| 1 |
#!/usr/bin/env perl |
| 2 |
|
| 3 |
# ==================================================================== |
| 4 |
# Written by Andy Polyakov <[email protected]> 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 |
# August 2011. |
| 11 |
# |
| 12 |
# Companion to x86_64-mont.pl that optimizes cache-timing attack |
| 13 |
# countermeasures. The subroutines are produced by replacing bp[i] |
| 14 |
# references in their x86_64-mont.pl counterparts with cache-neutral |
| 15 |
# references to powers table computed in BN_mod_exp_mont_consttime. |
| 16 |
# In addition subroutine that scatters elements of the powers table |
| 17 |
# is implemented, so that scatter-/gathering can be tuned without |
| 18 |
# bn_exp.c modifications. |
| 19 |
|
| 20 |
# August 2013. |
| 21 |
# |
| 22 |
# Add MULX/AD*X code paths and additional interfaces to optimize for |
| 23 |
# branch prediction unit. For input lengths that are multiples of 8 |
| 24 |
# the np argument is not just modulus value, but one interleaved |
| 25 |
# with 0. This is to optimize post-condition... |
| 26 |
|
| 27 |
$flavour = shift; |
| 28 |
$output = shift; |
| 29 |
if ($flavour =~ /\./) { $output = $flavour; undef $flavour; } |
| 30 |
|
| 31 |
$win64=0; $win64=1 if ($flavour =~ /[nm]asm|mingw64/ || $output =~ /\.asm$/); |
| 32 |
|
| 33 |
$0 =~ m/(.*[\/\\])[^\/\\]+$/; $dir=$1; |
| 34 |
( $xlate="${dir}x86_64-xlate.pl" and -f $xlate ) or |
| 35 |
( $xlate="${dir}../../perlasm/x86_64-xlate.pl" and -f $xlate) or |
| 36 |
die "can't locate x86_64-xlate.pl"; |
| 37 |
|
| 38 |
open OUT,"| \"$^X\" $xlate $flavour $output"; |
| 39 |
*STDOUT=*OUT; |
| 40 |
|
| 41 |
# In upstream, this is controlled by shelling out to the compiler to check |
| 42 |
# versions, but BoringSSL is intended to be used with pre-generated perlasm |
| 43 |
# output, so this isn't useful anyway. |
| 44 |
# |
| 45 |
# TODO(davidben): Enable this after testing. $addx goes up to 1. |
| 46 |
$addx = 0; |
| 47 |
|
| 48 |
# int bn_mul_mont_gather5( |
| 49 |
$rp="%rdi"; # BN_ULONG *rp, |
| 50 |
$ap="%rsi"; # const BN_ULONG *ap, |
| 51 |
$bp="%rdx"; # const BN_ULONG *bp, |
| 52 |
$np="%rcx"; # const BN_ULONG *np, |
| 53 |
$n0="%r8"; # const BN_ULONG *n0, |
| 54 |
$num="%r9"; # int num, |
| 55 |
# int idx); # 0 to 2^5-1, "index" in $bp holding |
| 56 |
# pre-computed powers of a', interlaced |
| 57 |
# in such manner that b[0] is $bp[idx], |
| 58 |
# b[1] is [2^5+idx], etc. |
| 59 |
$lo0="%r10"; |
| 60 |
$hi0="%r11"; |
| 61 |
$hi1="%r13"; |
| 62 |
$i="%r14"; |
| 63 |
$j="%r15"; |
| 64 |
$m0="%rbx"; |
| 65 |
$m1="%rbp"; |
| 66 |
|
| 67 |
$code=<<___; |
| 68 |
.text |
| 69 |
|
| 70 |
.extern OPENSSL_ia32cap_P |
| 71 |
|
| 72 |
.globl bn_mul_mont_gather5 |
| 73 |
.type bn_mul_mont_gather5,\@function,6 |
| 74 |
.align 64 |
| 75 |
bn_mul_mont_gather5: |
| 76 |
test \$7,${num}d |
| 77 |
jnz .Lmul_enter |
| 78 |
___ |
| 79 |
$code.=<<___ if ($addx); |
| 80 |
mov OPENSSL_ia32cap_P+8(%rip),%r11d |
| 81 |
___ |
| 82 |
$code.=<<___; |
| 83 |
jmp .Lmul4x_enter |
| 84 |
|
| 85 |
.align 16 |
| 86 |
.Lmul_enter: |
| 87 |
mov ${num}d,${num}d |
| 88 |
mov %rsp,%rax |
| 89 |
movd `($win64?56:8)`(%rsp),%xmm5 # load 7th argument |
| 90 |
lea .Linc(%rip),%r10 |
| 91 |
push %rbx |
| 92 |
push %rbp |
| 93 |
push %r12 |
| 94 |
push %r13 |
| 95 |
push %r14 |
| 96 |
push %r15 |
| 97 |
|
| 98 |
lea 2($num),%r11 |
| 99 |
neg %r11 |
| 100 |
lea -264(%rsp,%r11,8),%rsp # tp=alloca(8*(num+2)+256+8) |
| 101 |
and \$-1024,%rsp # minimize TLB usage |
| 102 |
|
| 103 |
mov %rax,8(%rsp,$num,8) # tp[num+1]=%rsp |
| 104 |
.Lmul_body: |
| 105 |
lea 128($bp),%r12 # reassign $bp (+size optimization) |
| 106 |
___ |
| 107 |
$bp="%r12"; |
| 108 |
$STRIDE=2**5*8; # 5 is "window size" |
| 109 |
$N=$STRIDE/4; # should match cache line size |
| 110 |
$code.=<<___; |
| 111 |
movdqa 0(%r10),%xmm0 # 00000001000000010000000000000000 |
| 112 |
movdqa 16(%r10),%xmm1 # 00000002000000020000000200000002 |
| 113 |
lea 24-112(%rsp,$num,8),%r10# place the mask after tp[num+3] (+ICache optimization) |
| 114 |
and \$-16,%r10 |
| 115 |
|
| 116 |
pshufd \$0,%xmm5,%xmm5 # broadcast index |
| 117 |
movdqa %xmm1,%xmm4 |
| 118 |
movdqa %xmm1,%xmm2 |
| 119 |
___ |
| 120 |
######################################################################## |
| 121 |
# calculate mask by comparing 0..31 to index and save result to stack |
| 122 |
# |
| 123 |
$code.=<<___; |
| 124 |
paddd %xmm0,%xmm1 |
| 125 |
pcmpeqd %xmm5,%xmm0 # compare to 1,0 |
| 126 |
.byte 0x67 |
| 127 |
movdqa %xmm4,%xmm3 |
| 128 |
___ |
| 129 |
for($k=0;$k<$STRIDE/16-4;$k+=4) { |
| 130 |
$code.=<<___; |
| 131 |
paddd %xmm1,%xmm2 |
| 132 |
pcmpeqd %xmm5,%xmm1 # compare to 3,2 |
| 133 |
movdqa %xmm0,`16*($k+0)+112`(%r10) |
| 134 |
movdqa %xmm4,%xmm0 |
| 135 |
|
| 136 |
paddd %xmm2,%xmm3 |
| 137 |
pcmpeqd %xmm5,%xmm2 # compare to 5,4 |
| 138 |
movdqa %xmm1,`16*($k+1)+112`(%r10) |
| 139 |
movdqa %xmm4,%xmm1 |
| 140 |
|
| 141 |
paddd %xmm3,%xmm0 |
| 142 |
pcmpeqd %xmm5,%xmm3 # compare to 7,6 |
| 143 |
movdqa %xmm2,`16*($k+2)+112`(%r10) |
| 144 |
movdqa %xmm4,%xmm2 |
| 145 |
|
| 146 |
paddd %xmm0,%xmm1 |
| 147 |
pcmpeqd %xmm5,%xmm0 |
| 148 |
movdqa %xmm3,`16*($k+3)+112`(%r10) |
| 149 |
movdqa %xmm4,%xmm3 |
| 150 |
___ |
| 151 |
} |
| 152 |
$code.=<<___; # last iteration can be optimized |
| 153 |
paddd %xmm1,%xmm2 |
| 154 |
pcmpeqd %xmm5,%xmm1 |
| 155 |
movdqa %xmm0,`16*($k+0)+112`(%r10) |
| 156 |
|
| 157 |
paddd %xmm2,%xmm3 |
| 158 |
.byte 0x67 |
| 159 |
pcmpeqd %xmm5,%xmm2 |
| 160 |
movdqa %xmm1,`16*($k+1)+112`(%r10) |
| 161 |
|
| 162 |
pcmpeqd %xmm5,%xmm3 |
| 163 |
movdqa %xmm2,`16*($k+2)+112`(%r10) |
| 164 |
pand `16*($k+0)-128`($bp),%xmm0 # while it's still in register |
| 165 |
|
| 166 |
pand `16*($k+1)-128`($bp),%xmm1 |
| 167 |
pand `16*($k+2)-128`($bp),%xmm2 |
| 168 |
movdqa %xmm3,`16*($k+3)+112`(%r10) |
| 169 |
pand `16*($k+3)-128`($bp),%xmm3 |
| 170 |
por %xmm2,%xmm0 |
| 171 |
por %xmm3,%xmm1 |
| 172 |
___ |
| 173 |
for($k=0;$k<$STRIDE/16-4;$k+=4) { |
| 174 |
$code.=<<___; |
| 175 |
movdqa `16*($k+0)-128`($bp),%xmm4 |
| 176 |
movdqa `16*($k+1)-128`($bp),%xmm5 |
| 177 |
movdqa `16*($k+2)-128`($bp),%xmm2 |
| 178 |
pand `16*($k+0)+112`(%r10),%xmm4 |
| 179 |
movdqa `16*($k+3)-128`($bp),%xmm3 |
| 180 |
pand `16*($k+1)+112`(%r10),%xmm5 |
| 181 |
por %xmm4,%xmm0 |
| 182 |
pand `16*($k+2)+112`(%r10),%xmm2 |
| 183 |
por %xmm5,%xmm1 |
| 184 |
pand `16*($k+3)+112`(%r10),%xmm3 |
| 185 |
por %xmm2,%xmm0 |
| 186 |
por %xmm3,%xmm1 |
| 187 |
___ |
| 188 |
} |
| 189 |
$code.=<<___; |
| 190 |
por %xmm1,%xmm0 |
| 191 |
pshufd \$0x4e,%xmm0,%xmm1 |
| 192 |
por %xmm1,%xmm0 |
| 193 |
lea $STRIDE($bp),$bp |
| 194 |
movq %xmm0,$m0 # m0=bp[0] |
| 195 |
|
| 196 |
mov ($n0),$n0 # pull n0[0] value |
| 197 |
mov ($ap),%rax |
| 198 |
|
| 199 |
xor $i,$i # i=0 |
| 200 |
xor $j,$j # j=0 |
| 201 |
|
| 202 |
mov $n0,$m1 |
| 203 |
mulq $m0 # ap[0]*bp[0] |
| 204 |
mov %rax,$lo0 |
| 205 |
mov ($np),%rax |
| 206 |
|
| 207 |
imulq $lo0,$m1 # "tp[0]"*n0 |
| 208 |
mov %rdx,$hi0 |
| 209 |
|
| 210 |
mulq $m1 # np[0]*m1 |
| 211 |
add %rax,$lo0 # discarded |
| 212 |
mov 8($ap),%rax |
| 213 |
adc \$0,%rdx |
| 214 |
mov %rdx,$hi1 |
| 215 |
|
| 216 |
lea 1($j),$j # j++ |
| 217 |
jmp .L1st_enter |
| 218 |
|
| 219 |
.align 16 |
| 220 |
.L1st: |
| 221 |
add %rax,$hi1 |
| 222 |
mov ($ap,$j,8),%rax |
| 223 |
adc \$0,%rdx |
| 224 |
add $hi0,$hi1 # np[j]*m1+ap[j]*bp[0] |
| 225 |
mov $lo0,$hi0 |
| 226 |
adc \$0,%rdx |
| 227 |
mov $hi1,-16(%rsp,$j,8) # tp[j-1] |
| 228 |
mov %rdx,$hi1 |
| 229 |
|
| 230 |
.L1st_enter: |
| 231 |
mulq $m0 # ap[j]*bp[0] |
| 232 |
add %rax,$hi0 |
| 233 |
mov ($np,$j,8),%rax |
| 234 |
adc \$0,%rdx |
| 235 |
lea 1($j),$j # j++ |
| 236 |
mov %rdx,$lo0 |
| 237 |
|
| 238 |
mulq $m1 # np[j]*m1 |
| 239 |
cmp $num,$j |
| 240 |
jne .L1st # note that upon exit $j==$num, so |
| 241 |
# they can be used interchangeably |
| 242 |
|
| 243 |
add %rax,$hi1 |
| 244 |
adc \$0,%rdx |
| 245 |
add $hi0,$hi1 # np[j]*m1+ap[j]*bp[0] |
| 246 |
adc \$0,%rdx |
| 247 |
mov $hi1,-16(%rsp,$num,8) # tp[num-1] |
| 248 |
mov %rdx,$hi1 |
| 249 |
mov $lo0,$hi0 |
| 250 |
|
| 251 |
xor %rdx,%rdx |
| 252 |
add $hi0,$hi1 |
| 253 |
adc \$0,%rdx |
| 254 |
mov $hi1,-8(%rsp,$num,8) |
| 255 |
mov %rdx,(%rsp,$num,8) # store upmost overflow bit |
| 256 |
|
| 257 |
lea 1($i),$i # i++ |
| 258 |
jmp .Louter |
| 259 |
.align 16 |
| 260 |
.Louter: |
| 261 |
lea 24+128(%rsp,$num,8),%rdx # where 256-byte mask is (+size optimization) |
| 262 |
and \$-16,%rdx |
| 263 |
pxor %xmm4,%xmm4 |
| 264 |
pxor %xmm5,%xmm5 |
| 265 |
___ |
| 266 |
for($k=0;$k<$STRIDE/16;$k+=4) { |
| 267 |
$code.=<<___; |
| 268 |
movdqa `16*($k+0)-128`($bp),%xmm0 |
| 269 |
movdqa `16*($k+1)-128`($bp),%xmm1 |
| 270 |
movdqa `16*($k+2)-128`($bp),%xmm2 |
| 271 |
movdqa `16*($k+3)-128`($bp),%xmm3 |
| 272 |
pand `16*($k+0)-128`(%rdx),%xmm0 |
| 273 |
pand `16*($k+1)-128`(%rdx),%xmm1 |
| 274 |
por %xmm0,%xmm4 |
| 275 |
pand `16*($k+2)-128`(%rdx),%xmm2 |
| 276 |
por %xmm1,%xmm5 |
| 277 |
pand `16*($k+3)-128`(%rdx),%xmm3 |
| 278 |
por %xmm2,%xmm4 |
| 279 |
por %xmm3,%xmm5 |
| 280 |
___ |
| 281 |
} |
| 282 |
$code.=<<___; |
| 283 |
por %xmm5,%xmm4 |
| 284 |
pshufd \$0x4e,%xmm4,%xmm0 |
| 285 |
por %xmm4,%xmm0 |
| 286 |
lea $STRIDE($bp),$bp |
| 287 |
|
| 288 |
mov ($ap),%rax # ap[0] |
| 289 |
movq %xmm0,$m0 # m0=bp[i] |
| 290 |
|
| 291 |
xor $j,$j # j=0 |
| 292 |
mov $n0,$m1 |
| 293 |
mov (%rsp),$lo0 |
| 294 |
|
| 295 |
mulq $m0 # ap[0]*bp[i] |
| 296 |
add %rax,$lo0 # ap[0]*bp[i]+tp[0] |
| 297 |
mov ($np),%rax |
| 298 |
adc \$0,%rdx |
| 299 |
|
| 300 |
imulq $lo0,$m1 # tp[0]*n0 |
| 301 |
mov %rdx,$hi0 |
| 302 |
|
| 303 |
mulq $m1 # np[0]*m1 |
| 304 |
add %rax,$lo0 # discarded |
| 305 |
mov 8($ap),%rax |
| 306 |
adc \$0,%rdx |
| 307 |
mov 8(%rsp),$lo0 # tp[1] |
| 308 |
mov %rdx,$hi1 |
| 309 |
|
| 310 |
lea 1($j),$j # j++ |
| 311 |
jmp .Linner_enter |
| 312 |
|
| 313 |
.align 16 |
| 314 |
.Linner: |
| 315 |
add %rax,$hi1 |
| 316 |
mov ($ap,$j,8),%rax |
| 317 |
adc \$0,%rdx |
| 318 |
add $lo0,$hi1 # np[j]*m1+ap[j]*bp[i]+tp[j] |
| 319 |
mov (%rsp,$j,8),$lo0 |
| 320 |
adc \$0,%rdx |
| 321 |
mov $hi1,-16(%rsp,$j,8) # tp[j-1] |
| 322 |
mov %rdx,$hi1 |
| 323 |
|
| 324 |
.Linner_enter: |
| 325 |
mulq $m0 # ap[j]*bp[i] |
| 326 |
add %rax,$hi0 |
| 327 |
mov ($np,$j,8),%rax |
| 328 |
adc \$0,%rdx |
| 329 |
add $hi0,$lo0 # ap[j]*bp[i]+tp[j] |
| 330 |
mov %rdx,$hi0 |
| 331 |
adc \$0,$hi0 |
| 332 |
lea 1($j),$j # j++ |
| 333 |
|
| 334 |
mulq $m1 # np[j]*m1 |
| 335 |
cmp $num,$j |
| 336 |
jne .Linner # note that upon exit $j==$num, so |
| 337 |
# they can be used interchangeably |
| 338 |
add %rax,$hi1 |
| 339 |
adc \$0,%rdx |
| 340 |
add $lo0,$hi1 # np[j]*m1+ap[j]*bp[i]+tp[j] |
| 341 |
mov (%rsp,$num,8),$lo0 |
| 342 |
adc \$0,%rdx |
| 343 |
mov $hi1,-16(%rsp,$num,8) # tp[num-1] |
| 344 |
mov %rdx,$hi1 |
| 345 |
|
| 346 |
xor %rdx,%rdx |
| 347 |
add $hi0,$hi1 |
| 348 |
adc \$0,%rdx |
| 349 |
add $lo0,$hi1 # pull upmost overflow bit |
| 350 |
adc \$0,%rdx |
| 351 |
mov $hi1,-8(%rsp,$num,8) |
| 352 |
mov %rdx,(%rsp,$num,8) # store upmost overflow bit |
| 353 |
|
| 354 |
lea 1($i),$i # i++ |
| 355 |
cmp $num,$i |
| 356 |
jb .Louter |
| 357 |
|
| 358 |
xor $i,$i # i=0 and clear CF! |
| 359 |
mov (%rsp),%rax # tp[0] |
| 360 |
lea (%rsp),$ap # borrow ap for tp |
| 361 |
mov $num,$j # j=num |
| 362 |
jmp .Lsub |
| 363 |
.align 16 |
| 364 |
.Lsub: sbb ($np,$i,8),%rax |
| 365 |
mov %rax,($rp,$i,8) # rp[i]=tp[i]-np[i] |
| 366 |
mov 8($ap,$i,8),%rax # tp[i+1] |
| 367 |
lea 1($i),$i # i++ |
| 368 |
dec $j # doesnn't affect CF! |
| 369 |
jnz .Lsub |
| 370 |
|
| 371 |
sbb \$0,%rax # handle upmost overflow bit |
| 372 |
xor $i,$i |
| 373 |
mov $num,$j # j=num |
| 374 |
.align 16 |
| 375 |
.Lcopy: # copy or in-place refresh |
| 376 |
mov (%rsp,$i,8),$ap |
| 377 |
mov ($rp,$i,8),$np |
| 378 |
xor $np,$ap # conditional select: |
| 379 |
and %rax,$ap # ((ap ^ np) & %rax) ^ np |
| 380 |
xor $np,$ap # ap = borrow?tp:rp |
| 381 |
mov $i,(%rsp,$i,8) # zap temporary vector |
| 382 |
mov $ap,($rp,$i,8) # rp[i]=tp[i] |
| 383 |
lea 1($i),$i |
| 384 |
sub \$1,$j |
| 385 |
jnz .Lcopy |
| 386 |
|
| 387 |
mov 8(%rsp,$num,8),%rsi # restore %rsp |
| 388 |
mov \$1,%rax |
| 389 |
|
| 390 |
mov -48(%rsi),%r15 |
| 391 |
mov -40(%rsi),%r14 |
| 392 |
mov -32(%rsi),%r13 |
| 393 |
mov -24(%rsi),%r12 |
| 394 |
mov -16(%rsi),%rbp |
| 395 |
mov -8(%rsi),%rbx |
| 396 |
lea (%rsi),%rsp |
| 397 |
.Lmul_epilogue: |
| 398 |
ret |
| 399 |
.size bn_mul_mont_gather5,.-bn_mul_mont_gather5 |
| 400 |
___ |
| 401 |
{{{ |
| 402 |
my @A=("%r10","%r11"); |
| 403 |
my @N=("%r13","%rdi"); |
| 404 |
$code.=<<___; |
| 405 |
.type bn_mul4x_mont_gather5,\@function,6 |
| 406 |
.align 32 |
| 407 |
bn_mul4x_mont_gather5: |
| 408 |
.Lmul4x_enter: |
| 409 |
___ |
| 410 |
$code.=<<___ if ($addx); |
| 411 |
and \$0x80108,%r11d |
| 412 |
cmp \$0x80108,%r11d # check for AD*X+BMI2+BMI1 |
| 413 |
je .Lmulx4x_enter |
| 414 |
___ |
| 415 |
$code.=<<___; |
| 416 |
.byte 0x67 |
| 417 |
mov %rsp,%rax |
| 418 |
push %rbx |
| 419 |
push %rbp |
| 420 |
push %r12 |
| 421 |
push %r13 |
| 422 |
push %r14 |
| 423 |
push %r15 |
| 424 |
|
| 425 |
.byte 0x67 |
| 426 |
shl \$3,${num}d # convert $num to bytes |
| 427 |
lea ($num,$num,2),%r10 # 3*$num in bytes |
| 428 |
neg $num # -$num |
| 429 |
|
| 430 |
############################################################## |
| 431 |
# Ensure that stack frame doesn't alias with $rptr+3*$num |
| 432 |
# modulo 4096, which covers ret[num], am[num] and n[num] |
| 433 |
# (see bn_exp.c). This is done to allow memory disambiguation |
| 434 |
# logic do its magic. [Extra [num] is allocated in order |
| 435 |
# to align with bn_power5's frame, which is cleansed after |
| 436 |
# completing exponentiation. Extra 256 bytes is for power mask |
| 437 |
# calculated from 7th argument, the index.] |
| 438 |
# |
| 439 |
lea -320(%rsp,$num,2),%r11 |
| 440 |
sub $rp,%r11 |
| 441 |
and \$4095,%r11 |
| 442 |
cmp %r11,%r10 |
| 443 |
jb .Lmul4xsp_alt |
| 444 |
sub %r11,%rsp # align with $rp |
| 445 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*num*8+256) |
| 446 |
jmp .Lmul4xsp_done |
| 447 |
|
| 448 |
.align 32 |
| 449 |
.Lmul4xsp_alt: |
| 450 |
lea 4096-320(,$num,2),%r10 |
| 451 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*num*8+256) |
| 452 |
sub %r10,%r11 |
| 453 |
mov \$0,%r10 |
| 454 |
cmovc %r10,%r11 |
| 455 |
sub %r11,%rsp |
| 456 |
.Lmul4xsp_done: |
| 457 |
and \$-64,%rsp |
| 458 |
