1238384Sjkim#!/usr/bin/env perl 2238384Sjkim# 3238384Sjkim# ==================================================================== 4238384Sjkim# Written by Andy Polyakov <appro@openssl.org> for the OpenSSL 5238384Sjkim# project. The module is, however, dual licensed under OpenSSL and 6238384Sjkim# CRYPTOGAMS licenses depending on where you obtain it. For further 7238384Sjkim# details see http://www.openssl.org/~appro/cryptogams/. 8238384Sjkim# ==================================================================== 9238384Sjkim# 10238384Sjkim# May 2011 11238384Sjkim# 12238384Sjkim# The module implements bn_GF2m_mul_2x2 polynomial multiplication used 13238384Sjkim# in bn_gf2m.c. It's kind of low-hanging mechanical port from C for 14238384Sjkim# the time being... Except that it has two code paths: code suitable 15238384Sjkim# for any x86_64 CPU and PCLMULQDQ one suitable for Westmere and 16291721Sjkim# later. Improvement varies from one benchmark and ��-arch to another. 17238384Sjkim# Vanilla code path is at most 20% faster than compiler-generated code 18238384Sjkim# [not very impressive], while PCLMULQDQ - whole 85%-160% better on 19238384Sjkim# 163- and 571-bit ECDH benchmarks on Intel CPUs. Keep in mind that 20238384Sjkim# these coefficients are not ones for bn_GF2m_mul_2x2 itself, as not 21238384Sjkim# all CPU time is burnt in it... 22238384Sjkim 23238384Sjkim$flavour = shift; 24238384Sjkim$output = shift; 25238384Sjkimif ($flavour =~ /\./) { $output = $flavour; undef $flavour; } 26238384Sjkim 27238384Sjkim$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); 28238384Sjkim 29238384Sjkim$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; 30238384Sjkim( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or 31238384Sjkim( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or 32238384Sjkimdie "can't locate x86_64-xlate.pl"; 33238384Sjkim 34264331Sjkimopen OUT,"| \"$^X\" $xlate $flavour $output"; 35264331Sjkim*STDOUT=*OUT; 36238384Sjkim 37238384Sjkim($lo,$hi)=("%rax","%rdx"); $a=$lo; 38238384Sjkim($i0,$i1)=("%rsi","%rdi"); 39238384Sjkim($t0,$t1)=("%rbx","%rcx"); 40238384Sjkim($b,$mask)=("%rbp","%r8"); 41238384Sjkim($a1,$a2,$a4,$a8,$a12,$a48)=map("%r$_",(9..15)); 42238384Sjkim($R,$Tx)=("%xmm0","%xmm1"); 43238384Sjkim 44238384Sjkim$code.=<<___; 45238384Sjkim.text 46238384Sjkim 47238384Sjkim.type _mul_1x1,\@abi-omnipotent 48238384Sjkim.align 16 49238384Sjkim_mul_1x1: 50238384Sjkim sub \$128+8,%rsp 51238384Sjkim mov \$-1,$a1 52238384Sjkim lea ($a,$a),$i0 53238384Sjkim shr \$3,$a1 54238384Sjkim lea (,$a,4),$i1 55238384Sjkim and $a,$a1 # a1=a&0x1fffffffffffffff 56238384Sjkim lea (,$a,8),$a8 57238384Sjkim sar \$63,$a # broadcast 63rd bit 58238384Sjkim lea ($a1,$a1),$a2 59238384Sjkim sar \$63,$i0 # broadcast 62nd bit 60238384Sjkim