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; 34 35$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 36( $xlate="${dir}arm-xlate.pl" and -f $xlate ) or 37( $xlate="${dir}../../perlasm/arm-xlate.pl" and -f $xlate) or 38die "can't locate arm-xlate.pl"; 39 40open OUT,"| \"$^X\" $xlate $flavour $output"; 41*STDOUT=*OUT; 42 43if ($output =~ /512/) { 44 $BITS=512; 45 $SZ=8; 46 @Sigma0=(28,34,39); 47 @Sigma1=(14,18,41); 48 @sigma0=(1, 8, 7); 49 @sigma1=(19,61, 6); 50 $rounds=80; 51 $reg_t="x"; 52} else { 53 $BITS=256; 54 $SZ=4; 55 @Sigma0=( 2,13,22); 56 @Sigma1=( 6,11,25); 57 @sigma0=( 7,18, 3); 58 @sigma1=(17,19,10); 59 $rounds=64; 60 $reg_t="w"; 61} 62 63$func="sha${BITS}_block_data_order"; 64 65($ctx,$inp,$num,$Ktbl)=map("x$_",(0..2,30)); 66 67@X=map("$reg_t$_",(3..15,0..2)); 68@V=($A,$B,$C,$D,$E,$F,$G,$H)=map("$reg_t$_",(20..27)); 69($t0,$t1,$t2,$t3)=map("$reg_t$_",(16,17,19,28)); 70 71sub BODY_00_xx { 72my ($i,$a,$b,$c,$d,$e,$f,$g,$h)=@_; 73my $j=($i+1)&15; 74my ($T0,$T1,$T2)=(@X[($i-8)&15],@X[($i-9)&15],@X[($i-10)&15]); 75 $T0=@X[$i+3] if ($i<11); 76 77$code.=<<___ if ($i<16); 78#ifndef __ARMEB__ 79 rev @X[$i],@X[$i] // $i 80#endif 81___ 82$code.=<<___ if ($i<13 && ($i&1)); 83 ldp @X[$i+1],@X[$i+2],[$inp],#2*$SZ 84___ 85$code.=<<___ if ($i==13); 86 ldp @X[14],@X[15],[$inp] 87___ 88$code.=<<___ if ($i>=14); 89 ldr @X[($i-11)&15],[sp,#`$SZ*(($i-11)%4)`] 90___ 91$code.=<<___ if ($i>0 && $i<16); 92 add $a,$a,$t1 // h+=Sigma0(a) 93___ 94$code.=<<___ if ($i>=11); 95 str @X[($i-8)&15],[sp,#`$SZ*(($i-8)%4)`] 96___ 97# While ARMv8 specifies merged rotate-n-logical operation such as 98# 'eor x,y,z,ror#n', it was found to negatively affect performance 99# on Apple A7. The reason seems to be that it requires even 'y' to 100# be available earlier. This means that such merged instruction is 101# not necessarily best choice on critical path... On the other hand 102# Cortex-A5x handles merged instructions much better than disjoint 103# rotate and logical... See (**) footnote above. 104$code.