| 1 |
dnl AMD64 mpn_sqr_basecase optimised for Intel Haswell. |
| 2 |
|
| 3 |
dnl Contributed to the GNU project by Torbjörn Granlund. |
| 4 |
|
| 5 |
dnl Copyright 2008, 2009, 2011-2013 Free Software Foundation, Inc. |
| 6 |
|
| 7 |
dnl This file is part of the GNU MP Library. |
| 8 |
dnl |
| 9 |
dnl The GNU MP Library is free software; you can redistribute it and/or modify |
| 10 |
dnl it under the terms of either: |
| 11 |
dnl |
| 12 |
dnl * the GNU Lesser General Public License as published by the Free |
| 13 |
dnl Software Foundation; either version 3 of the License, or (at your |
| 14 |
dnl option) any later version. |
| 15 |
dnl |
| 16 |
dnl or |
| 17 |
dnl |
| 18 |
dnl * the GNU General Public License as published by the Free Software |
| 19 |
dnl Foundation; either version 2 of the License, or (at your option) any |
| 20 |
dnl later version. |
| 21 |
dnl |
| 22 |
dnl or both in parallel, as here. |
| 23 |
dnl |
| 24 |
dnl The GNU MP Library is distributed in the hope that it will be useful, but |
| 25 |
dnl WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY |
| 26 |
dnl or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
| 27 |
dnl for more details. |
| 28 |
dnl |
| 29 |
dnl You should have received copies of the GNU General Public License and the |
| 30 |
dnl GNU Lesser General Public License along with the GNU MP Library. If not, |
| 31 |
dnl see https://www.gnu.org/licenses/. |
| 32 |
|
| 33 |
include(`../config.m4') |
| 34 |
|
| 35 |
C cycles/limb mul_2 addmul_2 sqr_diag_addlsh1 |
| 36 |
C AMD K8,K9 n/a n/a n/a |
| 37 |
C AMD K10 n/a n/a n/a |
| 38 |
C AMD bull n/a n/a n/a |
| 39 |
C AMD pile n/a n/a n/a |
| 40 |
C AMD steam ? ? ? |
| 41 |
C AMD bobcat n/a n/a n/a |
| 42 |
C AMD jaguar ? ? ? |
| 43 |
C Intel P4 n/a n/a n/a |
| 44 |
C Intel core n/a n/a n/a |
| 45 |
C Intel NHM n/a n/a n/a |
| 46 |
C Intel SBR n/a n/a n/a |
| 47 |
C Intel IBR n/a n/a n/a |
| 48 |
C Intel HWL 1.86 2.15 ~2.5 |
| 49 |
C Intel BWL ? ? ? |
| 50 |
C Intel atom n/a n/a n/a |
| 51 |
C VIA nano n/a n/a n/a |
| 52 |
|
| 53 |
C The inner loops of this code are the result of running a code generation and |
| 54 |
C optimisation tool suite written by David Harvey and Torbjörn Granlund, except |
| 55 |
C that the sqr_diag_addlsh1 loop was manually written. |
| 56 |
|
| 57 |
C TODO |
