| 1 |
dnl AMD64 mpn_mul_basecase optimised for Intel Haswell. |
| 2 |
|
| 3 |
dnl Contributed to the GNU project by Torbjörn Granlund. |
| 4 |
|
| 5 |
dnl Copyright 2003-2005, 2007, 2008, 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_1 mul_2 mul_3 addmul_2 |
| 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 1.77 1.86 - 2.15 |
| 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. |
| 55 |
|
| 56 |
C TODO |
| 57 |
C * Adjoin a mul_3. |
| 58 |
C * Further micro-optimise. |
| 59 |
|
| 60 |
define(`rp', `%rdi') |
| 61 |
define(`up', `%rsi') |
| 62 |
define(`un_param',`%rdx') |
| 63 |
define(`vp', `%rcx') |
| 64 |
define(`vn', `%r8') |
| 65 |
|
| 66 |
define(`un', `%rbx') |
| 67 |
|
| 68 |
define(`w0', `%r10') |
| 69 |
define(`w1', `%r11') |
| 70 |
define(`w2', `%r12') |
| 71 |
define(`w3', `%r13') |
| 72 |
define(`n', `%rbp') |
| 73 |
define(`v0', `%r9') |
| 74 |
|
| 75 |
ABI_SUPPORT(DOS64) |
| 76 |
ABI_SUPPORT(STD64) |
| 77 |
|
| 78 |
ASM_START() |
| 79 |
TEXT |
| 80 |
ALIGN(16) |
| 81 |
PROLOGUE(mpn_mul_basecase) |
| 82 |
FUNC_ENTRY(4) |
| 83 |
IFDOS(` mov 56(%rsp), %r8d ') |
| 84 |
push %rbx |
| 85 |
push %rbp |
| 86 |
push %r12 |
| 87 |
push %r13 |
| 88 |
push %r14 |
| 89 |
mov un_param, un C free up rdx |
| 90 |
neg un |
| 91 |
|
| 92 |
mov un_param, n C FIXME: share |
| 93 |
sar $2, n C FIXME: share |
| 94 |
|
| 95 |
test $1, R8(vn) |
| 96 |
jz L(do_mul_2) |
| 97 |
|
| 98 |
define(`w4', `%r9') |
| 99 |
define(`w5', `%r14') |
| 100 |
|
| 101 |
mov (vp), %rdx |
| 102 |
|
| 103 |
L(do_mul_1): |
| 104 |
test $1, R8(un) |
| 105 |
jnz L(m1x1) |
| 106 |
|
| 107 |
L(m1x0):test $2, R8(un) |
| 108 |
jnz L(m110) |
| 109 |
|
| 110 |
L(m100): |
| 111 |
mulx( (up), w5, w2) |
| 112 |
mulx( 8,(up), w1, w3) |
| 113 |
lea -24(rp), rp |
| 114 |
jmp L(m1l0) |
| 115 |
|
| 116 |
L(m110): |
| 117 |
mulx( (up), w3, w4) |
| 118 |
mulx( 8,(up), w1, w5) |
| 119 |
lea -8(rp), rp |
| 120 |
test n, n |
| 121 |
jz L(cj2) |
| 122 |
mulx( 16,(up), w0, w2) |
| 123 |
lea 16(up), up |
| 124 |
jmp L(m1l2) |
| 125 |
|
| 126 |
L(m1x1):test $2, R8(un) |
| 127 |
jz L(m111) |
| 128 |
|
| 129 |
L(m101): |
| 130 |
mulx( (up), w4, w5) |
| 131 |
lea -16(rp), rp |
| 132 |
test n, n |
| 133 |
jz L(cj1) |
| 134 |
mulx( 8,(up), w0, w2) |
| 135 |
lea 8(up), up |
| 136 |
jmp L(m1l1) |
| 137 |
|
| 138 |
L(m111): |
| 139 |
mulx( (up), w2, w3) |
| 140 |