=<<___ if ($i<15); 105 ror $t0,$e,#$Sigma1[0] 106 add $h,$h,$t2 // h+=K[i] 107 eor $T0,$e,$e,ror#`$Sigma1[2]-$Sigma1[1]` 108 and $t1,$f,$e 109 bic $t2,$g,$e 110 add $h,$h,@X[$i&15] // h+=X[i] 111 orr $t1,$t1,$t2 // Ch(e,f,g) 112 eor $t2,$a,$b // a^b, b^c in next round 113 eor $t0,$t0,$T0,ror#$Sigma1[1] // Sigma1(e) 114 ror $T0,$a,#$Sigma0[0] 115 add $h,$h,$t1 // h+=Ch(e,f,g) 116 eor $t1,$a,$a,ror#`$Sigma0[2]-$Sigma0[1]` 117 add $h,$h,$t0 // h+=Sigma1(e) 118 and $t3,$t3,$t2 // (b^c)&=(a^b) 119 add $d,$d,$h // d+=h 120 eor $t3,$t3,$b // Maj(a,b,c) 121 eor $t1,$T0,$t1,ror#$Sigma0[1] // Sigma0(a) 122 add $h,$h,$t3 // h+=Maj(a,b,c) 123 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 124 //add $h,$h,$t1 // h+=Sigma0(a) 125___ 126$code.=<<___ if ($i>=15); 127 ror $t0,$e,#$Sigma1[0] 128 add $h,$h,$t2 // h+=K[i] 129 ror $T1,@X[($j+1)&15],#$sigma0[0] 130 and $t1,$f,$e 131 ror $T2,@X[($j+14)&15],#$sigma1[0] 132 bic $t2,$g,$e 133 ror $T0,$a,#$Sigma0[0] 134 add $h,$h,@X[$i&15] // h+=X[i] 135 eor $t0,$t0,$e,ror#$Sigma1[1] 136 eor $T1,$T1,@X[($j+1)&15],ror#$sigma0[1] 137 orr $t1,$t1,$t2 // Ch(e,f,g) 138 eor $t2,$a,$b // a^b, b^c in next round 139 eor $t0,$t0,$e,ror#$Sigma1[2] // Sigma1(e) 140 eor $T0,$T0,$a,ror#$Sigma0[1] 141 add $h,$h,$t1 // h+=Ch(e,f,g) 142 and $t3,$t3,$t2 // (b^c)&=(a^b) 143 eor $T2,$T2,@X[($j+14)&15],ror#$sigma1[1] 144 eor $T1,$T1,@X[($j+1)&15],lsr#$sigma0[2] // sigma0(X[i+1]) 145 add $h,$h,$t0 // h+=Sigma1(e) 146 eor $t3,$t3,$b // Maj(a,b,c) 147 eor $t1,$T0,$a,ror#$Sigma0[2] // Sigma0(a) 148 eor $T2,$T2,@X[($j+14)&15],lsr#$sigma1[2] // sigma1(X[i+14]) 149 add @X[$j],@X[$j],@X[($j+9)&15] 150 add $d,$d,$h // d+=h 151 add $h,$h,$t3 // h+=Maj(a,b,c) 152 ldr $t3,[$Ktbl],#$SZ // *K++, $t2 in next round 153 add @X[$j],@X[$j],$T1 154 add $h,$h,$t1 // h+=Sigma0(a) 155 add @X[$j],@X[$j],$T2 156___ 157 ($t2,$t3)=($t3,$t2); 158} 159 160$code.=<<___; 161#include <openssl/arm_arch.h> 162 163.text 164 165.extern OPENSSL_armcap_P 166.globl $func 167.type $func,%function 168.align 6 169$func: 170___ 171$code.=<<___ if ($SZ==4); 172 ldr x16,.LOPENSSL_armcap_P 173 adr x17,.LOPENSSL_armcap_P 174 add x16,x16,x17 175 ldr w16,[x16] 176 tst w16,#ARMV8_SHA256 177 b.ne .Lv8_entry 178___ 179$code.=<<___; 180 stp x29,x30,[sp,#-128]! 181 add x29,sp,#0 182 183 stp x19,x20,[sp,#16] 184 stp x21,x22,[sp,#32] 185 stp x23,x24,[sp,#48] 186 stp x25,x26,[sp,#64] 187 stp x27,x28,[sp,#80] 188 sub sp,sp,#4*$SZ 189 190 ldp $A,$B,[$ctx] // load context 191 ldp $C,$D,[$ctx,#2*$SZ] 192 ldp $E,$F,[$ctx,#4*$SZ] 193 add $num,$inp,$num,lsl#`log(16*$SZ)/log(2)` // end of input 194 ldp $G,$H,[$ctx,#6*$SZ] 195 adr $Ktbl,.LK$BITS 196 stp $ctx,$num,[x29,#96] 197 198.Loop: 199 ldp @X[0],@X[1],[$inp],#2*$SZ 200 ldr $t2,[$Ktbl],#$SZ // *K++ 201 eor $t3,$B,$C // magic seed 202 str $inp,[x29,#112] 203___ 204for ($i=0;$i<16;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 205$code.