neg $num |
| 459 |
|
| 460 |
mov %rax,40(%rsp) |
| 461 |
.Lmul4x_body: |
| 462 |
|
| 463 |
call mul4x_internal |
| 464 |
|
| 465 |
mov 40(%rsp),%rsi # restore %rsp |
| 466 |
mov \$1,%rax |
| 467 |
|
| 468 |
mov -48(%rsi),%r15 |
| 469 |
mov -40(%rsi),%r14 |
| 470 |
mov -32(%rsi),%r13 |
| 471 |
mov -24(%rsi),%r12 |
| 472 |
mov -16(%rsi),%rbp |
| 473 |
mov -8(%rsi),%rbx |
| 474 |
lea (%rsi),%rsp |
| 475 |
.Lmul4x_epilogue: |
| 476 |
ret |
| 477 |
.size bn_mul4x_mont_gather5,.-bn_mul4x_mont_gather5 |
| 478 |
|
| 479 |
.type mul4x_internal,\@abi-omnipotent |
| 480 |
.align 32 |
| 481 |
mul4x_internal: |
| 482 |
shl \$5,$num # $num was in bytes |
| 483 |
movd `($win64?56:8)`(%rax),%xmm5 # load 7th argument, index |
| 484 |
lea .Linc(%rip),%rax |
| 485 |
lea 128(%rdx,$num),%r13 # end of powers table (+size optimization) |
| 486 |
shr \$5,$num # restore $num |
| 487 |
___ |
| 488 |
$bp="%r12"; |
| 489 |
$STRIDE=2**5*8; # 5 is "window size" |
| 490 |
$N=$STRIDE/4; # should match cache line size |
| 491 |
$tp=$i; |
| 492 |
$code.=<<___; |
| 493 |
movdqa 0(%rax),%xmm0 # 00000001000000010000000000000000 |
| 494 |
movdqa 16(%rax),%xmm1 # 00000002000000020000000200000002 |
| 495 |
lea 88-112(%rsp,$num),%r10 # place the mask after tp[num+1] (+ICache optimization) |
| 496 |
lea 128(%rdx),$bp # size optimization |
| 497 |
|
| 498 |
pshufd \$0,%xmm5,%xmm5 # broadcast index |
| 499 |
movdqa %xmm1,%xmm4 |
| 500 |
.byte 0x67,0x67 |
| 501 |
movdqa %xmm1,%xmm2 |
| 502 |
___ |
| 503 |
######################################################################## |
| 504 |
# calculate mask by comparing 0..31 to index and save result to stack |
| 505 |
# |
| 506 |
$code.=<<___; |
| 507 |
paddd %xmm0,%xmm1 |
| 508 |
pcmpeqd %xmm5,%xmm0 # compare to 1,0 |
| 509 |
.byte 0x67 |
| 510 |
movdqa %xmm4,%xmm3 |
| 511 |
___ |
| 512 |
for($i=0;$i<$STRIDE/16-4;$i+=4) { |
| 513 |
$code.=<<___; |
| 514 |
paddd %xmm1,%xmm2 |
| 515 |
pcmpeqd %xmm5,%xmm1 # compare to 3,2 |
| 516 |
movdqa %xmm0,`16*($i+0)+112`(%r10) |
| 517 |
movdqa %xmm4,%xmm0 |
| 518 |
|
| 519 |
paddd %xmm2,%xmm3 |
| 520 |
pcmpeqd %xmm5,%xmm2 # compare to 5,4 |
| 521 |
movdqa %xmm1,`16*($i+1)+112`(%r10) |
| 522 |
movdqa %xmm4,%xmm1 |
| 523 |
|
| 524 |
paddd %xmm3,%xmm0 |
| 525 |
pcmpeqd %xmm5,%xmm3 # compare to 7,6 |
| 526 |
movdqa %xmm2,`16*($i+2)+112`(%r10) |
| 527 |
movdqa %xmm4,%xmm2 |
| 528 |
|
| 529 |
paddd %xmm0,%xmm1 |
| 530 |
pcmpeqd %xmm5,%xmm0 |
| 531 |
movdqa %xmm3,`16*($i+3)+112`(%r10) |
| 532 |
movdqa %xmm4,%xmm3 |
| 533 |
___ |
| 534 |
} |
| 535 |
$code.=<<___; # last iteration can be optimized |
| 536 |
paddd %xmm1,%xmm2 |
| 537 |
pcmpeqd %xmm5,%xmm1 |
| 538 |
movdqa %xmm0,`16*($i+0)+112`(%r10) |
| 539 |
|
| 540 |
paddd %xmm2,%xmm3 |
| 541 |
.byte 0x67 |
| 542 |
pcmpeqd %xmm5,%xmm2 |
| 543 |
movdqa %xmm1,`16*($i+1)+112`(%r10) |
| 544 |
|
| 545 |
pcmpeqd %xmm5,%xmm3 |
| 546 |
movdqa %xmm2,`16*($i+2)+112`(%r10) |
| 547 |
pand `16*($i+0)-128`($bp),%xmm0 # while it's still in register |
| 548 |
|
| 549 |
pand `16*($i+1)-128`($bp),%xmm1 |
| 550 |
pand `16*($i+2)-128`($bp),%xmm2 |
| 551 |
movdqa %xmm3,`16*($i+3)+112`(%r10) |
| 552 |
pand `16*($i+3)-128`($bp),%xmm3 |
| 553 |
por %xmm2,%xmm0 |
| 554 |
por %xmm3,%xmm1 |
| 555 |
___ |
| 556 |
for($i=0;$i<$STRIDE/16-4;$i+=4) { |
| 557 |
$code.=<<___; |
| 558 |
movdqa `16*($i+0)-128`($bp),%xmm4 |
| 559 |
movdqa `16*($i+1)-128`($bp),%xmm5 |
| 560 |
movdqa `16*($i+2)-128`($bp),%xmm2 |
| 561 |
pand `16*($i+0)+112`(%r10),%xmm4 |
| 562 |
movdqa `16*($i+3)-128`($bp),%xmm3 |
| 563 |
pand `16*($i+1)+112`(%r10),%xmm5 |
| 564 |
por %xmm4,%xmm0 |
| 565 |
pand `16*($i+2)+112`(%r10),%xmm2 |
| 566 |
por %xmm5,%xmm1 |
| 567 |
pand `16*($i+3)+112`(%r10),%xmm3 |
| 568 |
por %xmm2,%xmm0 |
| 569 |
por %xmm3,%xmm1 |
| 570 |
___ |
| 571 |
} |
| 572 |
$code.=<<___; |
| 573 |
por %xmm1,%xmm0 |
| 574 |
pshufd \$0x4e,%xmm0,%xmm1 |
| 575 |
por %xmm1,%xmm0 |
| 576 |
lea $STRIDE($bp),$bp |
| 577 |
movq %xmm0,$m0 # m0=bp[0] |
| 578 |
|
| 579 |
mov %r13,16+8(%rsp) # save end of b[num] |
| 580 |
mov $rp, 56+8(%rsp) # save $rp |
| 581 |
|
| 582 |
mov ($n0),$n0 # pull n0[0] value |
| 583 |
mov ($ap),%rax |
| 584 |
lea ($ap,$num),$ap # end of a[num] |
| 585 |
neg $num |
| 586 |
|
| 587 |
mov $n0,$m1 |
| 588 |
mulq $m0 # ap[0]*bp[0] |
| 589 |
mov %rax,$A[0] |
| 590 |
mov ($np),%rax |
| 591 |
|
| 592 |
imulq $A[0],$m1 # "tp[0]"*n0 |
| 593 |
lea 64+8(%rsp),$tp |
| 594 |
mov %rdx,$A[1] |
| 595 |
|
| 596 |
mulq $m1 # np[0]*m1 |
| 597 |
add %rax,$A[0] # discarded |
| 598 |
mov 8($ap,$num),%rax |
| 599 |
adc \$0,%rdx |
| 600 |
mov %rdx,$N[1] |
| 601 |
|
| 602 |
mulq $m0 |
| 603 |
add %rax,$A[1] |
| 604 |
mov 8*1($np),%rax |
| 605 |
adc \$0,%rdx |
| 606 |
mov %rdx,$A[0] |
| 607 |
|
| 608 |
mulq $m1 |
| 609 |
add %rax,$N[1] |
| 610 |
mov 16($ap,$num),%rax |
| 611 |
adc \$0,%rdx |
| 612 |
add $A[1],$N[1] |
| 613 |
lea 4*8($num),$j # j=4 |
| 614 |
lea 8*4($np),$np |
| 615 |
adc \$0,%rdx |
| 616 |
mov $N[1],($tp) |
| 617 |
mov %rdx,$N[0] |
| 618 |
jmp .L1st4x |
| 619 |
|
| 620 |
.align 32 |
| 621 |
.L1st4x: |
| 622 |
mulq $m0 # ap[j]*bp[0] |
| 623 |
add %rax,$A[0] |
| 624 |
mov -8*2($np),%rax |
| 625 |
lea 32($tp),$tp |
| 626 |
adc \$0,%rdx |
| 627 |
mov %rdx,$A[1] |
| 628 |
|
| 629 |
mulq $m1 # np[j]*m1 |
| 630 |
add %rax,$N[0] |
| 631 |
mov -8($ap,$j),%rax |
| 632 |
adc \$0,%rdx |
| 633 |
add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0] |
| 634 |
adc \$0,%rdx |
| 635 |
mov $N[0],-24($tp) # tp[j-1] |
| 636 |
mov %rdx,$N[1] |
| 637 |
|
| 638 |
mulq $m0 # ap[j]*bp[0] |
| 639 |
add %rax,$A[1] |
| 640 |
mov -8*1($np),%rax |
| 641 |
adc \$0,%rdx |
| 642 |
mov %rdx,$A[0] |
| 643 |
|
| 644 |
mulq $m1 # np[j]*m1 |
| 645 |
add %rax,$N[1] |
| 646 |
mov ($ap,$j),%rax |
| 647 |
adc \$0,%rdx |
| 648 |
add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0] |
| 649 |
adc \$0,%rdx |
| 650 |
mov $N[1],-16($tp) # tp[j-1] |
| 651 |
mov %rdx,$N[0] |
| 652 |
|
| 653 |
mulq $m0 # ap[j]*bp[0] |
| 654 |
add %rax,$A[0] |
| 655 |
mov 8*0($np),%rax |
| 656 |
adc \$0,%rdx |
| 657 |
mov %rdx,$A[1] |
| 658 |
|
| 659 |
mulq $m1 # np[j]*m1 |
| 660 |
add %rax,$N[0] |
| 661 |
mov 8($ap,$j),%rax |
| 662 |
adc \$0,%rdx |
| 663 |
add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0] |
| 664 |
adc \$0,%rdx |
| 665 |
mov $N[0],-8($tp) # tp[j-1] |
| 666 |
mov %rdx,$N[1] |
| 667 |
|
| 668 |
mulq $m0 # ap[j]*bp[0] |
| 669 |
add %rax,$A[1] |
| 670 |
mov 8*1($np),%rax |
| 671 |
adc \$0,%rdx |
| 672 |
mov %rdx,$A[0] |
| 673 |
|
| 674 |
mulq $m1 # np[j]*m1 |
| 675 |
add %rax,$N[1] |
| 676 |
mov 16($ap,$j),%rax |
| 677 |
adc \$0,%rdx |
| 678 |
add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0] |
| 679 |
lea 8*4($np),$np |
| 680 |
adc \$0,%rdx |
| 681 |
mov $N[1],($tp) # tp[j-1] |
| 682 |
mov %rdx,$N[0] |
| 683 |
|
| 684 |
add \$32,$j # j+=4 |
| 685 |
jnz .L1st4x |
| 686 |
|
| 687 |
mulq $m0 # ap[j]*bp[0] |
| 688 |
add %rax,$A[0] |
| 689 |
mov -8*2($np),%rax |
| 690 |
lea 32($tp),$tp |
| 691 |
adc \$0,%rdx |
| 692 |
mov %rdx,$A[1] |
| 693 |
|
| 694 |
mulq $m1 # np[j]*m1 |
| 695 |
add %rax,$N[0] |
| 696 |
mov -8($ap),%rax |
| 697 |
adc \$0,%rdx |
| 698 |
add $A[0],$N[0] # np[j]*m1+ap[j]*bp[0] |
| 699 |
adc \$0,%rdx |
| 700 |
mov $N[0],-24($tp) # tp[j-1] |
| 701 |
mov %rdx,$N[1] |
| 702 |
|
| 703 |
mulq $m0 # ap[j]*bp[0] |
| 704 |
add %rax,$A[1] |
| 705 |
mov -8*1($np),%rax |
| 706 |
adc \$0,%rdx |
| 707 |
mov %rdx,$A[0] |
| 708 |
|
| 709 |
mulq $m1 # np[j]*m1 |
| 710 |
add %rax,$N[1] |
| 711 |
mov ($ap,$num),%rax # ap[0] |
| 712 |
adc \$0,%rdx |
| 713 |
add $A[1],$N[1] # np[j]*m1+ap[j]*bp[0] |
| 714 |
adc \$0,%rdx |
| 715 |
mov $N[1],-16($tp) # tp[j-1] |
| 716 |
mov %rdx,$N[0] |
| 717 |
|
| 718 |
lea ($np,$num),$np # rewind $np |
| 719 |
|
| 720 |
xor $N[1],$N[1] |
| 721 |
add $A[0],$N[0] |
| 722 |
adc \$0,$N[1] |
| 723 |
mov $N[0],-8($tp) |
| 724 |
|
| 725 |
jmp .Louter4x |
| 726 |
|
| 727 |
.align 32 |
| 728 |
.Louter4x: |
| 729 |
lea 16+128($tp),%rdx # where 256-byte mask is (+size optimization) |
| 730 |
pxor %xmm4,%xmm4 |
| 731 |
pxor %xmm5,%xmm5 |
| 732 |
___ |
| 733 |
for($i=0;$i<$STRIDE/16;$i+=4) { |
| 734 |
$code.=<<___; |
| 735 |
movdqa `16*($i+0)-128`($bp),%xmm0 |
| 736 |
movdqa `16*($i+1)-128`($bp),%xmm1 |
| 737 |
movdqa `16*($i+2)-128`($bp),%xmm2 |
| 738 |
movdqa `16*($i+3)-128`($bp),%xmm3 |
| 739 |
pand `16*($i+0)-128`(%rdx),%xmm0 |
| 740 |
pand `16*($i+1)-128`(%rdx),%xmm1 |
| 741 |
por %xmm0,%xmm4 |
| 742 |
pand `16*($i+2)-128`(%rdx),%xmm2 |
| 743 |
por %xmm1,%xmm5 |
| 744 |
pand `16*($i+3)-128`(%rdx),%xmm3 |
| 745 |
por %xmm2,%xmm4 |
| 746 |
por %xmm3,%xmm5 |
| 747 |
___ |
| 748 |
} |
| 749 |
$code.=<<___; |
| 750 |
por %xmm5,%xmm4 |
| 751 |
pshufd \$0x4e,%xmm4,%xmm0 |
| 752 |
por %xmm4,%xmm0 |
| 753 |
lea $STRIDE($bp),$bp |
| 754 |
movq %xmm0,$m0 # m0=bp[i] |
| 755 |
|
| 756 |
mov ($tp,$num),$A[0] |
| 757 |
mov $n0,$m1 |
| 758 |
mulq $m0 # ap[0]*bp[i] |
| 759 |
add %rax,$A[0] # ap[0]*bp[i]+tp[0] |
| 760 |
mov ($np),%rax |
| 761 |
adc \$0,%rdx |
| 762 |
|
| 763 |
imulq $A[0],$m1 # tp[0]*n0 |
| 764 |
mov %rdx,$A[1] |
| 765 |
mov $N[1],($tp) # store upmost overflow bit |
| 766 |
|
| 767 |
lea ($tp,$num),$tp # rewind $tp |
| 768 |
|
| 769 |
mulq $m1 # np[0]*m1 |
| 770 |
add %rax,$A[0] # "$N[0]", discarded |
| 771 |
mov 8($ap,$num),%rax |
| 772 |
adc \$0,%rdx |
| 773 |
mov %rdx,$N[1] |
| 774 |
|
| 775 |
mulq $m0 # ap[j]*bp[i] |
| 776 |
add %rax,$A[1] |
| 777 |
mov 8*1($np),%rax |
| 778 |
adc \$0,%rdx |
| 779 |
add 8($tp),$A[1] # +tp[1] |
| 780 |
adc \$0,%rdx |
| 781 |
mov %rdx,$A[0] |
| 782 |
|
| 783 |
mulq $m1 # np[j]*m1 |
| 784 |
add %rax,$N[1] |
| 785 |
mov 16($ap,$num),%rax |
| 786 |
adc \$0,%rdx |
| 787 |
add $A[1],$N[1] # np[j]*m1+ap[j]*bp[i]+tp[j] |
| 788 |
lea 4*8($num),$j # j=4 |
| 789 |
lea 8*4($np),$np |
| 790 |
adc \$0,%rdx |
| 791 |
mov %rdx,$N[0] |
| 792 |
jmp .Linner4x |
| 793 |
|
| 794 |
.align 32 |
| 795 |
.Linner4x: |
| 796 |
mulq $m0 # ap[j]*bp[i] |
| 797 |
add %rax,$A[0] |
| 798 |
mov -8*2($np),%rax |
| 799 |
adc \$0,%rdx |
| 800 |
add 16($tp),$A[0] # ap[j]*bp[i]+tp[j] |
| 801 |
lea 32($tp),$tp |
| 802 |
adc \$0,%rdx |
| 803 |
mov %rdx,$A[1] |
| 804 |
|
| 805 |
mulq $m1 # np[j]*m1 |
| 806 |
add %rax,$N[0] |
| 807 |
mov -8($ap,$j),%rax |
| 808 |
adc \$0,%rdx |
| 809 |
add $A[0],$N[0] |
| 810 |
adc \$0,%rdx |
| 811 |
mov $N[1],-32($tp) # tp[j-1] |
| 812 |
mov %rdx,$N[1] |
| 813 |
|
| 814 |
mulq $m0 # ap[j]*bp[i] |
| 815 |
add %rax,$A[1] |
| 816 |
mov -8*1($np),%rax |
| 817 |
adc \$0,%rdx |
| 818 |
add -8($tp),$A[1] |
| 819 |
adc \$0,%rdx |
| 820 |
mov %rdx,$A[0] |
| 821 |
|
| 822 |
mulq $m1 # np[j]*m1 |
| 823 |
add %rax,$N[1] |
| 824 |
mov ($ap,$j),%rax |
| 825 |
adc \$0,%rdx |
| 826 |
add $A[1],$N[1] |
| 827 |
adc \$0,%rdx |
| 828 |
mov $N[0],-24($tp) # tp[j-1] |
| 829 |
mov %rdx,$N[0] |
| 830 |
|
| 831 |
mulq $m0 # ap[j]*bp[i] |
| 832 |
add %rax,$A[0] |
| 833 |
mov 8*0($np),%rax |
| 834 |
adc \$0,%rdx |
| 835 |
add ($tp),$A[0] # ap[j]*bp[i]+tp[j] |
| 836 |
adc \$0,%rdx |
| 837 |
mov %rdx,$A[1] |
| 838 |
|
| 839 |
mulq $m1 # np[j]*m1 |
| 840 |
add %rax,$N[0] |
| 841 |
mov 8($ap,$j),%rax |
| 842 |
adc \$0,%rdx |
| 843 |
add $A[0],$N[0] |
| 844 |
adc \$0,%rdx |
| 845 |
mov $N[1],-16($tp) # tp[j-1] |
| 846 |
mov %rdx,$N[1] |
| 847 |
|
| 848 |
mulq $m0 # ap[j]*bp[i] |
| 849 |
add %rax,$A[1] |
| 850 |
mov 8*1($np),%rax |
| 851 |
adc \$0,%rdx |
| 852 |
add 8($tp),$A[1] |
| 853 |
adc \$0,%rdx |
| 854 |
mov %rdx,$A[0] |
| 855 |
|
| 856 |
mulq $m1 # np[j]*m1 |
| 857 |
add %rax,$N[1] |
| 858 |
mov 16($ap,$j),%rax |
| 859 |
adc \$0,%rdx |
| 860 |
add $A[1],$N[1] |
| 861 |
lea 8*4($np),$np |
| 862 |
adc \$0,%rdx |
| 863 |
mov $N[0],-8($tp) # tp[j-1] |
| 864 |
mov %rdx,$N[0] |
| 865 |
|
| 866 |
add \$32,$j # j+=4 |
| 867 |
jnz .Linner4x |
| 868 |
|
| 869 |
mulq $m0 # ap[j]*bp[i] |
| 870 |
add %rax,$A[0] |
| 871 |
mov -8*2($np),%rax |
| 872 |
adc \$0,%rdx |
| 873 |
add 16($tp),$A[0] # ap[j]*bp[i]+tp[j] |
| 874 |
lea 32($tp),$tp |
| 875 |
adc \$0,%rdx |
| 876 |
mov %rdx,$A[1] |
| 877 |
|
| 878 |
mulq $m1 # np[j]*m1 |
| 879 |
add %rax,$N[0] |
| 880 |
mov -8($ap),%rax |
| 881 |
adc \$0,%rdx |
| 882 |
add $A[0],$N[0] |
| 883 |
adc \$0,%rdx |
| 884 |
mov $N[1],-32($tp) # tp[j-1] |
| 885 |
mov %rdx,$N[1] |
| 886 |
|
| 887 |
mulq $m0 # ap[j]*bp[i] |
| 888 |
add %rax,$A[1] |
| 889 |
mov $m1,%rax |
| 890 |
mov -8*1($np),$m1 |
| 891 |
adc \$0,%rdx |
| 892 |
add -8($tp),$A[1] |
| 893 |
adc \$0,%rdx |
| 894 |
mov %rdx,$A[0] |
| 895 |
|
| 896 |
mulq $m1 # np[j]*m1 |
| 897 |
add %rax,$N[1] |
| 898 |
mov ($ap,$num),%rax # ap[0] |
| 899 |
adc \$0,%rdx |
| 900 |
add $A[1],$N[1] |
| 901 |
adc \$0,%rdx |
| 902 |
mov $N[0],-24($tp) # tp[j-1] |
| 903 |
mov %rdx,$N[0] |
| 904 |
|
| 905 |
mov $N[1],-16($tp) # tp[j-1] |
| 906 |
lea ($np,$num),$np # rewind $np |
| 907 |
|
| 908 |
xor $N[1],$N[1] |
| 909 |
add $A[0],$N[0] |
| 910 |
adc \$0,$N[1] |
| 911 |
add ($tp),$N[0] # pull upmost overflow bit |
| 912 |
adc \$0,$N[1] # upmost overflow bit |
| 913 |
mov $N[0],-8($tp) |
| 914 |
|
| 915 |
cmp 16+8(%rsp),$bp |
| 916 |
jb .Louter4x |
| 917 |
___ |
| 918 |
if (1) { |
| 919 |
$code.=<<___; |
| 920 |
xor %rax,%rax |
| 921 |
sub $N[0],$m1 # compare top-most words |
| 922 |
adc $j,$j # $j is zero |
| 923 |
or $j,$N[1] |
| 924 |
sub $N[1],%rax # %rax=-$N[1] |
| 925 |
lea ($tp,$num),%rbx # tptr in .sqr4x_sub |
| 926 |
mov ($np),%r12 |
| 927 |
lea ($np),%rbp # nptr in .sqr4x_sub |
| 928 |
mov %r9,%rcx |
| 929 |
sar \$3+2,%rcx |
| 930 |
mov 56+8(%rsp),%rdi # rptr in .sqr4x_sub |
| 931 |
dec %r12 # so that after 'not' we get -n[0] |
| 932 |
xor %r10,%r10 |
| 933 |
mov 8*1(%rbp),%r13 |
| 934 |
mov 8*2(%rbp),%r14 |
| 935 |
mov 8*3(%rbp),%r15 |
| 936 |
jmp .Lsqr4x_sub_entry |
| 937 |
___ |
| 938 |
} else { |
| 939 |
my @ri=("%rax",$bp,$m0,$m1); |
| 940 |
my $rp="%rdx"; |
| 941 |
$code.=<<___ |
| 942 |
xor \$1,$N[1] |
| 943 |
lea ($tp,$num),$tp # rewind $tp |
| 944 |
sar \$5,$num # cf=0 |
| 945 |
lea ($np,$N[1],8),$np |
| 946 |
mov 56+8(%rsp),$rp # restore $rp |
| 947 |
jmp .Lsub4x |
| 948 |
|
| 949 |
.align 32 |
| 950 |
.Lsub4x: |
| 951 |
.byte 0x66 |
| 952 |
mov 8*0($tp),@ri[0] |
| 953 |
mov 8*1($tp),@ri[1] |
| 954 |
.byte 0x66 |
| 955 |
sbb 16*0($np),@ri[0] |
| 956 |
mov 8*2($tp),@ri[2] |
| 957 |
sbb 16*1($np),@ri[1] |
| 958 |
mov 3*8($tp),@ri[3] |
| 959 |
lea 4*8($tp),$tp |
| 960 |
sbb 16*2($np),@ri[2] |
| 961 |
mov @ri[0],8*0($rp) |
| 962 |
sbb 16*3($np),@ri[3] |
| 963 |
lea 16*4($np),$np |
| 964 |
mov @ri[1],8*1($rp) |