| 58 |
C * Replace current unoptimised sqr_diag_addlsh1 loop; 1.75 c/l might be |
| 59 |
C possible. |
| 60 |
C * Consider splitting outer loop into 2, one for n = 1 (mod 2) and one for |
| 61 |
C n = 0 (mod 2). These loops could fall into specific "corner" code. |
| 62 |
C * Consider splitting outer loop into 4. |
| 63 |
C * Streamline pointer updates. |
| 64 |
C * Perhaps suppress a few more xor insns in feed-in code. |
| 65 |
C * Make sure we write no dead registers in feed-in code. |
| 66 |
C * We might use 32-bit size ops, since n >= 2^32 is non-terminating. Watch |
| 67 |
C out for negative sizes being zero-extended, though. |
| 68 |
C * Provide straight-line code for n = 4; then look for simplifications in |
| 69 |
C main code. |
| 70 |
|
| 71 |
define(`rp', `%rdi') |
| 72 |
define(`up', `%rsi') |
| 73 |
define(`un_param',`%rdx') |
| 74 |
|
| 75 |
|
| 76 |
ABI_SUPPORT(DOS64) |
| 77 |
ABI_SUPPORT(STD64) |
| 78 |
|
| 79 |
ASM_START() |
| 80 |
TEXT |
| 81 |
ALIGN(32) |
| 82 |
PROLOGUE(mpn_sqr_basecase) |
| 83 |
FUNC_ENTRY(3) |
| 84 |
|
| 85 |
cmp $2, un_param |
| 86 |
jae L(gt1) |
| 87 |
|
| 88 |
mov (up), %rdx |
| 89 |
mulx( %rdx, %rax, %rdx) |
| 90 |
mov %rax, (rp) |
| 91 |
mov %rdx, 8(rp) |
| 92 |
FUNC_EXIT() |
| 93 |
ret |
| 94 |
|
| 95 |
L(gt1): jne L(gt2) |
| 96 |
|
| 97 |
mov (up), %rdx |
| 98 |
mov 8(up), %rcx |
| 99 |
mulx( %rcx, %r9, %r10) C v0 * v1 W 1 2 |
| 100 |
mulx( %rdx, %rax, %r8) C v0 * v0 W 0 1 |
| 101 |
mov %rcx, %rdx |
| 102 |
mulx( %rdx, %r11, %rdx) C v1 * v1 W 2 3 |
| 103 |
add %r9, %r9 C W 1 |
| 104 |
adc %r10, %r10 C W 2 |
| 105 |
adc $0, %rdx C W 3 |
| 106 |
add %r9, %r8 C W 1 |
| 107 |
adc %r11, %r10 C W 2 |
| 108 |
adc $0, %rdx C W 3 |
| 109 |
mov %rax, (rp) |
| 110 |
mov %r8, 8(rp) |
| 111 |
mov %r10, 16(rp) |
| 112 |
mov %rdx, 24(rp) |
| 113 |
FUNC_EXIT() |
| 114 |
ret |
| 115 |
|
| 116 |
L(gt2): cmp $4, un_param |
| 117 |
jae L(gt3) |
| 118 |
define(`v0', `%r8') |
| 119 |
define(`v1', `%r9') |
| 120 |
define(`w0', `%r10') |
| 121 |
define(`w2', `%r11') |
| 122 |
|
| 123 |
mov (up), v0 |
| 124 |
mov 8(up), %rdx |
| 125 |
mov %rdx, v1 |
| 126 |
mulx( v0, w2, %rax) |
| 127 |
mov 16(up), %rdx |
| 128 |
mulx( v0, w0, %rcx) |
| 129 |
mov w2, %r8 |
| 130 |
add %rax, w0 |
| 131 |
adc $0, %rcx |
| 132 |
mulx( v1, %rdx, %rax) |
| 133 |
add %rcx, %rdx |