=".Loop_16_xx:\n"; 206for (;$i<32;$i++) { &BODY_00_xx($i,@V); unshift(@V,pop(@V)); } 207$code.=<<___; 208 cbnz $t2,.Loop_16_xx 209 210 ldp $ctx,$num,[x29,#96] 211 ldr $inp,[x29,#112] 212 sub $Ktbl,$Ktbl,#`$SZ*($rounds+1)` // rewind 213 214 ldp @X[0],@X[1],[$ctx] 215 ldp @X[2],@X[3],[$ctx,#2*$SZ] 216 add $inp,$inp,#14*$SZ // advance input pointer 217 ldp @X[4],@X[5],[$ctx,#4*$SZ] 218 add $A,$A,@X[0] 219 ldp @X[6],@X[7],[$ctx,#6*$SZ] 220 add $B,$B,@X[1] 221 add $C,$C,@X[2] 222 add $D,$D,@X[3] 223 stp $A,$B,[$ctx] 224 add $E,$E,@X[4] 225 add $F,$F,@X[5] 226 stp $C,$D,[$ctx,#2*$SZ] 227 add $G,$G,@X[6] 228 add $H,$H,@X[7] 229 cmp $inp,$num 230 stp $E,$F,[$ctx,#4*$SZ] 231 stp $G,$H,[$ctx,#6*$SZ] 232 b.ne .Loop 233 234 ldp x19,x20,[x29,#16] 235 add sp,sp,#4*$SZ 236 ldp x21,x22,[x29,#32] 237 ldp x23,x24,[x29,#48] 238 ldp x25,x26,[x29,#64] 239 ldp x27,x28,[x29,#80] 240 ldp x29,x30,[sp],#128 241 ret 242.size $func,.-$func 243 244.align 6 245.type .LK$BITS,%object 246.LK$BITS: 247___ 248$code.=<<___ if ($SZ==8); 249 .quad 0x428a2f98d728ae22,0x7137449123ef65cd 250 .quad 0xb5c0fbcfec4d3b2f,0xe9b5dba58189dbbc 251 .quad 0x3956c25bf348b538,0x59f111f1b605d019 252 .quad 0x923f82a4af194f9b,0xab1c5ed5da6d8118 253 .quad 0xd807aa98a3030242,0x12835b0145706fbe 254 .quad 0x243185be4ee4b28c,0x550c7dc3d5ffb4e2 255 .quad 0x72be5d74f27b896f,0x80deb1fe3b1696b1 256 .quad 0x9bdc06a725c71235,0xc19bf174cf692694 257 .quad 0xe49b69c19ef14ad2,0xefbe4786384f25e3 258 .quad 0x0fc19dc68b8cd5b5,0x240ca1cc77ac9c65 259 .quad 0x2de92c6f592b0275,0x4a7484aa6ea6e483 260 .quad 0x5cb0a9dcbd41fbd4,0x76f988da831153b5 261 .quad 0x983e5152ee66dfab,0xa831c66d2db43210 262 .quad 0xb00327c898fb213f,0xbf597fc7beef0ee4 263 .quad 0xc6e00bf33da88fc2,0xd5a79147930aa725 264 .quad 0x06ca6351e003826f,0x142929670a0e6e70 265 .quad 0x27b70a8546d22ffc,0x2e1b21385c26c926 266 .quad 0x4d2c6dfc5ac42aed,0x53380d139d95b3df 267 .quad 0x650a73548baf63de,0x766a0abb3c77b2a8 268 .quad 0x81c2c92e47edaee6,0x92722c851482353b 269 .quad 0xa2bfe8a14cf10364,0xa81a664bbc423001 270 .quad 