| 965 |
mov @ri[2],8*2($rp) |
| 966 |
mov @ri[3],8*3($rp) |
| 967 |
lea 8*4($rp),$rp |
| 968 |
|
| 969 |
inc $num |
| 970 |
jnz .Lsub4x |
| 971 |
|
| 972 |
ret |
| 973 |
___ |
| 974 |
} |
| 975 |
$code.=<<___; |
| 976 |
.size mul4x_internal,.-mul4x_internal |
| 977 |
___ |
| 978 |
}}} |
| 979 |
{{{ |
| 980 |
###################################################################### |
| 981 |
# void bn_power5( |
| 982 |
my $rptr="%rdi"; # BN_ULONG *rptr, |
| 983 |
my $aptr="%rsi"; # const BN_ULONG *aptr, |
| 984 |
my $bptr="%rdx"; # const void *table, |
| 985 |
my $nptr="%rcx"; # const BN_ULONG *nptr, |
| 986 |
my $n0 ="%r8"; # const BN_ULONG *n0); |
| 987 |
my $num ="%r9"; # int num, has to be divisible by 8 |
| 988 |
# int pwr |
| 989 |
|
| 990 |
my ($i,$j,$tptr)=("%rbp","%rcx",$rptr); |
| 991 |
my @A0=("%r10","%r11"); |
| 992 |
my @A1=("%r12","%r13"); |
| 993 |
my ($a0,$a1,$ai)=("%r14","%r15","%rbx"); |
| 994 |
|
| 995 |
$code.=<<___; |
| 996 |
.globl bn_power5 |
| 997 |
.type bn_power5,\@function,6 |
| 998 |
.align 32 |
| 999 |
bn_power5: |
| 1000 |
___ |
| 1001 |
$code.=<<___ if ($addx); |
| 1002 |
mov OPENSSL_ia32cap_P+8(%rip),%r11d |
| 1003 |
and \$0x80108,%r11d |
| 1004 |
cmp \$0x80108,%r11d # check for AD*X+BMI2+BMI1 |
| 1005 |
je .Lpowerx5_enter |
| 1006 |
___ |
| 1007 |
$code.=<<___; |
| 1008 |
mov %rsp,%rax |
| 1009 |
push %rbx |
| 1010 |
push %rbp |
| 1011 |
push %r12 |
| 1012 |
push %r13 |
| 1013 |
push %r14 |
| 1014 |
push %r15 |
| 1015 |
|
| 1016 |
shl \$3,${num}d # convert $num to bytes |
| 1017 |
lea ($num,$num,2),%r10d # 3*$num |
| 1018 |
neg $num |
| 1019 |
mov ($n0),$n0 # *n0 |
| 1020 |
|
| 1021 |
############################################################## |
| 1022 |
# Ensure that stack frame doesn't alias with $rptr+3*$num |
| 1023 |
# modulo 4096, which covers ret[num], am[num] and n[num] |
| 1024 |
# (see bn_exp.c). This is done to allow memory disambiguation |
| 1025 |
# logic do its magic. [Extra 256 bytes is for power mask |
| 1026 |
# calculated from 7th argument, the index.] |
| 1027 |
# |
| 1028 |
lea -320(%rsp,$num,2),%r11 |
| 1029 |
sub $rptr,%r11 |
| 1030 |
and \$4095,%r11 |
| 1031 |
cmp %r11,%r10 |
| 1032 |
jb .Lpwr_sp_alt |
| 1033 |
sub %r11,%rsp # align with $aptr |
| 1034 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*num*8+256) |
| 1035 |
jmp .Lpwr_sp_done |
| 1036 |
|
| 1037 |
.align 32 |
| 1038 |
.Lpwr_sp_alt: |
| 1039 |
lea 4096-320(,$num,2),%r10 |
| 1040 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*num*8+256) |
| 1041 |
sub %r10,%r11 |
| 1042 |
mov \$0,%r10 |
| 1043 |
cmovc %r10,%r11 |
| 1044 |
sub %r11,%rsp |
| 1045 |
.Lpwr_sp_done: |
| 1046 |
and \$-64,%rsp |
| 1047 |
mov $num,%r10 |
| 1048 |
neg $num |
| 1049 |
|
| 1050 |
############################################################## |
| 1051 |
# Stack layout |
| 1052 |
# |
| 1053 |
# +0 saved $num, used in reduction section |
| 1054 |
# +8 &t[2*$num], used in reduction section |
| 1055 |
# +32 saved *n0 |
| 1056 |
# +40 saved %rsp |
| 1057 |
# +48 t[2*$num] |
| 1058 |
# |
| 1059 |
mov $n0, 32(%rsp) |
| 1060 |
mov %rax, 40(%rsp) # save original %rsp |
| 1061 |
.Lpower5_body: |
| 1062 |
movq $rptr,%xmm1 # save $rptr, used in sqr8x |
| 1063 |
movq $nptr,%xmm2 # save $nptr |
| 1064 |
movq %r10, %xmm3 # -$num, used in sqr8x |
| 1065 |
movq $bptr,%xmm4 |
| 1066 |
|
| 1067 |
call __bn_sqr8x_internal |
| 1068 |
call __bn_post4x_internal |
| 1069 |
call __bn_sqr8x_internal |
| 1070 |
call __bn_post4x_internal |
| 1071 |
call __bn_sqr8x_internal |
| 1072 |
call __bn_post4x_internal |
| 1073 |
call __bn_sqr8x_internal |
| 1074 |
call __bn_post4x_internal |
| 1075 |
call __bn_sqr8x_internal |
| 1076 |
call __bn_post4x_internal |
| 1077 |
|
| 1078 |
movq %xmm2,$nptr |
| 1079 |
movq %xmm4,$bptr |
| 1080 |
mov $aptr,$rptr |
| 1081 |
mov 40(%rsp),%rax |
| 1082 |
lea 32(%rsp),$n0 |
| 1083 |
|
| 1084 |
call mul4x_internal |
| 1085 |
|
| 1086 |
mov 40(%rsp),%rsi # restore %rsp |
| 1087 |
mov \$1,%rax |
| 1088 |
mov -48(%rsi),%r15 |
| 1089 |
mov -40(%rsi),%r14 |
| 1090 |
mov -32(%rsi),%r13 |
| 1091 |
mov -24(%rsi),%r12 |
| 1092 |
mov -16(%rsi),%rbp |
| 1093 |
mov -8(%rsi),%rbx |
| 1094 |
lea (%rsi),%rsp |
| 1095 |
.Lpower5_epilogue: |
| 1096 |
ret |
| 1097 |
.size bn_power5,.-bn_power5 |
| 1098 |
|
| 1099 |
.globl bn_sqr8x_internal |
| 1100 |
.hidden bn_sqr8x_internal |
| 1101 |
.type bn_sqr8x_internal,\@abi-omnipotent |
| 1102 |
.align 32 |
| 1103 |
bn_sqr8x_internal: |
| 1104 |
__bn_sqr8x_internal: |
| 1105 |
############################################################## |
| 1106 |
# Squaring part: |
| 1107 |
# |
| 1108 |
# a) multiply-n-add everything but a[i]*a[i]; |
| 1109 |
# b) shift result of a) by 1 to the left and accumulate |
| 1110 |
# a[i]*a[i] products; |
| 1111 |
# |
| 1112 |
############################################################## |
| 1113 |
# a[1]a[0] |
| 1114 |
# a[2]a[0] |
| 1115 |
# a[3]a[0] |
| 1116 |
# a[2]a[1] |
| 1117 |
# a[4]a[0] |
| 1118 |
# a[3]a[1] |
| 1119 |
# a[5]a[0] |
| 1120 |
# a[4]a[1] |
| 1121 |
# a[3]a[2] |
| 1122 |
# a[6]a[0] |
| 1123 |
# a[5]a[1] |
| 1124 |
# a[4]a[2] |
| 1125 |
# a[7]a[0] |
| 1126 |
# a[6]a[1] |
| 1127 |
# a[5]a[2] |
| 1128 |
# a[4]a[3] |
| 1129 |
# a[7]a[1] |
| 1130 |
# a[6]a[2] |
| 1131 |
# a[5]a[3] |
| 1132 |
# a[7]a[2] |
| 1133 |
# a[6]a[3] |
| 1134 |
# a[5]a[4] |
| 1135 |
# a[7]a[3] |
| 1136 |
# a[6]a[4] |
| 1137 |
# a[7]a[4] |
| 1138 |
# a[6]a[5] |
| 1139 |
# a[7]a[5] |
| 1140 |
# a[7]a[6] |
| 1141 |
# a[1]a[0] |
| 1142 |
# a[2]a[0] |
| 1143 |
# a[3]a[0] |
| 1144 |
# a[4]a[0] |
| 1145 |
# a[5]a[0] |
| 1146 |
# a[6]a[0] |
| 1147 |
# a[7]a[0] |
| 1148 |
# a[2]a[1] |
| 1149 |
# a[3]a[1] |
| 1150 |
# a[4]a[1] |
| 1151 |
# a[5]a[1] |
| 1152 |
# a[6]a[1] |
| 1153 |
# a[7]a[1] |
| 1154 |
# a[3]a[2] |
| 1155 |
# a[4]a[2] |
| 1156 |
# a[5]a[2] |
| 1157 |
# a[6]a[2] |
| 1158 |
# a[7]a[2] |
| 1159 |
# a[4]a[3] |
| 1160 |
# a[5]a[3] |
| 1161 |
# a[6]a[3] |
| 1162 |
# a[7]a[3] |
| 1163 |
# a[5]a[4] |
| 1164 |
# a[6]a[4] |
| 1165 |
# a[7]a[4] |
| 1166 |
# a[6]a[5] |
| 1167 |
# a[7]a[5] |
| 1168 |
# a[7]a[6] |
| 1169 |
# a[0]a[0] |
| 1170 |
# a[1]a[1] |
| 1171 |
# a[2]a[2] |
| 1172 |
# a[3]a[3] |
| 1173 |
# a[4]a[4] |
| 1174 |
# a[5]a[5] |
| 1175 |
# a[6]a[6] |
| 1176 |
# a[7]a[7] |
| 1177 |
|
| 1178 |
lea 32(%r10),$i # $i=-($num-32) |
| 1179 |
lea ($aptr,$num),$aptr # end of a[] buffer, ($aptr,$i)=&ap[2] |
| 1180 |
|
| 1181 |
mov $num,$j # $j=$num |
| 1182 |
|
| 1183 |
# comments apply to $num==8 case |
| 1184 |
mov -32($aptr,$i),$a0 # a[0] |
| 1185 |
lea 48+8(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num] |
| 1186 |
mov -24($aptr,$i),%rax # a[1] |
| 1187 |
lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"] |
| 1188 |
mov -16($aptr,$i),$ai # a[2] |
| 1189 |
mov %rax,$a1 |
| 1190 |
|
| 1191 |
mul $a0 # a[1]*a[0] |
| 1192 |
mov %rax,$A0[0] # a[1]*a[0] |
| 1193 |
mov $ai,%rax # a[2] |
| 1194 |
mov %rdx,$A0[1] |
| 1195 |
mov $A0[0],-24($tptr,$i) # t[1] |
| 1196 |
|
| 1197 |
mul $a0 # a[2]*a[0] |
| 1198 |
add %rax,$A0[1] |
| 1199 |
mov $ai,%rax |
| 1200 |
adc \$0,%rdx |
| 1201 |
mov $A0[1],-16($tptr,$i) # t[2] |
| 1202 |
mov %rdx,$A0[0] |
| 1203 |
|
| 1204 |
|
| 1205 |
mov -8($aptr,$i),$ai # a[3] |
| 1206 |
mul $a1 # a[2]*a[1] |
| 1207 |
mov %rax,$A1[0] # a[2]*a[1]+t[3] |
| 1208 |
mov $ai,%rax |
| 1209 |
mov %rdx,$A1[1] |
| 1210 |
|
| 1211 |
lea ($i),$j |
| 1212 |
mul $a0 # a[3]*a[0] |
| 1213 |
add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3] |
| 1214 |
mov $ai,%rax |
| 1215 |
mov %rdx,$A0[1] |
| 1216 |
adc \$0,$A0[1] |
| 1217 |
add $A1[0],$A0[0] |
| 1218 |
adc \$0,$A0[1] |
| 1219 |
mov $A0[0],-8($tptr,$j) # t[3] |
| 1220 |
jmp .Lsqr4x_1st |
| 1221 |
|
| 1222 |
.align 32 |
| 1223 |
.Lsqr4x_1st: |
| 1224 |
mov ($aptr,$j),$ai # a[4] |
| 1225 |
mul $a1 # a[3]*a[1] |
| 1226 |
add %rax,$A1[1] # a[3]*a[1]+t[4] |
| 1227 |
mov $ai,%rax |
| 1228 |
mov %rdx,$A1[0] |
| 1229 |
adc \$0,$A1[0] |
| 1230 |
|
| 1231 |
mul $a0 # a[4]*a[0] |
| 1232 |
add %rax,$A0[1] # a[4]*a[0]+a[3]*a[1]+t[4] |
| 1233 |
mov $ai,%rax # a[3] |
| 1234 |
mov 8($aptr,$j),$ai # a[5] |
| 1235 |
mov %rdx,$A0[0] |
| 1236 |
adc \$0,$A0[0] |
| 1237 |
add $A1[1],$A0[1] |
| 1238 |
adc \$0,$A0[0] |
| 1239 |
|
| 1240 |
|
| 1241 |
mul $a1 # a[4]*a[3] |
| 1242 |
add %rax,$A1[0] # a[4]*a[3]+t[5] |
| 1243 |
mov $ai,%rax |
| 1244 |
mov $A0[1],($tptr,$j) # t[4] |
| 1245 |
mov %rdx,$A1[1] |
| 1246 |
adc \$0,$A1[1] |
| 1247 |
|
| 1248 |
mul $a0 # a[5]*a[2] |
| 1249 |
add %rax,$A0[0] # a[5]*a[2]+a[4]*a[3]+t[5] |
| 1250 |
mov $ai,%rax |
| 1251 |
mov 16($aptr,$j),$ai # a[6] |
| 1252 |
mov %rdx,$A0[1] |
| 1253 |
adc \$0,$A0[1] |
| 1254 |
add $A1[0],$A0[0] |
| 1255 |
adc \$0,$A0[1] |
| 1256 |
|
| 1257 |
mul $a1 # a[5]*a[3] |
| 1258 |
add %rax,$A1[1] # a[5]*a[3]+t[6] |
| 1259 |
mov $ai,%rax |
| 1260 |
mov $A0[0],8($tptr,$j) # t[5] |
| 1261 |
mov %rdx,$A1[0] |
| 1262 |
adc \$0,$A1[0] |
| 1263 |
|
| 1264 |
mul $a0 # a[6]*a[2] |
| 1265 |
add %rax,$A0[1] # a[6]*a[2]+a[5]*a[3]+t[6] |
| 1266 |
mov $ai,%rax # a[3] |
| 1267 |
mov 24($aptr,$j),$ai # a[7] |
| 1268 |
mov %rdx,$A0[0] |
| 1269 |
adc \$0,$A0[0] |
| 1270 |
add $A1[1],$A0[1] |
| 1271 |
adc \$0,$A0[0] |
| 1272 |
|
| 1273 |
|
| 1274 |
mul $a1 # a[6]*a[5] |
| 1275 |
add %rax,$A1[0] # a[6]*a[5]+t[7] |
| 1276 |
mov $ai,%rax |
| 1277 |
mov $A0[1],16($tptr,$j) # t[6] |
| 1278 |
mov %rdx,$A1[1] |
| 1279 |
adc \$0,$A1[1] |
| 1280 |
lea 32($j),$j |
| 1281 |
|
| 1282 |
mul $a0 # a[7]*a[4] |
| 1283 |
add %rax,$A0[0] # a[7]*a[4]+a[6]*a[5]+t[6] |
| 1284 |
mov $ai,%rax |
| 1285 |
mov %rdx,$A0[1] |
| 1286 |
adc \$0,$A0[1] |
| 1287 |
add $A1[0],$A0[0] |
| 1288 |
adc \$0,$A0[1] |
| 1289 |
mov $A0[0],-8($tptr,$j) # t[7] |
| 1290 |
|
| 1291 |
cmp \$0,$j |
| 1292 |
jne .Lsqr4x_1st |
| 1293 |
|
| 1294 |
mul $a1 # a[7]*a[5] |
| 1295 |
add %rax,$A1[1] |
| 1296 |
lea 16($i),$i |
| 1297 |
adc \$0,%rdx |
| 1298 |
add $A0[1],$A1[1] |
| 1299 |
adc \$0,%rdx |
| 1300 |
|
| 1301 |
mov $A1[1],($tptr) # t[8] |
| 1302 |
mov %rdx,$A1[0] |
| 1303 |
mov %rdx,8($tptr) # t[9] |
| 1304 |
jmp .Lsqr4x_outer |
| 1305 |
|
| 1306 |
.align 32 |
| 1307 |
.Lsqr4x_outer: # comments apply to $num==6 case |
| 1308 |
mov -32($aptr,$i),$a0 # a[0] |
| 1309 |
lea 48+8(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num] |
| 1310 |
mov -24($aptr,$i),%rax # a[1] |
| 1311 |
lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"] |
| 1312 |
mov -16($aptr,$i),$ai # a[2] |
| 1313 |
mov %rax,$a1 |
| 1314 |
|
| 1315 |
mul $a0 # a[1]*a[0] |
| 1316 |
mov -24($tptr,$i),$A0[0] # t[1] |
| 1317 |
add %rax,$A0[0] # a[1]*a[0]+t[1] |
| 1318 |
mov $ai,%rax # a[2] |
| 1319 |
adc \$0,%rdx |
| 1320 |
mov $A0[0],-24($tptr,$i) # t[1] |
| 1321 |
mov %rdx,$A0[1] |
| 1322 |
|
| 1323 |
mul $a0 # a[2]*a[0] |
| 1324 |
add %rax,$A0[1] |
| 1325 |
mov $ai,%rax |
| 1326 |
adc \$0,%rdx |
| 1327 |
add -16($tptr,$i),$A0[1] # a[2]*a[0]+t[2] |
| 1328 |
mov %rdx,$A0[0] |
| 1329 |
adc \$0,$A0[0] |
| 1330 |
mov $A0[1],-16($tptr,$i) # t[2] |
| 1331 |
|
| 1332 |
xor $A1[0],$A1[0] |
| 1333 |
|
| 1334 |
mov -8($aptr,$i),$ai # a[3] |
| 1335 |
mul $a1 # a[2]*a[1] |
| 1336 |
add %rax,$A1[0] # a[2]*a[1]+t[3] |
| 1337 |
mov $ai,%rax |
| 1338 |
adc \$0,%rdx |
| 1339 |
add -8($tptr,$i),$A1[0] |
| 1340 |
mov %rdx,$A1[1] |
| 1341 |
adc \$0,$A1[1] |
| 1342 |
|
| 1343 |
mul $a0 # a[3]*a[0] |
| 1344 |
add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3] |
| 1345 |
mov $ai,%rax |
| 1346 |
adc \$0,%rdx |
| 1347 |
add $A1[0],$A0[0] |
| 1348 |
mov %rdx,$A0[1] |
| 1349 |
adc \$0,$A0[1] |
| 1350 |
mov $A0[0],-8($tptr,$i) # t[3] |
| 1351 |
|
| 1352 |
lea ($i),$j |
| 1353 |
jmp .Lsqr4x_inner |
| 1354 |
|
| 1355 |
.align 32 |
| 1356 |
.Lsqr4x_inner: |
| 1357 |
mov ($aptr,$j),$ai # a[4] |
| 1358 |
mul $a1 # a[3]*a[1] |
| 1359 |
add %rax,$A1[1] # a[3]*a[1]+t[4] |
| 1360 |
mov $ai,%rax |
| 1361 |
mov %rdx,$A1[0] |
| 1362 |
adc \$0,$A1[0] |
| 1363 |
add ($tptr,$j),$A1[1] |
| 1364 |
adc \$0,$A1[0] |
| 1365 |
|
| 1366 |
.byte 0x67 |
| 1367 |
mul $a0 # a[4]*a[0] |
| 1368 |
add %rax,$A0[1] # a[4]*a[0]+a[3]*a[1]+t[4] |
| 1369 |
mov $ai,%rax # a[3] |
| 1370 |
mov 8($aptr,$j),$ai # a[5] |
| 1371 |
mov %rdx,$A0[0] |
| 1372 |
adc \$0,$A0[0] |
| 1373 |
add $A1[1],$A0[1] |
| 1374 |
adc \$0,$A0[0] |
| 1375 |
|
| 1376 |
mul $a1 # a[4]*a[3] |
| 1377 |
add %rax,$A1[0] # a[4]*a[3]+t[5] |
| 1378 |
mov $A0[1],($tptr,$j) # t[4] |
| 1379 |
mov $ai,%rax |
| 1380 |
mov %rdx,$A1[1] |
| 1381 |
adc \$0,$A1[1] |
| 1382 |
add 8($tptr,$j),$A1[0] |
| 1383 |
lea 16($j),$j # j++ |
| 1384 |
adc \$0,$A1[1] |
| 1385 |
|
| 1386 |
mul $a0 # a[5]*a[2] |
| 1387 |
add %rax,$A0[0] # a[5]*a[2]+a[4]*a[3]+t[5] |
| 1388 |
mov $ai,%rax |
| 1389 |
adc \$0,%rdx |
| 1390 |
add $A1[0],$A0[0] |
| 1391 |
mov %rdx,$A0[1] |
| 1392 |
adc \$0,$A0[1] |
| 1393 |
mov $A0[0],-8($tptr,$j) # t[5], "preloaded t[1]" below |
| 1394 |
|
| 1395 |
cmp \$0,$j |
| 1396 |
jne .Lsqr4x_inner |
| 1397 |
|
| 1398 |
.byte 0x67 |
| 1399 |
mul $a1 # a[5]*a[3] |
| 1400 |
add %rax,$A1[1] |
| 1401 |
adc \$0,%rdx |
| 1402 |
add $A0[1],$A1[1] |
| 1403 |
adc \$0,%rdx |
| 1404 |
|
| 1405 |
mov $A1[1],($tptr) # t[6], "preloaded t[2]" below |
| 1406 |
mov %rdx,$A1[0] |
| 1407 |
mov %rdx,8($tptr) # t[7], "preloaded t[3]" below |
| 1408 |
|
| 1409 |
add \$16,$i |
| 1410 |
jnz .Lsqr4x_outer |
| 1411 |
|
| 1412 |
# comments apply to $num==4 case |
| 1413 |
mov -32($aptr),$a0 # a[0] |
| 1414 |
lea 48+8(%rsp,$num,2),$tptr # end of tp[] buffer, &tp[2*$num] |
| 1415 |
mov -24($aptr),%rax # a[1] |
| 1416 |
lea -32($tptr,$i),$tptr # end of tp[] window, &tp[2*$num-"$i"] |
| 1417 |
mov -16($aptr),$ai # a[2] |
| 1418 |
mov %rax,$a1 |
| 1419 |
|
| 1420 |
mul $a0 # a[1]*a[0] |
| 1421 |
add %rax,$A0[0] # a[1]*a[0]+t[1], preloaded t[1] |
| 1422 |
mov $ai,%rax # a[2] |
| 1423 |
mov %rdx,$A0[1] |
| 1424 |
adc \$0,$A0[1] |
| 1425 |
|
| 1426 |
mul $a0 # a[2]*a[0] |
| 1427 |
add %rax,$A0[1] |
| 1428 |
mov $ai,%rax |
| 1429 |
mov $A0[0],-24($tptr) # t[1] |
| 1430 |
mov %rdx,$A0[0] |
| 1431 |
adc \$0,$A0[0] |
| 1432 |
add $A1[1],$A0[1] # a[2]*a[0]+t[2], preloaded t[2] |
| 1433 |
mov -8($aptr),$ai # a[3] |
| 1434 |
adc \$0,$A0[0] |
| 1435 |
|
| 1436 |
mul $a1 # a[2]*a[1] |
| 1437 |
add %rax,$A1[0] # a[2]*a[1]+t[3], preloaded t[3] |
| 1438 |
mov $ai,%rax |
| 1439 |
mov $A0[1],-16($tptr) # t[2] |
| 1440 |
mov %rdx,$A1[1] |
| 1441 |
adc \$0,$A1[1] |
| 1442 |
|
| 1443 |
mul $a0 # a[3]*a[0] |
| 1444 |
add %rax,$A0[0] # a[3]*a[0]+a[2]*a[1]+t[3] |
| 1445 |
mov $ai,%rax |
| 1446 |