| 134 |
mov %rdx, 24(rp) |
| 135 |
adc $0, %rax |
| 136 |
mov %rax, 32(rp) |
| 137 |
xor R32(%rcx), R32(%rcx) |
| 138 |
mov (up), %rdx |
| 139 |
mulx( %rdx, %rax, w2) |
| 140 |
mov %rax, (rp) |
| 141 |
add %r8, %r8 |
| 142 |
adc w0, w0 |
| 143 |
setc R8(%rcx) |
| 144 |
mov 8(up), %rdx |
| 145 |
mulx( %rdx, %rax, %rdx) |
| 146 |
add w2, %r8 |
| 147 |
adc %rax, w0 |
| 148 |
mov %r8, 8(rp) |
| 149 |
mov w0, 16(rp) |
| 150 |
mov 24(rp), %r8 |
| 151 |
mov 32(rp), w0 |
| 152 |
lea (%rdx,%rcx), w2 |
| 153 |
adc %r8, %r8 |
| 154 |
adc w0, w0 |
| 155 |
setc R8(%rcx) |
| 156 |
mov 16(up), %rdx |
| 157 |
mulx( %rdx, %rax, %rdx) |
| 158 |
add w2, %r8 |
| 159 |
adc %rax, w0 |
| 160 |
mov %r8, 24(rp) |
| 161 |
mov w0, 32(rp) |
| 162 |
adc %rcx, %rdx |
| 163 |
mov %rdx, 40(rp) |
| 164 |
FUNC_EXIT() |
| 165 |
ret |
| 166 |
|
| 167 |
L(gt3): |
| 168 |
|
| 169 |
define(`v0', `%r8') |
| 170 |
define(`v1', `%r9') |
| 171 |
define(`w0', `%r10') |
| 172 |
define(`w1', `%r11') |
| 173 |
define(`w2', `%rbx') |
| 174 |
define(`w3', `%rbp') |
| 175 |
define(`un', `%r12') |
| 176 |
define(`n', `%rcx') |
| 177 |
|
| 178 |
define(`X0', `%r13') |
| 179 |
define(`X1', `%r14') |
| 180 |
|
| 181 |
L(do_mul_2): |
| 182 |
push %rbx |
| 183 |
push %rbp |
| 184 |
push %r12 |
| 185 |
push %r13 |
| 186 |
push %r14 |
| 187 |
mov $0, R32(un) |
| 188 |
sub un_param, un C free up rdx |
| 189 |
push un |
| 190 |
mov (up), v0 |
| 191 |
mov 8(up), %rdx |
| 192 |
lea 2(un), n |
| 193 |
sar $2, n C FIXME: suppress, change loop? |
| 194 |
inc un C decrement |un| |
| 195 |
mov %rdx, v1 |
| 196 |
|
| 197 |
test $1, R8(un) |
| 198 |
jnz L(mx1) |
| 199 |
|
| 200 |
L(mx0): mulx( v0, w2, w1) |
| 201 |
mov 16(up), %rdx |
| 202 |
mov w2, 8(rp) |
| 203 |
xor w2, w2 |
| 204 |
mulx( v0, w0, w3) |
| 205 |
test $2, R8(un) |
| 206 |
jz L(m00) |
| 207 |
|
| 208 |
L(m10): lea -8(rp), rp |
| 209 |
lea -8(up), up |
| 210 |
jmp L(mlo2) |
| 211 |
|
| 212 |
L(m00): lea 8(up), up |
| 213 |
lea 8(rp), rp |
| 214 |
jmp L(mlo0) |
| 215 |
|
| 216 |
L(mx1): mulx( v0, w0, w3) |
| 217 |
mov 16(up), %rdx |
| 218 |
mov w0, 8(rp) |
| 219 |
xor w0, w0 |
| 220 |
mulx( v0, w2, w1) |
| 221 |
test $2, R8(un) |
| 222 |
jz L(mlo3) |
| 223 |
|
| 224 |
L(m01): lea 16(rp), rp |
| 225 |
lea 16(up), up |
| 226 |
jmp L(mlo1) |
| 227 |
|
| 228 |
ALIGN(32) |
| 229 |