0xc24b8b70d0f89791,0xc76c51a30654be30 271 .quad 0xd192e819d6ef5218,0xd69906245565a910 272 .quad 0xf40e35855771202a,0x106aa07032bbd1b8 273 .quad 0x19a4c116b8d2d0c8,0x1e376c085141ab53 274 .quad 0x2748774cdf8eeb99,0x34b0bcb5e19b48a8 275 .quad 0x391c0cb3c5c95a63,0x4ed8aa4ae3418acb 276 .quad 0x5b9cca4f7763e373,0x682e6ff3d6b2b8a3 277 .quad 0x748f82ee5defb2fc,0x78a5636f43172f60 278 .quad 0x84c87814a1f0ab72,0x8cc702081a6439ec 279 .quad 0x90befffa23631e28,0xa4506cebde82bde9 280 .quad 0xbef9a3f7b2c67915,0xc67178f2e372532b 281 .quad 0xca273eceea26619c,0xd186b8c721c0c207 282 .quad 0xeada7dd6cde0eb1e,0xf57d4f7fee6ed178 283 .quad 0x06f067aa72176fba,0x0a637dc5a2c898a6 284 .quad 0x113f9804bef90dae,0x1b710b35131c471b 285 .quad 0x28db77f523047d84,0x32caab7b40c72493 286 .quad 0x3c9ebe0a15c9bebc,0x431d67c49c100d4c 287 .quad 0x4cc5d4becb3e42b6,0x597f299cfc657e2a 288 .quad 0x5fcb6fab3ad6faec,0x6c44198c4a475817 289 .quad 0 // terminator 290___ 291$code.=<<___ if ($SZ==4); 292 .long 0x428a2f98,0x71374491,0xb5c0fbcf,0xe9b5dba5 293 .long 0x3956c25b,0x59f111f1,0x923f82a4,0xab1c5ed5 294 .long 0xd807aa98,0x12835b01,0x243185be,0x550c7dc3 295 .long 0x72be5d74,0x80deb1fe,0x9bdc06a7,0xc19bf174 296 .long 0xe49b69c1,0xefbe4786,0x0fc19dc6,0x240ca1cc 297 .long 0x2de92c6f,0x4a7484aa,0x5cb0a9dc,0x76f988da 298 .long 0x983e5152,0xa831c66d,0xb00327c8,0xbf597fc7 299 .long 0xc6e00bf3,0xd5a79147,0x06ca6351,0x14292967 300 .long 0x27b70a85,0x2e1b2138,0x4d2c6dfc,0x53380d13 301 .long 0x650a7354,0x766a0abb,0x81c2c92e,0x92722c85 302 .long 0xa2bfe8a1,0xa81a664b,0xc24b8b70,0xc76c51a3 303 .long 0xd192e819,0xd6990624,0xf40e3585,0x106aa070 304 .long 0x19a4c116,0x1e376c08,0x2748774c,0x34b0bcb5 305 .long 0x391c0cb3,0x4ed8aa4a,0x5b9cca4f,0x682e6ff3 306 .long 0x748f82ee,0x78a5636f,0x84c87814,0x8cc70208 307 .long 0x90befffa,0xa4506ceb,0xbef9a3f7,0xc67178f2 308 .long 0 //terminator 309___ 310$code.=<<___; 311.size .LK$BITS,.-.LK$BITS 312.align 3 313.LOPENSSL_armcap_P: 314 .quad OPENSSL_armcap_P-. 315.asciz "SHA$BITS block transform for ARMv8, CRYPTOGAMS by <appro\@openssl.org>" 316.align 2 317___ 318 319if ($SZ==4) { 320my $Ktbl="x3"; 321 322my ($ABCD,$EFGH,$abcd)=map("v$_.16b",(0..2)); 323my @MSG=map("v$_.16b",(4..7)); 324my ($W0,$W1)=("v16.4s","v17.4s"); 325my ($ABCD_SAVE,$EFGH_SAVE)=("v18.16b","v19.16b"); 326 327$code.