lea (,$a1,4),$a4 61238384Sjkim and $b,$a 62238384Sjkim sar \$63,$i1 # boardcast 61st bit 63238384Sjkim mov $a,$hi # $a is $lo 64238384Sjkim shl \$63,$lo 65238384Sjkim and $b,$i0 66238384Sjkim shr \$1,$hi 67238384Sjkim mov $i0,$t1 68238384Sjkim shl \$62,$i0 69238384Sjkim and $b,$i1 70238384Sjkim shr \$2,$t1 71238384Sjkim xor $i0,$lo 72238384Sjkim mov $i1,$t0 73238384Sjkim shl \$61,$i1 74238384Sjkim xor $t1,$hi 75238384Sjkim shr \$3,$t0 76238384Sjkim xor $i1,$lo 77238384Sjkim xor $t0,$hi 78238384Sjkim 79238384Sjkim mov $a1,$a12 80238384Sjkim movq \$0,0(%rsp) # tab[0]=0 81238384Sjkim xor $a2,$a12 # a1^a2 82238384Sjkim mov $a1,8(%rsp) # tab[1]=a1 83238384Sjkim mov $a4,$a48 84238384Sjkim mov $a2,16(%rsp) # tab[2]=a2 85238384Sjkim xor $a8,$a48 # a4^a8 86238384Sjkim mov $a12,24(%rsp) # tab[3]=a1^a2 87238384Sjkim 88238384Sjkim xor $a4,$a1 89238384Sjkim mov $a4,32(%rsp) # tab[4]=a4 90238384Sjkim xor $a4,$a2 91238384Sjkim mov $a1,40(%rsp) # tab[5]=a1^a4 92238384Sjkim xor $a4,$a12 93238384Sjkim mov $a2,48(%rsp) # tab[6]=a2^a4 94238384Sjkim xor $a48,$a1 # a1^a4^a4^a8=a1^a8 95238384Sjkim mov $a12,56(%rsp) # tab[7]=a1^a2^a4 96238384Sjkim xor $a48,$a2 # a2^a4^a4^a8=a1^a8 97238384Sjkim 98238384Sjkim mov $a8,64(%rsp) # tab[8]=a8 99238384Sjkim xor $a48,$a12 # a1^a2^a4^a4^a8=a1^a2^a8 100238384Sjkim mov $a1,72(%rsp) # tab[9]=a1^a8 101238384Sjkim xor $a4,$a1 # a1^a8^a4 102238384Sjkim mov $a2,80(%rsp) # tab[10]=a2^a8 103238384Sjkim xor $a4,$a2 # a2^a8^a4 104238384Sjkim mov $a12,88(%rsp) # tab[11]=a1^a2^a8 105238384Sjkim 106238384Sjkim xor $a4,$a12 # a1^a2^a8^a4 107238384Sjkim mov $a48,96(%rsp) # tab[12]=a4^a8 108238384Sjkim mov $mask,$i0 109238384Sjkim mov $a1,104(%rsp) # tab[13]=a1^a4^a8 110238384Sjkim and $b,$i0 111238384Sjkim mov $a2,112(%rsp) # tab[14]=a2^a4^a8 112238384Sjkim shr \$4,$b 113238384Sjkim mov $a12,120(%rsp) # tab[15]=a1^a2^a4^a8 114238384Sjkim mov $mask,$i1 115238384Sjkim and $b,$i1 116238384Sjkim shr \$4,$b 117238384Sjkim 118238384Sjkim movq (%rsp,$i0,8),$R # half of calculations is done in SSE2 119238384Sjkim mov $mask,$i0 120238384Sjkim and $b,$i0 121238384Sjkim shr \$4,$b 122238384Sjkim___ 123238384Sjkim for ($n=1;$n<8;$n++) { 124238384Sjkim $code.=<<___; 125238384Sjkim mov (%rsp,$i1,8),$t1 126238384Sjkim mov $mask,$i1 127238384Sjkim mov $t1,$t0 128238384Sjkim shl \$`8*$n-4`,$t1 129238384Sjkim and $b,$i1 130238384Sjkim movq (%rsp,$i0,8),$Tx 131238384Sjkim shr \$`64-(8*$n-4)`,$t0 132238384Sjkim xor $t1,$lo 133238384Sjkim pslldq \$$n,$Tx 134238384Sjkim mov $mask,$i0 135238384Sjkim shr \$4,$b 136238384Sjkim xor $t0,$hi 137238384Sjkim and $b,$i0 138238384Sjkim shr \$4,$b 139238384Sjkim pxor $Tx,$R 140238384Sjkim___ 141238384Sjkim } 142238384Sjkim$code.