L(mtop):mulx( v1, %rax, w0) |
| 230 |
add %rax, w2 C 0 |
| 231 |
mov (up), %rdx |
| 232 |
mulx( v0, %rax, w1) |
| 233 |
adc $0, w0 C 1 |
| 234 |
add %rax, w2 C 0 |
| 235 |
L(mlo1):adc $0, w1 C 1 |
| 236 |
add w3, w2 C 0 |
| 237 |
mov w2, (rp) C 0 |
| 238 |
adc $0, w1 C 1 |
| 239 |
mulx( v1, %rax, w2) |
| 240 |
add %rax, w0 C 1 |
| 241 |
mov 8(up), %rdx |
| 242 |
adc $0, w2 C 2 |
| 243 |
mulx( v0, %rax, w3) |
| 244 |
add %rax, w0 C 1 |
| 245 |
adc $0, w3 C 2 |
| 246 |
L(mlo0):add w1, w0 C 1 |
| 247 |
mov w0, 8(rp) C 1 |
| 248 |
adc $0, w3 C 2 |
| 249 |
mulx( v1, %rax, w0) |
| 250 |
add %rax, w2 C 2 |
| 251 |
mov 16(up), %rdx |
| 252 |
mulx( v0, %rax, w1) |
| 253 |
adc $0, w0 C 3 |
| 254 |
add %rax, w2 C 2 |
| 255 |
adc $0, w1 C 3 |
| 256 |
L(mlo3):add w3, w2 C 2 |
| 257 |
mov w2, 16(rp) C 2 |
| 258 |
adc $0, w1 C 3 |
| 259 |
mulx( v1, %rax, w2) |
| 260 |
add %rax, w0 C 3 |
| 261 |
mov 24(up), %rdx |
| 262 |
adc $0, w2 C 4 |
| 263 |
mulx( v0, %rax, w3) |
| 264 |
add %rax, w0 C 3 |
| 265 |
adc $0, w3 C 4 |
| 266 |
L(mlo2):add w1, w0 C 3 |
| 267 |
lea 32(up), up |
| 268 |
mov w0, 24(rp) C 3 |
| 269 |
adc $0, w3 C 4 |
| 270 |
inc n |
| 271 |
lea 32(rp), rp |
| 272 |
jnz L(mtop) |
| 273 |
|
| 274 |
L(mend):mulx( v1, %rdx, %rax) |
| 275 |
add %rdx, w2 |
| 276 |
adc $0, %rax |
| 277 |
add w3, w2 |
| 278 |
mov w2, (rp) |
| 279 |
adc $0, %rax |
| 280 |
mov %rax, 8(rp) |
| 281 |
|
| 282 |
lea 16(up), up |
| 283 |
lea -16(rp), rp |
| 284 |
|
| 285 |
L(do_addmul_2): |
| 286 |
L(outer): |
| 287 |
lea (up,un,8), up C put back up to 2 positions above last time |
| 288 |
lea 48(rp,un,8), rp C put back rp to 4 positions above last time |
| 289 |
|
| 290 |
mov -8(up), v0 C shared between addmul_2 and corner |
| 291 |
|
| 292 |
add $2, un C decrease |un| |
| 293 |
cmp $-2, un |
| 294 |
jge L(corner) |
| 295 |
|
| 296 |
mov (up), v1 |
| 297 |
|
| 298 |
lea 1(un), n |
| 299 |
sar $2, n C FIXME: suppress, change loop? |
| 300 |
|
| 301 |
mov v1, %rdx |
| 302 |
test $1, R8(un) |
| 303 |
jnz L(bx1) |
| 304 |
|
| 305 |
L(bx0): mov (rp), X0 |
| 306 |
mov 8(rp), X1 |
| 307 |
mulx( v0, %rax, w1) |
| 308 |
add %rax, X0 |
| 309 |
adc $0, w1 |
| 310 |
mov X0, (rp) |
| 311 |
xor w2, w2 |
| 312 |
test $2, R8(un) |
| 313 |
jnz L(b10) |
| 314 |
|
| 315 |
L(b00): mov 8(up), %rdx |
| 316 |
lea 16(rp), rp |