mulx( 8,(up), w0, w4) |
| 141 |
mulx( 16,(up), w1, w5) |
| 142 |
lea 24(up), up |
| 143 |
test n, n |
| 144 |
jnz L(gt3) |
| 145 |
add w0, w3 |
| 146 |
jmp L(cj3) |
| 147 |
L(gt3): add w0, w3 |
| 148 |
jmp L(m1l3) |
| 149 |
|
| 150 |
ALIGN(32) |
| 151 |
L(m1tp):lea 32(rp), rp |
| 152 |
L(m1l3):mov w2, (rp) |
| 153 |
mulx( (up), w0, w2) |
| 154 |
L(m1l2):mov w3, 8(rp) |
| 155 |
adc w1, w4 |
| 156 |
L(m1l1):adc w0, w5 |
| 157 |
mov w4, 16(rp) |
| 158 |
mulx( 8,(up), w1, w3) |
| 159 |
L(m1l0):mov w5, 24(rp) |
| 160 |
mulx( 16,(up), w0, w4) |
| 161 |
adc w1, w2 |
| 162 |
mulx( 24,(up), w1, w5) |
| 163 |
adc w0, w3 |
| 164 |
lea 32(up), up |
| 165 |
dec n |
| 166 |
jnz L(m1tp) |
| 167 |
|
| 168 |
L(m1ed):lea 32(rp), rp |
| 169 |
L(cj3): mov w2, (rp) |
| 170 |
L(cj2): mov w3, 8(rp) |
| 171 |
adc w1, w4 |
| 172 |
L(cj1): mov w4, 16(rp) |
| 173 |
adc $0, w5 |
| 174 |
mov w5, 24(rp) |
| 175 |
|
| 176 |
dec R32(vn) |
| 177 |
jz L(ret5) |
| 178 |
|
| 179 |
lea 8(vp), vp |
| 180 |
lea 32(rp), rp |
| 181 |
C push %r12 |
| 182 |
C push %r13 |
| 183 |
C push %r14 |
| 184 |
jmp L(do_addmul) |
| 185 |
|
| 186 |
L(do_mul_2): |
| 187 |
define(`v1', `%r14') |
| 188 |
C push %r12 |
| 189 |
C push %r13 |
| 190 |
C push %r14 |
| 191 |
|
| 192 |
mov (vp), v0 |
| 193 |
mov 8(vp), v1 |
| 194 |
|
| 195 |
lea (un), n |
| 196 |
sar $2, n |
| 197 |
|
| 198 |
test $1, R8(un) |
| 199 |
jnz L(m2x1) |
| 200 |
|
| 201 |
L(m2x0):xor w0, w0 |
| 202 |
test $2, R8(un) |
| 203 |
mov (up), %rdx |
| 204 |
mulx( v0, w2, w1) |
| 205 |
jz L(m2l0) |
| 206 |
|
| 207 |
L(m210):lea -16(rp), rp |
| 208 |
lea -16(up), up |
| 209 |
jmp L(m2l2) |
| 210 |
|
| 211 |
L(m2x1):xor w2, w2 |
| 212 |
test $2, R8(un) |
| 213 |
mov (up), %rdx |
| 214 |
mulx( v0, w0, w3) |
| 215 |
jz L(m211) |
| 216 |
|
| 217 |
L(m201):lea -24(rp), rp |
| 218 |
lea 8(up), up |
| 219 |
jmp L(m2l1) |
| 220 |
|
| 221 |
L(m211):lea -8(rp), rp |
| 222 |
lea -8(up), up |
| 223 |
jmp L(m2l3) |
| 224 |
|
| 225 |
ALIGN(16) |
| 226 |
L(m2tp):mulx( v1, %rax, w0) |
| 227 |
add %rax, w2 |
| 228 |
mov (up), %rdx |
| 229 |
mulx( v0, %rax, w1) |
| 230 |
adc $0, w0 |
| 231 |
add %rax, w2 |
| 232 |
adc $0, w1 |
| 233 |
add w3, w2 |
| 234 |
L(m2l0):mov w2, (rp) |
| 235 |
adc $0, w1 |
| 236 |
mulx( v1, %rax, w2) |
| 237 |
add %rax, w0 |
| 238 |
mov 8(up), %rdx |
| 239 |
adc $0, w2 |
| 240 |
mulx( v0, %rax, w3) |
| 241 |