=<<___; 143238384Sjkim mov (%rsp,$i1,8),$t1 144238384Sjkim mov $t1,$t0 145238384Sjkim shl \$`8*$n-4`,$t1 146238384Sjkim movq $R,$i0 147238384Sjkim shr \$`64-(8*$n-4)`,$t0 148238384Sjkim xor $t1,$lo 149238384Sjkim psrldq \$8,$R 150238384Sjkim xor $t0,$hi 151238384Sjkim movq $R,$i1 152238384Sjkim xor $i0,$lo 153238384Sjkim xor $i1,$hi 154238384Sjkim 155238384Sjkim add \$128+8,%rsp 156238384Sjkim ret 157238384Sjkim.Lend_mul_1x1: 158238384Sjkim.size _mul_1x1,.-_mul_1x1 159238384Sjkim___ 160238384Sjkim 161238384Sjkim($rp,$a1,$a0,$b1,$b0) = $win64? ("%rcx","%rdx","%r8", "%r9","%r10") : # Win64 order 162238384Sjkim ("%rdi","%rsi","%rdx","%rcx","%r8"); # Unix order 163238384Sjkim 164238384Sjkim$code.=<<___; 165238384Sjkim.extern OPENSSL_ia32cap_P 166238384Sjkim.globl bn_GF2m_mul_2x2 167238384Sjkim.type bn_GF2m_mul_2x2,\@abi-omnipotent 168238384Sjkim.align 16 169238384Sjkimbn_GF2m_mul_2x2: 170238384Sjkim mov OPENSSL_ia32cap_P(%rip),%rax 171238384Sjkim bt \$33,%rax 172238384Sjkim jnc .Lvanilla_mul_2x2 173238384Sjkim 174238384Sjkim movq $a1,%xmm0 175238384Sjkim movq $b1,%xmm1 176238384Sjkim movq $a0,%xmm2 177238384Sjkim___ 178238384Sjkim$code.=<<___ if ($win64); 179238384Sjkim movq 40(%rsp),%xmm3 180238384Sjkim___ 181238384Sjkim$code.=<<___ if (!$win64); 182238384Sjkim movq $b0,%xmm3 183238384Sjkim___ 184238384Sjkim$code.=<<___; 185238384Sjkim movdqa %xmm0,%xmm4 186238384Sjkim movdqa %xmm1,%xmm5 187291721Sjkim pclmulqdq \$0,%xmm1,%xmm0 # a1��b1 188238384Sjkim pxor %xmm2,%xmm4 189238384Sjkim pxor %xmm3,%xmm5 190291721Sjkim pclmulqdq \$0,%xmm3,%xmm2 # a0��b0 191291721Sjkim pclmulqdq \$0,%xmm5,%xmm4 # (a0+a1)��(b0+b1) 192238384Sjkim xorps %xmm0,%xmm4 193291721Sjkim xorps %xmm2,%xmm4 # (a0+a1)��(b0+b1)-a0��b0-a1��b1 194238384Sjkim movdqa %xmm4,%xmm5 195238384Sjkim pslldq \$8,%xmm4 196238384Sjkim psrldq \$8,%xmm5 197238384Sjkim pxor %xmm4,%xmm2 198238384Sjkim pxor %xmm5,%xmm0 199238384Sjkim movdqu %xmm2,0($rp) 200238384Sjkim movdqu %xmm0,16($rp) 201238384Sjkim ret 202238384Sjkim 203238384Sjkim.align 16 204238384Sjkim.Lvanilla_mul_2x2: 205238384Sjkim lea -8*17(%rsp),%rsp 206238384Sjkim___ 207238384Sjkim$code.=<<___ if ($win64); 208238384Sjkim mov `8*17+40`(%rsp),$b0 209238384Sjkim mov %rdi,8*15(%rsp) 210238384Sjkim mov %rsi,8*16(%rsp) 211238384Sjkim___ 212238384Sjkim$code.=<<___; 213238384Sjkim mov %r14,8*10(%rsp) 214238384Sjkim mov %r13,8*11(%rsp) 215238384Sjkim mov %r12,8*12(%rsp) 216238384Sjkim mov %rbp,8*13(%rsp) 217238384Sjkim mov %rbx,8*14(%rsp) 218238384Sjkim.Lbody_mul_2x2: 219238384Sjkim mov $rp,32(%rsp) # save the arguments 220238384Sjkim mov $a1,40(%rsp) 221238384Sjkim mov $a0,48(%rsp) 222238384Sjkim mov $b1,56(%rsp) 223238384Sjkim mov $b0,64(%rsp) 224238384Sjkim 225238384Sjkim mov \$0xf,$mask 226238384Sjkim mov $a1,$a 227238384Sjkim mov $b1,$b 228291721Sjkim call _mul_1x1 # a1��b1 229238384Sjkim