17,19,10); 52 $rounds=64; 53 $reg_t="w"; 54} 55 56$func="sha${BITS}_block_data_order"; 57 58($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30)); 59 60@X=map("$reg_t$_",(3..15,0..2)); 61@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27)); 62($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28)); 63 64sub BODY_00_xx { 65my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_; 66my $j=($i+1)&15; 67my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]); 68 $T0=@X[$i+3] if ($i<11); 69 70$code.=<<___ if ($i<16); 71#ifndef __ARMEB__ 72 rev @X[$i],@X[$i] // $i 73#endif 74___ 75$code.=<<___ if ($i<13 && ($i&1)); 76 ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ 77___ 78$code.=<<___ if ($i==13); 79 ldp @X[14],@X[15],[$inp] 80___ 81$code.=<<___ if ($i>=14); 82 ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`] 83___ 84$code.=<<___ if ($i>0 && $i<16); 85 add $a,$a,$t1 // h+=Sigma0(a) 86___ 87$code.=<<___ if ($i>=11); 88 str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`] 89___ 90# While ARMv8 specifies merged rotate-n-logical operation such as 91# 'eor x,y,z,ror#n', it was found to negatively affect performance 92# on Apple A7. The reason seems to be that it requires even 'y' to 93# be available earlier. This means that such merged instruction is 94# not necessarily best choice on critical path... On the other hand 95# Cortex-A5x handles merged instructions much better than disjoint 96# rotate and logical... See (**) footnote above. 97$code.=<<___ if ($i<15); 98 ror $t0,$e,#$Sigma1[0] 99 add $h,$h,$t2 // h+=K[i] 100 eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]` 101 and $t1,$f,$e 102 bic $t2,$g,$e 103 add $h,$h,@X[$i&15] // h+=X[i] 104 orr $t1,$t1,$t2 // Ch(e,f,g) 105 eor $t2,$a,$b // a^b, b^c in next round 106 eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e) 107 ror $T0,$a,#$Sigma0[0] 108 add $h,$h,$t1 // h+=Ch(e,f,g) 109 eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]` 110 add $h,$h,$t0 // h+=Sigma1(e) 111 and $t3,$t3,$t2 // (b^c)&=(a^b) 112 add $d,$d,$h // d+=h 113 eor $t3,$t3,$b // Maj(a,b,c) 114 eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a) 115 add $h,$h,$t3 // h+=Maj(a,b,c) 116 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 117 //add $h,$h,$t1 // h+=Sigma0(a) 118___ 119$code.=<<___ if ($i>=15); 120 ror $t0,$e,#$Sigma1[0] 121 add $h,$h,$t2 // h+=K[i] 122 ror $T1,@X[($j+1)&15],#$sigma0[0] 123 and $t1,$f,$e 124 ror $T2,@X[($j+14)&15],#$sigma1[0] 125 bic $t2,$g,$e 126 ror $T0,$a,#$Sigma0[0] 127 add $h,$h,@X[$i&15] // h+=X[i] 128 eor $t0,$t0,$e,ror#$Sigma1[1] 129 eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1] 130 orr $t1,$t1,$t2 // Ch(e,f,g) 131 eor $t2,$a,$b // a^b, b^c in next round 132 eor $t0,$t0,$e,ror#$Sigma1[2] // Sigma1(e) 133 eor $T0,$T0,$a,ror#$Sigma0[1] 134 add $h,$h,$t1 // h+=Ch(e,f,g) 135 and $t3,$t3,$t2 // (b^c)&=(a^b) 136 eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1] 137 eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1]) 138 add $h,$h,$t0 // h+=Sigma1(e) 139 eor $t3,$t3,$b // Maj(a,b,c) 140 eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a) 141 eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14]) 142 add @X[$j],@X[$j],@X[($j+9)&15] 143 add $d,$d,$h // d+=h 144 add $h,$h,$t3 // h+=Maj(a,b,c) 145 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 146 add @X[$j],@X[$j],$T1 147 add $h,$h,$t1 // h+=Sigma0(a) 148 add @X[$j],@X[$j],$T2 149___ 150 ($t2,$t3)=($t3,$t2); 151} 152 153$code.=<<___; 154#include "arm_arch.h" 155 156.text 157 158.globl $func 159.type $func,%function 160.align 6 161$func: 162___ 163$code.=<<___ if ($SZ==4); 164 ldr x16,.LOPENSSL_armcap_P 165 adr x17,.LOPENSSL_armcap_P 166 add x16,x16,x17 167 ldr w16,[x16] 168 tst w16,#ARMV8_SHA256 169 b.ne .Lv8_entry 170___ 171$code.=<<___; 172 stp x29,x30,[sp,#-128]! 