add %rax, w0 |
| 242 |
adc $0, w3 |
| 243 |
add w1, w0 |
| 244 |
L(m2l3):mov w0, 8(rp) |
| 245 |
adc $0, w3 |
| 246 |
mulx( v1, %rax, w0) |
| 247 |
add %rax, w2 |
| 248 |
mov 16(up), %rdx |
| 249 |
mulx( v0, %rax, w1) |
| 250 |
adc $0, w0 |
| 251 |
add %rax, w2 |
| 252 |
adc $0, w1 |
| 253 |
add w3, w2 |
| 254 |
L(m2l2):mov w2, 16(rp) |
| 255 |
adc $0, w1 |
| 256 |
mulx( v1, %rax, w2) |
| 257 |
add %rax, w0 |
| 258 |
mov 24(up), %rdx |
| 259 |
adc $0, w2 |
| 260 |
mulx( v0, %rax, w3) |
| 261 |
add %rax, w0 |
| 262 |
adc $0, w3 |
| 263 |
add w1, w0 |
| 264 |
lea 32(up), up |
| 265 |
L(m2l1):mov w0, 24(rp) |
| 266 |
adc $0, w3 |
| 267 |
inc n |
| 268 |
lea 32(rp), rp |
| 269 |
jnz L(m2tp) |
| 270 |
|
| 271 |
L(m2ed):mulx( v1, %rdx, %rax) |
| 272 |
add %rdx, w2 |
| 273 |
adc $0, %rax |
| 274 |
add w3, w2 |
| 275 |
mov w2, (rp) |
| 276 |
adc $0, %rax |
| 277 |
mov %rax, 8(rp) |
| 278 |
|
| 279 |
add $-2, R32(vn) |
| 280 |
jz L(ret5) |
| 281 |
lea 16(vp), vp |
| 282 |
lea 16(rp), rp |
| 283 |
|
| 284 |
|
| 285 |
L(do_addmul): |
| 286 |
push %r15 |
| 287 |
push vn C save vn in new stack slot |
| 288 |
define(`vn', `(%rsp)') |
| 289 |
define(`X0', `%r14') |
| 290 |
define(`X1', `%r15') |
| 291 |
define(`v1', `%r8') |
| 292 |
|
| 293 |
lea (rp,un,8), rp |
| 294 |
lea (up,un,8), up |
| 295 |
|
| 296 |
L(outer): |
| 297 |
mov (vp), v0 |
| 298 |
mov 8(vp), v1 |
| 299 |
|
| 300 |
lea 2(un), n |
| 301 |
sar $2, n |
| 302 |
|
| 303 |
mov (up), %rdx |
| 304 |
test $1, R8(un) |
| 305 |
jnz L(bx1) |
| 306 |
|
| 307 |
L(bx0): mov (rp), X0 |
| 308 |
mov 8(rp), X1 |
| 309 |
mulx( v0, %rax, w1) |
| 310 |
add %rax, X0 |
| 311 |
mulx( v1, %rax, w2) |
| 312 |
adc $0, w1 |
| 313 |
mov X0, (rp) |
| 314 |
add %rax, X1 |
| 315 |
adc $0, w2 |
| 316 |
mov 8(up), %rdx |
| 317 |
test $2, R8(un) |
| 318 |
jnz L(b10) |
| 319 |
|
| 320 |
L(b00): lea 16(up), up |
| 321 |
lea 16(rp), rp |
| 322 |
jmp L(lo0) |
| 323 |
|
| 324 |
L(b10): mov 16(rp), X0 |
| 325 |
lea 32(up), up |
| 326 |
mulx( v0, %rax, w3) |
| 327 |
jmp L(lo2) |
| 328 |
|
| 329 |
L(bx1): mov (rp), X1 |
| 330 |
mov 8(rp), X0 |
| 331 |
mulx( v0, %rax, w3) |
| 332 |
add %rax, X1 |
| 333 |
adc $0, w3 |
| 334 |
mulx( v1, %rax, w0) |
| 335 |
add %rax, X0 |
| 336 |
adc $0, w0 |
| 337 |
mov 8(up), %rdx |
| 338 |
mov X1, (rp) |
| 339 |
mulx( v0, %rax, w1) |
| 340 |
test $2, R8(un) |
| 341 |
jz L(b11) |