mov $lo,16(%rsp) 230238384Sjkim mov $hi,24(%rsp) 231238384Sjkim 232238384Sjkim mov 48(%rsp),$a 233238384Sjkim mov 64(%rsp),$b 234291721Sjkim call _mul_1x1 # a0��b0 235238384Sjkim mov $lo,0(%rsp) 236238384Sjkim mov $hi,8(%rsp) 237238384Sjkim 238238384Sjkim mov 40(%rsp),$a 239238384Sjkim mov 56(%rsp),$b 240238384Sjkim xor 48(%rsp),$a 241238384Sjkim xor 64(%rsp),$b 242291721Sjkim call _mul_1x1 # (a0+a1)��(b0+b1) 243238384Sjkim___ 244238384Sjkim @r=("%rbx","%rcx","%rdi","%rsi"); 245238384Sjkim$code.=<<___; 246238384Sjkim mov 0(%rsp),@r[0] 247238384Sjkim mov 8(%rsp),@r[1] 248238384Sjkim mov 16(%rsp),@r[2] 249238384Sjkim mov 24(%rsp),@r[3] 250238384Sjkim mov 32(%rsp),%rbp 251238384Sjkim 252238384Sjkim xor $hi,$lo 253238384Sjkim xor @r[1],$hi 254238384Sjkim xor @r[0],$lo 255238384Sjkim mov @r[0],0(%rbp) 256238384Sjkim xor @r[2],$hi 257238384Sjkim mov @r[3],24(%rbp) 258238384Sjkim xor @r[3],$lo 259238384Sjkim xor @r[3],$hi 260238384Sjkim xor $hi,$lo 261238384Sjkim mov $hi,16(%rbp) 262238384Sjkim mov $lo,8(%rbp) 263238384Sjkim 264238384Sjkim mov 8*10(%rsp),%r14 265238384Sjkim mov 8*11(%rsp),%r13 266238384Sjkim mov 8*12(%rsp),%r12 267238384Sjkim mov 8*13(%rsp),%rbp 268238384Sjkim mov 8*14(%rsp),%rbx 269238384Sjkim___ 270238384Sjkim$code.=<<___ if ($win64); 271238384Sjkim mov 8*15(%rsp),%rdi 272238384Sjkim mov 8*16(%rsp),%rsi 273238384Sjkim___ 274238384Sjkim$code.=<<___; 275238384Sjkim lea 8*17(%rsp),%rsp 276238384Sjkim ret 277238384Sjkim.Lend_mul_2x2: 278238384Sjkim.size bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2 279238384Sjkim.asciz "GF(2^m) Multiplication for x86_64, CRYPTOGAMS by <appro\@openssl.org>" 280238384Sjkim.align 16 281238384Sjkim___ 282238384Sjkim 283238384Sjkim# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, 284238384Sjkim# CONTEXT *context,DISPATCHER_CONTEXT *disp) 285238384Sjkimif ($win64) { 286238384Sjkim$rec="%rcx"; 287238384Sjkim$frame="%rdx"; 288238384Sjkim$context="%r8"; 289238384Sjkim$disp="%r9"; 290238384Sjkim 291238384Sjkim$code.=<<___; 292238384Sjkim.extern __imp_RtlVirtualUnwind 293238384Sjkim 294238384Sjkim.type se_handler,\@abi-omnipotent 295238384Sjkim.align 16 296238384Sjkimse_handler: 297238384Sjkim push %rsi 298238384Sjkim push %rdi 299238384Sjkim push %rbx 300238384Sjkim push %rbp 301238384Sjkim push %r12 302238384Sjkim push %r13 303238384Sjkim push %r14 304238384Sjkim push %r15 305238384Sjkim pushfq 306238384Sjkim sub \$64,%rsp 307238384Sjkim 308238384Sjkim mov 152($context),%rax # pull context->Rsp 309238384Sjkim mov 248($context),%rbx # pull context->Rip 310238384Sjkim 311238384Sjkim lea .Lbody_mul_2x2(%rip),%r10 312238384Sjkim cmp %r10,%rbx # context->Rip<"prologue" label 313238384Sjkim jb .Lin_prologue 314238384Sjkim 315238384Sjkim mov 8*10(%rax),%r14 # mimic epilogue 