mov %rdx,$A0[1] |
| 1447 |
adc \$0,$A0[1] |
| 1448 |
add $A1[0],$A0[0] |
| 1449 |
adc \$0,$A0[1] |
| 1450 |
mov $A0[0],-8($tptr) # t[3] |
| 1451 |
|
| 1452 |
mul $a1 # a[3]*a[1] |
| 1453 |
add %rax,$A1[1] |
| 1454 |
mov -16($aptr),%rax # a[2] |
| 1455 |
adc \$0,%rdx |
| 1456 |
add $A0[1],$A1[1] |
| 1457 |
adc \$0,%rdx |
| 1458 |
|
| 1459 |
mov $A1[1],($tptr) # t[4] |
| 1460 |
mov %rdx,$A1[0] |
| 1461 |
mov %rdx,8($tptr) # t[5] |
| 1462 |
|
| 1463 |
mul $ai # a[2]*a[3] |
| 1464 |
___ |
| 1465 |
{ |
| 1466 |
my ($shift,$carry)=($a0,$a1); |
| 1467 |
my @S=(@A1,$ai,$n0); |
| 1468 |
$code.=<<___; |
| 1469 |
add \$16,$i |
| 1470 |
xor $shift,$shift |
| 1471 |
sub $num,$i # $i=16-$num |
| 1472 |
xor $carry,$carry |
| 1473 |
|
| 1474 |
add $A1[0],%rax # t[5] |
| 1475 |
adc \$0,%rdx |
| 1476 |
mov %rax,8($tptr) # t[5] |
| 1477 |
mov %rdx,16($tptr) # t[6] |
| 1478 |
mov $carry,24($tptr) # t[7] |
| 1479 |
|
| 1480 |
mov -16($aptr,$i),%rax # a[0] |
| 1481 |
lea 48+8(%rsp),$tptr |
| 1482 |
xor $A0[0],$A0[0] # t[0] |
| 1483 |
mov 8($tptr),$A0[1] # t[1] |
| 1484 |
|
| 1485 |
lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift |
| 1486 |
shr \$63,$A0[0] |
| 1487 |
lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 | |
| 1488 |
shr \$63,$A0[1] |
| 1489 |
or $A0[0],$S[1] # | t[2*i]>>63 |
| 1490 |
mov 16($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1491 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1492 |
mul %rax # a[i]*a[i] |
| 1493 |
neg $carry # mov $carry,cf |
| 1494 |
mov 24($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1495 |
adc %rax,$S[0] |
| 1496 |
mov -8($aptr,$i),%rax # a[i+1] # prefetch |
| 1497 |
mov $S[0],($tptr) |
| 1498 |
adc %rdx,$S[1] |
| 1499 |
|
| 1500 |
lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift |
| 1501 |
mov $S[1],8($tptr) |
| 1502 |
sbb $carry,$carry # mov cf,$carry |
| 1503 |
shr \$63,$A0[0] |
| 1504 |
lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 | |
| 1505 |
shr \$63,$A0[1] |
| 1506 |
or $A0[0],$S[3] # | t[2*i]>>63 |
| 1507 |
mov 32($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1508 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1509 |
mul %rax # a[i]*a[i] |
| 1510 |
neg $carry # mov $carry,cf |
| 1511 |
mov 40($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1512 |
adc %rax,$S[2] |
| 1513 |
mov 0($aptr,$i),%rax # a[i+1] # prefetch |
| 1514 |
mov $S[2],16($tptr) |
| 1515 |
adc %rdx,$S[3] |
| 1516 |
lea 16($i),$i |
| 1517 |
mov $S[3],24($tptr) |
| 1518 |
sbb $carry,$carry # mov cf,$carry |
| 1519 |
lea 64($tptr),$tptr |
| 1520 |
jmp .Lsqr4x_shift_n_add |
| 1521 |
|
| 1522 |
.align 32 |
| 1523 |
.Lsqr4x_shift_n_add: |
| 1524 |
lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift |
| 1525 |
shr \$63,$A0[0] |
| 1526 |
lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 | |
| 1527 |
shr \$63,$A0[1] |
| 1528 |
or $A0[0],$S[1] # | t[2*i]>>63 |
| 1529 |
mov -16($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1530 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1531 |
mul %rax # a[i]*a[i] |
| 1532 |
neg $carry # mov $carry,cf |
| 1533 |
mov -8($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1534 |
adc %rax,$S[0] |
| 1535 |
mov -8($aptr,$i),%rax # a[i+1] # prefetch |
| 1536 |
mov $S[0],-32($tptr) |
| 1537 |
adc %rdx,$S[1] |
| 1538 |
|
| 1539 |
lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift |
| 1540 |
mov $S[1],-24($tptr) |
| 1541 |
sbb $carry,$carry # mov cf,$carry |
| 1542 |
shr \$63,$A0[0] |
| 1543 |
lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 | |
| 1544 |
shr \$63,$A0[1] |
| 1545 |
or $A0[0],$S[3] # | t[2*i]>>63 |
| 1546 |
mov 0($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1547 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1548 |
mul %rax # a[i]*a[i] |
| 1549 |
neg $carry # mov $carry,cf |
| 1550 |
mov 8($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1551 |
adc %rax,$S[2] |
| 1552 |
mov 0($aptr,$i),%rax # a[i+1] # prefetch |
| 1553 |
mov $S[2],-16($tptr) |
| 1554 |
adc %rdx,$S[3] |
| 1555 |
|
| 1556 |
lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift |
| 1557 |
mov $S[3],-8($tptr) |
| 1558 |
sbb $carry,$carry # mov cf,$carry |
| 1559 |
shr \$63,$A0[0] |
| 1560 |
lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 | |
| 1561 |
shr \$63,$A0[1] |
| 1562 |
or $A0[0],$S[1] # | t[2*i]>>63 |
| 1563 |
mov 16($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1564 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1565 |
mul %rax # a[i]*a[i] |
| 1566 |
neg $carry # mov $carry,cf |
| 1567 |
mov 24($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1568 |
adc %rax,$S[0] |
| 1569 |
mov 8($aptr,$i),%rax # a[i+1] # prefetch |
| 1570 |
mov $S[0],0($tptr) |
| 1571 |
adc %rdx,$S[1] |
| 1572 |
|
| 1573 |
lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1 | shift |
| 1574 |
mov $S[1],8($tptr) |
| 1575 |
sbb $carry,$carry # mov cf,$carry |
| 1576 |
shr \$63,$A0[0] |
| 1577 |
lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 | |
| 1578 |
shr \$63,$A0[1] |
| 1579 |
or $A0[0],$S[3] # | t[2*i]>>63 |
| 1580 |
mov 32($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1581 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1582 |
mul %rax # a[i]*a[i] |
| 1583 |
neg $carry # mov $carry,cf |
| 1584 |
mov 40($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1585 |
adc %rax,$S[2] |
| 1586 |
mov 16($aptr,$i),%rax # a[i+1] # prefetch |
| 1587 |
mov $S[2],16($tptr) |
| 1588 |
adc %rdx,$S[3] |
| 1589 |
mov $S[3],24($tptr) |
| 1590 |
sbb $carry,$carry # mov cf,$carry |
| 1591 |
lea 64($tptr),$tptr |
| 1592 |
add \$32,$i |
| 1593 |
jnz .Lsqr4x_shift_n_add |
| 1594 |
|
| 1595 |
lea ($shift,$A0[0],2),$S[0] # t[2*i]<<1 | shift |
| 1596 |
.byte 0x67 |
| 1597 |
shr \$63,$A0[0] |
| 1598 |
lea ($j,$A0[1],2),$S[1] # t[2*i+1]<<1 | |
| 1599 |
shr \$63,$A0[1] |
| 1600 |
or $A0[0],$S[1] # | t[2*i]>>63 |
| 1601 |
mov -16($tptr),$A0[0] # t[2*i+2] # prefetch |
| 1602 |
mov $A0[1],$shift # shift=t[2*i+1]>>63 |
| 1603 |
mul %rax # a[i]*a[i] |
| 1604 |
neg $carry # mov $carry,cf |
| 1605 |
mov -8($tptr),$A0[1] # t[2*i+2+1] # prefetch |
| 1606 |
adc %rax,$S[0] |
| 1607 |
mov -8($aptr),%rax # a[i+1] # prefetch |
| 1608 |
mov $S[0],-32($tptr) |
| 1609 |
adc %rdx,$S[1] |
| 1610 |
|
| 1611 |
lea ($shift,$A0[0],2),$S[2] # t[2*i]<<1|shift |
| 1612 |
mov $S[1],-24($tptr) |
| 1613 |
sbb $carry,$carry # mov cf,$carry |
| 1614 |
shr \$63,$A0[0] |
| 1615 |
lea ($j,$A0[1],2),$S[3] # t[2*i+1]<<1 | |
| 1616 |
shr \$63,$A0[1] |
| 1617 |
or $A0[0],$S[3] # | t[2*i]>>63 |
| 1618 |
mul %rax # a[i]*a[i] |
| 1619 |
neg $carry # mov $carry,cf |
| 1620 |
adc %rax,$S[2] |
| 1621 |
adc %rdx,$S[3] |
| 1622 |
mov $S[2],-16($tptr) |
| 1623 |
mov $S[3],-8($tptr) |
| 1624 |
___ |
| 1625 |
} |
| 1626 |
###################################################################### |
| 1627 |
# Montgomery reduction part, "word-by-word" algorithm. |
| 1628 |
# |
| 1629 |
# This new path is inspired by multiple submissions from Intel, by |
| 1630 |
# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford, |
| 1631 |
# Vinodh Gopal... |
| 1632 |
{ |
| 1633 |
my ($nptr,$tptr,$carry,$m0)=("%rbp","%rdi","%rsi","%rbx"); |
| 1634 |
|
| 1635 |
$code.=<<___; |
| 1636 |
movq %xmm2,$nptr |
| 1637 |
__bn_sqr8x_reduction: |
| 1638 |
xor %rax,%rax |
| 1639 |
lea ($nptr,$num),%rcx # end of n[] |
| 1640 |
lea 48+8(%rsp,$num,2),%rdx # end of t[] buffer |
| 1641 |
mov %rcx,0+8(%rsp) |
| 1642 |
lea 48+8(%rsp,$num),$tptr # end of initial t[] window |
| 1643 |
mov %rdx,8+8(%rsp) |
| 1644 |
neg $num |
| 1645 |
jmp .L8x_reduction_loop |
| 1646 |
|
| 1647 |
.align 32 |
| 1648 |
.L8x_reduction_loop: |
| 1649 |
lea ($tptr,$num),$tptr # start of current t[] window |
| 1650 |
.byte 0x66 |
| 1651 |
mov 8*0($tptr),$m0 |
| 1652 |
mov 8*1($tptr),%r9 |
| 1653 |
mov 8*2($tptr),%r10 |
| 1654 |
mov 8*3($tptr),%r11 |
| 1655 |
mov 8*4($tptr),%r12 |
| 1656 |
mov 8*5($tptr),%r13 |
| 1657 |
mov 8*6($tptr),%r14 |
| 1658 |
mov 8*7($tptr),%r15 |
| 1659 |
mov %rax,(%rdx) # store top-most carry bit |
| 1660 |
lea 8*8($tptr),$tptr |
| 1661 |
|
| 1662 |
.byte 0x67 |
| 1663 |
mov $m0,%r8 |
| 1664 |
imulq 32+8(%rsp),$m0 # n0*a[0] |
| 1665 |
mov 8*0($nptr),%rax # n[0] |
| 1666 |
mov \$8,%ecx |
| 1667 |
jmp .L8x_reduce |
| 1668 |
|
| 1669 |
.align 32 |
| 1670 |
.L8x_reduce: |
| 1671 |
mulq $m0 |
| 1672 |
mov 8*1($nptr),%rax # n[1] |
| 1673 |
neg %r8 |
| 1674 |
mov %rdx,%r8 |
| 1675 |
adc \$0,%r8 |
| 1676 |
|
| 1677 |
mulq $m0 |
| 1678 |
add %rax,%r9 |
| 1679 |
mov 8*2($nptr),%rax |
| 1680 |
adc \$0,%rdx |
| 1681 |
add %r9,%r8 |
| 1682 |
mov $m0,48-8+8(%rsp,%rcx,8) # put aside n0*a[i] |
| 1683 |
mov %rdx,%r9 |
| 1684 |
adc \$0,%r9 |
| 1685 |
|
| 1686 |
mulq $m0 |
| 1687 |
add %rax,%r10 |
| 1688 |
mov 8*3($nptr),%rax |
| 1689 |
adc \$0,%rdx |
| 1690 |
add %r10,%r9 |
| 1691 |
mov 32+8(%rsp),$carry # pull n0, borrow $carry |
| 1692 |
mov %rdx,%r10 |
| 1693 |
adc \$0,%r10 |
| 1694 |
|
| 1695 |
mulq $m0 |
| 1696 |
add %rax,%r11 |
| 1697 |
mov 8*4($nptr),%rax |
| 1698 |
adc \$0,%rdx |
| 1699 |
imulq %r8,$carry # modulo-scheduled |
| 1700 |
add %r11,%r10 |
| 1701 |
mov %rdx,%r11 |
| 1702 |
adc \$0,%r11 |
| 1703 |
|
| 1704 |
mulq $m0 |
| 1705 |
add %rax,%r12 |
| 1706 |
mov 8*5($nptr),%rax |
| 1707 |
adc \$0,%rdx |
| 1708 |
add %r12,%r11 |
| 1709 |
mov %rdx,%r12 |
| 1710 |
adc \$0,%r12 |
| 1711 |
|
| 1712 |
mulq $m0 |
| 1713 |
add %rax,%r13 |
| 1714 |
mov 8*6($nptr),%rax |
| 1715 |
adc \$0,%rdx |
| 1716 |
add %r13,%r12 |
| 1717 |
mov %rdx,%r13 |
| 1718 |
adc \$0,%r13 |
| 1719 |
|
| 1720 |
mulq $m0 |
| 1721 |
add %rax,%r14 |
| 1722 |
mov 8*7($nptr),%rax |
| 1723 |
adc \$0,%rdx |
| 1724 |
add %r14,%r13 |
| 1725 |
mov %rdx,%r14 |
| 1726 |
adc \$0,%r14 |
| 1727 |
|
| 1728 |
mulq $m0 |
| 1729 |
mov $carry,$m0 # n0*a[i] |
| 1730 |
add %rax,%r15 |
| 1731 |
mov 8*0($nptr),%rax # n[0] |
| 1732 |
adc \$0,%rdx |
| 1733 |
add %r15,%r14 |
| 1734 |
mov %rdx,%r15 |
| 1735 |
adc \$0,%r15 |
| 1736 |
|
| 1737 |
dec %ecx |
| 1738 |
jnz .L8x_reduce |
| 1739 |
|
| 1740 |
lea 8*8($nptr),$nptr |
| 1741 |
xor %rax,%rax |
| 1742 |
mov 8+8(%rsp),%rdx # pull end of t[] |
| 1743 |
cmp 0+8(%rsp),$nptr # end of n[]? |
| 1744 |
jae .L8x_no_tail |
| 1745 |
|
| 1746 |
.byte 0x66 |
| 1747 |
add 8*0($tptr),%r8 |
| 1748 |
adc 8*1($tptr),%r9 |
| 1749 |
adc 8*2($tptr),%r10 |
| 1750 |
adc 8*3($tptr),%r11 |
| 1751 |
adc 8*4($tptr),%r12 |
| 1752 |
adc 8*5($tptr),%r13 |
| 1753 |
adc 8*6($tptr),%r14 |
| 1754 |
adc 8*7($tptr),%r15 |
| 1755 |
sbb $carry,$carry # top carry |
| 1756 |
|
| 1757 |
mov 48+56+8(%rsp),$m0 # pull n0*a[0] |
| 1758 |
mov \$8,%ecx |
| 1759 |
mov 8*0($nptr),%rax |
| 1760 |
jmp .L8x_tail |
| 1761 |
|
| 1762 |
.align 32 |
| 1763 |
.L8x_tail: |
| 1764 |
mulq $m0 |
| 1765 |
add %rax,%r8 |
| 1766 |
mov 8*1($nptr),%rax |
| 1767 |
mov %r8,($tptr) # save result |
| 1768 |
mov %rdx,%r8 |
| 1769 |
adc \$0,%r8 |
| 1770 |
|
| 1771 |
mulq $m0 |
| 1772 |
add %rax,%r9 |
| 1773 |
mov 8*2($nptr),%rax |
| 1774 |
adc \$0,%rdx |
| 1775 |
add %r9,%r8 |
| 1776 |
lea 8($tptr),$tptr # $tptr++ |
| 1777 |
mov %rdx,%r9 |
| 1778 |
adc \$0,%r9 |
| 1779 |
|
| 1780 |
mulq $m0 |
| 1781 |
add %rax,%r10 |
| 1782 |
mov 8*3($nptr),%rax |
| 1783 |
adc \$0,%rdx |
| 1784 |
add %r10,%r9 |
| 1785 |
mov %rdx,%r10 |
| 1786 |
adc \$0,%r10 |
| 1787 |
|
| 1788 |
mulq $m0 |
| 1789 |
add %rax,%r11 |
| 1790 |
mov 8*4($nptr),%rax |
| 1791 |
adc \$0,%rdx |
| 1792 |
add %r11,%r10 |
| 1793 |
mov %rdx,%r11 |
| 1794 |
adc \$0,%r11 |
| 1795 |
|
| 1796 |
mulq $m0 |
| 1797 |
add %rax,%r12 |
| 1798 |
mov 8*5($nptr),%rax |
| 1799 |
adc \$0,%rdx |
| 1800 |
add %r12,%r11 |
| 1801 |
mov %rdx,%r12 |
| 1802 |
adc \$0,%r12 |
| 1803 |
|
| 1804 |
mulq $m0 |
| 1805 |
add %rax,%r13 |
| 1806 |
mov 8*6($nptr),%rax |
| 1807 |
adc \$0,%rdx |
| 1808 |
add %r13,%r12 |
| 1809 |
mov %rdx,%r13 |
| 1810 |
adc \$0,%r13 |
| 1811 |
|
| 1812 |
mulq $m0 |
| 1813 |
add %rax,%r14 |
| 1814 |
mov 8*7($nptr),%rax |
| 1815 |
adc \$0,%rdx |
| 1816 |
add %r14,%r13 |
| 1817 |
mov %rdx,%r14 |
| 1818 |
adc \$0,%r14 |
| 1819 |
|
| 1820 |
mulq $m0 |
| 1821 |
mov 48-16+8(%rsp,%rcx,8),$m0# pull n0*a[i] |
| 1822 |
add %rax,%r15 |
| 1823 |
adc \$0,%rdx |
| 1824 |
add %r15,%r14 |
| 1825 |
mov 8*0($nptr),%rax # pull n[0] |
| 1826 |
mov %rdx,%r15 |
| 1827 |
adc \$0,%r15 |
| 1828 |
|
| 1829 |
dec %ecx |
| 1830 |
jnz .L8x_tail |
| 1831 |
|
| 1832 |
lea 8*8($nptr),$nptr |
| 1833 |
mov 8+8(%rsp),%rdx # pull end of t[] |
| 1834 |
cmp 0+8(%rsp),$nptr # end of n[]? |
| 1835 |
jae .L8x_tail_done # break out of loop |
| 1836 |
|
| 1837 |
mov 48+56+8(%rsp),$m0 # pull n0*a[0] |
| 1838 |
neg $carry |
| 1839 |
mov 8*0($nptr),%rax # pull n[0] |
| 1840 |
adc 8*0($tptr),%r8 |
| 1841 |
adc 8*1($tptr),%r9 |
| 1842 |
adc 8*2($tptr),%r10 |
| 1843 |
adc 8*3($tptr),%r11 |
| 1844 |
adc 8*4($tptr),%r12 |
| 1845 |
adc 8*5($tptr),%r13 |
| 1846 |
adc 8*6($tptr),%r14 |
| 1847 |
adc 8*7($tptr),%r15 |
| 1848 |
sbb $carry,$carry # top carry |
| 1849 |
|
| 1850 |
mov \$8,%ecx |
| 1851 |
jmp .L8x_tail |
| 1852 |
|
| 1853 |
.align 32 |
| 1854 |
.L8x_tail_done: |
| 1855 |
add (%rdx),%r8 # can this overflow? |
| 1856 |
adc \$0,%r9 |
| 1857 |
adc \$0,%r10 |
| 1858 |
adc \$0,%r11 |
| 1859 |
adc \$0,%r12 |
| 1860 |
adc \$0,%r13 |
| 1861 |
adc \$0,%r14 |
| 1862 |
adc \$0,%r15 # can't overflow, because we |
| 1863 |
# started with "overhung" part |
| 1864 |
# of multiplication |
| 1865 |
xor %rax,%rax |
| 1866 |
|
| 1867 |
neg $carry |
| 1868 |
.L8x_no_tail: |
| 1869 |
adc 8*0($tptr),%r8 |
| 1870 |
adc 8*1($tptr),%r9 |
| 1871 |
adc 8*2($tptr),%r10 |
| 1872 |
adc 8*3($tptr),%r11 |
| 1873 |
adc 8*4($tptr),%r12 |
| 1874 |
adc 8*5($tptr),%r13 |
| 1875 |
adc 8*6($tptr),%r14 |
| 1876 |
adc 8*7($tptr),%r15 |
| 1877 |
adc \$0,%rax # top-most carry |
| 1878 |
mov -8($nptr),%rcx # np[num-1] |
| 1879 |
xor $carry,$carry |
| 1880 |
|
| 1881 |
movq %xmm2,$nptr # restore $nptr |
| 1882 |
|
| 1883 |
mov %r8,8*0($tptr) # store top 512 bits |
| 1884 |
mov %r9,8*1($tptr) |
| 1885 |
movq %xmm3,$num # $num is %r9, can't be moved upwards |
| 1886 |
mov %r10,8*2($tptr) |
| 1887 |
mov %r11,8*3($tptr) |
| 1888 |
mov %r12,8*4($tptr) |
| 1889 |
mov %r13,8*5($tptr) |
| 1890 |
mov %r14,8*6($tptr) |
| 1891 |
mov %r15,8*7($tptr) |
| 1892 |
lea 8*8($tptr),$tptr |
| 1893 |
|
| 1894 |
cmp %rdx,$tptr # end of t[]? |
| 1895 |
jb .L8x_reduction_loop |
| 1896 |
ret |
| 1897 |
.size bn_sqr8x_internal,.-bn_sqr8x_internal |
| 1898 |
___ |
| 1899 |
} |
| 1900 |