173 add x29,sp,#0 174 175 stp x19,x20,[sp,#16] 176 stp x21,x22,[sp,#32] 177 stp x23,x24,[sp,#48] 178 stp x25,x26,[sp,#64] 179 stp x27,x28,[sp,#80] 180 sub sp,sp,#4*$SZ 181 182 ldp $A,$B,[$ctx] // load context 183 ldp $C,$D,[$ctx,#2*$SZ] 184 ldp $E,$F,[$ctx,#4*$SZ] 185 add $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input 186 ldp $G,$H,[$ctx,#6*$SZ] 187 adr $Ktbl,K$BITS 188 stp $ctx,$num,[x29,#96] 189 190.Loop: 191 ldp @X[0],@X[1],[$inp],#2*$SZ 192 ldr $t2,[$Ktbl],#$SZ // *K++ 193 eor $t3,$B,$C // magic seed 194 str $inp,[x29,#112] 195___ 196for ($i=0;$i<16;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 197$code.=".Loop_16_xx:\n"; 198for (;$i<32;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 199$code.=<<___; 200 cbnz $t2,.Loop_16_xx 201 202 ldp $ctx,$num,[x29,#96] 203 ldr $inp,[x29,#112] 204 sub $Ktbl,$Ktbl,#`$SZ*($rounds+1)` // rewind 205 206 ldp @X[0],@X[1],[$ctx] 207 ldp @X[2],@X[3],[$ctx,#2*$SZ] 208 add $inp,$inp,#14*$SZ // advance input pointer 209 ldp @X[4],@X[5],[$ctx,#4*$SZ] 210 add $A,$A,@X[0] 211 ldp @X[6],@X[7],[$ctx,#6*$SZ] 212 add $B,$B,@X[1] 213 add $C,$C,@X[2] 214 add $D,$D,@X[3] 215 stp $A,$B,[$ctx] 216 add $E,$E,@X[4] 217 add $F,$F,@X[5] 218 stp $C,$D,[$ctx,#2*$SZ] 219 add $G,$G,@X[6] 220 add $H,$H,@X[7] 221 cmp $inp,$num 222 stp $E,$F,[$ctx,#4*$SZ] 223 stp $G,$H,[$ctx,#6*$SZ] 224 b.ne .Loop 225 226 ldp x19,x20,[x29,#16] 227 add sp,sp,#4*$SZ 228 ldp x21,x22,[x29,#32] 229 ldp x23,x24,[x29,#48] 230 ldp x25,x26,[x29,#64] 231 ldp x27,x28,[x29,#80] 232 ldp x29,x30,[sp],#128 233 ret 234.size $func,.-$func 235 236.align 6 237.type K$BITS,%object 238K$BITS: 239___ 240$code.=<<___ if ($SZ==8); 241 .quad 0x428a2f98d728ae22,0x7137449123ef65cd 242 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc 243 .quad 0x3956c25bf348b538,0x59f111f1b605d019 244 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118 245 .quad 0xd807aa98a3030242,0x12835b0145706fbe 246 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2 247 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1 248 .quad 0x9bdc06a725c71235,0xc19bf174cf692694 249 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3 250 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65 251 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483 252 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5 253 .quad 0x983e5152ee66dfab,0xa831c66d2db43210 254 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4 255 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725 256 .quad 0x06ca6351e003826f,0x142929670a0e6e70 257 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926 258 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df 259 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8 260 .quad 0x81c2c92e47edaee6,0x92722c851482353b 261 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001 262 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30 263 .quad 0xd192e819d6ef5218,0xd69906245565a910 264 .quad 0xf40e35855771202a,0x106aa07032bbd1b8 265 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53 266 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8 267 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb 268 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3 269 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60 270 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec 271 .quad 0x90befffa23631e28,0xa4506cebde82bde9 272 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b 273 .quad 0xca273eceea26619c,0xd186b8c721c0c207 274 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178 275 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6 276 .quad 0x113f9804bef90dae,0x1b710b35131c471b 277 .quad 0x28db77f523047d84,0x32caab7b40c72493 278 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c 279 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a 280 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817 281 .quad 0 // terminator 282___ 283$code.