316238384Sjkim mov 8*11(%rax),%r13 317238384Sjkim mov 8*12(%rax),%r12 318238384Sjkim mov 8*13(%rax),%rbp 319238384Sjkim mov 8*14(%rax),%rbx 320238384Sjkim mov 8*15(%rax),%rdi 321238384Sjkim mov 8*16(%rax),%rsi 322238384Sjkim 323238384Sjkim mov %rbx,144($context) # restore context->Rbx 324238384Sjkim mov %rbp,160($context) # restore context->Rbp 325238384Sjkim mov %rsi,168($context) # restore context->Rsi 326238384Sjkim mov %rdi,176($context) # restore context->Rdi 327238384Sjkim mov %r12,216($context) # restore context->R12 328238384Sjkim mov %r13,224($context) # restore context->R13 329238384Sjkim mov %r14,232($context) # restore context->R14 330238384Sjkim 331238384Sjkim.Lin_prologue: 332238384Sjkim lea 8*17(%rax),%rax 333238384Sjkim mov %rax,152($context) # restore context->Rsp 334238384Sjkim 335238384Sjkim mov 40($disp),%rdi # disp->ContextRecord 336238384Sjkim mov $context,%rsi # context 337238384Sjkim mov \$154,%ecx # sizeof(CONTEXT) 338238384Sjkim .long 0xa548f3fc # cld; rep movsq 339238384Sjkim 340238384Sjkim mov $disp,%rsi 341238384Sjkim xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER 342238384Sjkim mov 8(%rsi),%rdx # arg2, disp->ImageBase 343238384Sjkim mov 0(%rsi),%r8 # arg3, disp->ControlPc 344238384Sjkim mov 16(%rsi),%r9 # arg4, disp->FunctionEntry 345238384Sjkim mov 40(%rsi),%r10 # disp->ContextRecord 346238384Sjkim lea 56(%rsi),%r11 # &disp->HandlerData 347238384Sjkim lea 24(%rsi),%r12 # &disp->EstablisherFrame 348238384Sjkim mov %r10,32(%rsp) # arg5 349238384Sjkim mov %r11,40(%rsp) # arg6 350238384Sjkim mov %r12,48(%rsp) # arg7 351238384Sjkim mov %rcx,56(%rsp) # arg8, (NULL) 352238384Sjkim call *__imp_RtlVirtualUnwind(%rip) 353238384Sjkim 354238384Sjkim mov \$1,%eax # ExceptionContinueSearch 355238384Sjkim add \$64,%rsp 356238384Sjkim popfq 357238384Sjkim pop %r15 358238384Sjkim pop %r14 359238384Sjkim pop %r13 360238384Sjkim pop %r12 361238384Sjkim pop %rbp 362238384Sjkim pop %rbx 363238384Sjkim pop %rdi 364238384Sjkim pop %rsi 365238384Sjkim ret 366238384Sjkim.size se_handler,.-se_handler 367238384Sjkim 368238384Sjkim.section .pdata 369238384Sjkim.align 4 370238384Sjkim .rva _mul_1x1 371238384Sjkim .rva .Lend_mul_1x1 372238384Sjkim .rva .LSEH_info_1x1 373238384Sjkim 374238384Sjkim .rva .Lvanilla_mul_2x2 375238384Sjkim .rva .Lend_mul_2x2 376238384Sjkim .rva .LSEH_info_2x2 377238384Sjkim.section .xdata 378238384Sjkim.align 8 379238384Sjkim.LSEH_info_1x1: 380238384Sjkim .byte 0x01,0x07,0x02,0x00 381238384Sjkim .byte 0x07,0x01,0x11,0x00 # sub rsp,128+8 382238384Sjkim.LSEH_info_2x2: 383238384Sjkim .byte 9,0,0,0 384238384Sjkim .rva se_handler 385238384Sjkim___ 386238384Sjkim} 387238384Sjkim 388238384Sjkim$code =~ s/\`([^\`]*)\`/eval($1)/gem; 389238384Sjkimprint $code; 390238384Sjkimclose STDOUT; 391