| 342 |
|
| 343 |
L(b01): mov 16(rp), X1 |
| 344 |
lea 24(rp), rp |
| 345 |
lea 24(up), up |
| 346 |
jmp L(lo1) |
| 347 |
|
| 348 |
L(b11): lea 8(rp), rp |
| 349 |
lea 8(up), up |
| 350 |
jmp L(lo3) |
| 351 |
|
| 352 |
ALIGN(16) |
| 353 |
L(top): mulx( v0, %rax, w3) |
| 354 |
add w0, X1 |
| 355 |
adc $0, w2 |
| 356 |
L(lo2): add %rax, X1 |
| 357 |
adc $0, w3 |
| 358 |
mulx( v1, %rax, w0) |
| 359 |
add %rax, X0 |
| 360 |
adc $0, w0 |
| 361 |
lea 32(rp), rp |
| 362 |
add w1, X1 |
| 363 |
mov -16(up), %rdx |
| 364 |
mov X1, -24(rp) |
| 365 |
adc $0, w3 |
| 366 |
add w2, X0 |
| 367 |
mov -8(rp), X1 |
| 368 |
mulx( v0, %rax, w1) |
| 369 |
adc $0, w0 |
| 370 |
L(lo1): add %rax, X0 |
| 371 |
mulx( v1, %rax, w2) |
| 372 |
adc $0, w1 |
| 373 |
add w3, X0 |
| 374 |
mov X0, -16(rp) |
| 375 |
adc $0, w1 |
| 376 |
add %rax, X1 |
| 377 |
adc $0, w2 |
| 378 |
add w0, X1 |
| 379 |
mov -8(up), %rdx |
| 380 |
adc $0, w2 |
| 381 |
L(lo0): mulx( v0, %rax, w3) |
| 382 |
add %rax, X1 |
| 383 |
adc $0, w3 |
| 384 |
mov (rp), X0 |
| 385 |
mulx( v1, %rax, w0) |
| 386 |
add %rax, X0 |
| 387 |
adc $0, w0 |
| 388 |
add w1, X1 |
| 389 |
mov X1, -8(rp) |
| 390 |
adc $0, w3 |
| 391 |
mov (up), %rdx |
| 392 |
add w2, X0 |
| 393 |
mulx( v0, %rax, w1) |
| 394 |
adc $0, w0 |
| 395 |
L(lo3): add %rax, X0 |
| 396 |
adc $0, w1 |
| 397 |
mulx( v1, %rax, w2) |
| 398 |
add w3, X0 |
| 399 |
mov 8(rp), X1 |
| 400 |
mov X0, (rp) |
| 401 |
mov 16(rp), X0 |
| 402 |
adc $0, w1 |
| 403 |
add %rax, X1 |
| 404 |
adc $0, w2 |
| 405 |
mov 8(up), %rdx |
| 406 |
lea 32(up), up |
| 407 |
inc n |
| 408 |
jnz L(top) |
| 409 |
|
| 410 |
L(end): mulx( v0, %rax, w3) |
| 411 |
add w0, X1 |
| 412 |
adc $0, w2 |
| 413 |
add %rax, X1 |
| 414 |
adc $0, w3 |
| 415 |
mulx( v1, %rdx, %rax) |
| 416 |
add w1, X1 |
| 417 |
mov X1, 8(rp) |
| 418 |
adc $0, w3 |
| 419 |
add w2, %rdx |
| 420 |
adc $0, %rax |
| 421 |
add w3, %rdx |
| 422 |
mov %rdx, 16(rp) |
| 423 |
adc $0, %rax |
| 424 |
mov %rax, 24(rp) |
| 425 |
|
| 426 |
addl $-2, vn |
| 427 |
lea 16(vp), vp |
| 428 |
lea -16(up,un,8), up |
| 429 |
lea 32(rp,un,8), rp |
| 430 |
jnz L(outer) |
| 431 |
|
| 432 |
pop %rax C deallocate vn slot |
| 433 |
pop %r15 |
| 434 |
L(ret5):pop %r14 |
| 435 |
L(ret4):pop %r13 |
| 436 |
L(ret3):pop %r12 |
| 437 |
L(ret2):pop %rbp |
| 438 |
pop %rbx |
| 439 |
FUNC_EXIT() |
| 440 |
ret |
| 441 |
EPILOGUE() |