| 317 |
lea 16(up), up |
| 318 |
jmp L(lo0) |
| 319 |
|
| 320 |
L(b10): mov 8(up), %rdx |
| 321 |
mov 16(rp), X0 |
| 322 |
lea 32(up), up |
| 323 |
inc n |
| 324 |
mulx( v0, %rax, w3) |
| 325 |
jz L(ex) |
| 326 |
jmp L(lo2) |
| 327 |
|
| 328 |
L(bx1): mov (rp), X1 |
| 329 |
mov 8(rp), X0 |
| 330 |
mulx( v0, %rax, w3) |
| 331 |
mov 8(up), %rdx |
| 332 |
add %rax, X1 |
| 333 |
adc $0, w3 |
| 334 |
xor w0, w0 |
| 335 |
mov X1, (rp) |
| 336 |
mulx( v0, %rax, w1) |
| 337 |
test $2, R8(un) |
| 338 |
jz L(b11) |
| 339 |
|
| 340 |
L(b01): mov 16(rp), X1 |
| 341 |
lea 24(rp), rp |
| 342 |
lea 24(up), up |
| 343 |
jmp L(lo1) |
| 344 |
|
| 345 |
L(b11): lea 8(rp), rp |
| 346 |
lea 8(up), up |
| 347 |
jmp L(lo3) |
| 348 |
|
| 349 |
ALIGN(32) |
| 350 |
L(top): mulx( v0, %rax, w3) |
| 351 |
add w0, X1 |
| 352 |
adc $0, w2 |
| 353 |
L(lo2): add %rax, X1 |
| 354 |
adc $0, w3 |
| 355 |
mulx( v1, %rax, w0) |
| 356 |
add %rax, X0 |
| 357 |
adc $0, w0 |
| 358 |
lea 32(rp), rp |
| 359 |
add w1, X1 |
| 360 |
mov -16(up), %rdx |
| 361 |
mov X1, -24(rp) |
| 362 |
adc $0, w3 |
| 363 |
add w2, X0 |
| 364 |
mov -8(rp), X1 |
| 365 |
mulx( v0, %rax, w1) |
| 366 |
adc $0, w0 |
| 367 |
L(lo1): add %rax, X0 |
| 368 |
mulx( v1, %rax, w2) |
| 369 |
adc $0, w1 |
| 370 |
add w3, X0 |
| 371 |
mov X0, -16(rp) |
| 372 |
adc $0, w1 |
| 373 |
add %rax, X1 |
| 374 |
adc $0, w2 |
| 375 |
add w0, X1 |
| 376 |
mov -8(up), %rdx |
| 377 |
adc $0, w2 |
| 378 |
L(lo0): mulx( v0, %rax, w3) |
| 379 |
add %rax, X1 |
| 380 |
adc $0, w3 |
| 381 |
mov (rp), X0 |
| 382 |
mulx( v1, %rax, w0) |
| 383 |
add %rax, X0 |
| 384 |
adc $0, w0 |
| 385 |
add w1, X1 |
| 386 |
mov X1, -8(rp) |
| 387 |
adc $0, w3 |
| 388 |
mov (up), %rdx |
| 389 |
add w2, X0 |
| 390 |
mulx( v0, %rax, w1) |
| 391 |
adc $0, w0 |
| 392 |
L(lo3): add %rax, X0 |
| 393 |
adc $0, w1 |
| 394 |
mulx( v1, %rax, w2) |
| 395 |
add w3, X0 |
| 396 |
mov 8(rp), X1 |
| 397 |
mov X0, (rp) |
| 398 |
mov 16(rp), X0 |
| 399 |
adc $0, w1 |
| 400 |
add %rax, X1 |
| 401 |
adc $0, w2 |
| 402 |
mov 8(up), %rdx |
| 403 |
lea 32(up), up |
| 404 |
inc n |
| 405 |
jnz L(top) |
| 406 |
|
| 407 |
L(end): mulx( v0, %rax, w3) |
| 408 |
add w0, X1 |
| 409 |
adc $0, w2 |
| 410 |
L(ex): add %rax, X1 |
| 411 |
adc $0, w3 |
| 412 |
mulx( v1, %rdx, %rax) |
| 413 |
add w1, X1 |
| 414 |
mov X1, 8(rp) |