############################################################## |
| 1901 |
# Post-condition, 4x unrolled |
| 1902 |
# |
| 1903 |
{ |
| 1904 |
my ($tptr,$nptr)=("%rbx","%rbp"); |
| 1905 |
$code.=<<___; |
| 1906 |
.type __bn_post4x_internal,\@abi-omnipotent |
| 1907 |
.align 32 |
| 1908 |
__bn_post4x_internal: |
| 1909 |
mov 8*0($nptr),%r12 |
| 1910 |
lea (%rdi,$num),$tptr # %rdi was $tptr above |
| 1911 |
mov $num,%rcx |
| 1912 |
movq %xmm1,$rptr # restore $rptr |
| 1913 |
neg %rax |
| 1914 |
movq %xmm1,$aptr # prepare for back-to-back call |
| 1915 |
sar \$3+2,%rcx |
| 1916 |
dec %r12 # so that after 'not' we get -n[0] |
| 1917 |
xor %r10,%r10 |
| 1918 |
mov 8*1($nptr),%r13 |
| 1919 |
mov 8*2($nptr),%r14 |
| 1920 |
mov 8*3($nptr),%r15 |
| 1921 |
jmp .Lsqr4x_sub_entry |
| 1922 |
|
| 1923 |
.align 16 |
| 1924 |
.Lsqr4x_sub: |
| 1925 |
mov 8*0($nptr),%r12 |
| 1926 |
mov 8*1($nptr),%r13 |
| 1927 |
mov 8*2($nptr),%r14 |
| 1928 |
mov 8*3($nptr),%r15 |
| 1929 |
.Lsqr4x_sub_entry: |
| 1930 |
lea 8*4($nptr),$nptr |
| 1931 |
not %r12 |
| 1932 |
not %r13 |
| 1933 |
not %r14 |
| 1934 |
not %r15 |
| 1935 |
and %rax,%r12 |
| 1936 |
and %rax,%r13 |
| 1937 |
and %rax,%r14 |
| 1938 |
and %rax,%r15 |
| 1939 |
|
| 1940 |
neg %r10 # mov %r10,%cf |
| 1941 |
adc 8*0($tptr),%r12 |
| 1942 |
adc 8*1($tptr),%r13 |
| 1943 |
adc 8*2($tptr),%r14 |
| 1944 |
adc 8*3($tptr),%r15 |
| 1945 |
mov %r12,8*0($rptr) |
| 1946 |
lea 8*4($tptr),$tptr |
| 1947 |
mov %r13,8*1($rptr) |
| 1948 |
sbb %r10,%r10 # mov %cf,%r10 |
| 1949 |
mov %r14,8*2($rptr) |
| 1950 |
mov %r15,8*3($rptr) |
| 1951 |
lea 8*4($rptr),$rptr |
| 1952 |
|
| 1953 |
inc %rcx # pass %cf |
| 1954 |
jnz .Lsqr4x_sub |
| 1955 |
|
| 1956 |
mov $num,%r10 # prepare for back-to-back call |
| 1957 |
neg $num # restore $num |
| 1958 |
ret |
| 1959 |
.size __bn_post4x_internal,.-__bn_post4x_internal |
| 1960 |
___ |
| 1961 |
} |
| 1962 |
{ |
| 1963 |
$code.=<<___; |
| 1964 |
.globl bn_from_montgomery |
| 1965 |
.type bn_from_montgomery,\@abi-omnipotent |
| 1966 |
.align 32 |
| 1967 |
bn_from_montgomery: |
| 1968 |
testl \$7,`($win64?"48(%rsp)":"%r9d")` |
| 1969 |
jz bn_from_mont8x |
| 1970 |
xor %eax,%eax |
| 1971 |
ret |
| 1972 |
.size bn_from_montgomery,.-bn_from_montgomery |
| 1973 |
|
| 1974 |
.type bn_from_mont8x,\@function,6 |
| 1975 |
.align 32 |
| 1976 |
bn_from_mont8x: |
| 1977 |
.byte 0x67 |
| 1978 |
mov %rsp,%rax |
| 1979 |
push %rbx |
| 1980 |
push %rbp |
| 1981 |
push %r12 |
| 1982 |
push %r13 |
| 1983 |
push %r14 |
| 1984 |
push %r15 |
| 1985 |
|
| 1986 |
shl \$3,${num}d # convert $num to bytes |
| 1987 |
lea ($num,$num,2),%r10 # 3*$num in bytes |
| 1988 |
neg $num |
| 1989 |
mov ($n0),$n0 # *n0 |
| 1990 |
|
| 1991 |
############################################################## |
| 1992 |
# Ensure that stack frame doesn't alias with $rptr+3*$num |
| 1993 |
# modulo 4096, which covers ret[num], am[num] and n[num] |
| 1994 |
# (see bn_exp.c). The stack is allocated to aligned with |
| 1995 |
# bn_power5's frame, and as bn_from_montgomery happens to be |
| 1996 |
# last operation, we use the opportunity to cleanse it. |
| 1997 |
# |
| 1998 |
lea -320(%rsp,$num,2),%r11 |
| 1999 |
sub $rptr,%r11 |
| 2000 |
and \$4095,%r11 |
| 2001 |
cmp %r11,%r10 |
| 2002 |
jb .Lfrom_sp_alt |
| 2003 |
sub %r11,%rsp # align with $aptr |
| 2004 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*$num*8+256) |
| 2005 |
jmp .Lfrom_sp_done |
| 2006 |
|
| 2007 |
.align 32 |
| 2008 |
.Lfrom_sp_alt: |
| 2009 |
lea 4096-320(,$num,2),%r10 |
| 2010 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*$num*8+256) |
| 2011 |
sub %r10,%r11 |
| 2012 |
mov \$0,%r10 |
| 2013 |
cmovc %r10,%r11 |
| 2014 |
sub %r11,%rsp |
| 2015 |
.Lfrom_sp_done: |
| 2016 |
and \$-64,%rsp |
| 2017 |
mov $num,%r10 |
| 2018 |
neg $num |
| 2019 |
|
| 2020 |
############################################################## |
| 2021 |
# Stack layout |
| 2022 |
# |
| 2023 |
# +0 saved $num, used in reduction section |
| 2024 |
# +8 &t[2*$num], used in reduction section |
| 2025 |
# +32 saved *n0 |
| 2026 |
# +40 saved %rsp |
| 2027 |
# +48 t[2*$num] |
| 2028 |
# |
| 2029 |
mov $n0, 32(%rsp) |
| 2030 |
mov %rax, 40(%rsp) # save original %rsp |
| 2031 |
.Lfrom_body: |
| 2032 |
mov $num,%r11 |
| 2033 |
lea 48(%rsp),%rax |
| 2034 |
pxor %xmm0,%xmm0 |
| 2035 |
jmp .Lmul_by_1 |
| 2036 |
|
| 2037 |
.align 32 |
| 2038 |
.Lmul_by_1: |
| 2039 |
movdqu ($aptr),%xmm1 |
| 2040 |
movdqu 16($aptr),%xmm2 |
| 2041 |
movdqu 32($aptr),%xmm3 |
| 2042 |
movdqa %xmm0,(%rax,$num) |
| 2043 |
movdqu 48($aptr),%xmm4 |
| 2044 |
movdqa %xmm0,16(%rax,$num) |
| 2045 |
.byte 0x48,0x8d,0xb6,0x40,0x00,0x00,0x00 # lea 64($aptr),$aptr |
| 2046 |
movdqa %xmm1,(%rax) |
| 2047 |
movdqa %xmm0,32(%rax,$num) |
| 2048 |
movdqa %xmm2,16(%rax) |
| 2049 |
movdqa %xmm0,48(%rax,$num) |
| 2050 |
movdqa %xmm3,32(%rax) |
| 2051 |
movdqa %xmm4,48(%rax) |
| 2052 |
lea 64(%rax),%rax |
| 2053 |
sub \$64,%r11 |
| 2054 |
jnz .Lmul_by_1 |
| 2055 |
|
| 2056 |
movq $rptr,%xmm1 |
| 2057 |
movq $nptr,%xmm2 |
| 2058 |
.byte 0x67 |
| 2059 |
mov $nptr,%rbp |
| 2060 |
movq %r10, %xmm3 # -num |
| 2061 |
___ |
| 2062 |
$code.=<<___ if ($addx); |
| 2063 |
mov OPENSSL_ia32cap_P+8(%rip),%r11d |
| 2064 |
and \$0x80108,%r11d |
| 2065 |
cmp \$0x80108,%r11d # check for AD*X+BMI2+BMI1 |
| 2066 |
jne .Lfrom_mont_nox |
| 2067 |
|
| 2068 |
lea (%rax,$num),$rptr |
| 2069 |
call __bn_sqrx8x_reduction |
| 2070 |
call __bn_postx4x_internal |
| 2071 |
|
| 2072 |
pxor %xmm0,%xmm0 |
| 2073 |
lea 48(%rsp),%rax |
| 2074 |
mov 40(%rsp),%rsi # restore %rsp |
| 2075 |
jmp .Lfrom_mont_zero |
| 2076 |
|
| 2077 |
.align 32 |
| 2078 |
.Lfrom_mont_nox: |
| 2079 |
___ |
| 2080 |
$code.=<<___; |
| 2081 |
call __bn_sqr8x_reduction |
| 2082 |
call __bn_post4x_internal |
| 2083 |
|
| 2084 |
pxor %xmm0,%xmm0 |
| 2085 |
lea 48(%rsp),%rax |
| 2086 |
mov 40(%rsp),%rsi # restore %rsp |
| 2087 |
jmp .Lfrom_mont_zero |
| 2088 |
|
| 2089 |
.align 32 |
| 2090 |
.Lfrom_mont_zero: |
| 2091 |
movdqa %xmm0,16*0(%rax) |
| 2092 |
movdqa %xmm0,16*1(%rax) |
| 2093 |
movdqa %xmm0,16*2(%rax) |
| 2094 |
movdqa %xmm0,16*3(%rax) |
| 2095 |
lea 16*4(%rax),%rax |
| 2096 |
sub \$32,$num |
| 2097 |
jnz .Lfrom_mont_zero |
| 2098 |
|
| 2099 |
mov \$1,%rax |
| 2100 |
mov -48(%rsi),%r15 |
| 2101 |
mov -40(%rsi),%r14 |
| 2102 |
mov -32(%rsi),%r13 |
| 2103 |
mov -24(%rsi),%r12 |
| 2104 |
mov -16(%rsi),%rbp |
| 2105 |
mov -8(%rsi),%rbx |
| 2106 |
lea (%rsi),%rsp |
| 2107 |
.Lfrom_epilogue: |
| 2108 |
ret |
| 2109 |
.size bn_from_mont8x,.-bn_from_mont8x |
| 2110 |
___ |
| 2111 |
} |
| 2112 |
}}} |
| 2113 |
|
| 2114 |
if ($addx) {{{ |
| 2115 |
my $bp="%rdx"; # restore original value |
| 2116 |
|
| 2117 |
$code.=<<___; |
| 2118 |
.type bn_mulx4x_mont_gather5,\@function,6 |
| 2119 |
.align 32 |
| 2120 |
bn_mulx4x_mont_gather5: |
| 2121 |
.Lmulx4x_enter: |
| 2122 |
mov %rsp,%rax |
| 2123 |
push %rbx |
| 2124 |
push %rbp |
| 2125 |
push %r12 |
| 2126 |
push %r13 |
| 2127 |
push %r14 |
| 2128 |
push %r15 |
| 2129 |
|
| 2130 |
shl \$3,${num}d # convert $num to bytes |
| 2131 |
lea ($num,$num,2),%r10 # 3*$num in bytes |
| 2132 |
neg $num # -$num |
| 2133 |
mov ($n0),$n0 # *n0 |
| 2134 |
|
| 2135 |
############################################################## |
| 2136 |
# Ensure that stack frame doesn't alias with $rptr+3*$num |
| 2137 |
# modulo 4096, which covers ret[num], am[num] and n[num] |
| 2138 |
# (see bn_exp.c). This is done to allow memory disambiguation |
| 2139 |
# logic do its magic. [Extra [num] is allocated in order |
| 2140 |
# to align with bn_power5's frame, which is cleansed after |
| 2141 |
# completing exponentiation. Extra 256 bytes is for power mask |
| 2142 |
# calculated from 7th argument, the index.] |
| 2143 |
# |
| 2144 |
lea -320(%rsp,$num,2),%r11 |
| 2145 |
sub $rp,%r11 |
| 2146 |
and \$4095,%r11 |
| 2147 |
cmp %r11,%r10 |
| 2148 |
jb .Lmulx4xsp_alt |
| 2149 |
sub %r11,%rsp # align with $aptr |
| 2150 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*$num*8+256) |
| 2151 |
jmp .Lmulx4xsp_done |
| 2152 |
|
| 2153 |
.Lmulx4xsp_alt: |
| 2154 |
lea 4096-320(,$num,2),%r10 |
| 2155 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*$num*8+256) |
| 2156 |
sub %r10,%r11 |
| 2157 |
mov \$0,%r10 |
| 2158 |
cmovc %r10,%r11 |
| 2159 |
sub %r11,%rsp |
| 2160 |
.Lmulx4xsp_done: |
| 2161 |
and \$-64,%rsp # ensure alignment |
| 2162 |
############################################################## |
| 2163 |
# Stack layout |
| 2164 |
# +0 -num |
| 2165 |
# +8 off-loaded &b[i] |
| 2166 |
# +16 end of b[num] |
| 2167 |
# +24 inner counter |
| 2168 |
# +32 saved n0 |
| 2169 |
# +40 saved %rsp |
| 2170 |
# +48 |
| 2171 |
# +56 saved rp |
| 2172 |
# +64 tmp[num+1] |
| 2173 |
# |
| 2174 |
mov $n0, 32(%rsp) # save *n0 |
| 2175 |
mov %rax,40(%rsp) # save original %rsp |
| 2176 |
.Lmulx4x_body: |
| 2177 |
call mulx4x_internal |
| 2178 |
|
| 2179 |
mov 40(%rsp),%rsi # restore %rsp |
| 2180 |
mov \$1,%rax |
| 2181 |
|
| 2182 |
mov -48(%rsi),%r15 |
| 2183 |
mov -40(%rsi),%r14 |
| 2184 |
mov -32(%rsi),%r13 |
| 2185 |
mov -24(%rsi),%r12 |
| 2186 |
mov -16(%rsi),%rbp |
| 2187 |
mov -8(%rsi),%rbx |
| 2188 |
lea (%rsi),%rsp |
| 2189 |
.Lmulx4x_epilogue: |
| 2190 |
ret |
| 2191 |
.size bn_mulx4x_mont_gather5,.-bn_mulx4x_mont_gather5 |
| 2192 |
|
| 2193 |
.type mulx4x_internal,\@abi-omnipotent |
| 2194 |
.align 32 |
| 2195 |
mulx4x_internal: |
| 2196 |
mov $num,8(%rsp) # save -$num (it was in bytes) |
| 2197 |
mov $num,%r10 |
| 2198 |
neg $num # restore $num |
| 2199 |
shl \$5,$num |
| 2200 |
neg %r10 # restore $num |
| 2201 |
lea 128($bp,$num),%r13 # end of powers table (+size optimization) |
| 2202 |
shr \$5+5,$num |
| 2203 |
movd `($win64?56:8)`(%rax),%xmm5 # load 7th argument |
| 2204 |
sub \$1,$num |
| 2205 |
lea .Linc(%rip),%rax |
| 2206 |
mov %r13,16+8(%rsp) # end of b[num] |
| 2207 |
mov $num,24+8(%rsp) # inner counter |
| 2208 |
mov $rp, 56+8(%rsp) # save $rp |
| 2209 |
___ |
| 2210 |
my ($aptr, $bptr, $nptr, $tptr, $mi, $bi, $zero, $num)= |
| 2211 |
("%rsi","%rdi","%rcx","%rbx","%r8","%r9","%rbp","%rax"); |
| 2212 |
my $rptr=$bptr; |
| 2213 |
my $STRIDE=2**5*8; # 5 is "window size" |
| 2214 |
my $N=$STRIDE/4; # should match cache line size |
| 2215 |
$code.=<<___; |
| 2216 |
movdqa 0(%rax),%xmm0 # 00000001000000010000000000000000 |
| 2217 |
movdqa 16(%rax),%xmm1 # 00000002000000020000000200000002 |
| 2218 |
lea 88-112(%rsp,%r10),%r10 # place the mask after tp[num+1] (+ICache optimizaton) |
| 2219 |
lea 128($bp),$bptr # size optimization |
| 2220 |
|
| 2221 |
pshufd \$0,%xmm5,%xmm5 # broadcast index |
| 2222 |
movdqa %xmm1,%xmm4 |
| 2223 |
.byte 0x67 |
| 2224 |
movdqa %xmm1,%xmm2 |
| 2225 |
___ |
| 2226 |
######################################################################## |
| 2227 |
# calculate mask by comparing 0..31 to index and save result to stack |
| 2228 |
# |
| 2229 |
$code.=<<___; |
| 2230 |
.byte 0x67 |
| 2231 |
paddd %xmm0,%xmm1 |
| 2232 |
pcmpeqd %xmm5,%xmm0 # compare to 1,0 |
| 2233 |
movdqa %xmm4,%xmm3 |
| 2234 |
___ |
| 2235 |
for($i=0;$i<$STRIDE/16-4;$i+=4) { |
| 2236 |
$code.=<<___; |
| 2237 |
paddd %xmm1,%xmm2 |
| 2238 |
pcmpeqd %xmm5,%xmm1 # compare to 3,2 |
| 2239 |
movdqa %xmm0,`16*($i+0)+112`(%r10) |
| 2240 |
movdqa %xmm4,%xmm0 |
| 2241 |
|
| 2242 |
paddd %xmm2,%xmm3 |
| 2243 |
pcmpeqd %xmm5,%xmm2 # compare to 5,4 |
| 2244 |
movdqa %xmm1,`16*($i+1)+112`(%r10) |
| 2245 |
movdqa %xmm4,%xmm1 |
| 2246 |
|
| 2247 |
paddd %xmm3,%xmm0 |
| 2248 |
pcmpeqd %xmm5,%xmm3 # compare to 7,6 |
| 2249 |
movdqa %xmm2,`16*($i+2)+112`(%r10) |
| 2250 |
movdqa %xmm4,%xmm2 |
| 2251 |
|
| 2252 |
paddd %xmm0,%xmm1 |
| 2253 |
pcmpeqd %xmm5,%xmm0 |
| 2254 |
movdqa %xmm3,`16*($i+3)+112`(%r10) |
| 2255 |
movdqa %xmm4,%xmm3 |
| 2256 |
___ |
| 2257 |
} |
| 2258 |
$code.=<<___; # last iteration can be optimized |
| 2259 |
.byte 0x67 |
| 2260 |
paddd %xmm1,%xmm2 |
| 2261 |
pcmpeqd %xmm5,%xmm1 |
| 2262 |
movdqa %xmm0,`16*($i+0)+112`(%r10) |
| 2263 |
|
| 2264 |
paddd %xmm2,%xmm3 |
| 2265 |
pcmpeqd %xmm5,%xmm2 |
| 2266 |
movdqa %xmm1,`16*($i+1)+112`(%r10) |
| 2267 |
|
| 2268 |
pcmpeqd %xmm5,%xmm3 |
| 2269 |
movdqa %xmm2,`16*($i+2)+112`(%r10) |
| 2270 |
|
| 2271 |
pand `16*($i+0)-128`($bptr),%xmm0 # while it's still in register |
| 2272 |
pand `16*($i+1)-128`($bptr),%xmm1 |
| 2273 |
pand `16*($i+2)-128`($bptr),%xmm2 |
| 2274 |
movdqa %xmm3,`16*($i+3)+112`(%r10) |
| 2275 |
pand `16*($i+3)-128`($bptr),%xmm3 |
| 2276 |
por %xmm2,%xmm0 |
| 2277 |
por %xmm3,%xmm1 |
| 2278 |
___ |
| 2279 |
for($i=0;$i<$STRIDE/16-4;$i+=4) { |
| 2280 |
$code.=<<___; |
| 2281 |
movdqa `16*($i+0)-128`($bptr),%xmm4 |
| 2282 |
movdqa `16*($i+1)-128`($bptr),%xmm5 |
| 2283 |
movdqa `16*($i+2)-128`($bptr),%xmm2 |
| 2284 |
pand `16*($i+0)+112`(%r10),%xmm4 |
| 2285 |
movdqa `16*($i+3)-128`($bptr),%xmm3 |
| 2286 |
pand `16*($i+1)+112`(%r10),%xmm5 |
| 2287 |
por %xmm4,%xmm0 |
| 2288 |
pand `16*($i+2)+112`(%r10),%xmm2 |
| 2289 |
por %xmm5,%xmm1 |
| 2290 |
pand `16*($i+3)+112`(%r10),%xmm3 |
| 2291 |
por %xmm2,%xmm0 |
| 2292 |
por %xmm3,%xmm1 |
| 2293 |
___ |
| 2294 |
} |
| 2295 |
$code.=<<___; |
| 2296 |
pxor %xmm1,%xmm0 |
| 2297 |
pshufd \$0x4e,%xmm0,%xmm1 |
| 2298 |
por %xmm1,%xmm0 |
| 2299 |
lea $STRIDE($bptr),$bptr |
| 2300 |
movq %xmm0,%rdx # bp[0] |
| 2301 |
lea 64+8*4+8(%rsp),$tptr |
| 2302 |
|
| 2303 |
mov %rdx,$bi |
| 2304 |
mulx 0*8($aptr),$mi,%rax # a[0]*b[0] |
| 2305 |
mulx 1*8($aptr),%r11,%r12 # a[1]*b[0] |
| 2306 |
add %rax,%r11 |
| 2307 |
mulx 2*8($aptr),%rax,%r13 # ... |
| 2308 |
adc %rax,%r12 |
| 2309 |
adc \$0,%r13 |
| 2310 |
mulx 3*8($aptr),%rax,%r14 |
| 2311 |
|
| 2312 |
mov $mi,%r15 |
| 2313 |
imulq 32+8(%rsp),$mi # "t[0]"*n0 |
| 2314 |
xor $zero,$zero # cf=0, of=0 |
| 2315 |
mov $mi,%rdx |
| 2316 |
|
| 2317 |
mov $bptr,8+8(%rsp) # off-load &b[i] |
| 2318 |
|
| 2319 |
lea 4*8($aptr),$aptr |
| 2320 |
adcx %rax,%r13 |
| 2321 |
adcx $zero,%r14 # cf=0 |
| 2322 |
|
| 2323 |
mulx 0*8($nptr),%rax,%r10 |
| 2324 |
adcx %rax,%r15 # discarded |
| 2325 |
adox %r11,%r10 |
| 2326 |
mulx 1*8($nptr),%rax,%r11 |
| 2327 |
adcx %rax,%r10 |
| 2328 |
adox %r12,%r11 |
| 2329 |
mulx 2*8($nptr),%rax,%r12 |
| 2330 |
mov 24+8(%rsp),$bptr # counter value |
| 2331 |
mov %r10,-8*4($tptr) |
| 2332 |
adcx %rax,%r11 |
| 2333 |
adox %r13,%r12 |
| 2334 |
mulx 3*8($nptr),%rax,%r15 |
| 2335 |
mov $bi,%rdx |
| 2336 |
mov %r11,-8*3($tptr) |
| 2337 |
adcx %rax,%r12 |
| 2338 |
adox $zero,%r15 # of=0 |
| 2339 |
lea 4*8($nptr),$nptr |
| 2340 |
mov %r12,-8*2($tptr) |
| 2341 |