=<<___; 328.type sha256_block_armv8,%function 329.align 6 330sha256_block_armv8: 331.Lv8_entry: 332 stp x29,x30,[sp,#-16]! 333 add x29,sp,#0 334 335 ld1.32 {$ABCD,$EFGH},[$ctx] 336 adr $Ktbl,.LK256 337 338.Loop_hw: 339 ld1 {@MSG[0]-@MSG[3]},[$inp],#64 340 sub $num,$num,#1 341 ld1.32 {$W0},[$Ktbl],#16 342 rev32 @MSG[0],@MSG[0] 343 rev32 @MSG[1],@MSG[1] 344 rev32 @MSG[2],@MSG[2] 345 rev32 @MSG[3],@MSG[3] 346 orr $ABCD_SAVE,$ABCD,$ABCD // offload 347 orr $EFGH_SAVE,$EFGH,$EFGH 348___ 349for($i=0;$i<12;$i++) { 350$code.=<<___; 351 ld1.32 {$W1},[$Ktbl],#16 352 add.i32 $W0,$W0,@MSG[0] 353 sha256su0 @MSG[0],@MSG[1] 354 orr $abcd,$ABCD,$ABCD 355 sha256h $ABCD,$EFGH,$W0 356 sha256h2 $EFGH,$abcd,$W0 357 sha256su1 @MSG[0],@MSG[2],@MSG[3] 358___ 359 ($W0,$W1)=($W1,$W0); push(@MSG,shift(@MSG)); 360} 361$code.=<<___; 362 ld1.32 {$W1},[$Ktbl],#16 363 add.i32 $W0,$W0,@MSG[0] 364 orr $abcd,$ABCD,$ABCD 365 sha256h $ABCD,$EFGH,$W0 366 sha256h2 $EFGH,$abcd,$W0 367 368 ld1.32 {$W0},[$Ktbl],#16 369 add.i32 $W1,$W1,@MSG[1] 370 orr $abcd,$ABCD,$ABCD 371 sha256h $ABCD,$EFGH,$W1 372 sha256h2 $EFGH,$abcd,$W1 373 374 ld1.32 {$W1},[$Ktbl] 375 add.i32 $W0,$W0,@MSG[2] 376 sub $Ktbl,$Ktbl,#$rounds*$SZ-16 // rewind 377 orr $abcd,$ABCD,$ABCD 378 sha256h $ABCD,$EFGH,$W0 379 sha256h2 $EFGH,$abcd,$W0 380 381 add.i32 $W1,$W1,@MSG[3] 382 orr $abcd,$ABCD,$ABCD 383 sha256h $ABCD,$EFGH,$W1 384 sha256h2 $EFGH,$abcd,$W1 385 386 add.i32 $ABCD,$ABCD,$ABCD_SAVE 387 add.i32 $EFGH,$EFGH,$EFGH_SAVE 388 389 cbnz $num,.Loop_hw 390 391 st1.32 {$ABCD,$EFGH},[$ctx] 392 393 ldr x29,[sp],#16 394 ret 395.size sha256_block_armv8,.-sha256_block_armv8 396___ 397} 398 399$code.=<<___; 400.comm OPENSSL_armcap_P,4,4 401___ 402 403{ my %opcode = ( 404 "sha256h" => 0x5e004000, "sha256h2" => 0x5e005000, 405 "sha256su0" => 0x5e282800, "sha256su1" => 0x5e006000 ); 406 407 sub unsha256 { 408 my ($mnemonic,$arg)=@_; 409 410 $arg =~ m/[qv]([0-9]+)[^,]*,\s*[qv]([0-9]+)[^,]*(?:,\s*[qv]([0-9]+))?/o 411 && 412 sprintf ".inst\t0x%08x\t//%s %s", 413 $opcode{$mnemonic}|$1|($2<<5)|($3<<16), 414 $mnemonic,$arg; 415 } 416} 417 418foreach(split("\n",$code)) { 419 420 s/\`([^\`]*)\`/eval($1)/geo; 421 422 s/\b(sha256\w+)\s+([qv].*)/unsha256($1,$2)/geo; 423 424 s/\.\w?32\b//o and s/\.16b/\.4s/go; 425 m/(ld|st)1[^\[]+\[0\]/o and s/\.4s/\.s/go; 426 427 print $_,"\n"; 428} 429 430close STDOUT; 431