| 415 |
adc $0, w3 |
| 416 |
add w2, %rdx |
| 417 |
adc $0, %rax |
| 418 |
add %rdx, w3 |
| 419 |
mov w3, 16(rp) |
| 420 |
adc $0, %rax |
| 421 |
mov %rax, 24(rp) |
| 422 |
|
| 423 |
jmp L(outer) C loop until a small corner remains |
| 424 |
|
| 425 |
L(corner): |
| 426 |
pop un |
| 427 |
mov (up), %rdx |
| 428 |
jg L(small_corner) |
| 429 |
|
| 430 |
mov %rdx, v1 |
| 431 |
mov (rp), X0 |
| 432 |
mov %rax, X1 C Tricky rax reuse of last iteration |
| 433 |
mulx( v0, %rax, w1) |
| 434 |
add %rax, X0 |
| 435 |
adc $0, w1 |
| 436 |
mov X0, (rp) |
| 437 |
mov 8(up), %rdx |
| 438 |
mulx( v0, %rax, w3) |
| 439 |
add %rax, X1 |
| 440 |
adc $0, w3 |
| 441 |
mulx( v1, %rdx, %rax) |
| 442 |
add w1, X1 |
| 443 |
mov X1, 8(rp) |
| 444 |
adc $0, w3 |
| 445 |
add w3, %rdx |
| 446 |
mov %rdx, 16(rp) |
| 447 |
adc $0, %rax |
| 448 |
mov %rax, 24(rp) |
| 449 |
lea 32(rp), rp |
| 450 |
lea 16(up), up |
| 451 |
jmp L(com) |
| 452 |
|
| 453 |
L(small_corner): |
| 454 |
mulx( v0, X1, w3) |
| 455 |
add %rax, X1 C Tricky rax reuse of last iteration |
| 456 |
adc $0, w3 |
| 457 |
mov X1, (rp) |
| 458 |
mov w3, 8(rp) |
| 459 |
lea 16(rp), rp |
| 460 |
lea 8(up), up |
| 461 |
|
| 462 |
L(com): |
| 463 |
|
| 464 |
L(sqr_diag_addlsh1): |
| 465 |
lea 8(up,un,8), up C put back up at its very beginning |
| 466 |
lea (rp,un,8), rp |
| 467 |
lea (rp,un,8), rp C put back rp at its very beginning |
| 468 |
inc un |
| 469 |
|
| 470 |
mov -8(up), %rdx |
| 471 |
xor R32(%rbx), R32(%rbx) C clear CF as side effect |
| 472 |
mulx( %rdx, %rax, %r10) |
| 473 |
mov %rax, 8(rp) |
| 474 |
mov 16(rp), %r8 |
| 475 |
mov 24(rp), %r9 |
| 476 |
jmp L(dm) |
| 477 |
|
| 478 |
ALIGN(16) |
| 479 |
L(dtop):mov 32(rp), %r8 |
| 480 |
mov 40(rp), %r9 |
| 481 |
lea 16(rp), rp |
| 482 |
lea (%rdx,%rbx), %r10 |
| 483 |
L(dm): adc %r8, %r8 |
| 484 |
adc %r9, %r9 |
| 485 |
setc R8(%rbx) |
| 486 |
mov (up), %rdx |
| 487 |
lea 8(up), up |
| 488 |
mulx( %rdx, %rax, %rdx) |
| 489 |
add %r10, %r8 |
| 490 |
adc %rax, %r9 |
| 491 |
mov %r8, 16(rp) |
| 492 |
mov %r9, 24(rp) |
| 493 |
inc un |
| 494 |
jnz L(dtop) |
| 495 |
|
| 496 |
L(dend):adc %rbx, %rdx |
| 497 |
mov %rdx, 32(rp) |
| 498 |
|
| 499 |
pop %r14 |
| 500 |
pop %r13 |
| 501 |
pop %r12 |
| 502 |
pop %rbp |
| 503 |
pop %rbx |
| 504 |
FUNC_EXIT() |
| 505 |
ret |
| 506 |
EPILOGUE() |