jmp .Lmulx4x_1st |
| 2342 |
|
| 2343 |
.align 32 |
| 2344 |
.Lmulx4x_1st: |
| 2345 |
adcx $zero,%r15 # cf=0, modulo-scheduled |
| 2346 |
mulx 0*8($aptr),%r10,%rax # a[4]*b[0] |
| 2347 |
adcx %r14,%r10 |
| 2348 |
mulx 1*8($aptr),%r11,%r14 # a[5]*b[0] |
| 2349 |
adcx %rax,%r11 |
| 2350 |
mulx 2*8($aptr),%r12,%rax # ... |
| 2351 |
adcx %r14,%r12 |
| 2352 |
mulx 3*8($aptr),%r13,%r14 |
| 2353 |
.byte 0x67,0x67 |
| 2354 |
mov $mi,%rdx |
| 2355 |
adcx %rax,%r13 |
| 2356 |
adcx $zero,%r14 # cf=0 |
| 2357 |
lea 4*8($aptr),$aptr |
| 2358 |
lea 4*8($tptr),$tptr |
| 2359 |
|
| 2360 |
adox %r15,%r10 |
| 2361 |
mulx 0*8($nptr),%rax,%r15 |
| 2362 |
adcx %rax,%r10 |
| 2363 |
adox %r15,%r11 |
| 2364 |
mulx 1*8($nptr),%rax,%r15 |
| 2365 |
adcx %rax,%r11 |
| 2366 |
adox %r15,%r12 |
| 2367 |
mulx 2*8($nptr),%rax,%r15 |
| 2368 |
mov %r10,-5*8($tptr) |
| 2369 |
adcx %rax,%r12 |
| 2370 |
mov %r11,-4*8($tptr) |
| 2371 |
adox %r15,%r13 |
| 2372 |
mulx 3*8($nptr),%rax,%r15 |
| 2373 |
mov $bi,%rdx |
| 2374 |
mov %r12,-3*8($tptr) |
| 2375 |
adcx %rax,%r13 |
| 2376 |
adox $zero,%r15 |
| 2377 |
lea 4*8($nptr),$nptr |
| 2378 |
mov %r13,-2*8($tptr) |
| 2379 |
|
| 2380 |
dec $bptr # of=0, pass cf |
| 2381 |
jnz .Lmulx4x_1st |
| 2382 |
|
| 2383 |
mov 8(%rsp),$num # load -num |
| 2384 |
adc $zero,%r15 # modulo-scheduled |
| 2385 |
lea ($aptr,$num),$aptr # rewind $aptr |
| 2386 |
add %r15,%r14 |
| 2387 |
mov 8+8(%rsp),$bptr # re-load &b[i] |
| 2388 |
adc $zero,$zero # top-most carry |
| 2389 |
mov %r14,-1*8($tptr) |
| 2390 |
jmp .Lmulx4x_outer |
| 2391 |
|
| 2392 |
.align 32 |
| 2393 |
.Lmulx4x_outer: |
| 2394 |
lea 16-256($tptr),%r10 # where 256-byte mask is (+density control) |
| 2395 |
pxor %xmm4,%xmm4 |
| 2396 |
.byte 0x67,0x67 |
| 2397 |
pxor %xmm5,%xmm5 |
| 2398 |
___ |
| 2399 |
for($i=0;$i<$STRIDE/16;$i+=4) { |
| 2400 |
$code.=<<___; |
| 2401 |
movdqa `16*($i+0)-128`($bptr),%xmm0 |
| 2402 |
movdqa `16*($i+1)-128`($bptr),%xmm1 |
| 2403 |
movdqa `16*($i+2)-128`($bptr),%xmm2 |
| 2404 |
pand `16*($i+0)+256`(%r10),%xmm0 |
| 2405 |
movdqa `16*($i+3)-128`($bptr),%xmm3 |
| 2406 |
pand `16*($i+1)+256`(%r10),%xmm1 |
| 2407 |
por %xmm0,%xmm4 |
| 2408 |
pand `16*($i+2)+256`(%r10),%xmm2 |
| 2409 |
por %xmm1,%xmm5 |
| 2410 |
pand `16*($i+3)+256`(%r10),%xmm3 |
| 2411 |
por %xmm2,%xmm4 |
| 2412 |
por %xmm3,%xmm5 |
| 2413 |
___ |
| 2414 |
} |
| 2415 |
$code.=<<___; |
| 2416 |
por %xmm5,%xmm4 |
| 2417 |
pshufd \$0x4e,%xmm4,%xmm0 |
| 2418 |
por %xmm4,%xmm0 |
| 2419 |
lea $STRIDE($bptr),$bptr |
| 2420 |
movq %xmm0,%rdx # m0=bp[i] |
| 2421 |
|
| 2422 |
mov $zero,($tptr) # save top-most carry |
| 2423 |
lea 4*8($tptr,$num),$tptr # rewind $tptr |
| 2424 |
mulx 0*8($aptr),$mi,%r11 # a[0]*b[i] |
| 2425 |
xor $zero,$zero # cf=0, of=0 |
| 2426 |
mov %rdx,$bi |
| 2427 |
mulx 1*8($aptr),%r14,%r12 # a[1]*b[i] |
| 2428 |
adox -4*8($tptr),$mi # +t[0] |
| 2429 |
adcx %r14,%r11 |
| 2430 |
mulx 2*8($aptr),%r15,%r13 # ... |
| 2431 |
adox -3*8($tptr),%r11 |
| 2432 |
adcx %r15,%r12 |
| 2433 |
mulx 3*8($aptr),%rdx,%r14 |
| 2434 |
adox -2*8($tptr),%r12 |
| 2435 |
adcx %rdx,%r13 |
| 2436 |
lea ($nptr,$num),$nptr # rewind $nptr |
| 2437 |
lea 4*8($aptr),$aptr |
| 2438 |
adox -1*8($tptr),%r13 |
| 2439 |
adcx $zero,%r14 |
| 2440 |
adox $zero,%r14 |
| 2441 |
|
| 2442 |
mov $mi,%r15 |
| 2443 |
imulq 32+8(%rsp),$mi # "t[0]"*n0 |
| 2444 |
|
| 2445 |
mov $mi,%rdx |
| 2446 |
xor $zero,$zero # cf=0, of=0 |
| 2447 |
mov $bptr,8+8(%rsp) # off-load &b[i] |
| 2448 |
|
| 2449 |
mulx 0*8($nptr),%rax,%r10 |
| 2450 |
adcx %rax,%r15 # discarded |
| 2451 |
adox %r11,%r10 |
| 2452 |
mulx 1*8($nptr),%rax,%r11 |
| 2453 |
adcx %rax,%r10 |
| 2454 |
adox %r12,%r11 |
| 2455 |
mulx 2*8($nptr),%rax,%r12 |
| 2456 |
adcx %rax,%r11 |
| 2457 |
adox %r13,%r12 |
| 2458 |
mulx 3*8($nptr),%rax,%r15 |
| 2459 |
mov $bi,%rdx |
| 2460 |
mov 24+8(%rsp),$bptr # counter value |
| 2461 |
mov %r10,-8*4($tptr) |
| 2462 |
adcx %rax,%r12 |
| 2463 |
mov %r11,-8*3($tptr) |
| 2464 |
adox $zero,%r15 # of=0 |
| 2465 |
mov %r12,-8*2($tptr) |
| 2466 |
lea 4*8($nptr),$nptr |
| 2467 |
jmp .Lmulx4x_inner |
| 2468 |
|
| 2469 |
.align 32 |
| 2470 |
.Lmulx4x_inner: |
| 2471 |
mulx 0*8($aptr),%r10,%rax # a[4]*b[i] |
| 2472 |
adcx $zero,%r15 # cf=0, modulo-scheduled |
| 2473 |
adox %r14,%r10 |
| 2474 |
mulx 1*8($aptr),%r11,%r14 # a[5]*b[i] |
| 2475 |
adcx 0*8($tptr),%r10 |
| 2476 |
adox %rax,%r11 |
| 2477 |
mulx 2*8($aptr),%r12,%rax # ... |
| 2478 |
adcx 1*8($tptr),%r11 |
| 2479 |
adox %r14,%r12 |
| 2480 |
mulx 3*8($aptr),%r13,%r14 |
| 2481 |
mov $mi,%rdx |
| 2482 |
adcx 2*8($tptr),%r12 |
| 2483 |
adox %rax,%r13 |
| 2484 |
adcx 3*8($tptr),%r13 |
| 2485 |
adox $zero,%r14 # of=0 |
| 2486 |
lea 4*8($aptr),$aptr |
| 2487 |
lea 4*8($tptr),$tptr |
| 2488 |
adcx $zero,%r14 # cf=0 |
| 2489 |
|
| 2490 |
adox %r15,%r10 |
| 2491 |
mulx 0*8($nptr),%rax,%r15 |
| 2492 |
adcx %rax,%r10 |
| 2493 |
adox %r15,%r11 |
| 2494 |
mulx 1*8($nptr),%rax,%r15 |
| 2495 |
adcx %rax,%r11 |
| 2496 |
adox %r15,%r12 |
| 2497 |
mulx 2*8($nptr),%rax,%r15 |
| 2498 |
mov %r10,-5*8($tptr) |
| 2499 |
adcx %rax,%r12 |
| 2500 |
adox %r15,%r13 |
| 2501 |
mov %r11,-4*8($tptr) |
| 2502 |
mulx 3*8($nptr),%rax,%r15 |
| 2503 |
mov $bi,%rdx |
| 2504 |
lea 4*8($nptr),$nptr |
| 2505 |
mov %r12,-3*8($tptr) |
| 2506 |
adcx %rax,%r13 |
| 2507 |
adox $zero,%r15 |
| 2508 |
mov %r13,-2*8($tptr) |
| 2509 |
|
| 2510 |
dec $bptr # of=0, pass cf |
| 2511 |
jnz .Lmulx4x_inner |
| 2512 |
|
| 2513 |
mov 0+8(%rsp),$num # load -num |
| 2514 |
adc $zero,%r15 # modulo-scheduled |
| 2515 |
sub 0*8($tptr),$bptr # pull top-most carry to %cf |
| 2516 |
mov 8+8(%rsp),$bptr # re-load &b[i] |
| 2517 |
mov 16+8(%rsp),%r10 |
| 2518 |
adc %r15,%r14 |
| 2519 |
lea ($aptr,$num),$aptr # rewind $aptr |
| 2520 |
adc $zero,$zero # top-most carry |
| 2521 |
mov %r14,-1*8($tptr) |
| 2522 |
|
| 2523 |
cmp %r10,$bptr |
| 2524 |
jb .Lmulx4x_outer |
| 2525 |
|
| 2526 |
mov -8($nptr),%r10 |
| 2527 |
mov $zero,%r8 |
| 2528 |
mov ($nptr,$num),%r12 |
| 2529 |
lea ($nptr,$num),%rbp # rewind $nptr |
| 2530 |
mov $num,%rcx |
| 2531 |
lea ($tptr,$num),%rdi # rewind $tptr |
| 2532 |
xor %eax,%eax |
| 2533 |
xor %r15,%r15 |
| 2534 |
sub %r14,%r10 # compare top-most words |
| 2535 |
adc %r15,%r15 |
| 2536 |
or %r15,%r8 |
| 2537 |
sar \$3+2,%rcx |
| 2538 |
sub %r8,%rax # %rax=-%r8 |
| 2539 |
mov 56+8(%rsp),%rdx # restore rp |
| 2540 |
dec %r12 # so that after 'not' we get -n[0] |
| 2541 |
mov 8*1(%rbp),%r13 |
| 2542 |
xor %r8,%r8 |
| 2543 |
mov 8*2(%rbp),%r14 |
| 2544 |
mov 8*3(%rbp),%r15 |
| 2545 |
jmp .Lsqrx4x_sub_entry # common post-condition |
| 2546 |
.size mulx4x_internal,.-mulx4x_internal |
| 2547 |
___ |
| 2548 |
}{ |
| 2549 |
###################################################################### |
| 2550 |
# void bn_power5( |
| 2551 |
my $rptr="%rdi"; # BN_ULONG *rptr, |
| 2552 |
my $aptr="%rsi"; # const BN_ULONG *aptr, |
| 2553 |
my $bptr="%rdx"; # const void *table, |
| 2554 |
my $nptr="%rcx"; # const BN_ULONG *nptr, |
| 2555 |
my $n0 ="%r8"; # const BN_ULONG *n0); |
| 2556 |
my $num ="%r9"; # int num, has to be divisible by 8 |
| 2557 |
# int pwr); |
| 2558 |
|
| 2559 |
my ($i,$j,$tptr)=("%rbp","%rcx",$rptr); |
| 2560 |
my @A0=("%r10","%r11"); |
| 2561 |
my @A1=("%r12","%r13"); |
| 2562 |
my ($a0,$a1,$ai)=("%r14","%r15","%rbx"); |
| 2563 |
|
| 2564 |
$code.=<<___; |
| 2565 |
.type bn_powerx5,\@function,6 |
| 2566 |
.align 32 |
| 2567 |
bn_powerx5: |
| 2568 |
.Lpowerx5_enter: |
| 2569 |
mov %rsp,%rax |
| 2570 |
push %rbx |
| 2571 |
push %rbp |
| 2572 |
push %r12 |
| 2573 |
push %r13 |
| 2574 |
push %r14 |
| 2575 |
push %r15 |
| 2576 |
|
| 2577 |
shl \$3,${num}d # convert $num to bytes |
| 2578 |
lea ($num,$num,2),%r10 # 3*$num in bytes |
| 2579 |
neg $num |
| 2580 |
mov ($n0),$n0 # *n0 |
| 2581 |
|
| 2582 |
############################################################## |
| 2583 |
# Ensure that stack frame doesn't alias with $rptr+3*$num |
| 2584 |
# modulo 4096, which covers ret[num], am[num] and n[num] |
| 2585 |
# (see bn_exp.c). This is done to allow memory disambiguation |
| 2586 |
# logic do its magic. [Extra 256 bytes is for power mask |
| 2587 |
# calculated from 7th argument, the index.] |
| 2588 |
# |
| 2589 |
lea -320(%rsp,$num,2),%r11 |
| 2590 |
sub $rptr,%r11 |
| 2591 |
and \$4095,%r11 |
| 2592 |
cmp %r11,%r10 |
| 2593 |
jb .Lpwrx_sp_alt |
| 2594 |
sub %r11,%rsp # align with $aptr |
| 2595 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*$num*8+256) |
| 2596 |
jmp .Lpwrx_sp_done |
| 2597 |
|
| 2598 |
.align 32 |
| 2599 |
.Lpwrx_sp_alt: |
| 2600 |
lea 4096-320(,$num,2),%r10 |
| 2601 |
lea -320(%rsp,$num,2),%rsp # alloca(frame+2*$num*8+256) |
| 2602 |
sub %r10,%r11 |
| 2603 |
mov \$0,%r10 |
| 2604 |
cmovc %r10,%r11 |
| 2605 |
sub %r11,%rsp |
| 2606 |
.Lpwrx_sp_done: |
| 2607 |
and \$-64,%rsp |
| 2608 |
mov $num,%r10 |
| 2609 |
neg $num |
| 2610 |
|
| 2611 |
############################################################## |
| 2612 |
# Stack layout |
| 2613 |
# |
| 2614 |
# +0 saved $num, used in reduction section |
| 2615 |
# +8 &t[2*$num], used in reduction section |
| 2616 |
# +16 intermediate carry bit |
| 2617 |
# +24 top-most carry bit, used in reduction section |
| 2618 |
# +32 saved *n0 |
| 2619 |
# +40 saved %rsp |
| 2620 |
# +48 t[2*$num] |
| 2621 |
# |
| 2622 |
pxor %xmm0,%xmm0 |
| 2623 |
movq $rptr,%xmm1 # save $rptr |
| 2624 |
movq $nptr,%xmm2 # save $nptr |
| 2625 |
movq %r10, %xmm3 # -$num |
| 2626 |
movq $bptr,%xmm4 |
| 2627 |
mov $n0, 32(%rsp) |
| 2628 |
mov %rax, 40(%rsp) # save original %rsp |
| 2629 |
.Lpowerx5_body: |
| 2630 |
|
| 2631 |
call __bn_sqrx8x_internal |
| 2632 |
call __bn_postx4x_internal |
| 2633 |
call __bn_sqrx8x_internal |
| 2634 |
call __bn_postx4x_internal |
| 2635 |
call __bn_sqrx8x_internal |
| 2636 |
call __bn_postx4x_internal |
| 2637 |
call __bn_sqrx8x_internal |
| 2638 |
call __bn_postx4x_internal |
| 2639 |
call __bn_sqrx8x_internal |
| 2640 |
call __bn_postx4x_internal |
| 2641 |
|
| 2642 |
mov %r10,$num # -num |
| 2643 |
mov $aptr,$rptr |
| 2644 |
movq %xmm2,$nptr |
| 2645 |
movq %xmm4,$bptr |
| 2646 |
mov 40(%rsp),%rax |
| 2647 |
|
| 2648 |
call mulx4x_internal |
| 2649 |
|
| 2650 |
mov 40(%rsp),%rsi # restore %rsp |
| 2651 |
mov \$1,%rax |
| 2652 |
|
| 2653 |
mov -48(%rsi),%r15 |
| 2654 |
mov -40(%rsi),%r14 |
| 2655 |
mov -32(%rsi),%r13 |
| 2656 |
mov -24(%rsi),%r12 |
| 2657 |
mov -16(%rsi),%rbp |
| 2658 |
mov -8(%rsi),%rbx |
| 2659 |
lea (%rsi),%rsp |
| 2660 |
.Lpowerx5_epilogue: |
| 2661 |
ret |
| 2662 |
.size bn_powerx5,.-bn_powerx5 |
| 2663 |
|
| 2664 |
.globl bn_sqrx8x_internal |
| 2665 |
.hidden bn_sqrx8x_internal |
| 2666 |
.type bn_sqrx8x_internal,\@abi-omnipotent |
| 2667 |
.align 32 |
| 2668 |
bn_sqrx8x_internal: |
| 2669 |
__bn_sqrx8x_internal: |
| 2670 |
################################################################## |
| 2671 |
# Squaring part: |
| 2672 |
# |
| 2673 |
# a) multiply-n-add everything but a[i]*a[i]; |
| 2674 |
# b) shift result of a) by 1 to the left and accumulate |
| 2675 |
# a[i]*a[i] products; |
| 2676 |
# |
| 2677 |
################################################################## |
| 2678 |
# a[7]a[7]a[6]a[6]a[5]a[5]a[4]a[4]a[3]a[3]a[2]a[2]a[1]a[1]a[0]a[0] |
| 2679 |
# a[1]a[0] |
| 2680 |
# a[2]a[0] |
| 2681 |
# a[3]a[0] |
| 2682 |
# a[2]a[1] |
| 2683 |
# a[3]a[1] |
| 2684 |
# a[3]a[2] |
| 2685 |
# |
| 2686 |
# a[4]a[0] |
| 2687 |
# a[5]a[0] |
| 2688 |
# a[6]a[0] |
| 2689 |
# a[7]a[0] |
| 2690 |
# a[4]a[1] |
| 2691 |
# a[5]a[1] |
| 2692 |
# a[6]a[1] |
| 2693 |
# a[7]a[1] |
| 2694 |
# a[4]a[2] |
| 2695 |
# a[5]a[2] |
| 2696 |
# a[6]a[2] |
| 2697 |
# a[7]a[2] |
| 2698 |
# a[4]a[3] |
| 2699 |
# a[5]a[3] |
| 2700 |
# a[6]a[3] |
| 2701 |
# a[7]a[3] |
| 2702 |
# |
| 2703 |
# a[5]a[4] |
| 2704 |
# a[6]a[4] |
| 2705 |
# a[7]a[4] |
| 2706 |
# a[6]a[5] |
| 2707 |
# a[7]a[5] |
| 2708 |
# a[7]a[6] |
| 2709 |
# a[7]a[7]a[6]a[6]a[5]a[5]a[4]a[4]a[3]a[3]a[2]a[2]a[1]a[1]a[0]a[0] |
| 2710 |
___ |
| 2711 |
{ |
| 2712 |
my ($zero,$carry)=("%rbp","%rcx"); |
| 2713 |
my $aaptr=$zero; |
| 2714 |
$code.=<<___; |
| 2715 |
lea 48+8(%rsp),$tptr |
| 2716 |
lea ($aptr,$num),$aaptr |
| 2717 |
mov $num,0+8(%rsp) # save $num |
| 2718 |
mov $aaptr,8+8(%rsp) # save end of $aptr |
| 2719 |
jmp .Lsqr8x_zero_start |
| 2720 |
|
| 2721 |
.align 32 |
| 2722 |
.byte 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00 |
| 2723 |
.Lsqrx8x_zero: |
| 2724 |
.byte 0x3e |
| 2725 |
movdqa %xmm0,0*8($tptr) |
| 2726 |
movdqa %xmm0,2*8($tptr) |
| 2727 |
movdqa %xmm0,4*8($tptr) |
| 2728 |
movdqa %xmm0,6*8($tptr) |
| 2729 |
.Lsqr8x_zero_start: # aligned at 32 |
| 2730 |
movdqa %xmm0,8*8($tptr) |
| 2731 |
movdqa %xmm0,10*8($tptr) |
| 2732 |
movdqa %xmm0,12*8($tptr) |
| 2733 |
movdqa %xmm0,14*8($tptr) |
| 2734 |
lea 16*8($tptr),$tptr |
| 2735 |
sub \$64,$num |
| 2736 |
jnz .Lsqrx8x_zero |
| 2737 |
|
| 2738 |
mov 0*8($aptr),%rdx # a[0], modulo-scheduled |
| 2739 |
#xor %r9,%r9 # t[1], ex-$num, zero already |
| 2740 |
xor %r10,%r10 |
| 2741 |
xor %r11,%r11 |
| 2742 |
xor %r12,%r12 |
| 2743 |
xor %r13,%r13 |
| 2744 |
xor %r14,%r14 |
| 2745 |
xor %r15,%r15 |
| 2746 |
lea 48+8(%rsp),$tptr |
| 2747 |
xor $zero,$zero # cf=0, cf=0 |
| 2748 |
jmp .Lsqrx8x_outer_loop |
| 2749 |
|
| 2750 |
.align 32 |
| 2751 |
.Lsqrx8x_outer_loop: |
| 2752 |
mulx 1*8($aptr),%r8,%rax # a[1]*a[0] |
| 2753 |
adcx %r9,%r8 # a[1]*a[0]+=t[1] |
| 2754 |
adox %rax,%r10 |
| 2755 |
mulx 2*8($aptr),%r9,%rax # a[2]*a[0] |
| 2756 |
adcx %r10,%r9 |
| 2757 |
adox %rax,%r11 |
| 2758 |
.byte 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00 # mulx 3*8($aptr),%r10,%rax # ... |
| 2759 |
adcx %r11,%r10 |
| 2760 |
adox %rax,%r12 |
| 2761 |
.byte 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00 # mulx 4*8($aptr),%r11,%rax |
| 2762 |
adcx %r12,%r11 |
| 2763 |
adox %rax,%r13 |
| 2764 |
mulx 5*8($aptr),%r12,%rax |
| 2765 |
adcx %r13,%r12 |
| 2766 |
adox %rax,%r14 |
| 2767 |
mulx 6*8($aptr),%r13,%rax |
| 2768 |
adcx %r14,%r13 |
| 2769 |
adox %r15,%rax |
| 2770 |
mulx 7*8($aptr),%r14,%r15 |
| 2771 |
mov 1*8($aptr),%rdx # a[1] |
| 2772 |
adcx %rax,%r14 |
| 2773 |
adox $zero,%r15 |
| 2774 |
adc 8*8($tptr),%r15 |
| 2775 |
mov %r8,1*8($tptr) # t[1] |
| 2776 |
mov %r9,2*8($tptr) # t[2] |
| 2777 |
sbb $carry,$carry # mov %cf,$carry |
| 2778 |
xor $zero,$zero # cf=0, of=0 |
| 2779 |
|
| 2780 |
|
| 2781 |
mulx 2*8($aptr),%r8,%rbx # a[2]*a[1] |
| 2782 |
mulx 3*8($aptr),%r9,%rax # a[3]*a[1] |
| 2783 |
adcx %r10,%r8 |
| 2784 |