=<<___ if ($SZ==4); 284 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 285 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 286 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 287 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 288 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 289 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 290 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 291 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 292 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 293 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 294 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 295 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 296 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 297 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 298 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 299 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 300 .long 0 //terminator 301___ 302$code.=<<___; 303.size K$BITS,.-K$BITS 304.align 3 305.LOPENSSL_armcap_P: 306 .quad OPENSSL_armcap_P-. 307.asciz "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>" 308.align 2 309___ 310 311if ($SZ==4) { 312my $Ktbl="x3"; 313 314my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2)); 315my @MSG=map("v$_.16b",(4..7)); 316my ($W0,$W1)=("v16.4s","v17.4s"); 317my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b"); 318 319$code.=<<___; 320.type sha256_block_armv8,%function 321.align 6 322sha256_block_armv8: 323.Lv8_entry: 324 stp x29,x30,[sp,#-16]! 325 add x29,sp,#0 326 327 ld1.32 {$ABCD,$EFGH},[$ctx] 328 adr $Ktbl,K256 329 330.Loop_hw: 331 ld1 {@MSG[0]-@MSG[3]},[$inp],#64 332 sub $num,$num,#1 333 ld1.32 {$W0},[$Ktbl],#16 334 rev32 @MSG[0],@MSG[0] 335 rev32 @MSG[1],@MSG[1] 336 rev32 @MSG[2],@MSG[2] 337 rev32 @MSG[3],@MSG[3] 338 orr $ABCD_SAVE,$ABCD,$ABCD // offload 339 orr $EFGH_SAVE,$EFGH,$EFGH 340___ 341for($i=0;$i<12;$i++) { 342$code.=<<___; 343 ld1.32 {$W1},[$Ktbl],#16 344 add.i32 $W0,$W0,@MSG[0] 345 sha256su0 @MSG[0],@MSG[1] 346 orr $abcd,$ABCD,$ABCD 347 sha256h $ABCD,$EFGH,$W0 348 sha256h2 $EFGH,$abcd,$W0 349 sha256su1 @MSG[0],@MSG[2],@MSG[3] 350___ 351 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG)); 352} 353$code.=<<___; 354 ld1.32 {$W1},[$Ktbl],#16 355 add.i32 $W0,$W0,@MSG[0] 356 orr $abcd,$ABCD,$ABCD 357 sha256h $ABCD,$EFGH,$W0 358 sha256h2 $EFGH,$abcd,$W0 359 360 ld1.32 {$W0},[$Ktbl],#16 361 add.i32 $W1,$W1,@MSG[1] 362 orr $abcd,$ABCD,$ABCD 363 sha256h $ABCD,$EFGH,$W1 364 sha256h2 $EFGH,$abcd,$W1 365 366 ld1.32 {$W1},[$Ktbl] 367 add.i32 $W0,$W0,@MSG[2] 368 sub $Ktbl,$Ktbl,#$rounds*$SZ-16 // rewind 369 orr $abcd,$ABCD,$ABCD 370 sha256h $ABCD,$EFGH,$W0 371 sha256h2 $EFGH,$abcd,$W0 372 373 add.i32 $W1,$W1,@MSG[3] 374 orr $abcd,$ABCD,$ABCD 375 sha256h $ABCD,$EFGH,$W1 376 sha256h2 $EFGH,$abcd,$W1 377 378 add.i32 $ABCD,$ABCD,$ABCD_SAVE 379 add.i32 $EFGH,$EFGH,$EFGH_SAVE 380 381 cbnz $num,.Loop_hw 382 383 st1.32 {$ABCD,$EFGH},[$ctx] 384 385 ldr x29,[sp],#16 386 ret 387.size sha256_block_armv8,.-sha256_block_armv8 388___ 389} 390 391$code.=<<___; 392.comm OPENSSL_armcap_P,4,4 393___ 394 395{ my %opcode = ( 396 "sha256h" => 0x5e004000, "sha256h2" => 0x5e005000, 397 "sha256su0" => 0x5e282800, "sha256su1" => 0x5e006000 ); 398 399 sub unsha256 { 400 my ($mnemonic,$arg)=@_; 401 402 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o 403 && 404 sprintf ".inst\t0x%08x\t//%s %s", 405 $opcode{$mnemonic}|$1|($2<<5)|($3<<16), 406 $mnemonic,$arg; 407 } 408} 409 410foreach(split("\n",$code)) { 411 412 s/\`([^\`]*)\`/eval($1)/geo; 413 414 s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo; 415 416 s/\.\w?32\b//o and s/\.16b/\.4s/go; 417 m/(ld|st)1[^\[]+\[0\]/o and s/\.4s/\.s/go; 418 419 print $_,"\n"; 420} 421 422close STDOUT; 423