adox %rbx,%r9 |
| 2785 |
mulx 4*8($aptr),%r10,%rbx # ... |
| 2786 |
adcx %r11,%r9 |
| 2787 |
adox %rax,%r10 |
| 2788 |
.byte 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00 # mulx 5*8($aptr),%r11,%rax |
| 2789 |
adcx %r12,%r10 |
| 2790 |
adox %rbx,%r11 |
| 2791 |
.byte 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00 # mulx 6*8($aptr),%r12,%rbx |
| 2792 |
adcx %r13,%r11 |
| 2793 |
adox %r14,%r12 |
| 2794 |
.byte 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00 # mulx 7*8($aptr),%r13,%r14 |
| 2795 |
mov 2*8($aptr),%rdx # a[2] |
| 2796 |
adcx %rax,%r12 |
| 2797 |
adox %rbx,%r13 |
| 2798 |
adcx %r15,%r13 |
| 2799 |
adox $zero,%r14 # of=0 |
| 2800 |
adcx $zero,%r14 # cf=0 |
| 2801 |
|
| 2802 |
mov %r8,3*8($tptr) # t[3] |
| 2803 |
mov %r9,4*8($tptr) # t[4] |
| 2804 |
|
| 2805 |
mulx 3*8($aptr),%r8,%rbx # a[3]*a[2] |
| 2806 |
mulx 4*8($aptr),%r9,%rax # a[4]*a[2] |
| 2807 |
adcx %r10,%r8 |
| 2808 |
adox %rbx,%r9 |
| 2809 |
mulx 5*8($aptr),%r10,%rbx # ... |
| 2810 |
adcx %r11,%r9 |
| 2811 |
adox %rax,%r10 |
| 2812 |
.byte 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00 # mulx 6*8($aptr),%r11,%rax |
| 2813 |
adcx %r12,%r10 |
| 2814 |
adox %r13,%r11 |
| 2815 |
.byte 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00 # mulx 7*8($aptr),%r12,%r13 |
| 2816 |
.byte 0x3e |
| 2817 |
mov 3*8($aptr),%rdx # a[3] |
| 2818 |
adcx %rbx,%r11 |
| 2819 |
adox %rax,%r12 |
| 2820 |
adcx %r14,%r12 |
| 2821 |
mov %r8,5*8($tptr) # t[5] |
| 2822 |
mov %r9,6*8($tptr) # t[6] |
| 2823 |
mulx 4*8($aptr),%r8,%rax # a[4]*a[3] |
| 2824 |
adox $zero,%r13 # of=0 |
| 2825 |
adcx $zero,%r13 # cf=0 |
| 2826 |
|
| 2827 |
mulx 5*8($aptr),%r9,%rbx # a[5]*a[3] |
| 2828 |
adcx %r10,%r8 |
| 2829 |
adox %rax,%r9 |
| 2830 |
mulx 6*8($aptr),%r10,%rax # ... |
| 2831 |
adcx %r11,%r9 |
| 2832 |
adox %r12,%r10 |
| 2833 |
mulx 7*8($aptr),%r11,%r12 |
| 2834 |
mov 4*8($aptr),%rdx # a[4] |
| 2835 |
mov 5*8($aptr),%r14 # a[5] |
| 2836 |
adcx %rbx,%r10 |
| 2837 |
adox %rax,%r11 |
| 2838 |
mov 6*8($aptr),%r15 # a[6] |
| 2839 |
adcx %r13,%r11 |
| 2840 |
adox $zero,%r12 # of=0 |
| 2841 |
adcx $zero,%r12 # cf=0 |
| 2842 |
|
| 2843 |
mov %r8,7*8($tptr) # t[7] |
| 2844 |
mov %r9,8*8($tptr) # t[8] |
| 2845 |
|
| 2846 |
mulx %r14,%r9,%rax # a[5]*a[4] |
| 2847 |
mov 7*8($aptr),%r8 # a[7] |
| 2848 |
adcx %r10,%r9 |
| 2849 |
mulx %r15,%r10,%rbx # a[6]*a[4] |
| 2850 |
adox %rax,%r10 |
| 2851 |
adcx %r11,%r10 |
| 2852 |
mulx %r8,%r11,%rax # a[7]*a[4] |
| 2853 |
mov %r14,%rdx # a[5] |
| 2854 |
adox %rbx,%r11 |
| 2855 |
adcx %r12,%r11 |
| 2856 |
#adox $zero,%rax # of=0 |
| 2857 |
adcx $zero,%rax # cf=0 |
| 2858 |
|
| 2859 |
mulx %r15,%r14,%rbx # a[6]*a[5] |
| 2860 |
mulx %r8,%r12,%r13 # a[7]*a[5] |
| 2861 |
mov %r15,%rdx # a[6] |
| 2862 |
lea 8*8($aptr),$aptr |
| 2863 |
adcx %r14,%r11 |
| 2864 |
adox %rbx,%r12 |
| 2865 |
adcx %rax,%r12 |
| 2866 |
adox $zero,%r13 |
| 2867 |
|
| 2868 |
.byte 0x67,0x67 |
| 2869 |
mulx %r8,%r8,%r14 # a[7]*a[6] |
| 2870 |
adcx %r8,%r13 |
| 2871 |
adcx $zero,%r14 |
| 2872 |
|
| 2873 |
cmp 8+8(%rsp),$aptr |
| 2874 |
je .Lsqrx8x_outer_break |
| 2875 |
|
| 2876 |
neg $carry # mov $carry,%cf |
| 2877 |
mov \$-8,%rcx |
| 2878 |
mov $zero,%r15 |
| 2879 |
mov 8*8($tptr),%r8 |
| 2880 |
adcx 9*8($tptr),%r9 # +=t[9] |
| 2881 |
adcx 10*8($tptr),%r10 # ... |
| 2882 |
adcx 11*8($tptr),%r11 |
| 2883 |
adc 12*8($tptr),%r12 |
| 2884 |
adc 13*8($tptr),%r13 |
| 2885 |
adc 14*8($tptr),%r14 |
| 2886 |
adc 15*8($tptr),%r15 |
| 2887 |
lea ($aptr),$aaptr |
| 2888 |
lea 2*64($tptr),$tptr |
| 2889 |
sbb %rax,%rax # mov %cf,$carry |
| 2890 |
|
| 2891 |
mov -64($aptr),%rdx # a[0] |
| 2892 |
mov %rax,16+8(%rsp) # offload $carry |
| 2893 |
mov $tptr,24+8(%rsp) |
| 2894 |
|
| 2895 |
#lea 8*8($tptr),$tptr # see 2*8*8($tptr) above |
| 2896 |
xor %eax,%eax # cf=0, of=0 |
| 2897 |
jmp .Lsqrx8x_loop |
| 2898 |
|
| 2899 |
.align 32 |
| 2900 |
.Lsqrx8x_loop: |
| 2901 |
mov %r8,%rbx |
| 2902 |
mulx 0*8($aaptr),%rax,%r8 # a[8]*a[i] |
| 2903 |
adcx %rax,%rbx # +=t[8] |
| 2904 |
adox %r9,%r8 |
| 2905 |
|
| 2906 |
mulx 1*8($aaptr),%rax,%r9 # ... |
| 2907 |
adcx %rax,%r8 |
| 2908 |
adox %r10,%r9 |
| 2909 |
|
| 2910 |
mulx 2*8($aaptr),%rax,%r10 |
| 2911 |
adcx %rax,%r9 |
| 2912 |
adox %r11,%r10 |
| 2913 |
|
| 2914 |
mulx 3*8($aaptr),%rax,%r11 |
| 2915 |
adcx %rax,%r10 |
| 2916 |
adox %r12,%r11 |
| 2917 |
|
| 2918 |
.byte 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 # mulx 4*8($aaptr),%rax,%r12 |
| 2919 |
adcx %rax,%r11 |
| 2920 |
adox %r13,%r12 |
| 2921 |
|
| 2922 |
mulx 5*8($aaptr),%rax,%r13 |
| 2923 |
adcx %rax,%r12 |
| 2924 |
adox %r14,%r13 |
| 2925 |
|
| 2926 |
mulx 6*8($aaptr),%rax,%r14 |
| 2927 |
mov %rbx,($tptr,%rcx,8) # store t[8+i] |
| 2928 |
mov \$0,%ebx |
| 2929 |
adcx %rax,%r13 |
| 2930 |
adox %r15,%r14 |
| 2931 |
|
| 2932 |
.byte 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00 # mulx 7*8($aaptr),%rax,%r15 |
| 2933 |
mov 8($aptr,%rcx,8),%rdx # a[i] |
| 2934 |
adcx %rax,%r14 |
| 2935 |
adox %rbx,%r15 # %rbx is 0, of=0 |
| 2936 |
adcx %rbx,%r15 # cf=0 |
| 2937 |
|
| 2938 |
.byte 0x67 |
| 2939 |
inc %rcx # of=0 |
| 2940 |
jnz .Lsqrx8x_loop |
| 2941 |
|
| 2942 |
lea 8*8($aaptr),$aaptr |
| 2943 |
mov \$-8,%rcx |
| 2944 |
cmp 8+8(%rsp),$aaptr # done? |
| 2945 |
je .Lsqrx8x_break |
| 2946 |
|
| 2947 |
sub 16+8(%rsp),%rbx # mov 16(%rsp),%cf |
| 2948 |
.byte 0x66 |
| 2949 |
mov -64($aptr),%rdx |
| 2950 |
adcx 0*8($tptr),%r8 |
| 2951 |
adcx 1*8($tptr),%r9 |
| 2952 |
adc 2*8($tptr),%r10 |
| 2953 |
adc 3*8($tptr),%r11 |
| 2954 |
adc 4*8($tptr),%r12 |
| 2955 |
adc 5*8($tptr),%r13 |
| 2956 |
adc 6*8($tptr),%r14 |
| 2957 |
adc 7*8($tptr),%r15 |
| 2958 |
lea 8*8($tptr),$tptr |
| 2959 |
.byte 0x67 |
| 2960 |
sbb %rax,%rax # mov %cf,%rax |
| 2961 |
xor %ebx,%ebx # cf=0, of=0 |
| 2962 |
mov %rax,16+8(%rsp) # offload carry |
| 2963 |
jmp .Lsqrx8x_loop |
| 2964 |
|
| 2965 |
.align 32 |
| 2966 |
.Lsqrx8x_break: |
| 2967 |
sub 16+8(%rsp),%r8 # consume last carry |
| 2968 |
mov 24+8(%rsp),$carry # initial $tptr, borrow $carry |
| 2969 |
mov 0*8($aptr),%rdx # a[8], modulo-scheduled |
| 2970 |
xor %ebp,%ebp # xor $zero,$zero |
| 2971 |
mov %r8,0*8($tptr) |
| 2972 |
cmp $carry,$tptr # cf=0, of=0 |
| 2973 |
je .Lsqrx8x_outer_loop |
| 2974 |
|
| 2975 |
mov %r9,1*8($tptr) |
| 2976 |
mov 1*8($carry),%r9 |
| 2977 |
mov %r10,2*8($tptr) |
| 2978 |
mov 2*8($carry),%r10 |
| 2979 |
mov %r11,3*8($tptr) |
| 2980 |
mov 3*8($carry),%r11 |
| 2981 |
mov %r12,4*8($tptr) |
| 2982 |
mov 4*8($carry),%r12 |
| 2983 |
mov %r13,5*8($tptr) |
| 2984 |
mov 5*8($carry),%r13 |
| 2985 |
mov %r14,6*8($tptr) |
| 2986 |
mov 6*8($carry),%r14 |
| 2987 |
mov %r15,7*8($tptr) |
| 2988 |
mov 7*8($carry),%r15 |
| 2989 |
mov $carry,$tptr |
| 2990 |
jmp .Lsqrx8x_outer_loop |
| 2991 |
|
| 2992 |
.align 32 |
| 2993 |
.Lsqrx8x_outer_break: |
| 2994 |
mov %r9,9*8($tptr) # t[9] |
| 2995 |
movq %xmm3,%rcx # -$num |
| 2996 |
mov %r10,10*8($tptr) # ... |
| 2997 |
mov %r11,11*8($tptr) |
| 2998 |
mov %r12,12*8($tptr) |
| 2999 |
mov %r13,13*8($tptr) |
| 3000 |
mov %r14,14*8($tptr) |
| 3001 |
___ |
| 3002 |
}{ |
| 3003 |
my $i="%rcx"; |
| 3004 |
$code.=<<___; |
| 3005 |
lea 48+8(%rsp),$tptr |
| 3006 |
mov ($aptr,$i),%rdx # a[0] |
| 3007 |
|
| 3008 |
mov 8($tptr),$A0[1] # t[1] |
| 3009 |
xor $A0[0],$A0[0] # t[0], of=0, cf=0 |
| 3010 |
mov 0+8(%rsp),$num # restore $num |
| 3011 |
adox $A0[1],$A0[1] |
| 3012 |
mov 16($tptr),$A1[0] # t[2] # prefetch |
| 3013 |
mov 24($tptr),$A1[1] # t[3] # prefetch |
| 3014 |
#jmp .Lsqrx4x_shift_n_add # happens to be aligned |
| 3015 |
|
| 3016 |
.align 32 |
| 3017 |
.Lsqrx4x_shift_n_add: |
| 3018 |
mulx %rdx,%rax,%rbx |
| 3019 |
adox $A1[0],$A1[0] |
| 3020 |
adcx $A0[0],%rax |
| 3021 |
.byte 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00 # mov 8($aptr,$i),%rdx # a[i+1] # prefetch |
| 3022 |
.byte 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00 # mov 32($tptr),$A0[0] # t[2*i+4] # prefetch |
| 3023 |
adox $A1[1],$A1[1] |
| 3024 |
adcx $A0[1],%rbx |
| 3025 |
mov 40($tptr),$A0[1] # t[2*i+4+1] # prefetch |
| 3026 |
mov %rax,0($tptr) |
| 3027 |
mov %rbx,8($tptr) |
| 3028 |
|
| 3029 |
mulx %rdx,%rax,%rbx |
| 3030 |
adox $A0[0],$A0[0] |
| 3031 |
adcx $A1[0],%rax |
| 3032 |
mov 16($aptr,$i),%rdx # a[i+2] # prefetch |
| 3033 |
mov 48($tptr),$A1[0] # t[2*i+6] # prefetch |
| 3034 |
adox $A0[1],$A0[1] |
| 3035 |
adcx $A1[1],%rbx |
| 3036 |
mov 56($tptr),$A1[1] # t[2*i+6+1] # prefetch |
| 3037 |
mov %rax,16($tptr) |
| 3038 |
mov %rbx,24($tptr) |
| 3039 |
|
| 3040 |
mulx %rdx,%rax,%rbx |
| 3041 |
adox $A1[0],$A1[0] |
| 3042 |
adcx $A0[0],%rax |
| 3043 |
mov 24($aptr,$i),%rdx # a[i+3] # prefetch |
| 3044 |
lea 32($i),$i |
| 3045 |
mov 64($tptr),$A0[0] # t[2*i+8] # prefetch |
| 3046 |
adox $A1[1],$A1[1] |
| 3047 |
adcx $A0[1],%rbx |
| 3048 |
mov 72($tptr),$A0[1] # t[2*i+8+1] # prefetch |
| 3049 |
mov %rax,32($tptr) |
| 3050 |
mov %rbx,40($tptr) |
| 3051 |
|
| 3052 |
mulx %rdx,%rax,%rbx |
| 3053 |
adox $A0[0],$A0[0] |
| 3054 |
adcx $A1[0],%rax |
| 3055 |
jrcxz .Lsqrx4x_shift_n_add_break |
| 3056 |
.byte 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00 # mov 0($aptr,$i),%rdx # a[i+4] # prefetch |
| 3057 |
adox $A0[1],$A0[1] |
| 3058 |
adcx $A1[1],%rbx |
| 3059 |
mov 80($tptr),$A1[0] # t[2*i+10] # prefetch |
| 3060 |
mov 88($tptr),$A1[1] # t[2*i+10+1] # prefetch |
| 3061 |
mov %rax,48($tptr) |
| 3062 |
mov %rbx,56($tptr) |
| 3063 |
lea 64($tptr),$tptr |
| 3064 |
nop |
| 3065 |
jmp .Lsqrx4x_shift_n_add |
| 3066 |
|
| 3067 |
.align 32 |
| 3068 |
.Lsqrx4x_shift_n_add_break: |
| 3069 |
adcx $A1[1],%rbx |
| 3070 |
mov %rax,48($tptr) |
| 3071 |
mov %rbx,56($tptr) |
| 3072 |
lea 64($tptr),$tptr # end of t[] buffer |
| 3073 |
___ |
| 3074 |
} |
| 3075 |
###################################################################### |
| 3076 |
# Montgomery reduction part, "word-by-word" algorithm. |
| 3077 |
# |
| 3078 |
# This new path is inspired by multiple submissions from Intel, by |
| 3079 |
# Shay Gueron, Vlad Krasnov, Erdinc Ozturk, James Guilford, |
| 3080 |
# Vinodh Gopal... |
| 3081 |
{ |
| 3082 |
my ($nptr,$carry,$m0)=("%rbp","%rsi","%rdx"); |
| 3083 |
|
| 3084 |
$code.=<<___; |
| 3085 |
movq %xmm2,$nptr |
| 3086 |
__bn_sqrx8x_reduction: |
| 3087 |
xor %eax,%eax # initial top-most carry bit |
| 3088 |
mov 32+8(%rsp),%rbx # n0 |
| 3089 |
mov 48+8(%rsp),%rdx # "%r8", 8*0($tptr) |
| 3090 |
lea -8*8($nptr,$num),%rcx # end of n[] |
| 3091 |
#lea 48+8(%rsp,$num,2),$tptr # end of t[] buffer |
| 3092 |
mov %rcx, 0+8(%rsp) # save end of n[] |
| 3093 |
mov $tptr,8+8(%rsp) # save end of t[] |
| 3094 |
|
| 3095 |
lea 48+8(%rsp),$tptr # initial t[] window |
| 3096 |
jmp .Lsqrx8x_reduction_loop |
| 3097 |
|
| 3098 |
.align 32 |
| 3099 |
.Lsqrx8x_reduction_loop: |
| 3100 |
mov 8*1($tptr),%r9 |
| 3101 |
mov 8*2($tptr),%r10 |
| 3102 |
mov 8*3($tptr),%r11 |
| 3103 |
mov 8*4($tptr),%r12 |
| 3104 |
mov %rdx,%r8 |
| 3105 |
imulq %rbx,%rdx # n0*a[i] |
| 3106 |
mov 8*5($tptr),%r13 |
| 3107 |
mov 8*6($tptr),%r14 |
| 3108 |
mov 8*7($tptr),%r15 |
| 3109 |
mov %rax,24+8(%rsp) # store top-most carry bit |
| 3110 |
|
| 3111 |
lea 8*8($tptr),$tptr |
| 3112 |
xor $carry,$carry # cf=0,of=0 |
| 3113 |
mov \$-8,%rcx |
| 3114 |
jmp .Lsqrx8x_reduce |
| 3115 |
|
| 3116 |
.align 32 |
| 3117 |
.Lsqrx8x_reduce: |
| 3118 |
mov %r8, %rbx |
| 3119 |
mulx 8*0($nptr),%rax,%r8 # n[0] |
| 3120 |
adcx %rbx,%rax # discarded |
| 3121 |
adox %r9,%r8 |
| 3122 |
|
| 3123 |
mulx 8*1($nptr),%rbx,%r9 # n[1] |
| 3124 |
adcx %rbx,%r8 |
| 3125 |
adox %r10,%r9 |
| 3126 |
|
| 3127 |
mulx 8*2($nptr),%rbx,%r10 |
| 3128 |
adcx %rbx,%r9 |
| 3129 |
adox %r11,%r10 |
| 3130 |
|
| 3131 |
mulx 8*3($nptr),%rbx,%r11 |
| 3132 |
adcx %rbx,%r10 |
| 3133 |
adox %r12,%r11 |
| 3134 |
|
| 3135 |
.byte 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00 # mulx 8*4($nptr),%rbx,%r12 |
| 3136 |
mov %rdx,%rax |
| 3137 |
mov %r8,%rdx |
| 3138 |
adcx %rbx,%r11 |
| 3139 |
adox %r13,%r12 |
| 3140 |
|
| 3141 |
mulx 32+8(%rsp),%rbx,%rdx # %rdx discarded |
| 3142 |
mov %rax,%rdx |
| 3143 |
mov %rax,64+48+8(%rsp,%rcx,8) # put aside n0*a[i] |
| 3144 |
|
| 3145 |
mulx 8*5($nptr),%rax,%r13 |
| 3146 |
adcx %rax,%r12 |
| 3147 |
adox %r14,%r13 |
| 3148 |
|
| 3149 |
mulx 8*6($nptr),%rax,%r14 |
| 3150 |
adcx %rax,%r13 |
| 3151 |
adox %r15,%r14 |
| 3152 |
|
| 3153 |
mulx 8*7($nptr),%rax,%r15 |
| 3154 |
mov %rbx,%rdx |
| 3155 |
adcx %rax,%r14 |
| 3156 |
adox $carry,%r15 # $carry is 0 |
| 3157 |
adcx $carry,%r15 # cf=0 |
| 3158 |
|
| 3159 |
.byte 0x67,0x67,0x67 |
| 3160 |
inc %rcx # of=0 |
| 3161 |
jnz .Lsqrx8x_reduce |
| 3162 |
|
| 3163 |
mov $carry,%rax # xor %rax,%rax |
| 3164 |
cmp 0+8(%rsp),$nptr # end of n[]? |
| 3165 |
jae .Lsqrx8x_no_tail |
| 3166 |
|
| 3167 |
mov 48+8(%rsp),%rdx # pull n0*a[0] |
| 3168 |
add 8*0($tptr),%r8 |
| 3169 |
lea 8*8($nptr),$nptr |
| 3170 |
mov \$-8,%rcx |
| 3171 |
adcx 8*1($tptr),%r9 |
| 3172 |
adcx 8*2($tptr),%r10 |
| 3173 |
adc 8*3($tptr),%r11 |
| 3174 |
adc 8*4($tptr),%r12 |
| 3175 |
adc 8*5($tptr),%r13 |
| 3176 |
adc 8*6($tptr),%r14 |
| 3177 |
adc 8*7($tptr),%r15 |
| 3178 |
lea 8*8($tptr),$tptr |
| 3179 |
sbb %rax,%rax # top carry |
| 3180 |
|
| 3181 |
xor $carry,$carry # of=0, cf=0 |
| 3182 |
mov %rax,16+8(%rsp) |
| 3183 |
jmp .Lsqrx8x_tail |
| 3184 |
|
| 3185 |
.align 32 |
| 3186 |
.Lsqrx8x_tail: |
| 3187 |
mov %r8,%rbx |
| 3188 |
mulx 8*0($nptr),%rax,%r8 |
| 3189 |
adcx %rax,%rbx |
| 3190 |
adox %r9,%r8 |
| 3191 |
|
| 3192 |
mulx 8*1($nptr),%rax,%r9 |
| 3193 |
adcx %rax,%r8 |
| 3194 |
adox %r10,%r9 |
| 3195 |
|
| 3196 |
mulx 8*2($nptr),%rax,%r10 |
| 3197 |
adcx %rax,%r9 |
| 3198 |
adox %r11,%r10 |
| 3199 |
|
| 3200 |
mulx 8*3($nptr),%rax,%r11 |
| 3201 |
adcx %rax,%r10 |
| 3202 |
adox %r12,%r11 |
| 3203 |
|
| 3204 |
.byte 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 # mulx 8*4($nptr),%rax,%r12 |
| 3205 |
adcx %rax,%r11 |
| 3206 |
adox %r13,%r12 |
| 3207 |
|
| 3208 |
mulx 8*5($nptr),%rax,%r13 |
| 3209 |
adcx %rax,%r12 |
| 3210 |
adox %r14,%r13 |
| 3211 |
|
| 3212 |
mulx 8*6($nptr),%rax,%r14 |
| 3213 |
adcx %rax,%r13 |
| 3214 |
adox %r15,%r14 |
| 3215 |
|
| 3216 |
mulx 8*7($nptr),%rax,%r15 |
| 3217 |
mov 72+48+8(%rsp,%rcx,8),%rdx # pull n0*a[i] |
| 3218 |
adcx %rax,%r14 |
| 3219 |
adox $carry,%r15 |
| 3220 |
mov %rbx,($tptr,%rcx,8) # save result |
| 3221 |
mov %r8,%rbx |
| 3222 |
adcx $carry,%r15 # cf=0 |
| 3223 |
|
| 3224 |
inc %rcx # of=0 |
| 3225 |
jnz .Lsqrx8x_tail |
| 3226 |
|
| 3227 |
cmp 0+8(%rsp),$nptr # end of n[]? |
| 3228 |
jae .Lsqrx8x_tail_done # break out of loop |
| 3229 |
|
| 3230 |
sub 16+8(%rsp),$carry # mov 16(%rsp),%cf |
| 3231 |
mov 48+8(%rsp),%rdx # pull n0*a[0] |
| 3232 |
lea 8*8($nptr),$nptr |
| 3233 |
adc 8*0($tptr),%r8 |
| 3234 |
adc 8*1($tptr),%r9 |
| 3235 |
adc 8*2($tptr),%r10 |
| 3236 |
adc 8*3($tptr),%r11 |
| 3237 |
adc 8*4($tptr),%r12 |
| 3238 |
adc 8*5($tptr),%r13 |
| 3239 |
adc 8*6($tptr),%r14 |
| 3240 |
adc 8*7($tptr),%r15 |
| 3241 |
lea 8*8($tptr),$tptr |
| 3242 |
sbb %rax,%rax |
| 3243 |
sub \$8,%rcx # mov \$-8,%rcx |
| 3244 |
|
| 3245 |
xor $carry,$carry # of=0, cf=0 |
| 3246 |
mov %rax,16+8(%rsp) |
| 3247 |
jmp .Lsqrx8x_tail |
| 3248 |
|
| 3249 |
.align 32 |
| 3250 |
.Lsqrx8x_tail_done: |
| 3251 |
add 24+8(%rsp),%r8 # can this overflow? |
| 3252 |
adc \$0,%r9 |
| 3253 |
adc \$0,%r10 |
| 3254 |
adc \$0,%r11 |
| 3255 |
adc \$0,%r12 |
| 3256 |
adc \$0,%r13 |
| 3257 |
adc \$0,%r14 |
| 3258 |
adc \$0,%r15 # can't overflow, because we |
| 3259 |
# started with "overhung" part |
| 3260 |
# of multiplication |
| 3261 |
mov $carry,%rax # xor %rax,%rax |
| 3262 |
|
| 3263 |
sub 16+8(%rsp),$carry # mov 16(%rsp),%cf |
| 3264 |
.Lsqrx8x_no_tail: # %cf is 0 if jumped here |
| 3265 |
adc 8*0($tptr),%r8 |
| 3266 |
movq %xmm3,%rcx |
| 3267 |
adc 8*1($tptr),%r9 |
| 3268 |
mov 8*7($nptr),$carry |
| 3269 |
movq %xmm2,$nptr # restore $nptr |
| 3270 |
adc 8*2($tptr),%r10 |
| 3271 |
adc 8*3($tptr),%r11 |
| 3272 |
adc 8*4($tptr),%r12 |
| 3273 |
adc 8*5($tptr),%r13 |
| 3274 |
adc 8*6($tptr),%r14 |
| 3275 |
adc 8*7($tptr),%r15 |
| 3276 |
adc %rax,%rax # top-most carry |
| 3277 |
|
| 3278 |
mov 32+8(%rsp),%rbx # n0 |
| 3279 |
mov 8*8($tptr,%rcx),%rdx # modulo-scheduled "%r8" |
| 3280 |
|
| 3281 |
mov %r8,8*0($tptr) # store top 512 bits |
| 3282 |
lea 8*8($tptr),%r8 # borrow %r8 |
| 3283 |
mov %r9,8*1($tptr) |
| 3284 |
mov %r10,8*2($tptr) |
| 3285 |
mov %r11,8*3($tptr) |
| 3286 |
mov %r12,8*4($tptr) |
| 3287 |
mov %r13,8*5($tptr) |
| 3288 |
mov %r14,8*6($tptr) |
| 3289 |
mov %r15,8*7($tptr) |
| 3290 |
|
| 3291 |
lea 8*8($tptr,%rcx),$tptr # start of current t[] window |
| 3292 |
cmp 8+8(%rsp),%r8 # end of t[]? |
| 3293 |
jb .Lsqrx8x_reduction_loop |
| 3294 |
ret |
| 3295 |
.size bn_sqrx8x_internal,.-bn_sqrx8x_internal |
| 3296 |
___ |
| 3297 |
} |
| 3298 |
############################################################## |
| 3299 |
# Post-condition, 4x unrolled |
| 3300 |
# |
| 3301 |
{ |
| 3302 |
my ($rptr,$nptr)=("%rdx","%rbp"); |
| 3303 |
$code.=<<___; |
| 3304 |
.align 32 |
| 3305 |
__bn_postx4x_internal: |
| 3306 |
mov 8*0($nptr),%r12 |
| 3307 |
mov %rcx,%r10 # -$num |
| 3308 |
mov %rcx,%r9 # -$num |
| 3309 |
neg %rax |
| 3310 |
sar \$3+2,%rcx |
| 3311 |
#lea 48+8(%rsp,%r9),$tptr |
| 3312 |
movq %xmm1,$rptr # restore $rptr |
| 3313 |
movq %xmm1,$aptr # prepare for back-to-back call |
| 3314 |
dec %r12 # so that after 'not' we get -n[0] |
| 3315 |
mov 8*1($nptr),%r13 |
| 3316 |
xor %r8,%r8 |
| 3317 |
mov 8*2($nptr),%r14 |
| 3318 |
mov 8*3($nptr),%r15 |
| 3319 |
jmp .Lsqrx4x_sub_entry |
| 3320 |
|
| 3321 |
.align 16 |
| 3322 |
.Lsqrx4x_sub: |
| 3323 |
mov 8*0($nptr),%r12 |
| 3324 |
mov 8*1($nptr),%r13 |
| 3325 |
mov 8*2($nptr),%r14 |
| 3326 |
mov 8*3($nptr),%r15 |
| 3327 |
.Lsqrx4x_sub_entry: |
| 3328 |
andn %rax,%r12,%r12 |
| 3329 |
lea 8*4($nptr),$nptr |
| 3330 |
andn %rax,%r13,%r13 |
| 3331 |
andn %rax,%r14,%r14 |
| 3332 |
andn %rax,%r15,%r15 |
| 3333 |
|
| 3334 |
neg %r8 # mov %r8,%cf |
| 3335 |
adc 8*0($tptr),%r12 |
| 3336 |
adc 8*1($tptr),%r13 |
| 3337 |
adc 8*2($tptr),%r14 |
| 3338 |
adc 8*3($tptr),%r15 |
| 3339 |
mov %r12,8*0($rptr) |
| 3340 |
lea 8*4($tptr),$tptr |
| 3341 |
mov %r13,8*1($rptr) |
| 3342 |
sbb %r8,%r8 # mov %cf,%r8 |
| 3343 |
mov %r14,8*2($rptr) |
| 3344 |
mov %r15,8*3($rptr) |
| 3345 |
lea 8*4($rptr),$rptr |
| 3346 |
|
| 3347 |
inc %rcx |
| 3348 |
jnz .Lsqrx4x_sub |
| 3349 |
|
| 3350 |
neg %r9 # restore $num |
| 3351 |
|
| 3352 |
ret |
| 3353 |
.size __bn_postx4x_internal,.-__bn_postx4x_internal |
| 3354 |
___ |
| 3355 |
} |
| 3356 |
}}} |
| 3357 |
{ |
| 3358 |
my ($inp,$num,$tbl,$idx)=$win64?("%rcx","%edx","%r8", "%r9d") : # Win64 order |
| 3359 |
("%rdi","%esi","%rdx","%ecx"); # Unix order |
| 3360 |
my $out=$inp; |
| 3361 |
my $STRIDE=2**5*8; |
| 3362 |
my $N=$STRIDE/4; |
| 3363 |
|
| 3364 |
$code.=<<___; |
| 3365 |
.globl bn_scatter5 |
| 3366 |
.type bn_scatter5,\@abi-omnipotent |
| 3367 |
.align 16 |
| 3368 |
bn_scatter5: |
| 3369 |
cmp \$0, $num |
| 3370 |
jz .Lscatter_epilogue |
| 3371 |
lea ($tbl,$idx,8),$tbl |
| 3372 |
.Lscatter: |
| 3373 |
mov ($inp),%rax |
| 3374 |
lea 8($inp),$inp |
| 3375 |
mov %rax,($tbl) |
| 3376 |
lea 32*8($tbl),$tbl |
| 3377 |
sub \$1,$num |
| 3378 |
jnz .Lscatter |
| 3379 |
.Lscatter_epilogue: |
| 3380 |
ret |
| 3381 |
.size bn_scatter5,.-bn_scatter5 |
| 3382 |
|
| 3383 |
.globl bn_gather5 |
| 3384 |
.type bn_gather5,\@abi-omnipotent |
| 3385 |
.align 32 |
| 3386 |
bn_gather5: |
| 3387 |
.LSEH_begin_bn_gather5: # Win64 thing, but harmless in other cases |
| 3388 |
# I can't trust assembler to use specific encoding:-( |
| 3389 |
.byte 0x4c,0x8d,0x14,0x24 #lea (%rsp),%r10 |
| 3390 |
.byte 0x48,0x81,0xec,0x08,0x01,0x00,0x00 #sub $0x108,%rsp |
| 3391 |
lea .Linc(%rip),%rax |
| 3392 |
and \$-16,%rsp # shouldn't be formally required |
| 3393 |
|
| 3394 |
movd $idx,%xmm5 |
| 3395 |
movdqa 0(%rax),%xmm0 # 00000001000000010000000000000000 |
| 3396 |
movdqa 16(%rax),%xmm1 # 00000002000000020000000200000002 |
| 3397 |
lea 128($tbl),%r11 # size optimization |
| 3398 |
lea 128(%rsp),%rax # size optimization |
| 3399 |
|
| 3400 |
pshufd \$0,%xmm5,%xmm5 # broadcast $idx |
| 3401 |
movdqa %xmm1,%xmm4 |
| 3402 |
movdqa %xmm1,%xmm2 |
| 3403 |
___ |
| 3404 |
######################################################################## |
| 3405 |
# calculate mask by comparing 0..31 to $idx and save result to stack |
| 3406 |
# |
| 3407 |
for($i=0;$i<$STRIDE/16;$i+=4) { |
| 3408 |
$code.=<<___; |
| 3409 |
paddd %xmm0,%xmm1 |
| 3410 |
pcmpeqd %xmm5,%xmm0 # compare to 1,0 |
| 3411 |
___ |
| 3412 |
$code.=<<___ if ($i); |
| 3413 |
movdqa %xmm3,`16*($i-1)-128`(%rax) |
| 3414 |
___ |
| 3415 |
$code.=<<___; |
| 3416 |
movdqa %xmm4,%xmm3 |
| 3417 |
|
| 3418 |
paddd %xmm1,%xmm2 |
| 3419 |
pcmpeqd %xmm5,%xmm1 # compare to 3,2 |
| 3420 |
movdqa %xmm0,`16*($i+0)-128`(%rax) |
| 3421 |
movdqa %xmm4,%xmm0 |
| 3422 |
|
| 3423 |
paddd %xmm2,%xmm3 |
| 3424 |
pcmpeqd %xmm5,%xmm2 # compare to 5,4 |
| 3425 |
movdqa %xmm1,`16*($i+1)-128`(%rax) |
| 3426 |
movdqa %xmm4,%xmm1 |
| 3427 |
|
| 3428 |
paddd %xmm3,%xmm0 |
| 3429 |
pcmpeqd %xmm5,%xmm3 # compare to 7,6 |
| 3430 |
movdqa %xmm2,`16*($i+2)-128`(%rax) |
| 3431 |
movdqa %xmm4,%xmm2 |
| 3432 |
___ |
| 3433 |
} |
| 3434 |
$code.=<<___; |
| 3435 |
movdqa %xmm3,`16*($i-1)-128`(%rax) |
| 3436 |
jmp .Lgather |
| 3437 |
|
| 3438 |
.align 32 |
| 3439 |
.Lgather: |
| 3440 |
pxor %xmm4,%xmm4 |
| 3441 |
pxor %xmm5,%xmm5 |
| 3442 |
___ |
| 3443 |
for($i=0;$i<$STRIDE/16;$i+=4) { |
| 3444 |
$code.=<<___; |
| 3445 |
movdqa `16*($i+0)-128`(%r11),%xmm0 |
| 3446 |
movdqa `16*($i+1)-128`(%r11),%xmm1 |
| 3447 |
movdqa `16*($i+2)-128`(%r11),%xmm2 |
| 3448 |
pand `16*($i+0)-128`(%rax),%xmm0 |
| 3449 |
movdqa `16*($i+3)-128`(%r11),%xmm3 |
| 3450 |
pand `16*($i+1)-128`(%rax),%xmm1 |
| 3451 |
por %xmm0,%xmm4 |
| 3452 |
pand `16*($i+2)-128`(%rax),%xmm2 |
| 3453 |
por %xmm1,%xmm5 |
| 3454 |
pand `16*($i+3)-128`(%rax),%xmm3 |
| 3455 |
por %xmm2,%xmm4 |
| 3456 |
por %xmm3,%xmm5 |
| 3457 |
___ |
| 3458 |
} |
| 3459 |
$code.=<<___; |
| 3460 |
por %xmm5,%xmm4 |
| 3461 |
lea $STRIDE(%r11),%r11 |
| 3462 |
pshufd \$0x4e,%xmm4,%xmm0 |
| 3463 |
por %xmm4,%xmm0 |
| 3464 |
movq %xmm0,($out) # m0=bp[0] |
| 3465 |
lea 8($out),$out |
| 3466 |
sub \$1,$num |
| 3467 |
jnz .Lgather |
| 3468 |
|
| 3469 |
lea (%r10),%rsp |
| 3470 |
ret |
| 3471 |
.LSEH_end_bn_gather5: |
| 3472 |
.size bn_gather5,.-bn_gather5 |
| 3473 |
___ |
| 3474 |
} |
| 3475 |
$code.=<<___; |
| 3476 |
.align 64 |
| 3477 |
.Linc: |
| 3478 |
.long 0,0, 1,1 |
| 3479 |
.long 2,2, 2,2 |
| 3480 |
.asciz "Montgomery Multiplication with scatter/gather for x86_64, CRYPTOGAMS by <appro\@openssl.org>" |
| 3481 |
___ |
| 3482 |
|
| 3483 |
# EXCEPTION_DISPOSITION handler (EXCEPTION_RECORD *rec,ULONG64 frame, |
| 3484 |
# CONTEXT *context,DISPATCHER_CONTEXT *disp) |
| 3485 |
if ($win64) { |
| 3486 |
$rec="%rcx"; |
| 3487 |
$frame="%rdx"; |
| 3488 |
$context="%r8"; |
| 3489 |
$disp="%r9"; |
| 3490 |
|
| 3491 |
$code.=<<___; |
| 3492 |
.extern __imp_RtlVirtualUnwind |
| 3493 |
.type mul_handler,\@abi-omnipotent |
| 3494 |
.align 16 |
| 3495 |
mul_handler: |
| 3496 |
push %rsi |
| 3497 |
push %rdi |
| 3498 |
push %rbx |
| 3499 |
push %rbp |
| 3500 |
push %r12 |
| 3501 |
push %r13 |
| 3502 |
push %r14 |
| 3503 |
push %r15 |
| 3504 |
pushfq |
| 3505 |
sub \$64,%rsp |
| 3506 |
|
| 3507 |
mov 120($context),%rax # pull context->Rax |
| 3508 |
mov 248($context),%rbx # pull context->Rip |
| 3509 |
|
| 3510 |
mov 8($disp),%rsi # disp->ImageBase |
| 3511 |
mov 56($disp),%r11 # disp->HandlerData |
| 3512 |
|
| 3513 |
mov 0(%r11),%r10d # HandlerData[0] |
| 3514 |
lea (%rsi,%r10),%r10 # end of prologue label |
| 3515 |
cmp %r10,%rbx # context->Rip<end of prologue label |
| 3516 |
jb .Lcommon_seh_tail |
| 3517 |
|
| 3518 |
mov 152($context),%rax # pull context->Rsp |
| 3519 |
|
| 3520 |
mov 4(%r11),%r10d # HandlerData[1] |
| 3521 |
lea (%rsi,%r10),%r10 # epilogue label |
| 3522 |
cmp %r10,%rbx # context->Rip>=epilogue label |
| 3523 |
jae .Lcommon_seh_tail |
| 3524 |
|
| 3525 |
lea .Lmul_epilogue(%rip),%r10 |
| 3526 |
cmp %r10,%rbx |
| 3527 |
ja .Lbody_40 |
| 3528 |
|
| 3529 |
mov 192($context),%r10 # pull $num |
| 3530 |
mov 8(%rax,%r10,8),%rax # pull saved stack pointer |
| 3531 |
|
| 3532 |
jmp .Lbody_proceed |
| 3533 |
|
| 3534 |
.Lbody_40: |
| 3535 |
mov 40(%rax),%rax # pull saved stack pointer |
| 3536 |
.Lbody_proceed: |
| 3537 |
mov -8(%rax),%rbx |
| 3538 |
mov -16(%rax),%rbp |
| 3539 |
mov -24(%rax),%r12 |
| 3540 |
mov -32(%rax),%r13 |
| 3541 |
mov -40(%rax),%r14 |
| 3542 |
mov -48(%rax),%r15 |
| 3543 |
mov %rbx,144($context) # restore context->Rbx |
| 3544 |
mov %rbp,160($context) # restore context->Rbp |
| 3545 |
mov %r12,216($context) # restore context->R12 |
| 3546 |
mov %r13,224($context) # restore context->R13 |
| 3547 |
mov %r14,232($context) # restore context->R14 |
| 3548 |
mov %r15,240($context) # restore context->R15 |
| 3549 |
|
| 3550 |
.Lcommon_seh_tail: |
| 3551 |
mov 8(%rax),%rdi |
| 3552 |
mov 16(%rax),%rsi |
| 3553 |
mov %rax,152($context) # restore context->Rsp |
| 3554 |
mov %rsi,168($context) # restore context->Rsi |
| 3555 |
mov %rdi,176($context) # restore context->Rdi |
| 3556 |
|
| 3557 |
mov 40($disp),%rdi # disp->ContextRecord |
| 3558 |
mov $context,%rsi # context |
| 3559 |
mov \$154,%ecx # sizeof(CONTEXT) |
| 3560 |
.long 0xa548f3fc # cld; rep movsq |
| 3561 |
|
| 3562 |
mov $disp,%rsi |
| 3563 |
xor %rcx,%rcx # arg1, UNW_FLAG_NHANDLER |
| 3564 |
mov 8(%rsi),%rdx # arg2, disp->ImageBase |
| 3565 |
mov 0(%rsi),%r8 # arg3, disp->ControlPc |
| 3566 |
mov 16(%rsi),%r9 # arg4, disp->FunctionEntry |
| 3567 |
mov 40(%rsi),%r10 # disp->ContextRecord |
| 3568 |
lea 56(%rsi),%r11 # &disp->HandlerData |
| 3569 |
lea 24(%rsi),%r12 # &disp->EstablisherFrame |
| 3570 |
mov %r10,32(%rsp) # arg5 |
| 3571 |
mov %r11,40(%rsp) # arg6 |
| 3572 |
mov %r12,48(%rsp) # arg7 |
| 3573 |
mov %rcx,56(%rsp) # arg8, (NULL) |
| 3574 |
call *__imp_RtlVirtualUnwind(%rip) |
| 3575 |
|
| 3576 |
mov \$1,%eax # ExceptionContinueSearch |
| 3577 |
add \$64,%rsp |
| 3578 |
popfq |
| 3579 |
pop %r15 |
| 3580 |
pop %r14 |
| 3581 |
pop %r13 |
| 3582 |
pop %r12 |
| 3583 |
pop %rbp |
| 3584 |
pop %rbx |
| 3585 |
pop %rdi |
| 3586 |
pop %rsi |
| 3587 |
ret |
| 3588 |
.size mul_handler,.-mul_handler |
| 3589 |
|
| 3590 |
.section .pdata |
| 3591 |
.align 4 |
| 3592 |
.rva .LSEH_begin_bn_mul_mont_gather5 |
| 3593 |
.rva .LSEH_end_bn_mul_mont_gather5 |
| 3594 |
.rva .LSEH_info_bn_mul_mont_gather5 |
| 3595 |
|
| 3596 |
.rva .LSEH_begin_bn_mul4x_mont_gather5 |
| 3597 |
.rva .LSEH_end_bn_mul4x_mont_gather5 |
| 3598 |
.rva .LSEH_info_bn_mul4x_mont_gather5 |
| 3599 |
|
| 3600 |
.rva .LSEH_begin_bn_power5 |
| 3601 |
.rva .LSEH_end_bn_power5 |
| 3602 |
.rva .LSEH_info_bn_power5 |
| 3603 |
|
| 3604 |
.rva .LSEH_begin_bn_from_mont8x |
| 3605 |
.rva .LSEH_end_bn_from_mont8x |
| 3606 |
.rva .LSEH_info_bn_from_mont8x |
| 3607 |
___ |
| 3608 |
$code.=<<___ if ($addx); |
| 3609 |
.rva .LSEH_begin_bn_mulx4x_mont_gather5 |
| 3610 |
.rva .LSEH_end_bn_mulx4x_mont_gather5 |
| 3611 |
.rva .LSEH_info_bn_mulx4x_mont_gather5 |
| 3612 |
|
| 3613 |
.rva .LSEH_begin_bn_powerx5 |
| 3614 |
.rva .LSEH_end_bn_powerx5 |
| 3615 |
.rva .LSEH_info_bn_powerx5 |
| 3616 |
___ |
| 3617 |
$code.=<<___; |
| 3618 |
.rva .LSEH_begin_bn_gather5 |
| 3619 |
.rva .LSEH_end_bn_gather5 |
| 3620 |
.rva .LSEH_info_bn_gather5 |
| 3621 |
|
| 3622 |
.section .xdata |
| 3623 |
.align 8 |
| 3624 |
.LSEH_info_bn_mul_mont_gather5: |
| 3625 |
.byte 9,0,0,0 |
| 3626 |
.rva mul_handler |
| 3627 |
.rva .Lmul_body,.Lmul_epilogue # HandlerData[] |
| 3628 |
.align 8 |
| 3629 |
.LSEH_info_bn_mul4x_mont_gather5: |
| 3630 |
.byte 9,0,0,0 |
| 3631 |
.rva mul_handler |
| 3632 |
.rva .Lmul4x_body,.Lmul4x_epilogue # HandlerData[] |
| 3633 |
.align 8 |
| 3634 |
.LSEH_info_bn_power5: |
| 3635 |
.byte 9,0,0,0 |
| 3636 |
.rva mul_handler |
| 3637 |
.rva .Lpower5_body,.Lpower5_epilogue # HandlerData[] |
| 3638 |
.align 8 |
| 3639 |
.LSEH_info_bn_from_mont8x: |
| 3640 |
.byte 9,0,0,0 |
| 3641 |
.rva mul_handler |
| 3642 |
.rva .Lfrom_body,.Lfrom_epilogue # HandlerData[] |
| 3643 |
___ |
| 3644 |
$code.=<<___ if ($addx); |
| 3645 |
.align 8 |
| 3646 |
.LSEH_info_bn_mulx4x_mont_gather5: |
| 3647 |
.byte 9,0,0,0 |
| 3648 |
.rva mul_handler |
| 3649 |
.rva .Lmulx4x_body,.Lmulx4x_epilogue # HandlerData[] |
| 3650 |
.align 8 |
| 3651 |
.LSEH_info_bn_powerx5: |
| 3652 |
.byte 9,0,0,0 |
| 3653 |
.rva mul_handler |
| 3654 |
.rva .Lpowerx5_body,.Lpowerx5_epilogue # HandlerData[] |
| 3655 |
___ |
| 3656 |
$code.=<<___; |
| 3657 |
.align 8 |
| 3658 |
.LSEH_info_bn_gather5: |
| 3659 |
.byte 0x01,0x0b,0x03,0x0a |
| 3660 |
.byte 0x0b,0x01,0x21,0x00 # sub rsp,0x108 |
| 3661 |
.byte 0x04,0xa3,0x00,0x00 # lea r10,(rsp) |
| 3662 |
.align 8 |
| 3663 |
___ |
| 3664 |
} |
| 3665 |
|
| 3666 |
$code =~ s/\`([^\`]*)\`/eval($1)/gem; |
| 3667 |
|
| 3668 |
print $code; |
| 3669 |
close STDOUT; |